Machine learning with Plane Wave Imaging Tuomas Koskinen, Topias Tyystjärvi, Oskari Jessen-Juhler and Iikka Virkkunen Trueflaw Ltd Espoo, Uusimaa, 02330, Finland +358 50 563 85855 tuomas.koskinen@trueflaw.com Abstract Phased array ultrasonics have enabled the recording of ever-increasing amount of data from the inspection targets. With the latest advancements in total-focusing method with plane wave imaging the amount of data has increased exponentially when compared to conventional ultrasonic methods. As more data allows more reliable evaluation, the cost of evaluation also increases. Since there is more data for the inspector to evaluate, the inspector’s job becomes more difficult and laboursome with the modern technology. Moreover, as phased array techniques evolve to even more sophisticated approaches such as total focusing method and the latest form, plane wave imaging total focusing method (PWI-TFM) reading raw ultrasonic data is too convoluted for human inspectors. While the idea is to show pre-calculated image to the inspector, the data allows multiple different ways of presentation the data to the inspector, even though these representations are not normally used. Machine learning powered inspection enables the full use of all the data, while allowing the best possible presentation for the inspector. In this paper we demonstrate PWI-TFM inspection powered by machine learning model. The ML model is used to scan the data and present the flaw indications to the inspector. 1. Introduction The NDT industry is undergoing a significant digital transformation, with a growing emphasis on digital interfaces, storage, and offline analysis. In the realm of ultrasound inspection, this shift has led to the need for manual analysis of a substantially larger volume of data within a reasonable timeframe. Consequently, inspectors often find themselves merging multiple refraction angles to create a single image, with the danger of worsening the data quality or losing flaw indications altogether. Recent studies have also found that inspectors might unintentionally disregard some recorded data from the extreme angles, most like due to human error [1], [2]. Primarily because most of the data lacks any detectable indications, it is hard to keep focus constantly and only rarely actually findings something. In an attempt to save inspectors’ time on data analysis, the full potential of these advanced technologies is frequently underutilized. As a result, while these innovations offer increased sensitivity, they also introduce challenges and added 2 labour for inspectors. Which may be one of the reasons these new approaches are not actively adopted. Increasing computing power and digitalization in the ultrasonic inspection field has enabled significant advancements compared to single channel A-scans, which are still widely used today. While phased array, and more refined approach such as full-matrix- capture (FMC) and total-focusing method (TFM) [3] have been available for many years, their adoption to the inspections has been slow. The main benefit in TFM technique is the high image quality due to spatial resolution and the total focus in the region of interest. Moreover, same array response may have various different imaging modes[4]. These imaging modes can include half-skips and mode conversions to facilitate data analysis and enhance the image. The reconstruction in TFM also allows the image to be wider than the original probe, as the image is reconstructed based on the echoes received from the elements and calculated based on the transmitting element [5]. However, TFM has some significant drawbacks which have possibly hindered the wider adoption. Firstly, TFM has low acoustic power as only one element is transmitting ultrasound to the target. Especially for attenuative materials such as austenitic stainless-steel welds, the signal-to-noise ratio (SNR) has been poor. In addition, weld inspection is more difficult since the acoustic power cannot be directed at an angle, but it propagates as a spherical wave to the crack causing artefacts, which may even lead to false calls [4], [6]. Second drawback is that TFM produces a lot of data and processing and storing of this data becomes a bottle neck, especially using larger probes with many elements [5]. The acoustic power can be increased using Sparse Matrix Capture (SMC), which also reduces the number of processed signals, reducing the data processing burden [7], [8]. However, the aperture is not ideal for beam steering for weld inspections and the technique is still lacking the proper acoustic power for highly attenuative materials. Plane Wave Imaging (PWI) [9], [10], recently developed in medical field and later tested for NDT applications [5], [11] seems promising to fix the shortcomings of traditional TFM and SMC. PWI is combination of the best practices from traditional phased array and TFM. In PWI, plane ultrasonic wave-fronts are transmitted through the medium at various angles. These angles are then reconstructed and focused to different depths incrementally. The angled transmission is possible due to utilization of the whole element array instead of a single element as in traditional TFM, furthermore all the elements are used in the receiving as well to further enhance the data output [5], [11]. These images are then summed together to form a final image. This leads to high image quality with significantly lesser number of ultrasonic shots than for traditional TFM [5], [11]. Moreover, as all the elements can be used in transmitting, the aperture stays large allowing further beam steering and keeping the acoustic power as high as possible. As modern techniques allow more novel data evaluation, modern inspection equipment allows recording of data in significantly larger quantities, such as increased number of angles and higher resolution with minor effect to data recording time. Moreover, extensive and high bandwidth data output streams data outward more than human 3 inspector could evaluate in days. Thus, edge computing and machine learning (ML) has the potential to help the inspector with modern techniques and high data output. The immense potential of ML lies in its ability to harness the vast reservoir of available data and present the outcomes in a lucid and responsive manner. We observe that ML has already made significant inroads into various industries, powering applications like speech recognition [12] and self-driving cars [13] in real-time. However, within the NDT industry, ML solutions constitute only a fraction of the overall landscape and even scarcer real-time applications. Despite numerous academic examples showcasing the integration of ML in ultrasonic inspection [14]–[17] its widespread adoption by the industry remains in its infancy. In this paper, demonstrate manual inspection ready concept with plane-wave imaging total-focusing-method (PWI-TFM) with wavelet transform filtering and real-time ML annotation. The demonstration is for austenitic base material with two different channels using 5 MHz 64 element linear phased array probe with a rexolite wedge. 2. Experimental setup The theoretical background for PWI is well explained in [5]. In traditional TFM the time- of-flight from the transmitter with a spherical wave trough the target to the receiver is calculated one transmitter element at a time. Whereas, in PWI a set of 𝑄 plane waves at 𝑄 angles are transmitted and the back scattered signals for every element 𝑁 is recorded. This creates data matrix 𝑄 × 𝑁. An algorithm was proposed in [5] which allows focus in every point within the Region of Interest (ROI). In addition, this allows to take into account the half-skip modes for crack like flaw indications, reconstruct the image outside the probe aperture and take into account the mode conversions [5]. The major difference between the TFM and PWI is that in PWI the wave propagation angle is known in PWI. This requires significantly less operations than TFM algorithm, thus computational burden is more manageable. [5] The amplitude A in a pre-determined point P in the imaging area is calculated for the set of Q angles is obtained with the equation below. 𝐴(𝑃) = |∑ ∑ 𝑠𝑞𝑗(𝑡𝑞 𝑝 + 𝑡𝑗 𝑝) 𝑁 𝑗=1 𝑄 𝑞=1 | Where 𝑡𝑞 𝑝 is the time-of-flight of the plane wave at angle q to reach the point P. 𝑡𝑗 𝑝 is the time from the focusing point P to the receiving element j. N is the number of elements used in reception. 𝑠𝑞𝑗(𝑡) are transformed using Hilbert’s transform and summed together. As the angle q of the transmitted wave is known the time-of-flight calculation is a straight forward process. However, as the backscattered wave is cylindrical reception point has to be calculated using the traditional TFM approach with the Newton-Rhaphson algorithm. [5] Since plane waves are spatially limited, an effective area can be calculated from the transmission angle. This effective area can be used to apply the PWI algorithm only to 4 the points within the effective area. This helps to mitigate the effects of diffraction by the transducer edges or grating lobes [5]. Figure 1, shows the effective area at an angle and the image area in the specimen. As previous papers[5], [11] have used Hilbert’s transformation our reconstruction approached the algorithm without any transformation and with the wavelet transformation. In wavelet transformation the transformation happens on the time axle and does not affect the shape of the signal. The transformation is done with the help of a basis function, a “mother-wavelet” [18], [19]. This basis function has many options; however, Morlet wavelet seems to represent the ultrasonic signal typically the closest. Figure 1. Effective area smaller than the imaging area, which decreases as the angle increases. 2.1 Hardware setup and test specimens In the experiment we used similar setup as in [11]. 64 element linear phased array with 5 MHz central frequency. 19.56° rexolite wedge, provided by Dekra Industrial Finland and 64 element phased array ultrasonic device Peak LTPA. The probe was moved manually without any encoder and water was used as a coupling medium. The data was fed from the Peak ultrasound to edge-computing unit TrueflawBox for data reconstruction and ML annotation. To train and test the machine learning model 20 mm thick austenitic steel plates were used. Thermal fatigue cracks with varying sizes were manufactured to the steel plates. 2.2 Ultrasound setup and data reconstruction Three plane waves were used in transmission and reconstructing the image. Two longitudinal waves with 45° and 65° angles and a shear wave with 33° angle. The image area was set underneath the probe and slightly forward from the front of the probe. Figure 2 displays the imaging area and the channels used in the scan. 5 Figure 2. Longitudinal waves 45°, 65°, shear waves 33° and the image area displayed. The reconstruction was made similar way as stated before. However, the summation of the amplitudes was done without the Hilbert’s transform. As Hilbert’s transform loses the frequency component, we dictated the data is better to present to the ML model without this transform. As the images kept relatively high signal-to-noise ratio (SNR) without this transform, we decided to leave it out from the image to be shown for the inspector as well. As there were only two separate angles and a total of three different channels used, we decided that instead of summing these three channels together the channels would be presented with different colours to the inspector in a single image. Moreover, there has been multiple occasions that the set angles do not match between the theory and reality, causing unnecessary artefacts, this representation would prevent such behaviour and produce clearer image. 2.3 Machine learning In the plates there were 3 – 5 mm deep thermal fatigue cracks. In total of 3 cracks were used in training and two were left for testing of the model. While this is obviously not nearly enough flaw samples for proper ML model training it worked well as a demonstrator. The architecture was U-net type model, with the use of standard data augmentation, however, no virtual flaws were used in training. As the training data was originally too scarce the addition of virtual flaws would probably not have increased the model performance. As stated before, the channel data was not Hilbert transformed and were presented as three separate channels for the ML model. This would assure high data quality for the model, furthermore there would be no difficulties from artefacts caused by summation of the three separate channels together due to misalignment between the theoretical and real angles. Originally the input data was then reconstructed as stated before and 1024 × 512 resolution images were created for the model to analyse. While the results and performance were sufficient for the said resolution. The resolution should be decreased to allow more channels and faster throughput in the future. Thus, the resolution then would be reduced down to 256 × 128 with wavelet transformation. Wavelet transform proved to be effective and lossless method of decreasing the resolution, while preserving the waveform information as well. As a mother wavelet, Morlet wavelet was chosen with the same frequency as the probe. However, as the central frequency of the probe was 5 6 Mhz, especially for the welds it was detected that around 2 Mhz frequency would reach back to the probe due to material characteristics. 3. Results and discussion The results from the uncompressed 1024 × 512 resolution performed well, as the thermal fatigue crack could be detected and the edge caused no false calls. Figure 3 demonstrates the reconstructed response from three different channels, without transformation and the channels displayed in different colours. In Figure 4 ML performance is demonstrated with the crack and with the edge. Figure 3. Combined view of the three different plane wave channels. 1) The crack indication. 2) Mode conversion from 65° longitudinal wave. 3) Mode conversion from 45° longitudinal wave. 7 a) b) Figure 4. Combined view with ML annotation. ML highlights the crack indication in a) but doesn’t highlight the mode conversions as intended. b) shows the corner, which the ML model does not annotate. The performance reached was 20 frames per second (FPS), as the resolution was fairly large for the ML model. The 1024 × 512 resolution as a single channel would cause no computational burden, but with the three separate channels the on-line annotation was capped at 20 FPS. While this is sufficient for this purpose, as higher framerate is not required to evaluate the dynamics of the signal, performance could decrease when larger number of different channels are added. Furthermore, adding more channels means more possibilities for the ML model to detect the defects thus, adding more channels should be beneficial for reliability. To decrease the resolution, wavelet transformation was used. While the wavelet transformation produces fairly similar outcome as the Hilbert’s transform, the benefits for machine learning are more advantageous. Firstly, the wavelet transformation can be used as a convolutional layer, which means with GPU acceleration this operation is extremely fast. Secondly, the wavelet transform preserves the frequency information unlike the Hilbert’s transform allowing more data to be sent to ML model. Moreover, the convolution area of the signal can be chosen more freely saving computational time. Using the wavelet transform, the final resolution could be reduced to 256 × 128 which saves the memory considerably compared to the original state. However, this did not directly translate to final FPS, since more optimizing should be performed on the other areas as well. 8 a) b) c) Figure 5. Wavelet transformed data. a) The same data as in Figure 3, without ML annotation. b) ML annotated crack indication. c) The corner echo, which the ML model does not annotate. Figure 5 demonstrates the wavelet transformed images. While the resolution is 16 folds smaller than the original, the crack is still well visible. Although some noise seems to be highlighted with the wavelet, this is more of a probe issue rather than transformation issue. As the test sample set was too small to draw any performance related conclusions with the ML model, the performance stayed similar to the full resolution image. This indicates that the approach is viable to reduce the memory load and allow ever more channels and data to be fed for the ML model for higher reliability analysis. 4. Conclusions PWI-TFM operates well to identify crack like indications from austenitic material. The method works well without any transformations as only summation of the signals produces clear enough image for detection and image reconstruction. However, wavelet transform can be used as a viable option to Hilbert’s transformation while decreasing image resolution. The wavelet transformation preserves data integrity and increases computational efficiency. This allows faster and real-time interface with the manual inspection and enables further ML integration to the inspection procedure. Moreover, we suspect that wavelet transform has the opportunity to be used in multiple different ways due to its nature, such as noise reduction, tough it is yet to be explored. As multiple channels can be recorded with the modern techniques these should be fed separately to the ML model for the optimal performance. This reduces the dangers of artefacts and presents the best possible data to the ML model for to train and to evaluate. The final output on the other hand should be combined as a single image for easier viewing for the inspector while still maintaining information about the origin of the signals. Thus, the inspection procedure should be designed such a way that ML model gets the best and broadest amount of data, while the inspector is enable to view the best possible data for flaw evaluation trough annotation and filtering of the data. 9 Lastly, edge computing has the opportunity to increase the adoptability and versatility of ultrasonic inspection considerably. This enables similar approach to use inspection procedures more like an application rather than a piece of paper. This mindset has the opportunity to faster adopt newer techniques in the field, as also to reduce human errors and the required expertise to utilize these novel approaches. Acknowledgements We thank Peak NDT, Dekra Industrial Finland and VTT Technical Research Centre of Finland Ltd. 10 References [1] I. Virkkunen, T. Koskinen, and O. Jessen-Juhler, “Virtual round robin – A new opportunity to study NDT reliability,” Nuclear Engineering and Design, vol. 380, Aug. 2021, doi: 10.1016/j.nucengdes.2021.111297. [2] I. Vikkunen, T. Koskinen, and Oskar Siljama, “Virtual round robin 2 – phased array inspection of dissimilar metal welds,” NDT&E in review. [3] C. Holmes, B. W. Drinkwater, and P. D. Wilcox, “Post-processing of the full matrix of ultrasonic transmit–receive array data for non-destructive evaluation,” NDT & E International, vol. 38, no. 8, pp. 701–711, 2005, doi: https://doi.org/10.1016/j.ndteint.2005.04.002. [4] E. Iakovleva, S. Chatillon, P. Bredif, and S. Mahaut, “Multi-mode TFM imaging with artifacts filtering using CIVA UT forwards models,” AIP Conf Proc, vol. 1581, no. 1, pp. 72–79, Feb. 2014, doi: 10.1063/1.4864804. [5] L. Le Jeune, S. Robert, E. Lopez Villaverde, and C. Prada, “Plane Wave Imaging for ultrasonic non-destructive testing: Generalization to multimodal imaging,” Ultrasonics, vol. 64, pp. 128–138, Jan. 2016, doi: 10.1016/j.ultras.2015.08.008. [6] N. Portzgen, D. Gisolf, and D. J. Verschuur, “Wave equation-based imaging of mode converted waves in ultrasonic NDI, with suppressed leakage from nonmode converted waves,” IEEE Trans Ultrason Ferroelectr Freq Control, vol. 55, no. 8, pp. 1768–1780, 2008, doi: 10.1109/TUFFC.2008.861. [7] S. Bannouf, S. Robert, O. Casula, and C. Prada, “Data set reduction for ultrasonic TFM imaging using the effective aperture approach and virtual sources,” J Phys Conf Ser, vol. 457, no. 1, p. 12007, Aug. 2013, doi: 10.1088/1742- 6596/457/1/012007. [8] G. R. Lockwood, P.-C. Li, M. O’Donnell, and F. S. Foster, “Optimizing the radiation pattern of sparse periodic linear arrays,” IEEE Trans Ultrason Ferroelectr Freq Control, vol. 43, no. 1, pp. 7–14, 1996, doi: 10.1109/58.484457. [9] A. Austeng, C.-I. C. Nilsen, A. C. Jensen, S. P. Näsholm, and S. Holm, “Coherent plane-wave compounding and minimum variance beamforming,” in 2011 IEEE International Ultrasonics Symposium, 2011, pp. 2448–2451. doi: 10.1109/ULTSYM.2011.0608. [10] G. Montaldo, M. Tanter, J. Bercoff, N. Benech, and M. Fink, “Coherent plane- wave compounding for very high frame rate ultrasonography and transient elastography,” IEEE Trans Ultrason Ferroelectr Freq Control, vol. 56, no. 3, pp. 489–506, 2009, doi: 10.1109/TUFFC.2009.1067. [11] L. S. S. Pillarisetti, G. Raju, and A. Subramanian, “Sectorial Plane Wave Imaging for Ultrasonic Array-Based Angle Beam Inspection,” J Nondestr Eval, vol. 40, no. 3, Sep. 2021, doi: 10.1007/s10921-021-00813-6. [12] F. Mihelic and J. Zibert, Speech Recognition. Rijeka: IntechOpen, 2008. doi: 10.5772/93. [13] S. Ranjan and S. Senthamilarasu, Applied Deep Learning and Computer Vision for Self-Driving Cars: Build autonomous vehicles using deep neural networks and behavior-cloning techniques. Packt Publishing Ltd, 2020. [14] I. Virkkunen, T. Koskinen, O. Jessen-Juhler, and J. Rinta-aho, “Augmented Ultrasonic Data for Machine Learning,” J Nondestr Eval, vol. 40, p. 4, 2021, doi: 10.1007/s10921-020-00739-5. 11 [15] O. Siljama, T. Koskinen, · Oskari Jessen-Juhler, and · Iikka Virkkunen, “Automated Flaw Detection in Multi-channel Phased Array Ultrasonic Data Using Machine Learning,” J Nondestr Eval, vol. 40, p. 67, 2021, doi: 10.1007/s10921- 021-00796-4. [16] N. Munir, J. Park, H. J. Kim, S. J. Song, and S. S. Kang, “Performance enhancement of convolutional neural network for ultrasonic flaw classification by adopting autoencoder,” NDT and E International, vol. 111, Apr. 2020, doi: 10.1016/j.ndteint.2020.102218. [17] N. Munir, H.-J. Kim, S.-J. Song, and S.-S. Kang, “Investigation of deep neural network with drop out for ultrasonic flaw classification in weldments,” Journal of Mechanical Science and Technology, vol. 32, no. 7, pp. 3073–3080, 2018, doi: 10.1007/s12206-018-0610-1. [18] Y. Meyer, Wavelets and Operators. Cambridge University Press, 1992. [19] C. K. Chui, An Introduction to Wavelets. Academic Press, 1992.

Journal of Nondestructive Evaluation (2021) 40:24 https://doi.org/10.1007/s10921-021-00757-x The Effect of Different Flaw Data to Machine Learning Powered Ultrasonic Inspection Tuomas Koskinen1,2 · Iikka Virkkunen2 ·Oskar Siljama2 ·Oskari Jessen-Juhler1 Received: 3 August 2020 / Accepted: 30 January 2021 / Published online: 18 February 2021 © The Author(s) 2021 Abstract Previous research (Li et al., Understanding the disharmony between dropout and batch normalization by variance shift. CoRR abs/1801.05134 (2018). http://arxiv.org/abs/1801.05134 arXiv:1801.05134) has shown the plausibility of using amodern deep convolutional neural network to detect flaws from phased-array ultrasonic data. This brings the repeatability and effectiveness of automated systems to complex ultrasonic signal evaluation, previously done exclusively by human inspectors. The major breakthrough was to use virtual flaws to generate ample flaw data for the teaching of the algorithm. This enabled the use of raw ultrasonic scan data for detection and to leverage some of the approaches used in machine learning for image recognition. Unlike traditional image recognition, training data for ultrasonic inspection is scarce. While virtual flaws allow us to broaden the data considerably, original flaws with proper flaw-size distribution are still required. This is of course the same for training human inspectors. The training of human inspectors is usually done with easily manufacturable flaws such as side-drilled holes and EDM notches. While the difference between these easily manufactured artificial flaws and real flaws is obvious, human inspectors still manage to train with them and perform well in real inspection scenarios. In the present work, we use a modern, deep convolutional neural network to detect flaws from phased-array ultrasonic data and compare the results achieved from different training data obtained from various artificial flaws. The model demonstrated good generalization capability toward flaw sizes larger than the original training data, and the effect of the minimum flaw size in the data set affects the a90/95 value. This work also demonstrates how different artificial flaws, solidification cracks, EDM notch and simple simulated flaws generalize differently. Keywords NDT · Ultrasonic testing · Machine Learning · Image classification 1 Introduction Ultrasonic inspectors are commonly trained using simple artificial flaws, such as EDM notches and side-drilled holes. These two types offer a quick and cost-effective way of demonstrating where the flaw indication should appear, but their signal shape differs from a real service-induced crack, like amechanical or thermal fatigue crack. Inspectors can use reasoning to estimate real reflectors based on these simpli- fied signals. However, more than simple artificial flaws are B Tuomas Koskinen tuomas.koskinen@vtt.fi Iikka Virkkunen iikka.virkkunen@aalto.fi 1 VTT Technical Research Centre of Finland Ltd, VTT, P.O. Box 1000, 02044 Espoo, Finland 2 Aalto University, P.O. Box 11000, 00076 Aalto, Finland usually required for qualification of a technique, for exam- ple in nuclear power plants [7], to confirm performance in a representative setting. Difficulties arise when the inspection material is noisy and the inspector needs to use expert judgement to distin- guish flaws from structural noise. For example, an EDM notch might be found much more easily than a thermal fatigue crack in a dissimilar metal weld (DMW) inspection. While their signals can be distinguished from each other, a human inspector is not only looking for a specific reflector or thermal fatigue crack but also for an explanation for any unusual reflector. Therefore, while training and conducting the inspections, the inspector focuses on learning and detect- ingwhere the flaw indicationsmay appear and how they stand out compared to the surrounding noise. A human inspector can intuitively ignore possible artefacts in the artificial flaws and still successfully find real flaws in the inspection data. 123 24 Page 2 of 13 Journal of Nondestructive Evaluation (2021) 40 :24 For machine learning, the task is muchmore difficult. Due to the training process, the model can learn any and all fea- tures related to the training data; thus, the teaching data set determines the boundaries of the capability of the algorithm. This learning method is useful when the task is to determine specific features from images with high accuracy and the training data are largely available. For ultrasonic inspection, this is a problem since training data are not readily available and the detection probability for the algorithm needs to be high, while still avoiding false calls. The model may learn incidental features of the training data, i.e. it may overfit to the training flaws and fail to generalize to unseen flaw indica- tions. Conversely, underfitting may cause an excessive false call rate. Therefore, this paper aims to study how training data from different sources can be used to train ML algo- rithms to detect other flaw types and how the minimum flaw size in the training set affects the a90/95 value. 1.1 Effect of Different Kinds of Flaws and Artificial Flaws The flaw response for ultrasonic testing is highly related to the kind of reflector fromwhich the soundwaves are reflected back to the transducer. The characteristics that primarily affect the flaw response are the location and orientation of the crack, size of the crack, opening of the crack through the whole path and at the crack tip, fracture surface roughness and filling of the crack with a substance. An in-depth study of the flaw responses and crack characteristics has already been conducted by [9,10]. For the most representative flaw response signal, it is reasonable to assume that these charac- teristics should be met in order to achieve the best possible training data for machine learning as well. These character- istics are the main reason real cracks are preferred over EDM notches and side-drilled holes when conducting actual per- formance demonstrations for human inspectors. In addition, it is assumed that the larger the crack, the easier it is to detect. This should apply for the machine learning model as well. As the larger cracks are more critical, these types of cracks should be reliably found. 1.2 Teaching and Generalizing theMachine Learning Model Since humans can use their theoretical reasoning and target their focus on the relevant part of the data, it is possible (to a certain extent) to use simple flaws to teach and train human inspectors to find real flaws in inspection cases. Machine learning models lack this theoretical reasoning and imag- ination, and the training data must explicitly provide the variation that the models need to learn. The training data itself has a strong influence on training and generalizing the model. First of all, it is imperative to have enough flaw data for teaching. Secondly, the labelling of the data needs special attention to non-destructive testing (NDT). Labelling small flaws that are indistinguishable from the noise may cause the model to overfit on noise features and/or result in an excessive false call rate. Lastly, themodels can converge in training, even in the absence of generalizable features in the training data, as demonstrated by [26]. Overfitting can be mitigated by several approaches. The obvious first choice is to increase the amount of training and validation data. This ample data amount is seldom available for ultrasonic testing. The second choice for decreasing over- fitting is data augmentation when teaching data are scarce. Traditional data augmentation, where the image is rotated, reflected, scaled, cropped or translated, are common prac- tices to artificially increase the amount of available data [3,6]. These methods have been used successfully in NDT and ultrasonic inspection by [25]. For a weld scan, however, rota- tion of the flaw might be out of the question as cracks can form in a certain place and certain orientation for in-service inspections. Data augmentation through virtual flaws pre- sented in [23] has shown great promise as it allows scaling the flaws to represent smaller flaws and changing the location along the weld, allowing a larger variety of backgrounds for the flaw to reside in. In general, NDT data can be considered simple, thus there exist options for generating data other than virtual flaws. Alternative approaches for generating training data sets have been used in eddy current testing by [15] to generate an ample amount of data with an adaptive generation technique known as Output Space Filling (OSF) with an efficient computation time. Reference [1] used a similar approach by adding the Partial Least Squares (PLS) feature extraction to OSF and trained several machine learning models with this generated data. Reference [1] stated that this data generation method might be feasible for ultrasonic and thermographic testing. Further generalization can be obtained by tuning the hyperparameters. Batch size, for example, has a strong influ- ence on learning. Reference [14] showed that generally, the best generalization performance is achieved with a batch size of 2 to 32 and up to 64 with a batch normalization layer. The downside of using small batch sizes is that it slows down the teaching of the model. Thus, large batch sizes are preferred. Instead of further modifying or augmenting the teaching data or decreasing the performance of the model, there are also possibilities to affect the training of the model as well. Dropout is one of the most common approaches. Dropout works by zeroing out a certain amount of the layer’s out- put values at random during training. The number of values dropped out is determined by the dropout rate, which is usually between 10 and 50% of the layer’s output values. Essentially, this means that random variation is introduced to the output, and less significant features that are only present for the training data are valued less or cancelled out from the 123 Journal of Nondestructive Evaluation (2021) 40 :24 Page 3 of 13 24 finalmodel, thus leading to amore generalized representation of the task. During testing, the output values are allowed to work fully, but scaled down with the dropout rate to compen- sate all working output values [6,24]. Dropout has been used in ultrasonic inspection by [16] with successful results as the performance increase was significant compared to the neural network without the dropout for the A-signal classification. As overfitting is one of the major problems in teach- ing modern deep learning models, the difference between human image recognition and machine image recognition needs to be understood as well. Unlike a human inspector, a machine learning model does not know the actual concept of its task, since that is determined through the teaching data. Even though the data might look good enough for teaching humans and estimating the probability of detection (POD) curves [12,21], the data might contain artefacts from the arti- ficial flaw manufacturing process or poorly designed virtual flaw generation. This kind of teaching, i.e., with poor data, is demonstrated in [17], where distinguishing between wolves and huskies was based on the feature that wolves had snow in the background in the training data set. The effect of the poor data is the same for NDT.When the teaching data would have some feature such as an artefact from implantation, the flaw detection in themodel might focus on the implantation rather than the actual flaw characteristics. Due to these reasons, it is crucial for the NDT model to be actually tested with flaws where these kinds of artefacts do not exist or to map which features affect the decision the most with methods such as or similar to grad-CAM by [18] and LIME by [17]. Generalization of the model can be increased through adding the batch normalization layer introduced in [8]. Batch normalization drives to remove the covariate shift from the internal activations within the network. This has the effect of faster learning rates and increased accuracy.Asbatchnormal- ization works to generalize the model, it decreases the need for a dropout layer in some cases. In fact, the performance of the model might decrease drastically if a dropout and a batch normalization layers are used together. Reference [13] recommend the use of a dropout layer after all batch normal- ization layers on large data sets. On the other hand, Reference [5] reported a decrease in accuracy when both layers were used together. In general, it is recommended to use a batch normalization in the models first and then carefully observe the effect of an added dropout layer for the best possible result. Therefore, the main problem is teaching a model with too little real flaw data, while still keeping generalization to real flaws that the model has never seen before and still main- taining at least human-level performance. As the flaw data is scarce in NDT, virtual flaws present a way to mitigate the problem. However, the more diverse the data, the better, even with the virtual flaws. Hence, simulating the flaw responses for trainingmight be plausible to broaden the data efficiently. Simulated flaws have been previously used together with virtual flaw augmentation to calculate POD by [11]. While humans did not detect the difference between simulated and real flaw responses, the research showed that simulated flaws were slightly easier to detect. Thus, it might be assumed that an ML model could be able to adequately generalize to real flaws. 2 Materials andMethods Inspection data was gathered from scanning a DMW mock- up and generating flaw responses from CIVA simulations. The location of the flaws was the same for all flaw types, on the edge of the buffer zone, 7 to 10 mm to the carbon steel from the weld center. The scanned flaws were aug- mented with Trueflaw’s eFlaw [21] software, and data sets for machine learning purposes were created. 2.1 Scanned Samples For initial inspection data, a DMW pipe mock-up from SwedishQualificationCentreAB (SQC)was used. The spec- imenwas 32mm thickwith an outer diameter of 898mm.The specimen had implanted flaws and an EDM notch as defects. The original sample consisted of two “small” solidification flaws 2 mm and 3 mm in size. Two large solidification flaws, ofwhich 17mmwas tilted toward the carbon steel side and 26 mmwas straight oriented. There were two 6 mm sized flaws, an EDM notch and a solidification flaw. In total, six differ- ent flaws were available for training. In addition, the sample consisted of three axial solidification flaws with heights of 6, 17 and 26 mm and one axial EDM notch with a height of 6 mm. The flaw scanning was optimized for circumferential flaws; thus, the axial flaw indications were removed from the teaching and testing data sets with the eFlaw process. The inspection procedure was an optimized version of Zetec Inc.’s procedure C3467 Zetec OmniscanPA 03 Rev A. The inspection equipment that was used was Dynaray Lite with two Imasonic 1.5 MHz 32 element phased array probes in a wedge with a 7◦ roof angle set-up for TRL acquisition. The coupling was applied through a feed water system. In order to minimize data, only one scan line was utilized, with a 60◦ angle. The focal lawwas focused on the inner surface of the pipe and the probe positioned such that the best amplitude response from the flaws was achieved. Data recording was done at a 16-bit depth for best possible data quality. The schematic of the inspection procedure can be seen in Fig. 1. The scanned flaws were augmented with eFlaw software by scaling down the recorded amplitude, thus representing a wider size range of flaws similarly as in [22]. 123 24 Page 4 of 13 Journal of Nondestructive Evaluation (2021) 40 :24 CSSS Flaw area in cladding Fig. 1 Diagram of the inspection setup. TRL probe was situated on the carbon steel side of the test mock-up and focused on inner diameter of the pipe. The original flaws were situated on the cladding marked with black stripes on the image. Due to export control restrictions, the exact details of the test block’s dimensions, materials and weld cannot be made public 2.2 Simulation Set-up The same set-up was created in CIVA2019 simulation soft- ware with different-sized notches. For signal generation, the Hanning type was used, and for flaw response calculation, the Kirchhoff and GTD model was used as per CIVA guid- ance [4,20] in a similar simulation case, which is optimal for simulating reflection and diffraction echoes from crack like flaws. Unlike the paper by [20], the weld was modelled with orthotropic anisotropy. The buttering layer of the DMWwas modelledwith a polycrystalline cubic structure, with an aver- age grain size of 1.5µm to represent the simple simulation case. In order to reduce calculation time, only the flaw and the immediate surroundings of the flaw were simulated. The resolution of the simulation was aimed to be the same as for the scanned samples’ 2 mm scan step and 103.2 mm sound path. In total, six different-sized notches were simulated; at heights of 1 through 6 mm, the width of the flaw was three times the height. Just as with the scanned plate samples, the flaw responses were extracted from the simulation data and implemented in the pipe mock-up scan through the eFlaw software and more flaws were generated by scaling down the recorded amplitude from the simulated flaws to a total of 10,000 simulated flaws generated by the eFlaw augmen- tation. Figure 2a demonstrates the original simulated B-scan image from CIVA and Fig. 2b the pre-processed B-scan image for a better comparison. The raw simulated signal was used when implanting the flaw onto the scanned image with eFlaw and pre-processed for model training. Figure 2c shows the pre-processed simulated flaw image show to the model and Fig. 2d pre-processed scanned EDM notch implanted with eFlaw. The width of the simulated flaw matches well with the scanned one. Along the sound path, the simulated flaw is slightly longer; and after post-processing, the simu- lated flaw looks denser than the scanned EDM notch. As Fig. 2 demonstrates, the scanned 6 mm EDM notch and the simulated 6 mm EDM notch look different. This is because the aim was to use a simple simulation setup in CIVA to set a base-line for teaching data. The size of the flaw is accurate, and the implanted signal is plausible for human eye as well due to the accurate modelling of the flaw and model geometry. However, a closer simulation could be achieved with increased accuracy in the material and anisotropy parameters of the DMW as well as increased detail in the simulated signal representing the used probe more accurately as only the frequency was matched to rep- resent the scanned signal. 2.3 Training Data and Used Data Augmentation Reference [23] used only thermal fatigue flaws as scan input; thus, it is proven that it is viable to use thermal fatigue flaws as teachingmaterial to find thermal fatigue flaws. For this paper, we generated several different teaching data sets, where cer- tain flaw types were only shown during testing to investigate how well the model detects the completely new flaw type. In order to generate sufficient trainingdata from the six dif- ferent scanned flaws and six different simulated flaws, eFlaw software was used to augment the flaw locations and sizes within the training and testing data. These virtual flaws have been previously used successfully in training humans and evaluating POD by [12,19,21]. The indications of the six dif- ferent scanned flaws and simulated flawswere scaled down to represent smaller sizes up to 40% of the original indication. This allowed the generation of 7000 different variations for the scanned data to be used as training, validation and testing data with roughly 50% containing flaws and 50% without flaws. In addition, great care was taken to prevent the model from learning the virtual flaw introduction process by copy- ing and replacing the unflawed data as well within the set. A total of 10,000 images were created from the simulated flaws with the same method as for the scanned flaws. The raw RF signal was pre-processed by fully rectify- ing the signal to an absolute positive value. The scan data was processed for more efficient teaching purposes, thus the sound path was narrowed down to 2000 samples to repre- sent the inner diameter of the mock-up where the flaws were located. Theflaw image contained 480 scan steps in total. The B-scan dimension of 480× 2000 proved to be too slow to han- dle, as the whole data set could not fit into the GPU memory at the same time. In order to reduce the data set size, without losing information from the sound path, the original B-scan 123 Journal of Nondestructive Evaluation (2021) 40 :24 Page 5 of 13 24 Fig. 2 Comparison of the simulated 6 mm EDM notch signal and scanned 6 mm EDM notch implanted trough the eFlaw software. a Raw simulated RF signal with sound path of 2000 samples, b raw simulated signal post processed with max-pooling and rectified to absolute positive value, c simulated flaw implanted to the weld b-scan with eFlaw and d scanned 6 mm EDM notch implanted to the weld B-scan with eFlaw for comparison. The simulated flaw seems to be slightly longer and denser along the sound path and more symmetric than the reference scanned EDM notch was pre-processed by max-pooling the sound path with 1 4λ. This provided original data in the size of 480 × 118. To further optimize the image for machine learning, the image was normalized according to [3]. The image was reduced by the mean value and divided by the standard deviation. If the image was labelled as flawed, one flaw would be introduced in the image through the eFlaw process at a random loca- tion along the weld. Since only one weld was scanned as the background canvas, it later showed that themodel learned the weld pattern when it was shown the whole 480-sample-wide weld in a single training image. This led the model to overfit on the weld rather than detecting the actual flaw indications. This is clearly the wrong target as this would work only if the actual inspected weld would provide an exactly identi- cal weld image as that recorded from the mock-up, which is impossible. This wasmitigated by cropping the image area in half. Once the image size was 240 samples wide, it allowed the generation of images at multiple locations along the weld and maintaining a generalization on the clean weld, as the background kept changing. This meant that the model was shown a “new” clean weld with no flaws as for the flawed samples as well; thus, the initial image data was 240 × 118 samples in size. For a proper comparison to the previouslymentionedVRR data, themodel was adjusted to handle 48× 118 sized images to determine the proper location from the data. As the image was smaller than the original teaching data, the sound path was further cropped to 112 samples. This allowed randomly moving the crop window along the sound path for 6 samples, increasing the different backgrounds for training data. The variability of the images was further increased by a similar data augmentation used by [25]. The image was randomly flipped from left to right during training using the built in function from the Tensorflow package, but not rotated or scaled. To further validate that the taught model would not overfit, the images with no flaws were shown only 90% of the weld area. During testing the model would see the whole weld. The model was trained with the following flaw type combinations from (a) through (f), shown inTable 1 for solid- ification cracks and an EDM notch. The model taught with the simulated flaws was run with two different types of com- binations, (i) and (j) in Table 2. The tables show the amount of flaws available for training. The total number of images is doubled when the images without flaws is added to the data set. 20% of this said data set would be selected as the valida- tion set. In addition, the model was taught with only 6 mm solidification crack (g) consisting of 558 flaw images and only a 6 mm EDM notch (h) consisting of 599 flaw images not included in the tables. 2.4 UsedML Architecture A more refined deep neural network model was constructed based on [23]. To further enhance the accuracy, the dimen- sions of the latter convolutional layers and the dense layer were increased, and themax pooling layer with the batch nor- malization layer was added after each convolutional layer for increased generalization and overfitting reduction. The model architecture can be seen in Fig. 3. The optimized net- 123 24 Page 6 of 13 Journal of Nondestructive Evaluation (2021) 40 :24 Table 1 Flaw sets for training and flaw images Size Set a) All flaws b) Small c) Medium d) Large e) No larges f) No smalls Flaws total: 3442 1090 1157 1195 2247 2352 2 mm 3 mm 6 mm 6 mm (EDM) 17 mm 26 mm Table 2 Simulated flaw sets for training and flaw images Size Set i) All j) No smalls Flaws total: 4982 3360 1 mm 2 mm 3 mm 4 mm 5 mm 6 mm work structure was a result of trial and error by adjusting the dimensions of each layer. Dropout was left out of the model as batch normalization proved to be sufficient and using both dropout and batch normalization together seemed to yield variability in the results. The model was taught with the training data variations presented in Sect. 2.3. 2.5 Performance Evaluation POD and false calls were used to measure the performance of the model. The POD curve was a hit/miss POD calculated according to regular standard MIL-HDBK-1823a [2]. POD is a valid way to measure the performance of the model, since it is used for evaluating the performance of humans as well. In addition, this enables comparison between the model result and humanVRRdata, since theywere evaluated with the same data and standard. If the model were overly sensitive, it would show as false calls in the evaluation or if the model would overfit and constantly miss flaw types never seen before, this would easily be seen in the POD curve as erratic behaviour. The performance evaluationwas divided into two data sets from the virtual flaws described in Sect. 2.3. The first test data set would contain 4700 to 7000 samples, depending onwhich flaw set from Table 1 was used for training. Only the flaw types that were not used in training would be shown to the model to evaluate the generalization capability. The second data set would contain 1000 samples with all the available flaw types. Even though the same flaw types used in training are expected to be more easily found, they do not have an effect on finding the flaw types never shown to the model in training and are in the data set only to avoid flaw size gaps in the POD evaluation. 123 Journal of Nondestructive Evaluation (2021) 40 :24 Page 7 of 13 24 Fig. 3 Used optimized network structure Input shape: 48, 112, 1 Convolution 1 output: 48, 112, 96 Maxpool 1 output: 24, 28, 96 Batch norm 1 output: 24, 28, 96 Convolution 2 output: 24, 28, 80 Max pool 2 output: 12, 14, 80 Batch norm 2 output: 12, 14, 80 Convolution 3 output: 12, 14, 74 Max pool 3 output: 6, 7, 74 Batch norm 3 output: 6, 7, 74 Convolution 4 output: 6, 7, 30 Max pool 4 output: 1, 1, 30 Batch norm 4 output: 1, 1, 30 Flatten output: 30 Dense output: 30 Logits output: 1 After testing with the data set, which contained images from random locations of the DMW with or without a flaw, the model was shown the same ultrasonic weld image as a human would see in a traditional inspection. This was done by dividing the whole weld image to 48-sample-wide images with the location coordinate as metadata. The images would be shown to themodel, and themodelwould evaluatewhether or not the location contains a flaw. In case of a hit, the image centre line would be highlighted in green on the weld image. 3 Results The results have been divided into two sections: testing the generalization capability of the model; and comparison to human performance with similar ultrasonic data. 3.1 Testing the Generalization For POD calculations, the data was adjusted similarly as in [22] when there were no missed flaws; 0 to 0.2 mm sized misses were added to the calculation. When the data faced zero separation, a missed flaw was added with a size 0.1 mm larger than the smallest flaw found. These adjustments needed to be done for the POD calculation to converge some of the results. However, this has little effect on the final POD. The predictions for the differently trained models have been plotted in Fig. 4. The model was trained with the flaw combinations (a)–(j) described in Sect. 2.3 and in Tables 1 and 2. The tested flaws have not been shown to the said model before. This enables testing how well the model is capable of generalization when trained with different flaws and tested with completely different flaws. For POD hit/miss evaluation, all the indications scoring higher than 50% was considered as hits, and false calls when no flaw was in the data. If the prediction was less than 50% and flaw existed in the data, it was marked as a miss. The PODwas tested with a data set containing all the flaw types and 1000 samples. The PODs for different models can be found in Fig. 5, except for the simulated flaws, as the performance was so unreliable due to false calls that it was not comparable to other models. For cases (a), (b) and (e) in Fig. 5, adding zeros from size 0 to 0.2 mm and zero separation management have been used due to low or no misses. (a) and (e) gave no POD before the adjustment and (b) changed to a more conservative a90/95 value of 1.05 mm from 0.75 mm. As expected, the model trained with all the available flaw types provided perfect results with no missed flaws and zero false calls. This test was done to set the benchmark for other training data sets. When training with only the smallest flaws, the model generalizes well on the larger flaws. Predictions for the flaws can be seen in Fig. 4b. There is a slight deviation for the pre- dictions for the EDM notch and 17 mm solidification crack, which was slightly tilted compared to the 6 mm and 26 mm solidification cracks, both of which yielded almost perfect predictions but also had with three reported false calls. The POD for the model trained with only small flaws can be seen in Fig. 5b for which no misses on the 17 mm flaws were reported while few of the smaller flaws were missed. The POD is the same as for the case with all available flaws (a), since the smallest flaw available for training is the same. When the model was trained with the 6 mm solidification crack and the 6 mm EDM notch, the model managed to gen- eralize well on the larger flaws, while generalization towards the smaller solidification cracks was not impressive. As seen in Fig. 4c the model missed one 17 mm solidification crack for the larger testing set but found all with the smaller test set size for POD. This is well within the a90/95 limit, as the larger test set contained over 500 samples of virtually augmented 17 mm solidification cracks, which were not shown to the model during the training of the (c) set. The augmentation for this flaw group was from 2.4 to 6 mm, and the a90/95 value was 3.45 mm, which can be seen in Fig. 5c. The results for the model trained with only large flaws yielded similar results in Fig. 4d and the POD in Fig. 5d. Through the augmentation process, the smallest flaw for trainingwas 6.8mm. Themodel did poorly in finding smaller flaws, with the exception of the 6 mm solidification cracks, for which every flaw was found. This also has a decreasing effect on the POD and a90/95 values. 123 24 Page 8 of 13 Journal of Nondestructive Evaluation (2021) 40 :24 Fig. 4 Predictions vs. flaw size when testing with unseen flaws, with exception to a. Threshold for detection was set to 50% a all flaws used for training. All flaws found with no false calls. b Only small flaws, 2 and 3 mm used for training. All flaws found, 3 false calls. c Only medium flaws, 6 mm solidification crack and EDM notch used for training. One miss on 17 mm flaw, no false calls, poor performance on smaller flaws. d Only large flaws, 17 and 26 mm used for training. Only 6 mm solidification cracks found reliably, no false calls. e Large flaws, 17 mm and 26 mm removed for training. Some of the 17 mm flaws missed, no false calls. f Small flaws, 2 and 3 mm removed for training. Smaller flaws are not found with high consistency, no false calls. g Only 6 mm solidification crack used for training. Only largest flaws are found reliably, consistent misses on 6 mm EDM notch, no false calls. h Only 6 mm EDM notch used for training. Generalizes well on larger flaws and 6 mm solidification crack, missing constantly smaller flaws, no false calls. i Trained with all simulated flaws. 131 False calls, missing constantly every flaw type. j Small simulated flaws removed. 78 Calls, misses from every flaw type, slightly better performance compared to i When training without the large flaws, the model general- izeswellwith the larger flaws. Predictions for the flaws can be seen in Fig. 4e. There is a slight deviation for the predictions for the 17 mm solidification crack, as similarly seen when training only with the medium sized flaw in Fig. 4c, com- pared to 26 mm solidification cracks, which yielded perfect predictions. Again, the testing set for never before seen flaws was significantly larger than the testing set for PODmeasure- ment, which contained all the flaw types. Thus, missing 6 of the 17 mm flaws in a testing set containing over 500 flaws is within statistical limits. The POD for the model trained with- out the large flaws can be seen in Fig. 5e where no misses on 17 mm flaws were reported, which meant that the model made no misses or false calls. When the small flaws were excluded from training, the result was the same as when training with only the medium sized flaws in Fig. 4c. Predictions can be seen in Fig. 4f and the POD in Fig. 5f. 123 Journal of Nondestructive Evaluation (2021) 40 :24 Page 9 of 13 24 Fig. 5 POD when testing with all flaw types. a All flaws used for training. All flaws found with no false calls. 0 to 0.2 mm sized misses added for POD convergence. b Only small flaws, 2 and 3 mm used for training. Two 2 mm sized flaws missed with no false calls. 0 to 0.2 mm misses added for more conservative POD. c Only medium flaws, 6 mm solidification crack and EDM notch used for training. Smallest flaw size in training set was 2.4 mm and a90/95 was 3.45 mm. d Only large flaws, 17 and 26 mm used for training. As the 6 mm solidification cracks are found reliably, it improves the POD whereas the 6 mm EDM notches are mostly missed. e Large flaws, 17 mm and 26 mm removed from the training set. No false calls and 0 to 0.2 mm sized missed added for convergence. f Small flaws, 2 and 3 mm removed from training set. Smaller flaws are not found with high consistency, same a90/95 result as in c since smallest flaw size was the same. g Only 6 mm solidification crack used for training. Only largest flaws are found reliably, consistent misses on 6 mm EDM notch, the worst a90/95. h Only 6 mm EDM notch used for training. Generalizes well on larger flaws and 6 mm solidification crack, missing smaller flaws, but performs better than c and f with a90/95 of 2.55 When using only one flaw type, the training set tends to get dangerously small.When themodel was trainedwith only a 6 mm solidification crack, the result deteriorates considerably. All the smaller cracks are missed, EDM notches are barely detected and the model struggles to detect the larger 17 mm and 26 mm cracks without false calls. The predictions can be seen in Fig. 4g and the POD in Fig. 5g. However, the model trained with only a 6 mm EDM notch proved to perform well compared to the same-sized solidification crack. The predictions of the model trained with only a 6 mm EDM notch can be seen in Fig. 4h and the POD in Fig. 5h. There is a slight deviation for the 17mmsolidification crack, but 6mm and 26 mm are found consistently. The majority of the 3 mm solidification cracks are found, while the 2 mm solidification crack tends to go unnoticed, which has an improving effect on the POD. When training with only the EDM notch, the model achieves the best a90/95 value when there are no small flaws included in the training set. The results when training with all the simulated flaws can be seen in Fig. 4i and results for the training with simulated 123 24 Page 10 of 13 Journal of Nondestructive Evaluation (2021) 40 :24 flaws without the smaller flaws in Fig. 4j. A generalization to real defects proved to be not possible for the simple simu- lated flaw. Even though the flaw sizes ranged from 1 to 6mm, there were no indications of improvement, as false calls were an issue when tested with real flaws. The most concerning observationwas that,while themodelwas capable of general- ization to larger flaws fromsmaller flaws, thiswas not the case for the simulated flaws. The model kept constantly missing some of the largest flaws for both training data sets. The only difference was that when the smallest flaws were removed from the data set, performance increased by decreasing the false calls and a number ofmissed flaws. Unfortunately, there were still inconsistent misses on the large flaws as well. 3.2 Testing with FullWeld B-Scan and Comparison with Human Performance Instead of the previous training and testing, the ultrasonic B- scan image was divided into 48-sample-wide windows with a step of one sample and shown to the model in consecutive order moving from left to right. The centre line of the image would be highlighted as green if prediction would exceed over 50%. The results for VRR data performance can be seen in Fig. 6 for three different training methods: (a) model trained without smallest and largest flaws, (b) model trained without the 3 mm and 17 mm flaws and (c) model trained with only the largest flaws left in the training data. The grey colored prediction shows the centre line of themodelwindow with size 48 × 112 samples (window width 96 mm). When the model was presented with similar ultrasonic data as for humans, Sample 8 from [22], the model kept its performance well and acted predictably. When the model came across an edge of the flaw and moved forward it kept detecting it until the window had completely moved past the flaw. The exception can be seen in (a) and (b) in Fig. 6, where the 13.6 mm flaw is detected only by the edges of the flaw and not at all in the middle of the flaw. This might be due because when the window is on top of the large flaw and the training set has not consisted of enough large flaws over the size of the inspection window, the model cannot detect the flaw in those areas. In Fig. 6c, where the training data set has contained only the large flaws, the model is capable of detecting the large flaws in the middle as well, granted the detection is easy as the model has been taught with the same flaw type. When considering human performance, all human inspec- tors found all the large flaws, but the two 1.6 mm and 2.4 mm flaws were missed by a couple of inspectors. Those misses might have been caused by interpreting the flaw as not large enough to indicate a flaw. Model (a), which had no 2 mm and 26 mm flaw types in training, barely found the 1.6 mm flaws. Case (b) in Fig. 6 which was trained without the 3 mm and 17 mm flaw types got the perfect score on the small flaws. Case (c), which was trained with only the large flaw types, found the 3 mm flaw type and the second 2 mm flaw type easily while having difficulties with the other 2 mm flaw types. One of the reasons for these detections is that the same flaw is presented to the model multiple times as the window moves over it. Thus, there are more opportunities for finding the right features for detection compared to the test in Sect. 3.1, where detection was based only on a sin- gle attempt. This explains why the 1.6 mm flaw gets detected partially formodels (a) and (c). This shows that themodel has the potential for human-level performance, as these smaller flaws had not been shown to the model before. 4 Discussion Detection accuracy seems to be highly related to the smallest flaw size used in training. While the model is capable of finding larger flaws than it is used to train with, the detection probability decreases once the tested flaws start to be smaller than those used in the training data set. This is good for qualification purposes, as it can be shown that the model generalizes better in finding the larger flaws consistently, as they are also the most critical ones to be found. In addition, the model’s accuracy can be adjusted by using the flaw size range required for the task. While the flaws available for trainingwere limited, certain observations regarding the flaw type could be made. When themodel was trained with an EDMnotch within the training set or justwith theEDMnotch available, the generalization of the model was better than with the solidification cracks only. Themost drastic effect could be observed when training with only the 6 mm solidification crack and 6 mm EDM notch in Fig. 4g and h, respectively. The comparison of these two flaw responses in Fig. 7 shows that the flaws look completely dif- ferent. Also, the model trained with the EDM notch could find the solidification cracks. Whereas the model taught with only the 6 mm solidification crack struggled to find simi- lar cracks and kept constantly missing the same-sized EDM notch. This indicates that the model might have learned fea- tures related to the solidification crack, not the pure crack indication. While the data set size was small for the single flaw types, the two types performed completely differently with the same data size. This could also be observed when the model was trained with only the larger flaws in Fig. 4d, where the detection of the 6 mm EDM notch is clearly lower and separated from the 6 mm solidification crack, which was detected with high reliability. When comparing the performance to the whole weld image from the VRR data, the model showed consistent per- formance with similar results to the initial testing. The major observation was the performance drop for the larger 17 mm flaw, especially when the model was taught with the largest 123 Journal of Nondestructive Evaluation (2021) 40 :24 Page 11 of 13 24 Fig. 6 VRR test B-scan image, flaw locations are 220, 350, 590, 720, 820 and 880 mm along the scan axis and at a depth of 1100 samples on the sound path. Flaw sizes are 20.8 mm, 4.8 mm, 13.6 mm, 1.6 mm, 2.4 mm, 1.6 mm virtually augmented from 26, 6, 17, 2, 3 and 2 mm flaws respectively. Flaw predictions are highlighted as grey area. a Model trained without the smallest 2 mm and largest 26 mm flaw. Detection of the 17 mm flaw is unreliable on the middle of the flaw but more certain on the edges, 2 mm flaw at 720 mm is barely detected. b Model trained without the 3 mm and 17 mm flaws. Detection of 17 mm solidification crack is purely based on detecting from the edges. cModel trained with only the largest 17 and 26 mm flaws. Detection of the 17 mm solidification crack is reliable, but one of the two 2 mm flaws at 720 mm seems to be difficult to find Fig. 7 Comparison between 6mmEDMnotch (above) and 6mm solid- ification crack (below). The solidification crack clearly has two peaks whereas the EDM notch has a clear single peak flaws completely left out of the training set. While the large flaw was found, the detection relied mostly on detecting the flaw from its edges. This indicates that when the flaw is large enough (i.e.,wider than theobservationwindow), themodel’s performance decreases drastically, as it has not experienced a similar situation when training with only small flaws. There- fore, it is highly beneficial to have larger flaws in training to compensate for this performance decrease. This observa- tion is consistent with the results in Sect. 3.1 where some of the 17 mm flaws were unexpectedly missed when the model was taught without the large flaws. In addition, the 17 mm solidification crack was slightly tilted, thus it gave a slightly different flaw indication than the 26 mm solidification crack in Table 1. The effect of the smallest flaw in the data set could be seen with all real flaws. When small 2 and 3 mm flaws were included in the data set, the a90/95 value was 1.05 mm. This result might be overly optimistic, as the flaw type was the same for testing as for training, while the flaws themselves were new through the virtual flaws.Withmedium6mmflaws as the smallest flaws, the a90/95 rose to 3.45 mm and to 2.55 mm with the EDM notch. The better performance for the lone EDM notch can be explained for better generalization and focus on the real indication discussed above. When the smallest original flawwas 17mm, the a90/95 rose to 7.65mm. The decrease in performance can be explained bymissing the EDM-type flaws in large number. However, there is a clear link to the a90/95 value and the smallest available flaw for training. 123 24 Page 12 of 13 Journal of Nondestructive Evaluation (2021) 40 :24 The number of false calls seemed to increase as the small- est flaw size in the data set decreased. For these data sets and flaw types in Fig. 4a, b, the threshold for the increased number of false calls seemed to be when only the small 2 and 3 mm flaws were used, resulting in 3 false calls and a rise in prediction values where flaws did not exist. However, these flaws might have been still too large and clearly visible for a proper threshold determination. The effect of the small flaw size is seen more clearly with the simulated flaws, as the false calls are decreased by 40%when the simulated flaw sizes 1 and 2 mm are excluded from the training set. These said flaws were deemed undetectable by the human eye as well. Even though the simulated data did not provide reliable results compared to the real flaws, it needs to be noted that the simulation of the DMW case was largely simplified. It may be plausible to enhance the performance by simulating the subject in more detail, thus decreasing the false calls and improving generalization if the simulated flaw response represents the subject in greater detail. 5 Conclusions Modern deep learning models have proven highly efficient and reliable in image recognition tasks. It is clear that the same approach can be used for NDT applications such as ultrasonic inspection. However, as these models extract the features related to detection on their own, great care needs to be taken when designing a data set for training a machine learning model for ultrasonic inspection: • Smallest flaw size detected is related to the smallest flaw size available in the training data set. • Flaw types may generalize differently, e.g. solidifica- tion cracks generalized worse to EDM notches than vice versa. • Using small flaws that are nearly undetectable in training may lead to deteriorated model performance. Acknowledgements This paper was funded by the Finnish Radiation Safety Program SAFIR2022 and VTT Technical Research Centre of FinlandLtd. Trueflawcontributed the eFlaw augmentation, and the orig- inal weld sample was provided by SQC through the PIONIC program. All contributions are greatfully acknowledged. Funding Open access funding provided by Technical Research Centre of Finland (VTT). Data Availibility The training data set is made available for download at https://github.com/koomas/NDT_ML_Flaw. Compliance with Ethical Standards Conflict of interest The authors declare that they have no conflict of interest. Virkkunen is associated with Trueflaw Ltd., who supplied the eFlaw augmentation. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adap- tation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indi- cate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, youwill need to obtain permission directly from the copy- right holder. To view a copy of this licence, visit http://creativecomm ons.org/licenses/by/4.0/. References 1. Ahmed, S., Reboud, C., Lhuillier, P.E., Calmon, P., Miorelli, R.: An adaptive sampling strategy for quasi real time crack character- ization on eddy current testing signals. NDT E Int. 103, 154–165 (2019). https://doi.org/10.1016/j.ndteint.2019.02.001 2. Annis, C.: MILl-HDBK-1823a, Nondestructive Evaluation Sys- tem Reliability Assessment. Technical report. Department of Defence (2009). http://www.statisticalengineering.com/mh1823/ MIL-HDBK-1823A(2009).pdf 3. Chollet, F.: Deep Learning with Python, 1st edn. Manning Publi- cations Co., Greenwich (2017) 4. Extende: CIVA NDE 2017 User Manual (2017) 5. Garbin, C., Zhum, X., Marques, O.: Dropout vs. batch normaliza- tion: an empirical study of their impact to deep learning.Multimed. Tools Appl. (2020). https://doi.org/10.1007/s11042-019-08453-9 6. Hinton, G.E., Srivastava, N., Krizhevsky, A., Sutskever, I., Salakhutdinov R.: Improving neural networks by preventing co- adaptation of feature detectors. CoRR. abs/1207.0580 (2012). http://arxiv.org/abs/1207.0580 arXiv:1207.0580 7. Inspecta Sertifiointi Oy Finnish Methodology for Qualification of PSI/ISI NDT-Inspection Systems According to STUK YVL E.5 Scheme, 3rd issue (2019) 8. Ioffe, S., Szegedy, C.: Batch normalization: accelerating deep network training by reducing internal covariate shift. CoRR abs/1502.03167 (2015). http://arxiv.org/abs/1502.03167 arXiv:1502.03167 9. Kemppainen, M., Virkkunen, I.: Crack characteristics and their importance to NDE. J. Nondestruct. Eval. 30(3), 143–157 (2011). https://doi.org/10.1007/s10921-011-0102-z 10. Koskinen,A., Leskelä, E.: Differences in defect indications of three artificially produced defects in ultrasonic inspection. In: Baltica IX. International Conference on Life Management and Maintenance for Power Plants, No. 106, 11 July 2013 to 13 July 2013, pp 581– 602. VTT Technical Research Centre of Finland, Finland, VTT Technology (2013) 11. Koskinen, T., Virkkunen, I.: Hit/miss pod with model assisted and emulated flaws. In: 12th European Conference on Non-destructive Testing (ECNDT 2018), NDT.net, No. 8. e-J. Nondestruct. Test. (2018) 123 Journal of Nondestructive Evaluation (2021) 40 :24 Page 13 of 13 24 12. Koskinen, T.,Virkkunen, I., Papula, S., Sarikka, T.,Haapalainen, J.: Producing a pod curve with emulated signal response data. Insight 60(1), 42–48 (2018). https://doi.org/10.1784/insi.2018.60.1.42 13. Li, X., Chen, S., Hu, X., Yang, J.: Understanding the dishar- mony between dropout and batch normalization by variance shift. CoRR abs/1801.05134 (2018). http://arxiv.org/abs/1801.05134 arXiv:1801.05134 14. Masters, D., Luschi, C.: Revisiting small batch training for deep neural networks. CoRR abs/1804.07612 (2018). http://arxiv.org/ abs/1804.07612 arXiv:1804.07612 15. Miorelli, R., Artusi, X., Reboud, C.: An efficient adaptive database sampling strategy with applications to eddy current signals. Simul. Model. Pract. Theory 80, 75–88 (2018). https://doi.org/10.1016/j. simpat.2017.10.003 16. Munir, N., Kim, H.J., Song, S.J., Kang, S.S.: Investigation of deep neural network with drop out for ultrasonic flaw classification in weldments. J. Mech. Sci. Technol. 32(7), 3073–3080 (2018). https://doi.org/10.1007/s12206-018-0610-1 17. Ribeiro, M.T., Singh, S., Guestrin, C.: Why should I trust you? Explaining the predictions of any classifier. CoRR abs/1602.04938 (2016). http://arxiv.org/abs/1602.04938 arXiv:1602.04938 18. Selvaraju, R.R., Das, A., Vedantam, R., Cogswell, M., Parikh, D., Batra, D.: Grad-CAM: why did you say that? Visual expla- nations from deep networks via gradient-based localization. CoRR abs/1610.02391 (2016). http://arxiv.org/abs/1610.02391 arXiv:1610.02391 19. Svahn, P.H., Virkkunen, I., Zettervall, T., Snögren, D.: The use of virtual flaws to increase flexibility of qualification. In: 12th European Conference on Non-destructive Testing (ECNDT 2018), NDT.net, No. 8. e-J. Nondestruct. Test. (2018) 20. Szvai, S., Bzi, Z., Dudra, J., Mhsz, I.: Modelling of phased array ultrasonic inspection of a steam generator dissimilar metal weld. Procedia Struct. Integr. 2, 1015–1022 (2016). In: 21st European Conference on Fracture, ECF21, 20–24 June 2016, Catania, Italy. https://doi.org/10.1016/j.prostr.2016.06.130 21. Virkkunen, I., Miettinen, K., Packalén, T.: Virtual flaws for NDE training and qualification. In: 11th European Conference on Non- destructive Testing (ECNDT 2014), NDT.net. e-J. Nondestruct. Test. (2014) 22. Virkkunen, I., Koskinen, T., Jessen-Juhler, O.: Virtual round robin—a new opportunity to study NDT reliability (Submitted for review, 2020) 23. Virkkunen, I., Koskinen, T., Jessen-Juhler, O., Rinta-Aho, J.: Aug- mented ultrasonic data for machine learning. J. Nondestruct. Eval. (2021). https://doi.org/10.1007/s10921-020-00739-5 24. Wu, H., Gu, X.: Towards dropout training for convolutional neural networks. Neural Netw. 71, 1–10 (2015). https://doi.org/10.1016/ j.neunet.2015.07.007 25. Ye, J., Ito, S., Toyama, N.: Computerized ultrasonic imag- ing inspection: from shallow to deep learning. Sensors 18 (2018). https://pubmed.ncbi.nlm.nih.gov/30405086, https://www. ncbi.nlm.nih.gov/pmc/articles/PMC6263978/ 26. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning requires rethinking generalization. CoRR abs/1611.03530 (2016). http://arxiv.org/abs/1611.03530 arXiv:1611.03530 Publisher’s Note Springer Nature remains neutral with regard to juris- dictional claims in published maps and institutional affiliations. 123

Augmented Ultrasonic Data for Machine Learning Virkkunen, I.a, Koskinen, T.b, Jessen-Juhler, O.b, Rinta-aho, J.b aAalto University bVTT Technical Research Centre of Finland Ltd Abstract Flaw detection in non-destructive testing, especially in complex signals like ultrasonic data, has thus far relied heavily on the expertise and judgement of trained human inspectors. While automated systems have been used for a long time, these have mostly been limited to using simple decision automation, such as signal amplitude threshold. The recent advances in various machine learning algorithms have solved many similarly difficult classification problems, that have previously been considered intractable. For non-destructive testing, encouraging results have already been reported in the open literature, but the use of machine learning is still very limited in NDT applications in the field. Key issue hindering their use, is the limited availability of representative flawed data-sets to be used for training. In the present paper, we develop modern, very deep convolutional network to detect flaws from phased-array ultrasonic data. We make extensive use of data augmentation to enhance the initially limited raw data and to aid learning. The data augmentation utilizes virtual flaws - a technique, that has successfully been used in training human inspectors and is soon to be used in nuclear inspection qualification. The results from the machine learning classifier are compared to human performance. We show, that using sophisticated data augmentation, modern deep learning networks can be trained to achieve superhuman performance by significant margin. Keywords: Machine learning, NDT, Ultrasonic Inspection, Data augmentation, Virtual Flaws 1. Introduction Automated systems have long been used for flaw detection in various Non-destructive evaluation (NDE) systems. The au- tomated systems provide consistent results and do not show the variation commonly seen in human inspectors due to fatigue, stress or other factors. However, the traditional automated sys- tems have relied on simple decision algorithms such as a sig- nal amplitude threshold. In more demanding inspection cases, such as the typical ultrasonid inspections, the human inspec- tors achieve far superior inspection results than the simplistic automated systems. Consequently, in most of these inspections the data analysis are currently analyzed by human experts, even when the data acquisition is highly automated. Such analysis is time consuming to do and taxing for the personnel. The key problem with more sophisticated automation has been, that the work of the human inspector does not lend it- self to simple algorithmic description. The inspectors acquire their skill through years of training and utilize various signal characteristics in their judgement (e.g. the signal dynamics). Machine learning (ML) systems can be used to automate sys- tems, where direct algorithmic description is intractable. The recent improvements in ML algorithms and computational tools (GPU acceleration, in particular) have enabled more complex and powerful models that reach near human-level performance in tasks like image classification and machine translation. Early attemps to use machine learning for NDT flaw detec- tion and classification focused on using simple neural networks to classify various types of NDT data. Masnata and Sunser (1996) used a neural network with single hidden layer to clas- sify various flaw types (cracks, slag inclusions, porosity) from ultrasonic A-scans. Before learning, the A-scan was reduced to 24 pre-selected features using the Fischer discriminant analysis. Chen and Lee (1993) used wavelet decomposition, to obtain features from A-scans and reported goal classification, while the training and testing was done with limited data set. Yi and Yun (1998) similarly used shallow neural network to train flaw type classifier with a larger data set. Although in many cases this early work reported high classification accuracy, the results proved to be difficult to scale and to extend to new cases. One of the issues with developing ML-models for defect classification has been the limited availability of training data. Liu et al. (2002) used finite element simulation results to pro- vide artificial NDT signals to augment training data. With the increase in computational power, the used machine learning models have become more powerful. Many authors have reported good results with shallow models like support vector machines (SVM’s). While these models offer high clas- sification capability, they also require a pre-selected set of fea- tures to be extracted from the raw NDT signal. Fei et al. (2006) used wavelet packet decomposition of ultrasonic A-scans to train SVM for defect classification in a petroleum pipeline. Sam- bath et al. (2010) used neural network with two hidden layers to classify ultrasonic A-scans using a hand-engineered set of 12 features. Shipway et al. (2019) used random forests to detect cracks from fluorescent penetrant inspections (FPI). Cruz et al. (2017) used feature extraction based on principal component Preprint submitted to NDT & E International March 28, 2019 ar X iv :1 90 3. 11 39 9v 1 [ ee ss .S P] 2 6 M ar 2 01 9 analysis to train a shallow neural network to detect cracks from ultrasonic A-scans. He reported good classification analysis with only 5 extracted features, and computational efficiency that makes such classification feasible as on-line evaluation support for inspector during manual scanning. Kahrobaee et al. (2018) demonstrated the use of machine learning to achieve data fusion by learning separate classifi- cation networks from different NDT data and using a com- bined classifier with the results from these separate classifiers. It is often the case in inspection, that more than one inspec- tion method is used. Ability to take better advantage of the multiple data sources would thus be advantageous. Also, such approach could be used to discriminate between different flaw types, especially when the training data is too limited or sepa- rate to allow direct learning of classifiers to separate between similar flaw types. The machine learning classifiers have been used to wide variety of NDT signals and classification cases. Tong et al. (2018) used deep convolutional neural networks (CNNs) to de- tect subgrade defect from ground penetrating radar signals. For NDT methods, that provide image or image-like raw data, deep CNNs used for image classification have been applied with lit- tle modification. Dorafshan et al. (2018) used the AlexNet (Krizhevsky et al., 2017) deep CNN for detecting cracks in con- crete from visual inspection images. Convolutional networks have recently shown great success with various image classification tasks (Marcus, 2018). The convolutional architechture lets the networks to learn position independent classification. The recent deep architerctures have shown the ability to learn increasingly abstract representations in higher layers, which obliviates the need for hand-engineered features (Zhang et al., 2016). These features make the deep convolutional networks also interesting for the flaw detection in NDE signals. Recently Meng et al. (2017), Zhu et al. (2019) and Mu- nir et al. (2018b) used deep CNNs for defect classification in ultrasonic and EC-data. Meng et al. (2017) used deep neural networks with an SVM top layer for enhanced classification ca- pability. The classifier was used to classify voids and delam- ination flaws in carbon fibre composite material. Before pre- sented to the CNN, the raw A-scan data was decomposed using wavelet packet decomposition and the resulting coefficients re- organized into 32x16 feature matrix. Thus, the CNNs classified the A-scans separately. Munir et al. (2018a) used deep CNN’s to classify austenitic stainless steel welds. The training data was obtained from weld training samples containing artificial flaws (i.e. solidification flaws). The data-set was augmented by shifting the A-scans in time-domain and by introducing Gaussian noise to the signal. Zhu et al. (2019) used deep CNN’s to detect cracks in eddy current signal. Also, drop-out layer was used to estimate the confidence of the classification, which is an important oppor- tunity in using ML in field NDT, where the reliability require- ments are very high. This work is also notable in that the raw signal database was exceptionally representative with NDT in- dications representing plant data for various defect types (Udpa and Ramuhalli, 2009). In summary, the current state of the art for using machine learning in NDT classification may be seen to focus on two distinct aims. Firstly, modern shallow ML models (e.g. ran- dom forests) with advanced feature-engineering are used with the aim to develop computationally lightweight models that can be implemented on-line to aid inspector in manual inspection. Secondly, deep CNNs are used to learn from raw NDT signals without the need for explicit feature engineering. The recent work on deep models takes full advantage of recent advances in models developed for other industries and shows good results across different NDT fields. For ultrasonic testing, the existing machine learning mod- els have mostly involved classification the single A-scan level. This is a natural approach for many applications, such as the previously studied manual inspection (Cruz et al., 2017) or for C-scan style classification of large inspection analysis as done by Meng et al. (2017). However, in many inspection cases, mechanized inspection and electronic scanning using phased array ultrasonic systems provide rich data-set where adjacent A-scans can be analysed together to provide more information. Machine learning application to such data-sets have not been widely published. In the present work, we present application of deep CNN for phased array ultrasonic data, where number of adjacent A-scans are considered together for improved flaw detection capability. Common obstacle for using powerful ML models in NDE classification is, that the available flawed data tends to be scarce. Acquiring sufficient representative data-set would in many cases necessitate artificially manufacturing large set of flawed sam- ples, which quickly becomes infeasible. Data augmentation is commonly considered a key tool for successful application of ML for small data sets and some authors have used data augmentation (Munir et al., 2018a) for ultrasonic data. In the present work, we significantly expand on the previously pub- lished data augmentation schemes for ultrasonic inspection by using virtual flaws to generate augmented data sets. The use of virtual flaws enables generation of highly representative aug- mented data set for ML applications. Finally, the key requirement for adaptation of ML machine learning models in many industries, is to show how they com- pare with human inspectors. Especially in high-reliability in- dustries like the nuclear and aerospace industries, there’s com- mon requirement to employ best-available means to guarantee structural reliability. In practice, this would mean that the ML models would need to show performance exceeding that at- tained by the human inspectors or to show performance that meets the current requirements set for the traditional inspection systems (e.g. show required a90/95 performance, as commonly required in the aerospace industry). However, in many cases even the human inspection performance is not quantified and known with sufficient reliability to allow direct comparison to developed ML models. In present work, we used human perfor- mance data obtained from previous research (Virkkunen et al., 2017) and developed the machine learning models to work on comparable data thus enabling direct comparison between hu- man inspector and modern machine learning model. 2 1.1. Virtual flaws and data augmentation The problem with ultrasonic training of machine learning models is the scarcity of representative ultrasonic data. Sam- ples with real flaws are difficult to come by and in terms of nu- clear power plants can be contaminated making them challeng- ing to use. Mock-ups can be made with representative flaws, but production of such mock-ups is costly and time-consuming. The mock-ups also tend to be specific to a certain inspection case. Virtual flaws can be used to generate sufficient represen- tative flawed ultrasonic data from limited set of mock-ups and flaws (Virkkunen et al., 2014, 2016; Svahn et al., 2018; Kosk- inen et al., 2018). In essence, the flawed sample is scanned and the ultrasonic data recorded. From the recorded data the flaw signals are extracted by comparing the signal data point by data point to a selected flawless area. The flaw signal extracted this way is guaranteed to be representative, since it is recorded from an existing flaw. The extracted flaw signal can then be im- planted into different locations of the scan data, point by point, allowing the generation of new virtual flaws. In addition, the depth and length of the flaw can be altered and various other sig- nal modifications can be achieved. The flaw signals extracted can be moved to different samples. Flaw signals acquired with different ultrasonic parameters can be made compatible with different files. Using the virtual flaws augmented data gener- ation is virtually unlimited and ample representative training data can be generated for the training of ML models. The ap- proach has some similarity with synthesized learning cases used by Bansal et al. (2018). 1.2. Estimation of NDT performance and probability of detec- tion (POD) NDE is most valuable when used in area, where its expected reliability is very high. Consequently, measuring the perfor- mance of an NDE system and its reliability, in particular, is demanding. Demonstrating this high reliability requires high number of evaluation results on relevant targets and, thus, high number of test samples with representative flaws. Providing these flawed test samples is costly and thus different method- ologies have evolved to optimize the use of the available test blocks. Currently, the standard way to measure NDE performance is to define a probability of detection (POD) curve and, in par- ticular the smallest crack that can be found at level of sufficient confidence, typically 90% POD at 95% confidence (a90/95). Ex- perimentally, the POD curve is determined with test block trials and a set of standardized statistical tools (Annis, 2009; ASTM, 2012, 2015). . In this paper hit/miss method was selected due to nature of the test set-up. While signal amplitude can be used with fewer test blocks, it does not include the effects of inspector judge- ment on the NDE reliability. Especially in noisy inspection cases such as austenitic stainless steel welds, flaw detection re- lies on pattern recognition, not just signal amplitude and a clear threshold, thus the result is filtered by the inspector. This was observed also by Virkkunen and Ylitalo (2016). For the present study and comparing human and machine inspectors, it’s vital Figure 1: Scan set-up with Zetec pipe scanner, extension fixed to the right side for scanner mounting. to include the judgement effect and thus, the hit/miss approach was chosen. 2. Materials & Methods 2.1. NDT Data Inspected specimen for data-acquisition was a butt-weld in an austenitic 316L stainless steel pipe. Three thermal fatigue cracks with depths 1.6, 4.0 and 8.6 mm were implemented in the inner diameter of the pipe near the weld root by Trueflaw ltd. and scanned with ultrasonic equipment. An austenitic weld was chosen as a test specimen due to being common in the industry. In addition austenitic weld has increased inspection difficulty due to noise caused by the anisotropy of the weld structure. Inspection method used for data acquisition was Transmission Receive Shear (TRS) phased array, one of the common methods used in inspecting of austenitic and dissimilar metal welds. The scan was carried out by using Zetec Dynaray 64/64PR-Lite flaw detector linked to a PC. The probes used were a Imasonic 1.5 MHz 1.5M5x3E17.5-9 matrix probes with central frequency at 1.8 MHz, element dimensions 3.35 x 2.85 mm and element ar- rangement as 5 x 3 elements. Used wedge was ADUX577A used to produce a shear wave efficiently. One linear scan with no skew angles was utilized. The ultrasonic wave was focused to the inner surface of the pipe and the probe was positioned in a way that the beam would be focused directly to the manufac- tured cracks. Coupling was applied through a feed water system and the pipe was rotated underneath the probe to assure con- stant and even coupling between the probe and the pipe. Probe position was carefully monitored along the scan line by Zetec pipe scanner with 0.21 mm scan resolution. The specimen and the inspection procedure is described in more detail in Koski- nen et al. (2018). The specimen and the scanner can be seen in Figure 1. For data efficiency, only a signle angle was used. The cho- sen angle was the one, where the cracks were the most visible. In this case, this was the 45 angle. As only one scan line was ac- quired, the data was visualized and evaluated using B-scan im- ages. Since the crack locations and sizes were precisely know, the crack signals could be removed from the ultrasonic data to 3 create a blank canvas. Virtual flaw augmentation was used to broaden the representative sizes of the cracks. The virtual flaw software used was Trueflaw’s eFlaw. In this case, the eFlaw was used with an assumption that signal amplitude is the most sig- nificant feature of the crack signal from detection point of view. Similar assumption is used in the signal response POD estima- tion (â vs. a). The eFlaw was used to modify and scale down the original crack signal amplitude to represent different vari- ety of cracks with smaller sizes than the original. This allows creation of high amount of crack images required for POD esti- mation and for teaching datasets for ML algorithms. Details of the eFlaw technology are explained in Virkkunen et al. (2014, 2016); Svahn et al. (2018); Koskinen et al. (2018). The teaching data set was created in same way as for testing data set for human inspectors in previous paper Virkkunen et al. (2017). Once the teaching was finished, the ML algorithm was tested with the same data as human inspectors faced. Thus, the ML algorithm and human inspectors were given the exact same information with the same controlled environment and a POD curve was estimated based on the hit/miss results. 2.2. Training data and used data augmentation The single 45°scan line data containing signals from three manufactured thermal fatigue flaws was taken as the source data for training the machine learning model. This is the same data, that was used to generate human POD results in (Virkkunen et al., 2017). From this data, large number of data files were generated using the same algorithm as previously. The data contained 454 A-scans each containing 5058 samples with 16 bit depth. For machine learning purposes, the data was further pro- cessed, as follows; each A-scan was cut so that only the inter- esting area around the weld was included resulting in 454 x 454 point data. Then, the resolution of the ultrasonic data was down sampled to 256 x 256 points. Altogether 20000 variations were generated to be used as training and validation data. The data was stored in minibatches of 100 UT-images per file with accompanying true state infor- mation showing the included crack state present, if any. The data set also contained information, where virtual flaw process had been used to copy unflawed section to another location. This was done to avoid and to detect the possibility that the machine learning model would learn to notice the virtual flaw introduction process, instead of the actual flaws. 2.3. Used ML architecture The machine learning architecture used was based on the VGG16 network (Zhang et al., 2016). For ultrasonic data anal- ysis, the basic network was augmented with a first max-pooling layer, with pooling size adjusted to the wavelength of the ultra- sonic signal. This max-pooling layer had the effect of removing spectral information from the image so that the rest of the net- work was left with an envelope amplitude curve. The effect of this layer is shown in Figure 2. The training used binary cross entropy as the cost function and training was done using the RMSProp (Zeiler, 2012). Max pooling 7 x 1 Crack indication Figure 2: Max pooling was implemented as a first layer that removed the spec- tral informmation and reduced dimensionality of the data. 4 Previous work (Chen and Lee, 1993; Fei et al., 2006; Meng et al., 2017) typically extracted additional information from the spectral content of the A-scan data using, e.g., the wavelet de- composition. In this case, it was also considered to add addi- tional data layers obtained with wavelet decomposition. How- ever, the source data that was used for human inspectors was rectified, which made obtaining any useful information from the spectral content impossible. Since in this case, it was desir- able to use data, that was directly comparable to the data seen by the human inspectors it was decided to continue working with the rectified data. The data was read in the saved mini-batches, converted to 32 bit floating point numbers and normalized by subtracting the mean and dividing by standard deviation. A small value of 0.00001 was added to avoid division by zero. The size of the various layers were originally excessive, and as soon as successful training was obtained, the layer sizes were decreased step-by-step to obtain the most efficient network ca- pable of learning to classify the data. The full architecture (both initial trial and final) is shown in Figure 3. The network expe- rienced some sensitivity to initialization, and on repeated train- ing, the model sometimes failed to learn successfully. The computation was implemented with the Keras library (Chollet et al., 2015) using the TensorFlow back-end (Abadi et al., 2015). The chosen architecture does not make use of some of the recent features included in state of the art deep convolutional networks. The primary motivation for this was to keep the net- work as simple as possible while showing good flaw detection capability. Some of the considered, but not included, ML archi- tectural features are discussed in the following. Drop-out (Hinton et al., 2012) has been extensively used to prevent overfitting, and more recently to estimate prediction confidence (Zhu et al., 2019). In the present study, the model did not show susceptibility to overfitting. The likely reason for this is the high number of augmented images used for train- ing. Consequently, drop-out was not included and instead the training was stopped after sufficient performance was achieved. Training with smaller augmented data-sets could show overfit- ting and, consequently, make use of drop-out. Furthermore, even in the absense of overfitting, the use of drop-out to esti- mate prediction accuracy is an interesting option especially in case where multiple flaw types are classified within one model. Batch renormalization has shown to improve trainability of very deep networks (Ioffe and Szegedy, 2015). While the present network did show sensitivity to initialization values and some- times failed to train successfully, this did not present signifi- cant problem in this application. A simple re-try with different random starting values quickly resulted in successful training result. Channel-wise training (Chollet, 2017b) has been used to ease training and to improve training results in image classi- fication. In the present case, the interesting channel-wise infor- mation would be amplitude information (as used in the present analysis) and frequency-related information, such as the wavelet decomposed features used, e.g., by Chen and Lee (1993); Fei et al. (2006). However, in this case, it was of interest to use as- Input layer 256 x 256 pixels 16 bit depth Max pooling 7 x 1 Convolution 96, 3 x 3 Activation: relu Convolution 64, 3 x 3 Activation: relu Max pooling 2 x 8 Convolution 48, 3 x 3 Activation: relu Convolution 32, 3 x 3 Activation: relu Max pooling 3,4 Dense 14 Activation: relu Output layer, Dense 1 Activation: sigmoid Flatten Input layer 256 x 256 pixels 16 bit depth Max pooling 7 x 1 Convolution 128, 3 x 3 Activation: relu Convolution 128, 3 x 3 Activation: relu Max pooling 2 x 8 Convolution 128, 3 x 3 Activation: relu Convolution 128, 3 x 3 Activation: relu Max pooling 3,4 Dense 28 Activation: relu Output layer, Dense 1 Activation: sigmoid Flatten Optimized network Initial network Figure 3: The trained network stucture. Max pooling was implemented using Keras MaxPooling2D layer. Convolution layers were implemented using Keras Conv2D layer. The final dense layer was implemented with Keras Dense layer. 5 is the data that was used in previous research (Virkkunen et al., 2017) to estimate human POD performance. As this data was rectified, most of the spectral data was lost and could not be used. Extracting spectral features using wavelet decomposition as separate channels remains interesting option for further study and may improve flaw detection. 2.4. Performance evaluation In previous research (Virkkunen et al., 2017) an online tool for assessing inspector performance was developed. The tool presents randomly generated B-scan data with implemented vir- tual cracks and a possibility to change the software gain. In the normal mode the inspectors select the locations of the cracks and move on to the next image. In the learning mode feed- back from the previous image is provided before moving to the next image. The tool is publicly available at http://www. trueflaw.com/truepod and http://www.trueflaw.com/ truelearnpod. Not all images include cracks. The results are used to produce hit and miss POD-curve. In previous research, nine level-III ultrasonic inspection course attendees were ran- domly split into two groups to use the learning mode and the normal mode. Each inspector had time to practise with the tool during the course. Finally each inspector analysed 150 images and hit and miss POD-curve was generated. One inspector was excluded from the data due to excessive amount of false calls. For inspectors the best achieved a90/95 value was at 1 mm and under 20 false calls. Most inspectors rated between 1 - 2.5 mm a90/95 and under 30 false calls. The lower-end inspectore got a90/95 between 3.5 and 4.0 mm and the highest false call rates were above 180. The number of false calls did not correlate with inspection performance. While the online tool does not re- flect realistic inspection situation, it allows relatively rapid and cost-efficient gathering of relevant performance data. Inspec- tion is often done in suboptimal conditions, and requires skilled inspector. In addition, the rate at which flaws appear is low making the already repetitive work even more tiring. The target in this study is to assess the performance of the ML model with regard to inspector performance. In addition to the previous data, that included independent inspectors, a new data set was generated. To get direct comparison between the human inspectors and the ML model, a new set of 200 B-scan images not used in the training of the ML model was generated and a hit and miss POD-curve made for the ML model. A spe- cialized version of the previously used online tool for POD eval- uation was created with this data set. Human results were then obtained from 3 experienced inspectors from VTT. The same data-set was then given to the classifier network. This set-up enabled direct comparison of human and machine performance in a blind set-up. This data-set contained 200 images and 86 images with cracks. Both the humans and the ML-network had opportunity to train with similar data and similar set-up. For these data, the range of available inspectors is more limited, but the data is even more comparable. Validation loss Validation accuracy Va lid at io n lo ss /a cc ur ac y 0 0.2 0.4 0.6 0.8 1.0 Epoch 0 10 20 30 40 50 60 70 80 90 100 110 Figure 4: Validation loss and validation accuracy during training for 100 epochs. 3. Results 3.1. Training results The network was trained for 100 epochs of 10000 samples. This resulted in perfect classification: all cracks were correctly classified and no false calls were made. The evolution of the training accuracy is shown in Figure 4. The number of train- ing epochs was set by hand to stop slightly after perfect clas- sification score was achieved. During development, the results were evaluated against a separate validation set. The final result was then evaluated against a previously unseen verification set. Each set contained a 100 images, with roughly 50% cracks. 3.2. Comparison with human performance To evaluate the network performance against human per- formance, data set from previous work was utilized Virkkunen et al. (2017). In addition a new data set was generated for this purpose (section 2.4). The performance was evaluated using MIL-HDBK-1823a hit/miss analysis (Annis, 2009). The performance comparison is summarized in table 1. POD curve for the human inspectors and the ML network are shown in images 5 and 6, respectively. As noted in previous research, the cracks contained in the orig- inal data presented different challenge in relation to their size. This was primarily caused by the difference in relative ampli- tude. The same crack was difficult for both the human inspec- tors and the ML network. In the current data set, the small number of initial flaws as well as their difference caused some irregularities in the hit/miss performance, which the computed confidence bounds to be rather wide. For one inspector, the hits and misses did not show the expected crack size dependence. This may have been caused by excessive false calls for the in- spector. For the ML classifier, all the cracks were found. To get convergence for the POD curve, 30 misses of zero-sized cracks were added to all the results. This had the effect of improving slightly the a90/95 values of the human inspectors and provid- ing convergence for the ML-classifier even with all the cracks found. In future studies, wider selection of physical cracks are needed to avoid such problems. 6 Figure 5: Example POD curve from a human inspector. Note, that additional cracks were added at 0 crack length for comparability on ML-results. The data shows anomalous POD-a dependence due to differences in detectability of various natural cracks. In the future, this can be alleviated by additional cracks to better cover variability in natural cracks. Figure 6: POD curve the machine learning classifier. Note, that additional cracks were added at 0 crack length for convergence. Table 1: Comparison of performance from human inspectors and machine learning classifier. For ML classifier, all the cracks were found and smallest found crack is shown as a90/95 . Inspection a90/50 False calls Previous data 1 - 2.5 130 Inspector 1 3.0 36 Inspector 2 2.7 917 Inspector 3 5.6 2 ML classifier 0.9 0 4. Discussion The present study showed, that the current very deep ma- chine learning networks are powerful enough to achieve super- human performance on NDT-tasks previously considered in- tractable, such as crack detection in ultrasonic signals. This is, to the best of our knowlede, the first time that a direct com- parison is published between human inspectors and machine- learning classifiers. Achieving superhuman performance is an important milestone, since it indicates that the machine learn- ing networks can be used also in fields, where high reliability is sought after and regulatory requirements mandate the use of best available means, such as in the nuclear industry. Data augmentation is a well known technology in the ML literature and is commonly considered to be a key enabling technique when working with limited data sets Chollet (2017a). Data augmentation has also previously used for NDT appli- cations of ML (Munir et al., 2018b). In present study, exten- sive data augmentation was utilized using the previously devel- oped virtual flaw technology. This allowed generating training data, that incorporated many aspects of actual inspection, such as the detection of flaw signals from varying backgrounds and variations in probe contact, without extensive data base of real cracks. This can be expected to yield ML-models that gener- alize well to different real-world inspection cases. In addition, the virtual flaw technology has been used in training human in- spectors, and expected to be used in nuclear qualifications in the near future. The use and extensive validation of the vir- tual flaw technology in the case of human inspectors gives high confidence that the augmented data sets are relevant also for ML applications. The results from present study indicate, that such domain- specific and separately validated data-augmentation techniques enabling technique for succesfully applying machine learning in various NDE fields, where the data is scarse but performance requirements high. In previous work, the ML-classification of ultrasonic signal is usually applied at the single A-scan level. In contrast, our approach has been to train the network on full scan of 454 A- scan lines. This approach necessarily limits the applicability of the solution to mechanized or location-encoded inspections, where such coordinated combination of A-scans is available. The present work has some significant limitations. The raw data contained only three real cracks, that were then modified to give the total data set. This was similar for both the hu- 7 man inspectors and the machine learning solution. The natu- ral flaws exhibit significant variation and a set of three flaws is clearly insufficient to capture this variation. For example, the ASTM POD standard (ASTM, 2015) requires 30 cracks, which is chiefly to to capture this variation. Thus the network trained here is not expected to work as-is for more general crack de- tection tasks. Instead, future research will extend the source data using additional thermal fatigue cracks, simulated flaws and other interesting signal types. 5. Conclusions The following conclusions can be drawn from this study: • Deep convolutional neural networks are powerful enough to reach superhuman performance in detecting cracks from ultrasonic data • Data augmentation using virtual flaws is seen as key en- abling technique to train machine learning networks with limited flawed data 6. Data availability The used python code as well as the training data set is made available for download at https://github.com/iikka-v/ ML-NDT. 7. Acknowledgements The data augmentation using virtual flaws and the initial network & training was contributed by Trueflaw Ltd. Their contribution is gratefully acknowledged. 8. References References Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Cor- rado, G. S., Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M., Levenberg, J., Mané, D., Monga, R., Moore, S., Murray, D., Olah, C., Schuster, M., Shlens, J., Steiner, B., Sutskever, I., Talwar, K., Tucker, P., Vanhoucke, V., Vasudevan, V., Viégas, F., Vinyals, O., Warden, P., Watten- berg, M., Wicke, M., Yu, Y., and Zheng, X. (2015). TensorFlow: Large-scale machine learning on heterogeneous systems. Software available from ten- sorflow.org. Annis, C. (2009). Mil-hdbk-1823a, nondestructive evaluation system reliability assessment. Technical report. ASTM (2012). Standard practice for probability of detection analysis for hit/miss data. ASTM E2862-12, American Society for Testing and Mate- rials. ASTM (2015). Standard practice for probability of detection analysis for â ver- sus a data. ASTM E3023-15, American Society for Testing and Materials. Bansal, M., Krizhevsky, A., and Ogale, A. (2018). Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. Chen, H. and Lee, G. G. (1993). Neural networks for ultrasonic nde signal clas- sification using time-frequency analysis. In IEEE International Conference on Acoustics, Speech, and Signal Processing. Chollet, F. (2017a). Deep Learning with Python. Manning Publications Co., Greenwich, CT, USA, 1st edition. Chollet, F. (2017b). Xception: Deep learning with depthwise separable convo- lutions. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Chollet, F. et al. (2015). Keras. https://keras.io. Cruz, F. C., Simas Filho, E. F., Albuquerque, M. C., Silva, I. C., Farias, C. T., and Gouvea, L. L. (2017). Efficient feature selection for neural network based detection of flaws in steel welded joints using ultrasound testing. Ul- trasonics, 73:1–8. Dorafshan, S., Thomas, R. J., and Maguire, M. (2018). Comparison of deep convolutional neural networks and edge detectors for image-based crack de- tection in concrete. Construction and Building Materials, 186:1031–1045. Fei, C., Han, Z., and Dong, J. (2006). An ultrasonic flaw-classification system with wavelet-packet decomposition, a mutative scale chaotic ge- netic algorithm, and a support vector machine and its application to petroleum-transporting pipelines. Russian Journal of Nondestructive Test- ing, 42(3):190–197. Hinton, G. E., Srivastava, N., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. R. (2012). Improving neural networks by preventing co-adaptation of feature detectors. Ioffe, S. and Szegedy, C. (2015). Batch normalization: Accelerating deep net- work training by reducing internal covariate shift. Kahrobaee, S., Haghighi, M. S., and Akhlaghi, I. A. (2018). Improving nonde- structive characterization of dual phase steels using data fusion. Journal of Magnetism and Magnetic Materials, 458:317–326. Koskinen, T., Virkkunen, I., Papula, S., Sarikka, T., and Haapalainen, J. (2018). Producing a pod curve with emulated signal response data. Insight, 60(1):42–48. Krizhevsky, A., Sutskever, I., and Hinton, G. E. a. (2017). Imagenet classifica- tion with deep convolutional neural networks. Communications of the ACM, 60(6):84–90. Liu, S., Huang, J. H., Sung, J., and Lee, C. (2002). Detection of cracks using neural network and computational mechanics. Computer methods in applied mechanics and engineering, 191:2831 – 2845. Marcus, G. (2018). Deep learning: A critical appraisal. Masnata, A. and Sunser, M. (1996). Neural network classification of flaws detected by ultrasonic means. NDT & E International, 29(2):87–93. Meng, M., Chua, Y. J., Wouterson, E., and Ong, C. P. K. (2017). Ultrasonic signal classification and imaging system for composite materials via deep convolutional neural networks. Neurocomputing, 257:128–135. Munir, N., Kim, H. J., Park, J., Song, S. J., and Kang, S. S. (2018a). Convo- lutional neural network for ultrasonic weldment flaw classification in noisy conditions. Ultrasonics. Munir, N., Kim, H.-J., Song, S.-J., and Kang, S.-S. (2018b). Investigation of deep neural network with drop out for ultrasonic flaw classification in weldments. Journal of Mechanical Science and Technology, 32(7):3073– 3080. Sambath, S., Nagaraj, P., and Selvakumar, N. (2010). Automatic defect classi- fication in ultrasonic ndt using artificial intelligence. Journal of Nondestruc- tive Evaluation, 30(1):20–28. Shipway, N. J., Barden, T. J., Huthwaite, P., and Lowe, M. J. S. (2019). Au- tomated defect detection for fluorescent penetrant inspection using random forest. NDT & E International, 101:113–123. Svahn, P.-H., Virkkunen, I., Zettervall, T., and Snögren, D. (2018). The use of virtual flaws to increase flexibility of qualification. In 12th European Conference on Non-Destructive Testing (ECNDT 2018), number 8 in The e-Journal of Nondestructive Testing. NDT.net. Tong, Z., Gao, J., and Zhang, H. (2018). Innovative method for recognizing subgrade defects based on a convolutional neural network. Construction and Building Materials, 169:69–82. Udpa, L. and Ramuhalli, P. (2009). Steam generator management program: Automated analysis of array probe eddy current data. Technical Report 1018559, EPRI, Palo Alto, CA. Virkkunen, I., Haapalainen, J., Papula, S., Sarikka, T., Kotamies, J., and Hänninen, H. (2017). Effect of feedback and variation on inspection relia- bility. In 7th European-American Workshop on Reliability of NDE. German Society for Non-Destructive Testing. Virkkunen, I., Miettinen, K., and Packalén, T. (2014). Virtual flaws for nde training and qualification. 11th European Conference on Non-Destructive Testing (ECNDT 2014), October 6-10, 2014, Prague, Czech Republic. Virkkunen, I., Rönneteg, U., Grybäck, T., Emilsson, G., and Miettinen, K. (2016). Feasibility study of using eflaws on qualification of nuclear spent 8 fuel disposal canister inspection. International Conference on Non Destruc- tive Evaluation in Relation to Structural Integrity for Nuclear and Pressur- ized Components ; Conference date: 04-10-2016 Through 06-10-2016. Virkkunen, I. and Ylitalo, M. (2016). Practical experiences in pod determi- nation for airframe et inspection. International Symposium on NDT in Aerospace ; Conference date: 03-11-2016 Through 05-11-2016. Yi, W. and Yun, I.-s. (1998). The defect detection and non-destructive eval- uation in weld zone of austenitic stainless steel 304 using neural network- ultrasonic wave. KSMME Enternational Journal, 12(6):1150 – 1161. Zeiler, M. D. (2012). Adadelta: An adaptive learning rate method. Zhang, X., Zou, J., He, K., and Sun, J. (2016). Accelerating very deep con- volutional networks for classification and detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):1943–1955. Zhu, P., Cheng, Y., Banerjee, P., Tamburrino, A., and Deng, Y. (2019). A novel machine learning model for eddy current testing with uncertainty. NDT & E International, 101:104–112. 9

Journal of Nondestructive Evaluation (2019) 38:89 https://doi.org/10.1007/s10921-019-0628-z Comparison of â Versus a and Hit/Miss POD-Estimation Methods: A European Viewpoint Iikka Virkkunen1 · Tuomas Koskinen2 · Suvi Papula1 · Teemu Sarikka1 · Hannu Hänninen1 Received: 23 March 2019 / Accepted: 23 August 2019 © The Author(s) 2019 Abstract For estimating the probability of detection (POD) in non-destructive evaluation (NDE), there are two standard methods, the so-called â versus a approach and the hit/miss approach. The two approaches have different requirements for the quality and quantity of input data as well as for the underlying NDE method. There is considerable overlap between the methods, and they have different limitations, so it is of interest to study the differences arising from using each methodology. In particular, if the dataset is not ideal, the methodologies may exhibit different problems dealing with various limitations in the data. In this paper, a comparison between â versus a and hit/miss analysis was completed for two different data sets, a manual aerospace eddy-current inspection and a nuclear industry phased array ultrasonic weld inspection using a simplified online tool. It was found that the two standardmethods (â vs. a and hit/miss) may give significantly different results, if the true hit/miss decision is based on inspector judgement and not automated signal threshold. The true inspector hit/miss performance shows significant variance that is not attributable to signal amplitude. Model-assisted PODwas not able to model the inspector performance due to lack of representative amplitude threshold and difficulties in capturing true signal variance. The paper presents experience from practical cases and may be considered a European viewpoint. Keywords Non-destructive testing · NDT · NDE · Probability of detection · POD · Reliability 1 Introduction The best practices of estimating probability of detection (POD) in non-destructive evaluation (NDE) are now well established. The longstanding MIL-HDBK-1823A [1] is used extensively in the aerospace industry [2–4] and is now finding increasing use also in other areas, like the rail industry [5, 6] and Nuclear industry [7]. The methods have recently been standardized by ASTM [8, 9] and these standards are congruent with the current MIL-HDBK methodology. The standard practice offers two variant of POD curve estimation, the so-called â versus a approach and the hit/miss approach. The â versus a approach models, in simple terms, the NDE reliability as kind of measurement system problem, where the quantity to be measured (crack size a) give rise to measured signal (â) proportional to the measured quantity B Iikka Virkkunen iikka.virkkunen@aalto.fi 1 Aalto University, PL 14200, 00076 AALTO Espoo, Finland 2 VTT Technical Research Centre of Finland, PL 1000, 02044 VTT Espoo, Finland and the task is to determine the possible existence of the sig- nal with decreasing a (and thus decreasing â). The system has noise both on the signal (â varies due to factors other than a), which results in noisy â versus a relation. In addition, there’s noise, that is independent of a. Thus, the task is to find a deci- sion threshold (â value), that minimizes false calls from the noise and, in parallel, maximizes the number of cracks found (i.e. cracks with â above the threshold), given the variation in the â versus a relation. The ASTM-E3023 (andMIL-HDBK) solve this by fitting a linear function through the â versus a data, computing prediction intervals to take the notice and statistical uncertainty into account. The resulting best-fit and confidence limit lines are then compared to the set detec- tion threshold and the corresponding POD curves computed. Improvements to the classical Berens [10] model have been proposed to behave better with very limited data sets, e.g. by Syed Akbar Ali et al. [11], Syed Akbar Ali and Rajagopal [12] and Le Gratiet et al. [13]. For input, the â versus a analysis requires a set of represen- tative flaws (at least 30) and measurements of signal strength â and corresponding crack size a. In addition, noise indepen- dent of crack size needs to be evaluated either with additional 0123456789().: V,-vol 123 89 Page 2 of 13 Journal of Nondestructive Evaluation (2019) 38:89 measurements of crack-free samples or in connection with the same sample set measurement. Recently, the cost and lack of representative test pieces has been alleviated by using simulated inspection results in lieu of actual physical test samples and measurements. This is calledmodel-assisted POD (MA-POD), and is widely applied in different contexts. Typically, a simulation is used to provide â versus a data for the inspection case of interest. Formulating the computation involves several simplifications to the physical reality to make the modelling effort feasi- ble. These include simplified physical models to describe the inspection signal (e.g. wave propagation laws), simpli- fied material data (e.g. homogeneous and isotropic material instead of the actual inhomogeneous material) and simpli- fied flaw description (e.g. a simplified notch-like reflector instead of tortuous and branching crack). Due to these simpli- fications, the simulated data is normally free from noise and exhibits no variation for given flaw size and configuration. The variance in â versus a is introduced by varying flaw con- figuration parameters, e.g. flaw tilt or skew angle and location in relation to geometric features. The variation in flaw con- figuration then produces variation in â versus a dependence, which is then used to compute â versus a POD curves using standard methodology. Simulations enable computation of large number of cases and thus the statistical sampling error in the results can be decreased to arbitrary low values. The hit/miss approach, in contrast, does not deal with sig- nal values, but estimates the POD curve based on binary results, that is hits (correctly found cracks) andmisses (cracks not found in the inspection). Because the data contains less information (regarding the correlation between crack size and signal strength or “ease of detection”) more samples are needed for reliable POD determination (at least 60 [14]). Some statisticians have recommended sample sized over 300 for hit/miss [15], especially if the 95% confidence limits on POD curves are calculated according to MIL-HDBK-1823. There are also improved statistical methods proposed in the literature, allowing POD curves to be reliably determined from data sets of even as few as 50 hit/miss observations [16]. The POD curve is solved using generalized linear model and a chosen link function (typically logistic, but sometimes pro- bit), that gives the shape of the POD curve using maximum likelihood fit to the data. The corresponding confidence lim- its are the obtained by the likelihood-ratio method, where a likelihood surface near the maximum likelihood value is interrogated, POD curves with likelihoods corresponding to the chosen confidence interval computed and the lower (and upper) limit curves solved. The number of samples and the flaw size distribution in relation to the actual POD curve also affect the width of the confidence bounds [14]. For input, the hit/miss analysis requires a set of represen- tative flaws (at least 60) and hit/miss results for each crack. In addition, the hit/miss results should exhibit a range with “unlikely to find” cracks, a range with “likely to find” cracks and transition in between. Otherwise, the logistic (or pro- bit) model does not describe the data and, while a fit may in some cases be obtained, it does not describe the underlying probability of detection. In both cases, the basic assumptions underlying both POD models should be fulfilled: the POD should be an increasing function of the crack size and should reach 100% with suf- ficient crack size. If the data contains signs of violation of these assumptions (e.g. a miss with big crack length indicat- ing that the POD does not reach 100% even with large crack size), the standard models are not applicable and alternate model must be sought. Such alternate models exist, among others, for POD with limited maximum POD etc. With the two models available, the user has a choice of method to make. In many cases, the choice is predetermined by the available data, i.e. signal values or hit/miss data. However, especially when designing a POD determination project, both methods may be available and they may give different results. The difference and, indeed, the validity of eachmethodmay be difficult to judge beforehand and if (as is often the case) only one is completed the possible difference remains unknown. The two methods have different requirements for the underlying NDE method as well. The â versus an assumes, that the method can bemodelled by a single detection thresh- old and â versus a correlation. In many cases, the inspectors use information other than the signal strength to judge crack existence, and in such cases the â versus a does not describe the true performance of the system. Even more disturbingly, the measurement of the signal may also be affected by the inspector. For example, in manual EC inspection, the inspec- tor oftenmay receive spurious signals from small aberrations on the surface etc., and will compensate by doing repeated measurements and reporting the “correct” signal. Thus, the true noise is not recorded and is already filtered in the inspec- tor reporting. Similarly, when inspector judges a crack to be present, again repeated measurements are taken to find the “correct” signal strength. Again, the â versus a correlation is distorted by inspector judgement. Consequently, the â ver- sus a methodology is mostly applicable for highly automated systems, where human intervention is insignificant and sin- gle detection threshold fully describes the crack detection process. However, in practice it may be difficult to assert the absence of human intervention. In addition, the â versus a methodology requires fewer samples than the hit/miss and thus there may be a preference for using it even when the inspection is not fully automated. In contrast, the hit/miss analysis deals with the direct results of the inspection (i.e. hits and misses) and thus may incorporate information andvarianceof inspector judgement. Thus much less assumptions are made regarding the inspec- tion method or hit/miss judgement and the method is more 123 Journal of Nondestructive Evaluation (2019) 38:89 Page 3 of 13 89 robust in this respect. The inspection system can, in theory, be regarded as a “black box” and POD evaluation is done with the end results only. Thus, even if some of the aspects affecting inspector judgement are unknown, this does not jeopardize the validity of the analysis. At the same time, often the method is not a black box, and there may be significant information available, that describes the relevant “ease” of the detection (e.g. signal strength obtained), that is not incor- porated into the analysis and thus, in effect wasted. Thus the analysis may seem wasteful. Since there is considerable overlap between the twometh- ods, and since they have different limitations, it is of interest to study the differences arising from using each methodol- ogy. In particular, if the dataset is not ideal, themethodologies may exhibit different problems dealingwith limitations in the data. The source of variation in the POD estimation can be attributed to several distinct sources as follows: • Statistical sampling error: the error caused by limited set of available data for the approximation of the POD curve. • Measurement variation: for given crack and measurement set-up, there may be variation in the obtained â values due to operator variability, equipment calibration differences etc. • Configurational variation: for given nominal crack size, theremaybevariationduevariation in the crackorientation and location. • Variation in crack characteristics: natural cracks may exhibit differing â for same nominal crack size a. Various factors besides the crack size may affect the obtained sam- ple amplitude (e.g. crack path tortuosity, opening, surface roughness etc.). In some cases, this is modelled directly with so-called multivariate POD curves, where the POD curve is explicitly stated and modelled to be function of crack size and other chosen parameters. For standard anal- ysis, these other features are represented as variation in the â versus a correlation and affect the confidence bounds cal- culated for the POD. • Inspector judgement variation: for given â obtained from the inspection, inmany cases there is an element of inspec- tor judgement in translating the obtained signal strength to “cracks/no crack” -judgement. This may depend on the local data variance, noise surrounding the flaw, inspec- tor variability etc. Even when there’s an explicit detection threshold set, in manual inspection the inspector needs to separate the true indication from possible spurious signals. Thus the recorded signal may be affected by the inspector judgement and crack signals may be overlooked as arte- facts. Table 1 shows a comparison of â versus a and hit/miss in terms of how these sources of variation are handled. The statistical sampling error is explicitly handled with both â Table 1 Comparison of â versus a and hit/miss in terms covered sources of variation Source of variation MA-POD â versus a Hit/miss Statistical sampling error N/A YES YES Configurational variation YES YES YES Measurement variation NO YES YES Variation in crack characteristics NO YES YES Inspector judgement variation NO NO YES versus a methodology. The variation in crack characteristics and the measurement variation are directly measured in both methodologies and thus can be considered to be contained within the statistical scatter and confidence bounds, although there are differences, e.g.with the number of cracks used. The biggest difference is in the inspector judgement variation. In â versus a, this effect is assumed to be negligible, whereas for hit/miss the variation is included in the data. Thus, the main focus in selecting between â versus a and hit/miss is related to whether this effect can be assumed to be negligible. Finally, for the MA-POD case, the variation is assumed to come from configurational issues (flaw tilt, skew, etc.). The statistical error can be decreased by additional simulations, which is cheap in comparison to manufacturing physical test samples. On the other hand, variation in crack charac- teristics, sample microstructure and possible measurement system issues are not included. Inspections conducted by human inspectors are known to exhibit variability in inspector judgement. The variability is seen between different inspectors and between different inspections carried out by the same inspector (see e.g. [17, 18]). This variability resulting from inspector variance is often referred to as “human factors” effect. As shown in Table 1, in the hit/miss approach, inspector judgement is directly included in the results and thus any possible human factors that are present during the exercise are reflected in the POD results. However, inspection conditions during a POD exercise are seldom identical to the real inspection conditions, even when care is taken to make the exercise as representative as possible. Most notably, the number of flaw findings in a POD exercise is typically much higher than in normal inspections, whichmay affect inspector expectations. For the â versus a approach, the human factors are not, in general, included. When the â values are sourced from mod- elling or automated inspection systems, the human factors are not included in the analysis and need to be addressed separately, if the results are to be used in conditions, where human inspectors report â ormake flawdecisions. If the â val- ues are sourced from human inspectors, variation in human judgement may affect the results [19]. Over the years, several modifications to the traditional â versus a and hit/miss methodologies have been proposed to 123 89 Page 4 of 13 Journal of Nondestructive Evaluation (2019) 38:89 overcome some of its deficiencies. The statistical methodol- ogy and, in particular the computation of confidence bounds have been evaluated and alternate methods proposed [11–13, 16, 20–22]. In particular, the focus has been to obtain more robust confidence bounds, in case of small dataset using, e.g. the Bayesian approach. One of the key assumptions of the hit/miss approach is, that the probability of detection increasesmonotonicallywith increasing crack size a. Furthermore, it is typically assumed, that the POD reaches 100% at some crack size a. In reality, there may be error sources, that do not follow such depen- dence on crack size. Generazio [23–25] proposed alternate formulation based on design of experiment (the design of experiment probability of detection, DOEPOD). The DOE- POD model is based on extending the binomial view of hit/miss data. The main motivation for the DOEPOD model is, that using model-based POD estimation (e.g. ASTM E2862) assumes POD as a function of flaw size follows certain model. In particular, the POD is continuous, mono- tonically increasing function of flaw size a. This assumption may not always be justifiable, e.g. when the method sensitiv- ity varies for different flaw sizes due to different probes, beam focusing or for some other reason. TheDOEPODmodel does not assume functional relationship between POD and flaw size. Instead, the inspection results are grouped and anal- ysed, simply stated, as groups to make sure that the binomial 90/95% condition is fulfilled for certain flaw size ranges. Despite more recent formulations, the MIL- HDBK/ASTM methodologies [1, 8, 9] are still widely used and thus it is of interest to still study and better understand their limitations. In this paper, a comparison between â versus a and hit/miss analysis was completed for two different data sets. The first data set describes a manual aerospace eddy-current inspec- tion. In this data-set the signal strength is recorded bymanual process and thus there is potential for inconsistencies in the â versus a relation. At the same time, the inspection could also be analysed with â versus a methodology, since the pro- cedure does provide signal strength and process calibration and procedure definition are expected to minimize any such operator effects. The data set was arranged and gathered by Patria Aviation and The Finnish Defence Forces in condi- tions resembling true inspection as closely as possible by introducing test samples into representative locations on air- frames and testing them in proper locations. The second data-set represents a nuclear industry phased array ultrasonic weld inspection and was collected using a simplified online tool. The underlying data has limited real flaws. However, unlike traditional inspection records, the online analysis methodology provided direct opportu- nity to study the relationship between signal strength and hit/miss judgement with large number of artificially gener- ated flawed data images. These results were compared with Fig. 1 Crack size distribution model-assisted POD results generated for the same inspec- tion case. 2 Materials andMethods Two data sets were applied for this study, designated “EC” and “UT”. The EC data set was collected using manual eddy current representing aircraft body inspection. The inspectors were EN 4179 certified level 2 or level 3 inspectors. Each inspector completed the inspector using the normal equip- ment in his/her use (GE Mentor, Olympus Nortec 600, GE Phasec 3 or equivalent). A rotating probe was used and sys- tems were calibrated to aluminium reference standard with 0.5 mm artificial defect corresponding to 100% of display. The used frequencywas 500 kHz and scan rotation 1000 rpm. A set of cracked samples representing typical rivet hole configuration were prepared using mechanical fatigue load- ing. The existing crackswere characterized usingmicroscopy to define the true state of the samples. The small plate sam- ples with inspection targets (the rivet holes) were gathered to larger cassettes, which were attached to representative air- craft body location for inspection. Altogether 5 inspectors completed the inspection and reported both signal strength for each inspected hole and their judgement (crack/no crack). Thus the data provided input information for both â versus a and hit/miss analysis. The data set contained altogether 68 cracked locations, 480 inspection locations and 3360 inspec- tion results for 7 inspectors. The crack size distribution is shown in Fig. 3. More detailed description of the inspection set-up is given in [26] (Fig. 1). The second data-set represents a nuclear industry phased array ultrasonic weld inspection case. The use of POD methodology is not as common in nuclear industry as it is in aerospace industry. Thus, representative sample sets con- taining sufficient flaws for hit/miss analysis (or even for â vs. a analysis) are rare. The present data was gathered with simplified online tool described in the following. 123 Journal of Nondestructive Evaluation (2019) 38:89 Page 5 of 13 89 Fig. 2 Measured UT amplitude as a function of crack size. The cracks show significant variation of amplitude even with the small number of real flaws available For the nuclear inspection case, an austenitic stainless steel butt-weld mock-up representing primary circuit piping was available. The mock-up had three cracks (which is obviously far too small number for direct analysis). The mock-up was scanned with mechanized ultrasonic system and collected A- scans recorded in a data file for later analysis. To compensate for the insufficient number of real cracks in the mock-up, the data file was modified to include additional flaws. Also, for easy collection of hit/miss data, an online tool was created, that provided a simplified set of UT analysis tools necessary for this inspection case and provided tools for crack identi- fication and data gathering. The tool can be accessed online at (http://www.trueflaw.com/utpod/). The data file provided by the UT equipment was read and the rawUT-data extracted. The locations of the known cracks were compared with known un-cracked locations, and pure flaw signal was extracted by comparison. The flaw signal was then removed from the originalmeasurement data resulting in apparent clean mock-up data. This pre-processed data, with extracted flaw signals and cleaned, flawless, UT-signal pro- vided the source data for the online tool. The data extraction and details on the used data are provided in [1]. For this methodology, the â versus a relation is predeter- mined for each flaw. Furthermore, the number of different cracks is too small to allow analysis of variability of real â versus a relation was not available. Thus, a direct comparison of â versus a and hit/miss analyses was not possible with this data set. However, with this set-up (and the postulated â vs. a relation), it was possible to study directly the effect of the â to the POD of these inspections. For conventional â versus a analysis, this effect is implicitly assumed to be negligible and thus this allows direct estimation of the possible error caused by this assumption. Even with the very limited number of cracks available in the data set, the natural cracks exhibited significant variation in the maximum amplitude, as compared to the flaw depth. Figure 2 shows the measured UT amplitude as function of crack size. For comparison, similar inspection case was modelled using commercially available CIVA software. Artificial reflectors of different height, tilt, skew and location were introduced to the model and resulting expected signal strengths computed. This provided model-assisted â ver- sus a data for the same inspection case. Data were used to compute a corresponding â versus a MA-POD-curve for the case. Where applicable, the â versus a and hit/miss analyses were completed using Military handbook software [2]. Sev- eral software packages are available for performing these computations, but the MH1823-package is perhaps the most widely applied and thus it was chosen for this study. Where MH1823-package was not sufficient, in-house developed code was used to augment the analysis using the same math- ematical methodology. 3 Results For the data set EC, the actual hit/miss performance was very good and the inspectors detected very small cracks border- ing on the resolution of the microscopy used to confirm the real state of the samples. Some of the inspectors found all the cracks. Consequently, the requirement of having sepa- rate regions and transition in the data was not fulfilled. To alleviate this, and to obtain hit/miss results, a single virtual miss was added to small crack size (10 µm) that, if present, would probably be missed. This addition has little effect in cases where real misses exist, because the real misses have much larger crack size and thus dominate the POD fit. How- ever, for the cases with no misses, this allows the maximum likelihood fit to converge and provides sensible POD values and confidence bounds. In practice, the best estimate a90/95 is near the average between the first hit and the virtual miss and the 95% confidence bound is near the first hit. For inspector B, the data contained an outlying miss of a large crack (760 µm) much greater than the otherwise esti- mated a90/95. This significantly increased the computed a90/95 values and yet gave a90/95 estimates lower than the largest missed cracks. The a90/95 value for the â versus a analysis was not affected as much. The reported hit/miss data and corresponding â values were studied to see if the inspectors followed a consistent â threshold in their hit/miss assessment. All of the inspec- tors reported hits for lower â values than the highest â value reported for no-flaw and thus none of the inspectors based their hit/miss judgement on the â alone. Figure 3 shows the computed hit/miss POD curves and corresponding â versus a POD curves. The linear fits used to obtain the â versus a curves are shown in “Appendix”. Many of the â versus a curves show artificially high POD at zero crack length. In reality, cracks very near to size zero cannot be found with the studied methods, and the true POD 123 89 Page 6 of 13 Journal of Nondestructive Evaluation (2019) 38:89 Fig. 3 â versus a and hit/miss POD curves for the EC data set e d c b a 123 Journal of Nondestructive Evaluation (2019) 38:89 Page 7 of 13 89 Fig. 4 POD curves for the 7 inspectors as function maximum amplitude (â) 123 89 Page 8 of 13 Journal of Nondestructive Evaluation (2019) 38:89 Fig. 4 continued curve is expected to show zero probability of detection for crack size zero. Thus, some of the â versus a curves show unrealistically high POD at small crack sizes. This discrep- ancy is related to the reporting uncertainties of the â values, which result in somewhat unrealistic extrapolated â values at zero crack length. The effect is particularly notable for inspector E, where the reported â versus a values showed marked nonlinearity due to reporting discrepancies and con- sequent unrealistic â versus a fit. For the hit/miss results, inspectors B and E show similarly unrealistic POD at zero crack length. In the case of hit/miss, this is caused by the insufficient amount of misses in combi- nation with hits at greater crack sizes. The pattern displayed by these inspectors is somewhat inconsistentwith the precon- dition of the used hit/miss methodology, that POD increases with increasing crack size. However, for the hit/miss curves, the inconsistency is also reflected in the rapidly widening confidence bounds. For the dataset UT, results from 7 inspectors were avail- able. For these data, the hit/miss POD curve was computed similarly to the EC data set. In this case, clear region of “un- likely to find” and “likely to find” equivalent crack sizes were available and there was a clear transition in between. Thus, no additional conditioning was necessary and the hit/miss analyses were completed with the data “as-is”. For this data the number of real cracks was insufficient to establish traditional â versus a POD estimate. For the gener- ated UT-images, a linear â versus a relation was postulated 123 Journal of Nondestructive Evaluation (2019) 38:89 Page 9 of 13 89 Fig. 5 Model assisted â versus a linear model fit obtained for the same inspection case Fig. 6 Model assisted â versus a POD curve obtained for the UT inspection case for each crack and the images generated accordingly. Conse- quently, the data gives unique opportunity to study hit/miss in terms of â. The variation in hit/miss judgement of the inspectors as function of â represents the missing variation unaccounted for in the â versus a POD analysis. Figure 4 show hit/miss POD curves as function of â computed from the data. The POD variation as function of â shows the effect of the inspector judgement that uses features other than the ampli- tude to assess crack presence (e.g. signal as compared to local variation etc.). It is of interest to know, how this adaptive judgement compares with the simplistic amplitude thresh- old used in the traditional â versus a analysis. This can be obtained by selecting a threshold slightly above the highest noise peak in the data file. For the present data, this corre- sponds to amplitude of 4.9. As can be seen from Fig. 4, most of the inspectors show somewhat better performance than would be obtained with the simplified threshold. However, some of the inspectors missed flaws with amplitude signifi- cantly above the noise. To compare, the same UT inspection case was modelled andmodel-assisted POD curves completed. These are shown 123 89 Page 10 of 13 Journal of Nondestructive Evaluation (2019) 38:89 Table 2 Comparison of a90/95 values obtained from different sources Source a90/95 Inspector hit/miss a 3.7 b 1.1 c 1.9 d 1.6 e 2.4 f 2.5 g 3.7 Average 2.4 Simple threshold hit/miss 3.7 CIVA modelled â versus a 1.6 in Figs. 5 and 6. The â threshold was set to correspond to average simulated amplitude for 1 mm crack. This is to be compared to the noise amplitude in the measured data at the flaw locations,whichwere equivalent to expected signal from crack sizes 0.4–1.0 mm. Finally, with a90/95 values computed from three sources, i.e. inspector hit/miss results from amplitude-varied data, simple threshold hit/miss from amplitude-varied data and â versus a for simulated data, the different method can be compared directly. This comparison is shown in Table 2. 4 Discussion Based on the result of this study, both the â versus a and the hit/miss methodologies present a valid and standardized way to estimate POD curves and the a90/95 values. Neverthe- less, the results display some discrepancies. The correlation between â versus a and hit/miss for the EC dataset is shown in Fig. 7. There is overall correlation, but also significant varia- tion. For the very small a90/95 sizes, where inspectors found all or almost all cracks, the hit/miss analysis shows smaller (better) a90/95 values indicating, that the inspectors included factors other than the signal strength â for their judgement, e.g. signal stability in repeated measurements were cited. Conversely, for the larger a90/95 values, the â versus a results show smaller a90/95 values. This can be attributed to the â versus a methodology failing to account sufficiently to the larger missed cracks in the data. Furthermore, the â versus a is sensitive to variation in the â versus a relation, which in this case was also affected by inspector reporting practices. Values of â were read from the equipment screen and there may have been differences of accuracy between inspectors in this respect. This accuracy did not affect the inspector perfor- mance (as shown in hit/miss), but it did affect the confidence bounds obtained from â versus a and thus measured perfor- Fig. 7 Comparison of hit/miss and â versus a POD a90/95 values for the EC data-set. Dashed line shows the expected line, where results from both methodologies concur. The solid line shows regression line, which shows poor correlation (R2  0.36) mance. On one occasion, the reported â versus a relation showed significant non-linearity and the reliability of the â versus a was questionable, despite this non-linearity having no effect on actual inspector performance. In conclusion, the hit/miss method seems to better describe the present manual inspection case and caution is advised if using â versus a for such cases. In addition, it seems that the â overestimates the a90/95 for small crack sizes and underestimates it with larger a90/95 values, as compared to the hit/miss. For the UT dataset, direct observation on the effect of â to inspector judgement was obtainable. Here, the interest is mainly in establishing whether the inspector judgement pro- vides superior results to the simplistic amplitude threshold. As revealed by comparison in Table 2, none of the inspectors performedworse than a simple amplitude threshold and some inspectors showed quite significantly better performance. It is also noteworthy, that evenwith the small number of inspec- tors, the variance between inspectors is significant. Thus, the results indicate, that a simple amplitude threshold, properly applied, will underestimate the expected performance. Also, an amplitude threshold will not represent the true variance to be expected from inspections because it fails to capture the inspector variability in hit/miss assessment. The considerable overlap of hits and misses Fig. 4, as well as variation between inspectors indicates, that inspector judgement may be a significant source of uncertainty even when the variation in signal amplitude is accounted for. This uncertainty would not be addressed by an â versus a analysis even when the variation of measured signal strengths due to inspection conditions were taken into account as, e.g. by Bato et al. [19]. The issues seen in the hit/miss analysis, i.e. convergence problems with data-sets with insufficient misses and unre- alistic POD curves for data-sets with unclear separation of misses, primarily stem from the data not fitting the assump- tion of increasing POD with increasing flaw size. Thus, 123 Journal of Nondestructive Evaluation (2019) 38:89 Page 11 of 13 89 using alternatemethodology, such as theDOEPOD([23–25]) would solve these issues, albeit with an increased number of samples required. The modelled â versus a shows significantly better a90/95 than would be obtained with simple amplitude threshold and hit/miss analysis. The a90/95 obtained from modelling is highly dependent on the variance provided by the mod- elled cases and on the chosen threshold amplitude. As noted before, the modelled â response can only incorporate vari- ance from directly modelled flaw characteristics, such as orientation and thus fails to include variation in natural flaw characteristics and/or microstructural changes. This may make the MAPOD values overly optimistic. More impor- tantly, the amplitude threshold in present case was chosen to represent the perceivednoise level as is typical for such analy- sis. Further comparison with the experimental data revealed, that the selection was overly optimistic. Furthermore, the choice of amplitude threshold is critical determinant of the resulting a90/95 values and thus, proper choice of threshold is critical to the reliability of the whole assessment. Unfor- tunately, no guideline can be given for proper threshold selection for the present case, since the inspectors do not follow an amplitude threshold consistently. 5 Conclusions The following conclusions can be drawn from the study: • the two standard methodologies (â vs. a and hit/miss) may give significantly different results, if true hit/miss decision is to be based on inspector judgement (and not automated signal threshold), • true inspector hit/miss performance shows significant vari- ance that is not attributable to signal amplitude • MAPOD, as performed for the present study, is not able to model the inspector performance due to lack of repre- sentative amplitude threshold and difficulties in capturing true signal variance. Consequently, the â versus a approach can only be recom- mended for inspections, where a consistent signal threshold is enforced, e.g. by an automated system. Similarly,MAPOD can be recommended only where, in addition to the enforced signal threshold, the modelled flaw variance can be well jus- tified. In general, hit/miss approach is seen to be more robust and thus preferable, albeit may also exhibit issues for insuf- ficient data. Acknowledgements Open access funding provided by Aalto Univer- sity. The EC data set was arranged and gathered by Patria Aviation (Jouni Pirtola) and The Finnish Defence Forces (Ari Kivistö). Their support is gratefully acknowledged. Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecomm ons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. 123 89 Page 12 of 13 Journal of Nondestructive Evaluation (2019) 38:89 Appendix: â Versus a Linear Fit Results for the EC-Data A B C D E F G 123 Journal of Nondestructive Evaluation (2019) 38:89 Page 13 of 13 89 References 1. Charles Annis, P.E.: Statistical best-practices for building Proba- bility of Detection\n(POD) models. R package mh1823, version 4.3.2 (2016). http://StatisticalEngineering.com/mh1823/ 2. Underhill, P.R., Krause, T.W.: Eddy current analysis of mid-bore and corner cracks in bolt holes. NDT&E Int. 44, 513–518 (2011). https://doi.org/10.1016/j.ndteint.2011.05.007 3. Rummel,W.D.: Nondestructive evaluation—a critical part of struc- tural integrity. Procedia Eng. 86, 375–383 (2014). https://doi.org/ 10.1016/j.proeng.2014.11.051 4. Garza, J., Millwater, H.: Sensitivity of the probability of failure to probability of detection curve regions. Int. J. Press. Vessels Pip. 141, 26–39 (2016). https://doi.org/10.1016/j.ijpvp.2016.03.012 5. Carboni, M., Cantini, S.: A model assisted probability of detection approach for ultrasonic inspection of railway axles. In: 18th World Conference on Nondestructive Testing, 16–20 April 2012 (2012) 6. Carboni, M., Cantini, S.: Advanced ultrasonic “Probability of Detection” curves for designing in-service inspection intervals. Int. J. Fatigue 86, 77–87 (2016). https://doi.org/10.1016/j.ijfatigue. 2015.07.018 7. Gandossi, L.,Annis,C.: Probability ofDetectionCurves: Statistical Best-Practices. ENIQ report nr. 41, vol EUR 24429 EN. European Commission (2010) 8. ASTM: Standard Practice for Probability of Detection Analysis for Hit/Miss Data, vol. E2862-12. West Conshohocken, ASTM International (2012) 9. ASTM: Standard Practice for Probability of Detection Analysis for â Versus a Data, vol. ASTM-E3023. West Conshohocken, ASTM International (2015) 10. Berens,A.:NDE reliability data analysis. In: Lampman, S.R., Zorc, T.B. (eds.) ASM Metals Handbook, 9th edn. ASM, Ohio (1989) 11. Syed Akbar Ali, M., Kumar, A., Rao, P., Tammana, J., Balasubra- maniam,K.,Rajagopal, P.:Bayesian synthesis for simulation-based generation of probability of detection (PoD) curves.Ultrasonics 84, 210–222 (2018). https://doi.org/10.1016/j.ultras.2017.11.004 12. SyedAkbarAli,M.S., Rajagopal, P.: Probability of detection (PoD) curves based onweibull statistics. J. Nondestr. Eval. (2018). https:// doi.org/10.1007/s10921-018-0468-2 13. Le Gratiet, L., Iooss, B., Blatman, G., Browne, T., Cordeiro, S., Goursaud, B.: Model assisted probability of detection curves: new statistical tools and progressive methodology. J. Nondestr. Eval. (2017). https://doi.org/10.1007/s10921-016-0387-z 14. Annis, C., Gandossi, L., Martin, O.: Optimal sample size for prob- ability of detection curves. Nucl. Eng. Des. 262, 98–105 (2013). https://doi.org/10.1016/j.nucengdes.2013.03.059 15. Knopp, J.G., Zeng,L.,Aldrin, J.: Considerations for statistical anal- ysis of nondestructive evaluation data: hit/miss analysis. E J. Adv. Maint. 4(3), 105–115 (2012) 16. Harding, C.A., Hugo, G.R.: Statistical analysis of probability of detection hit/miss data for small data sets. In: AIP Conference Pro- ceedings. AIP, pp 1838–1845 (2003) 17. McGrath, B.: Programme for the assessment of NDT in industry. PANI 3. Serco Assurance, UK (2008) 18. D’Agostino, A., Morrow, S., Franklin, C., Hughes, N.: Review of Human Factors Research in Nondestructive Examination. NRC, Rockville (2017) 19. Bato, M.R., Hor, A., Rautureau, A., Bes, C.: Impact of human and environmental factors on the probability of detection during NDT control by eddy currents.Measurement 133, 222–232 (2019). https://doi.org/10.1016/j.measurement.2018.10.008 20. Ben Abdessalem, A., Jenson, F., Calmon, P.: Quantifying uncer- tainty in parameter estimates of ultrasonic inspection system using Bayesian computational framework. Mech. Syst. Signal Process. 109, 89–110 (2018). https://doi.org/10.1016/j.ymssp.2018.02.037 21. Yusa, N., Chen, W., Hashizume, H.: Demonstration of probability of detection taking consideration of both the length and the depth of a flaw explicitly. NDT E Int. 81, 1–8 (2016). https://doi.org/10. 1016/j.ndteint.2016.03.001 22. Seuaciuc-Osorio, T., Ammirato, F.: Materials Reliability Program: Development of Probability of Detection Curves for Ultrasonic Examination of Dissimilar Metal Welds MRP-262, 3rd edn. EPRI, Charlotte (2017) 23. Generazio, E.R.: Design of experiments for validating probability of detection capability of NDT systems and for qualification of inspectors. Mater. Eval. 67(6), 730–738 (2009) 24. Generazio, E.R.: Validating design of experiments for determining probability of detection capability for fracture critical applications. Mater. Eval. 69(12), 1399–1407 (2011) 25. Generazio, E.R.: Directed Design of Experiments for Validating Probability of Detection Capability of NDE Systems (DOEPOD) (2015) 26. Virkkunen, M., Ylitalo, M.: Practical experiences in POD determi- nation for airframe ET inspection. In: 2016/11/3 (2016) Publisher’s Note Springer Nature remains neutral with regard to juris- dictional claims in published maps and institutional affiliations. 123

7th European-American Workshop on Reliability of NDE 1 License: http://creativecommons.org/licenses/by/3.0/ Effect of Feedback and Variation on Inspection Reliability Iikka VIRKKUNEN 1, Jonne HAAPALAINEN 2, Suvi PAPULA 3, Teemu SARIKKA 3, Juha KOTAMIES 4, Hannu HÄNNINEN 1 Trueflaw Ltd., Espoo, Finland 2 VTT, Espoo, Finland 3 Aalto University, Espoo, Finland 4 Metropolia University of Applied Sciences, Helsinki, Finland 5 Aalto University, Contact e-mail: iikka@trueflaw.com Abstract. The reliability of Non-destructive testing (NDT) is an on-going challenge. The consequences of failed inspections can be dire, and thus the requirements for NDT reliability are very high. The work is technically demanding and requires skilled use of the available equipment and keen judgement to properly discern flaw signals from noise. Somewhat paradoxically, the work is also very tedious and repetitive. Most of the inspected targets do not contain any flaws but the inspectors need to be constantly alert for the possibility. The recent studies in human factors have brought advances in (among other things) improved readability of inspection procedures and procedures of reviews and redundant inspections widely used in order to improve overall inspection reliability. In present paper, the effect of feedback and variation on inspector performance is studied. To test this, a small empirical study was completed. An online tool was created with simplified UT set-up: B-scan image of the data and software gain control and tools to indicate cracks with point and click. The system generates random flawed data images on the fly. The user than analyses the images indicates found flaws by clicking them. After 150 images have been analysed (many of them without flaws), the system uses the provided hits and misses to compute a POD curve and confidence bounds using standard (ASTM E2862) techniques. Additional "learning" version of the tool was created. In this "learning mode", after user requests next image, the system shows results of the current images (i.e. hits, misses and false calls in the current image). This set-up thus provides the inspector with direct feedback of his success and should thus better facilitate learning this particular inspection task. The tool was presented to small group of inspector in level-III inspector training and results gathered from trainees both before and after training (and with and without feedback). The results from this small group of inspectors indicated, that direct feedback on achieved reliability can quickly improve POD values. However, the study group was small and thus the results need further investigation and confirmation. 2 Introduction The reliability of Non-destructive testing (NDT) is an on-going challenge. The consequences of failed inspections can be dire, and thus the requirements for NDT reliability are very high. The work is technically demanding and requires skilled use of the available equipment and keen judgement to properly discern flaw signals from noise. At the same time, the work is also very tedious and repetitive. Most of the inspected targets do not contain any flaws but the inspectors need to be constantly alert for the possibility. Finally, there's variation and challenges in the working conditions of the actual inspection: the inspections are often completed in awkward positions, uncomfortable high temperatures and the inspection target may not offer sufficient coupling with the inspection instruments. Significant effort has been put into securing the performance and reliability of these inspections. Firstly, the inspections are codified into detailed procedures, for the inspector to follow. These written procedures enhance the repeatability and consistency over different inspectors and conditions. Numerous standards now exist for basic inspections [e.g. 1,2]. In addition, more detailed (and comprehensive) procedures for especially demanding or critical inspections are used, e.g. in the aerospace industry and the nuclear industry [e.g. 3]. The use of common standard practices and procedures greatly improves both the reliability and predictability of the inspection. Furthermore, it provides a common understanding of generic inspection capabilities and expected performance. Fixed procedures are also necessary precondition for meaningful measurement of expected performance. However, as has been shown in various round-robin excercises [4], a common procedure is not sufficient to guarantee the needed performance. In the aerospace industry, rather conservative "default" performance levels have been adopted, that may be applied in absence of additional evidence [5]. In addition, methodology for quantitative assessment of actual performance has been developed [6]. In the nuclear industry, NDT qualification and performance demonstration were developed and are now required around the world for nuclear inspections [3,7]. The central idea in these is, that the performance of the procedure and then the performance of individual inspectors applying this procedure are verified with combination of technical justification and practical trials, where the inspection procedure is applied on test samples with known flaws and its performance evaluated. These performance demonstrations have significantly improved the expected inspection performance. However, some high-profile failures in qualification inspection [8] have shown, that the actual inspection performance of a qualified inspector using procedure does not necessarily reflect the performance apparently demonstrated during qualification. These cases have led to increasing emphasis to "human factors" of inspection performance, i.e. factors affecting the inspection procedure but outside the traditional scope of performance of the technical inspection system and procedure [9, 10]. To account for the various separately studied aspects of inspection reliability, an integraded model has been proposed: "holistic approach to inspection reliability" [10] which offers a generic framework or synthesis for addressing various aspects of inspection reliability and their links and interdependencies. In the holistic approach, the inspection reliability is divided into four elements, as shown in figure 1. 3 Fig. 1. Elements of inspection qualification according to the holistic approach for inspection qualification [10]. The holistic approach has brought significant contribution to the field by making explicit the interconnectedness of various aspects of inspection reliability and the fallacy of estimating actual inspection performance based on demonstrated technical performance alone. However, the model, in its current form, has some shortcomings as well. Since the division to various components is done along empirical lines, all the components are deeply interconnected and addressing some aspect will potentially affect all the other components as well. This interconnectedness will make it difficult to use the approach for making quantitative estimates of inspection performance (even if all the components could be quantified in isolation, which at present is not possible, the combined performance is not apparent). Also, recent findings [9] indicate, that the inspectors do not necessarily learn or improve in terms of reliability (since the procedure is very detailed and kept constant, this is almost by design). Furthermore, there's some evident contradictions inherent in the current approach. The procedures are made very detailed and adherence to the procedure is required of inspectors. The underlying assumption here is, that if procedure fully defines the inspection and the inspectors fully follow the procedure, their performance is constant and predictable. In contradiction, it is well known that inspector performances vary significantly, and this is why personnel qualifications are required in addition to procedure qualifications in the nuclear sector. The procedures expect certain explicit (and implicit) preconditions to be met (e.g. surface quality of the inspection target, accessibility etc.). In practice, these are often violated, and the inspectors are put in contradictory position: they need to adhere to the procedure, which is impossible or counterproductive and at the same time obtain the performance attributed to the procedure which is impossible following the procedure. It is unreasonable to expect that all the future inspection conditions could be predetermined at the time of the writing of the procedure (at least for many inspections), and thus the predetermined nature of the procedure conflicts with the varying inspection targets. (In the holistic framework this is expressly included in "application parameters", but the conflict with the procedural approach is not stated.) Diagnostic system Intristic capability Human factors Application parameters Organizational context 4 In summary, the recent advances in the holistic approach of inspection reliability have brought into light several longstanding handicaps of the current paradigm for inspection qualification. In essence, the paradigm states, that inspection reliability is the property of the inspection system or procedure, it may be measured in qualification or other ways and it may be reduced or compromised in actual inspections by unforeseen conditions (application parameters) or human factors. As stated above, this paradigm leaves both learning and adaptation outside the scope of the discussion (in practice, to be addressed when writing the next procedure or designing the next system). Excluding learning and adaptation from using a set procedure induces combination of conflicting requirements to the inspectors and necessitates various "corrective" elements to move from the "system performance" to the "actual performance" as shown in the holistic approach. 1. Learning-centered paradigm for inspection performance To address the limitations of the holistic approach, an alternate viewpoint on NDE reliability is developed. This is not meant to replace the holistic approach, but to complement it and to offer alternate approach for improving NDE reliability. As a starting point, we take the viewpoint, that the expected NDE performance is dependent on the inspector (in the context of the available tools, i.e. inspection system, procedure and inspection target). Different inspectors may show different performance using the same procedure and inspection target, as is well known. Since procedure is the same, these differences result from differences in the inspector proficiency. To obtain optimal performance from different inspectors, the inspector performance and/or the procedure needs to be adapted to the current situation. Thus, equivalently, the inspector performance can be seen as result of inspector learning. In summary, the expected inspector performance is a function of the available tools (procedures) and inspector adaptation (both to the tools and the task at hand). Successful adaptation is the result of inspector learning. Thus, primary way to improve inspection perfor- mance is to improve inspector learning to adapt procedure to existing conditions. Likewise, the primary way to monitor and control the expected performance is to test adaptation by testing performance. Learning, in this context, is not necessarily beneficial. It is possible for inspectors to learn practices that lower the performance. Also, it does not necessitate conscious deliberation. It is expected that learning is continuous and on-going during the working life of the inspector. Thus, the expected performance will vary continuously and may improve or worsen. In fact, re-qualifications often fail which is notorious evidence of changing performance after qualification. 2. The necessary conditions for learning NDT performance is now seen primarily as a result of inspector learning (i.e. acquired ability to adapt tools and behaviour to the task at hand for optimal performance). Improving performance is thus primarily matter of improving the conditions for inspector learning. Likewise, estimating expected performance is primarily matter of testing obtained ability in different settings. Consequently, it is of interest to study the necessary preconditions of learning and how they are present in current NDT inspection qualification framework. The necessary conditions for learning can be summarized as in Figure 2 [11,12]. 5 Fig. 2. Necessary conditions for learning adapted to the NDE context. By comparing the necessary condition of learning with the holistic model of NDE reliability we note, that the models overlap significantly. Both include explicitly the significance of the underlying physical phenomenon (Intristic capability and Underlying model). The "Application parameters" corresponds with the "Variation" in the learning model. Whereas the application parameters in the holistic model are seen to threaten NDE reliability, variation is both challenge and a prerequisite for learning the necessary adaptation. The biggest difference is with the "human factor" and "organizational context" parts, which are implicit in the learning model. Feedback, a necessary part in the learning model is not represented in the holistic model. Thus, the role of feedback may be under-represented in the current viewpoint. Available feedback for inspectors is scarce within the current setting. The qualification exercise gives some limited feedback on pass/fail basis. Training also gives feedback on the signal-response of the used method from training samples. However, feedback on reliability is especially scarce. 3. Experimental study on the effect of feedback and variation on inspection reliability As noted, the most significant difference to prevailing model is the significance of feedback and variation to inspection reliability. To test this hypothesis, a small empirical study was completed. A typical nuclear industry mock-up component (stainless steel tube butt-weld) was scanned with phased-array ultrasonic inspection set-up. The mock-up contained three artificially induced thermal fatigue cracks. The acquired data was analysed and the flaw signal was extracted using the eFlaw technology [13]. Flaw indications were then removed from the data to provide a clean defect-free data, where the extracted flaws could be re- introduced at various locations at will. An online tool was created to provide facilities for the inspectors to identify possible flaws. The online tool includes a simplified UT set-up with B- scan image of the data and soft-ware gain control and tools to indicate cracks with point and click (Figure 3.). The system generates random flawed data images on the fly by re- introducing extracted flaw signals at different locations in the data (also, the data is rotated and flipped to disable direct comparison of background noise and using the difference as hint for flaw detection). The user than analyses the image (possibly changes the gain to get more confidence on the results) and indicates found flaws by clicking them. After the image is analysed, the user requests next image by clicking a button. After 150 images have been analysed (many of them without flaws), the system uses the provided hits and misses to Underlying model FeedbackVariation 6 compute a POD curve and confidence bounds. Since the mock-up contained only three cracks, the data provided limited variation of real cracks to the inspectors. To augment this, flaw amplitudes were artificially altered to produce a wider set of equivalent flaw sizes. For this study, additional "learning" version of the tool was created. In this "learning mode", after user requests next image, the system shows results of the current images (i.e. hits, misses and false calls in the current image). This set-up thus provides the inspector with direct feedback of his success and should thus better facilitate learning this particular inspection task. Fig. 3. The simplified inspection view used to gather hit/miss data in feedback mode. The green rectangles show true crack locations and equivalent sizes. Blue rectangle shows user-indicated crack location. A simple amplitude threshold just above the highest noise peak in the data was used to generate a POD curve. This curve represents a performance attainable via a simple amplitude rule, optimized for this data and provides a useful reference to compare student performance against. This POD curve is shown in Figure 4. The curve shows a90/95 value of 3.75 mm. 7 Fig. 4. POD curve generated from the automated system with a fixed amplitude threshold set just above the highest noise-peak in the data. The set-up was tested during a level-III inspector training course. 9 students available were randomly divided in two groups (A and B). In the beginning of the course, all inspectors used the tool to get a base-line result before the course. After this baseline was recorded, group A continued training with the system in the normal mode during the week-long course. Group B had similar amount of training using the "learning mode". Finally, at the end of the course, final POD curve was obtained from all the inspectors. The final number of full training sessions varied as the students took different time to do the tests. The students were not penalized for false calls. However, one person made so many false calls (1290), that they effectively obscured any information about true performance and thus this person was excluded from further study. The final number of results are summarized in Table 1. Table 1. Final number of full POD exercises Inspector Version Number of full results A B C D E F G H Non-learn mode only Non-learn mode only Non-learn mode only Non-learn mode only Learn-mode Learn-mode Learn-mode Learn-mode 1 1 1 2 3 3 3 4 4. Results The first-trial results are shown in Figure 5. Most of the inspectors reached better than reference (Figure 4) performance. False call rates varied significantly, and there's no clear correlation between false call rates and a90/95. In Figure 6, the final results are presented, after learning trials and full POD trials as listed in Table 1. The learning-mode students show 8 significant improvement with one significant outlier. A more detailed analysis of the POD results for this inspector revealed, that the learning significantly improved the small hits. However, the number of "big misses" and the largest missed cracks were not decreased to the same degree, and thus the improvement did not translate into improved a90/95 results. Fig. 5. First trial POD results from the students. Fig. 6. Final trial POD results from the students. The learning-mode students show significant improvement with one significant outlier. 4. Conclusions The limited evidence gathered during this study indicates, that direct feedback helps inspectors improve reliability, as measured by the POD curve and a90/95 values. This is to be compared with previous evidence showing no improvement on reliability during normal Normal mode Learn mode F al se c al ls 0 20 40 60 80 100 120 140 160 180 200 a90/95 0 1 2 3 4 5 6 7 8 9 10 Normal mode Learn mode F al se c al ls 0 20 40 60 80 100 120 140 160 180 200 a90/95 0 1 2 3 4 5 6 7 8 9 10 9 work experience [9]. The results also indicate, that virtual flaws can be used successfully to give more direct feedback than previously available. References [1] SFS standardi ultraäänitarkastukselle [2] SFS standardi pyörrevirtastandardille [3] Anon. 2007. The European methodology for qualification of non-destructive testing, Third Issue. ENIQ Report nr. 31, EUR 22906 EN, ISSN 1018-5593. [4] Lemaintre, P., Koblé, T. D. & Doctor, S. R. 1996. "Summary of the PISC round robin test results on wrought and cast austenitic steel weldments, part III: cast-to-cast capability study". International Journal of Pressure Vessels and Piping (69) pp. 33-44. [5] Anon. 2008. Nondestructive Evaluation Requirements for Fracture-Critical Metallic Components. NASA- STD-5009. National Aeronautics and Space Administration Washington, DC 20546-000. 28 pp. [6] Anon. 2009. Nondestructive Evaluation System Reliability Assessment. Department of Defence Handbook. MIL-HDBK-1823A. 171 p. [7] Cowfer, C.D. 1991. Basis / Background for ASME Code Section XI proposed Appendix VIII: Ultrasonic examination performance demonstration. Nuclear Engineering and Design, 131, pp. 313 - 317. [8] Anderson, M., Diaz, A. & Doctor, S. 2012. Evaluation of Manual Ultrasonic Examinations Applied to Detect Flaws in Primary System Dissimilar Metal Welds at North Anna Power Station. PNNL-21546, Pacific Northwest National Laboratory, Richland, Washington. ADAMS Accession No. ML12200A216. [9] Bertovic, Marija. 2016. Human Factors in Non-Destructive Testing (NDT): Risks and Challenges of Mechanised NDT. Bundesanstalt für Materialforschungn und -prüfung (BAM), Berlin, Germany. ISBN 978-3- 9817502-7-0. [10] Müller, C., Bertovic, M., Pavlovic, M., Kantzler, D., Ewert, U, Pitkänen, J & Ronneteg, U. 2013. Paradigm shift in the Holistic Evaluation of the Reliability fo NDE Systems. Materials Testing, 55 (4), pp. 261-269. [11] Marton, F. 2015. Necessary Conditions for Learning. Taylor & Francis, New Your, U.S.A. ISBN 978-1- 315-81687-6. [12] Sterman [13] Koskinen, T., Virkkunen, I., Papula, S., Sarikka, T. & Haapalainen, J. 2017. Producing a POD curve with emulated signal response data. Insight. Accepted for publication.

8th International Symposium on NDT in Aerospace, November 3-5, 2016 Practical Experiences in POD Determination for Airframe ET Inspection Virkkunen, I.1 and Ylitalo, M.2 1 Trueflaw Ltd., e-mail: iikka@trueflaw.com 2Patria Aviation Oy Abstract Evaluation of NDT reliability has received increasing emphasis in recent times. In particular, quantifying the probability of detection (POD) attained in routine inspections have become more widespread. Although there are good guidelines and standards for POD determination, the process is still far from trivial. Various choices made during the experimental set-up may have significant effect on the results. Also, the cracked samples used are often limited necessitating various compromises in the analysis. Patria performed a set of POD studies for eddy-current inspections performed on various parts of typical metal airframe. The project included manufacturing of cracked samples, organizing the inspection of these samples and final analysis of the results. Several inspectors from different organizations took part in the exercise. The project was done in collaboration with Finnish and international partners. The data showed various unlikely events (small hits, big misses and poor separation), which necessitated adjustment for the standard methodology. Contrary to expectation, the false call rate did not show significant correlation with the inspection performance. When the â vs. a and hit/miss analyses could be directly compared, they showed surprisingly poor correlation and caution is advised in using â vs. a analysis for manual inspections such as the ones shown here. Keywords: Probability of detection (POD), Eddy current inspection (ET) 1. Introduction The best practices of estimating probability of detection (POD) in non-destructive evaluation (NDE) are now well established. The venerable MIL-HDBK-1823A (most recent release from 2009) [1] is used extensively in the aerospace industry and is now finding increasing use also in other areas, like the rail industry and even nuclear industry. The methods have recently been standardized by ASTM (ASTM-E2862 [2] and ASTM-E3023 [3]) and these standards are congruent with the current MIL-HDBK methodology. Despite the now standardized methodology and significant tradition in POD determination, the process still offers some practical challenges. The requirements for cracked test pieces are sometimes difficult or costly to fulfill, the statistical analysis may prove demanding and, perhaps most importantly, justifying that the various assumptions behind the methodology are fulfilled to sufficient extend may prove challenging. The standard practice offers two variant of POD curve estimation, the â vs. a approach and the hit/miss approach. The â vs. a approach models, in simple terms, the NDE reliability as kind of measurement system problem, where the quantity to be measured (crack size a) give rise to measured signal (â) proportional to the measured quantity and the task is to determine the possible existence of the signal with decreasing a (and thus decreasing â). The system has noise both related to the signal and independent of the signal. That is, â varies due to factors other than a, like crack orientation and tortuosity, which results in noisy â vs. a relation. In addition, there's noise, that is independent of a, e.g. electric noise on the signal path. Thus, the task is to find a decision threshold (â value), that minimizes false calls from the noise and, in parallel, maximizes the number of cracks found (i.e. cracks with â above the threshold), given the variation in the â-vs-a relation. This is done by fitting a linear function through the â-vs-a data, computing prediction intervals to take the noise and statistical uncertainty into account. The resulting best-fit and confidence limit lines are then compared to the set detection threshold and the corresponding POD curves computed. For input, the â vs. a analysis requires a set of representative flaws (at least 40) and measurements of signal strength â and corresponding crack size a. In addition, noise independent of crack size needs to be evaluated either with additional measurements of crack-free samples or in connection with the same sample set measurement. The hit/miss approach, in contrast, does not deal with signal values, but estimates the POD curve based on binary results, that is hits (correctly found cracks) and misses (cracks not found in the inspection). Because the data contains less information (regarding the correlation between crack size and signal strength or "ease of detection") more samples are needed for reliable POD determination. The POD curve is solved using generalized linear model and a chosen link function (typically logit), that gives the shape of the POD curve using maximum likelihood fit to the data. The corresponding confidence limits are then obtained by the likelihood-ratio method, where a likelihood surface near the maximum likelihood value is interrogated, POD curves with likelihoods corresponding to the chosen confidence interval computed and the lower (and upper) limit curves resolved. For input, the hit/miss analysis requires a set of representative flaws (at least 60) and hit/miss results for each crack. In addition, the hit/miss results should exhibit a range with "unlikely to find" cracks, a range with "likely to find" cracks and transition in between. Otherwise, the model does not describe the data and, while a fit may in some cases be obtained, it does not describe the underlying probability of detection. In both cases, the basic assumptions underlying both POD models should be fulfilled: the POD should be an increasing function of the crack size and should reach 100% with sufficient crack size. If the data contains signs of violation of these assumptions (e.g. a miss with big crack length indicating that the POD does not reach 100% even with large crack size), the standard models are not applicable and alternate model must be sought. Patria performed a set of POD studies for eddy-current inspections performed on various parts of typical metal airframe. The project included manufacturing of cracked samples, organizing the inspection of these samples and final analysis of the results. Some of the cracked samples were provided by collaborating organizations. Several inspectors took part in the exercise from different organizations. The project was done in collaboration with Finnish and international partners. This study provided an opportunity to study various practical aspects of the POD determination process and to compare POD results obtained in different settings. 2. Materials and methods The study was divided in three cases, as described in Table 1. Each case had different set of cracked samples and was completed in one "go". The analyses were completed with the openly available mh1823 software package [4]. Table 1. Summary of studied inspection cases. Case Description Cracks Inspectors A Typical fillet 49 7 B Rivet hole 58 11 C Rivet hole 68 7 3. Results and discussion For each case, the practical difficulties obtained were somewhat different and are analyzed on case-by-case basis below. 3.1 Case A For case A a hit/miss analysis was completed. The sample set contained somewhat smaller number of cracks (49) than required by the MIL-HDBK [1]. However, the crack sizes were well distributed in terms of hits and misses and showed no adverse behavior in the statistical analysis. Thus, hit/miss analysis was deemed appropriate for the data. The lack of sufficient cracked samples increases the uncertainty of the analysis and thus the reported a90/95 values are expected to be greater than what would be obtained with additional number of samples. The cracks used were produced using mechanical fatigue. Produced cracks were inspected using automated ET and selected cracks were destructively examined. A typical obtained POD curve is shown in Figure 1. The data shows good separation between crack sizes likely to be missed, a transition zone and crack sizes likely to be found. On several cases, the inspectors also found some very small cracks. This unlikely hit significantly changed the confidence bounds and, paradoxically, increased the size of the computed a90/95. Such curves were re-analyzed with the small hit changed to miss (thus "worsening" the inspection behavior) and smaller a90/95 values were obtained. Figure 1. Typical POD curve for case A. The curve includes an unlikely small hit, which widens the confidence bounds and paradoxically decreases the measured performance. With this hit changed to miss, the a90/95 value decreased by 14%. The inspectors also showed strong variation in false call rates. Interestingly, the false call rate was not correlated with inspection performance (as measured by the a90/95). Figure 2. shows the obtained a90/95 values in comparison to the false call rate of the inspectors. Figure 2. False call rate as a function of obtained a90/95 values. Negative correlation would be expected, but there's no clear correlation observed. The overall performance was not quite as good as was hoped. This was attributed partly to significant time pressure during the inspection. 3.2 Case B For case B a hit/miss analysis was completed. The sample set contained somewhat smaller number of cracks (58) than required by the MIL-HDBK [1]. However, the crack sizes were well distributed in terms of hits and misses and showed no adverse behavior in the statistical analysis. Thus, hit/miss analysis was deemed appropriate for the data. A typical obtained POD curve is shown in Figure 3. As in the case A, the data shows good separation between crack sizes likely to be missed, a transition zone and crack sizes likely to be found. The overall results were significantly better than for Case A. Figure 3. Typical POD curve for case B. In case of one inspector, a single large (or larger than other) crack was missed. This unlikely event significantly changed the maximum likelihood curve and widened the confidence bounds. The resulting a90/95 value was quite large for this data set, but with the single big miss changed F al se c al ls a90/95 to hit, decreased by 52%. Furthermore, with the one outlier, the confidence bounds do not seem to sufficiently cover the variability: the biggest miss is still significantly over the computed a90/95 value and would be highly unlikely, according to the computed POD curve. Thus, the single big miss effectively calls to question the applicability of the POD model used. A comparison is shown in Figure 4. Figure 4. Atypical POD curve, where a single big miss has significantly altered the obtained POD curve. For comparison, curves computed with the same data except the biggest miss changed to hit are shown in light-blue. With the modified data, the computed a90/95 decreased by 52%. Again, the inspectors also showed strong variation in false call rates and the false call rate was not correlated with inspection performance (as measured by the a90/95). Figure 5. shows the obtained a90/95 values in comparison to the false call rate of the inspectors. Figure 5. False call rate as a function of obtained a90/95 values. Negative correlation would be expected, but there's no clear correlation observed. F al se c al ls a90/95 3.3 Case C For case C both â vs a and a hit/miss analysis were completed. The sample set contained 68 cracks. However, the inspection performance in this case was better than expected and most inpsectors found most of the cracks and some inspectors found all of the cracks. Paradoxically, this good performance caused numerical difficulties with the POD curve determination (for the hit/miss analysis) since now the crack sizes were not well distributed in terms of hits and misses, despite large population of small crack sizes. To obtain maximum likelihood estimates for the hit/miss analysis, a single small miss was added to the data. (Had there been such a small crack in the samples, it would have likely been missed by both the inspection under investigation and the previous after-manufacturing inspection. Thus assuming such a miss is not unrealistic.) This single miss allowed the maximum likelihood estimate to converge and provided a90/95 results. In practice, the induced single miss forced the POD curve to steep curve between the artificial miss and the smallest found crack and the lower-limit estimate near the second-smallest crack missed. Thus the shape of the POD curve does not carry much information, but the obtained a90/95 results are justifiable. A typical POD curve is shown in Figure 6. Figure 6. Typical hit/miss POD curve for case C. For case C, results enabled both â vs. a and hit/miss analyses to be completed and allowed direct comparison between the two methodologies. Figure 7. shows typical POD curve obtained from â vs. a analysis. Figure 7. Typical â vs. a POD curve for case C.(Same inspector as for Figure 6 for direct comparability.) To compare the POD values obtained from â vs. a and hit/miss analysis, the values were compared inspector-by-inspector. The results are shown in Figure 8. Although both results were obtained with standard methodology, they present significant variation and the overall correlation is not very good. For the very small a90/95 sizes (where inspectors found all or almost all the cracks), the hit/miss analysis shows smaller (better) a90/95 values indicating, that the inspectors included factors other than the signal strength â for their judgement (e.g. signal stability in repeated measurements were sited). Conversely, for the larger a90/95 values, the â vs. a results show smaller a90/95 values. This can be attributed to the â vs. a methodology failing to account sufficiently to the larger missed cracks in the data. Furthermore, the â vs. a is sensitive to variation in the â vs. a relation, which in this case was also affected by inspector reporting practices. Values of â were read from the equipment screen and there may have been differences of accuracy between inspectors in this respect. This accuracy did not affect the inspector performance (as shown in hit/miss), but it did affect the confidence bounds obtained from â vs. a and thus measured performance. On one occasion, the reported â vs. a relation showed significant non-linearity and the reliability of the â vs. a was questionable (despite this non- linearity having no effect on actual inspector performance). In conclusion, the hit/miss method seems to better describe the present manual inspection case and caution is advised if using â vs. a for such cases. Figure 8. Comparison of â vs. a and hit/miss a90/95 results. Finally, case C was, on the surface, very similar in inspection arrangement with case A. However, inspectors showed significantly better performance in Case C. This can be attributed to smaller time-pressure during case C, possible learning from earlier cases and differences in samples. This indicates, that rather small changes in the set-up or inspector feedback can have significant impact to the obtained POD performance. 4. Conclusions Three separate POD exercises were completed, each showing separate experimental challenges and solutions. The following conclusions may be drawn from this study: o Contrary to expectation, the false call rate did not show significant correlation with the inspection performance. o The data showed various unlikely events (small hits, big misses and poor separation), which necessitated adjustment for the standard methodology. This shows, that the standard POD methodology can not be used as a "black box", and must be accompanied by careful analysis of the underlying technical and statistical factors leading to the obtained POD values. o â vs. a and hit/miss analyses showed surprisingly poor correlation and caution is adviced in using â vs. a analysis for manual inspections such as the ones shown here. o Small changes in the inspection set-up (e.g. time pressure) can have significant impact to the obtained POD performance. Acknowledgements This work was supported by several partners who provided test pieces and inspection results for the study. Their support is gratefully acknowledged. In particular, the authors wish to thank RUAG aviation, Switzerland for providing some of the test samples for this study. a vs . â a 9 0 /9 5 hit/miss a90/95 References 1. Anon. 2009. Nondestructive Evaluation System Reliability Assessment. Department of Defense Handbook. MIL-HDBK-1823A. 171 p. 2. Anon. 2012. Standard Practice for Probability of Detection analysis for Hit/Miss Data. American Society for Testing and Materials, ASTM E2862-1 3. Anon. 2015. Standard Practice for Probability of Detection Analysis for â Versus a Data. American Society for Testing and Materials, ASTM-E3023 4. Annis, C., 2015. mh1823 R software package, version 4.3.2. Available online: http://statisticalengineering.com/mh1823/mh1823-algorithms.html

FEASIBILITY STUDY OF USING EFLAWS ON QUALIFICATION OF NUCLEAR SPENT FUEL DISPOSAL CANISTER INSPECTION Iikka Virkkunen 1 , Ulf Ronneteg 2 , Göran Emilsson 2 , Thomas Grybäck 2 , Kaisa Miettinen 1 1 Trueflaw Ltd., Finland, 2 Svensk Kärnbränslehantering AB, Sweden ABSTRACT The Swedish KBS-3 design for the disposal of spent fuel is based on the encapsulation of the fuel in canisters that consist of cast iron inserts and an outer 5-cm-thick shell of copper. To verify that the copper canisters fulfil the requirements, an extensive program for quality control is under development. In this program, the use of non-destructive testing (NDT) is vital, and it is therefore very important to develop reliable NDT methods, demonstrate their capability and, finally, qualify the NDT procedures. The qualification of NDT procedures traditionally requires representative mock-ups with a significant number of representative flaws. For the copper canister weld inspections, the expected flaw types and their NDT responses are well characterized due to extensive research [1]. However, there are limited possibilities for the artificial manufacturing of representative flaws with pre-defined sizes. In addition, the data collection is highly automated and reliable, and the main focus for qualification is thus in the data analysis. In present report, the use of the eFlaw technology has been developed and evaluated for the purpose of qualifying the weld inspection. The eFlaw technology allows extracting flaw signals from the existing NDT data files and introducing them back into the data files in different locations. Thus, the existing NDT data files can be used to produce a high number of "flawed" data files for training and qualification purposes. The copper canister inspection is fully automated and thus well suited for the method. The previously available technology has been developed further to be applicable to the copper canister inspection. In particular, the ability to use a single dataset containing flawed and unflawed regions for the flaw extraction has been developed and evaluated. With eFlaw, a limited number of collected data files can be used to produce a virtually unlimited number of blind data files for training and qualification purposes. With good access to variable training data, the data evaluators can improve their skills, become more confident and perform better. For qualification, each candidate can receive an individual blind data file, thereby improving the confidence level of the personnel qualification and, thus, its reliability. Furthermore, the time- consuming collection of ultrasonic data for each data set is avoided. M or e in fo a bo ut th is a rt ic le : ht tp :// w w w .n dt .n et /? id = 22 53 2 INTRODUCTION SKB is developing nondestructive testing (NDT) methods for inspecting copper canisters that are for the final disposal of spent nuclear fuel [1]. The methods and procedures will be qualified according to the Swedish requirements, which follow the guidelines published by the European Network for Inspection and Qualification (ENIQ). At the same time, the inspection requirements and conditions for the canisters are markedly different from typical nuclear inspections previously qualified in Sweden. Thus, it is expected that the qualification will need to be adapted significantly to suit the needs of this particular inspection. The ENIQ publishes a methodology document and set of recommended practices for conducting inspection qualification [2]. These form a flexible set of guidelines and are explicitly meant to be adapted for varying legal, regulatory and technical requirements. In the ENIQ methodology, the qualification consists, in simple terms, of • input information, which details the inspection target, • technical justification, which justifies the applicability of the chosen inspection technology, • open trials, which show the capability of the system on representative, known mock- ups, and finally, • blind trials, which show the capability of the inspection personnel on representative mock-ups containing defects unknown to the inspector. Both the open and blind trials traditionally require representative mock-ups. The manufacture of these representative mock-ups containing flaws is challenging. Often, the availability of representative mock-ups with a sufficient number of defects is one of the key challenges in qualification. This is one of the key reasons for having the technical justification in such an important role in the ENIQ methodology. In addition to the qualification mock-ups, the inspectors will need additional samples for training and method development. The challenge in providing representative mock-ups is common for all qualifications. For the canister inspection, there are additional challenges due to the rather unique materials and flaw types that are not well represented in traditional artificial flaw manufacturing methods. Furthermore, the canister inspection requires a highly specialized and fixed setup that will collect the inspection data. These requirements place some limitations on the way in which data collection can be done during the qualification exercise. Ideally, the qualification mock-ups should contain a sufficient number of various flaws to give a reliable (in statistical terms) picture of the capabilities of the inspection. Because the reliability requirements of the inspections are high, it follows that the number of flaws needed to provide statistical assurance of meeting the reliability is also high. For example, the ASTM-E2862 standard for probability of detection analysis [3] requires 60 flaws to define the 90% POD flaw size with 95% confidence. The mock-ups should also provide a sufficient number of defect-free areas to provide ample opportunity for detecting excessive false-call rates. Sufficient mock-ups should be available for training, open trials and blind trials. For the blind trials, there is also the risk of inspectors "learning the mock-ups"; thus, additional mock-ups are needed to provide sufficient variation in qualification trials and avoid developing dependence between supposedly independent qualification trials. These factors compound and cause the number of required mock-ups and flaws to become unmanageable for the traditional approach. Modern automated inspections, such as the highly specialized system applied in canister inspection, provide new opportunities to work around these traditional limitations. With automated inspection, the data gathering and analysis are separated into distinct steps. This separation allows new possibilities for training and qualification. Because the analysis now operates, essentially, on pre- recorded data, the need for different physical training samples and training data sets is also separated. The data gathering can be developed and qualified on physical samples, and the more demanding data analysis can be completed on a separate data set. The needs for these two steps are quite different. For data gathering, a representative sample is needed, but the problematic need for a high number of representative flaws primarily concerns data analysis. Consequently, being able to modify the gathered data sets to include non-existing virtual flaws offers several significant advantages: the number of physical test blocks and flaws can be reduced, the number of flaws in the data can be increased to give statistically significant results and the number of different data sets available can be increased such that every trainee or qualification candidate receives a fresh data set. This is the central idea of the eFlaw technology developed by Trueflaw [4]: a pure flaw signal is extracted from flawed data set and then re-introduced into various locations in the data set to provide a virtually unlimited number of different data files for training and qualification purposes. In this feasibility study, the previously developed eFlaw technology was further adapted for the SKB canister weld inspection case, and the resulting data files were evaluated for applicability for the intended training and qualification purposes. Previously, the flaw extraction had been done with two separate data files that were acquired before and after the flaw production. This process was not possible in this case because the flaws of interest are manufacturing flaws inadvertently created during the welding trials and cannot be manufactured artificially for a ready-made mock-up. Thus, the technology was developed to extract the flaw signal by comparing two different areas in the data file (flawed and unflawed). In addition, one of the challenges for such technology is the obscurity of the data file formats produced by the used ultrasonic inspection equipment. Consequently, significant adaptation was also necessary to facilitate use of this technology with the file formats presently applied by SKB. MATERIALS AND METHODS The preliminary ultrasonic inspection of the canister weld is performed using a linear array from the top of the copper lid, as shown in figure 1. Figure 1. Ultrasonic inspection setup of the canister weld. [1] The phased array ultrasonic data files consist of a number of individual ultrasonic channels that apply electronic scanning using different inspection angles and focus depths according to table 1. Table 1. Ultrasonic inspection setup of the canister weld. Channel Region Angle Depth range Aperture (elements) Focus depth +20° root Root 20° 40-65 mm 50 60 mm +12° root Root 12° 40-75 mm 50 60 mm 0° root Root 0° 51-76 mm 50 60 mm -12° root Root -12° 50-80 mm 40 67 mm -20° root Root -20° 55-75 mm 28 67 mm +35° shallow Upper part excl. root 35° 30-63 mm 42 47 mm +35° deep Lower part excl. root 35° 57-88 mm 55 75 mm +25° shallow Upper part excl. root 25° 30-62 mm 39 45 mm +25° deep Lower part excl. root 25° 57-88 mm 50 75 mm 0° shallow Entire weld excluding surface 0° 30-90 mm 32 60 mm SKB provided the data file from weld "FSW105" for this study. In the file, one root indication is clearly seen at the inspection channel "-20° root" at the circumferential position of 15°.The circumferential section 240°-300° is free of indications. Thus, the data in the "-20° root" ultrasonic channel at the circumferential position of 12°-18° were designated as the defect data, and the circumferential position 240°-300° were designated as the non-defect data volume to be used as the reference for the flaw signal extraction. The target UT data (corresponding with channel "-20° root") were extracted from the .uvdata file format. The data were then processed to extract the pure flaw signal, as described below. Then, the extracted flaw signal was re-introduced to the original data in different locations to produce a modified UT data set. Finally, the modified UT data were written back to the original .uvdata file to produce a modified .uvdata suitable for use in training or for evaluation purposes. Each of these steps is described in more detail below. Available UT channels Each channel contains an array of A-scans (i.e. recorded UT amplitudes as function of time). The corresponding data for each channel were extracted, and a C-plot was generated from the data (i.e. projection of the data in parallel to the scanning surface). These are shown in figure 2. The horizontal axis shows the circumferential scans around the canister weld from 0 ... 360°, and the Y-axis shows the radial direction. Each pixel represents a maximum value of the A-scan in the relevant position. The raw data is stored in 16 bit values (i.e., possible values range from 0 to 65536). However, true measured values use only a fraction of the possible range, and the colors are thus scaled to range from 0 to 6% of the possible range (roughly equal to 24 dB "software gain"). The analysis was completed on the UT data corresponding to the "-20° root" channel in the .uvdata file provided. The data consisted of 1834 circumferential positions (0.2°/step, approximately 360°), with 25 a-scans per position and 150 data points per a-scan. Each data point is a 16-bit unsigned integer value. An overview of the target data is shown in figure 2. The designated flaw region and the unflawed region are shown in more detail in figures 3 and 4, respectively. Figure 2. C-plot from the target ultrasonic channel "-20° root" Figure 3. Detailed view of the designated flawed region (12° – 18°, scanlines 60-90) Figure 4. Detailed view of the designated unflawed region (250° – 256°, scanlines 1250-1310) Flaw signal extraction To extract the flaw signal, the difference between the flawed region and the corresponding unflawed region was taken data point by data point. Due to the variation in the noise, some of these differences were negative. Consequently, the difference signals were stored as signed values. Due to the difference in the noise level between the designated flawed region and the unflawed region, the extracted signal could not be used directly. When this was tried, a noticeable difference in the noise level of the modified data was observed. This difference gives away the flaw location and thus interferes with the use of the data file for the intended purpose. To alleviate this problem, a histogram filter was applied to the extracted flaw signal, thereby reducing the noise level but leaving the flaw signal intact. The filter is depicted in figure 5. The threshold was adjusted until no adverse effect was noticed to 10%. By using this filter, the difference in the noise level was removed, and good data files were obtained. Filtering the extracted signal potentially disturbs the flaw signal and may affect data analysis. Thus, a filter with minimal impact on the signal was sought and the threshold 0° 90° 180° 270° 360° Circumference R a d ia l s c a n 0% 2% 4% 6% 8% 0% 2% 4% 6% 8% 0% 2% 4% 6% 8% value set on trial-and-error basis to avoid disturbing the signal. Some sizing techniques (most notably the various dB-drop techniques) may be sensitive to filtering. To re-introduce the flaw signal to another location in the data file, a process roughly inverse to the extraction was applied. That is, the extracted flaw signal was summed, data point by data point, to the chosen location. Figure 5. Histogram filter applied to the flaw data to remove noise effects on re- introduction of the flaw signal. A threshold value is determined as a percentage of the maximum peak value present in the data. Values above this value are not modified. Values below this value are multiplied by linearly decreasing factor resulting in a histogram filter as shown. (Negative values are similarly adjusted.) Threshold is set high for illustration, true threshold was adjusted to data. RESULTS With the tool set developed as described above, three modified data files were generated for detailed evaluation of the data. Figure 6 show C-scans from three generated data files. The images show that the flaw can be re-introduced to designated locations, and no noise difference or other artefacts are observable. To further investigate possible artifacts introduced by the data modification, the UT data were analyzed at the a-scan level. Figure 7 shows a-scans from the original flaw, the original signal at the re-introduction location and finally, the modified a-scan with the flaw introduced. The images indicate that the a-scans do not show any noticeable artifacts resulting from the modification process. Figure 6. C-scan from modified data file showing the original flaw and the three times re-introduced flaw at locations 150°, 230° and 270°. 0° 90° 180° 270° 360° Circumference R a d ia l s c a n 0% 2% 4% 6% 8% a) b) c) Figure 7. A scan from the: a) original flaw location, b) re-introduction location before modification, and c) after modification. Finally, the flaw location was compared at the sector level to ensure that the modification did not induce any cutoff in the flaw signal or other similar artifacts. Figure 8 shows the comparison for a single sectorial position. The figure shows that there is no cutoff or other visible artefacts. a) b) c) Figure 8. Sector (B) scan from the original flaw location (a), re-introduction location before modification (b) and after modification (c). DISCUSSION The flaw signal was extracted from the designated location and applied to different locations to form modified data files with additional flaws. The resulting data files were carefully evaluated for any possible artefacts caused by the modification operation. The flaw signal integrated well with the existing signal, and no artefacts were noted. Consequently, the data file modification was successful in all accounts. The current analysis was limited to a single .uvdata file and a single flaw signal. Thus, it is appropriate to discuss the extensibility of these results to wider applications with different data files and flaw signals. The current process of flaw extraction, which uses a flawed region and an unflawed reference region, worked quite well for the present data file and flaw signal. However, due to the different noise level between the locations, some additional filtering was necessary to remove the effects of this noise difference from the extracted flaw signal. The flaw in question has a very good signal-to-noise ratio, and the filtering could thus easily be done without affecting the flaw signal. However, if the signal-to- noise ratio were very small, it would become increasingly difficult to differentiate between the flaw signal and the noise signal. Thus, setting the filtering parameters may need manual adjusting when the process is applied to flaws with a very different signal-to-noise ratio. In addition, for flaws with a very low signal-to-noise region, the reference region must be chosen to have a very similar noise level, and/or more sophisticated filtering may be necessary for good results. This is the price for extracting the flaw signal from a single file (and not files acquired before and after flaw production). In summary, the feasibility study developed and documented here indicates that the eFlaw technology is well suited for training and inspection of the SKB canister. The chosen reference defect could be extracted successfully and re-introduced to different locations. The method is expected to be widely applicable to different data files, channels and flaws with sufficient signal-to-noise ratio. 0% 2% 4% 6% 8% CONCLUSIONS The following conclusions can be drawn from this feasibility study: • a flaw signal could be extracted successfully from the flawed and unflawed areas; • the extracted flaw signal could be re-introduced without introducing visible artefacts to the data; • the data file thus generated is applicable for training and qualification purposes; and • the developed capability is expected to be widely applicable to different data files, channels and flaws. REFERENCES 1) Ronneteg U, Grybäck T, 2015. Non-destructive of canister components and welds, Svensk Kärnbränslehantering AB. available online (2016-04-19): http://www.skb.se/wp- content/uploads/2016/03/1434744-Non-destructive-testing-of-canister-components-and- welds.pdf. 2) Anon. 2007. The European methodology for qualification of non-destructive testing, Third Issue. ENIQ Report nr. 31, EUR 22906 EN, ISSN 1018-5593. 3) Anon. 2012. Standard Practice for Probability of Detection analysis for Hit/Miss Data. American Society for Testing and Materials, ASTM E2862-12. 4) Virkkunen, I., Miettinen, K. and Packalén, T. Virtual flaws for NDE training and qualification. 11th European Conference on Non-Destructive Testing (ECNDT 2014), October 6-10, 2014, Prague, Czech Republic.

7th International Symposium on NDT in Aerospace – Tu.3.A.2 1 License: http://creativecommons.org/licenses/by/3.0/ CRACKED SAMPLES FOR AIRFRAME COMPONENTS Iikka VIRKKUNEN 1, Jorma PATRONEN 2, Marko YLITALO 2 1 Trueflaw Ltd., Espoo, Finland 2 Patria Aviation, Halli, Finland Abstract. Evaluation of NDT performance and reliability is an important part of any NDT system. Traditionally, the challenge has been limited supply of representative cracked samples, that would allow accurate evaluation of the achieved performance. This paper describes production of cracked samples in airframe component representative of typical rivet hole configuration. The flaws were produced using local thermal fatigue loading and are representative of in-service fatigue cracks. The manufactured flaws were inspected using typical eddy current inspection. The flaws gave inspection indication comparable to those expected from service- induced cracks. Introduction Demonstration of NDE reliability is an important part of any inspection system. Capability is commonly assessed in terms of probability of detection (POD) as function of crack size. The largest flaw, that could be missed during inspection is of particular importance. In the case of airframe rivet holes, it is sometimes assumed, that cracks with size below the detection capability of the NDE system may exist in the component even after clean inspection result. The possible existence of such cracks can be dealt with by boring rivet holes to larger size and thus removing possibly existing cracks. However, this procedure is limited by the amount of material that can be removed without other adverse effects. Consequently, it is particularly important to have accurate information on the detection capability of the NDE system to avoid leaving cracks in the component on the one hand and to avoid removing unnecessary material on the other. To demonstrate performance and to find out the detection capabilities of an NDE system, inspection is performed on a set of cracked test samples. These inspections should replicate the actual inspection as closely as possible. In particular, the cracks should be representative of service induced cracks. The determination of POD curve is codified in ASTM standards [1] and handbooks [2]. In general, crack sizes ranging from undetectable to highly detectable should be included to get meaningful POD curve. Traditionally, it's been challenging to manufacture the necessary cracked samples for performance demonstration. Various techniques including mechanical fatigue of plate samples with crack initiators or stress concentrators have been used. As the detection targets for the NDE systems decrease, the production of controlled cracks has become increasingly difficult. For number of years, Trueflaw has manufactured cracks for NDE purposes using thermal fatigue loading. These have been used extensively in the nuclear industry 2 performance demonstrations for over a decade. The cracks have also been used in aero- engine components for method development [3] and POD determination. Trueflaw uses in-situ thermal fatigue loading to grow cracks for NDE purposes. This approach has number of advantages: the flaws are natural (thermal)fatigue cracks and as such highly representative of service-induced cracks. The location and size is highly controllable and range of cracks with different sizes can be produced, as is needed for a POD determination. No artificial initiators or stress concentrators are needed and cracks can be produced to actual components and geometries. This enables far greater representativeness for the POD inspections than would be possible for more traditional small samples. So far, the use of Trueflaw cracks in the aerospace industry has been limited to engine components. The reason is, that the most interesting materials for airframe components, i.e. aluminum, has been very difficult for thermal fatigue crack manufacturing due to the very high thermal conductivity of the material. Recently, the crack manufacturing technology was developed to allow crack production also in aluminum thus enabling the use of these cracks also for airframe components. Materials and methods In present study, cracked evaluation samples representing a typical airframe component, i.e. a rivet hole was manufactured. Similar samples were manufactured in titanium and aluminum. Titanium is used both in the aero engines and in airframe and its thermal properties are closer to the previously cracked aero-engine materials. Thus, it was used as a reference material. Aluminum, on the other hand, is more interesting material for airframe components and more directly shows the potential for these cracks on the airframe components. Due to the different thermal conductivity and other material properties, the cracks produced in aluminum show somewhat different properties. The used sample geometries are shown in Figure 1. Figure 1. Sample geometry. Crack manufacturing by thermal fatigue During crack manufacturing, the sample is locally heated and cooled repeatedly. The heating is done by high frequency induction and cooling by water spray. Rapid heating and cooling cause uneven temperature distribution which induces compressive and tensile surface stresses onto the sample, respectively. The alternating thermal stresses result in gradually accumulating fatigue damage, crack initiation and crack growth. The loading is stopped periodically and the sample inspected in order to follow the crack growth. The 3 process is highly repeatable and thus crack properties that are not directly observable (e.g. crack depth) can be reliably assessed using a separate validation sample that is destructively examined. Results Table 1 shows the generated cracked samples and associated crack sizes. Figures 2 and 3 show typical microscopic images of generated cracks in titanium and aluminum, respectively. Table 1. Generated cracks and crack sizes. # Flaw ID Material Length (mm) 1 187BFB2859 Titanium 0.2 2 189BFB2862 Titanium 1.4 3 190BFB2865 Titanium 1.1 4 191BFB2869 Titanium 0.3 5 194BFB2871 Titanium 1.4 6 195BFB2872 Titanium 0.3 7 196BFB2873 Titanium 0.6 8 210BFB2891 Aluminum 0.4 9 212BFB2901 Aluminum 2.2 10 218BFB2907 Aluminum 1.3 11 221BFB2908 Aluminum 2.1 12 222BFB2910 Aluminum 0.1 Figure 2. Microscopic image of typical crack in titanium. 4 Figure 3. Microscopic image of typical crack in aluminum. The manufactured cracks were first inspected using fluorescent dye penetrant (FPI) at Trueflaw. The samples were subsequently sent to Patria for representative eddy current inspection and further evaluation. At Patria, inspection more closely representing the actual inspection for airframe components were conducted. Figures 4 and 5 show typical FPI-images for titanium and aluminum samples, respectively. Figures 6 and 7 show corresponding EC-images. Figure 4. FPI image of typical crack in titanium. The crack is the same as that depicted in Figure 2 (Trueflaw crack number 196BFB2873). 5 Figure 5. FPI image of typical crack in aluminum. The crack is the same as that depicted in Figure 3 (Trueflaw crack number 218BFB2907). Figure 6. EC image of typical crack in titanium. The crack is the same as that depicted in Figures 2 and 4 (Trueflaw crack number 196BFB2873). 6 Figure 7. EC image of typical crack in aluminum. The crack is the same as that depicted in Figure 3 (Trueflaw crack number 218BFB2907). The inspection results on various used EC techniques is collected in Table 2 and depicted graphically in Figure 8. The number of cracks generated for this preliminary study is insufficient for full ASTM-E2862 POD analysis. However, the results demonstrate, that an array of cracks ranging from undetectable to clearly detectable can now be produced in materials and components representative of typical airframe components. Table 2. Summary of inspection results. # ID Crack length Inspector Hit rate (mm) #1 #2 #3 #4 Titanium 1 187BFB2859 0.2 miss miss miss miss 0 % 2 195BFB2872 0.3 miss miss miss hit 25 % 3 191BFB2869 0.3 hit miss hit miss 50 % 4 196BFB2873 0.6 hit hit hit hit 100 % 5 190BFB2865 1.1 hit hit hit hit 100 % 6 194BFB2871 1.4 hit hit hit hit 100 % 7 189BFB2862 1.4 hit hit hit hit 100 % Aluminium 8 222BFB2910 0.1 miss miss miss 0 % 9 210BFB2891 0.4 hit hit hit 100 % 10 218BFB2907 1.3 hit hit hit 100 % 11 221BFB2908 2.1 hit hit hit 100 % 12 212BFB2901 2.2 hit hit hit 100 % 7 Figure 8. Summary of inspection results.The solid line shows the detection as percentage from trials in Table 2. The dashed line shows rough estimate 95% confidence bound based on the binomial model (true POD above this line would show the better-than-given results at 95% confidence, assuming POD increases with increasing crack size). The data is insufficient for true POD analysis and thus this is likely to seriously underestimate the true POD. Discussion Cracked sample plates representing typical airframe rivet hole configuration were manufactured from titanium and aluminum materials. The samples had thermal fatigue cracks with different sizes manufactured to them. The cracks were characterized with optical microscopy and subjected to inspections representative of actual in-service inspections of comparable components. The microscopic characterization shows that cracks initiate from the hole surface and grow following the local microstructure as expected for natural fatigue cracks. The cracks initiate at the hole corner and extend to both directions. The cracks in titanium show, in general, smaller crack opening than those in aluminum. This can be caused by both different material properties and different loading conditions. The smaller yield strength of the aluminum increases crack tip plastic zone size and increases crack tip blunting and crack opening for similar stresses. In the future, this could be alleviated to some extent by decreasing the loading (with corresponding increase in the crack production time). The aluminum samples also showed some increased oxidation during the loading. This can be attributed to both the increased temperature and the interaction between the deformation and oxidation (i.e. the oxide film may break during tensile loading and expose new material to the environment). The oxidation also caused conductivity changes, which caused phase angle change in eddy current inspection. This tendency was reduced using corrosion inhibits in the cooling medium. Nevertheless, mechanical removal of affected oxide layer was necessary after production. Although these issues were reduced by the use of inhibit, some additional development is necessary to fully remove oxidation effects. 0% 20% 40% 60% 80% 100% 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 Crack size (mm) Percentage detected 95% lower bound 8 The eddy current inspection showed detectable indications for all but the smallest cracks. The signals were comparable to those expected of a service-induced cracks in titanium. In some of the aluminum cracks, the eddy current signals differed slightly in phase angle and in form from service-induced cracks. The cracks were still detectable, as shown in Table 2. Consequently, the cracks were evaluated positively for further use in airframe NDE performance demonstration use. The current sample set was too small for proper POD analysis. Nevertheless, the results already give better indication on the capabilities of the NDE techniques used and can be used to focus further effort on POD determination to interesting range of crack sizes. Conclusions Cracks were successfully manufactured to titanium and aluminum samples representative of typical airframe rivet hole configuration. This represents the first application of thermal fatigue cracks for NDE performance demonstration in airframe components and first evaluation of developed technique in typical aerospace aluminum materials. The manufactured cracks showed characteristics representative of service-induced cracks and were positively evaluated for use in airframe NDE performance demonstration. Acknowledgements The support and collaboration of Ari Kivistö and the Finnish Defence Forces Logistics Command is gratefully acknowledged. References [1] Anon. 2012. Standard Practice for Probability of Detection analysis for Hit/Miss Data. American Society for Testing and Materials, ASTM E2862-12. [2] Anon. 2009. Nondestructive Evaluation System Reliability Assessment. Department of Defence Handbook. MIL-HDBK-1823A. 171 p. [3] Virkkunen I., Kempainen M., Ostermeyer H., Paussu R. and Dunhill T., 2009. Grown cracks for NDT development and qualification. Insight Vol 51 No 5 May 2009, 5 p. [4] Kemppainen, M., Virkkunen, I. 2012. Production of Real Flaws in Probability of Detection (POD-) Samples for Aerospace Applications. 4th International Symposium on NDT in Aerospace, 13th – 15th November 2012, Augsburg, Germany. 2012-11-24

Virtual flaws for NDE training and qualification Iikka VIRKKUNEN1, Kaisa MIETTINEN1, Tapani PACKALÉN2 1 Trueflaw Ltd., Espoo, Finland, e-mail: iikka.virkkunen@trueflaw.com, kaisa.miettinen@trueflaw.com 2 Inspecta Sertifiointi Oy, Vantaa, Finland Abstract A typical qualification mock-up contains limited number of realistic flaws due to time and cost limitations. Increasing automated inspections offer new possibilities to overcome these limitations. In automated inspections, the data gathering and analysis phases are separated. This allows the introduction of flaw signal into the data, that were not present in the original data. On the other hand, the representativeness of these "virtual flaws" need to be confirmed. This paper introduces newly developed technique that allows introduction of realistic virtual flaws to data-sets of automated ultrasonic inspection data. The sample is first scanned before flaw introduction to establish baseline signal. Then, flaws are manufacture to the sample, and it is re-scanned. The difference in the datasets reveals the crack signal, which can be extracted. After the crack signal is separated, the it can be re- applied to an another location or locations. The new virtual crack data differs from the actual crack data, as the separated crack signal is superimposed to the (possibly different) background noise. At the same time, the signal is still realistic, since it is acquired from realistic crack and transferred unaltered. The data manipulation is undetectable from the manipulated. Keywords: Ultrasonic Testing (UT), personnel training and certification, Qualification 1. Introduction One long standing issue in NDE training and qualification is the availability and cost of relevant flawed test blocks. The test blocks and, in particular, the defects they contain should be representative to the actual inspection challenges that the inspectors are likely to encounter during field operation. However, manufacturing realistic test blocks and flaws is time consuming and costly. Consequently, the number and quality of test blocks is often limited. In training, where the required number of test blocks is high this leads to reduced representativeness of both samples and flaws they contain. In particular, EDM notches are often used instead of real cracks. In qualification, the use of representative samples and flaws is considered indispensable. Consequently, the number of flaws is small to limit the overall cost of test blocks. This is becoming increasingly problematic now that there's more demand to get quantitative performance data from qualification. This, in turn, necessitates statistically significant number of cracks and trials. Also, as the number of qualified inspectors and re-qualifications increase, growing number of data-sets are needed. Over the years, there has been number of approaches to overcome this. For training, there has been systems that mimic or manipulate the operation of standard inspection machine and introduce "virtual flaws" into the machine screen during operation [1]. However, these have not found widespread use, possibly due to complexity and cost of such systems. Traditionally, the data gathering and analysis is done simultaneously and the operator can move the inspection probe arbitrarily to study a possible indication. The operator can also change probes to get more information regarding the source of the indication. All this has made it rather difficult to simulate actual inspection conditions with virtual instruments and flaws. However, in recent years, there has been a growing trend in use of automated inspections. Simultaneously, the use of multi-channel systems (phased array ultrasonic inspection or eddy current array probes) has dramatically increased the amount of information 11th European Conference on Non-Destructive Testing (ECNDT 2014), October 6-10, 2014, Prague, Czech Republic obtained from a single scan. The automated inspections greatly improve the reproducibility of the inspection and allow later re-analysis of the data. With automated inspection, the data gathering and analysis are separated to distinct steps. The gathered data can be analysed in its entirety and possible indications can be compared to other areas in the data. Also, number of post-processing and data merging options are available for the operator to improve detection and characterization of indications. This separation also allows new possibilities for training and qualification. Since the analysis now operates, essentially, on pre-recorded data the need for different physical training samples and training data-sets are also separated. The data gathering can be trained and qualified on physical samples while the more demanding data analysis can be completed on separate (possibly unrelated) data set. The needs for the two steps are quite different. For the data gathering, representative sample is needed, but the costly need for high number of representative flaws primarily concerns data analysis. Consequently, being able to modify the gathered data-sets to include non-existing virtual flaws offer several significant advantages: the number of physical test blocks and flaws could be reduced, the number of flaws in the data could be increased to give statistically significant results and the number of different data-sets available could be increased so that every trainee or qualification candidate receives a fresh data set. Simulation is nowadays routinely used as part of training and qualification. In qualification, its important use is in technical justification, where calculations can show the applicability and possible limitations of the chosen inspection procedure. Simulations could also be used, to provide data-sets for performance evaluation [2,3]. Recently Wirdelius & Persson [2] used simulation to provide virtual test data and also used detection criteria specified in the procedure to produce a full simulated POD curve. The challenge for using simulation for performance demonstration is, of course, proper validation of the simulation to make sure that the performance demonstrated on validation cases is descriptive of the actual inspection. Another possibility to provide flawed data-sets is to manipulate existing, actually acquired data-set to introduce new flaws to the data. The flaws themselves can be actually acquired as part of some other data and can be introduced to numerous locations in the target data-set. This approach retains most of the advantages of the simulation approach, but since both the flawless data-set and the flaw signals are really acquired data from actual realistic flaws, there is better inherent representativeness to actual data. This is the approach taken in this paper. As discussed above, the idea of using virtual data to train and qualify NDT procedures is by no means new. However, the present approach offers some advantages over the previously used ones, namely the better representativeness to the actual data. Recently EPRI [4] has done similar work, but there is no more detailed description available on the open literature. Furthermore, the approach described here has some unique features of its own. The main challenge for this kind of approach is to make sure, that the manipulation done to the data files does not introduce any distinct features or artefacts that would affect the analysis of the inspector. Firstly, the flaw indication should be copied entirely, and no part of the flaw data should be left out. Secondly, only the flaw data should be copied - variations in noise etc. should not be copied together with the flaw. Finally, the embedding of the flaw signal to the acquired data should not introduce any new features such as abrupt changes in signal strength etc. that would indicate artificial changes. 2. Virtual flaws for NDE training and qualification The approach taken here is then to: scan the sample in unflawed condition, introduce required flaws to the sample using standard Trueflaw technology, re-scan the sample in flawed condition, extract the flawed signal from the two data-sets and re-introduce the flaw signal(s) to the unflawed data to various locations to produce any number of flawed data sets. The principle is illustrated in Figure 1. Figure 1. Principle of virtual flaw extraction and flawed data set generation Since the only physical difference between the unflawed mock-up scan and the flawed mock- up scan is the introduction of the flaws, the flawed signal can be extracted, in principle, by taking the difference between the data files. In practice, there's random noise that will cause some additional difference to be registered. Consequently, the extraction is limited to areas around locations where the signal is above noise threshold. The process is not very sensitive to the cut-off point (as long as the whole flaw signal is included). The difference between two similar random noise signals is similar in character to the noise itself and thus does not introduce visible artifacts to the data. In addition, there may be slight changes in the sample location between the two scans, which will cause differences in the signal. Thus, it may be necessary to introduce small adjustments to the data location to allow for optimal extraction. After successful data extraction, the flaw signal can be superposed to the clean mock-up to various locations. Since the flaw signal is not altered or modified (just translated), the signal is still representative of the actual flaw signal. 3. Materials and methods To test the process in practice, a pilot study was completed for small scale sample. The sample chosen was a simple stainless steel plate sample with weld in the middle. The sample was scanned with qualified procedure in clean condition. The procedure has been qualified by the Finnish qualification body (Inspecta Sertifiointi Oy). Then, a flaw was produced to the fusion line of the crack and the sample was re-scanned with the same procedure. Figure 2. shows the sample and scanning set-up. The scanning was done by Ville Lehtinen, DEKRA Industrial Oy. Examination was carried out using qualified phased array ultrasonic technique based on Zetec’s PDI qualification. Used instrument was Omniscan PA (32/128), Manual Pipe Scanner and Zetec PDI transducers (TRS-technique). The probes used were 1.5 MHz matrix probes (3 x 5 elements in each probe). The sample was scanned with three scan lines. Figure 2. Sample and scanning set-up. After scanning, the flaw signal was extracted as described in section 2 and re-introduced to the clean data set with 20 mm offset. The small size of the sample did not allow for multiple flaws to be introduced, so in this pilot study, only one flaw was superimposed to the flawless data set. 4. Results and discussion The resulting data file was compared with the flawless and flawed data sets. The data sets were given to several inspectors with request to try and find possible problems or artifacts in the modified data. The data compared favorably and no artifacts were reported in the modified data. Figures 3 - 5 show the raw (un-merged) data in various ways. Figure 6 shows a sample A-scan from the image. Initially, there was a concern, that although the modifications are indistinguishable from the sectional typical analysis images, it might be possible to notice difference if the inspector went down to analyze the individual A-scan images. However, in analyzing several A-scan images, no problems were seen and the modified signal integrated well with the A-scan images in all cases. Figures 7 and 8 show the merged images. It can be seen, that the merge images are produced from the modified data as expected and no trace of the modification can be seen. This is to be expected, since all the raw data images showed good results and since the merge process has a smoothing effect on the data and thus is somewhat forgiving to small problems in the data. This smoothing effect improves the inspection by reducing the effect of natural problems occurring during the normal scanning of the data, but it also affects favorably possible artificial problems from data manipulation. Figure 3. The three data-sets compared; from left to right: the modified data, flawless data and flawed data. The flaw signal has been moved in the modified data. Also, the flaw signal is superimposed to the local noise and thus looks a bit different from the original flaw signal (as it should). Figure 4. C-scan showing that the flaw signal has been moved in all the scan lines. Figure 5. Sectional scan showing that the flaw signal has been successfully modified in all sample angles. Figure 6. Sample A-scans showing how the extracted flaw signal superimposes nicely to the level of individual A-scans, even when the noise level of the unflawed sample at the specified location is different from the original location. Figure 7. Merged data set from the top showing that the modified data file produces the expected merge images as well. Figure 8. Merge data from sector scan showing that the the modified data file produces expected merge images and that the flaw signal integrates well to local noise. 4. Conclusions The pilot project showed, that the data manipulation and introduction of virtual flaws is an effective technology that can now be used to simultaneously reduce cost and improve the statistical quality of qualification. It also offers interesting opportunities for training of data analysis. Furthermore, the technology could be used in ways described in Burkhardt et al. [1] as more integral part of NDT quality improvement. Acknowledgements The modified data sets were reviewed and commented on by several finnish NDT experts, whose support is gratefully acknowledged: Ville Lehtinen (DEKRA Industrial Oy), Jonne Haapalainen (VTT), Esa Leskelä (VTT) and Ari Koskinen (VTT). References 1. Burkhardt, G, Fisher, J. & Peterson, E. 2004. System and Method for Nondestructive Testing Simulation. US. patent 2004/0117133 A1. 2. Wirdelius, H. & Persson, G. 2012. Simulation Vased Validation of the Detection CApacity of an Ultrasonic Inspection Procedure. International Journal of Fatigue, 41, pp. 23 - 29. 3. Rodriguez, R., Fernández, M., Domíngue, J. & Biezma, M. 2014. Detection of defects in metallic specimens supported by ultrasound propagation simulations. Materials Testing, 56 (5), pp. 386-392. 4. Anon. 2013. Nondestructive Evaluation: Virtual Mockups -- Feasibility Study into Electronic Implantation of Flaw Responses into Previously Recorded Ultrasonic Data. Available online: http://www.epri.com/abstracts/Pages/ProductAbstract.aspx?ProductId=00000000000102 5222

NDT&E International 55 (2013) 96–101Contents lists available at SciVerse ScienceDirectNDT&E International0963-86 http://d n Corr E-m jingwanjournal homepage: www.elsevier.com/locate/ndteintDiscussion on modeling of thermal fatigue cracks in numerical simulation based on eddy current signalsJing Wang a,n, Noritaka Yusa b, Hongliang Pan a, Mika Kemppainen c, Iikka Virkkunen c, Hidetoshi Hashizume b a School of Mechanical and Power Engineering, East China University of Science and Technology, 130 Meilong, Xuhui, Shanghai 200237, China b Department of Quantum Science and Energy Engineering, Graduate School of Engineering, Tohoku University, 6-6 Aramaki Aza Aoba, Aoba-Ku, Sendai, Miyagi 980-8579, Japan c Trueflaw Ltd., Tillinmäentie 3, tila A113, 02330 Espoo, Finlanda r t i c l e i n f o Article history: Received 15 September 2012 Received in revised form 19 December 2012 Accepted 6 January 2013 Available online 23 January 2013 Keywords: Thermal fatigue crack Numerical modeling Finite element method Eddy current testing95/$ - see front matter & 2013 Elsevier Ltd. A x.doi.org/10.1016/j.ndteint.2013.01.012 esponding author. Tel.: þ81 22 795 7906; fax ail addresses: jwang@karma.qse.tohoku.ac.jp, g8604@gmail.com (J. Wang).a b s t r a c t This study evaluates modeling of thermal fatigue cracks by the finite element method from the view point of eddy current testing. Five artificial thermal fatigue cracks introduced into type 304 stainless steel plates were prepared for the research. Eddy current signals were gathered by a differential type plus point probe and subsequent destructive tests were performed to confirm the true profiles of the cracks. Numerical simulation based on the results of eddy current testing and destructive tests were carried out to show how the thermal fatigue cracks should be modeled in numerical simulations. The results of the numerical simulations revealed that thermal fatigue cracks tend to be much less conductive than stress corrosion cracks if they are assumed to have uniform conductivity inside. The results also imply that taking consideration of magnetization induced by the thermal fatigue process enables eddy current signals to be analyzed more quantitatively. & 2013 Elsevier Ltd. All rights reserved.1. Introduction Thermal fatigue is attributed to uneven temperature distribu- tion, and inevitably occurs in components where temperatures alternate, e.g., a T-joint with mixture of hot and cold fluids [1]. With the cumulative damage of thermal fatigue, cracks of the type called the thermal fatigue crack (TFC) are widely generated in industrial processes. Therefore, carrying out active non-destructive inspection for TFCs has significant importance for safety and for ensuring continuity of industrial production. In general, TFCs have a small crack opening and many branches [2]. Whereas ultrasonic testing (UT) is the most popular method for sizing defects of components [3,4], the accuracy of measurement of UT tends to be affected by the minute properties of the crack [5,6]. Furthermore carrying out UT is time-consum- ing, and utilization of a couplant is necessary. These facts indicate that the application of non-ultrasonic-based nondestructive test- ing together with UT should provide more quantitative evaluation of TFCs. Eddy current testing (ECT) is one of the most promising non- destructive testing methods from this point of view. Advanced computational physics for electromagnetic fields enable quantitativell rights reserved. : þ81 22 795 6319.evaluation of crack profiles from ECT signals [7,8]; a few studies have even reported success in quantitative evaluation of stress corrosion cracks (SCC) [9,10]. However, recent studies have pointed out that the accuracy of evaluation strongly depends on how a crack is modeled [11–15]. This indicates that evaluating the modeling of TFCs is essential to discuss the applicability of ECT to the evaluation of TFCs. However, studies on the nondestructive evaluation of TFCs have been carried out mainly from the viewpoint of UT [16,17]. This study is conducted on the basis of the above background and aims to evaluate the modeling of TFCs in an eddy current numerical simulation. Five TFCs artificially introduced into austenitic stainless steel plates were investigated. The eddy current testing utilized a plus point probe driven at three frequencies. Subsequent destructive tests were carried out to reveal the profile of the TFCs. Based on the results of the eddy current testing and destructive tests, modeling of TFCs was evaluated in numerical simulations.2. Materials and methods 2.1. Preparation of thermal fatigue crack samples Three type 304 austenitic stainless steel plates were prepared for this study. Each plate measured 250 mm long, 150 mm wide, 25 mm thick and had a chemical composition listed in Table 1. Thermal fatigue loading was introduced by alternate heating and J. Wang et al. / NDT&E International 55 (2013) 96–101 97cooling. Heating was applied with high frequency induction, and cooling with a water spray. Maximum temperatures used were below 400 degrees, and minimum temperatures were around 10 degrees. Two TFCs were introduced to each of two of the plates, and one TFC to the third. Hereafter, these five cracks are denoted as TFC1-5. These artificial TFCs were regarded as favorable representatives of service-induced TFCs since previous study has proved that TFCs produced by controlled thermal fatigue loading have quite similar performance to service-induced forms, accord- ing to results of metallographic analysis and UT [16,17].2.2. Eddy current testing and destructive tests Eddy current testing was conducted to measure eddy current signals using a differential type plus point probe illustrated in Fig. 1. The plus point probe consists of two differentially connected square coils with inner lengths, thicknesses, and widths of 5.0, 3.0, and 0.2 mm, respectively. Since the depths of the flaws were estimated to be around 4 mm and an earlier study demonstrated eddy current signals gathered using a differential plus point probe shows dependency on flaw depth up whose depth is up to the twice of the depth of penetration [14], this study adopted an exciting frequency of 50 kHz that provides aFig. 1. The differential type plus point probe Table 1 Chemical composition of the plates. C Si Mn P S Ni Cr Co Content 0.019 0.420 1.750 0.028 0.001 8.170 18.360 0.110 Unit: %.skin depth of approximately 1.94 mm, In order to discuss the dependency of modeling on frequency two additional frequencies of 100 and 400 kHz were also utilized. An XYZ stage controlled by a PC positioned the probe during the inspection, and the probe scanned along lines parallel to a crack. The specimens were destroyed to reveal the true profile of the TFCs after the eddy current testing. The destructive tests evalu- ated both the cross-section at the center of the cracks and the boundary profiles to confirm the three-dimensional geometry of the cracks. 2.3. Numerical simulations Finite element simulations were performed to show the numerical modeling of TFCs from the viewpoint of eddy current testing. The simulations modeled a TFC as a planar region with a constant width, and evaluated the electrical properties of the crack that reproduced the measured eddy current signals. The difference between the measured and simulated signals was calculated by e¼ S 5 i ¼ 1 9Zmea i Zsim i 9 ð1Þ where Zmea and Zsim denote measured and simulated eddy current signals, respectively and subscript i denotes signals measured at the ith scanning point. Thus, comparison between signals of measurement and simulation were carried out by considering the trajectory of signals when the probe scanned directly above the crack. In the study, a total of five scanning points were considered. The maximum value of trajectory could be gathered at the third scanning point; the other four scanning points were uniformly and symmetrically distributed beside the thirdutilized to measure eddy current signals. Fig. 2. Model of thermal fatigue in numerical simulation: (a) Model A with constant width and uniform conductivity; (b) model B, considering electromag- netic variation in the vicinity of a thermal fatigue crack to illustrate conductive variation. Fig. 3. Trajectories of the eddy current signals measured using the plus point probe. (a) 50 kHz, (b) 100 kHz, (c) 400 kHz. J. Wang et al. / NDT&E International 55 (2013) 96–10198scanning point so that the signal trend could be clearly expressed by the values given at these points. The simulations were conducted by commercial software, Comsol Multiphysics and its AC/DC module. The governing equation is ðjoso2eÞAþr  ðm1r  AÞ ¼ Je ð2Þ where o is the angular frequency, s is the conductivity, e is the permittivity, A is the magnetic vector potential, m is the perme- ability and Je is the current density of the exciter. A curl element was utilized in the numerical model. The size of the computa- tional domain was 400400400 mm. The boundary condition was imposed so that the tangential component of the magnetic vector potential was zero. The total number of elements in the model was about 200,000. The mesh was sufficiently fine so that the error caused by the mesh was only 0.08% of the signal. Fig. 2 illustrates the two numerical models utilized in this study. Both the models have a constant width, and the boundary profile of a crack is as correctly modeled as possible on the basis of the results of the destructive test. Model A, shown in Fig. 2(a) models a crack as a region with a constant width and uniform conductivity inside. The width was set to be as 0.01, 0.02, 0.05, 0.10, 0.20, 0.50, or 1.00 mm; the conductivity was assumed to be 0.0, 0.1, 0.2, 0.5, 1.0, 2.0, 5.0, or 10.0% of the base material’s conductivity (1.35 MS/m). Another model, model B shown in Fig. 2(b), assumes that the conductivity in a deep part of a crack is larger than that near the surface [18]. The upper half of a crack has a constant width and uniform conductivity identical to the ones that minimized Eq. (1) using model A. The adoption of optimized width and conductivity in the upper half of a crack is related to two reasons. First minimum initial difference between measured and simulated signal could be reached using optimized width and conductivity compared with other widths and conductivities. Furthermore, the optimized width is the most similar one to the result of optical microscope tests. The lower half of a crack has, in contrast, greater conductivity, whereas the width is the same as that of the upper half of the crack. The conductivity of the lower half was set to be 0.5, 1.0, 2.0, 5.0, or 10.0% of the base material’s conductivity. The possibility of magnetization induced due to the thermal fatigue process, in the vicinity of a TFC, was modeled by two magnetic layers beside the crack. These two layers had the same boundary profile with the flaw and their conductivities were same with that of the base material. The relative permeability andwidth of the layers were assumed to be 1, 2, 3, 4, 5 and 1, 2, 5, 10 times of the width that minimized Eq. (1) using model A, respectively. Furthermore, these layers had identical conductivity and relative permeability for each simulation. Fig. 6. Geometric model of simulation. Boundary profiles of the cracks are revealed by a destructive test. Table 2 Maximum lengths and depths of the cracks. Max. length (mm) Max. depth (mm) TFC-1 9.7 3.5 J. Wang et al. / NDT&E International 55 (2013) 96–101 993. Results and discussion The trajectories of the eddy current signals measured by the eddy current instruments are shown in Fig. 3. The signals shown in the figure were those measured along a scanning line where the maximum signal was obtained, and they were all calibrated so that the maximum signal due to an artificial rectangular slit of 20.0 mm long, 0.5 mm wide, and 5.0 mm deep had an amplitude of 10 V and phase of 45 degrees. Signals of simulation in this paper were also calibrated by the same slit. The results of optical microscope tests are shown in Fig. 4. This reveals that all the cracks are generated in one direction except TFC-5. An obvious cluster is shown in Fig. 4(e). An example of cross-sectional profiles of one of the cracks, TFC-3, is presented in Fig. 5. Pictures with higher resolutions are available at http:// jsm.or.jp/jsm/at/scc/index_eng.htm. Profiles of other cracks did not differ significantly from those in the figure. That is, no cluster was observed in the direction of the depth of a crack. Fig. 6 demonstrates a geometric model of a numerical simulation. The boundary profiles of the cracks, namely the crack profile on the plane parallel to the crack, revealed by the destructive tests are also shown in Fig. 6. Table 2 summarizes the maximum lengthsFig. 4. Results of optical microscope tests. (a) TFC-1; (b) TFC-2; (c) TFC-3; (d) TFC- 4; (e) TFC-5. Fig. 5. Examples of the cross-sectional profile of the crack (TFC-3). TFC-2 11.7 4.1 TFC-3 22.1 6.5 TFC-4 14.4 4.1 TFC-5 6.5 3.1 Table 3 Evaluated crack parameters without considering electromagnetic variation in the vicinity of the TFC (model A). Crack Freq. (kHz) Conductivity (%)a Width (mm) e TFC-1 50 0.0 0.01 1.75 100 0.1 0.10 1.48 400 0.1 0.10 1.14 TFC-2 50 0.0 0.01 2.79 100 0.0 0.01 1.57 400 0.0 0.05 1.45 TFC-3 50 0.1 0.10 2.15 100 0.1 0.10 1.82 400 0.2 0.20 2.55 TFC-4 50 0.0 0.01 3.16 100 0.1 0.10 2.62 400 0.1 0.10 2.74 TFC-5 50 0.1 0.05 2.11 100 0.2 0.05 1.77 400 2.0 0.50 1.68 a Conductivity with respect to the base material.and depths of the cracks evaluated from the results of optical microscope tests and destructive tests. The conductivity and width of the cracks minimizing Eq. (1) using model A are summarized in Table 3, together with the value of e that the conductivity and width provide. The values of conductivity show that a TFC is much less conductive than an SCC, as earlier publications have proved [15,19]. However, the conductivity of TFC-5 is relatively higher, especially when a frequency of 400 kHz is employed. The most plausible reason would be that the cluster of TFC-5, as Fig. 4(e) shows, increases the area of the contact between the fracture surfaces of cracks. The results of Table 3 indicate that the conductivity and width tend to become larger with the increase in frequency. Whereas it is plausible that this is due to the shallower depth of penetration and larger crack opening near the surface of plate; validations using more number of specimens are indispensable for further discussion. It should be noted, however, the resistance defined as width divided Table 4 Evaluated crack parameters considering electromagnetic variation in the vicinity of TFC (model B) when a frequency of 50 kHz is used. Crack Conductivitya of conductive edge Width of magnetic layer (mm) Relative permeability of magnetic layer e TFC-1 1.0 0.1 4 1.48 TFC-2 5.0 0.1 5 2.53 TFC-3 10.0 1.0 2 1.58 TFC-4 5.0 0.1 5 2.89 TFC-5 5.0 0.5 2 1.66 a Conductivity with respect to the base material. Fig. 7. Trajectories of the measured signal and simulated signal generated by two mod of modeling TFC by model A. Simulation-2 presents the results of modeling TFC by mo J. Wang et al. / NDT&E International 55 (2013) 96–101100by the conductivity is almost unchanged, which agrees well with the previous study on modeling of SCC [15,19]. Table 4 shows results obtained by considering electromagnetic variation in the vicinity of TFCs when 50 kHz was utilized, namely results obtained by model B. The fact that the lower half of a crack was more conductive would be due to the partial closure. The closure is assumed to be caused by the tortuosity of the crack path and compression stress around the tip of the crack [20,21]. A real crack opening would be smaller than the crack opening demonstrated by metallographic analysis in Fig. 5 since etching would increase the crack opening by rounding its fracture surface. It has been shown that the width of the magnetic layer tends to increase along with the size of the crack opening. Inels of TFC when a frequency of 50 kHz is used. Simulation-1 presents the results del B. (a) TFC-1; (b) TFC-2; (c) TFC-3; (d) TFC-4; (e) TFC-5. J. Wang et al. / NDT&E International 55 (2013) 96–101 101practice, magnetization is caused by plastic deformation during the propagation of the crack since the base material is type 304 austenitic stainless steel. It has been demonstrated that type 304 austenitic stainless steel becomes magnetic after the generation of plastic deformation [22]. This reasonably explains why the range of the magnetic layer is related to the size of the crack. Fig. 7 illustrates the trajectories of the measured signal and simulated signal produced by the above two models of TFC at a frequency of 50 kHz. The results from more than five scanning points are shown in the figure for better presentation of the trajectories. Combined with the values of e shown in Tables 2 and 3, they indicate that modeling of TFC could be developed to a certain extent by considering magnetizations induced by the thermal fatigue process. However, the result of TFC-3 in Fig. 7 does not show obvious improvement in reconstruction of the eddy current signal. That means it is not always necessary to consider magnetization of modeling of TFCs for more accurate reconstruction.4. Conclusion This study evaluated modeling of the TFC in a numerical simulation from the viewpoint of eddy current testing. Five artificial TFCs introduced into type 304 austenitic stainless steel plates were prepared for the research. Signals were gathered by a differential type plus point probe using several frequencies. Each crack was simulated as two kinds of models. Both models had a constant width and true profile revealed by destructive test. Model A had uniform conductivity. Then, the conductivity and width of the TFC were evaluated by this model. The results demonstrated that the resistance defined as width divided by the conductivity is almost unchanged, even though the conduc- tivity and width become larger with an increase in frequency. The values of the conductivity indicate that TFC is much less conductive than a stress corrosion crack, and thus the latter should be modeled as a more conductive region according to previous research. Furthermore, model B, considering magnetization induced due to the thermal fatigue process was also used to simulate a TFC. The more conductive lower part of the TFC would be caused by the partial closure and compression stress around the tip of the crack. The magnetic layer would be due to the plastic deformation since the base material was a type 304 austenitic stainless steel. In general, modeling of TFCs could be developed further by considering magnetizations induced by the thermal fatigue pro- cess. However, one of the reconstructive results in Fig. 7 indicates that considering magnetization in the vicinity of a TFC does not always enable more accurate reconstruction.Acknowledgment This study was supported by a project launched by the Japan Society of Maintenology aiming at the enhancement of NDT & E ofstress corrosion cracks [23]. The results of optical microscope tests can be found on the website http://jsm.or.jp/jsm/at/scc/ index_eng.htm.References [1] Chapuliot S, Gourdin C, Payen T, Magnaud JP, Monavon A. Hydro-Thermal- Mechanical analysis of thermal fatigue in a mixing tee. Nucl Eng Des 2005;235:575–96. [2] Wale, J. Crack characterization for in-service inspection planning—an update. SKI reference 14.43-200543105, ISRN SKI-R-06/24-SE, SKI, Stockholm, Swe- den; 2006: 24. ISSN 1104-1374. [3] Whittle M. A review of worldwide practice and experience in the qualifica- tion of ultrasonic inspections of nuclear components over the past two decades. Insight: Non-Destr Test Cond Monit 2009;51(3):140–50. [4] Ahmed S, Saka M. A new ultrasonic angle-beam technique for sensitive evaluation of closed cracks. NDT&E Int 1999;33:261–71. [5] Saka M, Salam Akanda M. Ultrasonic measurement of the crack depth and crack opening stress intensity factor under a no-load condition. J Nondestr Eval 2004;23(2):49–63. [6] Kemppainen M, Virkkumen I, Packalen T, Sillanpaa J, Paussu R. Importance of crack opening in UT inspection qualification. In: Proceedings of the sixth international conference on NDE in relation to structural integrity for nuclear and pressurised components, 8–19 October: Budapest, Hungary; 2007. p. 93–105. [7] Bowler JR. Review of eddy current inversion with application to nondestruc- tive evaluation. Int J Appl Electromagn Mech 1997;8(1):3–16. [8] Auld BA, Moulder JC. Review of advances in quantitative eddy current nondestructive evaluation. J Nondestr Eval 1999;18(1):3–36. [9] Yusa N, Chen Z, Miya K. Sizing of stress corrosion cracking on austenitic stainless piping in a nuclear power plant from eddy current NDT signals. Nondestr Testing Eval 2005;20(2):103–14. [10] Yusa N, Chen Z, Miya K, Uchimoto T, Takagi T. Large-scale parallel computa- tion for the reconstruction of natural stress corrosion cracks from eddy current testing signals. NDT&E Int 2003;36:449–59. [11] Yusa N, Perrin S, Mizumo K, Miya K. Numerical modeling of general cracks from the viewpoint of eddy current simulations. NDT&E Int 2007;40:577–83. [12] Badics Z, Matsumoto Y, Aoki K, Nakayasu F, Kurokawa A. Finite element models of stress corrosion cracks (SCC) in 3-D eddy current NDE problem. In: Collins R, Dover WD, Bowler JR, Miya K, editors. Nondestructive testing of materials. IOS Press; 1995. p. 21–9. [13] Chen Z, Aoto K, Miya K. Reconstruction of cracks with physical closure form signals of eddy current testing. IEEE Trans Magn 2000;36:1018–22. [14] Yusa N, Huang H, Miya K. Numerical evaluation of the ill-posedness of eddy current problems to size real cracks. NDT&E Int 2007;40:185–91. [15] Yusa N, Miya K. Discussion on the equivalent conductivity and resistance of stress corrosion cracks in eddy current simulations. NDT&E Int 2009;42:9–15. [16] Kemppainen M, Virkkunen I, Pitkänen J, Paussu R, Hänninen H. Advanced flaw production method for in-service inspection qualification mock-ups. Nucl Eng Des 2003;224:105–17. [17] Kemppainen M, Virkkunen I. Crack characteristics and their importance to NDE. J Nondestr Eval 2011;30(3):143–57. [18] Yusa N, Hashizume H. Four-terminal measurement of the distribution of electrical resistance across stress corrosion cracking. NDT&E Int 2011;44:544–6. [19] Yusa N, Hashizume H. Evaluation of stress corrosion cracking as a function of its resistance to eddy currents. Nucl Eng Des 2009;239:2713–8. [20] Kane A, Doquet V. Problems related to thermal fatigue of stainless steel: interactions of orthogonal cracks networks under biaxial tension and influence of stress biaxiality on 3D mode I crack growth. In: 15th European conference on fracture, ECF15: Stockholm France; 2004. [21] Marci G, Packman P. The effect of the plastic wake zone on the conditions for fatigue crack propagation. Int J Fracture 1980;16(2):133–53. [22] Li H, Chen Z, Li Y, Takagi T, Uchimoto T, Chigusa N, et al. Dependence of deformation-induced magnetic field on plastic deformation for SUS304 stainless steel. Int J Appl Electromagn Mech 2012;38:17–26. [23] Yusa N, Miya K, Komura I, Chen Z. A project aiming at the enhancement of NDT&E of stress corrosion cracking. Int J Appl Electromagn Mech 2012;33:1587–90.

5th European-American Workshop on Reliability of NDE – Lecture 7 1 Lizenz: http://creativecommons.org/licenses/by/3.0/de/ Quantifying NDE Reliability from ENIQ Qualification Information Iikka VIRKKUNEN *, Mika KEMPPAINEN *, Kaisa MIETTINEN * * Trueflaw Ltd., Espoo, Finland Abstract. The current trend towards risk-informed inspection planning, increasing requirements on plant safety and aging of power plants increase the importance of quantifying nondestructive evaluation (NDE) reliability in the Nuclear industry. At the same time, there's large body of work already done to ensure NDE reliability in the form of inspection qualification. In Europe, this mostly takes the form of European network for inspection qualification (ENIQ) -style qualification. However, attempts to infer quantitative NDE reliability information from existing qualification data have met with limited success. In particular, numerous approaches have been tried to estimate probability of detection (POD )curves based on qualification data. These include the MIL-HDBK-1823A statistical approach, Bayesian approach and others. Unfortunately, this work has, to date, been largely unsuccessful due to lack of data or test pieces for statistical analysis, or in some cases due to improper distribution of available data. The present paper introduces an alternate approach for estimating NDE reliability from existing qualifications. Instead of focusing on the actual inspection results gathered from the qualification trials (which are few in number), the approach focuses on the qualification requirements, i.e. test piece trials and related pass-fail criteria. A set of relevant performance criteria (i.e. POD curve or sizing error) is then tested against the qualification requirements to determine the pass probability for an inspector having such a performance. The pass probabilities can be calculated (e.g.) using Monte Carlo method. The highest performance that will likely fail qualification can be used as a lower limit estimate of performance for inspectors who will pass the qualification. 1. Introduction Nondestructive evaluation (NDE) is one of the key tools for ensuring continued safe and reliable operation of nuclear power plants. Thus, the required reliability for NDE is high. The flaws that NDE needs to find are quite challenging. Service induced cracks may be tight, small and otherwise difficult to detect. Yet, the expected reliability of inspections is very high. After some early round robin exercised showed [1,2] clearly insufficient performance, the nuclear industry has taken steps to assert sufficient NDE performance. In essence, it is now required that sufficient performance is demonstrated (qualified) before using any NDE method or procedure for in service inspection of nuclear power plant. Also, it was noted, that significant variability between inspectors exists and thus also used personnel must be qualified for nuclear inspections. 2 Two distinct qualification schemes or frameworks are currently widely used for nuclear NDE qualification. In the U.S., the american society of mechanical engineers (ASME) code adopted a statistical screening approach [3]. The code defines a required sample set and pass/fail criteria for inspection qualification. The inspector is, in principle, allowed to use any method necessary to make the inspection, as long as he or she is able to complete the inspection of defined sample and pass the set criteria. The ASME criteria were developed so that an inspector with unacceptable performance would fail the test with high probability whereas inspector with acceptable performance would pass the test with high probability. Such a test is always a compromise between number of samples or inspections and accuracy of the test. A set of "power curves" was developed to quantify this idea and to define sample set size and pass fail criteria that would provide good compromise between number of samples required and demonstrated reliability. Sometimes quoted example of the criteria is that if an examiner identifies 90% of the flaws in the specimen test set and does not exceed 10% false call rate, the examiners probability of passing the qualification is 90%. [3, 4] At the same time, the European network for inspection qualification (ENIQ) took somewhat different approach. In ENIQ, it was postulated, that statistical evidence alone would not be sufficient to guarantee the required level of performance and that the number of test samples necessary for such demonstration would be prohibitively large. Furthermore, the European context necessitated greater flexibility for requirements due to varying technical situation and authority requirements in various countries. Consequently, the ENIQ developed a more flexible framework for inspection qualification. In this framework, a separate input information document is created for each qualification case. The input information defines the inspection requirements, expected flaw types and the structural integrity context of the inspection. Next, a technical justification (TJ) is written for the NDE procedure to be qualified. This justification includes the physical reasoning and possibly references to other evidence that demonstrates high-expected reliability for the inspection. The TJ is also used to select test blocks and defects to test the reliability. This allows use of "worst case" defects to reduce needed sample set. Finally, the qualification is completed with open practical trials for procedure qualification and blind practical trials for personnel qualification. The TJ and practical trials together demonstrate that the NDE system has required performance. Neither the ASME qualification nor ENIQ qualification provide quantitative data on the attained performance levels. The results are given as pass/fail or acceptable/unacceptable. In recent years, there's been increasing need for quantitative performance data for NDE. In part, this is due to advances in risk-informed in service inspection (RI-ISI). The rationale behind RI-ISI is that inspections should be focused to components and locations where they are most beneficial. Likewise, the inspection intervals should be chosen with maximum expected benefit from the inspection. Unsurprisingly, the expected NDE reliability has significant influence in these calculations and the optimal inspection strategy varies depending on inspection reliability. In particular, it's often necessary to demonstrate very high inspection reliability to get significant advantage for performing NDE in the RI- ISI calculations. Quantifying inspection reliability is also becoming increasingly important for the qualification itself. It's now over 30 years since the first inspection qualification programs started. Consequently, there's also an ever-increasing body of completed qualifications. These include qualification done tens of years between them and with wildly different NDE equipment. It's difficult, in particular fort the ENIQ type qualifications, to enforce and demonstrate consistent requirements across qualifications. Furthermore, there's increasing interest to take advantage of already completed work on previous qualifications or 3 qualifications done elsewhere. This is also complicated by the lack of quantitative measures that would make qualifications comparable across time and between countries. Finally, in the ENIQ methodology, the input information is used to define scope and performance target for the inspection procedure. Structural integrity significance and previous degradation data, when available, are used to determine the inspection scope. This allows, at least in principle, for the NDE to focus on significant damage modes and work with reasonable detection targets. However, it's currently difficult to assess how well the qualification really addresses these requirements and what is the confidence level that qualified inspectors meet the targets set in the input information. There's been significant effort to address this increasing need for quantitative performance demonstration data. Given the great body of available qualification data, the focus has mostly been to extract quantitative data from existing qualification data. However, this has proven quite challenging. Currently, the work has culminated in defining POD-curve, as required by the RI-ISI calculations, from completed or planned qualification exercise. There's a series of ENIQ reports [5-7] detailing various approaches that show the extent of this challenge. At first, it was recognized, that since in an ENIQ qualification the performance is demonstrated with combination of technical justification and practical trials, any quantitative measure should take credit for both parts. It was a challenge to quantify the value of the TJ and to combine this with practical trials. Gandossi et al. [5,6] noted that as it is, the adequacy of the TJ is judged by an expert judgement in the qualification body. Consequently, the problem was, in essence, quantifying this expert judgement. However, it proved difficult for the experts to provide quantitative assessment of the TJs'. The quantitative evaluation was finally improved by introducing the concept of "equivalent test blocks". The experts were asked to compare the TJ to a practical trial and asses the number of test blocks required to provide same confidence of NDE performance that the TJ alone provides. The quantified expert judgement and information from practical trials were then combined to provide final estimate for the quantified NDE reliability using Bayesian inference. However, even the improved expert judgements contained significant variability and thus the reliability of the judgement remained somewhat questionable. Furthermore, when the expert judgement and practical trial information was combined, often the TJ did not show significant contribution to the overall demonstrated performance (and sometimes even provided negative contribution). Over the same period of time, the aerospace industry developed their own methodology for demonstrating NDE performance. This methodology relied more heavily on statistical evidence and developed advanced methodology to extract POD information from limited set of cracked samples [8-9]. The most recent version is documented in the MIL-HDBK-1823A and recently standardised as the ASTM E2862. Sadly, attempts to apply this approach to extract POD curves from nuclear industry qualification data have met with limited success. The most obvious problem has been lack of samples: the MIL- HDBK-1823A hit/miss analysis requires 60 samples. For nuclear qualification, it's uncommon to have more than 20 flaws in qualification. There are also more subtle problems. The ASTM E2862 directly states, that POD cannot be modelled (as continuous function of discontinuity size) if all discontinuities are found (or if none are found). In nuclear qualification it is expected that all the cracks will be found. Furthermore, the use of various worst-case defect locations makes the assumption of monotonously increasing POD for the test set questionable. Consequently, it would require significant changes to current qualification practices for the ASTM E2862 to be applicable; most notably increased number of cracks and cracks with low to medium probability of detection. 4 There's also some other approaches developed in the aerospace industry. These have not, to our knowledge, been applied in the nuclear industry nor do they seem to present a viable solution for the problem at hand. Some are summarized here because they provide alternate solution to similar problems or expose criticism within the aerospace industry. A set of independent hit/miss inspections can be modelled with binomial distribution (with certain limitations). It follows that finding 29 cracks out of 29 cracked samples is consistent with 90% lower limit POD estimate at 95% confidence level (i.e., the true POD is >= 90% with 95% confidence, or conversely there is a 5% chance that the true POD is < 90%). Thus, the 29/29 requirement has been used in some cases in the aerospace industry [13]. This is, perhaps, the closest equivalent the aerospace industry uses to the nuclear industry qualification (ASME-type qualification, in particular). The requirement does not provide POD curve, but it provides statistical assurance (at 95% confidence level) of sufficient performance (90% POD). The DOEPOD model [10-12] is also based on the binomial view of hit/miss data. The main motivation for the DOEPOD model is, that using model-based POD estimation (e.g. ASTM E2862) assumes POD as a function of flaw size follows certain model. In particular, the POD is continuous, monotonically increasing function of flaw size a. This assumption may not always be justifiable, e.g. when the method sensitivity varies for different flaw sizes due to different probes, beam focusing or for some other reason. The DOEPOD model does not assume functional relationship between POD and flaw size. Instead, the inspection results are grouped and analysed, simply stated, as groups to make sure that the 29/29 condition is fulfilled for certain flaw size and flaw sizes above it. In summary, it can be said that the previous attempts to quantify NDE performance based on ENIQ qualifications have not proven successful. Furthermore, the approaches tried so far have clear problems or incompatibilities and thus success with these seems improbable. At the same time, it's generally agreed that qualification has significantly improved NDE reliability. Thus, it should also be possible to quantify the improvement. There's no doubt that the vast amount of qualification data contains valuable information about the NDE performance. The topic of this paper is to present an alternate approach to quantifying NDE reliability based on ENIQ qualification data. 2. Qualification as a screening test The root cause for many of the problems in quantifying NDE performance from ENIQ qualification is, that the qualification was designed to be a screening test and not a POD experiment or performance evaluation. Consequently, it's closer to the 29/29 requirement than it is to ASTM E2862 exercise. In order to get quantitative data from the current nuclear inspection, the problem set must be changed to reflect the character of the qualification. The chosen problem set is then: given that the inspector has passed qualification, what's the lower limit performance we can expect from this inspector? To make the problem easier to solve, we first solve the inverse problem: given, that the inspector has certain performance, what's the probability that he will pass the qualification. With this information, the inverse problem can be solved iteratively, given certain limitations for inspector performance. The approach can be seen as an extension to the ASME power curve approach or to the 29/29 demonstration criteria. For further simplification, we first solve the problem concerning the blind test results only. The role and significance of the TJ and open trials is discussed in section 5. 5 Solving probability of passing a qualification, given the inspector performance is rather straightforward task. Sometimes, the inspection task and the pass/fail criteria are simple enough to be solved analytically. Often, the pass/fail criteria contain multiple overlapping criteria, and forming analytical solutions is cumbersome. Consequently, numerical solution via Monte Carlo simulation was chosen for this work. The Monte Carlo simulation is easy to formulate, and solves in couple of seconds. The numerical Monte Carlo solution easily accommodates various extended pass/fail criteria and false calls. (See paragraph 4. for details of the simulation.) This intermediary result has certain practical significance. In the ENIQ process, input information is used to review possible degradation mechanisms and their possible structural significance. With this information, the input information sets the goal or target for qualification, i.e. the performance required for the qualified personnel. Probability of passing the qualification with this given performance is a measure of the confidence level provided by the qualification (assuming different performance targets are comparable; see paragraph 3. for further discussion). Any inspector, who has lower performance than specified, has lower probability of passing the qualification. Thus, the probability of the target performance passing can be seen as the risk that we are willing to accept for any inspector up to this performance level qualifying. In many of the previous approaches, false calls do not significantly alter the outcome of the analysis. They can often be handled by simply removing false call data from the analysis. In contrast, the false call rate of the inspector and related pass-fail criteria significantly affect the probability of passing the qualification. The qualification body cannot, in principle, separate the inspectors that passed due to high false call rate from those that passed due to the required skill. The amount of blank samples (opportunities to make false calls) and the number of allowed false calls affect the probability of passing qualification as well as the "optimal" level of false calls that maximize candidates chance to pass. The discussion here focuses on probability of detection. However, the approach is easily extendible to sizing or other performance criteria. 3. POD vs. a dependence Different approaches for solving probability of detection as a function of crack size ( POD (a) ) require different assumptions concerning the POD(a) behaviour. The present approach is no different. The MIL-HDBK-1823A approach requires that POD be continuous and increasing function of a. In addition, it is sometimes assumed that the limiting POD for large cracks sizes is 100%. These assumptions are integral in the success of the approach and significantly improve the amount of information that can be extracted from limited sample data. At the same time, both of these assumptions have been questioned and may be difficult to justify in some cases. The DOEPOD approach specifically addresses the assumption that POD is an increasing function of a and provides assurance of POD without this assumption. For present purpose, different assumptions regarding POD(a) dependence can be made. If information about this dependence is available, it can be used to provide better estimate for the expected POD. In cases where such information is not available, simplified models can be used, at the expense of justified performance. The present approach enforces one additional requirement for whatever POD(a) dependence is assumed. Since the required/postulated performance is used as the lower 6 limit performance estimate, any chosen POD(a) dependence must provide a uniquely comparable set of performance criteria. For example, Picture 1 shows a set of two POD curve estimations. Both of these curves show an area where POD surpasses the other curve and, consequently, neither is ambiguously better. In other words, is the inspector allowed to compensate lower-than-required performance for certain flaw sizes with better-than-require performance for other flaw sizes? Pic. 1. Two possible POD curves neither of which is unambiguously better than the other. For this reason, four models of POD(a) dependence are included, each of which is useful under certain conditions. Firstly, the usual normal-cfd functional dependence is used. This is a commonly used approximation for POD(a) dependence. As shown above, the curves are not, in general, unambiguously comparable and thus additional discretion is needed from the operator to choose a comparable subset of curves. Secondly, the normal-cfd function with additional maximum POD is provided. This provides additional flexibility, when the assumption of POD limiting to 1 cannot be justified. Both of these models require significant assumption regarding the POD(a) dependence, which may not be available. The third model is a stepwise curve with adjustable POD at top level. It is conventional to use detection target and minimum POD to determine inspection requirements in the input information. This functional form corresponds to such target definition. Stepwise POD curve is also sometimes used in the RI-ISI analysis [14]. The stepwise POD curve assumes equal POD over significant flaw size range. If POD is an increasing function of flaw size, then the candidate may compensate less-than- required performance at the detection target with higher-than-required performance on the bigger flaws. To calculate POD at detection target, the fourth model "counts" only cracks at the smallest size. Other flaw sizes are not used for calculating the probability of passing qualification. If, on the other hand, some of the bigger flaw sizes are missed, then the test is considered a failure. This is similar to including only cracks with size matching the detection target in the flaw set. 4. Software implementation To provide a practical implementation for this method, a web-based software code was written and made publicly available at http://www.trueflaw.com/qualificationhelper. The implementation has a user interface usable with modern web browser and a computation engine running on the server that computes the computationally intensive Monte Carlo 7 results. The back-end is written in Objective-C and is heavily optimized to provide responsive user interface. The front-end provides fields to input given inspector performance (as outlined above). Two sets of performance criteria can be input: one for lower limit (unacceptable) performance and one for upper limit (satisfactory) performance. The lower limit performance can be used to iteratively solve the performance that can be reliably expected from an inspector that has passed the qualification. The upper limit performance can be used to iteratively solve the performance, which can be expected to pass the qualification (i.e. the requirements seen by the inspector). Fields are provided to input qualification test setup information: number of cracks and corresponding crack sizes, number of blanks (for false call analysis) and pass/fail criteria: misses allowed and false calls allowed. The front end then gathers this data and sends it to the server for analysis. The server calculates the corresponding pass probabilities and reports them back to the front end, which shows them to the user. The tool also has similar analysis implemented for crack sizing performance. This is not considered in present paper. The back-end receives the candidate performance data (POD as function of crack size and false call probability), trial information and pass/fail criteria. It then calculates the corresponding pass probability using ordinary Monte Carlo analysis. The algorithm used for the Monte Carlo simulation is outlined below in pseudo code: repeat for N Monte Carlo trials: for each test block (crack or blank): draw a random number R between 0.0 ... 1.0 if R < given false call rate and test block contains a crack, report detection if R < given false call rate and test block is a blank, report false call if R > given false call rate and test block is a blank, report correct clean block if R > given false call rate and test block contains a crack of size a, draw another random number S between 0.0 ... 1.0 if S < POD(a), report detection else, report miss compare reported misses and false calls with given criteria and determine the trial to be pass or fail report number of passes / number of trials as the pass probability 5. The role of the technical justification As noted in section 1, in the ENIQ methodology the evidence for sufficient performance is formed by the technical justification and the practical trials. So far, the discussion has only considered practical trials. Consequently, it may be argued that the present discussion gives overly pessimistic view of the true performance of the inspection system. Thus, the TJ should be considered as well. The role of the TJ in ENIQ qualification is somewhat controversial. It's generally agreed that the TJ is important and valuable part of the ENIQ qualification. However, there are marked differences in the way the TJ is valued both between qualification bodies and between qualification cases. Thus, the contribution of TJ is presented in four different ways: a) TJ as necessary but insufficient condition One approach is to say, that the role of the TJ is to set the limits of applicability for the inspection in question. It is thus necessary and important part of the qualification. However, in this approach, it's not considered sufficient to show performance. Thus, the reliability of the inspection should be valued on the practical trials alone, and the TJ seen as a precondition. 8 b) TJ as necessary and sufficient condition Second view states, that the TJ in itself provides assurance of NDE reliability. In this case, the contribution should be quantifiable. To use TJ's contribution in connection with present framework, its contribution should be quantified in terms of equivalent hits. That is, the TJ adds test blocks, which are automatically found, and this improves the statistical confidence limits. Such quantification is similar to what was done by Gandossi et al. [5,6], and it has proven challenging to do in practice. c) TJ as justification for lower statistical confidence level The basic premise of the ENIQ qualification is, that practical trials alone will not provide sufficient statistical confidence and thus the TJ is needed to provide sufficient overall confidence to the NDE performance. This notion offers another way to quantify the effect of the TJ: the TJ can be used to justify that lower statistical confidence can be considered sufficient. This would be another way to quantify the expert judgement applied in current ENIQ qualifications. It can also be calculated after the fact, by analysing a qualification case in connection with the stated detection target and POD requirement and calculating the confidence level at which this is attained. d) TJ as justification for POD(a) dependence Finally, the TJ can be used to justify certain form of POD(a) dependence. Better knowledge about this dependence can increase the amount of information that can be extracted from inspection data. Depending on the qualification case, each one of these may be justified. The list is not exhaustive and some approaches can be used in combination (especially c and d). In any case, the decision is a form of expert judgement applied to the qualification case. 6. Example analysis of qualification data To show current approach in practical context, an example analysis was completed with real data from a Finnish qualification. Due to confidentiality of the qualification data, the analysis is completed on open sample data. The blind sample data can be assumed to be similar enough for present purposes. Also, the exact details of the case in question are left undisclosed. In first analysis, the shape of the POD curve is assumed known and POD is assumed to reach 100%. Picture 2 shows screen capture with POD curves giving 90% pass probability (performance which vendors need to have to pass qualification with high probability), and 4% pass probability. The inspectors are assumed to have 0% false call rate, to simplify the analysis. The curves were found by manually iterating to find acceptable risk that a passed inspector does not, in fact, have better than desired capability. 9 Pic. 2. Example analysis with assumed known POD curve shape reaching 100%. POD curves giving 90% pass probability (green curve) and 4% pass probability (red curve). The shape and slope of the curve is assumed to be known. Second analysis assumes the shape of the POD curve to be known, but the lower- limit POD is expected to level off at 90% (Picture 3a; for the following analysis only the POD curves and resulting pass probabilities are shown; qualification information is as shown in Picture 2). Now the decrease in maximum POD must be compensated with higher POD for the lower flaw sizes, which is seen as a slight shift of curves to the left. Thirdly, it is assumed, that the Input information states detection target of 1 mm and assumes step-wise POD-curve with maximum POD levelling off at 90% (Picture 3b). Now it can be seen, that the qualification has 39% probability of passing inspectors at the 90% limit. At the same time, 99% POD is needed for the vendor to be confident of his possibility to pass the qualification. The risk of passing at the lower limit curve can be decreased by lowering the required POD. For 71% maximum POD, the risk of passing is 5%. Alternately, the detection threshold can be increased above the smallest crack size in the test, which decreases the the pass probability to zero (the stepwise function assumes 0% POD below the threshold, so crack sizes are automatically missed. Since the pass/fail criteria do not allow any misses, this fails the qualification.) Finally, it is assumed, that the POD may, or may not increase after 1 mm detection target (Picture 3c.). Thus the confidence of POD at this level, is only affected by cracks near this size. Thus the amount of flaws is reduced significantly, and there's 73% risk of passing POD of 90%. Conversely, there's 5% risk of passing POD of 35%. (It should be noted, that this is quite conservative assumption.) 10 a) b) c) Pic. 3. Example analysis with different assumed known POD curves. See text for details. (Different stepwise POD-curves are drawn with slight offset to aid readability.) 7. Conclusions The approach presented here provides an alternate way to quantify the performance guaranteed by existing ENIQ-qualification (with given the risk that a passed inspector does not, in fact, have better than desired capability). It can be applied on existing information and it allows for different ways to take advantage of the technical justification included in the ENIQ qualification. 11 The example analysis here shows, that with limited sample set the level of risk for passing lower than desired capability associated with the practical trials alone is somewhat high (even assuming zero false call rate), and the technical justification is important to justify that this level of practical demonstration is sufficient to show that the required performance is met in practice. Acknowledgements The authors wish to thank Jennifer R. Brown for her most valuable comments on the manuscript. References [1] Lemaintre, P., Koblé, T. D. & Doctor, S. R. 1996. "Summary of the PISC round robin test results on wrought and cast austenitic steel weldments, part III: cast-to-cast capability study". International Journal of Pressure Vessels and Piping (69) pp. 33-44. [2] Lemaitre, P., Iacono, I. & Vergucht, P. (eds.) 1999. Results of the Destructive Examination of the ENIQ Pilot Study: Defect Catalogue. ENIQ Report 19, EUR 19024 EN, JRC Petten, The Netherlands [3] Cowfer, C.D. 1991. Basis / Background for ASME Code Section XI proposed Appendix VIII: Ultrasonic examination performance demonstration. Nuclear Engineering and Design, 131, pp. 313 - 317. [4] Becker, F. L. 2001. Examination Effectiveness Based on Performance Demonstration Results for Flaws at the Clad-to-Base Metal Interface, Proc. 3rd Int. Conf. NDE in Relation to Structural Integrity for Nuclear and Pressurized Components, Seville, Spain, November 14–16. [5] Gandossi, L., Simola, K. A. 2007. Bayesian Framework For Quantitative Modelling of the ENIQ Methodology for Qualification of Non-Destructive Testing. European Communities, DG JRC, Institute for Energy, EUR22675EN, ISSN 1018-5593. [6] Gandossi, L., Simola, K., Shepherd, B. 2010. Application of a Bayesian Model for the Quantification of the European Methodology for Qualification of Non-Destructive Testing. International Journal of Pressure Vessels and Piping, 87, pp. 111-116. [7] Gandossi, L., Annis, C. 2010. Probability of Detection Curves: Statistical Best-Practices. European Communities, DG JRC, Institute for Energy, ENIQ Report nr. 41, EUR 24429 EN, ISBN 978-92-79- 16105-6, ISSN 1018-5593. [8] Anon. 2009. Nondestructive Evaluation System Reliability Assessment. Department of Defence Handbook. MIL-HDBK-1823A. 171 p. [9] Anon. 2012. Standard Practice for Probability of Detection analysis for Hit/Miss Data. American Society for Testing and Materials, ASTM E2862-12. [10] Generazio, E.R. 2009. Design of Experiemnts for Validating Probability of Detection Capability of NDT Systems and for Qualification of Inspectors. Materials Evaluation, June, pp. 730 – 738. [11] Generazio, E.R. 2011. Validating Design of Experiments for Determining Probability of Detection Capability for Fracture Critical Applications. Materials Evaluation, december, pp. 1399 – 1407. [12] Generazio, E.R. 2011. Binomial Test Method for Determining Probability of Detection Capability for Fracture Critical Applications. NASA/TP–2011–217176. National Aeronautics and Space Adminstration, Langley Research Center. 33 p. [13] Rummel, W.D., Matzkanin, G.A. 1997. Nondestructive Evaluation (NDE) Capabilities Data Book. 3rd. ed. Nondestructive Testing Information Analysis Center. NTIAC-DB-97-92. NTIAC-1997-A286978. 625 p. [14] Shepherd, B., Gandossi, L., Simola, K. 2009. Link Between Risk-Informed In-Service Inspection And Inspection Qualification. ENIQ Report No 36, European Communities, DG JRC, Institute for Energy, EUR 23928 EN, ISSN 1018-5593.

Effect of the different artificial flaw type in qualification purposes from the point of view of the inspection company Krunoslav Markulin, Matija Vavrous, INETEC - Institute for Nuclear Technology Ltd. Mika Kemppainen, Iikka Virkkunen, Kaisa Miettinen, Trueflaw Ltd. Raimo Paussu, Jani Pirinen, Fortum Power and Heat Oy ABSTRACT A comprehensive in-service inspection programme of the primary circuit components assures their safe and reliable operation with resultant benefit to the overall nuclear safety. Qualification of in- service inspection of nuclear power plant primary components is a powerful tool which provides confidence that a given in-service inspection is fit for its purpose. Through the process of qualification the estimate of the defect detection and sizing capabilities including defect characterization is provided. For that purpose, representative specimens with different type of defects are manufactured in order to make sure that in-service inspection will have capabilities to successfully address potential degradation and deterioration of primary circuit components. Moreover, in order to reliably simulate real inspection conditions, qualification specimens need to fully correspond to the intended areas of control, they need to be made of the same or similar material from which the subject component is made of and finally need to have the same constructional elements including the welded joints. Implementation of the defects in such specimens is of great importance and could affect the chemical composition, structure, physical and mechanical properties which later on affect the inspection system capabilities to locate, characterize, size, evaluate and adjudge defect types and detectable defect sizes with previously identified accuracy. This article will introduce defects manufactured in three different ways to be used in qualification purposes and their comparison from the view point of the inspection company, i.e. detectability and sizing capability including advantages and disadvantages of the different defect types will be presented and analyzed coupled with experience from the actual specimens. INTRODUCTION In nuclear industry periodical maintenance through application of non-destructive testing is essential activity for ensuring the safe and reliable operation of plant components and nuclear power plant as a whole unit. Maintenance activities as in-service inspections and non-destructive testing performed during outages are important to efficiency of outage and entire operation of the plant. Constant demand is that maintenance and in-service inspections is performed in more efficient manner with less expense of time. In order to keep high level of quality of in-service inspection and high level of reliability of non-destructive testing, one of the tools used are qualification proceedings. Through qualification process, validation is performed if selected in-service inspection with related non-destructive testing methods has sufficient capabilities to meet requirements defined with applicable standards and codes. Within qualification process, practical trials are key segment to verify inspection capability. To make practical trials worthwhile, used qualification samples must be representative. Representative covers similarity to inspected component and defects that will realistically represent types of degradation that exists on actual component in real operating conditions. Methods of creating, manufacturing or implementing the artificial defects that will be used for qualification and assessment of inspection is critical because on success of selection and implementation of defects success of qualification will be affected. DEFECTS FOR QUALIFICATION Nowadays, number of different flaw manufacturing techniques are available to the flawed test pieces needed for qualification. Over the years, more techniques have become available and each technique has developed. At the same time, both the inspection techniques and qualification practices have developed. This has resulted in increase of the quality of the test flaws and test pieces. At first, most samples contained mechanically machined flaws, i.e. flat bottom holes, saw cuts and later EDM (electro-discharge machined) notches. These can be easily manufactured to tight tolerances. However, as NDE methods and requirements developed, it soon became apparent that their representativeness was not good enough for service induced cracks. The techniques were developed to decrease the opening to minimum and to introduce surface roughness and out-of-plane surface forms. EDM notches are widely used, e.g., for signal calibration due to their very high repeatability and cost effectiveness. EDM notches are sometimes the only flaw production technology applicable. Also, the EDM notches do not cause any changes in surrounding microstructures that might affect the inspection. To offer better representativeness, various welded flaw simulations have been developed. These are manufactured, in simple terms, by either implanting an existing crack by welding to the test piece, or by inducing cracking of the weld by carefully chosen weld parameters. The welded in flaws offer better simulation of the tortuous crack paths of real service-induced cracks. They can be manufactured with low cost, especially during test block manufacturing. In particular, for large cracks they offer good compromise between representativeness and cost. Finally, to offer even better representativeness and to avoid microstructural alterations that might affect the inspection, new techniques have been developed. These employ a true damage mechanism, which is accelerated and controlled to allow artificial production of service induced flaws. These cracks offer very good representativeness for wide range of service induced cracks and NDT techniques. They also do not cause microstructural alterations, because there's no welding involved. Especially, the delicate crack tips are retained and crack opening is typically much smaller than it is for other techniques. Figure 1 shows the development of different flaw manufacturing techniques on time scale. In many cases, a combination of different flaw manufacturing types is used for a given qualification project. Figure 1. Development of flaw manufacturing techniques for NDE training and qualification. DESCRIPTION OF SAMPLES WITH DEFECTS For present study, four different samples were selected, each containing different artificial flaws. The first three (designated YC033, YC035 and YC036), the representative component was control rod drive housing. For the fourth sample (designated YC015), the representative component was reactor pressure vessel and, in particular, the inner cladding for the RPV wall. The different samples are shown in Figures 2 and 3. Example surface images for thermal fatigue flaw type is shown in Figure 4. YC033 (pipe / flange) Test block is a cut off from a real control rod housing. This is the worst case considering flaw manufacturing population. There is very limited access to inside surface of weld. Grinding of openings and seeing inside was not so easy. Welding of solidification crack was possible only by welding using mirror. Tiny hands of lady welder made it possible to get inside and see the mirror while welding and feeding filler metal. The test blocks contains solidification cracks and EDM notches. Due to the manufacturing difficulties, the quality of the solidification cracks is questionable. YC035 (small pipe / reducer) Test block is a cut off from real control rod housing. The test block contains thermal fatigue cracks and EDM notches. YC036 (small pipe / flange) Test block is a cut off from real control rod housing. The test block contains solidification cracks and EDM notches.             Figure 2. Test pieces YC033, YC035 and YC036, which represent various inspection targets in the control rod drive housing.   Figure 3. Test pieces YC015 and YC030, which represent the reactor pressure vessel cladding area. Figure 4. Example thermal fatigue flaw surface image found in one of the test pieces. INSPECTION COMPANY ASSESSMENT A pulse-echo and phased-array probes with nominal frequencies ranging from 1 to 4 MHz were used for nondestructive testing of defects manufactured for qualification of automated examination of different nuclear power plant primary components. Nondestructively tested qualification specimens were typically made of austenitic titanium stabilized stainless steel, designated by 08X18H10T similar to AISI 321 or W 1.4541, or 22K material (carbon steel), or low alloyed chrome-molybdenum vanadium steel of grade 12X2MΦA or 15X2MΦA. Consequently pulse-echo and phased-array probes were designed for examination of these materials. The available specimens had the range of different defect values and length dimensions provided the good source for determination of the detection and measurement accuracy, with the great number of defects manufactured in three different ways which is sufficient for extensive analysis. The following observations have been drawn after extensive ultrasonic data analysis of numerous specimens. As the material surrounding the EDM notch is left in the original state and the width of EDM notch is usually much greater than the width of true crack, an ultrasonic echoes from such geometric reflectors provides clear tip signals without noise that may affect the depth sizing and consequently result in the easiest detectability and sizing capability of NDE procedure when compared to the other defect types used for qualification purposes. Figure 5 shows the ultrasonic response from such geometric reflector of nominal depth 5 mm located on opposite surface to the scanning one. A clear tips signals are typical for such reflectors and represent 1 - upper tip response seen as OD surface reflection, 2 - corner trap response and 3 – upper tip response seen directly.     Figure 5. Typical ultrasonic echo from the EDM notch However, because ultrasonic echo received from EDM notch is not well correlated with the echo received from true crack, its use in evaluation of inspection system capabilities in true crack detection and length and depth sizing is questionable. The reason for that lay in the fact that the ultrasonic echoes received from true cracks which have rough surfaces tightly closed or irregular surfaces with extremely sharp edges with specific crack profile cannot be well compared to smoothly machined EDM notch profile having constant width and depth dimensions. On the other hand, the constant width and depth dimensions is the main advantage of this type of reflectors as it result in constant and repeatable ultrasonic echo along EDM notch length. This advantage is very useful in calibration and calibration verification of ultrasonic testing systems. Further advantage of using the EDM notches, which are standardized across the industry, is the easier comparison of the results of qualifications and later on inspections performed at different sites. Echoes from in-situ produced thermal fatigue cracks reflectors also provides very clear tip signals without surplus noise, as shown in Figure 6, and as a result detection probability and sizing ability is very similar to one achieved in ultrasonic examination of EDM notches. In opposite to EDM notches profile, the in-situ produced thermal fatigue cracks, as well as welded solidification cracks, have profile which is not constant in depth and width along each defect length, so the ultrasonic echoes received from such defects, especially in application of automated examinations are more comparable to echoes received from true cracks, so inspection company realize these as crack like flaws.   Figure 6. Typical ultrasonic echo from the in-situ produced thermal fatigue cracks, side and top view Because an automated ultrasonic examination has to be performed with qualified scan pattern parameters, for example scan, index, or overlap, these are subject to qualification. An index resolution is parameter highly affected by the defect profile and critical defect length previously specified by the plant owner. From the index dimension, a total length of defect is determined by observing and counting sweeps having proper, discernible echo-dynamic signal as seen on Figure 7. Figure 7. Bidirectional scan, example defect depth profile and thermal fatigue defect top view     DEFECT    DEPTH D EFEC T LEN G TH   DE FE CT    L EN GT H   The index dimension is closely correlated to the root-mean-square (RMS) value of depth sizing errors from ultrasonic data. As index dimension increases, defect depth estimation decreases, as well as RMS value of depth sizing errors increases because of defect depth changes along the defect length. In order to estimate defect depth with previously identified accuracy, inspection company NDE procedure has to take above in account during the procedure qualification if the EDM notch is only defect type available. Clearly, this is very important advantage of both the in-situ produced thermal fatigue cracks and welded solidification cracks over the EDM notches when used for qualification and assessment of inspection. Regardless the shapes of the in-situ produced thermal fatigue cracks and welded solidification cracks are crack like, though ultrasonic echoes are not exactly the same to echoes from true cracks. Echoes from welded solidification cracks are pretty noisy, shown in Figure 8, most probably because of the way they are manufactured, by either implanting an existing crack by welding to the test piece, or by inducing cracking of the weld by carefully chosen weld parameters. This surplus noise requires additional efforts for inspection company to distinguish tips from unwanted noise and, as a result do not accurately represent types of degradation that exists on actual component in real operating conditions.     Figure 8. Typical ultrasonic echo from the welded solidification cracks Examples of surplus noise can be seen on aid piece flaws where used method of creation of flaws is solidification combined with use of aid piece. This way flaws simulating lack of fusion or thermal fatigue cracking can be achieved. In case that aid piece flaw is simulating lack of fusion defects, aid piece is tack welded against straight surface tightly and pressed tightly surface to surface. Aid piece ends are welded carefully and also the rest of opening. Figure 9 presents response from aid piece lack of fusion flaw. Figure 9. Typical ultrasonic echo from flaws produced by aid piece simulating lack of fusion defect Lower and upper tip are present and seen without noise from the side of straight cut surface. The other direction response has higher level of noise due to implanting welding. Alternatively, similar response is received when aid piece is used for creating a flaw that simulates thermal fatigue cracking. Similar approach to produce fatigue crack into straight cut surface by welding and bending aid piece weld produces crack which is fixed tightly face to face and welded into opening. Figure 10 presents response from aid piece thermal fatigue flaw. Figure 10. Typical ultrasonic echo from flaw produced by aid piece simulating thermal fatigue defect Response from such crack is noisier but it includes the tortuous crack path and crack opening can be controlled by manipulation. The noise is low while scanning from straight cut side and higher from the side of opening welding. The in-situ produced thermal fatigue cracks have been in great extent detected with no material noise similar to one seen at welded solidification cracks and/or other changes visible which could affect the inspection system sizing capabilities and requires additional precautions to be taken during the analysis in order not to miss or misinterpret the defect. Figure 11. Comparison of ultrasonic echo from the in-situ produced thermal fatigue cracks with the EDM notch An additional challenge that Inspection Company has to deal with was in certain number of defects a detection of crack tips which found to be hardly detectable, thus when examination was performed with phased array probes, detection and sizing were found to be more efficient. On the other hand, the approach to qualification which assures more challenging conditions than real operating conditions are welcome and will make sure that in-service inspection will have capabilities to successfully address all potential degradation and deterioration of primary circuit components. CONCLUSIONS Based on achieved results through experimental testing and qualification proceedings characteristically responses for each flaw type used during qualifications were obtained. For each artificial flaw group there is an area of application within qualification sample. Certain flaw types like EDM notch can provide solution due to it’s repeatability and ease of manufacturing while flaws made by complex techniques like thermal fatigue provide flaws that are more representative of actual existing degradations from the standpoint of flaws size and crack tips, or can provide a challenge for inspection applications due to difficulty to adequately detect and size such flaws. Selection of adequate flaw types within test sample is crucial step and overall success of test sample will depend on it. Component that test sample is simulating also plays a role in selection and manufacture of defects. Qualification as a verification process to assess capability of selected inspection techniques is highly defendant on the tests sample and flaws within it, and only if all aspects of artificial flaws, both from component and degradation side and inspection technique side are taken into account will the test sample be successfully prepared and consequently, qualification will be successful and reflect problems and challenges that can occur during actual application of in-service inspection in real conditions. REFERENCES 1) FNS-262714 LO1-K854-963-50, “INPUT INFORMATION: Reactor pressure vessel”, O. Hietanen, R. Korhonen, A. Neuvonen, J. Pirinen, 25 June 2009 2) FNS-280989 LO1-K854-963-60, “INPUT INFORMATION: CRD Protection Pipe”, A. Neuvonen, J. Pirinen, 13 April 2010 3) YVL GUIDE YVL 3.8, „NUCLEAR POWER PLANT PRESSURE EQUIPMENT: In-service inspection with non-destructive testing methods“, 3rd edition, STUK - Radiation and Nuclear Safety Authority, Finland, 22 September 2003 4) Technical Justification for the Loviisa Power Plant Control Rod Drive Protection Pipe Examination, M. Vavrous, K. Markulin, June 20, 2012 5) Technical Justification for the Loviisa Power Plant Reactor Pressure Vessel Examination, M. Vavrous, A. Matoković, June 20, 2012

  1     NEEDS FOR MASSIVE MOCK UPS AND CHALLENGES IN MANUFACTURING OF THEM FOR QUALIFICATION PURPOSES Author: Jani Pirinen, Fortum Power and Heat Oy Co-authors: Raimo Paussu, Veijo Nikula, Fortum Power and Heat Oy, Iikka Virkkunen, Mika Kemppainen, Kaisa Miettinen Trueflaw Ltd. ABSTRACT Fortum operates two VVER-440 nuclear power plants units in Loviisa, Finland. The main focus of the in service inspections of the primary circulation components are in ensuring the reliable and safe operation of the both units. These components are reactor pressure vessel, pressurizer, and steam generators including piping, valves and pumps linked to them. The in-service inspections of components are closely following ASME Section XI requirements. Inspection objects of the components are qualified according to Finnish qualification rules, closely following ENIQ recommended practices. Over the years, Fortum has gathered experience on inspection qualifications and on fabrication of own mock ups and defects. Primary components have significant safety importance. The importance of the integrity of the primary components is the main reason to invest in the development of qualification. This includes development of the manufacturing techniques, purchasing relevant material for the mock ups and finally improving the qualification of the inspection system. Also the fact that main components are effectively not replaceable has to be considered when evaluating the needs for massive mock ups. Massive mock ups are needed for the above mentioned components. It has been said, that the qualification of the inspection system will be as good as the used test blocks. The goodness or effectiveness can be estimate with several parameters. Probable the most significant factor is the authenticity of the mock up and more precisely the geometry and the materials of the test block. As known the main goal in qualification is to perform the capability of the selected inspection system (includes used manipulator, probes, data etc.) and that the inspection system fulfills requirements set up for the inspection. Main focus in this paper is in the challenges in manufacturing massive mock ups, both in general and with respect to two different cases with similar challenges. Also reasons for use of the mock ups have been discussed. To have more practical view of the challenges, these are presented via two on-going cases. Challenges in manufacturing of the massive mock up consist mainly from the way to handle the test pieces; is the mock up transportable to work shop. Is it needed to move the flaw manufacturing systems to the mock up. Preparing defects to emergency cooling nozzle inner radius in full-scale reactor pressure vessel and bottom head of the pressurizer for qualification purposes will be presented as example cases. INTRODUCTION Fortum operates two VVER-440 nuclear power plants units in Loviisa, Finland. The main focus of the in service inspections of the primary circulation components are in ensuring the reliable and safe operation of the both units. These components are reactor pressure vessel, pressurizer, and steam generators including piping, valves and pumps linked to them. Also these primary components have significant safety importance. The in-service inspections of components are closely following ASME Section XI requirements. Inspection objects of the components are qualified according to Finnish qualification rules, closely following ENIQ recommended practices. Over the years, Fortum has gathered experience on inspection qualifications and on fabrication with subcontractors of own mock ups and defects. Tight cooperation with Finnish companies Mekelex (EDM machining) and Trueflaw Ltd (thermal fatigue cracks) has started already early in their starting history and the cooperation still continues and develops. The importance of the integrity of the primary components is the main reason to invest in the development of qualification. This includes development of the manufacturing techniques, purchasing relevant material for the mock ups and finally improving the qualification of the inspection system. Also the fact that main components are effectively not replaceable has to be considered when evaluating the needs for massive mock ups.   2     The main goal in qualification process is to verify the capability of the selected inspection system (includes used manipulator, probes, data etc.) and personnel to do inspections with qualified inspection procedure. During the qualification process, the inspection vendor selects techniques for inspection and also justifies selections in technical justification. Selections shall fulfill the requirements presented in the input information. Functionality of the selected inspection system is performed in open trials and witnessed by the qualification body Capability of the Inspection personnel to do data analyzing is performed with blind trial. In Blind trials the data is collected with selected inspection system. Blind-trials are also witnessed by the qualification body. WHY MASSIVE MOCK-UPS ARE NEEDED In Loviisa power plant cases the principle to use as authentic and representative mock-ups as possible, has been followed as often as possible. The geometry, size and materials of the mock-up should be as close to real situation as possible. In minimum, in the open trials have used at least geometrically as representative mock-ups as possible. Massive mock-ups for qualification process are usually needed in the cases where inspection objects are in the massive components like pressurizer, reactor pressure vessel etc. Modeling is seen also as a potential option. The number of the flaws for qualification can be increased with modeling, but it does not replace the need to use mock-ups. The inspection performance especially in the open trials is needed to verify the inspection system in operation. Performing the actual sound paths is also one of the advantages while using the full scale massive mock-ups. Also, manipulator design results and further development needs can be discovered in testing due to the representative inspection conditions. The goodness or effectiveness of the qualification process can be estimate with several parameters. Probably the most significant factor is the authenticity of the mock up and more precisely the geometry and the materials of the test block. The lack of representative materials is generally the reason that massive test blocks are not used. Years ago Fortum was able to purchase authentic full scale primary components which have been used for qualification purposes in resent year. These components include steam-generator, reactor pressure vessel, reactor pressure vessel head and all the internals. The largest mock-up project up to date was the manufacturing of the collector nozzle to steam generator (dissimilar metal weld). Other large mock-ups were manufactured even before to main circulation pipe inspection qualification purposes. In these components more authentic materials were used . In contrast e.g. to the bottom head mock-up of pressurizer where western type of materials were used. Fortum has been able to use authentic mock-ups for qualification in those cases. Materials and geometry could not be closer to real situation. After decision to use the massive mock-ups in qualification purposes manufacturing activities were started. For first massive mock up case was selected the qualification of the emergency cooling nozzle inner radius (TH-NIR). Before the manufacturing of the flaws to the reactor pressure vessel (RPV) TH-NIR, all manufacturing systems had to be redesigned to be movable. In the first phase the manufacturing started with pressurizer bottom head which was assumed to be simpler than the full scale reactor pressure vessel. Bottom head of the pressurizer was seen as an excellent place for training and testing of the defect manufacturing systems. In addition, the development work of the defects manufacturing systems would not be needed during the defect manufacturing for TH nozzle inner radius. In that second phase the target was to use advanced manufacturing method without time consuming development work. Needed practical training was one of the reasons to manufacture the full scale mock-up for qualification of the pressurizer bottom head inspections. In that case the manufacturing had to start by purchasing the bottom head of the pressurizer. INSPECTION OBJECTS TO BE QUALIFIED Example cases of this paper are related to the qualification of the inspection objects in pressurizer bottom head and emergency cooling nozzle in reactor pressure vessel. Both of these components inspection objects will be externally inspected with ultrasonic examination with long sound path. In bottom head mock-up consists several inspection objects where to manufacture the defects. The most critical inspection object of the bottom head of the pressurizer is the inner radius of surge line nozzle,   3     corrosion sleeve and its fixing weld in addition to dissimilar weld of nozzle. Also fixing welds of thermal sleeve holders that are located near the critical inner radius area are important. In the rector pressure vessel there are four emergency cooling nozzles where to manufacture defects and one inspection object in each nozzle in proportion to the pressurizer inspection objects. Inspection objects are presented in Figure 1 through to Figure 3. Figure 1. Bottom head of the pressurizer on left and example from the nozzles on right Figure 2. Front view from the reactor, TH-NIR in the middle on the left side and challenging inspection environment on right Figure 3. Crossection from the emergency cooling nozzle, inspection object is nozzle inner radius (TH- NIR) REQUIREMENTS FOR THE MOCK-UPS Any special requirements as compared to "smaller" mock ups were not placed. Test block design and selection of flaw types to be manufactured are performed by the utility based on the input data for qualification of the inspections. ASME XI principles have been applied in designing of the mock-ups. Same artificial defect types have been used to massive-mocks up as to other mock-ups. Detection targets and critical flaw sizes have also guided the selection of flaw fabrication methods. Used artificial defect types are   4     thermal fatigue cracks, electro discharge machining EDM notches, solidification cracks and implantation defects. As the different types of defects have their advantages and disadvantages in detection and sizing, whole arsenal of manufacturing methods was desirable The amount of the defects is mostly limited due ASME requirements of the grading unit, which should be at least 75mm. Current requirement has been followed in designing of the defect population. Also requirements presented in the input information for the qualification have been followed. These requirements are, among the other things, the detection target, specific and postulated defect types and assumed locations of them. SPECIAL CHARACTERISTIC OF THE MANUFACTURING THE MASSIVE MOCK-UPS   The challenges in manufacturing massive mock ups can be discussed in general and with respect to two different cases with similar challenges. Also challenges can be described as expected and unexpected difficulties. Expected challenges consist mostly from logistic and handling issues of the mock-up and manufacturing equipment. Difference between the cases consists from the situations where either the mock- up or equipment is moved. Other challenges consist from handling of the mock-ups. Smaller full size mock- ups can be moved and rotated to optimal position for manufacturing. Respectively transportable massive mock-ups e.g. bottom-head of the pressurizer can be moved to manufacturing system at workshop, but used manufacturing system shall be assembled to optimum position for manufacturing. Similar challenges consists in cases where manufacturing has done near mock-up e.g. full size reactor pressure vessel which is not transportable to the subcontractors due to the limitations of the workshops. Respectively field conditions set requirements for manufacturing systems. These are mainly consists from the limited yearly time window and also limited space around the reactor pressure vessel (RPV). RPV and other main components are stored to hall which used as a welding hall during the building time of the Loviisa power plant. Since then the hall has been modified to storage hall. From the utility point of view, when deep surface or sub-surface cracks are needed to be fabricated into finished component surface, welding of solidification cracks is the easiest, cheapest and often representative enough as fatigue cracks, compared to other type of flaw options, to be used in ferrite and austenitic structures. Shallow solidification cracks are hard to produce as surface breaking. Logistic and handling issues, limited yearly time window among other challenges are presented in more detail via two similar cases. Challenges in flaw manufacturing are also presented. MANUFACTURING OF THE PRESSURIZER BOTTOM HEAD MOCK-UP FOR INSPECTION QUALIFICATION   The simulation of the real inspection situation requires the same thickness as the bottom has. Also The frame structure around the bottom and the legs were designed at Loviisa NPP to simulate the inspection circumstances under the bottom. The solution was to order bottom from Germany with the same dimensions and shape as the pressurizer bottom. Welding of the cladding and manufacturing of the frame structure were done in Finland, Figure 4. Figure 4. Half-finished bottom on the left and frame structure on the right.   5     Expected challenges were considered from the transportation and handling issues point of view in the workshops. Heavy lift transportation itself isn't a problem thus the pressurizer bottom head with the frames weight about 7 tons. Moving and handling the mock-up inside the workshop were assumed to be challenging, because a small fork lift would not be able to lift samples and bridge crane was not available. Moving of test block inside the workshop was settled by using U-profiles beams as a rails with roller set to roll along guides on the floor. Expected challenges also observed to cause from undersized working facilities. Work shop facilities were updated to increase the work safety by improving ventilation and assembling curtains from floor to roof to limit grinding dust. Unexpected challenges caused from the repair and grinding of the cladding. The work took about four months of production time and caused extra costs. SAW welding of cladding layer on inner surface of bottom was a challenge for the welding company. Poor connection of parallel, adjacent strip welded passes and many welding defects in cladding caused long grinding job before the bonding of cladding could be inspected. Also due to the disturbing grinding work behind the curtains, other test blocks were not possible to be manufactured at the same time. As mentioned, several types of artificial flaws have been manufactured. Used methods have been solidification cracks, aid-piece flaws, thermal flaws and EDM-notches. The deep cracks were produced by welding solidification cracks and shallow cracks with thermal fatigue by Trueflaw Ltd. Also narrow EDM notches with shape of crack front have been used. Unexpected challenges did not occur during the flaw manufacturing by welding methods. Moisture were expected to occur during the flaw manufacturing and it removed with heating the mock-up. Remove of the hydrogen were daily time-consuming task. Anyhow work was necessary to do to avoid unexpected flaws. Manufacturing of the EDM-notches did not cause either an unexpected challenges. Typical challenges which are expected also during normal EDM machining. This were e.g erosion of the electrode, re assembling of the electrode due to malfunction and impropriate parameters for current task. Usually parameters are needed to adjust during the manufacturing depending the material properties and size of the manufactured flaw size. To have acceptable process and tolerances to flaws in EDM machining, the head of the EDM machine shall be sturdy assembled. During the manufacturing process the head is not allowed to vibrate. Assembly system of the EDM machining is presented in the Figure 5. Figure 5. Assembly of the EDM machining system for YP-pressurizer bottom head flaw manufacturing. Adjustment movements are marked with arrows. Fluid  pool     6     For Trueflaw, the pressurizer bottom was the largest mock-up to date. This caused some need for special arrangements. In particular, moving the heavy mock-up from transportation vehicle into suitable location at the shop floor and back to transportation vehicle required special attention (see Figure 6). The production location and transportation was well planned ahead and went accordingly. For flaw production, the mock-up size caused some additional challenges. The mock-up could not be rotated to optimal location for accessibility. Thus, the designated flaw locations were out of reach for normal flaw production tools. Special tooling with extended arm length was developed for this application. The extended arm also resulted in increased tool flexibility, which necessitated some additional fixtures to allow reliable flaw production. Although the need for further tooling was foreseen, it was also recognized that the specific needs may change due to changes in flaw location or sample orientation. Thus, the tooling was completed when the sample was already in place and tools could be adapted to the exact requirements of the situation. Overall, combination of good preparation and adaptability made producing cracks to massive mock-up successful without major extra effort. Figure 6. The Massive mock-up rolled into Trueflaw facilities. Modification of the assembly system during the manufacturing is also expected. Manufacturing process is also a learning process and observed limitations in the design are addressed. As mentioned the mock-up cannot move to optimum position, the manufacturing system shall be moved. In some cases the assembly is not flexible enough due to the temporary assembly of the machining head. MANUFACTURING OF THE DEFECTS TO TH-NOZZLE INNER RADIUS INSPECTION QUALIFICATION Based to experiences form the flaw manufacturing to the pressurizer bottom head mock-up and other mock- ups during the resent year, flaw manufacturing to authentic RPV was ready to begin. In this time the flaw manufacturing systems were moved away from workshop to the field conditions. Manufacturing was done in the storage hall at the Loviisa power plant area. In field conditions more detail design for manufacturing arrangements and flaw placement. Thus the mock-up is lying down, all manufacturing systems have their optimal grades which to use. Also the work safety issues, manufacturing of the defect or documentation procedure were observed more detail than in workshop manufacturing. First preparation work was done already on the 2008. Then the thermal shields from the nozzles were removed. During that work the firs unexpected challenge occurred. Machining work were notable larger and took longer than expected due that the thermal shields were fully welded to RPV. Assumption was that there is only sealing weld and screw fixing. Work with RPV continued on 2012 after manufacturing of the pressurizer. Trueflaw started the flaw manufacturing work at the end of the summer 2012. Unexpected challenges were observed and also solved during the manufacturing processes due to the environment. The weather conditions were expected to cause challenges during the manufacturing of the defects to reactor pressure vessel. One of the biggest challenges was that the manufacturing time window had   7     to be scheduled to the summer / autumn time. Used storage hall for manufacturing is equipped only with the power supply and is without heating system. In generally manufacturing systems worked as expected. Mostly challenges were related to the auxiliary systems which are usually permanently installed to function in work shop. Thermal fatigue flaws and EDM-notches are manufactured with remote control and surveillance system via mobile phone net. To ensure reliable connection to net took more time than expected. During the EDM manufacturing similar requirements were not needed than during the fully remote controlled thermal fatigue flaw manufacturing. For Trueflaw, the most significant challenge was moving the production to remote location and inside the reactor pressure vessel. Although the flaw manufacturing machines ("flawmakers") were designed to be modular and transportable, this had never been tried before. Also, all the flawmakers are connected to central control computer and central database. This provided several possible configurations for remote operation and it was decided to use single central database for all machines while using separate control computers for both locations. This allowed remote monitoring and control of the units while still supporting continued operation in case of intermittent network errors. The flawmakers are also designed to monitor possible error conditions and stop in case of error. The remote machines were connected to central database and control computers through 3G mobile data network. The arrangement was "dry-tested" in the shop floor to make sure it operated as planned. The dry-tested arrangement was then moved to production location. The arrangements at the site were good and thus getting the machines into location and re-connecting the machines went well. Figure 7 shows the production machines in the location, inside the reactor pressure vessel. After production started, network connectivity provided additional challenges. While the mobile network had worked well during testing, in actual location the link was more error prone and network hardware experienced intermittent crashes that required rebooting the devices. The remote location also increased the occurrences of some (rather harmless) error conditions. This combined with network problems caused excessive machine down time and slowed flaw production in the beginning. It took numerous trials until robust network configuration could be established and the final system included both better hardware as well as local link-status monitoring and automatic recovery. Furthermore, as the final locations of the flaws were decided, it became evident that some of the planned production optimizations could not be used. Figure 7. Flawmakers on location at Loviisa, inside reactor pressure vessel. After the initial problems, the production advanced smoothly. However, the combined effect of the early delays provided its own problems. Over time, the approaching winter caused outside temperature to drop (sometimes to -20°C or below). This caused the working conditions to deteriorate. Furthermore, the flawmakers started to experience cooling water freezing problems. These were remedied by local heating and addition of anti-freezing agent to the cooling water. Despite these final problems due to outside temperature, the flaw production was completed successfully and the production quality for the flaws is as good as it would have been in the shop.   8     Manufacturing activities continued again on the summer 2013 with solidification cracks. After that all EDM notches have been manufactured. Changes in material properties caused an unexpected challenges during the openings for welded flaws. Due to the hardness differences from expected the used manufacturing tools were needed to be changed. Local heating was also needed to avoid moisture based hydrogen problems. Similar challenges in assembly of the EDM machining occurred as during the machining of the pressurizer bottom head. Also expected normal machining challenges at the beginning of the machining had to be solved. After optimal parameters were reached the position of the EDM head has been the most challenging task to do. In this case the modification needs could not be excluded to predesigning. General view from the assembly of the EDM manufacturing system and assembly of the EDM machining head are presented in the Figure 8. Parts of the system is located to outside of the reactor pressure vessel, the air filter of the local ventilation and fluid collector (see the Figure 9). There can be seen the plugged nozzle which is also used as the pool for fluid. Figure 8. General view from the assembly of the manufacturing system on left and assembly of the EDM machining head on the right Figure 9. General view from the assembly of the manufacturing system outside of the reactor pressure vessel. In the front right is fluid collector and on the left side behind is the air filter of the local ventilation CONCLUSION Massive mock ups are needed to verify inspection procedures with qualification process to ensure integrity of the components. As mention the geometry and used material should be as authentic as possible for all inspection mock-ups. Especially authenticity is probably most effective factor in the open-trials where selected inspection technique is demonstrated to fulfill determined requirements. Under the circumstances the massive-mock ups are used and manufactured for qualification purposes of the primary components inspection objects. In addition to verified inspection procedure, the quality of inspection system and most of all the nuclear safety are increased   9     At the moment Fortum has under construction with the subcontractors the biggest mock-up than before manufactured –full scale Reactor Pressure Vessel. RPV mock-up will be used for emergency cooling nozzle inner radius inspections. Later plans for the RPV-mock up for qualification purposes have been also considered. Many kind of artificial flaw types have been manufactured to the mock ups, more detail designing were done compared to the previous manufacturing projects. In both presented cases, the mock-up could not be moved to the optimum position for manufacturing. Also sector locations were chosen in design according to the limitations of the used manufacturing system. Experience gathered in resent year from manufacturing methods were notable advantage for designing and manufacturing of the mock-ups. During the manufacturing processes both expected and unexpected challenges were met. Unexpected problems were caused mainly from auxiliary systems. Movable flaw manufacturing systems were developed without any massive extra effort. Afterward can be even said that it were easier than expected. Logistics and handling issues related to the moving of the component or the manufacturing systems were solved in both cases. Also challenges with the auxiliary systems were met. Other challenges in the field condition were mostly caused from the environment of the facilities due to the limited time window for all year defect manufacturing. Although manufacturing were the first step in the process to have accepted mock-up for qualification. Finger print work will be also challenging, due the size of the mock-ups. There are only few companies which are able to give the feedback from the manufactured artificial flaws when scanning is performed externally. At the same time there are only few companies which can be qualified to do the actual inspection. This is one of the challenge which is still open. Qualification process continues and more detail results of the quality of the mock up will be received in near future after finger print and open trials. Fortum and subcontractors are more capably to carry out similar projects due to the gained experience from the presented cases.

NUMERICAL MODELING OF THERMAL FATIGUE CRACK FROM VIEW POINT OF EDDY CURRENT TESTING Jing WANG1 ∗, Noritaka YUSA1, Mika KEMPPAINEN2, Iikka VIRKKUNEN2 and Hidetoshi HASHIZUMET1 1 Department of Quantum Science and Energy Engineering, Graduate School of Engineering, Tohoku University, 6-6 Aramaki Aza Aoba, Aoba-Ku, Sendai, Miyagi 980-8579, 2 Trueflaw Ltd, Tillinmäentie 3, tila A113, 02330 Espoo, Finland Abstract Eddy current testing (ECT) is one of electromagnetic nondestructive method. It was initially regarded as one kind of qualitative measurement during the inspection of natural cracks, such as stress corrosion crack (SCC), thermal fatigue crack (TFC) et al. On the other hand, ultrasonic testing (UT) is widely used for quantitatively detecting cracks. However performance of UT is inevitably affected by complexity of natural cracks, e.g. fracture surface roughness and so on [1]. Therefore, enhancement of inspecting natural cracks by other non- destructive method is demanded. ECT has become a favorable candidate for the aim with the aid of great progress of electromagnetic computational technology [2]. In fact quite a few studies have been reported success in quantitative evaluation of SCC and also they pointed out that modeling of crack is the issue for the accuracy measurement [3,4,5]. Based on this, modeling of TFC should be emphasized for the accurate eddy current inspection. However, only few studies have been discussed on the matter [6,7]. Further researches by more specimens are needed. The present study discusses numerical modeling of thermal fatigue crack (TFC) from view point of eddy current testing. Four TFCs artificially introduced into type SUS316 stainless steel plates are prepared for the study. Eddy current signal are gathered by a differential type plus point probe with three frequencies of 100 kHz, 200 kHz and 400 kHz. Subsequent destructive test is carried out to reveal the profile of cracks. TFCs are modeled as a region with constant width, uniform conductivity and real profile in numerical simulation by software Comsol Multiphysics 4.4. The basic geometry of simulation is shown as figure.1. With consideration on previous studies [6,7], numerical modeling of TFC in type SUS304 stainless steel and Inconel 600 plates, results of the study demonstrate that TFC should be modeled as an almost nonconductive region in general no matter how the frequency is used. Figure 1. Basic geometry of simulation References [1] M. Kemppainen, I. Virkkunen, Crack characteristics and their importance to NDE, Journal of Nondestructive Evaluation 30(3)(2011), 143-157. ∗ Corresponding author. Phone:81-8096300309, Fax:81-227957906, E-mail address:jingwang8604@gmail.com [2] BA. Auld, JC. Moulder, Review of advances in quantitative eddy current nondestructive evaluation, Journal of Nondestructive Evaluation 18(1)(1999), 3-36. [3] N. Yusa, Z. Chen and K. Miya, Sizing of stress corrosion cracking on austenitic stainless piping in a nuclear power plant from eddy current NDT signals, Nondestructive Testing and Evaluation 20(2)(2005), 103-114. [4] N. Yusa, Z. Chen, K. Miya, T. Uchimoto and T. Takagi, Large-scale parallel computation for the reconstruction of natural stress corrosion cracks from eddy current testing signals. NDT&E International 36 (2003), 449-459. [5] Z. Badics, Y. Matsumoto, K. Aoki, F. Nakayasu and A. Kurokawa, Finite element models of stress corrosion cracks (SCC) in 3-D eddy current NDE problem. Nondestructive testing of materials. IOS Press (1995), 21-29. [6] J. Wang, N. Yusa, H. L. Pan, M. Kemppainen, I. Virkkunen and H. Hashizume, Discussion on modeling of thermal fatigue cracks in numerical simulation based on eddy current signals, NDT&E International 55 (2013), 96-101. [7] J. Wang, N. Yusa, H. Pan, M. Kemppainen, I. Virkkunen and H. Hashizume, Modeling of thermal fatigue crack for enhancement of electromagnetic nondestructive evaluation of nuclear power plant, The 21st International Conference on Nuclear Engineering. (2013). Acknowledgements This study was conducted as part of the Aging Management Project for System Safety of Nuclear Power Plants commissioned by Nuclear Regulation Authority of Japan.

Production of Real Flaws in Probability of Detection (POD-) Samples for Aerospace Applications Mika KEMPPAINEN*, Iikka VIRKKUNEN* *Trueflaw Ltd., Espoo, Finland (+358 45 635 4414, mika@trueflaw.com) Abstract. Assessment and demonstration of NDE reliability is important part of the inspection system. Capability is commonly assessed in terms of probability of detection (POD) as function of crack size (target size). For performance demonstration, number of cracks (with different size) is inspected and POD estimated from acquired inspection results. In order to provide reliable capability demonstration, the demonstration should closely mimic real life inspection. The cracks used in demonstration should closely resemble real-life defects. The component geometry, material and other features that may affect inspection reliability should also be representative of real inspection. Historically, it has been challenging to produce relevant test samples with high number of representative defects. Trueflaw Ltd. produces real thermal fatigue cracks to NDE-applications. These cover different inspection techniques (UT, EC, X-ray, penetrant, magnetic particle, etc.) and whole range of inspection objects from nuclear to aerospace applications. In this paper two aerospace cases are presented, where the clients needed real flaws in their POD-samples. Flaws were to be produced in-situ to the original samples. The first case was produced for a jet engine manufacturer in Germany. In this case the sample was a turbine disk supplied by the client. Material of the disk was Inconel alloy. Totally 60 flaws were produced in the original sample by Trueflaw’s in-situ crack growth process. Flaw sizes and locations were specified by the client and followed in the production. After delivery of the sample, the client used it in his POD-trials to assess the capability of his inspectors. The other case was for Rolls-Royce plc, UK, requiring a range of defects in the actual component, as opposed to representative test pieces. By using the Trueflaw method the cracks could be placed accurately in the complex geometry. In total 14 flaws were produced in the original POD-sample by Trueflaw’s in-situ crack growth process. In this paper the successful results of flaw manufacturing in the two cases are presented and discussed. Also, the probability of detection results of the first case is analysed and discussed. 1. Introduction Assessment and demonstration of NDE reliability is an important part of an inspection system. Capability is commonly assessed in terms of probability of detection (POD) as function of crack size (target size). For performance demonstration, number of cracks (with different size) is inspected and POD estimated from acquired inspection results. In order to provide reliable capability demonstration, the demonstration should closely mimic real life inspection. The cracks used in demonstration should closely resemble real-life defects. The 4th International Symposium on NDT in Aerospace 2012 - Th.1.A.2 License: http://creativecommons.org/licenses/by/3.0/ 1 component geometry, material and other features that may affect inspection reliability should also be representative of real inspection. Historically, it has been challenging to produce relevant test samples with high number of representative defects. The structural significance of flaws increases, in general, with increasing crack size. Thus, any available NDE system is developed to find as small cracks as possible as the value of the inspection increases with smaller detected cracks. At the same time, the acquired signal response and the ease of detection for many NDE techniques decreases with decreasing flaw size. Thus, the reliability of crack detection and hence the value of the inspection decreases, in general, with decreasing flaw size. Consequently, it is of interest to find the smallest flaw sizes, where the inspections can still provide sufficiently reliable results. Many NDE methods are used to detect flaw sizes in a range, where the probability of detection is still high, but significantly below one. This is where the POD curves and, in particular, the lower limit POD curve is used to plan inspections. Significant amount of data is needed to reliably estimate the POD as function of crack depth. Data is needed both in terms of repeated inspections and in terms of different flawed samples inspected. Both the inspections and cracked samples are costly. Thus the main challenge in POD determination is to extract maximum amount of information from limited data and to get realistic lower limit POD curve. The best practices for attaining a POD curve are well established in the aerospace industry and are thoroughly documented in MIL-HDBK-1823 [1]. The MIL-HDBK-1823 approach is based on seminal work by Berens [2] and relies on rather sophisticated statistical methods to attain the lower limit POD curve for a given data. These methods significantly reduce the amount of needed data, as compared to simpler, previously used, methods. Nevertheless, preferably 60 and no less than 40 cracked samples with various crack sizes are needed to reliably estimate the POD curve. Due to this high number of samples required (among other things), the MIL-HDBK statistical approach for POD determination has not so far found widespread use outside the aerospace industry, albeit there's been recent increase of interest for it's application in the nuclear industry [3]. In addition to the sheer number of cracks, determining POD curve contains other challenges: the validity of the obtained POD curve depends on the representativeness of the used test samples and inspection procedures. Manufacturing representative test pieces has traditionally been quite problematic. The test specimens should reflect the structural types that the NDE process will encounter in application with respect to geometry, material, part processing, surface condition, and, to the extent possible, target characteristics. In practice, producing realistic cracks in samples with realistic geometry and material has rarely been achieved. Dye penetrant inspection and eddy-current inspection are commonly used for the detection of service-induced cracks in aerospace applications. Flaws to be detected are very small, as compared, e.g., to the nuclear applications. Both dye penetrant and eddy-current inspections are slow to perform and suffer from subjective interpretation of the results. Weekes et al. [4] show the results of POD studies for dye penetrant and eddy-current inspections. New techniques based on active thermography, such as eddytherm and thermosonics (sonic IR, ultrasound-stimulated thermography), have the advantage of rapidness and automation in the interpretation of the results. Both of these techniques have shown high sensitivity in detecting small cracks. Weekes et al [4] reports the probability of detection (POD) for eddytherm and compared the results to similar POD studies with dye penetrant inspection, eddy-current inspection and thermosonics (see Figure 2, where the tabulated results of Weekes et al have been drawn to show the POD90(a) curves). The aim of a POD study is to determine the lower limit probability of detecting cracks as a function of crack size. Weekes et al [4] showed that, in addition to that 2 eddytherm is very rapid, it has very high sensitivity to small cracks. The results show that, when applied and analyzed automatically, eddytherm can detect sub-millimeter cracks with a high degree of confidence, as shown in Figure 1. Furthermore, the comparison results of Weekes et al [4] show that both eddytherm and thermosonics have quite similar capability to detect sub-millimeter cracks, Figure 2. Hence, the choice between these two techniques is recommended to be based on their respective practicalities (e.g., area to be inspected, requirement for non-contacting inspection, etc). Comparison to more conventional NDE methods; dye penetrant and eddy- current inspections (see Figure 2), showed that dye penetrant can detect even smaller cracks than eddytherm, but may miss bigger ones due to over-washing. In eddy-current inspection the sensitivity can exceed eddytherm’s sensitivity in tightly controlled inspections, but a manual inspection (most of the practical cases are performed manually) is typically less sensitive than eddytherm. Figure 1 POD curves of Eddytherm for Steel, Titanium and Waspaloy. [4] Figure 2 Curves estimated based on a90 values given by Weekes et al. [4] 3 Trueflaw’s unique technology has been used widely in training and qualifying inspectors for inspecting nuclear power plants (NPP’s). Trueflaw’s technique has been used to produce true thermal fatigue cracks to samples provided by clients. In each case the flaw characteristics (length, depth, opening, location, skew, tilt etc.) were determined beforehand. Also the tolerances used have been tight and they have been successfully reproduced in all cases. Samples have been full-scale mock-ups, actual samples from NPP’s, and different simplified training samples. In all samples, flaws were needed without any additional disturbances to the ready-made samples than the natural crack. The published cases (see e.g., [5]) show that Trueflaw’s technique fulfils the requirements in the nuclear engineering applications. Trueflaw’s technique has been used only in few development cases for aerospace applications, but never in aerospace POD samples. The aim of this study was to determine if Trueflaw could produce flaws to full-scale aerospace POD samples. Normally, the POD studies are done using small flat samples. Those samples do not include the challenge provided by the actual geometry. Hence, Trueflaw technology would allow aerospace companies do POD studies including the demanding challenge of actual geometries. 2. Materials and Methods In this paper two aerospace cases are presented, where the clients required real flaws in their POD-samples. Flaws were produced to real samples supplied by the clients. These samples were actual components that were similar to the ones used in the actual location. Trueflaw Ltd. has a unique process where the flaws can be placed accurately, in-situ to the customer’s complex samples. See more detailed description in [5, 7, 8]. Trueflaw utilizes patented technology where the flaws are induced by using high frequency induction heating and water spray cooling. By repeated heating – cooling cycles, the flaws are created to the ready-made samples without any additional preparation or modification of the samples. The flaws produced are real thermal fatigue cracks with accurately controlled location and flaw characteristics (length, depth, opening, etc.). Flaws are used in different NDE-applications from nuclear to aerospace industries covering different inspection techniques (UT, EC, X-ray, penetrant, magnetic particle, etc.). Flaws are used to assess the performance of the technique used in inspecting or monitoring the true condition of a component. 2.1 Sample A - German Jet Engine Manufacturer The first case presented is flaw production in the POD sample of a German jet engine manufacturer (it is not allowed to release the name of the client). The sample was a turbine disc made of Inconel alloy. Flaws were produced in this sample by Trueflaw’s in-situ crack growth process. Flaws were produced in the locations and with characteristics specified by the client (see Table 1). 2.2 Sample B – Rolls-Royce Plc. UK The other case was for Rolls-Royce plc., UK, requiring a range of defects in the real component, as opposed to representative test pieces. Rolls-Royce plc. was developing a novel inspection technique for the inspection of defects under coatings. Normally, POD studies are carried out with tens of flat test pieces with mechanical fatigue cracks. These samples are small and do not represent the actual geometry. So, as the normal way is to use mechanical fatigue flaws in very simple samples, Rolls-Royce plc. saw here the opportunity 4 to have real flaws in their actual component. They saw that the advantage of Trueflaw’s technique is the ability of placing flaws freely into any geometry and therefore carry out the POD study taking in account geometric issues. Rolls-Royce is developing techniques for inspecting through coatings, including thermal NDE techniques. For developing the new technology, they needed a sample with known crack population. The inspection target was a seal fin specimen of a turbine disc, which is covered with a wear coating while in use. Cracks were needed under the coating in the same way that they would appear during use (see figure 1). Figure 3 A schematic illustration of the seal fin region [5]. Due to complicated geometry, it would have been impossible for Rolls-Royce plc to use mechanical fatigue to generate cracks in the real component. Furthermore, in this case using flat test pieces would have been so different from the actual geometry that it would have yielded any POD study meaningless [6]. By using the Trueflaw method the cracks could be placed accurately in the complex geometry. Flaws were produced in the actual POD sample by Trueflaw’s in-situ crack growth process to allow inspection of all interesting locations. Table 1 Samples and number of flaws produced in this study. Sample code Number of flaws Location of flaws Flaw characteristics Sample A 60 Including hole corners, fillet radii, etc. Range of crack sizes suitable for POD demonstration, Sample B 14 Several locations of the sample representative crack characteristics 3. Results 3.1 Sample A - German Jet Engine Manufacturer Flaws were produced in different locations of the disc sample. This Inconel alloy disc sample is used in POD studies to determine the performance of NDE inspectors. The results of the flaw production are reported in the following. 5 Trueflaw’s capability to produce required number of flaws in to real specimen was shown in this work. An example of the produced flaws is shown in the following Figure 4. The produced POD sample is used for assessing the performance of the inspectors. Figure 4 An example picture of one of flaws in the sample A (fluorescent dye penetrant image). 3.2 Sample B - Rolls-Royce plc. UK Previously, Trueflaw manufactured flaws for Rolls-Royce plc., UK, for their development sample [5]. In that development work, it was shown that Trueflaw’s technology has the potential of producing flaws to actual components. Rolls-Royce plc. used this first development sample in series of trials while developing their new inspection technology. The current POD sample builds on the earlier work. The POD sample is used to assess the performance of the new inspection technology and inspectors. The following Figure 5 shows a typical example of produced flaws. Figure 5 An example picture of one of the flaws in the sample B (microscopic image on the left, fluorescent dye penetrant image on the right). 4. Conclusions Trueflaw placed the flaws in the planned locations with specified flaw characteristics. The specified flaw characteristics were flaw sizes (length, depth) and openings varying between different flaws (releasing detailed flaw characteristics is prohibited). Trueflaw placed the flaws in the actual components supplied by the clients. Flaw manufacturing for both of the cases described above were successful. 6 The results show that, the problem of not being able to produce real flaws to true geometries and size of inspection samples (as indicated, e.g., in ref. 4) is overcome. It was shown that flaws could be produced to wide range of geometries, in different flaw locations and crack sizes. Results show, that the Trueflaw method can be used to manufacture cracks to real components for use in POD studies. This allows more realistic POD curves to be determined that include the effect of sample geometry. These flawed samples are now used in assessment of probability of detection performance of different inspectors and inspection techniques. Acknowledgements The authors want to acknowledge the support from the industrial partners and their permission to publish results in this paper. References [1] MIL-HDBK-1823, 1999. Department of Defense handbook: Nondestructive Evaluation System Reliability Assessment. MIL-HDBK-1823, 30th April 1999. [2] Berens, A. P. NDE Reliability Data Analysis in: Metals Handbook, 9th ed. Vol 17. [3] Gandossi, L. & Annis, C. 2010. Probability of Detection Curves: Statistical Best-Practices, ENIQ report No 41. European Commission Joint Research Centre Institute for Energy, EUR 24429 EN, ISBN 978-92-79- 16105-6. [4] Weekes, B., Almond, D.A., Cawley, P. and Barden, T., 2012. Eddy-current induced thermography – probability of detection study of small fatigue cracks in steel, titanium and nickel-based superalloy. NDT&E International 49 (2012) pp. 47-56. [5] Virkkunen I., Kempainen M., Ostermeyer H., Paussu R. and Dunhill T., 2009. Grown cracks for NDT development and qualification. Insight Vol 51 No 5 May 2009, 5 p. [6] Oral communication with client representative, 2012-05-25. [7] Kemppainen, M., Virkkunen, I., Ostermeyer, H. and Gribi, M., 2009. Development and production of test specimens to evaluate an inspection issue. Materialprüfung mechanischer Komponenten in Kernkraftwerken, Nuklearforum Schweiz, November 2009, Olten. 2009-11-19. [8] Kemppainen M., 2006. Realistic artificial flaws for NDE qualification – novel manufacturing method based on thermal fatigue. Dissertation for the degree of Doctor of Science in Technology, Espoo, Finland, (Available online from: http://lib.tkk.fi/Diss/2006/isbn9512282631/) 7

Proceedings of the ASME 2012 Pressure Vessels & Piping Division Conference PVP2012 July 15-19, 2012, Toronto, Ontario, CANADA 1 Copyright © 2012 by ASME PVP2012-78477 NEW FLAWS FOR QUALIFICATION OF CAST STAINLESS STEEL INSPECTION Iikka Virkkunen Trueflaw Ltd. Tillinmäentie 3 A 113, 02330 Espoo Finland Tel: +358 45 6354415 Email: iikka.virkkunen@trueflaw.com Doug Kull EPRI NDE-Center 1300 West WT Harris Blvd. Charlotte, North Carolina 28262-7097 USA Email: dkull@epri.com Mika Kemppainen Trueflaw Ltd. Tillinmäentie 3 A 113, 02330 Espoo Finland Email: mika.kemppainen@trueflaw.com ABSTRACT For decades, cast austenitic stainless steels (CASS) have presented a challenge for inspection. However, recent advanced inspection technologies have shown promise in inspecting CASS materials with wall thicknesses that were once considered impossible. Before being applied on larger scale, these new inspection methods must be proven to be effective at identifying discontinuities in CASS material. This presents a problem of its own. Several traditional flaw manufacturing methods cannot be applied to CASS due to the disruption of the parent material. Excavation and welding changes the cast material microstructure and thus significantly affects the inspection results. At the same time, due to the significant wall thickness and inspection limitations, the required qualification flaws can be quite large. Until recently, modern flaw manufacturing techniques, that do not require welding, have not been applied to flaws of this size. In this paper, recent developments will be presented on the manufacturing of thermal fatigue cracks in centrifically CASS material. The presented developments make it possible to use real cracks for demonstrating the effectiveness of CASS inspection techniques. The results also contain first published UT data on this kind of thermal fatigue cracks in CASS and reveal new insight on the inspectability of this difficult material. INTRODUCTION Cast austenitic stainless steel (CASS)materials are used extensively in reactor coolant pressure boundary components including pipes, fittings, valve bodies and pump castings in pressurized-water reactors (PWRs) throughout the world [1, 2]. The use of CASS materials for these components is motivated by the favorable corrosion resistance properties as well as the relatively low cost. The service record of CASS components has also been good thus far. It has been shown that CASS material exposed to high temperatures for prolonged periods of time can be susceptible to thermal aging. The microstructure of cast stainless steels contains 5-25% of ferrite phase (with balance austenite phase). Over a period of time in normal reactor temperatures (290°C / 550°F), the ferrite phase goes through spinodal decomposition Copyright © 2012 by ASME 2 and a brittle α' -phase precipitates inside the former ferrite grains. The brittle precipitates decrease the fracture toughness of the material and the high toughness characteristic to austenitic stainless steels is eventually lost. The effect of aging depends, among other things, on the ferrite content of the material. Data suggests that increased ferrite content can be associated with more pronounced embrittlement. The initial Charpy impact energy for austenitic stainless steels is above 200 J/cm2. For fully aged samples, the lower bound estimates are 20, 25 and 30 J/cm2 (for CF-8M, CF-8 and CF-8A, CF3 and CF-3A, respectively). [2, 3] From inspection point of view, the most significant effect of thermal aging is that it decreases the size of the inspection target. The reduction in fracture toughness decreases the critical flaw size and thus the allowable flaw size for these components. E.g., Cicero et al. (2009) calculated critical flaw sizes in the range of 10 - 20 mm (0.4 - 0.8"; 40 - 80% through-wall ) for aged valve component with remaining Charpy impact energy of 65 J/cm2 [4]. While higher ferrite content has been found to increase the affects of thermal aging, it is also attributed to the beneficial effect of reducing sensitization. Sensitization is caused by precipitation of chromium-rich carbides and the resulting depletion of chromium at austenitic grain boundaries. In CASS the chromium carbides preferentially precipitate on the ferrite- austenite interfaces and not in the austenitic grain boundaries. The carbides form on the chromium-rich ferrite side of the boundary and do not cause the problematic chromium depleted zone [5]. Cast stainless steel components are used in safety-related systems and there is evidence that shows that CASS materials can be susceptible to thermal fatigue under certain conditions [6]. Consequently, there has been an increased interest in developing a reliable in-service inspection method for CASS material. The non-destructive evaluation of cast stainless steel components has proven quite challenging. The large and irregular grain size of some CASS materials greatly affects the propagation of ultrasonic waves causing attenuation, beam deflection and scattering. During the PISC-III study, the then-current state of the art inspection methods were evaluated against cast stainless steel samples. In the PISC-III round robin summary Lemaintre et al. (1996) [7] conclude that "...the detection performance, in general, was satisfactory. The length sizing performance in general was poor, whereas the performance for depth sizing was very poor". For two fatigue cracks in centrifugally cast steel weld, the probability of detection was < 0.5. The poor performance observed lead to significant development efforts to facilitate the inspection of these important CASS components. In particular, synthetic aperture focusing (SAFT) techniques were utilized to decrease the noise level in the material by several authors [1,8, 9]. In parallel, the use of eddy current (EC) techniques were studied [6], although the ID access required by the EC inspection would clearly impose problems for several locations. More recently, phased array ultrasonic techniques have been studied for inspection of cast stainless steel components [8, 10]. In summary, it's generally believed that current phased- array ultrasonic systems hold the best potential for the reliable inspection of cast stainless steel components. However, before the systems are applied on a larger scale, these new inspection methods must be proven to be effective at identifying and characterizing discontinuities in CASS material. This presents a problem of its own. To be able to demonstrate NDE reliability, the test blocks must be representative of what is currently installed in the nuclear fleet. Since the main feature making the inspection so challenging is the material microstructure, it's important that the samples contain representative microstructures. Also, the flaws to be used should be representative of postulated in-service induced flaws. The complex grain structure and postulated fatigue damage mechanism eliminates several traditional flaw manufacturing methods in cast stainless steels base material. Because fatigue flaw faces are rough and tortuous, the EDM notches used in the PISC studies are not considered ideal for evaluating NDE techniques in CASS material. Also, the weld implanted flaws traditionally applied are problematic in components without joints due to the excavation required to introduce the simulated flaw. It is typically possible to detect the cavity and weld metal in the middle of the base material more easily than the actual flaw. Cracks have also been manufactured by mechanical loads. However, this is limited to simple shapes and, due to the heavy wall thickness, requires rather heavy loading equipment. More recently, flaw manufacturing techniques based on controlled thermal fatigue have become available. These overcome many of the deficiencies in the more traditional flaw manufacturing techniques. The material microstructure is not disturbed, since there's no welding involved in the process. Also, the thermal loads can be applied to heavy sections or complex shapes like elbows or pump casings. However, due to the significant wall thickness and inspection limitations, the required flaws for CASS can be quite large. Until recently, modern flaw manufacturing techniques have not been applied to flaws of this size. In this paper, recent developments will be presented on the manufacturing of thermal fatigue cracks in CASS material. The work was performed by EPRI and Trueflaw and is still on-going. MATERIALS AND METHODS To develop test blocks for cast austenitic stainless steels, three samples were provided by EPRI. The samples were centrifically cast CF-8M grade stainless steel with a varied equiaxed grain structure. Figure 1 shows a section of the material, used for this study, which has been polished and etched to show the grain structure. Two of the samples were used for actual test blocks whereas the third was reserved for development. Figures 2 and 3 show the sample geometry of the components as well as the flaw locations. Copyright © 2012 by ASME 3 FIG. 1 EXAMPLE OF THE SAMPLES GRAIN STRUCTURE FIG. 2 SAMPLE GEOMETRY AND FLAW LOCATIONS FOR 15 mm TARGET DEPTH CRACKS FIG. 3 SAMPLE GEOMETRY AND FLAW LOCATIONS FOR 30 mm TARGET DEPTH CRACKS. In each test block, two flaws were produced by in-situ controlled thermal fatigue. Controlled thermal fatigue flaws have been available since early 2000's from Trueflaw. In recent years the technology has matured, tried and tested. Capability of the technique to produce realistic, representative flaws has been analyzed by comparing the crack characteristics to the characteristics measured from service-induced flaws. This comparison has been made against measured values from service-induced flaws reported by Wåle [11]. The comparison indicated the flaws produced by the new technique are representative of several types of service-induced flaws. To ensure reliable crack production and to know the depth of the produced cracks, each different crack produced is first validated destructively. That is, a crack is grown with a predetermined set of parameters and destructively examined to reveal the depth. Then, the same process can be repeated any number of times to produce number of similar cracks with known depth. This process is followed specifically for each material and flaw size. Thermal fatigue loading is characteristically greatest at the surface and decays with distance from the surface. Consequently, crack production is fastest near the surface and gets slower with increasing crack depth. Thus, the significant challenge for this project was that the cracks needed to be much deeper than previously manufactured. The first sample contained two cracks with target depth of 15 mm (0.59"). One of these cracks will remain intact while the other will be used as a validation crack, and will be destructively examined at later time. Sample 2 was designed to have two cracks, with one serving as validation crack for the other after destructive examination. The target depth for the second set cracks was 30 mm (1.18"). A significant amount of development was necessary, because flaws of this depth had previously never been produced with Trueflaws techniques. During the development, numerous production tests were done and destructively examined. This included a number of partial validations, where crack growth rates from several depths were tested in samples with EDM-notch starter of known depth. After the four cracks were grown in the two samples, the cracks were documented, photographed and surface features measured. Both samples were sent to EPRI for non-destructive evaluation. After thorough evaluation, the validation cracks will be destructively examined to confirm the true crack depths. RESULTS Figures 4 through 7 show penetrant images from the cracks produced for this study. It is suspected that the large grain structure caused significant tortuosity to all cracks. Additionally, some level of secondary cracking was present in all cases but was most significant in conjunction with the deeper cracks. Figures 8 through 9 show selected crack characteristics measured from the sample surface. Copyright © 2012 by ASME 4 FIG. 4 PENETRANT IMAGES OF PRODUCED CRACKS (15 mm TARGET DEPTH). FIG. 5 PENETRANT IMAGES OF PRODUCED CRACKS (15 mm TARGET DEPTH). FIG. 6 PENETRANT IMAGES OF PRODUCED CRACKS (30 mm TARGET DEPTH). FIG. 7 PENETRANT IMAGES OF PRODUCED CRACKS (30 mm TARGET DEPTH). FIG. 8 CRACK OPENING MEASURED FROM THE SURFACE (15 mm TARGET DEPTH). FIG. 9 CRACK OPENING MEASURED FROM THE SURFACE (30 mm TARGET DEPTH). Preliminary results from the NDE evaluation indicate that the 15 mm cracks in Sample 1 are readily detectable with techniques utilizing low frequency ultrasonic transducers. Currently, data has been collected using phased array search units ranging in frequencies between 500 kHz and 1.5 MHz. It has been noted that the data collected with lower refracted angles and lower frequencies experience less affects from the varying grain structure. Figures 10 through 13 show the signal characteristics of the flaws using search units with different frequencies. In some cases, the beam redirection has been noticeably different when scanning with the same search unit in opposite directions. The extraneous cracking around the intentional flaws has also been identifiable using the UT techniques and in some cases it is believed to partially mask the response from the primary target. Work is currently underway to subject Sample 2 to the same UT evaluation used on Sample 1. Copyright © 2012 by ASME 5 FIG. 10 SAMPLE 1 USING 1.50MHZ SEARCH UNIT (C-Scan) FIG. 11 SAMPLE 1 USING 1.00MHZ SEARCH UNIT (C-Scan) FIG. 12 SAMPLE 1 USING 0.75MHZ SEARCH UNIT (C-Scan) FIG. 13 SAMPLE 1 USING 0.50MHZ SEARCH UNIT (C-Scan) DISCUSSION In view of the production tests and partial validations, the development of crack growth into cast stainless steel was a success. The final depth will be revealed, after the destructive examination is complete. Further development is needed to eliminate the secondary cracking experienced with the deeper flaws. Secondary cracks could be avoided in the production tests and partial validations, and thus can be overcome with some additional development. The ultrasonic data images above show the C-scan images for a 30° angle with four different search unit frequencies. It can be seen that the noise level increases as the search unit frequency is increased up to 1.5MHz. The noise levels decrease because increased wavelengths of the pulsed energy from a low frequency search units interact less with the individual grain boundaries of course grained material. In this case the 30° 0.75MHz search unit images the two flaws in Sample 1. The higher frequency search units have too much noise and the lower frequency search unit does not seem to have enough resolution to image the first flaw. Additional NDE work is scheduled for this study throughout 2012. Once all of the data is collected on the two specimens, at full thickness, the samples OD will be machined to reduce the thickness by 25%. Ultrasonic data, matching that which has already been collected, will be acquired and compared before taking additional material off of the OD surface. By reducing the wall thickness of the mockup the crack height percentage will increase. After several iterations of taking data and removing additional material it is hoped that conclusions can be made on the effects of CASS material thickness and the inspectability. Once the sample thickness has been reduced enough that the cracks represent 80 to 90% through wall the validation cracks will be destructively analyzed to determine the true depth. CONCLUSIONS In view of the production tests and partial validations, the development of crack growth into cast stainless steel was a success. The final depth will be revealed, after the destructive examination is complete and final judgment will then be available. With the increased NDE technology now available it seems that meaningful inspections of CASS components has increased, but additional work is needed to quantify its true capability. It is important to have realistic mockups to evaluate new NDE techniques and controlled thermal fatigue flaws have shown promise for this purpose. Additional work is required to develop methods of creating deeper flaws and determine if the individual grain structure of the base material has any effect on the crack growth. REFERENCES 1. Diaz, A. A., Harris, R.V. & Doctor, S.R. 2008. "Field Evaluation of Low-Frequency SAFT-UT on Cast Stainless Steel and Dissimilar Metal Weld Components". United States Nuclear Regulatory Commission. NUREG/CR-6984. 104 p. Available online: http://pbadupws.nrc.gov/docs/ML0900/ML090020419.pdf (referred 2012-02-08). 2. Lee, S, Kuo, P. T., Wichman, K. & Chopra, O. 1997. "Flaw evaluation of thermally aged cast stainless steel in light- water reactor applications". International Journal of Pressure Vessels and Piping (72) pp. 37-44. 3. Chopra, O. K. 1994. "Estimation of Fracture Toughness of Cast Stainless Steels during Thermal Aging in LWR Systems". United States Nuclear Regulatory Commission. NUREG/CR–4513. 83 p. Available online: http://pbadupws.nrc.gov/docs/ML0523/ML052360554.pdf (referred 2012-02-08). 4. Cicero, S., Setién, J. & Gorrochategui, I. 2009. "Assessment Of Thermal Aging Embrittlement In A Cast Stainless Steel Copyright © 2012 by ASME 6 Valve And Its Effect On The Structural Integrity". Nuclear engineering and Design (239) pp. 16 - 22. 5. Jaske, C. E. & Shah, V. N. 1990. "Life assessment Procedures for Major LWR Components". United States Nuclear Regulatory Commission. NUREG/CR-5314, Vol 3. Available online: http://pbadupws.nrc.gov/docs/ML0403/ML040360127.pdf (referred 2012-02-08) 6. Diaz, A.A., Mathews, R. A., Hixon, J. & Doctor, S.R. 2007. "Assessment of Eddy Current Testing for the Detection of Cracks in Cast Stainless Steel Reactor Piping Components". United States Nuclear Regulatory Commission. NUREG/CR-6929. Available online: http://pbadupws.nrc.gov/docs/ML0706/ML070670477.pdf (referred 2012-02-08) 7. Lemaintre, P., Koblé, T. D. & Doctor, S. R. 1996. "Summary of the PISC round robin test results on wrought and cast austenitic steel weldments, part III: cast-to-cast capability study". International Journal of Pressure Vessels and Piping (69) pp. 33-44. 8. Anderson, M. T., Crawford, S. L., Cumblidge, S. E., Denslow, K. M., Diaz, A. A. & Doctor, S. R. 2007. "Heavy- Walled Cast Stainless Steel Piping Welds Using Advanced Low-Frequency Ultrasonic Methods". United States Nuclear Regulatory Commission. NUREG/CR-6933. Available online: http://www.nrc.gov/reading-rm/doc- collections/nuregs/contract/cr6933/ (referred 2012-02-08) 9. Spies, M. & Rieder, H. 2010. "Synthetic Aperture Focusing of Ultrasonic Inspection Data to Enhance the Probability of Detection of Defects in Strongly Attenuating Materials". NDT&E International (43) pp. 425-431. 10. Mahaut, S., Godefroit, J.-L., Roy, O. & Cattiaux, G. 2004. "Application of Phased Array Techniques to Coarse Grain Components Inspection". Ultrasonics (42) pp. 791-796. 11. Wåle, J. 2006. "Crack Characterisation for In-Service Inspection Planning - an Update". SKI reference 14.43- 200543105, ISRN SKI-R-06/24-SE, SKI, Stockholm, Sweden.

DETECTION AND SIZING SURFACE BREAKING DEFECTS IN NODULAR CAST IRON INSERT   Jorma Pitkänen *, Aarne Lipponen**, Matti Sarkimo**, Iikka Virkkunen***, Posiva Oy Olkiluoto Finland, VTT Espoo Finland, Trueflaw Ltd., Espoo, Finland ABSTRACT The planned disposal depth is about 420 m below ground surface in the bedrock. The nodular cast iron insert will be used as the inner component of the nuclear fuel disposal canister. The outer cover of the canister is a copper tube of 50 mm nominal thickness. The nodular cast iron insert is cast around several steel channels. The insert is the load carrying part of the canister structure. The basic dimensioning calculations are performed for normal operating conditions and in some upset conditions. The basic mechanical design load for the canister is 45MPa external pressure. The design load consists of the hydrostatic pressure of the groundwater, the swelling pressure of the bentonite buffer around the canister and of the pressure of glaciations of 2 to 3 km. The cast iron insert is checked in design pressure load cases to have a reasonable margin in general membrane stresses when comparing to the material design strength (yield strength) in the design temperature. Secondly, the structure is checked in postulated upset load conditions - 5 cm rock shear through the canister position - to have a reasonable margin against failure. This rock shear is setting requirement for the most critical defect size for the surface breaking cracks having dimensions of 4.5 mm (depth) x 27 mm (length) in circumference direction. For surface inspection are used 70° TRL (Transmission-receiving-longitudinal, 2MHz) probes. This measurement is carried out in direct contact using water as a couplant between the inspection specimen and probe. Probes are focused from 5 mm until 40 mm in depth. The focus point is in 20 mm depth. Four main directions for the inspection are used (2 in axial and 2 in circumference directions). This technique is a simple method for evaluating the surface volume until the nearest corner of the steel cassettes. The method is time consuming but can be accelerated by increasing the probes and its construction. The detectability of the 70° TRL probes has been tested against to thermally induced artificial cracks made by Trueflaw ltd. The depth of cracks varied between 1 to 6 mm which around the critical defect size. Also other ultrasonic techniques (TOFD, shear and longitudinal angle probes, creeping wave probes) have been tested to evaluate the size of the cracks. The results of the measurement have been reported. 1. INTRODUCTION Design loads for the canister structure are mechanical loads (pressure, local forces or forced displacements), thermal loads (varying temperature in time or position), chemical loads (chemical around the canister environment, including bacteria-induced chemical loads) and radiation load (radiation embrittlement). Loading phenomena are grouped into the following sets: • Handling loads • Incidents and accidents in the operation phase • Internal loads • External mechanical loads • External chemical loads. From the mechanical point of view, the most severe load cases are the isostatic pressure under the glacial period and the rock shear deformation. The strength values are mainly based on either tension or compression tests depending on the load case type. The only load case that may locally lead to significant yielding and plasticity of the insert is the rock shear case. Rock shear is, however, a “displacement-controlled load” that causes secondary stresses only, according to ASME nomenclature. If the load is secondary, the possible local yielding or cracking leads to decreasing stiffness and increasing deformation in the structure and, consequently, the load would decrease. That is why additional safety factors are not needed in displacement-controlled load cases. The maximum allowable surface defect size on the cylinder surface is a 4.5 mm deep and 27 mm long reference defect laying in a circumferential orientation. This damage tolerance analysis is the design basis load case for the canister insert for close-to-surface volumes, Raiko et al. 2010 /1/. The reference canister withstands the specified loads with an applicable safety margin even if the material has the allowable size defects mentioned above. The rock shear scenarios are shown in Figure 1 /2/. Figure 1. Rock shear is a “displacement-controlled load and the maximum allowable surface defect size on the cylinder surface is a 4.5 mm deep and 27 mm long reference defect laying in a circumferential orientation based on the fracture mechanics computations. Possible sensitive area on the surface is surrounded with the red line area. There are also in the insert one another near surface area, where can exist larger defect area - consisting of several pores forming a larger area. The surface and near surface area will be inspected with TRL70-2MHz probes using typical 4 directions (0°, 90°, 180°, 270°), Figure 2. The acceptance -rejection process of the NDT- inspections for the insert is shown in Figure 2 on the right. According to this process the defect will be detected and sized as shown in Figure 2 in three phases: raw evaluation, advanced evaluation and evaluation deviation. This is the base when the detectability was studied in a real situation. This time the defects were real cracks which were manufactured on the surface. The surface is in typical condition as it has been until now. This is the object of this study. Figure 2. Surface and near surface inspection of VVER type of insert using TRL70-2MHz probes (left) and the acceptance and rejection process based on the NDT measurements (right) 2 DEFECT MANUFACTURING BY THERMAL FATIGUE Controlled thermal fatigue cracks have been available since early 2000's from Trueflaw. In recent years the technology has matured, tried and tested. Capability of the technique to produce realistic, representative flaws has been analyzed by comparing the crack characteristics to the characteristics measured from service-induced cracks. This comparison has been made against measured values from service-induced flaws reported by Wåle /3/. The comparison indicated the flaws produced by the new technique are representative of several types of service-induced cracks. Current project represents the first instance of crack production to nodular cast iron. Due to it's two-phase microstructure, nodular cast iron presents new challenges for crack manufacturing. The nodules made crack initiation particularly easy and thus secondary cracking could not be avoided, especially for the larger cracks. As always, to ensure reliable crack production and to know the depth of the produced cracks, each different crack produced is first validated destructively. That is, a crack is grown with a predetermined set of parameters and destructively examined to reveal the depth. Then, the same process can be repeated any number of times to produce number of similar cracks with known depth. This process is followed specifically for each material and flaw size. In this case, several validation trials were required to produce the desired cracking. Figure 3 shows example fracture surface image from a validation crack. Figure 3. Example fracture surface image from a validation crack. After validation, the actual cracks to the component were manufactured. Figures 4 – 5 show example surface image and PT image from a produced crack. Figure 4. Example surface image from a produced crack. Graphite nodules are readily observable in the surface and seen to affect the crack growth. Figure 5. Example PT image from a produced crack. For all the cracks, the sufrace opening was measured. Figure 6 shows example measurement graph. The mean opening for the crack in question was measured to be 46.6 µm. Altogether five cracks were produced. The crack sizes are shown in Table N1. Figure 6. Example surface opening profile from a produced crack. Table1. Manufactured crack dimensions. Trueflaw flaw ID Size (l x a) 229BBB1385 32.7 x 5.3 142BBB1346 24 x 4.3 c089BBB1299 12.0 x 1.7 143BBB1374 5.2 x 0.9 213BBB1375 3.8 x 0.9 3 SURFACE DEFECT DETECTION BY ULTRASONIC TECHNIQUES Several methods were used to study surface defect detection and sizing. Some of the methods were applied manually and some using mechanized inspection. Surface wave technique (90°, two probes in similar configuration like in TOFD-method) was not used because the surface condition was not sufficient good and there were several secondary cracks near the actual crack as mentioned already in crack manufacturing part. The applied methods are named in Table 2. More advanced methods like SAFT /4/ or sampling phased array /5/ are also usable for surface breaking defect detection but those methods have not yet applied. Table 2. The applied ultrasonic inspection methods to detection of surface breaking defects Technique Detection Depth  Sizing Length  Sizing T55  (conposite) X X not  applied TRL-­‐70  -­‐  2  MHz X X not  applied WSY-­‐70  -­‐  2MHz X -­‐ not  applied PA  (angular  scanning X X not  applied Technique Detection Depth  Sizing Length  Sizing T55  (conposite) X X x TRL-­‐70  -­‐  2  MHz X X x WSY-­‐70-­‐2  MHz X -­‐ x PA  (angular  scanning X X x TOFD X X x Material  property  probe x -­‐ x Manual  detection  and  sizng Mechanized  detection  and  sizng 3.1 Angle probe detection Angle probe measurements can be done typically with single element or dual element probe. In this case the applied single element probe was shear wave probe having 55° angle of incidence. In order to evaluate the depth of a crack the tip diffraction can be used. In the crack three different areas can be distinguished /6/: • Corner, which corresponds to CMOD (crack mouth opening displacement) • Crack face, which corresponds to fracture surface of the crack • Crack tip, which corresponds to effective area of crack tip opening (plastic zone), CTOD (crack tip opening displacement) and variable loading conditions All these areas have own impact to the ultrasonic response. The size of the crack tip area is actually dependent on the load at the crack tip. Under tensile load the area can be estimated to be 3-4 x crack tip opening width (plastic zone) where crack tip affects. The possible echoes from crack face have to be taken in consideration also in evaluation of surface breaking defect, Figure 7. Crack tip diffraction techniques has been described in /7/ The depth sizing is very simple applying the formula 1 )cos(αtcd Δ= (1) where d is the depth of a crack, c is the sound velocity of the inspected material, Δt is time of flight to crack tip and α is the angle of incidence. The shear wave angle probe is usable there where grain size is small enough, but for instance in austenite materials shear wave probe can be applied successfully using short pulses either with high damped probes or as in our case composite probes. In case of cast iron there can be also echoes from voids and porosities originating from casting near crack tip which can cause errors in sizing. TRL-probe applies longitudinal wave which is less sensitive to larger grains in general, but in fact the wavelength vs. grain size is the actual parameter. In TRL probe near area of the signal has low noise level which is good for detection of surface defects. This is one of the main reasons for the use of TR-type of probes for surface defect detection. The TRL probe is especially good for detection of crack tip signals, which is well known in austenitic inspections. In these measurements the crack tip signals were applied for sizing. The calibration of TRL probe is more complicated compared to single element angle probe. Using TRL probe the angle of incidence varies a little in different depths. This can have an affect the sizing results. The angle variation can be gained in calibration using holes of varying depths. In this study the angle variation was not measured. 3.2 TOFD Measuring the amplitude of the reflected signal can be an unreliable method of sizing defects because the amplitude strongly depends on the orientation of the crack. Instead of amplitude, TOFD (Time Of Flight Diffraction) uses the time of flight of an ultrasonic pulse to determine the position of a reflector/8/. In a TOFD system, a pair of probes is turned against each other. One of the probes transmits an ultrasonic pulse that is received by the other probe. In undamaged material, the signals picked up by the receiver probe are from two waves: first one that travels along. TOFD technique is well explained in /9/. Figure 7. Sizing applying angle probe: Shear wave probe T55°- 4MHz (single element probe, left) and longitudinal wave probe TRL70-2MHz (dual element probe, right) the surface and the other one that reflects from the back wall. When a crack is present, there is a diffraction of the ultrasonic wave from the tip(s) of the crack. Using the measured time of flight of the pulse, the depth of a crack tip can be calculated by simple trigonometry. TOFD technique uses normally longitudinal waves for detection and sizing. The main principle is to use crack tip signals in order to receive the dimensions of the crack. The depth of surface breaking crack can be estimated according to following way /10/: tSctcd Δ+Δ= 4)(5.0 2 (2) Where d is the depth of a surface breaking crack, c is the sound velocity of the longitudinal wave in the inspected material, Δt is time difference between from lateral wave time of flight to crack tip, 2S is the separation between probes, which is in these measurements 16 mm. The calibration curve shows a good sizing capability for TOFD technique, as shown on the right in Figure 8. Figure 8. The measurement setup for surface breaking crack applying TOFD technique (left), calibration for measurement using notches having varying depths (1, 2, 4, 6, 10 mm) and corresponding calibration curve (right). 3.3 Phased array technique Phased array probes consist of an array of elements. Driving electrically these elements sound field is produced into the material. Some of phased array parameter measurements is described in /11/. The principle is well known but the industrialized phased array systems came into use at the beginning of 2000. These systems were able to do manual or mechanized inspection. The best performance of the phased array system can be achieved by the angular scanning often called also sector scanning for sizing. There are several ways to use sectorial scanning. In this we are concentrated to sectorial scanning and the defects are on the same side as the probe. This makes the detection of the small surface defects more difficult, because the angle of incidence must be large. And the elements have a certain aperture, which affects the angular sensitivity of the probe in the larger angle range. This is depending mainly on the size of the single element in the active part of the phased array. The crack tip detection is clearly enhanced by focusing the ultrasound near the tip /12/. Figure 9. Phased array probe measurement of the cracks (on the left) and calibration results from the Creeping wave probe reflections from cracks of varying sizes. 3.4 Creeping wave probe Creeping waves has been studied in many applications for crack detection successfully. The use is mostly applied to secondary creeping, which means crack detection in the opposite side as the probe is. In this studied case the cracks are in the same side as the probe and so called primary creeping are used for the detection of surface breaking cracks. The creeping wave probe is not applicable for sizing. The primary creeping can be calibrated using notches of varying depths. There are critical discussions of the existence of creeping waves (Blashan & Ginzel,2004). In spite of these wave type differences the method is usable for crack detection. In this study the term of creeping wave is used. The creeping wave probe can be calibrated either on shear wave or longitudinal wave velocity. In the measurement longitudinal wave calibration was applied. The primary creeping wave detectability of the manufactured 5 cracks was demonstrated manually as shown in Figure 9. All manufactured 5 cracks were clearly detectable as a minimum S/N 17 dB. The detection of creeping wave probe is not affected even if the crack is not oriented perpendicular to the sound beam of the probe. This is a good property as well as the high detectability of small cracks (depth of the crack is small). The characteristics of the creeping wave probe is discussed more detailed in Erhard study /13/ 3.5 Material property probe Surface (Rayleigh-waves) and leaky-Rayleigh waves are widely used in acoustic microscopy to characterize different materials and thin surface films and coatings. Much less is reported about the application of the same wave types in the low-frequency range. However, the same theoretical background and principles on which the high-frequency acoustic lenses are based can be utilized in low-frequency immersion and contact transducers. The low-frequency (2-15 MHz) transducers technique used for creating Rayleigh and leaky-Rayleigh-waves on the surface of a material can be applied for surface defect detection. The information received from material properties are measured with a special ultrasonic probe optimized for surface measurement and with a 0° longitudinal wave probe. The technique is based on combination of three factors: using back scattered ultrasonic signals and induced leaky Rayleigh wave information (1), and simple statistical data analysis (2) in combination with optimized ultrasonic transducer (3). The back scattered ultrasonic signal is a measure of the amount of geometrical reflectors such as micro-­‐pores, inclusions, precipitations, segregations, micro-­‐cracks and cracks as well as of back-­‐scattering from phase boundaries during fatigue damaging and increase of degradation inside the material /14/. The leaky Rayleigh wave component is sensitive to surface properties as known from normal Rayleigh wave probes. Especially cracks cause strong effect on the leaky Rayleigh wave. If the crack is deep it cancels the leaky Rayleigh wave signal totally. Figure 10. The principle of material property probe (left) and images (C-, B- and A-scans) from a crack measurement (right). 4 MEASUREMENTS Both mechanized and manual inspections were carried out applying real size mock up, which was cut from actual nodular cast iron insert component. To this mock up were manufactured artificial cracks as described earlier. All the inspection methods were applied to these cracks. The main emphasis was the detection and sizing of the cracks. The results will be discussed crack wise. The cracks were quite tight according to the crack manufacturer and this was also seen in the measurements. Other factor was the surface roughness which is clearly affecting the detection and sizing capability. The specified roughness Ra was 12.3 µm, which is clearly too rough for the ultrasonic measurements. Figure 11. The results from the ultrasonic sizing measurements applied to the crack 1 Crack 1 was according to manufacturing data (33 mm in length x 5.3 mm in depth) the deepest crack and similar results were estimated using applied ultrasonic methods. The crack size in depth and length is deeper than allowable defect size. The detectability was clear using all methods. The depth sizing gave a little deeper depth (5.4 - 6.1 mm) than the manufacturing data assumed, Figure 11. The difference was less than 1 mm. The length sizing gave 33 mm using TOFD and Material property probes, which was same as surface size estimated by the manufacturer. Creeping wave probe gave longer length (37 mm) and TRL probe gave clearly shorter length (10 mm) than the actual length of the crack. Crack 2 was according to manufacturing data (24 mm in length x 4.3 mm in depth) the second deepest crack and similar results were estimated using applied ultrasonic methods. The crack size in depth and length is similar to allowable defect size. The detectability was clear using all methods. The depth sizing gave a little deeper depth (4.5 - 5.1 mm) than the manufacturing data assumed, Figure 12. The difference was less than 1 mm. The length sizing gave 23-24 mm using TOFD and Material property probes, which was same as surface size estimated by the manufacturer. Creeping wave probe gave only little bit shorter length (19.9 mm) and TRL probe gave clearly shorter length (11.6 mm) than the actual length of the crack, Figure 12 (left). Crack 3 was according to manufacturing data (12 mm in length x 1.7 mm in depth) middle deep crack and similar results were estimated using applied ultrasonic methods. The crack size in depth and length is clearly smaller than allowable defect size. The detectability was clear using all methods. The depth sizing gave a little deeper depth (1.5 - 2.0 mm) than the manufacturing data assumed, Figure 12. The difference was less than 1 mm. The depth sizing was clearly more difficult compared to crack 1 and crack 2. The length sizing gave 11.0 mm using creeping wave probe and 10.4 mm material property probes, when the actual length was 12.0 mm estimated by the manufacturer. TOFD probe gave shorter length (6.0 mm) than the actual length of the crack and using TRL probe the length sizing was not successful, Figure 12 (right). Figure 12. The results from the ultrasonic sizing measurements applied to crack 2(left) and crack 3(right). Crack 4 was according to manufacturing data (5.2 mm in length x 0.9 mm in depth) small crack and detectability was clearly difficult. The crack size in depth was less than 1 mm, which is clearly less than allowable depth (4.5 mm) and in length in the range of sound field size (6 dB). The crack was not detectable in the data of mechanized measurements except in the data of the material property probe. Detection was not so clear also using the material property probe, but the length sizing could be done giving length 5.4 mm. The depth sizing was not successful, Figure 13 (right). Using manual inspection of creeping wave probe the crack could be detected clearly with S/N 17 dB. Crack 5 was according to manufacturing data (3.8 mm in length x 0.9 mm in depth) smallest crack and detectability was clearly difficult. The crack size in depth was less than 1 mm, which is clearly less than allowable depth (4.5 mm) and in length in the range of sound field size (6 dB). The crack was not detectable in the data of mechanized measurements except in the data of the material property probe. Detection was not so clear also using the material property probe, but the length sizing could be done giving length 3.5 mm. The depth sizing was not successful, Figure 13 (left). Using manual inspection of creeping wave probe the crack could be detected clearly with S/N 24 dB. Figure 13. The results from the ultrasonic sizing measurements applied to crack 4 (left) and crack 5 (right). 5 SUMMARY AND CONCLUSIONS The nuclear fuel disposal canister contains nodular cast iron insert. the insert is the load bearing component is the structure. One loading possibility is rock shear case, which according present fracture mechanics computations has given allowable defect size in circumference direction 27 mm in length x 4.5 mm in depth. In this study several ultrasonic measurement techniques has been applied to detect and size artificial cracks on the surface of real size mock up. These cracks are very tight and thus more difficult for ultrasonic testing. In the mock up the surface roughness according to manufacturing specification was 12.3 µm, which is clearly too rough for the ultrasonic measurements. This is already noticed and the manufacturing specification has been changed to have smaller surface roughness. The detectability was sufficient to detect not allowable and allowable surface breaking crack size. Even clearly smaller crack than allowable (12 mm in lenght and 1.7 mm in depth) was clearly detected. Only the smallest cracks could not be found clearly. The surface roughness and tightness had of course affect on the detectability of those small cracks. The sizing of the cracks could be applied using several methods. The simplest methods for depth sizing were clearly TOFD and PA applying angular scanning. The best method to size the crack lenght was the use of the material property probe. The TOFD technique was also applicable for lenght sizing even though the smallest defect were not successfully sized. REFERENCES 1 Raiko, H., Sandström, R., Rydén H., Johansson, M. 2010. Design analysis report for the canister. SKB report TR-10-28, Swedish Nuclear Fuel and Waste Management Co. ISSN 1404- 0344. 2 Raiko, H., 2012. Structural Design of Disposal Canister. Report POSIVA 2012-13, Posiva Oy. (To be published). 3 Wåle, J. "Crack Characterisation for In-Service Inspection Planning - an Update". SKI reference 14.43-200543105, ISRN SKI-R-06/24-SE, SKI, Stockholm, Sweden. 2006. 4 Bulaninov, A., 2005. Der getakte Gruppenstrahler (Sampling Phased Array). Dissertation des grades des doktors der Ingeniurwissenschaften der Naturwissenschaflich-Technischen Fakultät III Chemie, Pharmazie, Biound Werkstoffwissenschaften der Universität des Saarlandes, 11. August, 128 p. 5 Pitkänen 2006,SAFT – Is it a Tool for Improved Sizing in Ultrasonic Testing, ECNDT 2006 - Poster 211, 25th - 29th September 2006, 13p. 6 Pitkänen J, Laukkanen, A., Kemppainen, M. & Virkkunen, I, 2007, Einwirkung der Spannung auf Ultraschallanzeigen bei der Detektierung und Größenbestimmung von Rissen (Effect of Stress on Ultrasonic Response in Detection and Sizing of Cracks), MP Materials testing 4(2007)6 pp. 299-309. 7 Gruber, G., 1980. Defect Identification and Sizing by the Ultrasonic Satellite-Pulse Technique. Deutsche Gesellschaft für zerstörungsfreie Prüfung, Jahrestagung, Göttingen, 12th -14th May 8 p. 8 Silk M. G., 1987. Changes in Ultrasonic Defect Location and Sizing. NDT International, 20(1987)1 pp. 9-14. 9 Charlesworth J. P. & Temple J. A. G., 2001, Engineering application of ultrasonic Time-of Flight diffraction, Second Edition, Research Studies Press LTD, 254 p. 10 SFS-EN 583-6, 2008, Non-destructive testing. Ultrasonic examination. Time-of-flight diffraction technique as a method for detection and sizing of discontinuities 26p. 11 Davis, M & Moles M., 2006 Resolving capabilities of phased array sectorial scans (S-scans) on diffracted tip signals, Insight 48(2006)4 April, 7p. 12 Dupond, O., Bredif, P., Poidevin, C. & De Mathan, N, 2004, Advanced phased array transducer detection of closed crack tip diffraction, NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, 6.-8. December, London UK, pp. 724 – 733. 13 Erhard, A., 1983, Untersuchungen zur Ausbreitung von Longitudinalwellen an Oberflächen bei der Materialprüfung mit Ultraschall, Forschungsbericht 88, BAM Berlin, 32 p 14 Pitkänen, J; Kauppinen, P & Jeskanen, H , 2004, Materials degradation and altering influence on ultrasonic scattering by the light statistical approach in austenitic thermal loaded materials, AIP Conference Proceedings, Volume 700 Review of progress in Quantitative Nondestructive Evaluation. Vol. 23. D.O.Thompson and D.E.Chimenti (Eds.). American Institute of Physics. Melville, New York (2004), 1363 - 1367

NUMERICAL MODELING OF THERMAL FATIGUE CRACK FROM VIEW POINT OF EDDY CURRENT TESTING Jing WANG1 ∗, Noritaka YUSA1, Mika KEMPPAINEN2, Iikka VIRKKUNEN2 and Hidetoshi HASHIZUMET1 1 Department of Quantum Science and Energy Engineering, Graduate School of Engineering, Tohoku University, 6-6 Aramaki Aza Aoba, Aoba-Ku, Sendai, Miyagi 980-8579, 2 Trueflaw Ltd, Tillinmäentie 3, tila A113, 02330 Espoo, Finland Abstract Eddy current testing (ECT) is one of electromagnetic nondestructive method. It was initially regarded as one kind of qualitative measurement during the inspection of natural cracks, such as stress corrosion crack (SCC), thermal fatigue crack (TFC) et al. On the other hand, ultrasonic testing (UT) is widely used for quantitatively detecting cracks. However performance of UT is inevitably affected by complexity of natural cracks, e.g. fracture surface roughness and so on [1]. Therefore, enhancement of inspecting natural cracks by other non- destructive method is demanded. ECT has become a favorable candidate for the aim with the aid of great progress of electromagnetic computational technology [2]. In fact quite a few studies have been reported success in quantitative evaluation of SCC and also they pointed out that modeling of crack is the issue for the accuracy measurement [3,4,5]. Based on this, modeling of TFC should be emphasized for the accurate eddy current inspection. However, only few studies have been discussed on the matter [6,7]. Further researches by more specimens are needed. The present study discusses numerical modeling of thermal fatigue crack (TFC) from view point of eddy current testing. Four TFCs artificially introduced into type SUS316 stainless steel plates are prepared for the study. Eddy current signal are gathered by a differential type plus point probe with three frequencies of 100 kHz, 200 kHz and 400 kHz. Subsequent destructive test is carried out to reveal the profile of cracks. TFCs are modeled as a region with constant width, uniform conductivity and real profile in numerical simulation by software Comsol Multiphysics 4.4. The basic geometry of simulation is shown as figure.1. With consideration on previous studies [6,7], numerical modeling of TFC in type SUS304 stainless steel and Inconel 600 plates, results of the study demonstrate that TFC should be modeled as an almost nonconductive region in general no matter how the frequency is used. Figure 1. Basic geometry of simulation References [1] M. Kemppainen, I. Virkkunen, Crack characteristics and their importance to NDE, Journal of Nondestructive Evaluation 30(3)(2011), 143-157. ∗ Corresponding author. Phone:81-8096300309, Fax:81-227957906, E-mail address:jingwang8604@gmail.com [2] BA. Auld, JC. Moulder, Review of advances in quantitative eddy current nondestructive evaluation, Journal of Nondestructive Evaluation 18(1)(1999), 3-36. [3] N. Yusa, Z. Chen and K. Miya, Sizing of stress corrosion cracking on austenitic stainless piping in a nuclear power plant from eddy current NDT signals, Nondestructive Testing and Evaluation 20(2)(2005), 103-114. [4] N. Yusa, Z. Chen, K. Miya, T. Uchimoto and T. Takagi, Large-scale parallel computation for the reconstruction of natural stress corrosion cracks from eddy current testing signals. NDT&E International 36 (2003), 449-459. [5] Z. Badics, Y. Matsumoto, K. Aoki, F. Nakayasu and A. Kurokawa, Finite element models of stress corrosion cracks (SCC) in 3-D eddy current NDE problem. Nondestructive testing of materials. IOS Press (1995), 21-29. [6] J. Wang, N. Yusa, H. L. Pan, M. Kemppainen, I. Virkkunen and H. Hashizume, Discussion on modeling of thermal fatigue cracks in numerical simulation based on eddy current signals, NDT&E International 55 (2013), 96-101. [7] J. Wang, N. Yusa, H. Pan, M. Kemppainen, I. Virkkunen and H. Hashizume, Modeling of thermal fatigue crack for enhancement of electromagnetic nondestructive evaluation of nuclear power plant, The 21st International Conference on Nuclear Engineering. (2013). Acknowledgements This study was conducted as part of the Aging Management Project for System Safety of Nuclear Power Plants commissioned by Nuclear Regulation Authority of Japan.

NEW FLAWS FOR QUALIFICATION OF CAST STAINLESS STEEL INSPECTION Iikka Virkkunen, Trueflaw Ltd., Finland, Doug Kull, EPRI NDE-Center, USA, Mika Kemppainen, Trueflaw Ltd., Finland ABSTRACT For decades, cast austenitic stainless steels (CASS) have presented a challenge for inspection. However, recent advanced inspection technologies have shown promise in inspecting CASS materials with wall thicknesses that were once considered impossible. Before being applied on larger scale, these new inspection methods must be proven to be effective at identifying discontinuities in CASS material. This presents a problem of its own. Several traditional flaw manufacturing methods cannot be applied to CASS due to the disruption of the parent material. Excavation and welding changes the cast material microstructure and thus significantly affects the inspection results. At the same time, due to the significant wall thickness and inspection limitations, the required qualification flaws can be quite large. Until recently, modern flaw manufacturing techniques, that do not require welding, have not been applied to flaws of this size. In this paper, recent developments will be presented on the manufacturing of thermal fatigue cracks in centrifugally CASS material. The presented developments make it possible to use real cracks for demonstrating the effectiveness of CASS inspection techniques. INTRODUCTION Cast austenitic stainless steel (CASS) materials are used extensively in reactor coolant pressure boundary components including pipes, fittings, valve bodies and pump castings in pressurized-water reactors (PWRs) throughout the world. The main coolant line and branch lines of many plant types contain cast components. In Table 1. a summary of main coolant line and branch line materials is shown for those plant types that include cast stainless steels. For many manufacturers, both wrought austenitic stainless steel (WASS) and CASS has been used. For Framatome plants, the use of CASS is concentrated on the elbows, and there's currently a program to replace these cast elbows over time [1]. The surge lines are generally fabricated from WASS except in some Combustion Engineering plants where they are fabricated from CASS. Table 1. Main coolant line and branch line materials according to plant type [2] Plant type Piping Elbows Westinghouse WASS, CASS CASS, bent pipe Framatome WASS, CASS CASS, bent pipe B&W WASS CASS, bent pipe Combustion WASS, CASS (CF-8M) CASS (CF-8M) The use of CASS materials for these components is motivated by the favorable corrosion resistance properties as well as the relatively low cost. The service record of CASS components has also been good thus far. It has been shown that CASS material exposed to high temperatures for prolonged periods of time can be susceptible to thermal aging. Thermal aging causes the decrease in material ductility and critical flaw sizes. For inspection, this has the effect of decreasing the size of the inspection target. The reduction in fracture toughness decreases the critical flaw size and thus the allowable flaw size for these components. E.g., Cicero et al. (2009) calculated critical flaw sizes in the range of 10 - 20 mm (0.4 - 0.8"; 40 - 80% through-wall ) for aged valve component with remaining Charpy impact energy of 65 J/cm2. [3,4] Cast stainless steel components are used in safety-related systems and there is evidence that shows that CASS materials can be susceptible to thermal fatigue under certain conditions [5]. Consequently, there has been an increased interest in developing a reliable in-service inspection method for CASS material. The non-destructive evaluation of cast stainless steel components has proven quite challenging. The large and irregular grain size of some CASS materials greatly affects the propagation of ultrasonic waves causing attenuation, beam deflection and scattering. During the PISC-III study, the then-current state of the art inspection methods were evaluated against cast stainless steel samples and concluded that the detection performance, in general, was satisfactory, whereas the performance for depth sizing was very poor. For two fatigue cracks in centrifugally cast steel weld, the probability of detection was below 50%. [6] The poor performance observed lead to significant development efforts to facilitate the inspection of these important CASS components. In particular, synthetic aperture focusing (SAFT) techniques were utilized to decrease the noise level in the material by several authors [7,8,9]. In parallel, the use of eddy current (EC) techniques were studied [4], although the ID access required by the EC inspection would clearly impose problems for several locations. More recently, phased array ultrasonic techniques have been studied for inspection of cast stainless steel components [8, 10]. In summary, it's generally believed that current phased-array ultrasonic systems hold the best potential for the reliable inspection of cast stainless steel components. However, before the systems are applied on a larger scale, these new inspection methods must be proven to be effective at identifying and characterizing discontinuities in CASS material. To be able to demonstrate NDE reliability, test blocks representative of what is currently installed in the nuclear fleet are needed. Since the main feature making the inspection so challenging is the material microstructure, it's important that the samples contain representative microstructures. Also, the flaws to be used should be representative of postulated in-service induced flaws. The complex grain structure and postulated fatigue damage mechanism eliminates several traditional flaw manufacturing methods in cast stainless steels base material. Because fatigue flaw faces are rough and tortuous, the EDM notches used in the PISC studies are not considered ideal for evaluating NDE techniques in CASS material. Also, the weld implanted flaws traditionally applied are problematic in components without joints due to the excavation required to introduce the simulated flaw. It is typically possible to detect the cavity and weld metal in the middle of the base material more easily than the actual flaw. Cracks have also been manufactured by mechanical loads. However, this is limited to simple shapes and, due to the heavy wall thickness, requires rather heavy loading equipment. More recently, flaw manufacturing techniques based on controlled thermal fatigue have become available. These overcome many of the deficiencies in the more traditional flaw manufacturing techniques. The material microstructure is not disturbed, since there's no welding involved in the process. Also, the thermal loads can be applied to heavy sections or complex shapes like elbows or pump casings. However, due to the significant wall thickness and inspection limitations, the required flaws for CASS can be quite large. Until recently, Trueflaw flaw manufacturing techniques have not been applied to flaws of this size. In this paper, recent developments will be presented on the manufacturing of thermal fatigue cracks in CASS material. The work was performed by EPRI and Trueflaw and is still on-going. MATERIALS AND METHODS To develop test blocks for cast austenitic stainless steels, three samples were provided by EPRI. The samples were centrifugally cast CF-8M grade stainless steel with a varied equiaxed grain structure. Figure 1 shows a section of the material, used for this study, which has been polished and etched to show the grain structure. Two of the samples were used for actual test blocks whereas the third was reserved for development. Figures 2 and 3 show the sample geometry of the components as well as the flaw locations. Figure 1. Example of the samples grain structure Figure 2. Sample geometry and flaw locations for cracks with 15 mm target depth Figure 3. Sample geometry and flaw locations for cracks with 30 mm target depth In each test block, two flaws were produced by in-situ controlled thermal fatigue. Controlled thermal fatigue flaws have been available since early 2000's from Trueflaw. In recent years the technology has matured, tried and tested. Capability of the technique to produce realistic, representative flaws has been analyzed by comparing the crack characteristics to the characteristics measured from service-induced flaws. This comparison has been made against measured values from service-induced flaws reported by Wåle [11]. The comparison indicated the flaws produced by the new technique are representative of several types of service-induced flaws. To ensure reliable crack production and to know the depth of the produced cracks, each different crack produced is first validated destructively. That is, a crack is grown with a predetermined set of parameters and destructively examined to reveal the depth. Then, the same process can be repeated any number of times to produce number of similar cracks with known depth. This process is followed for each material and flaw size. Thermal fatigue loading is characteristically greatest at the surface and decays with distance from the surface. Consequently, crack production is fastest near the surface and gets slower with increasing crack depth. The significant challenge for this project was that the cracks needed to be much deeper than previously manufactured. The first sample contained two cracks with target depth of 15 mm (0.59"). One of these cracks will remain intact while the other will be used as a validation crack, and will be destructively examined at later time. Sample 2 was designed to have two cracks, with one serving as validation crack for the other after destructive examination. The target depth for the second set cracks was 30 mm (1.18"). A significant amount of development was necessary, because flaws of this depth had previously never been produced with Trueflaws techniques. During the development, numerous production tests were done and destructively examined. This included a number of partial validations, where crack growth rates from several depths were tested in samples with EDM-notch starter of known depth. After the four cracks were grown in the two samples, the cracks were documented, photographed and surface features measured. Both samples were sent to EPRI for non- destructive evaluation. After thorough evaluation, the validation cracks will be destructively examined to confirm the true crack depths. RESULTS Figure 4 shows fracture surface image from a preliminary thermal fatigue flaw production test completed. The fracture surfaces show beach marks and other features typical for fatigue cracking. Also the effect of the coarse microstructure is evident. The fracture surface is markedly uneven. Although the overall shape of the crack is semi-elliptic, the crack tip also shows waviness, which indicates that the local microstructural features cause crack acceleration and retardation. Figures 5 through 8 show penetrant images from the cracks produced for this study. It is suspected that the large grain structure caused significant tortuosity to all cracks. Figures 9 and 10 show the crack surface roughness measured from the surface. Typical Ra values for similar cracks in wrought stainless steels range from 15 to 65 µm. Some level of secondary cracking was present in all cases but was most significant in conjunction with the deeper cracks. Figure 4. Fracture surface image from a production trial. The crack fracture surface is seen as darker area at the lower part of the image. The strong influence of coarse microstructure is evident from the fracture surface. Figure 5. Red dye penetrant image of a produced crack (15 mm target depth). Figure 6. Penetrant image of produced crack (15 mm target depth). Figure 7. Penetrant images of produced crack (30 mm target depth). Figure 8. Penetrant images of produced crack (30 mm target depth). Figure 9. Crack roughness (Ra) measured from surface (15 mm target depth). Figure 10. Crack roughness (Ra) measured from surface (30 mm target depth). Preliminary results from the NDE evaluation indicate that the 15 mm cracks in Sample 1 are readily detectable with techniques utilizing low frequency ultrasonic transducers. Currently, data has been collected using phased array search units ranging in frequencies between 500 kHz and 1.5 MHz. It has been noted that the data collected with lower refracted angles and lower frequencies experience less affects from the varying grain structure. Figures 11 through 14 show the signal characteristics of the flaws using search units with different frequencies. In some cases, the beam redirection has been noticeably different when scanning with the same search unit in opposite directions. The extraneous cracking around the intentional flaws has also been identifiable using the UT techniques and in some cases it is believed to partially mask the response from the primary target. Work is currently underway to subject Sample 2 to the same UT evaluation used on Sample 1. Figure 11. Sample 1. using 1.5 MHz search unit (C-scan) Figure 12. Sample 1 useing 1.00 MHz search unit (C-scan) Figure 13. Sample 1 using 0.75 MHz search unit (C-scan) Figure 14. Sample 1 using 0.50 MHz search unit (C-scan) DISCUSSION In view of the production tests and partial validations, the development of crack growth into cast stainless steel was a success. The final depth will be revealed, after the destructive examination is complete. Further development is needed to eliminate the secondary cracking experienced with the deeper flaws. Secondary cracks could be avoided in the production tests and partial validations, and thus can be overcome with some additional development. The crack characteristics measured from the sample surface indicate that cracks produced in the cast material exhibit greater crack path tortuosity, which is reflected in the greater Ra values measured from the surface. The measured Ra values range from 50 to 300 µm and more whereas the typical Ra values for similar cracks in WASS are in the range of 15 ... 65 µm. This increase in Ra is readily explained by the large grain structure of the material. The growing cracks find the weakest path through the microstructure and thus the grain size markedly affects the crack path tortuosity. The ultrasonic data images above show the C-scan images for a 30° angle with four different search unit frequencies. It can be seen that the noise level increases as the search unit frequency is increased up to 1.5MHz. The noise levels decrease because increased wavelengths of the pulsed energy from a low frequency search units interact less with the individual grain boundaries of course grained material. In this case the 30° 0.75MHz search unit images the two flaws in Sample 1. The higher frequency search units have too much noise and the lower frequency search unit does not seem to have enough resolution to image the first flaw. Additional NDE work is scheduled for this study throughout 2012. Once all of the data is collected on the two specimens, at full thickness, the samples OD will be machined to reduce the thickness by 25%. Ultrasonic data, matching that which has already been collected, will be acquired and compared before taking additional material off of the OD surface. By reducing the wall thickness of the mockup the crack height percentage will increase. After several iterations of taking data and removing additional material it is hoped that conclusions can be made on the effects of CASS material thickness and the inspectability. Once the sample thickness has been reduced enough that the cracks represent 80 to 90% through wall the validation cracks will be destructively analyzed to determine the true depth. CONCLUSIONS In view of the production tests and partial validations, the development of crack growth into cast stainless steel was a success. The final depth will be revealed, after the destructive examination is complete and final judgment will then be available. Also, further analysis of the crack characteristics is made possible by the destructive examination. With the increased NDE technology now available it seems that meaningful inspections of CASS components has increased, but additional work is needed to quantify its true capability. It is important to have realistic mockups to evaluate new NDE techniques and controlled thermal fatigue flaws have shown promise for this purpose. Additional work is required to develop methods of creating deeper flaws and determine if the individual grain structure of the base material has any effect on the crack growth. REFERENCES 1) Chockie, A. "Summary Report. 3rd International Workshop on the Future Directions for the Inspection of Cast Austenitic Stainless Steel Piping". January 28 - 29, 2011 Seattle, WA. 2011. 2) IAEA. "Assessment and management of ageing of major nuclear power plant components important to safety - Primary piping in PWRs". IAEA-TECDOC-1361, International atomic energy agency, Vienna. ISBN 92–0–108003–4. 2003. 3) Lee, S, Kuo, P. T., Wichman, K. & Chopra, O. "Flaw evaluation of thermally aged cast stainless steel in light-water reactor applications". International Journal of Pressure Vessels and Piping (72) pp. 37-44. 1997. 4) Cicero, S., Setién, J. & Gorrochategui, I. "Assessment Of Thermal Aging Embrittlement In A Cast Stainless Steel Valve And Its Effect On The Structural Integrity". Nuclear engineering and Design (239) pp. 16 - 22. 2009. 5) Diaz, A.A., Mathews, R. A., Hixon, J. & Doctor, S.R. "Assessment of Eddy Current Testing for the Detection of Cracks in Cast Stainless Steel Reactor Piping Components". United States Nuclear Regulatory Commission. NUREG/CR-6929. Available online: http://pbadupws.nrc.gov/docs/ML0706/ML070670477.pdf (referred 2012-02-08). 2007. 6) Lemaintre, P., Koblé, T. D. & Doctor, S. R. "Summary of the PISC round robin test results on wrought and cast austenitic steel weldments, part III: cast-to-cast capability study". International Journal of Pressure Vessels and Piping (69) pp. 33-44. 1996. 7) Diaz, A. A., Harris, R.V. & Doctor, S.R. "Field Evaluation of Low-Frequency SAFT- UT on Cast Stainless Steel and Dissimilar Metal Weld Components". United States Nuclear Regulatory Commission. NUREG/CR-6984. 104 p. Available online: http://pbadupws.nrc.gov/docs/ML0900/ML090020419.pdf (referred 2012-02-08). 2008. 8) Anderson, M. T., Crawford, S. L., Cumblidge, S. E., Denslow, K. M., Diaz, A. A. & Doctor, S. R. "Heavy-Walled Cast Stainless Steel Piping Welds Using Advanced Low-Frequency Ultrasonic Methods". United States Nuclear Regulatory Commission. NUREG/CR-6933. Available online: http://www.nrc.gov/reading- rm/doc-collections/nuregs/contract/cr6933/ (referred 2012-02-08). 2007. 9) Spies, M. & Rieder, H. "Synthetic Aperture Focusing of Ultrasonic Inspection Data to Enhance the Probability of Detection of Defects in Strongly Attenuating Materials". NDT&E International (43) pp. 425-431. 2010. 10) Mahaut, S., Godefroit, J.-L., Roy, O. & Cattiaux, G. "Application of Phased Array Techniques to Coarse Grain Components Inspection". Ultrasonics (42) pp. 791-796. 2004. 11) Wåle, J. "Crack Characterisation for In-Service Inspection Planning - an Update". SKI reference 14.43-200543105, ISRN SKI-R-06/24-SE, SKI, Stockholm, Sweden. 2006.

access key: U1M3IYEAT Differences in Ultrasonic Indications – Thermal Fatigue Cracks and EDM Notches Ari KOSKINEN 1, Jonne HAAPALAINEN 1, Iikka VIRKKUNEN 2, Mika KEMPPAINEN2 1 VTT Technical Research Centre of Finland; Espoo, Finland Phone: +358 20 722 4027, Fax: +358 20 722 7002; ari.koskinen@vtt.fi, jonne.haapalainen@vtt.fi 2 Trueflaw Ltd.; Espoo, Finland; iikka.virkkunen@trueflaw.com, mika.kemppainen@trueflaw.com Abstract Different types of artificial defects are used for a qualification of NDT methods. The representativeness of these defects, as compared to service induce flaws, is crucial when the performance of NDT method or inspector is evaluated. Ultrasonic indications are highly dependent on defect characteristics like roughness, crack opening, tilt and branching. These characteristics are even more significant than the defect size. To develop performance and reliability of ultrasonic inspection methods and procedures more research on different type of discontinuity indications is needed. In this study, indications from different types of discontinuities are measured using scanning acoustic microscope (SAM). The inspection item is austenitic stainless steel pipe with thermal fatigue cracks and similar size of EDM notches. Clear difference can be seen between indications from the EDM notches and from the thermal fatigue cracks. Keywords: Ultrasound, ultrasonic inspection, scanning acoustic microscope (SAM), thermal fatigue, indication, discontinuity, defect, qualification 1 Introduction Ultrasonic inspection is widely used NDT method for structural integrity inspections in power plants as well as in other industrial areas. Real components with real flaws are usually not available to be used in the practical tests and therefore during the qualification process of any non-destructive method as well as ultrasonic inspection method different test samples are needed to prove in practice the effectiveness of the testing system. The representativeness of used test sample is often crucial and therefore used test samples should have similar material, dimensions, geometry etc. as the real component. Also the defects that are expected to exist in real life should be represented in the test sample (figure 1.). Usually it is very difficult and expensive to produce real defects in the test structures and therefore artificial defects are often used as substitutes. Figure 1. Indications from similar size of discontinuities produced by electric discharge machining (on the left) and thermal fatigue (on the right). Same equipment, probe, probe angle and gain used with both measurements. Typical types of artificial defects are flat bottom holes, electric discharge machining (EDM) notches, welded defects and different types of fatigued defects. 2 Scanning Acoustic Microscope (SAM) Scanning Acoustic Microscopy works by directing focused sound from a transducer at a small point on a target object. Sound hitting the object is either scattered, absorbed, reflected (scattered at 180°) or transmitted (scattered at 0°). It is possible to detect the scattered pulses travelling in a particular direction. A detected pulse informs of the presence of a boundary or object. The `time of flight' of the pulse is defined as the time taken for it to be emitted by an acoustic source, scattered by an object and received by the detector, which is usually coincident with the source. The time of flight can be used to determine the distance of the inhomogeneity from the source given knowledge of the speed through the medium. Based on the measurement, a value is assigned to the location investigated. The transducer (or object) is moved slightly and then insonified again. This process is repeated in a systematic pattern until the entire region of interest has been investigated. Often the values for each point are assembled into an image of the object. The contrast seen in the image is based either on the object's geometry or material composition. The resolution of the image is limited either by the physical scanning resolution or the width of the sound beam (which in turn is determined by the frequency of the sound). 3 Stainless steel pipe sample 3.1 Thermal Fatigue Crack Representative controlled artificial cracks can be produced in-situ with thermal fatigue. The use of grown cracks based on the thermal fatigue production process has increased markedly during the last few years. Furthermore, the amount of different applications has become larger thus covering today most of the NDE inspection techniques and targets in the nuclear field. Capability of this technique to produce realistic, representative flaws has been analysed by comparing the crack characteristics to the characteristics measured from service-induced flaws. This comparison has been made against measured values from service-induced flaws reported by Wåle. In general, thermal fatigue flaws are representative for most of the service- induced flaws, when used as a reflector for different NDE development, training and qualification purposes. In present study, a tube sample with three thermal fatigue flaws were used. The produced flaws were characterized by surface microscopy. Figure 1. shows example surface image from one of the cracks. The measured crack surface opening profile and roughness are shown in Figure 2. Figure 1. Example surface microscopy image. a) b) Figure 2. Measured crack surface opening (a) and Rz roughness (b). The depth of these cracks is known through destructive examination of a validation cracks produced with the same local sequence. 3.2 Electric Discharge Machining (EDM) 4 Ultrasonic testing and results Test setup Conventional Scanning acoustic microscope (SAM) Phased Array? 5 Conclusions

J Nondestruct Eval (2011) 30:143–157 DOI 10.1007/s10921-011-0102-z Crack Characteristics and Their Importance to NDE Mika Kemppainen · Iikka Virkkunen Received: 23 September 2010 / Accepted: 13 May 2011 / Published online: 2 June 2011 © The Author(s) 2011. This article is published with open access at Springerlink.com Abstract The reliability of non-destructive evaluation de- pends on multitude of different factors. Consequently, it is difficult to assess the performance of the system. The only practical way to overcome this complexity and asses inspec- tion reliability is using practical trials. In practical trials the inspection is performed on a known, flawed sample and reli- ability is judged by comparing the acquired inspection result with known state of the sample. However, confirming (and demonstrating), that the practical trials and, in particular, the used flaws are representative to postulated inspection case is, at present, challenging. In this paper, the open literature is reviewed and compiled to the extent necessary for providing a starting point for confirming and demonstrating representativeness of flaws used in the practical trials. The available information on es- sential flaw parameters for various NDE techniques is re- viewed. The used measurement methods for each parameter is discussed and the available in-service data summarized. Finally, a simple procedure is proposed for confirming and demonstrating the representativeness of used flaws. Keywords Nondestructive evaluation · NDE · Qualification · NDE reliability · Nondestructive testing · NDT · Crack characterization · Crack characteristics 1 Introduction The reliability of non-destructive evaluation depends on multitude of different factors. These range from physical as- pects of the used technology (e.g, wavelength of ultrasound) M. Kemppainen · I. Virkkunen () Trueflaw Ltd., Tillinmäentie 3, tila A113, 02330 Espoo, Finland e-mail: iikka.virkkunen@trueflaw.com to application issues (e.g. probe coupling or scanning cov- erage) and human factors (e.g. inspector training and stress or time pressure during inspection). Due to this complexity, the only practical way to asses inspection reliability and to confirm that the inspection procedure functions as intended is by using practical trials. In practical trials the inspection is performed on a known, flawed sample and reliability is judged by comparing the acquired inspection result with known state of the sample. In order to get reliable perfor- mance data from practical trials, it is paramount, that the inspection arrangement and flawed sample used are repre- sentative to real inspection situation. However, confirming (and demonstrating), that the used artificial flaws are representative to postulated inspection case is, at present, challenging. The methods used to define and measure various crack characteristics vary. Thus, com- piling cohesive justification for used flaws is laborious and difficult. Furthermore, there is no standard or best practice to follow on how to compile and present such information. In this paper, the open literature is reviewed and compiled to the extent necessary for providing a starting point for con- firming and demonstrating representativeness of used artifi- cial flaws. The available information on essential parameters for various NDE techniques is reviewed. The used measure- ment methods for each parameter is discussed and compared and the available in-service data summarized. Finally, a sim- ple procedure is proposed for confirming and demonstrating the representativeness of used flaws. 2 Representativeness of Flaws Representativeness, in this context, means, that the used ar- tificial cracks give similar response in the used NDE than actual, service-induced cracks would. This, of course, is not trivial to confirm or to demonstrate. 144 J Nondestruct Eval (2011) 30:143–157 One approach, that has been used, is to directly compare NDE response acquired from artificial flaws and service- induced cracks. This approach, however, has several signif- icant drawbacks and thus cannot be recommended. Firstly, the result becomes overly linked to the used NDE method. If the method is developed or changed even slightly, the re- sponse may change (this is generally the aim and reason for the development) and thus the comparison must be re-done every time. Secondly, representative service-induced cracks are not generally available; this is the motivation for the use of artificial flaws in the first place. If representative service- induced cracks of known size were available, artificial flaws would not be needed as these service-induced cracks could be used in stead. Thirdly, service-induced cracks, as any natural phenomena, exhibit wide range of variance in their characteristics. Thus, confirming similar signal-response to limited set of service-induced cracks (when available) does not confirm expected behavior if the actual service-induced cracks differ from the used sample due to natural variance. The second approach that can be used, is to show repre- sentativeness implicitly by used production method. That is, the artificial flaw is produced by the same damage mech- anism that is postulated to be active in service and thus, it is argued, the response must be similar. This is far bet- ter than comparing signal response since it does not require availability of representative service-induced cracks (only the postulated cracking mechanism must be known). How- ever, it is not always possible to produce artificial cracks to relevant components with the postulated damage mecha- nism. Furthermore, it is a bit unclear how broadly “damage mechanism” may be defined while still retaining sufficient representativeness. Still, where available, this approach pro- vides a simple and solid justification for representativeness of used cracks. The third and the recommended approach is to use met- allographic crack characteristics as the basis for representa- tiveness. The wide variety of natural cracks is broken down to determine discrete set of crack characteristics. The mag- nitudes of such essential characteristics are measured from service-induced cracks and from artificial cracks. Then, the used NDE methods are studied to define the list of character- istics that are essential for this particular method. Finally, the values of essential crack characteristics (for particular NDE- method) are compared between literature data from service induced cracks (with postulated defect mechanism(s)) and artificial cracks. Representativeness is confirmed by demon- strating that the used artificial cracks are similar to postu- lated service-induced cracks in terms of the defined essential crack characteristics. This approach allows solid justifica- tion for representativeness of used cracks while retaining as much flexibility in flaw production as possible. By narrow- ing or widening the selection of the essential parameters, used cracks can be chosen to serve wider range of NDE- methods and/or crack types. Also, if certain characteristic can’t be reproduced by the used flaws, this approach brings this into light so that it can be addressed by, e.g., technical justification or additional set of test samples. 3 Essential Characteristics for NDE Wide variety of different NDE methods is available and new ones are constantly developed. Thus a comprehensive list of essential crack characteristics for every method cannot be compiled by one party but the final definition of essen- tial characteristics to be considered must be done in collab- oration with the inspection vendor when all the details of the used inspection technique are know. However, to pro- vide general background work for such analysis, informa- tion available on the open literature is summarized below. Also, if there is uncertainty about the significance of cer- tain crack characteristic, it may always be added to the list of characteristics to be reproduced by the used artificial flaws. Thus, representativeness of used cracks can always be shown, even in the absence of detailed information from used NDE-method (with the possible risk of using “overly representative” cracks). In the current paper, the focus is on ultrasonic techniques. The reason is that of all the critical inspection techniques used to detect and size real cracks, ultrasonic inspections are most widely used. Furthermore, sensitivity of the ultrasonic techniques to different crack characteristics is known and recognized by the international NDE community. 3.1 Flaw Properties Affecting Ultrasonic Detection and Sizing There is a lot of experience confirming the difficulty to reli- ably detect and size service-induced thermal fatigue cracks as reported, e.g., by Edwards et al. [6, 7], Pirson et al. [15], and Gauthier [8]. The difficulty of the inspection is caused by typical characteristics of cracks, which affect, e.g., prop- agation, reflection, diffraction, transmission, attenuation and diffusion of ultrasonic energy [2, 10]. Such flaw character- istics have been stated to be, amongst others, location, ori- entation and size of a crack (e.g., [19]), the opening of a crack and crack tip (e.g., [1, 20, 24]), the remaining residual stresses in the material (e.g., [8, 11]), fracture surface rough- ness (e.g., [13, 20]), plastic zone (e.g., [16]), and filling of the crack with some substance (e.g., [2]). 3.1.1 Effect of Fracture Surface Roughness on Detection and Sizing The fracture surface of a realistic flaw is not ideally planar, but it has natural irregularities. The general effects of the fracture surface roughness of a reflector on the spatial dis- tribution of scattered waves are well known and have been J Nondestruct Eval (2011) 30:143–157 145 Fig. 1 Polar plots of scattered amplitude distributions from surfaces of different roughness values (σ ), when a 2 MHz monochromatic wave is incident at 30 ◦ [13] widely studied both practically and theoretically. The prin- ciple is that by increasing the reflection surface roughness, the forward scattered high amplitude is decreased and en- ergy is redistributed into a more widely spread diffuse field [13, 20]. Figure 1 shows an example of the effect of surface roughness (from smooth to very rough) to the distribution of the scattered field. From a smooth surface a strong coherent field will arise because of the interference between all the scattered wavelets from all parts of the surface. Enhanced surface roughness destroys the summation as the phase of the wave varies with position along the flaw surface. Thus, the strength of the coherent field is reduced and a diffuse, widely scattered field of varying phase will be generated. Increased surface roughness reduces the detection sensi- tivity of the methods relying on specular signal. Off-specular signals may be increased because of the diffuse field. This effect was seen for certain inspection geometries in the PISC-II exercise, where small rough flaws were found to have higher detectability than the smooth flaws of the same size. Increased frequency of the used probe increases the scattering effect due to fracture surface roughness [2]. Fur- thermore, if the surface has a regularly shaped fracture sur- face roughness profile, reflection of the incident sound wave may favor certain directions thereby decreasing the testing repeatability and reliability [9]. Effect of surface roughness on the scattered energy can be used qualitatively in flaw characterization to separate smooth planar flaws, rough pla- nar flaws and volumetric flaws [3, 13]. Theoretical models and experimental work on the effect of surface roughness have shown quite good agreement with smaller surface roughness, but with increased values this is not the situation. For example, Ogilvy [13] found an agree- ment between model predictions and experimental results to be typically within 3 dB, except for very rough surface where the theoretical prediction results did not agree with the experimental results. The effect of fracture surface roughness on the detectabil- ity becomes advantageous, when the flaw is tilted. When the incident wave is normal to the reflection surface, the specular signal is detected and increased surface roughness decreases the amplitude hence decreasing the detectability. When the flaw is tilted, the increased surface roughness al- lows reflection of ultrasonic energy in different directions, as theoretically modeled by Ogilvy [13]. This means that by increased tilt the detection changes from specular to off- specular field. For example, Toft [18] showed experimen- tally that increase of the fracture surface roughness of the flaw decreases the signal amplitude of well-oriented flaws (Fig. 2). Furthermore, increase of the tilting angle decreases the signal amplitude, but with increasing surface roughness the rate of amplitude decrease is diminished. Through tilt- ing the flaw, the signal amplitudes from rough flaws exceed those from smooth flaws as detection moves out from the main lobe of specular energy to the diffuse field. The level of misorientation, after which the detection is enhanced, depends on the flaw size and shape. The larger is the flaw, the smaller is the degree of mis-orientation beyond which the surface roughness will enhance detection. It must be noted that, since the diffuse field amplitude will never exceed the amplitude of the coherent field, the detectability can be enhanced only with sufficiently large misorientation. The model predictions showed this to be more than 20◦ from normal angle (0◦) [13]. However, there are studies that do not clearly indicate the effect. For example, Yoneyama et al. [24] studied the effect with conventional 45◦ and focusing 45◦ and conventional 60◦ and 70◦ probes. In the experimental work they used mechanical fatigue cracks with maximum surface rough- ness values (Ry) varying from 40 µm to 70 µm and av- erage values (Rz) from 34.6 µm to 55.9 µm (larger with deeper cracks). The difference in crack corner echo height was compared to the echo from a smooth corner. Authors 146 J Nondestruct Eval (2011) 30:143–157 Fig. 2 Signal levels of rough and smooth flaws as a function of flaw misorientation [18] found some differences between the echo heights from the smooth corner and corners of different fatigue cracks, but they considered them negligible and drew a conclusion that the difference of surface roughness between an EDM notch and a mechanical fatigue crack may be neglected when con- ducting flaw detection. Amplitude differences at different locations of flaws are used to size flaws with techniques relying on amplitude changes at flaw extremities [13]. With smooth flaws, the signal amplitude drops clearly near the edges of the flaw. However, for rougher flaws there is not a clear plateau re- gion from which the decibel drop could be measured, but the signal amplitude fluctuates across the flaw. The principle and difference between smooth and rough flaws is shown in Fig. 3. Surface roughness of the flaw affects also sizing tech- niques relying on the time-differences from flaw extremities [13]. Two clearly separate pulses are detected from a smooth flaw, from which the flaw size can be calculated. These dis- tinct pulses arise from edge diffraction. From a rough flaw, a continuous pulse may be obtained which is a superposition of the edge-diffracted pulses and diffuse scattering from all parts of the surface. Roughness may cause a loss of distinct diffracted pulses and hinder the timing measurements. The principle of the effect is shown in Fig. 3. Fig. 3 Principle of the effect of flaw surface roughness on the ampli- tude- and time-based sizing techniques 3.1.2 Effect of Opening and Residual Stresses on Detection and Sizing Basically, with increasing crack opening the obtained ultra- sonic echo amplitude increases and when the crack is closed, echo amplitude decreases. In the following the effect of flaw opening to the obtained ultrasonic response is introduced. Furthermore, the condition of residual stress is treated sepa- rately as an individual factor affecting the flaw opening and obtained ultrasonic response. The recorded amplitude changes are related to the re- flection surface movements. If the crack is open through its full length, the whole surface from the crack opening corner to the crack tip can vibrate freely as excited by the incident ultrasonic energy. If the crack is closed partly or through its whole length, the fracture surfaces pressed to- gether hinder the free oscillation of the reflection surface. In this case, fracture surface heights touching each other al- low transmission of the ultrasonic energy. Hence, the ultra- sonic wave does not meet any boundary of different acoustic impedances. Vice versa, if the crack is open, there is a clear boundary of different acoustic impedances where reflection and scatter of the ultrasonic energy occur. By a continued increase of the flaw opening, a saturation level of the echo amplitude is reached where the reflection from the fracture surface is at maximum. For a mechanical fatigue crack, Iida et al. [11] reported a saturation level to be at 10 µm width of the crack surface opening. When closing a crack, it is defined acoustically closed, when a large change in the ul- trasonic signal is observed. However, according to Ibrahim et al. [10], the signal does not necessarily disappear com- pletely. The effect of flaw opening to the detection sensitivity of different mechanical fatigue cracks has been studied, e.g., by Yoneyama et al. [24]. Yoneyama et al. [24] loaded three different mechanical fatigue cracks by mechanical tensile and compressive loads to study the effect of crack opening on the obtained crack corner echo height. In their results, the echo height from crack opening significantly changed J Nondestruct Eval (2011) 30:143–157 147 Fig. 4 Relation between fatigue crack surface opening width and echo height of three different sizes of cracks (3.7 mm, 6.1 mm and 8.2 mm) [24] with the changed surface opening width. Furthermore, echo height with the same opening width increased with increas- ing crack depth (Fig. 4) finally being, with the largest open- ing widths (>18 µm), approximately the same as what was obtained from an EDM-notch. The authors attributed their results to the effect of crack opening width, but they did not discuss the reasons for this. The results may only reveal the effect of different cross-sectional areas taking part to the re- flection of the acoustic energy. In the studies of detection sensitivity of mechanical fa- tigue cracks the effect of material condition to the inter- dependence of the opening of the crack and obtained ul- trasonic echo amplitude height have been studied, e.g., by Becker et al. [2]. Ultrasonic amplitude height obtained from mechanically loaded fatigue cracks showed different behav- ior with cold-worked and annealed materials. In the cold- worked condition, the amplitude height decreases gradually, when the crack is closed (compressive loading is increased). After annealing the sample, the change in amplitude height is much faster, i.e., the slope of amplitude change vs. ap- plied load is steeper than that of the cold-worked sample. This was attributed to the material condition as in the cold- worked material the crack tip closes first and after that grad- ually the rest of the crack, and the last part to close is the crack mouth. This is seen in the change of the ultrasonic amplitude first as a slow decrease, when the less reflective crack tip closes, and then as faster decrease when the rest, more reflective parts of the crack close. In the annealed ma- terial, the whole crack closes practically at the same time. This difference was attributed to the compressive residual stresses present near the crack tip of the cold-worked spec- imen causing immediate increase of the compressive stress from the beginning of the compressive loading. In the an- nealed material sample there is no stress near the crack tip as unloaded. Furthermore, the total signal drop was with the cold-worked material 32 dB and with the annealed material 22 dB, and both were reached at the yield stress of the ma- terial. Becker et al. [2] attribute the difference in the total amplitude drop and the magnitude of the drop to the force the fracture surfaces bear. In the annealed material much less force is available, because the yield strength of the annealed material is lower, being half of that of the cold-worked ma- terial. As a separated factor of the crack opening, the present stress state affects the echo amplitudes obtained from the crack opening corner, fracture surface and crack tip. Hence, the results from the open literature indicated here do not only show the crack opening differences, but also the effect of the present stress state. The effect of the stress state is handled as the effect of residual stresses in the following. This is the case also during the in-service inspection, where the present stresses are not dynamic but static. The possible present residual stresses affect the de- tectability of a flaw, as tensile residual stress opens and compressive residual stress closes the flaw. Studies on the effect of residual stresses have been performed experimen- tally with external mechanical loading, e.g., by Iida et al. [11], Yoneyama et al. [24], Becker et al. [2], Ibrahim et al. [10] and Denby et al. [5]. Theoretical studies on this topic have been published, e.g., by Temple (1985) and Wirdelius (1992). Denby et al. [5] mentioned that the reflection from fatigue flaws is most seriously affected by the compressive stresses, of all the flaws. Furthermore, flaw tips of the ther- mal fatigue flaws are considered to be the most challenging ones as the flaw tips are surrounded by a plastic zone already under compression. The detectability is affected so, that the increasing tensile stresses (opening the crack) increase the echo amplitude and increasing compressive stresses (closing the crack) decrease the amplitude. With high enough loading, there will be a plateau in the amplitude value both in tension and compres- sion. Similar phenomenon has been found for carbon steels [11] and austenitic stainless steels [2]. With high enough tensile loading the echo amplitude obtained from a mechani- cal fatigue crack may be almost identical to the one obtained from an EDM-notch [24]. The effect on the echo amplitude, and detection sensitiv- ity, may be dependent on the direction of the load change, i.e., increasing or decreasing loading. As, in addition to plateau values obtained, a clear hysteresis may be recorded in the change of the ultrasonic amplitude during cyclic load- ing of a crack. This was shown by Becker et al. [2] with a mechanical fatigue crack in an austenitic stainless steel. During cyclic loading between tension and compression, the first loading to tension did not change the height of the sig- nal amplitude from the unloaded condition. However, by the followed high enough compressive stress, the ampli- tude height dropped markedly, finally stabilizing at a lower plateau at loads well over the yield stress of the material. 148 J Nondestruct Eval (2011) 30:143–157 During the second load cycle (unloading-tension-unloading- compression), while unloading from compression, the sig- nal remained in the lower plateau, beginning to increase at a lower force than required to reach the plateau. In max- imum tension, the same maximum value of the amplitude height was reached as in the first cycle. When the sample was loaded again in compression, lower force was needed to keep the lower plateau. Becker et al. [2] attributed the seen hysteresis in the change of the amplitude height to the plastic deformation taking place during the first compressive loading cycle. The detection sensitivity is differently affected by loading with mechanical fatigue and thermal fatigue cracks. Even Fig. 5 Different stress conditions affecting the obtained ultrasonic echo height from a thermal fatigue crack in AISI 304 type austenitic stainless steel [2] though the basic phenomenon of ultrasonic echo amplitude change under cyclic loading is similar, there are some im- portant differences. Typically, thermal fatigue cracks show low values of echo amplitude. Becker et al. [2] reported that thermal fatigue cracks showed low values of echo amplitude already when no external load was applied. As Fig. 5 shows, by application of an external load the thermal fatigue crack followed the theoretical behavior of reflection amplitude as a function of crack opening (the theoretical amplitude curve is shown in Fig. 6). That is, when the crack is under tension the plateau of high amplitude is obtained and under com- pression the low amplitude plateau of reflection is observed. When no external load is applied, the obtained echo heights lie in the lower part of the steep slope region between the plateaus. Consequently, unloaded thermal fatigue cracks are very sensitive to changes in the loading conditions. This is a result of the crack tightness and rough fracture surfaces al- lowing the surface heights to be, partly already as unloaded, in contact. Apart from this, mechanical fatigue cracks ex- hibit higher echo amplitudes as unloaded, hence being more detectable. Although unloaded thermal fatigue cracks were unde- tectable (with 50% DAC criterion), the application of the tensile stress equal to the yield strength made most of them detectable. Detection of cracks with such tightness in ser- vice conditions substantially depends on stress condition. If cracks are filled with water, according to Becker et al. [2], they can be undetectable under any realistic stress condi- tions as a result of the better acoustic energy transfer ability of water than air. The obtained ultrasonic echo height, and detection sen- sitivity, under different stress conditions is also affected by the fracture surface roughness (i.e., planarity) of the flaw. Fig. 6 Theoretical ultrasonic reflection coefficient as a function of distance between parallel stainless steel plates separated by air and water (2.25 MHz, transverse wave at 45◦) [2] J Nondestruct Eval (2011) 30:143–157 149 Fig. 7 Energy reflection coefficient of 10 MHz longitudinal waves incident at (a) 20◦ and (b) 30◦ to the normal on the model crack under compressive stress [17] Fig. 8 Energy reflection coefficient for longitudinal waves at normal incidence (0◦) on rough cracks (three different rms values) under (a) 60 MPa and (b) 160 MPa compressive stress [17] The more planar the crack is, the bigger is the difference be- tween the high and low plateau values (i.e., the smaller is the value of low plateau amplitude). This indicates that the non-planar geometry of the cracks may prevent their full clo- sure. Under compressive loading, the smoother cracks can be acoustically more closed than the rougher ones having higher peaks on their fracture surface. This phenomenon has been reported for both carbon steel and austenitic stainless steel [2, 10]. There are also theoretical studies on the impact of com- pressive loads to the crack detection including calculation of reflection coefficients for different fracture surface rough- ness [17] and evaluation of echo amplitudes as a function of tilt angle [21] of flaws under different stress states. The smaller scale roughness shows higher effect on the reflection coefficient than the larger scale roughness with different tilt- ing angles for the same probing frequency and applied load (Fig. 7). With smaller roughness the higher amount of con- tact points allows a higher amount of energy transmission through the crack. The use of higher frequency reduces the influence of the compressive stress (Fig. 8). Theoretical evaluations on detection sensitivity of Wird- elius [21] show that the background pressure affects differ- ently the obtained echo amplitude, when the flaw tilt angle is varied. In the specular reflection (Fig. 9a), the signal level is decreased until the background pressure reaches 200 MPa. In normal incidence (Fig. 9b), the tip-diffracted signal is dropped from the signal level of an open flaw markedly already with 50 MPa background pressure and increase to 200 MPa decreases the amplitude only a bit more. The effect of stress state and tilt angle to detection sensi- tivity was experimentally studied by Ibrahim et al. [10] re- sulting in differences in the echo heights of different incident beam angles under different compressive stresses. Ibrahim et al. [10] studied pulse-echo response from crack open- ing corners of three different mechanical fatigue cracks and, hence, their results are not directly comparable to the theo- retical results given in Fig. 9. However, the tendency shown in Fig. 10, that the highest detectability is achieved with 45◦ probe, was also shown by Wirdelius et al. [21]. The good re- sults of 45◦ probe were attributed to the favored orientation of large portion of small facets on the flaw fracture surface. Sizing of the flaw, similarly than detection but probably with even stronger influence, is affected by the combined effect of the flaw opening width, residual stresses and ma- terial condition. However, there are no studies showing the 150 J Nondestruct Eval (2011) 30:143–157 Fig. 9 Theoretical evaluation of pulse-echo signal responses for a penny-shaped flaw (depth 60 mm, diameter 4 mm) at different tilts and under different background pressures, with a zero degree, 3 MHz longitudinal wave probe. Different line types indicate different stresses: 200 MPa (- - -), 50 MPa (-·-·-) and unloaded open crack (—). Tilts were (a) 0◦ (parallel to scanning surface), (b) 90◦ (perpendicular to the scanning surface) and (c) 30◦ [21] Fig. 10 Maximum signal responses from three different fatigue cracks as a function of compressive stress with three different 5 MHz angled transverse wave probes [10] effect of different flaw tip openings under different loading conditions, as was the case with detection sensitivity stud- ies. Hence, all the sizing sensitivity studies are connected to different opening widths at the flaw mouth. The effect of material condition to the obtained ultra- sonic echo height from the flaw tip, i.e., sizing sensitivity, under different loading conditions has been studied, e.g., by Becker et al. [2]. The reported difference in the rate of change of the ultrasonic response from loaded crack tips of cold-worked and annealed materials was attributed to the different condition of residual stresses in the material. This was seen as slower change of obtained amplitude height with the cold-worked material and faster change with the annealed material. This difference was attributed to the first closing crack tip with the cold-worked material, while with the annealed material the faster amplitude drop indicates that the whole crack closes practically at the same time. The growth of a fatigue crack always induces a plastic zone in the material around the crack tip, affecting the siz- ing sensitivity. With austenitic stainless steels the radius of the crack tip plastic zone can be from some hundred mi- crometres to some millimetres, depending on the loading used during the crack growth. The residual stresses inside the plastic zone are compressive, caused by the plastic ten- sile loads in front of the crack tip during crack growth. Such compressive stresses around the crack tip promote closure of the crack tip. In addition to the plastic zone around the crack tip, a plastic wake forms at the fracture surfaces of the crack in any material during fatigue crack growth. By annealing the material, the plastic areas are stress relieved resulting in stress-free material. Consequently, the echo amplitude from the crack and crack tip will behave differently during load- ing. However, although there are differences between cold- worked and annealed materials, the fatigue crack tips in both materials are very tight and sharp. They give, already as un- loaded, a very weak ultrasonic response making the flaw sizing a difficult task. In case some loads are applied; the crack tip response is changed and if the loads are compres- sive they will result in marked difficulties in flaw sizing as the obtained amplitude is decreased. The effect of residual stress on the flaw sizing sensitiv- ity and accuracy has been studied, e.g., by Iida et al. [11] and Temple [17]. In these studies, test specimens containing different types of flaws were mechanically loaded. Studies were performed with austenitic and ferritic steels under dif- ferent stress conditions both with metallic fracture surfaces and surfaces covered with oxide layers. Ultrasonic measure- ment methods used were, amongst others, based on crack tip reflection and diffraction. J Nondestruct Eval (2011) 30:143–157 151 Fig. 11 Interaction between compressive stress and relative strength of crack tip diffracted signal [17] The sizing sensitivity is reduced as the obtained ampli- tude height from the crack tip is decreased with increasing compressive stress. However, the lateral scanning graphs of flaw tip reflection show similar shape under different stress states. On the contrary, the flaw tip echo height is increased by increased tensile stress. According to Iida et al. [11] the sizing accuracy of DAC% method showed more sensitivity to stress changes than the amplitude drop method. The flaw tip reflection is measurable for deeper flaws, but that may not be the case with shallow flaws. Iida et al. [11] mentioned the threshold depth to be 0.9 mm; with the smaller flaws they could not measure the tip reflection echo even under tensile loading. The sizing sensitivity is reduced by increased compres- sive stresses also with flaw tip diffraction based sizing tech- niques. With these methods the strength of the crack tip diffracted signal show clear correlation to applied compres- sive stress. Temple [17] has published results of correlation between theoretical calculations and experimental studies of stress vs. flaw tip diffraction (Fig. 11), showing reduction in the strength of the diffracted signal with increasing com- pressive stress. However, in the studies of Temple [17] the diffracted signal was not completely lost even with the max- imum applied compressive stress. The maximum loss of sig- nal strength was about 13 dB with maximum applied stress of 260 MPa. More recently Packalén et al. [14] and Kemppainen et al. [12] studied the effect of crack opening on ultrasonic sizing on a set of artificial cracks that were carefully de- structively analyzed after inspection. They confirmed that the magnitude of the crack opening affects the amplitude of the crack tip signal in that smaller opening gives smaller signal amplitude. This has direct consequence on sizing per- formance: with decreasing signal amplitude, the likelihood of misinterpretation in identifying of the crack tip signal increases. Similarly, with increasing noise amplitude (for flaws near weld), the likelihood of misinterpretation in iden- tifying crack tip signal increases. 3.1.3 Effect of Crack Orientation on Detection and Sizing Detection is changed from specular to off-specular field as the tilt of the flaw increases [13]. Especially with flaws hav- ing rough fracture surfaces, the effect of tilt angle on the detectability is pronounced. When detecting flaws, the inspection performed in nor- mal (0◦) position from the opposite surface, where the surface-opening vertical flaw is located, does not provide very good results. In this position the ultrasonic wave hits first the flaw tip. The tip of a natural flaw is tight and does not provide significant reflection surface. Instead, the flaw may diffract quite a big portion of the ultrasonic energy, re- flecting back only a smaller amount of it. Larger tilt angles provide higher echo amplitudes, when the fracture surface and the opening corner become more “visible” for the ultra- sonic beam [1]. This is due to the combination of the raising amplitude of the corner echo and the fracture surface reflec- tion and scattering. The different incident beams may cause a self-shadowing phenomenon, which may occur when a wave is incident onto a rough surface at a sufficiently tilted angle. Then part of the surface is not directly “illuminated” by the incoming wave, but it is shaded by other parts of the surface. Self-shadowing spoils the phase coherence of the adjacent surface scattered waves and, thus, the amplitude of the overall scattered field may be diminished [13]. Occurrence of self-shadowing de- pends on the surface profile and is therefore difficult to take into account precisely. In some cases a small flaw may produce higher response than a similarly tilted large flaw [2]. As a consequence, it is possible that when the flaw grows, repeated inspections would show decreasing signal amplitude. The degree of mis- orientation affects the strength of the diffracted ultrasonic signal from the flaw tip [18]. Figure 12 shows results ob- tained with three different 0◦ probes (different wave modes and frequencies). All the graphs show that by changing the angle of the incident of ultrasonic wave, the obtained ampli- tude height may be remarkably changed. 3.1.4 Effect of Oxide Film on Detection and Sizing During in-service inspection the tested flaw may be filled with air, water and/or have an oxide layer on its fracture 152 J Nondestruct Eval (2011) 30:143–157 Fig. 12 Pulse-echo response of a mechanical fatigue crack tip for (a) 5 MHz longitudinal waves, (b) horizontally polarized 2.25 MHz transverse waves and (c) vertically polarized 2.25 MHz transverse waves. Solid lines show results of theoretical modelling and crosses are measured data [18] surfaces entailing different acoustic impedances and, hence, affecting the reflection and transmission of ultrasound. Ac- cording to Crutzen et al. [4], the presence of corrosion prod- ucts in the flaw enhances its transparency during ultrasonic inspection decreasing detectability. Filling the crack with oxide or water results in higher sensitivity to amplitude drop, when the crack is closed as with an air-filled crack [2]. That is, the transmission of ultrasound through the crack occurs earlier if the crack is filled with oxide or water. The sen- sitivity to tightness is ranked to be highest with water-filled cracks, second highest with oxide-filled cracks, while metal- lic, air-filled cracks are the least sensitive. The theoretical calculations (pulse-echo, 45◦, 4 MHz) of Temple [17] show similarly that a narrow (opening 2 µm) water-filled flaw gives 9.6 dB lower signal than air-filled one. Similar drop is obtained with a 4 µm wide flaw in 2 MHz inspection. Some authors report opposite results from their studies. For example, Iida et al. [11] reported that if an oxide layer is present on the fracture surfaces, the changes in the ul- trasonic echo heights, when closing the crack, are remark- ably reduced. This was explained as a cause of air present in the crack hindering changes of the reflection coefficient of the sound pressure. Iida et al. [11] made a conclusion from their results that, even if compressive stresses are present in a component during the shutdown of a plant, the oxide films present on the fracture surfaces do not reduce the detectabil- ity of cracks. During flaw sizing the oxide film grown on the fracture surfaces affects the sizing capability. This is due to the oxide layer on the fracture surfaces holding the metallic surfaces separate. The difference in the impedances of the metal and the oxide affects the obtained ultrasonic amplitude from the oxide filled flaws, as shown, e.g., by Iida et al. [11]. 3.1.5 Summary of Crack Characteristics Essential for UT In summary, the characteristics relevant for crack represen- tativeness for UT techniques in general are: 1. location and orientation of the crack, 2. size of the crack, 3. opening of the crack through the whole path and at crack tip, 4. fracture surface roughness, 5. filling of the crack with some substance (e.g., water or oxide). 4 Measurement Methods for Crack Characteristics A consistent set of measurement methods is needed to facili- tate meaningful comparison of crack characteristics between service-induced cracks and various artificial cracks. While the importance and effect of several crack characteristics on NDE has been studied in the open literature (see paragraph 3), in most of these studies the parameters were not directly measured on actual service induced cracks. The first com- prehensive framework for measuring crack characteristics for NDE was published by Wåle et al. [22, 23]. He mea- sured comprehensive set of available images from service- induced cracks and, to facilitate this work, defined a set of measurement methods to be used. To date, this remains the most comprehensive data published on the crack characteris- tics of service-induced cracks. It must be noted, that the raw material, from which this information was extracted, was of- ten of bad quality. The images were from failure analysis re- ports and taken to clarify reasons for failure—not to describe cracking for in-service inspection. Furthermore, all the mea- surements were done manually from paper images. Conse- quently, the measurements were rather laborious and partly inaccurate. This limited the possible measurement methods as well as the number of measured cracks. In 2007, Trueflaw developed a more advanced set of mea- surement methods to overcome some of the limitations in the original Wåle methods resulting from the limitations mentioned above. This development was done in connec- tion to a national development project where (among other J Nondestruct Eval (2011) 30:143–157 153 Fig. 13 Crack opening at surface data from service induced cracks. Bars show measured range and line shows median value. Values from [22, 23] Fig. 14 Crack opening at midway data from service induced cracks. Bars show measured range and line shows median value. Values from Wåle [22, 23] Fig. 15 Crack opening close to crack tip data from service induced cracks. Bars show measured range and line shows median value. Values from Wåle [22, 23] Fig. 16 Crack branching data from service induced cracks. Bars show measured range and line shows median value. Values from Wåle [22, 23] 154 J Nondestruct Eval (2011) 30:143–157 Fig. 17 Crack correlation length data from service induced cracks. Bars show measured range and line shows median value. Values from Wåle [22, 23] Fig. 18 Crack intersections data from service induced cracks. Bars show measured range and line shows median value. Values from Wåle [22, 23] Fig. 19 Crack surface roughness data from service-induced cracks. Bars show measured range and line shows median value. Values from Wåle [22, 23] Fig. 20 Crack turns/mm data from service-induced cracks. Bars show measured range and line shows median value. Values from Wåle [22, 23] J Nondestruct Eval (2011) 30:143–157 155 aims) a wide set of different artificially produced cracks were characterized in an effort to provide dataset on artificial crack characteristics comparable to the Wåle data [14]. Care was taken to retain comparability to the Wåle measurements in all possible measurements. Simultaneously, the measure- ment methods were developed to be more suited on auto- matic measurements with much improved repeatability and accuracy than what could have been reached manually. This measurement methodology is described in more detail in the Appendix. This methodology (where applicable) is currently used by Trueflaw to document all manufactured cracks and it is presented here as a proposal for unified set of measurement methods for crack characteristics. It should be noted, that not all the measured characteristics presented may not be equally significant for NDE; some were included in the orig- inal Wåle reports for other purposes (e.g. leak-rate estima- tion) and they are included for completeness. 5 Available Data on Service-Induced Cracks To this date, the most comprehensive published set of quan- titative data on crack characteristics of service-induced cracks remains the work by Wåle [22, 23]. The data is re- visualized here for easy reference in Figs. 13–20. 6 Process for Confirming Artificial Flaw Representativeness A procedure for confirming and demonstrating representa- tiveness of selected artificial flaws is outlined the following: 1. Determine essential parameters. Consult paragraph 3 and inspection vendors to determine set of essential crack characteristics for the current in- spection method. If uncertainty exists on whether a cer- tain parameter is essential, it should be included to be conservative. 2. Check relevant data from service-induced cracks. Assign target values and/or range for essential parame- ters based on available data from service-induced cracks. Data from paragraph 5 can be used. 3. Confirm that measured data from artificial cracks con- forms to assigned criteria. Crack characteristics for Trueflaw cracks can be obtained from manufacturing documentation. If the essential char- acteristics from used artificial cracks match the criteria, the representativeness is confirmed. Otherwise, the dis- crepancy must be justified by, e.g. technical justification. Open Access This article is distributed under the terms of the Cre- ative Commons Attribution Noncommercial License which permits any noncommercial use, distribution, and reproduction in any medium, provided the original author(s) and source are credited. Appendix: Measurement Methods for Crack Characteristics This annex describes the methods used by Trueflaw to mea- sure various crack characteristics from digital images. These methods were developed to be comparable with the methods used by Wåle [23] but to overcome some of the limitations in these methods. A.1. Crack Detection, Centerline and Opening Measurements Measurements are done on digital image. The image plane is selected so, that the long side of the image corresponds to 0◦ orientation of the crack. Measurement is done for each ver- tical image line. The line is smoothed, and background sub- tracted. The lowest brightness point is considered to be the crack. If this lowest point (signal) exceeds the mean devia- tion from background (noise) by a certain limit, it is consid- ered a valid crack. The S/N limit is adjustable and reported for each measurement individually. The crack edges are then measured from the unsmoothed brightness curve as width of the half-maximum. The crack centerline is the arithmetic mean between the crack edges. The image resolution (num- ber of measurement lines per millimeter) is reported for each measurement image. A.2. Crack Roughness, Intersections and Correlation Length For these measurements, the crack length is divided to sec- tions with a typical length of 1.5 mm (Wåle [23] used 1– 2 mm). For each section, a line is fitted to the crack cen- terline data obtained. Fitting is done by the least squares method. For each line, the roughness values Ra and Rz as well as number of intersections, and simplified correlation length are calculated. The Ra is calculated according to fol- lowing formula: Ra = ∑ |e| l , (A.1) where e is the difference between the fitted line and crack centerline and l is the measurement length. The Rz is calculated according to following formula: Rz = ∑5 i=1 pi − ∑5 j=1 vj 5 , (A.2) where pi is the ith highest peak on the measurement length and vj is the j th lowest valley on the measurement length. 156 J Nondestruct Eval (2011) 30:143–157 The number of intersections is defined as a point, where two preceding lines are below the fitted line and two following lines above it or vice versa. The correlation length is defined as the length of the measurement range divided by twice the number of intersections. A.3. Crack Turns To calculate the number of macroscopic turns in the crack, the crack centerline curve is divided to sections being typ- ically 0.5 mm. For each section, a line is fitted with the least square method. Whenever the angle between succes- sive lines differs by more than 30◦, a turn in the crack is identified. A.4. Orientation To calculate the crack orientation, a line is fitted to the crack centerline curve with the least squares method. The reported orientation is the angle of this line against the picture orien- tation. A.5. Macroscopic Shape, Branching, Microstructure, Discontinuities Macroscopic shape, branching, microstructure and discon- tinuities are manually calculated from the image with the same terms and methods as used by Wåle [23]. The shape is determined visually by one of the words: straight, winding, bend, bilinear or branched. The reported value for branching is the number of branches (greater than five grain diameters) per mm crack length. Microstructure is described with, e.g., one of the words: equi-axed grains, column formed grains (weld metal), cold worked, cast microstructure. The value reported for discontinuities is the number of discontinuities manually counted for the image. References 1. Ahmed, S.R., Saka, M.: A sensitive ultrasonic approach to NDE of tightly closed small cracks. J. Press. Vessel Technol. 120, 384–392 (1998) 2. Becker, F.L., Doctor, S.R., Heasler, P.G., Morris, C.J., Pitman, S.G., Selby, G.P., Simonen, F.A.: Integration of NDE reliability and fracture mechanics—Phase I report. NUREG/CR-1696 PNL- 3469, vol. 1, 170 p. (1981) 3. Crutzen, S.J., Jehenson, P., Nichols, R.W., McDonald, N.: The ma- jor results of the PISC II RRT. Nucl. Eng. Des. 115, 7–21 (1989) 4. Crutzen, S., Lemaitre, P., Iacono, I.: Realistic defects suitable for ISI capability evaluation and qualification. In: Proceedings of the 14th International Conference on NDE in the Nuclear and Pres- sure Vessel Industries, 24–26 September 1996, Stockholm, Swe- den, pp. 153–163 (1996) 5. Denby, D., Duncumb, A.C.: The effects of stress on the ultra- sonic detectability of defects. In: Proceedings of the Conference of Nondestructive Testing in the Fitness-for-Purpose Assessment of Welded Constructions, pp. 73–81. The Weldiung Institute, Cam- bridge (1984) 6. Edwards, R., Gruber, G., Watson, P.: Fabrication of performance demonstration initiative specimens with controlled flaws. In: Pro- ceedings on 13th International Conference on NDE in the Nuclear and Pressure Vessel Industries, 22–25 May 1995, Kyoto, Japan, pp. 167–176 (1995) 7. Edwards, R.L., Watson, P.D., Gruber, G.J.: Fabrication of speci- mens with controlled flaws for procedure development and per- sonnel training and qualification. In: Proceedings on 12th Inter- national Conference on NDE in the Nuclear and Pressure Vessel Industries, 11–13 October 1993, Philadelphia, Pennsylvania, USA (1993), pp. 93–100 8. Gauthier, V.: Thermal fatigue cracking of safety injection sys- tem pipes non destructive testing inspections feedback. In: Pro- ceedings of NEA/CSNI Specialists’ Meeting on: Experiences with Thermal Fatigue in LWR Piping Caused by Mixing and Stratifica- tion, 8–10 June, Paris, France (1998), pp. 436–453 9. Green, E.R.: Worst-case defects affecting ultrasonic inspection re- liability. Mater. Eval. 47, 1401–1407 (1989) 10. Ibrahim, S.I., Whittaker, V.N.: The influence of crack topogra- phy and compressive stresses on the ultrasonic detection of fatigue cracks in submerged arc welds. Br. J. NDT, September, pp. 233– 240 (1981) 11. Iida, K., Takumi, K., Naruse, A.: Influence of stress condition on flaw detectability and sizing accuracy by ultrasonic inspection. In: The Ninth International Conference on Nondestructive Evaluation in the Nuclear Industry, 25–28 April, Tokyo, Japan, pp. 563–567 (1988) 12. Kemppainen, M., Virkkunen, I., Packalén, T., Sillanpää, J., Paussu, R.: Importance of crack opening in UT inspection qual- ification. In: Proceedings of the 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pres- surised Components, 8–19 October, Budapest, Hungary, p. 93– 105 (2007) 13. Ogilvy, J.A.: Model for the ultrasonic inspection of rough defects. Ultrasonics 27, 69–79 (1989) 14. Packalén, T., Sillanpää, J., Kemppainen, M., Virkkunen, I., Paussu, R.: The influence of the crack opening in the UT inspec- tion qualification. In: Proceedings of the 6th International Confer- ence on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, 8–10 October, Budapest, Hungary, pp. 463–470 (2007) 15. Pirson, J., Roussel, G.: Emergency core cooling system pipe crack incident at the Tihange unit 1 plant. In: Proceedings of NEA/CSNI Specialists’ Meeting on: Experiences with Thermal Fatigue in LWR Piping Caused by Mixing and Stratification, 8–10 June, Paris, France, pp. 103–114 (1998) 16. Saka, M., Fukuda, Y.: NDT of closed cracks by ultrasonic propa- gation along the crack surface. NDT E Int. 24(4), 191–194 (1991) 17. Temple, J.A.G.: The effects of stress and crack morphology on time-of-flight diffraction signals. Int. J. Press. Vessels Piping 19, 185–211 (1985) 18. Toft, M.W.: Experimental studies of ultrasonic reflection from var- ious types of misoriented defect. In: Proceedings of 21st Annual British Conference on Non-Destructive Testing – NDT, vol. 86, 193–206 (1986) 19. Waites, C., Whittle, J.: The status of performance demonstration and evaluation developments. Insight 40(12), 810–813 (1998) 20. Wirdelius, H., Osterberg, E.: Study of defect characteristics essen- tial for NDT testing methods ET, UT and RT. SKI Project Number 98267, SKI Report 00:42, October 2000, Sweden, 50 p. (2000) 21. Wirdelius, H.: Probe model implementation in the null field ap- proach to crack scattering. J. Nondestruct. Eval. 11(1), 29–39 (1992) J Nondestruct Eval (2011) 30:143–157 157 22. Wåle, J., Ekström, P.: Crack Characterisation for In-service In- spection Planning, SKI Projekt 14.4-940389, 94164 SAQ/FoU- Rapport 95/70, SAQ Kontroll AB, Stockholm, Sweden, 84 p. (1995) 23. Wåle, J.: Crack characterization for in-service inspection planning – an update. SKI reference 14.43-200543105, ISRN SKI-R-06/24- SE, SKI, Stockholm, Sweden (2006) 24. Yoneyama, H., Senoo, M., Miharada, H., Uesugi, N.: Comparison of echo heights between fatigue crack and EDM notch. In: Pro- ceedings of 2nd International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurized Components, 24– 26 May, 2000, New Orleans, Louisiana, USA, 8 p. (2000)

Proposed improvements for use of different qualification defect types - three generations of defects Iikka Virkkunen, Trueflaw Ltd. Mika Kemppainen, Trueflaw Ltd. Raimo Paussu, Fortum Nuclear Services Ltd. Jani Pirinen, Fortum Nuclear Services Ltd. Petri Luostarinen, Fortum Loviisa Powerplant INTRODUCTION The reliability of non-destructive examination depends on multitude of different factors. These range from physical aspects of the used technology (e.g, wavelength of ultrasound) to application issues (e.g. probe coupling or scanning coverage) and human factors (e.g. inspector training and stress or time pressure during inspection). To assess the vital issue of NDE reliability, qualification of used methods is nowadays required in most countries for nuclear inspections. The exact manner in which the qualification is implemented varies, but in most cases qualification includes practical trials to some extend to verify and confirm that the inspection procedure functions as intended. In practical trials the inspection is performed on a known, flawed test piece and reliability is judged by comparing the acquired inspection result with known state of the test piece. However, practical trials themselves have proven somewhat challenging to arrange. In particular, it is rather challenging to show that the inspection arrangement and flawed test piece used are representative to real inspection situation. In practice, quality and applicability of the qualification is heavily dependent on the quality of the used test pieces. Over the years, number of different approaches has been used to produce the flawed test pieces needed. Both the inspection techniques and qualification practices have developed. Consequently, both the quality and requirements for test pieces have developed. The first flawed samples contained mechanically machined flaws, i.e. flat bottom holes, saw cuts and later EDM (electro-discharge machined) notches. We call these "1st generation flaws" (later "1G"). These can be easily manufactured to tight tolerances. However, as NDE methods and requirements developed, it soon became apparent that their representativeness was not good enough to estimate true NDE performance. Mechanical notches are still used, for example, for signal calibration where representativeness to true cracks is not important. To improve on the mechanical notches, various welded flaw simulations have been developed. We call these "2nd generation flaws" (later "2G"). These are manufactured, in simple terms, by either implanting an existing crack by welding to the test piece, or by inducing cracking of the weld by carefully chosen weld parameters (see, e.g.,[1-2]). Most of the flawed test pieces currently in use apply 2nd generation flaws (see, e.g., [3]) These 2G flaws offer closer approximation to real cracks than 1G flaws, as the crack propagation is tortuous and some other crack parameters resemble true, service-induced cracks. Welded flaw simulations have been extensively used in inspection qualifications during the last decades. However, significant discrepancies remain between true, service- induced cracks and welded flaw simulation. The weld metal introduced in the production process may affect inspection in an unpredictable manner. Also, the flaw characteristics tend to differ from natural cracks, for example, the crack opening tends to be bigger and realistic cracks tip conditions are not attained. Not surprisingly, as NDE methods and requirements have developed, need for better flaws increased especially in cases that are sensitive to weld material (e.g. austenitic materials or EC-inspections) or critical crack characteristics (e.g. TOFD-sizing). To further improve from the welded flaw simulations, techniques were developed to grow real cracks without welding. These methods can be called "grown cracks" due to the fact that they rely on the same natural growth mechanism that might be responsible for growth of the service-induced cracks. We call these "3rd generation" flaws (later "3G"). With these methods, natural crack growth is accelerated and controlled to facilitate production of cracks. The use of test blocks with natural grown cracks has several advantages: the performance of the NDE system is shown with minimal uncertainties and inspectors get experience on true cracks and knowledge what to expect during ISI's. Also, there's no room for discussions about the validity of the samples. On the other hand, manufacturing cost of 3G cracks is generally greater than 2G defects. The third generation flaws have generally been available since early 2000's. In recent years such technology has matured, tried and tested [4-8]. Capability of the technique to produce realistic, representative flaws has been analysed by comparing the crack characteristics to the characteristics measured from service-induced flaws. This comparison has been made against measured values from service-induced flaws reported by Wåle [9,10]. Comparison done, e.g. in [8] for the crack opening values, has indicated that the flaws produced by the new technique are very representative for most of the service-induced flaws, when used as a reflector for different NDE development, training and qualification purposes. However, the technology is still much less used than first and second generation flaws and thus it's still a "newcomer" in many ways: inspectors have generally not seen these cracks to great extent. Production methods to produce grown cracks have been under significant scrutiny and they have now been validated and qualified for use in inspection qualification in various countries. SELECTING ARTIFICIAL DEFECTS Up until late 1990's the choice to make for test pieces was, in simple terms, to choose optimal combination of first and second generation flaws. Now, however, the range of flaw types has increased and the challenge is to choose optimal combination of first, second and third generation of flaws. There's basically two frameworks which offer guidance to selecting flaws: the ENIQ methodology and ASME code (see Whittle [11] for critical review of ASME and ENIQ). These may be used separately or in conjunction to perform a qualification. The ASME code (Section XI, Division 1, Appendix VIII, first published in 1989) gives direct guidelines and requirements for the test pieces used (See, e.g. [12] for review on various aspects of the ASME code). The demonstrated performance is based on statistical screening approach; if an examiner identifies 90% of the flaws in the specimen test set and does not exceed 10% false call rate, the examiners propability of passing the qualification is 90%. If the examiner identifies only 50% of the flaws and has false call rate of 30%, passing propability is reduced to 1% [13,14]. Based on the ASME acceptance criteria, the inspector pass probability for various POD's and false call rates can be calculated with monte carlo simulation. Results for such simulations are illustrated in Figure 1. It should be noted, that the curves describe pass probability on per test basis. If candidate is allowed to retake the test, then the combined probability of passing the test differs from the curves shown here (i.e. passing the test becomes significantly easier). However, the pass probabilities shown may give pessimistic view on actual inspector POD: data published from actually gathered performance demonstration test results ([15]) indicates that actual POD values demonstrated by the candidates far exceed estimates given above. Figure 1. Monte-Carlo simulation for ASME Section XI, Appendix VIII, Supplement 2 detection acceptance criteria (Table VIII-S2-1) with 10 flawed grading units (minimum required). Curves show different false call rates (FCR). Monte-Carlo simulation used 100000 random samples per point (21 million random samples in total). Appendix VIII includes number of supplements each for different (but generic) inspection type. The flaw types, size distributions and number of flaws are detailed for each inspection type in the corresponding supplement. For example, supplement 10 (dissimilar metal piping welds) gives the following guidance on the flaw types: ... (a) At least 60% of the flaws shall be cracks and the remainder shall be alternative flaws. Specimens with IGSCC shall be used when available. Alternative flaws shall meet the following requirements: (1) Alternative flaws, if used, shall provide crack-like reflective characteristics and shall only be used when implantation of cracks would produce spurious reflectors that are uncharacteristic of service-induced flaws. ... All in all Appendix VIII considers three types of flaws: i) weld implanted cracks (mechanical fatigue, thermal fatigue or SCC are implicitly assumed to be weld-implanted), i.e., 2G ii) Alternate flaws (presumably tight notches, "1.5G") and iii) Notches (1G). That is, 3G cracks are not considered by the Appendix VIII and/or they are implicitly considered equivalent to 2G weld-implanted cracks. This conclusion is supported by number of papers written soon after the initial release of the Appendix VIII ([16,2]). At the same time, the wording does suggest strong preference to natural cracks and does recognize possible problems with implantation. The problems with sample manufacturing and "less than perfect" test pieces is also identified by Becker [17] in review of Appendix VIII implementation. The main advantages of ASME qualification are its simplicity of use and generality. On the other hand, it's been criticized [11] for being overly general, providing insufficient verification for inspection reliability and for being costly (due to its heavy reliance on practical trials). The ENIQ methodology was developed to overcome the perceived shortcomings of the ASME qualification scheme. Its should be noted, that due to it's European roots, the ENIQ methodology gives general guidelines for inspection qualification. Each country can then adapt suitable national implemenation to use. The actual implementations vary considerably. 0  %   10  %   20  %   30  %   40  %   50  %   60  %   70  %   80  %   90  %   100  %   40  %   50  %   60  %   70  %   80  %   90  %   100  %   P ro b ab il it y   to  p as s   te st   POD  of  candidate   FCR=0%   FCR=5%   FCR=10%   FCR=15%   FCR=20%   FCR=25%   FCR=30%   FCR=35%   FCR=40%   FCR=45%   The starting point of the qualification for the ENIQ is the input information dossier. This contains information about the cracks that are expected in the component. The crack growth mechanisms as well as the critical flaw sizes are defined in the input information. The input information is typically prepared by the plant operator, who has best information on the possible damage mechanisms. Based on this input information, the inspection procedure is defined, usually by the inspection vendor. When these two are available, a technical justification (TJ) is prepared. It takes the relevant data from input information and inspection procedure and defines the most important parameters for successful inspection. The applicability and performance of the chosen procedure is then justified using previous experimental evidence, modelling parametric studies etc. Finally, guidance is given for the test blocks to be used for open and blind trials in qualification. So, in broad terms, the generality and simplicity of the ASME code has been replaced with an adaptive approach. For each component, the expected damage mechanisms are assessed and the inspection reliability is assessed on theoretical considerations. The role of practical trials is then, to confirm that the technical justification works as expected (and not to provide statistical evidence of the reliability of the inspection). With use of the technical justification, the test blocks can focus on testing the most important challenges of the inspection and the amount of needed test blocks and defects can be reduced. According to ENIQ, the amount of defects can be further reduced by using worst case -defects. In this case, the most difficult defects from inspection point of view are defined and tested for in the open and blind trials. The ENIQ methodology has it's own shortcomings that generally mirror those of the ASME code. Whereas for ASME code the statistics give very clear evidence on the proven performance, for the ENIQ the proven performance is more obscure since it's based on both technical justification and practical trials. The recent advances in risk-informed in service inspection (RI-ISI) underline the problem with qualitative demonstrated performance. Some recent advances to quantify demonstrated performance include Bayesian model with relative weights and probabilities assigned to TJ and practical trials based on expert judgement [18- 20]. If worst case defects are used to reduce number of test blocks, then reliance on technical justification increases further. Also, qualification is not just a final test of a frozen inspection method, but rather an iterative learning experience were inspectors develop and tune their inspection procedures to increase inspection reliability. Consequently, using worst-case defects focuses development resources on the difficult (but possibly improbable or impossible) defects on the expence of the more likely (but possibly easier) defects. The ENIQ documents do not give much guidance for flaw selection in test pieces, since this is, generally, considered to be an issue determined in the technical justifications and by the QB. Consequently, the methodology requires very high level of expertise available on damage mechanisms, inspection technology and flaw manufacturing when preparing the technical justification. Due to the complementing advantages and disadvantages of the ASME code and ENIQ they can be successfully used in conjunction. This is rather counter-intuitive, since ENIQ was develop to overcome the perceived weaknesses in ASME. In practice, this means using ENIQ to define the inspection case and critically design and study the inspection procedure (i.e. doing the input information and TJ), and using ASME as practical guidelines to required flaw populations etc. In fact, this approach is used as basis for grading in Finnish qualification guide documents [21]. It's still ENIQ, but with use of the know-how embedded in the ASME code. The disadvantage of using both ASME and ENIQ this way is, of course, that the cost of test blocks is higher than could be with ENIQ worst-case defects. On the other hand, extended practical trials provide increased confidence on the assessment done in the TJ. Table 1. compares the use of ENIQ and ASME. Table 1. Comparison of ASME and ENIQ qualification ASME ENIQ ENIQ(+ASME) Confidence provided well-defined but limited performance qualitatively-defined but possibly higher performance possibly high performance with well-defined lower bound Application General and simple Adaptive but complex Adaptive and simple Performance evaluated on Various types of flaws in the test block Expected or worst case defects in the case Expected or worst case defects in the case Cost High cost of test blocks High cost of TJ High cost of TJ + High cost of test blocks As for selecting optimal combination of 1G, 2G and 3G flaws, neither of the available guidelines help in (or even recognize) the choice between 2G and 3G flaws. CASE STUDIES: LOVIISA QUALIFICATIONS AND FLAW SELECTION In the following, three case studies of qualifications completed (or in progress) for the Loviisa power plant are presented. The Loviisa power plant is WWER-440 type PWR. The qualifications were done according to Finnish regulations [22], which closely follow the ENIQ methodology. Fortum has constantly developed their qualifications and actively searched to improve qualification practices and representativeness of used test pieces. Also, there's been a number of qualifications completed and hence significant experience gathered during recent years. For each case, the inspection target and input information is summarized; the considered flaw types and scope of inspection is documented. The input information considers three types of defects: specific defects, postulated defects and unspecified defects. The used test pieces and artificial flaws are described. Specific defects are defects which the damage mechanisms of potential defects are well known and defects have been detected in the inspection objects in question or in the similar structures either at Loviisa unit 1 or 2 or in other VVER-440 units outside Finland. The damage mechanisms of postulated defects are known and initiation and/or growing of the defects is assessed to be possible in the inspection object. Defect types have been observed in other locations in piping or components at Loviisa unit 1 or 2 or in other nuclear power plants (VVER, PWR, BWR), but not in the location to be inspected. Unspecified defects are defect types which have not been detected, nor are they postulated in the inspection object, or damage mechanisms are not identified. Due to confidentiality issues on the blind test pieces, not all the information can be published. The cases include use of different 1G, 2G and 3G flaw manufacturing techniques and to reveal their corresponding advantages. The decision on the used flaws is explained. Case 1. Steam generator collector dissimilar metal weld (1G, 2G and 3G defects) (2007 - 2009) Inspection of Steam Generator DMW is qualified for UT inspections. Inspection volumes are presented in Figure 2. The defect types specified in the input information are presented in Table 2. Detection target for personnel qualification in circumferential direction is 17 mm deep and 51 mm long defect and in axial direction 19 deep and 57 long defect. System detection target is 6x18mm for circumferential direction and in axial direction 7x21mm. Figure 2. Steam generator collector dissimilar metal weld geometry and inspection target. Table 2. Defect information for the steam generator DMW. Defect type Mechanism Specific defects Stress corrosion cracking, manufacturing defects (lack of fusion, slag lines, porosity) Postulate defects Stress corrosion cracking, Environmentally assisted fatigue cracking, Fatigue cracking Unspecified defects - The test pieces use 1G, 2G and 3G defects. 1G -EDM notches are used because they can be easily and affordably produced to various locations and sizes. This allows manufacturing of wide variety of flaws that cover all potentially interesting configurations. Especially the configurations that are difficult for NDE, but not perhaps likely to occur in real inspection are studied with notches. They are also used in locations or configurations, where other flaw manufacturing technologies are not viable, e.g. due to manufacturing constraints or high cost. However, the notches are not considered representative to true in-service cracks defined in the input information and thus they are not considered sufficient alone. 2G defects (welded solidification cracks) were used in the dissimilar metal weld (DMW) region to produce deep cracks. These are affordable to produce and they provide better correspondence with the defined crack properties than EDM-notches. On the other hand, the weld material introduced causes differences in the NDE response (change in noise level etc.) which disturb the qualification. Also, the flaw types are still different from the types specified in the input information. Consequently, they were not considered to be sufficient alone. Also, the 2G flaws were considered inappropriate for the base material, due to disturbances caused by the weld material. Yet, it was decided that they provide good compromise between cost of manufacturing and flaw representativeness for the deep cracks. 3G defects (in-situ produced thermal fatigue cracks) were used in the base material locations and in the buffer-fusion line for smaller defects. The 3G cracks offer good representativeness with the defect types specified in the input information. They can be well used in the base material and other areas, where welding would disturb the qualification. On the other hand, the cost of production increases proportionately to the size of the flaws and thus they were not used for big flaw sizes. Also, some of the NDE-worst-case defects could not be readily produced (e.g., flaws with tilt). Consequently, the 3G cracks alone were not considered sufficient and they were used in concert with 1G and 2G flaws. Case 2. RPV nozzle (1G, 2G and 3G defects) (2002 - 2010) Inspection of inner corner area of Reactor pressure vessel nozzle is qualified for UT and ET inspections. Inspection volume is presented in Figure 3. The defect information from the input information is summarized in Table 3. Qualification of inspection contains near area inspections (30 mm from inside surface). Figure 3. Inspection volume of RPV nozzle inner corner area. Inspection volume marked with A-B-C- D. Table 3. Defect information for the RPV nozzle. Defect type Mechanism Specific defects Subsurface volumetric (slag, porosity) and planar (LOF) welding defects, solidification cracking. Postulate defects Fatigue cracks in nozzle inside radius, under-clad fatigue cracks Unspecified defects Transverse fatigue cracks Slice of nozzle piece with width of 100 mm was used as open test piece in the first qualification trial in 2002. 1G EDM notches were used due to their ease of production: EDM flaws can be accurately produced, positioned, tilted, skewed and shaped. The notches are suitable for UT examination (often used as worst case reflector due to their specular reflection characteristics) and also for ET examination. However, the EDM notches were not considered to be representative for the defect types defined in the input information. 2G welded radial solidification cracks are used to simulate deep, open to surface and subsurface cracks for UT examination. One of the specific defect types in the input information was solidification cracking, so the flaw type offers good representativeness for these flaws. On the other hand, the extra weld material introduced in the process produces extra noise around the cracks, which disturbs UT. For ET, the welding disturbs strongly the examination of cracks and thus solidification cracks are not proper for qualifying ET examination. Radial 3G thermal fatigue cracks were used for qualifying ET examination of cladding surface area where welding would have disturbed the qualification. For the shallow cracks needed for ET qualification, the production cost of 3G flaws was smaller than for the deeper cracks needed for UT. Also they are representative to the postulated defects Case 3. Base material inspection, steam generator collector threaded hole (1G and 3G defects) UT inspection with phased array technique was qualified for steam generator collector threaded hole. The defect types and scanning area is shown in Figure 4. The defect types from the input information are summarized in Table 4. Scanning is done from surfaces on the flange top and inner wall side. Figure 4. Inspection volume of threaded holes of steam collector flange area (marked A-B-C-D) Table 4. Defect information for the steam generator collector threaded hole. Defect type Mechanism Specific defects Manufacturing defects (hot cracks, lack of fusion, slag inclusion) Postulate defects Fatigue cracks, Stress corrosion cracks Unspecified defects Thermal fatigue cracks 1G EDM notches were used, again, to cover wide variety of difficult to detect and/or difficult/costly to manufacture configurations such as very deep flaws and tilted flaws. Again, they were not considered sufficient alone due to lack of similarity to flaw types specified in the input information. 2G cracks were not used in this case. All the flaws are in base material and thus the weld material introduced by 2G methods were considered too disturbing to allow their use. 3G defects were used to get representativeness to the flaw types specified in the input information. Due to limitations in production time and cost constraints, 3G defects were not used for all locations. DISCUSSION AND CONCLUSIONS The above case studies demonstrate how the selection of different flaw types requires much expertise and experience beyond the guidelines given in ASME or ENIQ. On the other hand, quite clear and practical guidelines readily emerge from the practical case studies. These can be summarized as follows: 1G notches are best suited for studying range of different flaw configurations due to their ease of manufacturing and low cost. They can also be used to complement other flaw types in cases where production methodology limits possibilities. However, their response differs from real service-induced cracks found in service and thus they are not sufficient alone. 2G flaws are best used in cases, where they offer good compromise between producibility, representativeness and cost. They are more representative than the 1G flaws and offer significant cost advantage over 3G flaws in big flaw sizes. However, they are not suited to cases where the weld material introduced disturbs the inspection, e.g. in austenitic stainless steel base material UT (case 3) or in EC inspection (case 2). 3G cracks are important in cases where 2G cracks are not suitable due to weld disturbance. They offer better representativeness to real cracks but this advantage is weighted against cost of production. Since the cost of 3G cracks is proportional to crack size, they have been especially used in smaller crack sizes and /or smaller wall thicknesses. REFERENCES 1) Pherigo, G. L., Pherigo, A. L. 1994. Using Flaw Implants to Qualify Nuclear NDE Personnel. 12th International Conference on NDE in the Nuclear and Pressure Vessel Industries; Philadelphia, Pennsylvania; United States; 11-13 Oct. 1993. pp. 81-84. 2) Pherigo, G. L., 1992. Flawed Specimen Design and Manufacture for ASME Section XI, Appendix VII-VIII. 11th International Conference on NDE in the Nuclear and Pressure Vessel Industries; Albuquerque, New Mexico; USA; 30 Apr.-2 May 1992. pp. 225-230. 3) Horácek, L., Zdárek, J. 1996. Inspection qualification programme for VVER reactors and review of round robin test results. Nuclear Engineering and Design, 182, pp. 193–198 4) Kemppainen, M. 2006. Realistic artificial flaws for NDE qualification – novel manufacturing method based on thermal fatigue, Dissertation for the degree of Doctor of Science in Technology, Espoo, Finland, 2006. (Available online from: http://lib.tkk.fi/Diss/2006/isbn9512282631/) 5) Virkkunen, I., Kemppainen, M., Ostermeyer, H., Paussu, R., Dunhill, T. 2009. Grown cracks for NDT development and qualification, InSight, 2009 5, to be published. 6) Paussu, R., Virkkunen, I., Kemppainen, M. 2007. Utility aspect of applicability of different flaw types for qualification test pieces, Proceedings of the 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8 th – Oct 10th. Budapest, Hungary. p. 85-91. 7) Packalén, T., Sillanpää, J., Kemppainen, M., Virkkunen, I. and Paussu, R. 2007. The influence of the crack opening in the UT inspection qualification, Proceedings of the 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8th – Oct 10th. Budapest, Hungary. p. 463-470. 8) Kemppainen, M., Virkkunen, I., Packalén, T., Sillanpää , J., Paussu, R. 2007. Importance of crack opening in UT inspection qualification, Proceedings of the 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8 th – Oct 10th. Budapest, Hungary. 2007. p. 93-105. 9) Wåle, J., Ekström, P. 1995. Crack Characterisation for In-Service Inspection Planning. SKI-project 14.4-940389, SAQ/FoU report 95/07, SAQ Kontroll Ab, Stockholm, Sverige. 10) Wåle, J. 2006. Crack Characterisation for In-Service Inspection Planning- An update. SKI reference 14.43-200543105, ISRN SKI-R-06/24-SE, SKI, Stockholm, Sweden. 11) Whittle, M. J. 2009. A Review of Worldwide Practice and Experience in the Qualification of Ultrasonic Inspections of Nuclear Components Over the Past Two Decades. Insight, 51 (3), pp. 140 - 150. 12) Hedden, O, Cowfer, D., Batey, J., Spanner, J. and Becker, L. 2002. Overview of the Impact of Ultrasonic Examination Performance Demonstration on the ASME Boiler and Pressure Vessel Code. Journal of Pressure Vessel Technology, 124, pp. 254 - 260. 13) Cowfer, C.D. 1991. Basis / Background for ASME Code Section XI proposed Appendix VIII: Ultrasonic examination performance demonstration. Nuclear Engineering and Design, 131, pp. 313 - 317. 14) Serpan, C. Z., Mayfield, M.E., Muscara, J. 1997. US Nuclear Regulatory Commission Research for Primary System Integrity Regulations. Nuclear Engineering and Design, 171, pp. 1 - 14. 15) Becker, F. L. 2001. Examination Effectiveness Based on Performance Demonstration Results for Flaws at the Clad-to-Base Metal Interface, Proc. 3rd Int. Conf. NDE in Relation to Structural Integrity for Nuclear and Pressurized Components, Seville, Spain, November 14–16. 16) Spanner, J., Carr, F., Becker, F.L., Ammirato, F., Huffman, K. 1993. Performance demonstration for inservice inspection of nuclear power plant components. 1993 Pressure Vessels and Piping Conference, Denver, CO, USA, pp. 47-51. 17) Becker, F. L. 2001. Performance Demonstration -- 25 Years of Progress, Proc. 3rd Int. Conf. NDE in Relation to Structural Integrity for Nuclear and Pressurized Components, Seville, Spain, November 14–16. 18) Gandossi, L., Simola, K. 2005. Framework for the quantitative modelling of the European methodology for qualification of non-destructive testing. International Journal of Pressure Vessels and Piping, 82, pp. 814–824 19) Gandossi, L., Simola, K. A. 2007. Bayesian Framework For Quantitative Modelling of the ENIQ Methodology for Qualification of Non-Destructive Testing. European Communities, DG JRC, Institute for Energy, EUR22675EN, ISSN 1018-5593. 20) Gandossi, L., Simola, K., Shepherd, B. 2010. Application of a Bayesian Model for the Quantification of the European Methodology for Qualification of Non- Destructive Testing. International Journal of Pressure Vessels and Piping, 87, pp. 111-116. 21) Anon. 2007. The qualification of non-destructive in-service inspections of Finnish nuclear power plants. Qualification of personnel. SP8, Inspecta, Finland. available online: http://www.inspecta.fi/sfs/sertifiointipalvelut/docs/sp8_eng.pdf referred on 2010-09-13) 22) Anon. 2003. Nuclear power plant pressure equipment. In-service inspection with non-destructive testing methods. GUIDE YVL 3.8. STUK, Finland. available online: http://www.edilex.fi/stuklex/en/lainsaadanto/saannosto/YVL3-8 referred on 2010-09- 13)

Development and production of test specimens to evaluate an inspection issue Dr. Sc. (Tech) Mika Kemppainen, Dr. Sc. (Tech) Iikka Virkkunen Trueflaw Ltd., Espoo (FI) Dr. Henner Ostermeyer Kernkraftwerk Brokdorf, E.ON Kernkraft GmbH, Brokdorf (D) Dr. Markus Gribi Kernkraftwerk Beznau, Axpo AG, Döttingen 93 Dr. Mika Kemppainen, Dr. Iikka Virkkunen, Dr. Henner Ostermeyer, Dr. Markus Gribi 94 Abstract Assessing the reliability of non-destructive evaluation (NDE) is quite challenging, because it depends on a multi- tude of different factors ranging from physical aspects of the used technology to human factors. Generally, the re- liability can only be assessed via practical trials where the inspection is performed on a known, flawed sample and reliability is judged by comparing the acquired inspection result with the known state of the sample. Consequently, it is vital to have representative test specimens and representative cracks in order to get reliable information about NDE reliability. Over the years, several techniques have been developed to manufacture representative flawed test samples. The development has led to increasing representativeness in test specimens. In recent years the most significant trend has been the increased use of “grown cracks”, which currently offer the best representativeness available. However, the available techniques are still not used to their full potential so that additional guidance and experience are needed to improve practical trials. 1. Introduction The reliability of non-destructive evaluation (NDE) depends on a multitude of different factors. These range from physical aspects of the used technology (e. g. wavelength of ultrasound) to application issues (e. g. probe coupling or scanning coverage) and human factors (e. g. inspector training and stress or time pressure during inspection). Due to this complexity, the only practical way to assess inspection reliability and to confirm that the inspection procedure functions as intended is by using practical trials. In practical trials the inspection is performed on a known, flawed sample and reliability is judged by comparing the acquired inspection result with the known state of the sample. However, the practical trials themselves have proven somewhat challenging to arrange. In particular, it is paramount that the inspection arrangement and flawed sample used are representative of a real inspection situation. Representative flawed samples are vital in order to assess NDE reliability and to get necessary feedback for improving NDE. The first flawed samples used applied mechanically-machined flaws, i. e. flat bottom holes, saw cuts and later electro discharge machining (EDM) notches. They can be easily manufactured to tight tolerances. However, as NDE methods and requirements evolved, it soon became apparent that their representativeness was not good enough to estimate true NDE performance. Mechanical notches are still used in samples used, for example, for signal calibration where representativeness to true cracks is not important. Also, use of notches to estimate NDE performance is still justified in cases where no other flaw type can be manufactured. Here, of course, the justification becomes increasingly difficult, as there is no direct information about the true performance of the inspection. If an indication is actually found, the test samples with notches provide no help to explain it and additional information is then required. Sometimes this is covered by additional technical justification to justify the difference between NDE response obtained and the performance seen with unrealistic notches. Eventually, get- ting confidence in unclear indication issues may require a post-inspection open trial with realistic, natural cracks. To improve on the mechanical notches, various welded flaw simulations have been developed. They are manufac- tured, in simple terms, by either implanting an existing flaw by welding to the sample, or by inducing cracking of the weld by carefully chosen weld parameters. These simulations offer closer approximation to real cracks, as the crack propagation is tortuous and some other crack parameters resemble true service-induced cracks. Welded flaw simulations have been extensively used in inspection qualifications during the last decades. However, signifi- cant discrepancies remain between true, service-induced cracks and welded flaw simulation as the weld metal introduced in the production process may affect inspection in an unpredictable manner. Also, the flaw character- istics tend to differ from natural cracks, for example, the crack opening tends to be bigger and most importantly realistic crack tip conditions are not attained. Consequently, inferring real world performance is more difficult and the inspectors get less experience with true cracks. On the contrary, inspectors get experience, which may lead to unrealistic confidence on the inspectors skills or capability of the technique. Also, the technical justification required to address the differences to natural cracks makes the samples more case-specific. There’s a risk in redoing the qualification if the justification later proves invalid. Development and production of test specimens to evaluate an inspection issue 95 Not surprisingly, as NDE methods and requirements have developed, need for better flaws increased especially in cases that are sensitive to weld material (e. g. austenitic materials or eddy current [EC] inspections) or critical crack characteristics (e. g. TOFD-sizing). To further improve from the welded flaw simulations, techniques were developed to induce real cracks without welding. These methods are called “grown cracks” due to the fact that they rely on the same natural growth mechanism that might be responsible for growth of the service-induced cracks. With these methods, crack growth is accelerated and controlled to facilitate production of cracks. The use of test blocks with natural grown cracks has several advantages: the performance of the NDE system is shown with minimal uncertainties and inspectors get experience on true cracks and know what to expect during in-service inspections (ISIs). There is no room for discussion about the validity of the samples. In recent years such technology has matured and has been tried and tested [1]–[5]. Capability of the technique to produce realistic, representative flaws has been analysed by comparing the crack characteristics with the characteristics measured from service-induced flaws. This comparison has been made against measured values from service-induced flaws reported by Wåle [6, 7]. Comparison done, e. g. in [5] for the crack opening values, has indicated that the flaws produced by the new technique are very representative for most of the service-induced flaws, when used as a reflector for different NDE development, training and qualification purposes. The following figure (Fig. 1) shows a comparison of some often used defect types: EDM notches, welded crack simulations and grown cracks as well as a true crack from the literature. Production methods to produce grown cracks have been under significant scrutiny and they have now been validated and qualified for use in inspection qualification in various countries. Due to the investigations done on grown cracks, there is now much new information available on applicability of different artificial defects [2, 3]. Various essential characteristics of both service-induced cracks [6, 7] and arti- ficially produced grown cracks have been studied and documented [5]. Consequently, a sound body of technical information is now available for use in inspection qualification. However, this information is currently not used to its full potential. The nature of the technology is to accelerate a natural damage mechanism in a controlled way and hence cracks grow only as allowed by nature. In practice the crack growth follows the weakest path through the microstructure in the intended location. This is clearly an advantage of the technology. However, if one has applied other tech- nologies such as EDM notches or welded flaws earlier, and placed them freely, there may be a need to change the way of designing flaw specifications to follow more the natural crack growth. Such cases have been, for example, straight cracks specified in the fusion line of a winding weld where the true crack growth follows the shape of Fig. 1: EDM notch (a), welded crack simulation (b) [3], grown crack (c) [3] and true crack (d) from the literature [8] Dr. Mika Kemppainen, Dr. Iikka Virkkunen, Dr. Henner Ostermeyer, Dr. Markus Gribi 96 the fusion line. Or, for example, a skew of the flaw has been specified in geometry, where no skew can exist for a natural crack. So, as with all new technologies, though there is a lot of potential, one should be aware that in applying the new technology to new cases, different aspects showing inspection technique related limitations and a respective need for improvement in applying the technology may arise. Such things are part of the learning curve of applying the new technology and they also show that there is a need to change part of the current way of specifying defects used. 1.1 Current state of the art The controllability of the flaw manufacturing technology is one crucial aspect when selecting flaw pro duction technology for NDE qualification purposes. This is due to the fact that only by controlled manufacturing can one get repeatedly known flaw sizes and reproduce essential flaw characteristics. The most controllable of the flaw manufacturing techniques is the electro discharge machining. However, it does not reproduce any realistic characteristics for the flaws. For the other techniques, representativeness is better, but a variation can be seen in the production controllability from bad to good. Representativeness and controllability of different flaw manufacturing techniques are shown in the following figure (Fig. 2). Production of EDM flaws may meet limitations in the specimen geometry and weight due to challenges in handling and access of the intended location. Furthermore, all the flaws produced by EDM are surface breaking. With this technique, flaws can be produced in most of the metallic materials. The welding-based procedures have the limitation that they induce extra weld material in the specimen. Fur- thermore, welding-based flaw manufacturing, if applied to ready-made specimens, is done on purpose-machined grooves, or it has to be done during the manufacturing of the specimen. Hence, these techniques cannot be applied to ready-made specimens without modifying the specimen with an additional weld. By welding-based techniques, flaws can be produced in all the materials that can be welded. There has been considerable development on grown crack production techniques relying on stress corrosion cracking. Developed stress corrosion cracking techniques can be applied to limited size and shape of specimens due to the fact that an external loading is needed to create the stresses. Such external stresses are not to induce full-scale mock-ups with complex geometry. Furthermore, due to the nature of the damage mechanism itself, control of the flaw sizes is quite rough. All the cracks grown by stress corrosion techniques are surface-breaking. This technique is limited to materials where stress corrosion damage mechanisms can occur. Fig. 2: Representativeness vs. controllability of different flaw types SCC = Stress Corrosion Cracking CIP = Cold Isostatic Pressing of a notch Development and production of test specimens to evaluate an inspection issue 97 The nature of the thermal fatigue crack growth technology allows surface-breaking cracks to be produced. Size of produced natural cracks can be controlled accurately. Furthermore, the location where the crack is to be manufactured must be attainable (i. e. the loading tool must fit the location). This prevents crack production, for example, to the inner diameter (ID) of very small tubes. Currently the smallest tube ID where cracks have been produced is about 16 mm. While this technique is applicable to a wide variety of materials, there are also some materials that present a challenge. Currently cracks cannot be manufactured, e. g., to copper and aluminium. The use of grown cracks based on the thermal fatigue production process has increased markedly during the last few years. Furthermore, the amount of different applications has become larger thus covering today most of the NDE inspection techniques and targets in the nuclear field. 1.2 ENIQ methodology Today, rules and requirements for the inspection qualification are introduced increasingly in the nuclear industry in different countries. Very rapid development in the requirements for the inspection qualification can be seen in many countries. Though they still vary from country to country, an effort to harmonize different require- ments has been made. One common guideline is the set of recommended practices published by the European Network for Inspection Qualification (ENIQ) [9] – [18] giving general guidelines for inspection qualification. ENIQ methodology is, to a certain degree, followed by most European countries. The starting point of the ENIQ methodology is the “input information” dossier that defines the inspection case and includes information about component and flaw types expected under service conditions. Input information dos- sier is typically prepared by the plant operator having the best information on the pertinent damage mechanisms. Crack growth mechanisms as well as critical flaw sizes are defined in the input information. The inspection proce- dure is defined based on this input information, usually by the inspection vendor. When these two are available, a technical justification is prepared. This document takes the relevant data from input information and inspection procedure and defines the most important parameters for successful inspection. With use of the technical justifi- cation, test blocks can focus on testing the most important challenges of the inspection. Different aspects of the practical trials are accomplished by the technical justification. However, the technical justification dossier has also some problems related to test block manufacturing and defect selection. The problem is that the guidance to test block manufacturing comes only in the end of the technical justification. Consequently, the defect specifications tend to be dominated by inspection considerations and information relating to crack characteristics to be expected in real life is not well preserved. In fact, the ENIQ methodology does not currently facilitate selection of defects very well. Also, when the test blocks and defects are tightly coupled with the technical justification, they become specific to the qualification at hand. This is more widely discussed in the reference [19]. ENIQ methodology mentions use of test blocks with defects in three phases: laboratory samples, open trials and blind tests. The purpose of the laboratory samples is to give background information and supporting evidence for the technical justification. The purpose of the open trials is to show that the technique is able to achieve the performance defined in the technical justification. The purpose of the blind trials is to demonstrate that the personnel are able to correctly apply the technique and judge its results. To fulfil these purposes, the defects in all test blocks should give a representative response (in terms of essential characteristics) as compared to the defects defined in the input information. 1.2.1 Current way of selecting different flaw types (under ENIQ methodology) In a national example of ENIQ-based methodology followed in Finland for the qualification, the justification of the used flaws relies heavily on the experience of the involved parties. There is no clear set of requirements for flaws. Instead, the applicability of defect types for each case is defined in discussion between the qualification body and the operator, that is, requirements are redetermined for each case. When limited experience is available, defects chosen may be rejected during fingerprinting, for example, based on unrealistic response or unacceptable disturbances. Dr. Mika Kemppainen, Dr. Iikka Virkkunen, Dr. Henner Ostermeyer, Dr. Markus Gribi 98 However, to clarify the process of using mock-ups with representative flaws, there should be clear requirements set for the flaw characteristics in the ENIQ-recommended practises, as well as in the authority requirements. Such requirements should be for the essential flaw characteristics as, for example, crack opening, crack tip radius, fracture surface roughness, amount of turns and flaw tilt angle. The most important for the range of values of different characteristics is that it should be based on real measured values from service-induced flaws. Quantified requirements would induce measurement of essential characteristics of each flaw type used, together with statistical analysis on the reproducibility and manufacturing tolerance of the characteristics in question. 2. Application case examples Two different, real-world application cases are presented. These show the benefits obtained by using advanced realistic crack types. In both cases, an earlier qualified and accepted inspection technique was applied with an assumption that it should give reliable results. However, in both cases an indication obtained from the actual inspection, had left unclear the origin of the indication. Therefore, additional studies of the inspection issue had to be performed with realistic natural cracks produced in the actual specimens. To solve the unclear NDE indication obtained in both cases a post-inspection open trial was done with realistic grown cracks. These were produced in real specimens with the material and geometry similar to the actual inspection target. 2.1 E.ON reactor pressure vessel nozzle inspection issue In the nuclear power plant of E.ON in Brokdorf, Germany, non-destructive inspection of dissimilar welds is an important part of the inspection program in refuelling outages in nuclear power plants. The inspection of the in- ner weld surface in the reactor pressure vessel head nozzles of German PWR plants is complicated by geometrical constriction. This dissimilar metal weld is accessible only through a 1 mm thick gap, which the eddy current probe must pass through. For this inspection a new eddy current technique had to be developed. Due to the geometrical limitations, the probe design had to ensure an extremely flat probe. The qualification of the inspection technique was performed with a test specimen made of a real nozzle using EDM notches as simulation of cracks according to applicable rules. Qualification was accepted before applying the inspection technique for inspections during the outage of the plant in 2007. During the ISI in 2007, an indication was found close to the austenitic side of the dissimilar weld in one nozzle. The signal was not within the phase range of defects detected in the qualification and the signature was totally different from the signal of notches. The circumferential extent was small in respect to the length of the weld. So, the indication was not classified as a defect signal, nor was it a clear geometrical indication. It was decided to make further investigation to find out the reason of the signal. One of the points to study in this investigation was to find out the difference between notch signals and the signals of real cracks. The next aim was to develop a visual technique able to inspect the inner weld surface through the 1 mm gap. A test specimen was made using an original nozzle made of vintage material. Due to need of realistic cracks and existing sample, in-situ grown cracks had to be used. Trueflaw was ordered to manufacture cracks in this new specimen as well as to make different EDM notches and notch fields as references. E.ON supplied an original nozzle to Trueflaw to be used as a test block. Part of the test block area was marked for validation. Trueflaw produced validation cracks of intended size to this area. The area containing the validation cracks was then cut out from the tube using EDM and the cracks destructively examined to reveal the true crack depth. As the Trueflaw technology was used by E.ON for the first time, E.ON and a consultant expert of the authority (TÜV) visited Trueflaw to follow the progress. Subsequent to the accepted validation result, the final cracks were manufactured and the sample supplied to E.ON. Development and production of test specimens to evaluate an inspection issue 99 There was a requirement for the crack to be a natural one with realistic opening profile. Due to this requirement, in addition to the normal crack validation, a cross-sectional sample was made to measure the crack opening profile in the depth direction. The next figure (Fig. 3) shows a picture of the cross-sectional sample together with the opening profile measured. With the manufactured cracks, the eddy current system qualification was repeated, and the phase range for defects could basically be verified with the signal being reduced at the edges. With a crack having a secondary crack close to it, it could be proved that no phase shift occurs, when more than one crack is in the area of influence of the probe. The new developed visual inspection technique (using special optical components and a CCD-chip together with an optical fibre lighting) was qualified as well. The applicability of the very small high resolution video probe (to be used in the gap of around 1 mm width) for the detection of cracks even from problematic view angles could be verified clearly with the natural cracks delivered by Trueflaw. In the 2008 outage a second inspection of the reactor head penetration with the optimized qualification of the eddy current inspection and the visual inspection was made. The signal was found unchanged. It could be con- firmed that the reason for the indication was of geometrical nature. In conclusion, a crack in the component could be excluded. Furthermore, inspection processes were requalified and accepted for further use. With the real cracks used, a complex problem that otherwise would have led to extended discussions and technical justifications with, for example, the need of disassembling the control rod drive mechanism, was solved. Fig. 3: Crack opening profile in depth direction together with the measured opening profile values Dr. Mika Kemppainen, Dr. Iikka Virkkunen, Dr. Henner Ostermeyer, Dr. Markus Gribi 100 2.2 Axpo CRDH inspection issue In the nuclear power plants operated by the Axpo an EC inspection technique is used to inspect dissimilar metal weld in the reactor pressure vessel upper head penetrations. Axpo had successfully qualified the inspection tech- nique before the planned outage of the year 2008. Qualification has been carried out by samples with EDM notches. However, during ISI of the head penetrations in the outage 2008, problems arose due to unanticipated geometry of the target area. The ID contour, contrary to the expectation, was not smooth. It had a small counterbore, that prohibited the eddy current testing (ET) probe to have constant contact. As a result, a disturbing indication along the circumference was obtained. This indication, caused by the poor coupling of the eddy current probe, could mask possible real indications. The regulator did not accept the inspection results, and required the plant to optimize the inspection technique to overcome the geometric restriction present. Consequently, when the optimization of the technique has been done, both the qualification and the performed in-service inspections have to be redone. The upper head shall be re- placed in about four years, so the developed system will have a short useful life and the effort planned accordingly. So, it was decided first to apply a qualified moulding technique on three penetrations to determine the ID contour. This investigation showed that there is an inner diameter difference between 1.5 and 2 mm and that the counter- bores have different steepness. The next figure (Fig. 4) shows a drawing of the geometry based on the moulding result. The next step was to manufacture a sample representing the geometry of the real penetrations. Realistic flaws were needed to achieve reliable inspection results. The sample had to contain realistic cracks in different positions in the weld area. Furthermore, as the sample was already finished, the applied production technology had to deal with this. The plant decided to use Trueflaw cracks based on the criteria that realistic cracks were needed and they had to be put in a ready-made sample. Two different crack depths were specified: 1 mm and 3 mm. The 1 mm depth was for the ET inspection and 3 mm for the additional UT sizing technique. As the moulding technique was also used for the in-service inspections, this sample was to be used as a verification test of the technique’s capability to show that there are no cracks in the penetrations inspected by the moulding technique. The next figures (Fig. 5 to 8) show examples of penetrant testing (PT) done (by Trueflaw) and the corresponding moulding inspection result. Moulding inspection results are photographs of the replica taken from the sample. Fig. 4: Example of a penetration ID contour Development and production of test specimens to evaluate an inspection issue 101 In the near future the sample shall be inspected by the ISI technique used in 2008. The plant expects that it will be possible to show that cracks would have been detected even with the non-optimized inspection technique and that the already performed inspections of year 2008 can be accepted. And the authority questioned performance of the inspection techniques would be confirmed. Thus the real cracks used in an additional open development sample could solve a complex problem that otherwise would have led to extensive rework with a part that will soon be replaced. However, if a complete new qualification should be required, further samples with Trueflaw cracks will be used. Fig. 5 and 6: PT indication with corresponding moulding indication (with secondary crack) Fig. 7 and 8: PT indication with corresponding moulding indication Dr. Mika Kemppainen, Dr. Iikka Virkkunen, Dr. Henner Ostermeyer, Dr. Markus Gribi 102 3. Conclusions Assessing NDE reliability through representative practical trials remains quite a challenging task. Over the years, the techniques to produce representative flawed samples have been developed. This has led to increasing repre- sentativeness in test specimens. In recent years the most significant trend has been the increased use of “grown cracks” which currently offer the best representativeness available. The available techniques are not currently used to their full potential. Consequently, additional guidance and experience is needed to improve practical trials. In particular, clear guidelines for justifying used defects and requirements for defects are needed in the ENIQ. Also, sharing practical experience between users, like the two cases presented here, is important to promote efficient use of different flaw types. The current experience of using modern grown cracks ranges from full qualification cases with both open and blind trials, NDE development for new inspection technologies, reference and calibration samples, and samples for training of inspectors. Grown cracks have been used for various NDE techniques covering ultrasonic, eddy current, dye penetrant, magnetic particle, visual and radiographic testing techniques (both film-based and digital radiography). Also grown cracks have proved valuable in developing emerging techniques such as thermosonic and x-ray diffraction inspection techniques. In conclusion, the use of the more advanced defect manufacturing techniques gives inspectors experience on real cracks and allows tuning the procedure to find and size real cracks. The test blocks are more generic and can be used for various NDT methods and procedures, thus avoiding rework when methods or requirements change or new information becomes available. 4. References [1] Kemppainen, M.: Realistic artificial flaws for NDE qualification – novel manufacturing method based on thermal fatigue, Dissertation for the degree of Doctor of Science in Technology, Espoo, Finland, 2006. (Available online from: http://lib.tkk.fi/Diss/2006/isbn9512282631/) [2] Virkkunen, I., Kemppainen, M., Ostermeyer, H., Paussu, R., Dunhill, T.: “Grown cracks for NDT development and qualification”, InSight, 2009 5, to be published. [3] Paussu, R., Virkkunen, I., Kemppainen, M.: “Utility aspect of applicability of different flaw types for qual- ification test pieces”, Proceedings of the 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8th – Oct 10th, Budapest, Hungary, 2007, p. 85 – 91. [4] Packalén, T., Sillanpää, J., Kemppainen, M., Virkkunen, I., and Paussu, R.: “The influence of the crack opening in the UT inspection qualification”, Proceedings of the 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8th – Oct 10th, Budapest, Hungary, 2007, p. 463 – 470. [5] Kemppainen, M., Virkkunen, I., Packalén, T., Sillanpää , J., Paussu, R.: “Importance of crack opening in UT inspection qualification”, Proceedings of the 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8th – Oct 10th, Budapest, Hungary, 2007, p. 93 – 105. [6] Wåle, J., Ekström, P.: Crack Characterisation for In-Service Inspection Planning. SKI-project 14.4-940389, SAQ/FoU report 95/07, SAQ Kontroll Ab, Stockholm, Sverige, 1995. [7] Wåle, J.: Crack Characterisation for In-Service Inspection Planning – An update. SKI reference 14.43-200543105, ISRN SKI-R-06/24-SE, SKI, Stockholm, Sweden, 2006. Development and production of test specimens to evaluate an inspection issue 103 [8] Hänninen, H., Hakala, J.: “Pipe Failure Caused by Thermal Loading in BWR Water Conditions”, International Journal of Pressure Vessels and Piping, Vol 9, pp. 445 – 455. [9] Chapman, B., Seldis, T., Eriksson, A. (Eds.): European Methodology For Qualification Of Non-Destructive Testing, Third Issue. ENIQ report nr. 31, EUR22906EN, Luxemburg, Office for Official Publications of the European Communities, 2007. ISSN 1018-5593. [10] Eriksson, A., Whittle, J. (Eds.): ENIQ Recommended Practice 1, Influential / Essential Parameters, Issue 2. ENIQ report nr. 24, EUR21751, Luxemburg, Office for Official Publications of the European Communities, 2005. [11] Lemaitre, P. (Ed.): ENIQ Recommended Practice 2, Recommended Contents for a Technical Justification. EUR 18099 EN, Luxemburg, Office for Official Publications of the European Communities, 1998. [12] Lemaitre, P. (Ed.): ENIQ Recommended Practice 3, Strategy Document for Technical Justification. EUR 18100 EN, Luxemburg, Office for Official Publications of the European Communities, 1998. [13] Lemaitre, P. (Ed.): ENIQ Recommended Practice 4, Recommended Contents for the Qualification Dossier. EUR 18685 EN, Luxemburg, Office for Official Publications of the European Communities, 1999. [14] Lemaitre, P. (Ed.): ENIQ Recommended Practice 5, Guidelines for the Design of Test Pieces and Conduct of Test Piece Trials. EUR 18686 EN, Luxemburg, Office for Official Publications of the European Communities, 1999. [15] Lemaitre, P. (Ed.): ENIQ Recommended Practice 6, The use of modelling in inspection qualification. EUR 19017 EN, Luxemburg, Office for Official Publications of the European Communities, 1999. [16] European commission: ENIQ Recommended Practice 7, Recommended General Requirements for a Body Operating Qualification Of Non-Destructive Tests. EUR 20395 EN, Luxemburg, Office for Official Publications of the European Communities, 2002. [17] Chapman, R., Chapman, V., Eriksson, A., Simola, K., Waites, C., Whittle, J. (Eds): ENIQ Recommended Practice 8, Qualification Levels and Qualification approaches, Issue 1. EUR 21761 EN, Luxemburg, Office for Official Publications of the European Communities, 2005. [18] Cueto-Felgueroso, C., Simola, K., Gandossi, L.: ENIQ Recommended Practice 9, Verification and Validation of Structural Reliability Models And Associated Software To Be Used In Risk-Informed In-Service Inspection Programmes. EUR 22228 EN, Luxemburg, Office for Official Publications of the European Communities, 2007. [19] Virkkunen, I., Kemppainen, M., Koskinen, A.: “Recent Advances in Artificial Cracks for NDT Development and Qualification”, Proceedings of the 7th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, May 12th – Oct 14th, Yokohama, Japan, 2009, to be published.

ADVANCES IN PRODUCTION OF REALISTIC CRACKS TO NDT DEVELOPMENT AND QUALIFICATION PURPOSES OF STEAM GENERATOR TUBES I. Virkkunen1, M. Kemppainen1, J. M. Tchilian2 and J. Martens3 1Trueflaw Ltd., Espoo, Finland, Email: iikka.virkkunen@trueflaw.com 2AREVA Nuclear Power Plant sector, Paris, France 3AREVA NP Intercontrole, Paris, France Abstract Realistic defects are needed for steam generator tube inspections when developing new NDT methods or assessing the performance and reliability of methods and procedures used. Furthermore, realistic defects give the most reliable results in assessing service-related reliability of steam generator tubes by, for example, burst or leak tests. It is crucial to have representative defects as the defect characteristcs has marked effect on the results both in NDE, burst and leak tests. Representativeness should be to the actual service-induced defects, and the evaluation should be based on the essential defect characteristics. In this paper realworld application cases are presented about crack production to steam generator tubes. Crack production technique used is based on controlled thermal fatigue process creating natural cracks. Such cracks have been produced in Alloy 690 and austenitic stainless steel steam generator tubes. These cracks have been used, for example, for advanced NDT qualification purposes of a new build nuclear power plant. Paper presents results of the destructive tests perfomed after validation tests of the crack manufacturing in the Alloy 690 and austenitic stainless steel. These results are shown for both of the materials with measured essential crack characteristics. In addition to metallographic analysis, the paper presents the results of performed NDT inspections for the Alloy 690. Results have been obtained with an advanced inspection technique developed and used for today's inspections of steam generator tubes in nuclear power plants. 1. Introduction Nuclear power plant steam generator tubes have been found susceptible to cracking in various conditions. These tubes are an integral part of the reactor coolant pressure boundary. The steam generator tubes are also relied upon to isolate the radioactive fission products in the primary coolant from the secondary system. Primary coolant can leak through cracks to secondary coolant and ultimately, severe cracking can lead to tube rupture. On some reactor types, this can lead to release of radioactivity to the environment outside the reactor containment through the pressure relief valves in the secondary system [1]. Consequently, cracking in stream generator tubes is of significant concern and can lead to leaks and ultimately compromise the safe operation of the plant. [1,2]. The mitigation of steam generator degradation can be classified to three stages. Firstly, effective in-service inspection is used to detect cracking. When cracking is detected, the tube in question can be plugged to prevent further damage. The detection and characterization of steam generator damage, however, remains challenging and better methods are continuously developed to improve inspection (e.g., [3]). Secondly, if a through wall crack is present, the leak rate expected from such a crack must be known in order to assess the safety-significance of possible leaks (see, e.g. [4]). Finally, the limits of structural integrity of a cracked tube must be known in order to properly set inspection goals and risk estimates (see, e.g., [5]). To this end, the cracked tubes are often modelled and experimental rupture-tests performed ([6]) to confirm the expected behaviour. In summary, the expected behaviour of degraded (i.e. cracked) tubes must be estimated and investigated. That is, for all the mitigation efforts mentioned above, it is necessary to understand and estimate the behaviour of cracked steam generator tubes in service conditions. Experimental studies are necessary to validate the mitigation techniques used. For such experimental studies, e.g., to improve, and qualify NDE methods, samples with cracks representative of the expected service-induced cracking are needed. However, manufacturing of such cracked samples has been a major challenge. In order to get representative experimental evidence, essential crack characteristics must be well represented in the test samples; e.g. crack opening and tortuosity, crack surface roughness and crack tip radius for NDE and leak studies and crack tip stress intensity for integrity assessment. In practice, this necessitates the use of test specimens with controlled natural cracks. This paper describes a new technology to produce cracked test samples that can be used to gain information on the behaviour of cracked steam generator tubing. So far, the technology has been used mainly for NDE training and qualification. This paper present some early application cases of the production technology for steam generator tube use. 2. Trueflaw Technology Trueflaw produces defects using controlled natural thermal fatigue damage process. The defects are grown in much the same way as could occur during service conditions. However, the growth is accelerated to make production times practical and controlled to enable predetermined flaw parameters. Flaw production is done in-situ to ready-made sample. Cyclic thermal fatigue loading is induced locally by alternating heating and water spray cooling, as described by Kemppainen [7]. Loading is based on pure thermal loading and there is no welding, machining, or mechanical treatment applied. No artificial initiators of any kind are used and the material microstructure is not disturbed in the process. More detailed information on the properties and use of produced cracks has been presented earlier [8-11]. 2.1 Characteristics of the growth procedure Manufacturing of real cracks has traditionally been restricted to simple component shapes and small components. Providing the stress for crack growth mechanically becomes impractical when the sample geometry becomes more complicated. Also, accurate control of induced stress in complex shapes during crack growth is very difficult and it is not possible to limit the stress only to desired areas where defects are needed. In contrast, thermal loading can be applied to local areas in complicate sample geometries. Since only a limited area is stressed at any given time, the needed equipment is relatively light. Furthermore, the ability to locate and control the stressed area enables accurate control over crack growth location and crack parameters. 2.2 Confirming essential flaw parameters without supplementary NDE inspections Often, the dimensions and other characteristics of the cracks must be known. Some of the parameters, like surface length, can be readily measured. However, for many important defect parameters, depth in particular, this is not possible from the final sample. In Trueflaw technology, these parameters are known by destructive validation. When there is no experience on he specified crack, the intended crack is first produced to a representative validation sample. This sample needs to have similar material and similar local geometry, but can be simplified and smaller compared to the actual test sample. This validation crack is destructively examined to reveal the true crack depth and other specified parameters (e.g, crack opening, surface roughness etc.). Then, using the same procedure, a similar crack can be produced to any number of actual test samples repeatably. The Trueflaw technology has been available on the market since 2001. During that time, the technology has been developed immensely. It has been applied to numerous different materials and component geomeries. It has been used to solve various problems requiring information about response of cracked components in nuclear and conventional energy production, as well as in aerospace and railroad industries. It has been tried and tested. Most of the applications to date have been on inspection training and qualification samples. There has also been several application cases for using such cracks in developing existing and new NDE techniques (see, e.g., [11]). 3. Application Cases Most of the application cases for Trueflaw cracks to date have been related to inspection training and qualification samples. The use of the technology for steam generator tube cracking is still quite uncommon. Below, two example cases are presented, which show some of the potential uses of this technology to steam generator tubes. The first case details inspection qualification on steam generator tubes manufactured from Inconel 690. The second case shows crack production validation for a more traditional stainless steel tube material. 3.1 Crack production in Inconel 690 steam generator tubes Inconel 690 steam generator tubes are used in the modern EPR reactor type that is currently under construction in Olkiluoto (later "OL3"), Finland. This material has been selected due to its high resistance to stress corrosion cracking that has caused problems in older reactor types (made from stainless steel or Inconel 600). According to Finnish regulations, all in-service inspections to be performed to the new power plant must be qualified. In qualification, the performance of the inspection technique to be applied is shown on realistic defects. Hence, also the SG-tubes were qualified according to Finnish regulations. The qualification done for OL-3 was the first to be completed on Inconel 690 tubes with real cracks. Consequently, no cracked specimens were available nor had such specimens ever had been manufactured. Areva NP approached Trueflaw to manufacture the cracked samples. Figure 1. shows drawing of the tube sample used in qualification. Figure 1 Tube sample drawing. The material is Inconel 690. 3.1.1 Development and validation of crack production to Inconel 690 steam generator tubes Crack production to Inconel 690 steam generator tubes posed several challenges. Firstly, due to the thin wall section and correspondingly small crack depth targets, the tolerances for crack production were set very tight. Hence, the first target was to obtain successful small crack initiation to this new material with the thin wall section. After several trials, successful circumferential cracking was obtained (Figure 2). However, rather unexpected difficulties were met when developing technique for axial cracking. As it turned out, the material was highly anisotropic with respect to cracking. Presumably this is caused by the microstructural texture resulting from tube manufacturing. It was noticed, that the samples strongly preferred circumferential cracking over axial cracking. Consequently, no axial cracking could be produced without accompanying circumferential cracking. Figure 2 Small crack initiation in Alloy 690 steam generator tube; a dye penetrant indication. The specified cracking included long and shallow cracks, i.e. the cracking extended up to several millimeters around the circumference but only a fraction of a millimeter in depth direction. Consequently, the production was met with a problem of through wall cracking. That is, successful crack initiation was obtained, but as soon as the surface length of the crack was extended, cracks rapidly grew through-wall. To overcome this problem, it was decided to produce the desired dimensions with non-contiguous cracking. Figure 3 shows such a non- contiguous crack pattern produced. Figure 3 Circumferential cracking in steam generator tube samples. The validation samples were then sent to Intercontrole for inspection (see paragraph 3.1.2). The purpose of these inspections was to confirm that the produced samples are suitable for the intended use and to give the inspectors early experience on the cracked samples. Finally, the validation samples were destructively examined by Trueflaw in order to confirm the depth of the cracks. Figure 4. shows an image of destructive examination of a validation sample. Figure 4 Destructive examination of produced validation cracks in the SG tube sample. 3.1.2 Inspection results Three samples were inspected with eddy current inspection by Intercontrole. The samples are detailed in Table 1. Inspection was carried out with two different probes: bobbin probe and rotating probe (Figure 5). Table 1 Samples in eddy current inspection. All cracks were circumferential. The depth values for samples 1 and 2 are measured from destructive examination of sample 1. Other values are directly measured on the sample. # Cracks Note 1 OD crack, 20% depth, 100° angular extension destructively examined 2 OD crack, 20% depth, 140° angular extension 3 through-wall crack, 48° angular extension Figure 5 The probes used in the inspection were ET Bobbin probe, Zetec 610UL (upper image) and ET rotating probe, Zetec 3 coil 610PP with 1 +point coil PP11A, 1 pancake coil 115MR Ø 3 mm (medium range) and 1 pancake coil 080 HF Ø 2 mm (High frequency) (lower image). The EC response of cracks in each sample with two different probes can be seen in Figures 6-14. Three different pictures are shown for each sample: first the result of bobbin probe then results of rotating probe signals and finally topographic picture of the rotating probe response. The crack signal is marked to each image. The ET bobbin probe clearly detects single indication for all the inspected samples. For samples 1 and 2 the indication is correctly interpreted as OD defect and for sample 3 the indication is correctly interpreted as through-wall defect. The results for the rotating probe are summarized in Table 2. The sizing results from the rotating probe agree very well with the true crack sizes. For the sample 1, where direct destructive results are available, the depth sizing is within 5%-units of the true depth, which corresponds to about 0.05 mm. For the sample 2, which was produced similarly to sample 1, the inspection results are within expected crack depth variance ±0.3 mm. Furthermore, the EC inspection indicated that the thermal loading had caused slight local denting of the tube. However, this did not disturb the inspection. Table 2 Results for the rotating probe inspection. Sample Amplitude (V) Phase (OD plane, °) Depth (%) Angular extension (°) 1 0,17 87 15 115 2 0,26 79 45 140 3 19,07 35 100 140 Figure 6 ET probe results for sample 1 for the bobbin probe. Figure 7 ET probe results for sample 1 for the rotating probe. Figure 8 ET probe results for sample 1 for the rotating probe. Figure 9 ET probe results for sample 2 for the bobbin probe. Figure 10 ET probe results for sample 2 for the rotating probe. Figure 11 ET probe results for sample 2 for the rotating probe. Figure 12 ET probe results for sample 3 for the bobbin probe. Figure 13 ET probe results for sample 3 for the rotating probe. Figure 14 ET probe results for sample 3 for the rotating probe. 3.1.3 Conclusions Cracks were successfully produced to Inconel 690 samples in circumferential direction. Samples strongly preferred circumferential crack growth over axial crack growth. The cracked samples showed that the specified crack sizes can be reliably detected and characterized with the used probe design, even with the complex non-contiguous cracking morphology. The inspection method was successfully qualified for OL3. 3.2 Crack production in stainless steel steam generator tubes The Research Institute of Nuclear Power Operation (RINPO), China, approached Trueflaw in late 2008 for possibility to produce cracked steam generator tube samples to be used for burst tests to qualify the extreme load of defective tube. The tube in question is austenitic stainless steel tube, typical for WWER reactors. It was decided to do production trials to investigate the possibility and suitability of Trueflaw cracks in this case. For this use RINPO send Trueflaw sample tube to be used in the production trials. The interesting crack type was axial cracking. 3.2.1 Production trials Trueflaw completed altogether nine production tests for this tube. With these tests, it was shown that axial cracking can be repeatably initiated to austenitic stainless steel tubes and size controlled to produce axial cracks with known size. During the limited testing, however, the method was not conclusively validated. The follow-up for this work is currently considered. Typical initiation images are shown in Figure 15. Produced cracks are very tight and thus PT image show very poor contrast. Corresponding images from destructive examination are shown in Figure 16. It can be seen, that the crack is continuous and follows the long-but-shallow aspect ratio specified for these cracks. Figure 15 Typical initiation images for axial cracks in austenitic stainless steel steam generator tubes. Figure 16 Destructive examination results for azial cracks in austenitic stainless steel steam generator tubes. 3.2.2 Conclusions Cracks were successfully produced to stainless steel tube samples in both axial and circumferential directions. It was shown that Trueflaw technology is well applicable to the intended use. 4. Conclusions Cracked steam generator tube samples can be used in various applications where the real response of cracked tubes needs to be known. These include NDE qualification, validation of safety calculations in burst tests and leak rate tests. The use of cracked sample is still quite uncommon. However, the technology for the more widespread use of cracked samples is now available and some successful early experience has already been obtained. 5. Acknowledgements The authors wish to gratefully acknowledge the possibility to publish the application cases presented here courtesy of Areva NP and the Research Institute of Nuclear Power Operation (RINPO). 6. References 1 Anon. (Ed.) "Assessment and management of ageing of major nuclear power plant components important to safety: Steam generators". IAEA-TECDOC -981. International Atomic Energy Agency, Austria, 1997. 2 Anon. (Ed.) "Strategy for Assessment of WWER Steam Generator Tube Integrity". IAEA- TECDOC-1577. International Atomic Energy Agency, Austria, 2007. 3 Kim, D., Eom, H. "Ultrasonic NDE discrimination with the gradient descent algorithm and SAFT image processing", NDT&E International, 42, 2009, p. 250-250. 4 Taljat, B., Kosel, F. "Sensitivity and response surface analysis of cracked tube parameters in elastoplastic domain", Int. J. Pres. Ves. & Piping, 62, 1995, p. 161-166. 5 Cizelj, L., Mavko, B. "On the risk-based steam generator lifetime optimisation", Theorethical and Applied Fracture mechanics, 23, 1995, p. 129-137. 6 Hwang, S., Namgung, C., Jung, M., Kim, H, Kim, J. "Rupture pressure of wear degraded alloy 600 steam generator tubings", Journal of Nuclear Materials, 373, 2008, p. 71-74. 7 Kemppainen, M., Realistic artificial flaws for NDE qualification – novel manufacturing method based on thermal fatigue, Dissertation for the degree of Doctor of Science in Technology, Espoo, Finland, 2006. (Available online from: http://lib.tkk.fi/Diss/2006/isbn9512282631/) 8 Paussu, R., Virkkunen, I., Kemppainen, M., "Utility aspect of applicability of different flaw types for qualification test pieces", Proceedings of the 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8 th – Oct 10th. Budapest, Hungary. 2007. p. 85-91. 9 Packalén, T., Sillanpää, J., Kemppainen, M., Virkkunen, I. and Paussu, R., "The influence of the crack opening in the UT inspection qualification", Proceedings of the 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8th – Oct 10th. Budapest, Hungary. 2007. p. 463-470. 10 Kemppainen, M., Virkkunen, I., Packalén, T., Sillanpää , J., Paussu, R., "Importance of crack opening in UT inspection qualification", Proceedings of the 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8 th – Oct 10th. Budapest, Hungary. 2007. p. 93-105. 11 Virkkunen, I., Kemppainen, M., Ostermeyer, H., Paussu, R., Dunhill, T., "Grown cracks for NDT development and qualification", InSight, 5, 2009.

Inspection Qualification II Recent Advances in Artificial Cracks for NDT Development and Qualification M. Kempainen, I. Virkkunen, Trueflaw Ltd., Finland; A. Koskinen, VTT, Finland ABSTRACT Defects are needed to develop new NDT methods and to assess the performance and reliability of used inspection methods and procedures. It is crucial to have representative defects in order to have an accurate and realistic assessment of the performance of NDT. Representativeness should be to the actual service-induced defects that the NDT method is used to evaluate. While various techniques have been used to create such defects, all conventional techniques seem to have some shortcomings that limit true assessment of the NDT performance. Currently, the used procedures and requirements do not promote efficient use of available defects. This paper describes use and selection of artificial defects for NDE qualification. Finally, real-world application cases are presented showing the use of such cracks. INTRODUCTION Inspection qualification is nowadays quite well established in the nuclear industry. While the requirements vary from country to country, most countries do require inspections to be qualified prior to actual inspection. The European Network for Inspection Qualification (ENIQ) has published a comprehensive set of recommended practices that cover most aspects of inspection qualification [1- 10]. Most European countries follow the ENIQ methodology in inspection qualification to a varying degree. The methodology as well as the national requirements for inspection qualification have developed very rapidly in recent years. One of the key issues in inspection qualification is the production of relevant test blocks to show the performance. Traditionally, there has been severe limitations in manufacturing test blocks with defects. In particular, the techniques available for producing representative defects to test blocks have had limitations. Also, the ENIQ methodology currently gives very limited guidance to this vital subject. Recent advances in artificial cracks In recent years new possibilities for producing real, natural cracks for qualification test blocks have emerged, tried and tested [11,12]. These techniques have developed and matured over the recent years and many of the traditional limitations of defect manufacturing have been overcome. The production methods to produce these grown cracks have been under significant scrutiny and they have now been validated and qualified for use in inspection qualification in various countries. Due to the investigations done on grown cracks, there's now much new information on applicability of different artificial defects [13-15]. Also, various characteristics of both natural cracks [16] and artificial cracks have been studied and documented. Consequently, a sound body of technical information is now available for use in inspection qualification. However, this information is currently not used to its full potential. In fact, the ENIQ methodology does not currently facilitate selection of defects very well. ENIQ methodology The ENIQ methodology gives general guidelines for inspection qualification. Each european country has a set of authority requirements and current practices that define their implementation of the general ENIQ methodology. These implementations vary from country to country. Figure 1 - General ENIQ flow chart [3] The general ENIQ flow chart is shown in Figure 1. In general terms, the starting point of the qualification is the input information dossier. This contains information about the cracks that are expected in the component. The crack growth mechanisms as well as the critical flaw sizes are defined in the input information. The input information is typically prepared by the plant operator, who has best information on the possible damage mechanisms. Based on this input information, the inspection procedure is defined, usually by the inspection vendor. When these two are available, a technical justification is prepared. It has been said, that the technical justification is the most important document in the qualification. This document takes the relevant data from input information and inspection procedure and defines the most important parameters for successful inspection. The applicability and performance of the chosen procedure is then justified using previous experimental evidence, modelling parametric studies etc. Finally, guidance is given for the test blocks to be used for open and blind trials in qualification. With use of the technical justification, the test blocks can focus on testing the most important challenges of the inspection and the amount of needed test blocks and defects can be reduced. According to ENIQ, the amount of defects can be further reduced by using worst case -defects. In this case, the most difficult defects from inspection point of view are defined and tested for in the open and blind trials. It is then argued, that if the inspection method performs well even in the worst case defects, performance in other cases can be expected to be sufficient as well. Clearly, the use of technical justification has some significant advantages. However, the process also has some problems related to test block manufacturing and defect selection. Information about cracks to be found is defined in the input information. However, as shown in the the flow chart (Figure 1.), the guidance to test block manufacturing comes only in the end of the technical justification. This leads to long information chain from the input information to the final defect specification. Consequently, the defect specifications tend to be dominated by inspection considerations and information relating to crack characteristics to be expected in real life is not preserved well. Also, the sequential process described in the flow chart can generate rather long timetables. In particular, the test block manufacturing takes time and it would be beneficial to be able to start it before the entire technical justification is finalized. ENIQ document also suggest this with the "shortcut" path. Furthermore, the worst-case specification may lead to unnatural defect specifications and tuning the inspection method with unlikely or impossible defect types. Finally, when the test blocks and defects are tightly coupled with the technical justification, they become specific to the qualification at hand. If the inspection method is developed in the future or if additional methods are needed to support the primary method, new test blocks are needed. The ENIQ methodology covers most aspects of inspection qualification with great detail. The methodology document and the associated recommended practices cover 236 pages, in total. However, in all the ENIQ documentation, there's one paragraph (22 lines) that gives instructions on the manufacturing of defects. It's safe to say that defect manufacturing hasn't been emphasized in ENIQ work so far. New version of this document is being prepared, which gives more guidance on how to select defects. The ENIQ methodology mentions use of test blocks with defects in three phase: laboratory samples, open trials and blind tests. The purpose of the laboratory samples is to give background information and supporting evidence for the technical justification. These samples are also used to fine-tune the inspection procedure to maximize its performance. The purpose of the open trials is to show, that the technique is able to achieve the performance defined in the technical justification. In open trials, the location and size of the defects are known and thus the NDE performance can readily be assessed by all involved parties. Successful open trials alone are not considered sufficient to demonstrate real-life performance. It is necessary to demonstrate, that the NDE personnel can apply the technique/procedure correctly. Furthermore, most NDE techniques include human judgement, which may vary between inspectors and which may be influenced by the knowledge of the defect types in the open trials. The purpose of the blind trials is to demonstrate, that the personnel is able to correctly apply the technique and judge its results (even when the correct answer is unknown). To fulfil this purpose, the defects in all test blocks should give representative response (in terms of essential parameters) as compared to the defects defined in the input information. The most obvious way to realize this is, of course, to use test blocks with natural grown defects. This has several advantages: the performance of the NDE system is shown with minimal uncertainties and inspectors get experience on true cracks and know what to expect. There's no room for discussions about the validity of the samples. However, producing realistic flaws may not always be possible. So, the next option is to use semi-realistic crack simulations and then provide technical justification on how to address the shortcomings of the defects in the qualifications. This route has the advantage that the semi-realistic cracks are typically easier and less costly to manufacture. However, inferring real world performance becomes more difficult and the inspectors get less experience with true cracks. On the contrary, inspectors get experience, which is not natural and which may lead to unrealistic confidence on the inspectors skills or capability of the technique. Also, the technical justification makes the samples more case specific. Furthermore, there's a risk in re-doing the qualification if the justification later proves invalid. Finally, even the use of clearly unrealistic notches can be justified for some cases. Here, of course, the justification becomes increasingly difficult as there's no direct information about the true performance of the inspection. Finally, if an indication is actually found, the test samples with notches provide no help to explain it and additional information is then required. Figure 2. shows comparison of some of the often used defect types: grown cracks, welded crack simulations and EDM notches as well as true crack from the literature. a) b) c) d) Figure 2 - Grown crack (a), welded crack simulation (b) and EDM notch (c) and true crack (d) from the literature [17] Unfortunately, the ENIQ does not currently give clear guidelines on how to justify use of any of the available defect types or requirements for used defects. Consequently, these requirements and justifications need to be re-determined for each specific case with the experience available. Application of ENIQ The ENIQ methodology provides general guidance to inspection qualification. However, each qualification body using it seems to have slightly different implementation of it. For example, the system used in Finland [18] follows roughly the following flow chart (Figure 3). Figure 3 - Schematic flow chart of Finnish qualification The main difference to ENIQ flow chart (Figure 1) is that the test blocks are defined based on input information and general features of the inspection procedure. The advantage of the described implementation is, that there's more direct link to the input information and the resulting test blocks are more generic. The test blocks can then often be applied for different qualifications and even totally different inspection methods. Also, the more parallel process improves the time needed. The justification of the used flaws relies heavily on the experience of the involved parties. There's no clear set of requirements for flaws. In stead the applicability of defect types for each case are defined in discussion between the qualification body and the operator. When limited experience is available, defects may be rejected during fingerprinting, e.g., based on unrealistic response or unacceptable disturbances. APPLICATION CASE EXAMPLES VTT research sample VTT is currently developing a monitoring system for online monitoring of material degradation in the primary circuit of a NPP. This may be needed, for example, if an indication is found in primary circuit during normal outage and the part can not be replaced or repaired during that outage, online monitoring system will be able to monitor the indication during the next operation period and confirm the safe operation of the NPP. Currently, a pilot monitoring system is designed, constructed and tested as a part of a project called RAKEMON. RAKEMON is a project in a national research programme on nuclear power plant safety 2007 – 2010, which is mainly funded by State Nuclear Waste Management Fund VYR and Technical Research Centre of Finland VTT. The results of this part of the RAKEMON project will also be used as a part of IAEA coordinated research programme on advanced, surveillance, diagnostics and prognostics techniques used for health monitoring of systems, structures, and components in nuclear power plants. To confirm the proper operation of the pilot monitoring system it needs to detect the changes in the indications originally detected in the inspected pipe. With more advanced methods even the changes in material itself could be detected and the remaining lifetime of a component could be estimated. Only a real grown crack will behave like a real crack in a real component and therefore thermal fatigue cracks made by Trueflaw were chosen for these pilot monitoring system tests. Use of EDM notches in this case was rejected since their use would severely limit the reliability and representativeness of the test results. Indications from the EDM notches are very different from the ones that come from the thermal fatigue cracks and also the behaviour at elevated temperatures would differ from the real defects. It is understood that using realistic defects from the start will give an advantage when the pilot monitoring system is finally ready for power plant scale tests. The measurements with pilot system will be started during 2009 and they will last until the end of 2010. The final results will be published as a VTT research report and they will also be used and reported as a part of the IAEA CRP final report. Ringhals Alloy 600 repair Ringhals ordered a test block from Trueflaw to verify the feed water nozzle safe-end inspection. Crack sizing using TOFD was qualified. In this case, the inspection was done from ground pit with depth of 3.9 mm. Inspection was performed from the cracked side. Inspection was performed by WesDyne TRC. It was decided to use Trueflaw cracks, to avoid any welding to the specimen. Trueflaw manufactured three cracks to Inconel 600 mock-up provided by the client. As per normal Trueflaw procedure the case required manufacturing of validation cracks, which were destructively examined to reveal the true crack depth. After production of validation cracks, cracks of known size could be produced. By mutual discussions, it was decided, that Ringhals would do initial qualification with the validation cracks. The NDE results of the initial qualification on the validation cracks showed that the inspection procedure can meet the inspection target and set tolerances (Table 1). Destructive depth Inspection depth 3.2 mm 2.9 mm 2.5 mm 2.3 mm 2.3 mm 2.6 mm Table 1 - Inspection results on TOFD sizing of Inconel 600 cracks NOK qualification projects Nuclear Power plant Beznau (NOK) in Switzwerland has a huge NDT inspection qualification program started last year. During this program, NOK decided to use realistic cracks in most of the cases. The main reason for selecting realistic cracks was to avoid any additional measurements and technical justification to show that the applied inspection techniques would also work on cracks. Also any discussions on the relevance of the qualifications due to unrealistic defects used were avoided. NOK ordered several test samples from Trueflaw to do inspection qualification. In this paper, two cases are presented: baffle bolts of the pressure vessel internals and RPV bolts. Former was for ultrasonic inspections and the latter for eddy current inspections. Totally there has been more than 20 Baffle bolts and four RPV bolts where cracks have been produced. Part of the work is still ongoing, but currently totally about 90 cracks have been produced during these two project cases. Sample geometries are shown in Figure 4 for both cases. Figure 4 - NOK test pieces for the Baffle bolt and RPV qualification For all cases, the project has followed similar process: first technical applicability was evaluated in cooperation with the supplier and NOK. Then set defect sizes were validated with full report of the destructive validation results and validation was accepted by the client. Finally, the actual cracks were produced to the open and blind test samples. The work is still ongoing and the final results are not available yet. There may be more information published after the qualifications have been done and analyses of the procedures performed. CONCLUSIONS In conclusion, the use of the more advanced defect manufacturing techniques is often beneficial. It gives inspectors more experience on real cracks and allows tuning the procedure to find and size real cracks. The test blocks would be more generic and could be used for various NDT methods and procedures. Using grown cracks reduces re-work when methods or requirements change or new information becomes available. Clear guidelines for justifying used defects and requirements for defects are needed in the ENIQ to promote efficient use of different flaw types and efficient use of gained experience. ACKNOWLEDGEMENTS The industrial applications shown in this paper were provided by Claes-Göran Bengtsson (Vattenfall), Markus Gribi (NOK) and Ari Koskinen (VTT). Their cooperation is gratefuly acknowledged. The insight on the practical application of ENIQ in the Finnish qualification system was kindly provided by Tapani Packalén (Inspecta Sertifiointi). He's help is gratefully acknowledged. REFERENCES 1) Chapman, B., Seldis, T., Eriksson, A. (Eds.) European Methodology For Qualification Of Non- Destructive Testing, Third Issue. ENIQ report nr. 31, EUR22906EN, Luxemburg, Office for Official Publications of the European Communities, 2007. ISSN 1018-5593 2) Eriksson, A., Whittle, J. (Eds.) ENIQ Recommended Practice 1, Influential / Essential Parameters, Issue 2. ENIQ report nr. 24, EUR21751, Luxemburg, Office for Official Publications of the European Communities, 2005 3) Lemaitre, P. (Ed.) ENIQ Recommended Practice 2, Recommended Contents for a Technical Justification. EUR 18099 EN, Luxemburg, Office for Official Publications of the European Communities, 1998 4) Lemaitre, P. (Ed.) ENIQ Recommended Practice 3, Strategy Document for Technical Justification. EUR 18100 EN, Luxemburg, Office for Official Publications of the European Communities, 1998 5) Lemaitre, P. (Ed.) ENIQ Recommended Practice 4, Recommended Contents for the Qualification Dossier. EUR 18685 EN, Luxemburg, Office for Official Publications of the European Communities, 1999 6) Lemaitre, P. (Ed.) ENIQ Recommended Practice 5, Guidelines for the Design of Test Pieces and Conduct of Test Piece Trials. EUR 18686 EN, Luxemburg, Office for Official Publications of the European Communities, 1999 7) Lemaitre, P. (Ed.) ENIQ Recommended Practice 6, The use of modelling in inspection qualification. EUR 19017 EN, Luxemburg, Office for Official Publications of the European Communities, 1999 8) ENIQ Recommended Practice 7, Recommended General Requirements for a Body Operating Qualification Of Non-Destructive Tests. EUR 20395 EN, Luxemburg, Office for Official Publications of the European Communities, 2002 9) Chapman, R., Chapman, V., Eriksson, A., Simola, K., Waites, C., Whittle, J. (Eds) ENIQ Recommended Practice 8, Qualification Levels and Qualification approaches, Issue 1. EUR 21761 EN, Luxemburg, Office for Official Publications of the European Communities, 2005 10) Cueto-Felgueroso, C., Simola, K., Gandossi, L. ENIQ Recommended Practice 9, Verification and Validation of Structural Reliability Models And Associated Software To Be Used In Risk- Informed In-Service Inspection Programmes. EUR 22228 EN, Luxemburg, Office for Official Publications of the European Communities, 2007 11) Kemppainen, M., Realistic artificial flaws for NDE qualification – novel manufacturing method based on thermal fatigue, Dissertation for the degree of Doctor of Science in Technology, Espoo, Finland, 2006. (Available online from: http://lib.tkk.fi/Diss/2006/isbn9512282631/) 12) Virkkunen, I., Kemppainen, M., Ostermeyer, H., Paussu, R., Dunhill, T. " Grown cracks for NDT development and qualification", InSight, 2009 5, to be published 13) Paussu, R., Virkkunen, I., Kemppainen, M., "Utility aspect of applicability of different flaw types for qualification test pieces", 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8 th – Oct 10th. Budapest, Hungary. 2007 14) Packalén, T., Sillanpää, J., Kemppainen, M., Virkkunen, I. and Paussu, R., "The influence of the crack opening in the UT inspection qualification", 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8 th – Oct 10th. Budapest, Hungary. 2007 15) Kemppainen, M., Virkkunen, I., Packalén, T., Sillanpää , J., Paussu, R. "Importance of crack opening in UT inspection qualification", 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8 th – Oct 10th. Budapest, Hungary. 2007 16) Wåle, J., Ekström, P. Crack Characterisation for In-Service Inspection Planning. SKI-project 14.4-940389, SAQ/FoU report 95/07, SAQ Kontroll Ab, Stockholm, Swerige. 1995 17) Hänninen, H. Hakala, J., "Pipe Failure Caused by Thermal Loading in BWR Water Conditions", International Journal of Pressure Vessels and Piping, Vol 9, pp. 445-455 18) Tapani Packalén, private communication

Insight Vol 51 No 5 May 2009 1 TEST BLOCKS Defects are needed to develop new NDT methods and to assess the performance and reliability of used methods and procedures. It is crucial to have representative defects in order to have an accurate and realistic assessment of the performance of NDT. Representativeness should be to the actual service-induced defects that the NDT method is used to evaluate. While various techniques have been used to create such defects, all conventional techniques seem to have some shortcomings that limit true assessment of the NDT performance. This paper describes recent developments of defect manufacturing technology based on controlled thermal fatigue. It is shown that most of the traditional limitations can be overcome using the currently available technology. Finally, three real-world application cases are presented showing the use of such cracks. 1. Introduction The real performance and reliability of used NDT techniques and procedures should be known in order to effectively use them. Without this knowledge, it is difficult to select the correct methods and inspection targets, let alone determine correct inspection intervals. Also, dependable performance information that highlights potential targets for further improvement is necessary for the development of better NDT techniques. To provide this crucial information, various performance demonstration and qualification procedures have been established and are under development(1,2). One of the key challenges in assessing NDT performance is the production of relevant test blocks with which the performance can be tested. These test blocks should contain defects identical to those expected in real inspection, but with known and predetermined location, size and other properties. Producing such controlled natural defects has been, and still is, quite a difficult task. Consequently, a number of defect simulation techniques have been developed, each with their virtues and limitations. As proposed in an ENIQ working document(2), there is essentially four classes of defects currently available (numbering and highlights added): 1. Implanted defects where a pre-existing defect is attached to the testpiece. The attachment usually takes the form of a weld in a machined recess. The technique has the benefits that there is flexibility in the type of defect that can be included and that the insert can be carefully accessed prior to insertion. The main disadvantages are that the insertion process may produce artefacts which either give away the implant’s position or make the inspection response unrealistic in some way. An example of this latter affect is implants into an austenitic weld where the implant material will not form a continuous part of the weld and the attachment welds may significantly influence the performance of the inspection being qualified in an unknown manner. 2. Weld doping or weld modification where for instance crack prone material is added to the weld to promote localised weld cracking. Other examples include introduction of porosity or slag. The main advantage over the previous process is that there are no insert attachment welds. The main disadvantages are that the final size of the defects and their character would need to be confirmed by supplementary inspection. This means that there is a risk of comparing one inspection method with another rather than comparing one inspection against known flaw parameters. Another disadvantage is that the doping process can influence the material properties of the weld in the immediate vicinity of the defect, potentially affecting the inspection in an unpredictable manner. 3. Machined defects where a defect can consist of a cut or machined void. Electro discharge machining (EDM) is perhaps the most relied upon technology in this area where a shaped electrode is used to erode the testpiece. The process is most suitable for production of surface defects, although it is possible to use in combination with welding to produce buried defects. The main advantages of this method are that it tends to be relatively inexpensive, the resulting defect parameters are known to fairly tight tolerances at fabrication and the parent material is left largely unmodified apart from the presence of the machined slot. Disadvantages are that it is difficult or impossible to produce any of the characteristic roughness expected from plant defects and that using standard implantation techniques, the tip radius is likely to be large compared to many crack species. 4. Grown defects where cracking is initiated and propagated into testpieces in much the same way as would occur in plant, simply accelerated to make fabrication times practical. The main processes used for this class of defect are fatigue, thermal fatigue and stress corrosion cracking. There are limitations associated with each growth method, but this option has the advantages of realism and avoidance of attachment welds. The main disadvantage aside from restrictions in the implant process is likely to be reliance upon a supplementary inspection to confirm critical flaw parameters. In the schemes already discussed, this limitation can be minimised by using the defects in parametric specimens and then destructively examining some or all of the specimens once the qualification process is complete. Conventionally, methods 1-3 are applied for performance demonstration and qualification. Most qualifications in effect currently rely on defects produced with these three defect types. As is evident from the above quote, finding suitable defects for performance demonstration is a rather demanding task. Obviously, the problems in using defects that ‘affect the inspection in an unpredictable manner’ make it difficult if not impossible to infer real-world performance from the performance demonstration data acquired using unsuitable defects. This effectively undermines the practical value of the whole exercise. As stated by the ENIQ Grown cracks for NDT development and qualification I Virkkunen, M Kempainen, H Ostermeyer, R Paussu and T Dunhill Iikka Virkkunen and Mika Kempainen are with Trueflaw Ltd, Espoo, FI-02330, Finland. Tel: +35 8456 354415; Email: iikka.virkkunen@ trueflaw.com Henner Ostermeyer is with EON Kernkraft GmbH, Germany. Raimo Paussu is with Fortum Nuclear Services, Finland. Tony Dunhill is with Rolls-Royce plc, UK. working document(2): Some plant defects when inspected with techniques generally used in plant present a very significant challenge to testpiece design and testpiece defect fabrication. Examples are the qualification of ultrasonic inspections of austenitic or Inconel weld metal and inspections for intergranular stress corrosion cracking in or near stainless steel welds. In both cases, the conventionally applied testpiece defect manufacturing processes have been shown to introduce unrealistic defects with significant manufacturing artefacts. For more reliable performance demonstration and qualification, further development is needed to get realistic test defects that allow true observation of real world performance. In particular, the development of type 4 defect manufacturing(2) which gives realistic defects and avoids any attachment welds would be needed to overcome their traditional disadvantages – namely restrictions in growth procedure and reliance upon a supplementary inspection to confirm critical flaw parameters. Trueflaw produces type 4 grown cracks using thermal fatigue cracking mechanism. The purpose of this paper is to present the current status of the crack manufacturing technology and how some of the limitations mentioned above have been overcome. Furthermore, the paper presents some application examples of how this technology has been used to solve real-world problems in different fields. 2. Trueflaw crack manufacturing technology Trueflaw produces defects using natural thermal fatigue damage process. The defects are grown in much the same way as could occur during in-service condition. However, the growth is accelerated to make production times practical and controlled to enable predetermined flaw parameters. Flaw production is done in-situ to ready-made samples. Cyclic thermal fatigue loading is induced locally by alternating heating and water spray cooling, as described by Kemppainen(3). The loading is based on pure thermal loading and there is no welding, machining, or mechanical treatment applied. No artificial initiators of any kind are used and the material microstructure is not disturbed in the process. More detailed information on the properties and use of produced cracks has been presented earlier(4,5,6). 2.1 Restrictions in growth procedure Manufacturing of grown defects has traditionally been restricted to simple component shapes and small components. The reason has been that crack growth, in general, requires stress to provide a driving force. Providing the required stress mechanically becomes impractical when material thickness increases or geometry becomes more complicated. Huge mechanical loading equipment would be needed to generate sufficient stress to most components of any practical interest. Even if such equipment was available, accurate control of induced stress in complex shapes during crack growth would be very difficult and it would not be possible to limit the stress only to areas where defects are needed. In contrast, thermal loading can be applied to local areas in heavy components. Since only a limited volume is stressed at any given time, the needed equipment is relatively light. Furthermore, the ability to locate and control the stressed volume enables accurate control over flaw growth location and essential flaw parameters. Consequently, thermal fatigue cracks can be grown in components of any size or shape. 2.2 Confirming critical flaw parameters without supplementary inspections In order to use test defects to assess NDT performance, the true parameters of the defects must be known. Otherwise, the error in NDT results cannot be accurately determined and the true reliability of the NDT remains unclear. Some of the parameters of test defects can be readily measured, for example, surface length. However, most important defect parameters, defect depth in particular, are not directly observable. It may be argued that reliable defect depth information is the most important challenge on many defect manufacturing techniques. Defect manufacturing techniques in general and grown cracks in particular conventionally rely on ‘supplementary inspection’ to give defect depth information(2). Consequently, reliability is assessed by comparing one inspection method with another rather than comparing inspection to-be-qualified against accurately known flaw parameters. This is unacceptable. The alternate route, suggested by the ENIQ working document(2), is to destructively examine the defects once the qualification is complete. While this has been successfully done in some special cases(7), it is not generally considered a feasible option. Test blocks with realistic geometry are far too expensive to manufacture to allow this kind of qualification. Furthermore, any additional qualification, re-qualification and method development would require a new set of test blocks. To overcome this very significant shortcoming, Trueflaw has developed an alternative way to verify critical flaw parameters, and in particular the flaw depth. This approach retains the credibility of destructive examination and avoids the expensive and problematic destruction of valuable test blocks. The key feature of this approach is the development of a highly repeatable crack growth procedure. Because of the repeatability, not all the cracks need to be destructively examined. In simple terms, the procedure is as follows: first, the desired crack depth is produced in a representative validation sample. This sample needs to have similar material and similar local geometry, but can be simplified and smaller compared to the actual test block. This validation crack is destructively examined to reveal the true crack depth and other desired parameters (crack opening, surface roughness and so on can be measured at this stage). Then, using the same procedure, a similar crack is produced in the actual test block. Due to very good repeatability, this crack has the same depth and other essential parameters as the destructively examined validation crack. Finally, all the destructive validation cracks are analysed to give an estimate on the process variability and a tolerance is determined to given crack depth values. 2.3 Crack growth repeatability The repeatability of the crack production is most important as this is reflected in the crack growth tolerance that can be given for the crack depth. Thus, significant effort has been made to assess and analyse process repeatability. Throughout its history, Trueflaw has manufactured and destructively examined altogether 215 validation cracks to date. The work is ongoing and new data is added continuously. The data spans a wide variety of materials, component geometries and crack sizes. As an example, Figure 1 presents the current validation data for austenitic stainless steel base material. The data includes a wide variety of different austenitic stainless steel base material samples and crack sizes. The maximum error in this data set is ±0.4 mm. The standard crack depth tolerance given for produced cracks is ±1.0 mm, due to practical client requirements. It is seen that the process variability is well within the given tolerance. 2.4 Limitations of thermal fatigue crack growth The nature of the described thermal fatigue crack growth technology allows only surface-breaking cracks to be manufactured. Furthermore, the location where the crack is to be manufactured must be attainable (ie, the loading tool must fit to the location). This prevents crack production to, for example, inner diameter (ID) of very small tubes. Currently, the smallest tube ID where cracks have been produced is about 16 mm. 2 Insight Vol 51 No 5 May 2009 Insight Vol 51 No 5 May 2009 3 While the technique is applicable to a wide variety of materials, there is also some materials that present a challenge. Currently cracks can not be manufactured, for example, to copper and aluminium. 3. Case examples on application of Trueflaw technology Three cases using Trueflaw cracks are presented in the following to give an overview on the application possibilities. The three cases were selected to give a diverse selection of non-trivial applications. The cases span different materials, component geometries and crack sizes. 3.1 E.ON reactor pressure vessel head nozzles The non-destructive inspection of dissimilar welds is an important part of the inspection programme in refuelling outages in nuclear power plants. The inspection of the inner weld surface in the reactor pressure vessel head nozzles of German PWR plants is complicated by geometrical constriction. This dissimilar weld is accessible only through a 1 mm-thick gap, through which the eddy current probe must pass. For this inspection, a new eddy current technique had to be developed. Due to the geometrical limitations, the probe design had to ensure an extremely flat probe. The qualification of the inspection technique was performed with a test specimen made of a real nozzle using EDM notches as simulation of cracks according to applicable rules. During the inspection in 2007 an indication was found close to the austenitic side of the dissimilar weld in one nozzle. The signal was not within the phase range of defects found in the qualification and the signature was totally different from the signal of notches. So, the indication was not classified as a defect signal. Nevertheless, it was decided to make further investigation to find out the reason for the signal. One of the points to study in this investigation was to find out the difference between notch signals and the signals of real cracks. The next aim was to develop a visual technique able to inspect the inner weld surface through the 1 mm gap. A new test specimen was made using again an original nozzle. Trueflaw was ordered to manufacture cracks in this new specimen and as well to make different EDM notches and notch fields as a reference. E.ON supplied an original nozzle to Trueflaw to be used as a test block. Part of the test block was marked to be used for validation. Trueflaw produced validation cracks of desired size in this area. Figure 2 shows an example image from a validation crack with measured crack opening on the surface. The area containing the validation cracks was then cut out from the tube using electric discharge machining (EDM) and the cracks destructively examined to reveal the true crack depth. During the production, E.ON and consultant expert of the authority (TÜV) visited Trueflaw to follow the progress. Subsequent to the accepted validation result, the final cracks were manufactured and the sample supplied to E.ON. With the manufactured cracks, the eddy current system qualification was repeated, and the phase range for defects could be basically verified but reduced at the edges. It could be proved that, due to a crack with secondary crack close to it, no phase shift occurs when more then one crack is in the area of influence of the probe. The new developed visual inspection technique (using special optical components and CCD chip together with an optical fibre lighting) was as well qualified with the natural cracks from Trueflaw. In the 2008 outage, a second inspection with the optimised qualification and the visual inspection was made. It could be shown that the reason for the indication was of geometrical nature. A crack in the component could be excluded. 3.2 Fortum steam generator primary collector Fortum Ltd, Loviisa Powerplant (Finland) conducted an ultrasonic qualification for a VVER steam generator primary collector during the 2008 summer outage. A schematic illustration of the steam generator is presented in Figure 3. The area of interest is cracking in M48x5 threaded holes of the primary collector flange. The inspection is done using phased array UT with scanning from top and inner diameter (ID) surface of the primary collector. It was decided to use a component removed from a similar powerplant, that never went to operation, as a test block for the qualification. Figure 4 shows the qualification test block. The flaw types to be detected are shown in Figure 5. Fortum provided Trueflaw with the target flaw sizes and locations for this very challenging geometry. The flange is a forged ring fabricated from Ti-stabilised austenitic stainless steel. Since the material and geometric conditions are unique, new validation for each crack size was required for reliable flaw production. A material sample was cut out from the test block for validation purposes. A simplified validation sample was machined that replicated the local Figure 1. Example of validation data for austenitic stainless steel Figure 2. Example validation crack in E.ON nozzle. The crack was imaged using a camera microscope and is shown with respective crack opening measured from the image Figure 3. Loviisa powerplant horizontal steam generator geometry conditions of the threaded hole bottom cup. Trueflaw produced validation cracks for all the desired flaw sizes and locations and supplied a destructive evaluation report to Fortum. After accepted validation, the production of the actual qualification defects was done and test block supplied to Fortum. Figure 6 shows an example of a fracture surface from this validation. Thermal fatigue cracks were supplemented with a selection of EDM notches in different locations. The open trials on the test block were performed during the 2008 summer outage. All the defects were successfully detected in open trials with UT examination (even the small sizes). Loviisa now has a reliable inspection procedure that is tested with real cracks. 3.3 Rolls-Royce seal fin sample Rolls-Royce wished to study the effectiveness of novel NDT methods in detecting cracks under conductive coatings and needed a sample with a known crack population. The component chosen for this work was the seal fin region of a turbine disc. In use this component is covered with a wear coating (TBT406). The task was to create a realistic testpiece containing cracks under the coating and by using the Trueflaw method the cracks could be placed in the required position at the tips of the fin. Figure 7 shows a schematic illustration of the seal fin sample. The fin sample was provided by Rolls-Royce. The material and geometry were both new to Trueflaw. Consequently, part of the fin sample was dedicated to production test and validation. In this case, all the critical flaw parameters were directly observable and depth validation was not necessary. It was expected that the crack opening would affect the NDT methods to be studied. Consequently, several production trials were completed to allow production of a variety of crack openings. Furthermore, while doing the production tests it became evident that different fin locations in the sample had different responses to fatigue loading. Numerous cracks were produced at different locations on the seal fin sample to allow determination of inspection capabilities in Figure 4. Loviisa test block from vintage material Figure 5. Flaw types to be detected in Loviisa primary collector qualification Figure 6. Fracture surface from a destructively examined validation crack corresponding to crack type 3 in Figure 3 Figure 7. Schematic image of the seal fin sample 4 Insight Vol 51 No 5 May 2009 Insight Vol 51 No 5 May 2009 5 all interesting locations. A sample red dye penetrant test (PT) image from a crack produced in the fin sample is shown in Figure 8. Characterisation was done by using penetrant testing with magnified digital imaging to measure the crack size and the crack opening. Some NDT (thermosonics) was carried out prior to the coating being applied. Following verification of the cracks using fluorescent penetrant the part was coated and a sample crack cut out. X-ray computer tomography was used to visualise the crack under the coating to verify the coating material had not entered the crack. This can be seen in Figure 9(a), showing the coated fin. Figure 9(b) shows the X-ray image with the tip of the fin removed making the crack beneath clearly visible. This sample is now being used in a series of trials to establish if an inspection method is possible. 3. Conclusions The most significant disadvantages traditionally associated with realistic grown defects have been overcome by developments in the thermal fatigue crack growth process as shown in this paper. The developed validation procedure has solved the traditional problem of reliance upon a supplementary inspection to confirm critical flaw parameters for grown cracks. A similar validation approach could be used with any repeatable crack growth process. Thermal fatigue cracks have been successfully used in numerous practical applications ranging from qualification to development and testing of novel NDT methods. This is shown by the various real-world application cases presented in this paper. The technology is reliable and mature. Acknowledgements The industrial usage examples were kindly provided by E.ON (Mr Henner Ostermeyer), Fortum (Mr Raimo Paussu) and Rolls- Royce plc (Mr Tony Dunhill). Their contribution is gratefully acknowledged. Figure 8. PT indication of seal fin crack Figure 9(a). X-ray image of seal fin showing coating is uncracked Figure 9(b). Seal fin tip removed from X-ray image showing subsurface crack References 1. Anon, ‘The European methodology for qualification of non- destructive testing’, Second Issue, EUR 17299 EN, published by the European Commission, Brussels-Luxembourg, 1997. 2. ENIQ working document (private communication), to be published. 3. M Kemppainen, ‘Realistic artificial flaws for NDE qualification – novel manufacturing method based on thermal fatigue’, Dissertation for the degree of Doctor of Science in Technology, Espoo, Finland, 2006. (Available online from: http://lib.tkk.fi/ Diss/2006/isbn9512282631/) 4. R Paussu, I Virkkunen and M Kemppainen, ‘Utility aspect of applicability of different flaw types for qualification test pieces’, 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Budapest, Hungary, 8-10 Oct 2007. 5. T Packalén, J Sillanpää, M Kemppainen, I Virkkunen and R Paussu, ‘The influence of the crack opening in the UT inspection qualification’, 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Budapest, Hungary, 8-10 Oct 2007. 6. M Kemppainen, I Virkkunen, T Packalén, J Sillanpää and R Paussu, ‘Importance of crack opening in UT inspection qualification’, 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Budapest, Hungary, 8-10 Oct 2007. 7. R Krajcovic and J Plasek, ‘Eddy current inspection of WWER steam generator tubes – sensitivity of bobbin probe technique’, 9th European Conference on NDT, Berlin, 25-29 September 2006. DGZfP e.V. (Available online from: http://www.ndt.net/ article/ecndt2006/doc/Th.3.1.4.pdf)

Grown cracks for NDT development and qualification Iikka Virkkunen, Mika Kempainen Trueflaw Ltd. Espoo, FI-02330, Finland +358456354415 iikka.virkkunen@trueflaw.com Henner Ostermeyer EON Kernkraft GmbH, Germany Raimo Paussu Fortum Nuclear Services, Finland Abstract Defects are needed to develop new NDT methods and to assess the performance and reliability of used methods and procedures. It is crucial to have representative defects in order to have an accurate and realistic assessment of the performance of NDT. Representativeness should be to the actual service-induced defects that the NDT method is used to evaluate. While various techniques have been used to create such defects, all conventional techniques seem to have some shortcomings that limit true assessment of the NDT performance. This paper describes recent developments of defect manufacturing technology based on controlled thermal fatigue. It is shown, that most of the traditional limitations can be overcome using the currently available technology. Finally three real-world application cases are presented showing the use of such cracks. 1. Introduction The real performance and reliability of used NDT techniques and procedures should be known in order to effectively use them. Without this knowledge, it is difficult to select the correct methods and inspection targets, let alone determine correct inspection intervals. Also, dependable performance information that highlights potential targets for further improvement is necessary for the development of better NDT techniques. To provide this crucial information, various performance demonstration and qualification procedures have been established and are under development (1,2). One of the key challenges in assessing NDT performance is the production of relevant test blocks which the performance can be tested with. These test blocks should contain defects identical to those expected in real inspection, but with known and predetermined location, size and other properties. Producing such controlled natural defects has been, and still is, quite a difficult task. Consequently, a number of defect simulation techniques have been developed, each with their virtues and limitations. As proposed in an ENIQ working document(2), there is essentially four classes of defects currently available (numbering and highlights added): 2 1. Implanted Defects where a pre-existing defect is attached to the testpiece. The attachment usually takes the form of a weld in a machined recess. The technique has the benefits that there is flexibility in the type of defect that can be included and that the insert can be carefully accessed prior to insertion. The main disadvantages are that the insertion process may produce artefacts which either give away the implants position or make the inspection response unrealistic in some way. An example of this latter affect is implants into austenitic weld where the implant material will not form a continuous part of the weld and the attachment welds may significantly influence the performance of the inspection being qualified in an unknown manner. 2. Weld doping or weld modification where for instance crack prone material is added to weld to promote localised weld cracking. Other examples include introduction of porosity or slag. The main advantage over the previous process is that there are no insert attachment welds. Main disadvantages are that the final size of the defects and their character would need to be confirmed by supplementary inspection. This means that there is a risk of comparing one inspection method with another rather than comparing one inspection against known flaw parameters. Another disadvantage is that the doping process can influence the material properties of the weld in the immediate vicinity of the defect potentially affecting the inspection in an unpredictable manner. 3. Machined Defects where a defect can consist of a cut or machined void. Electro Discharge Machining (EDM) is perhaps the most relied upon technology in this area where an shaped electrode is used to erode the testpiece. The process is most suitable for production of surface defects, although it is possible to use in combination with welding to produce buried defects. The main advantages of this method is that it tends to be relatively inexpensive, the resulting defect parameters are known to fairly tight tolerances at fabrication and the parent material is left largely unmodified apart from the presence of the machined slot. Disadvantages are that it is difficult or impossible to produce any of the characteristic roughness expected from plant defects and that using standard implantation techniques, the tip radius is likely to be large compared to many crack species. 4. Grown Defects where cracking is initiated and propagated into testpieces in much the same way as would occur in plant, simply accelerated to make fabrication times practical. The main processes used for this class of defect are fatigue, thermal fatigue and stress corrosion cracking. There are limitations associated with each growth method, but this option has the advantages of realism and avoidance of attachment welds. The main disadvantage aside from restrictions in the implant process is likely to be reliance upon a supplementary inspection to confirm critical flaw parameters. In the schemes already discussed, this limitation can be minimized by using the defects in parametric specimens and then destructive examining so or all of the specimens once the qualification process is complete. Conventionally, methods 1-3 are applied for performance demonstration and qualification. Most qualifications in effect currently rely on defects produced with these three defect types. 3 As is evident from the above quote, finding suitable defects for performance demonstration is rather demanding task. Obviously, the problems in using defects that "affect the inspection in an unpredictable manner" make it difficult if not impossible to infer real-world performance from the performance demonstration data acquired using unsuitable defects. This effectively undermines the practical value of the whole exercise. As stated by the ENIQ working document (2): Some plant defects when inspected with techniques generally used in plant present a very significant challenge to testpiece design and testpiece defect fabrication. Examples are the qualification of ultrasonic inspections of austenitic or Inconel weld metal and inspections for intergranular stress corrosion cracking in or near stainless steel welds. In both cases, the conventionally applied testpiece defect manufacturing processes have been shown to introduce unrealistic defects with significant manufacturing artefacts. For more reliable performance demonstration and qualification, further development is needed to get realistic test defects that allow true observation of real world performance. In particular, the development of type 4 defect manufacturing(2) which gives realistic defects and avoids any attachment welds would be needed to overcome their traditional disadvantages - namely restrictions in growth procedure and reliance upon a supplementary inspection to confirm critical flaw parameters. Trueflaw produces type 4 grown cracks using thermal fatigue cracking mechanism. The purpose of this paper, is to present the current status of the crack manufacturing technology and how some of the limitations mentioned above have been overcome. Furthermore, the paper presents some application examples of how this technology has been used to solve real-world problems in different fields. 2. Trueflaw crack manufacturing technology Trueflaw produces defects using natural thermal fatigue damage process. The defects are grown in much the same way as could occur during in-service condition. However, the growth is accelerated to make production times practical and controlled to enable predetermined flaw parameters. Flaw production is done in-situ to ready-made sample. Cyclic thermal fatigue loading is induced locally by alternating heating and water spray cooling, as described by Kemppainen(3). The loading is based on pure thermal loading and there is no welding, machining, or mechanical treatment applied. No artificial initiators of any kind are used and the material microstructure is not disturbed in the process. More detailed information on the properties and use of produced cracks has been presented earlier(4,5,6). 2.1 Restrictions in growth procedure Manufacturing of grown defects has traditionally been restricted to simple component shapes and small components. The reason has been that crack growth, in general, requires stress to provide a driving force. Providing the required stress mechanically becomes impractical when material thickness increases or geometry becomes more 4 complicated. Huge mechanical loading equipment would be needed to generate sufficient stress to most components of any practical interest. Even if such equipment was available, accurate control of induced stress in complex shapes during crack growth would be very difficult and it would not be possible to limit the stress only to areas where defects are needed. In contrast, thermal loading can be applied to local areas in heavy components. Since only a limited volume is stressed at any given time, the needed equipment is relatively light. Furthermore, the ability to locate and control the stressed volume enables accurate control over flaw growth location and essential flaw parameters. Consequently, thermal fatigue cracks can be grown to components of any size or shape. The only remaining limitation is, that the surface where crack is to be produced must be accessible to the loading tool. Currently, this prevents using the technology inside very small inner diameter pipes (current limiting diameter is 16 mm). 2.2 Confirming critical flaw parameters without supplementary inspections In order to use test defects to assess NDT performance, the true parameters of the defects must be known. Otherwise, the error in NDT results cannot be accurately determined and the true reliability of the NDT remains unclear. Some of the parameters of test defects can be readily measured, e.g., surface length. However, most important defect parameters, defect depth in particular, are not directly observable. It may be argued, that reliable defect depth information is the most important challenge on many defect manufacturing techniques. Defect manufacturing techniques in general and grown cracks in particular conventionally rely on "supplementary inspection" to give defect depth information(2). Consequently, reliability is assessed by comparing one inspection method with another rather than comparing inspection to-be-qualified against accurately known flaw parameters. This is unacceptable. The alternate route, suggested by the ENIQ working document(2) is to destructively examine the defects once the qualification is complete. While this has been successfully done in some special cases(7), it's not generally feasible. Test blocks with realistic geometry are far too expensive to manufacture, to allow this kind of qualification. Furthermore, any additional qualification, re- qualification and method development would require new set of test blocks. To overcome this very significant shortcoming, Trueflaw has developed an alternative way to verify critical flaw parameters, and in particular the flaw depth. This approach retains the credibility of destructive examination and avoids the expensive and problematic destruction of valuable test blocks. The key feature of this approach is the development of a highly repeatable crack growth procedure. Because of the repeatability, not all the cracks need to be destructively examined. In simple terms, the procedure is as follows: first, the desired crack depth is produced to representative validation sample. This sample needs to have similar material and similar local geometry, but can be simplified and smaller compared to the actual test block. This validation crack is destructively examined to reveal the true crack depth and other 5 desired parameters (crack opening, surface roughness etc. can be measured at this stage). Then, using the same procedure, a similar crack is produced to the actual test block. Due to very good repeatability this crack has the same depth and other essential parameters as the destructively examined validation crack. Finally, all the destructive validation cracks are analyzed to give an estimate on the process variability and a tolerance is determined to given crack depth values. 2.3 Crack growth repeatability The repeatability of the crack production is most important as this is reflected in the crack growth tolerance that can be given for the crack depth. Thus, significant effort has been made to assess and analyze process repeatability. Throughout it's history, Trueflaw has manufactured and destructively examined altogether 215 validation cracks up to date. The work is ongoing and new data is added continuously. The data spans wide variety of materials, component geometries and crack sizes. As an example, Figure 1. presents the current validation data for austenitic stainless steel base material. The data includes wide variety of different austenitic stainless steel base material samples and crack sizes. The maximum error in this data set is ±0.4 mm. The standard crack depth tolerance given for produced cracks is ±1.0 mm, due to practical client requirements. It is seen, that the process variability is well within the given tolerance. Figure 1. Example of validation data for austenitic stainless steel 3. Case examples on application of Trueflaw technology Three cases using Trueflaw cracks are presented in the following to give an overview on the application possibilities. The three cases were selected to give a diverse selection of 6 non-trivial applications. The cases span different materials, component geometries and crack sizes. 3.1 E.On reactor pressure vessel head nozzles The non-destructive inspection of dissimilar welds is an important part of the inspection program in refuelling outages in nuclear power plants. The inspection of the inner weld surface in the reactor pressure vessel head nozzles of german PWR plants is complicated by geometrical constriction. This dissimilar weld is accessible only through a 1 mm thick gap, which the eddy current probe must pass trough. For this inspection a new eddy current technique had to be developed. Due to the geometrical limitations the probe design had to ensure an extremely flat probe. The qualification of the inspection technique was performed with a test specimen made of a real nozzle using EDM notches as simulation of cracks according to applicable rules. During the inspection in 2007 an indication was found close to the austenitic side of the dissimilar weld in one nozzle. The signal was not within the phase range of defects found in the qualification and the signature was totally different from the signal of notches. So the indication was not classified as a defect signal. Nevertheless it was decided to make further investigation to find out the reason of the signal. One of the points to study in this investigation was to find out the difference between notch signals and the signals of real cracks. Next aim was to develop a visual technique able to inspect the inner weld surface through the 1 mm gap. A new test specimen was made using again an original nozzle. Trueflaw was ordered to manufacture cracks in this new specimen and as well to make different EDM notches and notch fields as a reference. E.On supplied an original nozzle to Trueflaw to be used as a test block. Part of the test block was marked to be used for validation. Trueflaw produced validation cracks of desired size to this area. Figure 2 shows example image from a validation crack with measured crack opening on the surface. The area containing the validation cracks was then cut out from the tube using electric dicharge machining (EDM) and the cracks destructively examined to reveal the true crack depth. During the production, E.On and consultant expert of the authority (TÜV) visited Trueflaw to follow the progress. Subsequent to the accepted validation result, the final cracks were manufactured and sample supplied to E.On. 7 Figure 2. Example validation crack in E.On nozzle. Crack was imaged using a camera microscope and shown with respective crack opening measured from the image. With the manufactured cracks, the eddy current system qualification was repeated, and the phase range for defects could be basically verified but reduced at the edges. Due to a crack with secondary crack close to it could be proved, that no phase shift occurs, when more then one crack is in the area of influence of the probe. The new developed visual inspection technique (using special optical components an ccd chip together with an optical fibre lighting) was as well qualified with the natural cracks from Trueflaw. In the outage 2008 a second inspection with the optimized qualification and the visual inspection was made. It could be shown that the reason for the indication was of geometrical nature. A crack in the component could be excluded. 3.2 Fortum steam collector primary collector Fortum Ltd., Loviisa Powerplant (Finland) conducted an ultrasonic qualification for VVER steam generator primary collector during the 2008 summer outage. A schematic illustration of the steam generator is presented in Figure 3. 8 Figure 3. Loviisa powerplant horizontal steam generator The area of interest is cracking in M48x5 threaded holes of the primary collector flange. The inspection is done using phased array UT with scanning from top and inner diameter (ID) surface of primary collector. It was decided to use component removed from a similar poverplant, that never went to operation as a test block for the qualification. Figure 4. shows the qualification test block. The flaw types to be detected are shown in Figure 5. Figure 4. Loviisa test block from vintage material 9 Figure 5. Flaw types to be detected in Loviisa primary collector qualification Fortum provided Trueflaw with the target flaw sizes and locations for this very challenging geometry. Flange is a forged ring fabricated from Ti-stabilised austenitic stainless steel. Since the material and geometric conditions are unique, new validation for each crack size was required for reliable flaw production. Material sample was cut out from the test block for validation purposes. A simplified validation sample was machined, that replicated the local geometry conditions of the threaded hole bottom cup. Trueflaw produced validation cracks for all the desired flaw sizes and locations and supplied destructive evaluation report to Fortum. After accepted validation, the production of the actual qualification defects were done and test block supplied to Fortum. Figure 6 shows an example of a fracture surface from this validation. Thermal fatigue cracks were supplemented with a selection of EDM-notches in different locations. Figure 6. Fracture surface from a destructively examined validation crack corresponding to crack type 3 in Figure 3. 10 The open trials on the test block were performed during the 2008 summer outage. All the defects were successfully detected in open trials with UT examination (even the small sizes). Loviisa now has a reliable inspection procedure that is tested with real cracks. 3.2 Rolls-Royce seal fin sample Rolls-Royce wished to study the effectiveness of novel NDT methods in detecting cracks under conductive coatings and needed a sample with a known crack population. The component chosen for this work was the seal fin region of a turbine disc. In use this component is covered with a wear coating (TBT406). The task was to create a realistic test piece containing cracks under the coating and by using the Trueflaw method the cracks could be placed in the required position at the tips of the fin. Figure 7. shows a schematic illustration of the seal fin sample. Figure 7. Schematic image of the seal fin sample The fin sample was provided by Rolls-Royce. The material and geometry were both new to Trueflaw. Consequently, part of the fin sample was dedicated to production test and validation. In this case, all the critical flaw parameters were directly observable and depth validation was not necessary. It was expected, that the crack opening would affect the NDT methods to be studied. Consequently, several production trials were completed to allow production of a variety of crack openings. Furthermore, while doing the production tests it became evident, that different fin locations in the sample had different response to fatigue loading. 11 Numerous cracks were produced at different locations on the seal fin sample to allow determination of inspection capabilities in all interesting locations. A sample red dye penetrant test (PT) image from a crack produced to the fin sample is shown in Figure 8. Figure 8. Schematic image of the seal fin sample Characterisation was done by using penetrant testing with magnified digital imaging to measure the crack size and the crack opening. Some uncoated NDT (Thermosonics) was carried out prior to the coating was applied. This sample is now being used in a series of trials to establish if an inspection method is possible. There is a risk that the coating may have entered the wider cracks and to understand this some coated cracks will be cut from the disc segment to perform micro x-ray computer tomography analysis. 3. Conclusions The most significant disadvantages traditionally associated with realistic, grown defects have been overcome by developments in the thermal fatigue crack growth process as shown in this paper. The developed validation procedure has solved the traditional problem of reliance upon a supplementary inspection to confirm critical flaw parameters for grown cracks. Similar validation approach could be used with any repeatable crack growth process. Thermal fatigue cracks have been successfully used in numerous practical applications ranging from qualification to development and testing of novel NDT methods. This is shown by the various real-world application cases presented in this paper. The technology is reliable and mature. ! 12 Acknowledgements The industrial usage examples were kindly provided by E.On (Mr. Henner Ostermeyer), Fortum (Mr. Raimo Paussu) and Rolls-Royce (Mr. Tony Dunhill). Their contribution is gratefully acknowledged. References 1. Anon. 'The European methodology for qualification of non-destructive testing', Second Issue, EUR 17299 EN, published by the European Commission, Brussels- Luxembourg, 1997. 2. ENIQ working document (private communication), to be published 3. M Kemppainen, 'Realistic artificial flaws for NDE qualification – novel manufacturing method based on thermal fatigue', Dissertation for the degree of Doctor of Science in Technology, Espoo, Finland, 2006. (Available online from: http://lib.tkk.fi/Diss/2006/isbn9512282631/) 4. R Paussu, I Virkkunen and M Kemppainen, 'Utility aspect of applicability of different flaw types for qualification test pieces', 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8 th – Oct 10th. Budapest, Hungary. 2007. 5. T Packalén, J Sillanpää, M Kemppainen, I Virkkunen and R Paussu, 'The influence of the crack opening in the UT inspection qualification', 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8 th – Oct 10th. Budapest, Hungary. 2007. 6. M Kemppainen, I Virkkunen, T Packalén, J Sillanpää and R Paussu, 'Importance of crack opening in UT inspection qualification', 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, Oct 8 th – Oct 10th. Budapest, Hungary. 2007. 7. R Krajcovic and J Plasek, 'Eddy current inspection of WWER steam generator tubes - sensitivity of bobbin probe technique', 9th European Conference on NDT, Berlin, 25-29.9.2006. DGZfP e. V. (available online from: http://www.ndt.net/article/ecndt2006/doc/Th.3.1.4.pdf)

UTILITY ASPECT OF APPLICABILITY OF DIFFERENT FLAW TYPES FOR QUALIFICATION TEST PIECES Raimo Paussu, Fortum Nuclear Services Ltd, Finland Iikka Virkkunen, Mika Kemppainen, Trueflaw Ltd, Finland ABSTRACT The objective is to give the utility perspective for the basis of selection and fabrication of realistic flaws for different qualification cases. Good knowledge about the applicability of flaw types for different purposes is important. National project has been carried out for getting information of essential crack features considering ultrasonic response and finally defining the metallographic features of fabricated cracks. Fortum Nuclear Services is an expert organization supporting Loviisa NPP. Trueflaw is an advanced crack manufacturer. Wide experience of VVER components and the knowledge of component failures, failure mechanisms, defect morphology, manufacturing technology and welding give good background also for design and fabrication of test pieces with representative flaw population of realistic flaws. The basic ideas for flaw design and selection of crack and flaw fabrication methods are described in the paper. The key issue of reliable ISI inspections at Loviisa NPP is to select the potential failure areas for inspection program and to detect cracks in early stage, before they grow to critical size during an inspection interval. Compilation of comprehensive and technically sound input data for the qualification cases is one of the key issues of the utility organizations, and the best way to guarantee conditions for successful qualification, and fabrication of test pieces with proper flaws. The main objective of test piece design is to use test pieces with several cracks, weighing benefits and disadvantages of different crack types for each qualification case. Utility experience and views are discussed and conclusions are drawn based on results of qualification cases and current national project about the applicability of different crack production techniques, realistic simulation of studied crack types and NDE results. INTRODUCTION Council of State of Finland has recently granted for both VVER-440 units of Loviisa Power Plant an operating license for additional 20 years, totally for 50 years operation. Extension of operation life is essentially based on assessments of plant life management (PLIM) program, and analysis and development of identified degradation phenomenon of components and structures. The material research program of Fortum is tightly connected and targeted to support licensing and plant life management process of Loviisa NPP. The purpose of material R&D program, managed by Fortum Nuclear Services (FNS) is to produce the material knowledge for recognition of degradation issues in time, and create technical preparedness for application of methods and knowledge needed for safety assessments and trouble-free operation of Loviisa NPP. Reliable ISI has an important role in confirmation of structural integrity of components and piping. Compilation of the RI-ISI programs for Loviisa NPP is the new requirement of Finnish Authority. The selection of risky areas into the ISI program would reduce the total risk. The qualification of inspection systems and inspection personnel is needed for reliable performance of ISI program. One subtask of the material R&D program is to create preparedness for manufacturing test pieces with relevant cracks and reflectors needed for different qualification cases of Loviisa NPP. The co-operation with Trueflaw Ltd. strengthens and completes the capabilities of FNS for test piece and flaw manufacturing, and vice versa. Test piece fabrication and trials for qualification of UT inspection systems started in 1996 at Loviisa NPP. The qualification rules were established in 2000 and the roles of participants were defined to follow the qualification requirements of guide 3.8 of the Finnish Authority /1/ and the ENIQ Methodology /2/. Inspecta Certification is independent and permanent member of qualification body in Finland using case by case nominated, outside expert members /3/. The Authority approves the input data for qualification and all the qualification activities are delegated to the qualification body (QB). The Authority supervises the activity of QB. Training of UT personnel for detection of stress corrosion cracking started in the middle of 80's, when this cracking phenomenon was identified. Additional training of UT personnel for crack detection was the requirement of the Finnish Authority. In the beginning of 90's, test pieces with service induced cracks were rented from Sweden for practical training, and unofficial qualification tests started. Most of the blind and open test pieces with crack simulations, used for personnel training and qualification in Finland today, are produced by FNS together with Trueflaw. Training today includes theoretical training for using proper inspection techniques and practical training for flaw detection of typical crack types using test pieces with crack simulations. The official personnel qualification started in 1997 using blind test pieces with crack simulations. Pilot qualification for flaw sizing was organized in 2000. Inspecta Certification as an independent organization is responsible for the qualification of NDE personnel in Finland. DESIGN BASIS FOR SELECTION OF FLAW POPULATION AND FLAW TYPES National qualification rules in Finland define guidelines for test piece fabrication for open and blind trials describing the amount and sizes of defects needed /3/. The input data for qualification case with detection target and defect types to be detected shall be considered when selecting defect types and sizes for test pieces. Exact rules are not given so far for the defect types to be applied simulating different degradation mechanisms defined in input data for qualification. Common documentation requirements for test pieces and flaws are now given in revised national qualification rules. The Supplements of mandatory Appendix VIII of ASME Code Section XI are relevant regulation for design of flaw size and type population for the test pieces for different qualification cases and applications. Clear rules are given in those Supplements for the amount of different types of cracks and size distribution in flaw population that should be present in test pieces, and how other type of reflectors can be used. Fabrication of test pieces and flaws is executed in Finland with financing of the utility. New open test pieces for qualification cases of Loviisa NPP are designed together with QB and FNS, taking into account the needs of Inspection Company and the Utility. With respect to blind test pieces, the Utility doesn’t know the details of flaw sizes, types and amount to be fabricated. This will guarantee the confidentiality and independency of QB for executing qualification without influence of the Utility. For blind test pieces, QB and FNS will design the flaw size population together. In many cases FNS will propose the flaw simulations to be applied, and the design is reviewed together before its approval by QB. Flaw manufacturing can be audited by QB. EXPERIENCE OF DESIGN AND MANUFACTURING OF TEST PIECES AND FLAWS From flaw manufacturing perspective, the most crucial aspect is the condition of test piece. Producing flaws in the finished test piece is the most limiting case concerning the possibilities to apply different crack simulation methods. The easiest case is to fabricate a new test piece and to utilize all relevant crack and reflector simulations. Features and applications of crack simulations of some degradation mechanisms and reflectors simulating welding defects and planar defects are described below based on experience gained from trials and qualifications of inspection systems. Thermally produced cracks Thermally produced cracks are initiated and grown with heating and cooling cycling without any initiator on surface. Thermal fatigue crack production will not change the structure of material, and cracking will follow the weakest path in material. The advanced crack producing technique of Trueflaw is beneficial to certain qualification cases. The best advantage is the possibility to produce cracks into finished test piece without any disturbing effects on NDE response. One additional benefit is the capability to manipulate crack opening along the crack face on purpose, if needed. Typical application cases are the cracks of base material and corners (simulating fatigue cracks) and cracks further from weld toe (simulating crack location of stress corrosion cracking), cracks transverse to weld and cracks for ET and VT applications. Thermal fatigue crack is applied to qualification cases in ferritic steel, austenitic stainless steel and Inconel base materials and welds, buttering and cladding. Thermal fatigue cracks are postulated crack type in many qualification cases of piping at Loviisa NPP. Experience gained from qualification cases shows that these cracks are suitable. Mechanically produced crack simulation Mechanically produced crack simulations are produced using welded aid piece and manually cycled with bending the aid piece up to fracture, to create crack surface on weld groove. The shape of crack front is finished on aid piece before welding the fracture surfaces together. Crack simulation can be tilted, skewed and fabricated to defined size. Crack simulation can be produced in different materials and boundaries as surface and subsurface cracks. Typically crack surfaces are rather rough and crack tips can be fitted tight, but crack opening on surface is often more open. Mechanically produced crack simulations are most convenient to a new test piece to be welded, or to finished test pieces with small crack sizes. Fatigue cracks are the postulated crack type in the most qualification cases of Loviisa NPP. Experience of qualification cases shows that these simulated cracks are working. Welded solidification cracks Solidification cracks are produced into narrow opening by welding with proper filler material to produce mixing of melt, sensitive for cracking. Solidification crack has blunt tip and it’s typically located in the middle of welded pass. Crack opening is wide in austenitic, narrower in Inconel and quite narrow in ferritic weld. Crack can be straight, winding or tortuous and have interruptions in longitudinal and depth direction. The first trials and applications were executed at the end of 90’s into root surface of thick walled butt weld of piping by opening the finished root and by welding the solidification crack. In those applications, the dendrite structure of welded opening with crack is opposite compared to weld structure around the opening. This metallographic difference may be detectable in test piece with UT examination. UT examination of RPV cladding of Loviisa NPP in 2002 was one qualification case when solidification cracks have been applied in test pieces. Solidification cracks were welded into strip welded cladding as surface cracks with crack front shape and as subsurface cracks with elliptical shape. Before the commissioning of Loviisa unit 2, real solidification cracks were found in cladding as connected to slag lines, located in corner between weld beads, in the overlay layer welded on slag line (due to delayed solidification caused by slag line). Thin ligaments above cracks were broken when cladding was stressed (during pressure tests). Subsurface solidification cracks were applied in 2006 in test pieces of thick walled primary collector weld of a steam generator to simulate hot cracks in the weld volume. Deep openings into collector weld were machined and grinded from outside surface. Welded solidification cracks were produced into weld volume with same dendrite direction structure as in SAW weld of primary collector. Real subsurface hot cracks were detected in the volume of primary collector welds of steam generators during outage in 1980. Hot cracks were repaired in the upper weld and left in the lower weld of primary collector to be re-inspected with UT examination. Welding trials and real applications in Inconel buttering and weld started in 2006. Cracking of buttering and weld was simulated with solidification cracks. Short, deep transverse cracks in wide buttering were difficult to control on the bottom of opening, but both surface and subsurface cracks were produced. Feedback is not given whether the crack behavior resembles the service induced crack, whether the opening edge will be revealed or disturb detection and sizing of the crack. Circular, subsurface solidification cracks were produced into weld volume of Inconel test pieces in 2006 and 2007. The test piece with circular cracks will be inspected later this year. The influence of the direction of dendrite structure in Inconel weld will be found out. Development work will be continued in this field. Welding trials and real applications in ferritic base material and welds started in 2006. Two cracks in ferritic base material were also included into the national "Crack opening project" for getting feedback from UT response. Many applications can be seen for this kind of crack production in ferritic materials and welds. Underclad cracks, cracks near the buttering boundary and cracks in ferritic welds will be the application areas. Development work will be continued in this field. Simulation of planar defects Lack of fusion defects (LOF) are defined as specific defect type in many qualification cases of Loviisa NPP. LOF defect can be simulated by welding aid piece on weld groove with tilted orientation or between weld beads when a new test piece has to be manufactured (see NESC III /4/), or in the finished test pieces by opening weld area, fitting and welding aid piece and filling the opening carefully. LOF defect can be actually very tight and not detectable with x-ray. Aid piece can be welded partly transparent to ultrasound, if needed. Other an aid piece application is to simulate cracks by finishing aid pieces to shape of crack front, simulating surface crack or to elliptical shape of subsurface cracks, and by welding surfaces tightly together. Deep planar crack simulation with tilt and skew can be produced, especially in cases when accurate location on boundaries of buttering and cladding is needed. Large planar reflectors are suitable application case. Surface roughness and the opening between surfaces can be varied, if needed. Typical applications are ferritic and austenitic welds, dissimilar welds, buttering and cladding boundaries and positions between buttering and cladding layers simulating LOF defects and cracks in qualification cases. Reflectors are simulating quite effectively the real defects, partly being the worst cases for qualification. EDM reflectors EDM reflectors can be adapted in cases of finished test pieces as additional surface and subsurface reflectors, when exact positions, orientations and depth sizes are needed. Narrow notches can simulate LOF defects and also cracks with electrodes having crack front shape. Notches of PISC type A are no longer used due to wide opening and round reflecting surfaces (the last application was in the late 90's). Notches can be easily tilted and skewed, accurately positioned and produced to defined sizes. Shape and size of notch can be documented with replicas and by preserving the electrode used for finishing the notch. Even heavy and large size test pieces are possible for EDM production with special arrangements using proper equipment and skilled partner. Notches are used in base material, transverse to weld defects and dissimilar welds in some qualification cases of Loviisa NPP to complete the flaw population in addition to cracks. RECENT DEVELOPMENT WORK FOR QUALIFICATION Needs for qualification test pieces with relevant cracks and new application areas are considered annually. R&D project plan for crack development work includes crack trials needed for producing the pieces for actual and future qualification cases. Flexible planning enables crack trials, checking of NDE response and control of crack sizes with destructive test by using simplified test pieces. National crack opening project 2006-2007 This project started from feedback of difficulties with sizing. The idea was to manufacture cracks with different crack openings using thermally and mechanically produced cracks and welded solidification cracks. The aim was to adapt the qualified ultrasonic inspection procedures (manual and phased array) for sizing of cracks. Also other sizing techniques were used. The targets of national “Crack opening project” realized during 2006-2007 were to get - real data about UT response versus crack opening - information about crack opening values critical for sizing - evidence of sizing accuracy of crack simulations - information about features of crack simulations - experience from destructive examination of cracks using comprehensive reporting of cracks according to SKI recommendations /5/ - feedback for fabrication of crack simulations - certification for flaw manufacturing Project is a good example of flexible co-operation in Finland, when each participant is ready to use their resources into the same useful target. The summary report of the project will be finalized together with FNS, Trueflaw and Inspecta Certification (QB). Development work for Loviisa NPP specific applications ongoing Development work of FNS is concentrated in 2007 to widen the fabrication capacity and on training for producing deep fatigue cracks for qualification of sizing with phased array technique. Welding of solidification cracks into deep and narrow openings has been developed for Inconel weld and also for non-nuclear application. Producing cracks to buttering interface in ferritic material has been the main development task together with Trueflaw. The purpose is to manufacture test blocks for flange of control rod drive nozzle in reactor pressure vessel head and for dissimilar weld of steam generator. DISCUSSION Development work for producing different types of cracks in different types of materials and test pieces has given experience how to apply different flaw manufacturing processes. Natural cracks are rare at Loviisa NPP and test pieces with field cracks are not available, therefore test pieces with crack simulations are needed. The Utility is responsible to manufacture test pieces with flaws. Design of the test pieces for qualification cases of Loviisa NPP is carried out by QB and FNS together. Opinions and needs of Inspection Company and Utility are considered in case of open test pieces. No information of flaw population of blind test pieces is given to the Utility and Inspection Company. This will guarantee the confidentiality and independency of QB for executing qualification without influence of the Utility. One important duty of QB is to execute the fingerprint examination of the flaws to confirm the relevance of crack and flaw simulations for qualification case and confirm their sizes. QB will fingerprint the flaws with phased array UT examination and in many cases collect the inspection data for qualification of data analysts. Feedback from fingerprint examination and qualification trials to flaw manufacturers has been minor so far in Finland. One way to improve crack simulations is to give the chance to see the NDE response of fingerprint examination and assess all the details of scanned data together with QB, and also later to hear all feedback gained from qualification. In many qualification cases of Loviisa NPP, it was expected that Inspection Company has experience and evidence from cracks - unfortunately minor information has been available. In Finnish approach, the most effort has put to produce relevant blind test pieces. During blind trials, difficulties have been faced in many qualifications - proposed inspection technique has not been able to detect, characterize and size the cracks properly. National rules and ENIQ Recommendations contain check lists of QB for review of content of procedures and technical justification and for supervision of practical trials. Those check lists are tools for review - just filling the lists is not the main thing. The main task of QB is to get confidence and clear evidence, that the inspection system proposed by the Inspection Company will detect the defined flaws reliably and all the qualification targets are achieved. The national project “Crack opening project” was realized with thermal fatigue cracks and with crack simulations. Thermal fatigue cracks were produced with the controlled process like in the real test pieces. The crack simulations in the project were produced with variation of details in manufacturing work to get feedback from destructive testing for further improvement work and to see the UT response. CONCLUSIONS Theoretical and practical training and qualification of NDE personnel with cracks is the most effective way to improve the reliability of ISI inspections. Personnel qualifications and national trials have shown that the knowledge of NDE personnel about the features of the cracks should still be improved. With better understanding of cracks, the detection and sizing will be improved. The Qualification Bodies have an important mandate when taking care of qualifications. QB should review the inspection system and assess the capabilities of the proposed technique. When problems are faced during practical trials, QB should react fast and demand the improvement of inspection technique. Otherwise the ISI might be utilized with unsatisfactory performance and purpose of the qualified and reliable ISI would be ruined, and investments on safety would be useless. More attention is needed for open test pieces. The real improvement can be achieved only when the Inspection Company can test the capability of inspection system and optimize the inspection technique by using relevant cracks and reflectors of open test pieces. The effective way for the manufacturer to improve crack simulations would be the chance to see the NDE response and results of fingerprint examination of QB and assess all the details of scanned data together with QB. Also the feedback from qualification should be reviewed together with QB to get all the feedback from the cracks used for qualification. The targets of national crack opening project have been achieved. Real data for evidence and decision making for future has been received. Destructive testing of cracks is reported using recommendations /5/ and preparedness for future investigations of cracks has been achieved. Thermally produced cracking by Trueflaw was certified by QB. Advanced crack opening measurement program was also created by Trueflaw. The same software would be used for crack documentation and to scan crack features from representative old failure analyses. Crack features of representative service induced cracks should be analyzed and measured from earlier failure reports to increase knowledge and to be able to improve crack production and simulate field cracks better in test pieces. New failure reports should include all the necessary information, needed for qualification activities. REFERENCES 1) Radiation and Nuclear Safety Authority (STUK), Guide YVL 3.8, Nuclear power plant pressure equipment, In-service inspection with non-destructive testing methods, Helsinki, 22 September 2003 2) ENIQ Report nr.31, European Methodology for Qualification of non-destructive testing - third issue, Petten, EUR 22906 EN, 2007 3) Inspecta Certification, Guides for Qualification of periodical inspections of nuclear power plants, Espoo, 2007 http://www.inspecta.fi/sfs/english/certification_services/person_certification.php?m=m3 4) NESC III Inspection Task Group, Destructive Examination Report, RRT Dissimilar Metal Weld Component, Petten, EUR 22607 EN, 2006 5) Jan Wåle, SKI Report 2006:24, Crack characterization for In-service Inspection Planning - An update, May 2006

THE INFLUENCE OF THE CRACK OPENING IN THE UT INSPECTION QUALIFICATION. Tapani Packalén, Juha Sillanpää, Inspecta Certification Oy Mika Kemppainen, Iikka Virkkunen, Trueflaw Ltd. Raimo Paussu, Fortum Nuclear Services Oy ABSTRACT The character of cracks is an important parameter in manufacturing of test pieces for UT qualification. What is the influence of crack opening in the sizing of the crack? In these studies we (Trueflaw and Fortum Nuclear Services) have manufactured 22 cracks by different production methods. Capes of cracks have been manipulated for different values. Sizing of the real dimensions of cracks has been made by destructive tests. These values are basis for the evaluation of the accuracy of different techniques. This presentation will focus on NDT aspect. The purpose of these studies is to find dependency in test results and opening of cracks. We also try to find out differences in accuracy between various techniques. Are some techniques more sensitive for the influence of crack opening? The test pieces have been inspected by nine inspectors from Finnish NDT vendors (Polartest, Inspecta, VTT). Inspectors used different techniques and procedures for flaw size evaluation. Three inspectors made sizing (size defining) by traditional manual method. They used different techniques and probes. The selection of height of defect was more or less experiential. Two inspectors used manual phase array technique and four inspectors made sizing (size defining) by mechanizesd phase array technique. All inspections were blind tests. In the evaluations of results it has been concentrated in the used techniques and procedures. Influences of human elements have been left totally out. INTRODUCTION One important detail in qualification process is to select right defects to the test specimens. Qualification defect types must be realistic compared to postulated defect types. Influence of crack opening has been one explanation for unsuccessful NDT- results. Is it true reason? On the other hand, it is unfair to inspectors if there are defects what is unrealistic to detect. When qualification body selects the defects to be used in qualification project, it must have know-how about behavior of ultrasonic with different crack types. In this project we hope to find out some responses to these questions. AIM OF THE WORK The aim of the work was to find out relation between crack opening and ultrasonic response. In these studies we had three different types of crack and three different types of ultrasonic techniques. We hope to find out limitations of different techniques and limitation for opening when crack tip can’t be detected. EXPERIMENTAL This study includes NDT results from four companies and nine inspectors. All inspection companies used their own NDT equipment. Used inspection methods included manual inspection, phased array inspection and mechanized inspection. Manual inspections were conducted using Finish ISI procedure ISI511 and conventional flaw detectors units (Krautkrämer USN series, Epoch IV). Used probes were WSY60-2, WSY70-2, ADEPT60, MWK55-2 and MWB45-5. Phased array inspections were conducting using OMW-610.424 procedure with Omniscan units. In mechanized inspection were used the same probes as in manual inspection. The analysis was made from pre-collected data with possibility to use B- and C-screens for the evaluation of crack height. All the cracks were also evaluated by scanning acoustic microscope (SAM). These measurements were performed with smaller samples cut near the crack face. Scanning measured reflection and shadowing of cracks. Frequencies used were 5 MHz and 10 MHz (Figure 1). Six test specimens were produced for this study. Three different flaw types were used: solidification cracks, welded fatigue surfaces and thermal fatigue cracks. The samples and flaw types are documented in detail by Kemppainen et al. [2]. RESULTS All results y = 0,7382x + 0,9077 R2 = 0,3391 0 2 4 6 8 10 12 14 2 3 4 5 6 7 8 9 10 True depht M es ur ed d ep ht Figure 1. All results 0,0 1,0 2,0 3,0 4,0 5,0 6,0 7,0 8,0 9,0 10,0 B385 B389 B364 B367 F103 B405 F101 B396 B395 S102 B398 S101 F104 S103 F102 S104 Crack D ep ht m m True depht ISI511 Mekan. Phasearray Figure 2. Average depths of cracks measured by different techniques Crack Crack Opening Depth (mm) Deviation (mm) Type Number Tip (µm) True ISI Mechan Phase Total ISI Mechan Phase Thermal B405 4,8 5,7 6,0 5,3 5,9 1,7 3,6 0,3 0,9 B395 0,8 6,2 6,1 5,2 6,1 1,8 3,6 0,9 0,6 B396 0,1 5,8 5,0 5,1 6,2 1,4 2,3 0,4 0,6 B398 5,5 6,5 5,0 5,6 6,3 1,1 2,8 0,3 0,5 B364 1,5 4,0 3,5 3,5 4,6 1,1 1,3 0,6 1,1 B367 2,2 4,3 2,3 3,5 4,6 1,4 0,6 0,9 1,4 B385 0,4 3,2 6,0 3,2 3,6 0,7 0,9 0,2 0,5 B389 0,4 3,3 2,2 3,1 3,4 0,7 0,8 0,1 0,5 Fatique F101 3,9 5,7 5,1 4,1 5,4 1,0 0,9 1,1 1,0 F102 0,5 8,1 7,0 6,2 6,8 2,6 3,2 3,2 1,8 F103 0,3 5,4 5,0 6,0 5,9 0,9 1,1 0,8 0,5 F104 0,7 7,8 5,0 6,9 6,4 1,6 0,8 2,8 0,5 Solidif. S101 9,1 6,9 6,9 5,8 5,9 1,9 2,5 1,7 1,7 S102 12,5 6,4 5,1 5,0 5,5 1,6 2,4 0,8 1,6 S103 14,9 8,0 6,1 6,3 6,7 2,0 2,7 1,5 2,1 S104 5,7 9,0 6,8 6,9 6,5 2,4 2,4 2,8 2,3 Table 1. Measured average crack depths and deviation from true value Influence of crack opening to deviation 0,0 2,0 4,0 6,0 8,0 10,0 12,0 14,0 16,0 B3 96 F1 03 B3 89 B3 85 F1 02 F1 04 B3 95 B3 64 B3 67 F1 01 B4 05 B3 98 S1 04 S1 01 S1 02 S1 03 O pe ni ng µ m 0,0 0,5 1,0 1,5 2,0 2,5 3,0 D ev ia tio n m m Deviation Opening Figure 3. Opening of crack tip and deviation (mean value) 0,0 20,0 40,0 60,0 80,0 100,0 120,0 140,0 160,0 B3 89 F1 02 B3 96 B3 85 B3 95 F1 01 B4 05 B3 64 S1 02 B3 98 S1 01 B3 67 F1 03 F1 04 S1 03 S1 04 O pe ni ng µ m 0,0 0,5 1,0 1,5 2,0 2,5 3,0 D ev ia tio n m m Deviation Opening Figure 4. Opening of crack tip area and deviation (mean value) Deviation on different crack type 0,0 0,5 1,0 1,5 2,0 2,5 3,0 B3 64 B3 67 B3 85 B3 89 B3 95 B3 96 B3 98 B4 05 F1 01 F1 02 F1 03 F1 04 S1 01 S1 02 S1 03 S1 04 De via tio n Figure 5. Average deviation on different crack types. (thermal magna, fatigue yellow and solidificion blue) Figure 6. Correllation between opening and UT amblitude in crack tip. Left crack B396 and right F102. Ficure 7. NDT size measurements from different cracks (B396 thermal, F102 fatigue and S104 soludification) Figure 8. Reflection of thermal fatigue cracks in acoustic microscope. Correlations between the crack opening and the ultrasonic response. DISCUSSION Crack tip opening affects the amplitude of the crack tip signal (Figure 6). Smaller opening gives smaller signal amplitude. In figure 8 no correlation can be seen with the SAM signal amplitude and crack face opening measured from metallographic examination. . Flaw sizing results shown in Figures 2 and 3 are in line with general expectations. Small defects were sized quite well and deep defects were undersized. The same behavior is seen in many studies. In Table 1, it can be seen, that increasing crack size is associated with increasing sizing error (undersizing). In order to understand the reason for this, it is necessary to discus the sources of sizing error present in this study. All sizing techniques used utilize crack tip signal to determine crack size. The error in sizing may thus come from two distinct sources. Firstly, when the inspector has correctly identified the crack tip, there is some error associated with the procedure of correctly locating the source of the signal within the sample. For present discussion, this error is termed "measurement error". It is typically reasonably small, in the order of ±1 mm and comes from sources like error in calibration, resolution of the device used etc. The error may be assumed to be centered around the true location of the signal and to be normally distributed. This error is not affected by crack characteristics. The other possible source of error is the error in correctly identifying the crack tip signal. For present discussion, this type of error is termed "interpretation error". When the inspector fails to correctly identify the crack tip signal, he is, in fact locating the singal from some other source. This source may be signal from other parts of the crack, besides the tip or from microstructural noise (and not from crack at all). This error can not be assumed to be centered around the true depth of the crack and nor can it be assumed to be normally distributed. The crack characteristics have influence on the interpretation error. Firstly, the strength of the crack tip signal affects the likelihood of misinterpretation. The stronger the tip signal is, the easier it is to correctly distinguish from microstructural noise etc. The crack opening, and crack tip opening in particular affect the strength of the crack signal. Secondly, the artificial flaws may have features other than the deepest tip that give strong signal and thus may easily be misinterpreted as the crack tip. For example, strong twists or branches with secondary tips may act this way. It should be noted, that other factors affect the interpretation error, as well. For example, if the inspected component contains welds or other strong sources of microstructural noise, the likelihood of interpretation error increases. Consequently, to correctly asses the effect of crack characteristics on the sizing reliability, one needs to asses whether the error includes interpretation error and if so, what is it, that the inspector has, in fact, located in stead of the crack tip. To this end, all the NDT results from these flaws (as reported in 2) were plotted on the cross section images from the cracks. With this visualization, it was possible to determine, not only which results included interpretation error but also, whether the inspectors in these cases had studied random noise or some other feature of the crack. When reading these images, it should be noted that the thermal fatigue cracks were introduced to base material so the microstructural noise was very low. In contrast. All solidification cracks and welded fatigue surfaces were introduced near or within weld material and thus had much higher level of noise. The analysis of these images is revealing. For thermal fatigue cracks, both the mechanized and phased array inspections (green and blue lines) have correctly identified the crack tip. The error is small and centered around the true crack depth. However, the most of the manual inspectors (red lines) have not correctly identified the crack tip even in this simple geometry and in absence of weld. The error signals in manual inspection cannot be linked to any feature of the crack and thus it is assumed, that the inspectors have located signals from microstructural noise. In crack 218AGB389, the two erroneous manual inspection results are near a place, where the crack opening suddenly decreases, but there's not enough data to determine whether this is an isolated case. In case of welded fatigue surfaces, there are errors too big to be attributed to measurement error in all inspection types (manual, mechanized and phased array). This may be attributed to the weld present, which increases noise level and thus the likelihood of interpretation error. In this case, however, the erroneous interpretations are connected with strong twists in the flaw. Thus, it may be concluded, that when the inspectors have failed to correctly identify the crack tip they have, in most cases, located signal coming from a twist in the crack. This lead, on average, to undersizing of the flaws. Curiously, it also decreases the average error in cases, where the inspetor has failed to correctly identify the crack tip. This can be attributed to the fact, that welded fatigue surfaces (with the exception of 139AHF103) have a characteristic twist quite near (within 2 mm of) the true crack depth which many inspectors have, incorrectly, interpreted as the crack tip. Consequently, although the event of failing to correctly determine the crack tip is much more common in welded fatigue flaws than in thermal fatigue cracks, the average error in sizing is nearly the same (in case of mechanized and phased array inspections) or smaller (in manual inspection) that in case of thermal fatigue flaws. In case of solidification cracks, yet another behaviour can be observed. The flaws are characteristically heavily branched and have numerous similar crack tips, which are metallographically similar. Consequently, the main source of interpretation error is not that the inspector has located signals from random noise or signals from twists or other crack features. In stead, the inspector has picked the wrong (that is, other than the deepest) tip to analyze. The tip that the inspector has picked is, however, likely to be near the true deepest tip, and thus the average error in this case also reasonably small and the inspectors are much more likely to undersize than oversize the crack. For further studies, it would be interesting to investigate all crack types in more realistic components and with welds, where microstructural noise would be significant. This might reveal more about the influence of crack characteristics to in service inspection accuracy. CONCLUSIONS The following conclusions can be drawn from this study: - Crack opening affects the amplitude of the crack tip signal. Smaller opening gives smaller signal amplitude. - With decreasing signal amplitude, the likelihood of misinterpretation in identifying crack tip signal increases. - With increasing noise amplitude (for flaws near weld), the likelihood of misinterpretation in identifying crack tip signal increases. - For thermal fatigue flaws included in this study (in base material), mechanized and phased array inspections correctly identified the crack tip signal and sizing accuracy is thus good. - In manual inspection (for all flaw types) the likelihood of misinterpretation in identifying crack tip signal is high REFERENCES 1) Kemppainen, M., Realistic artificial flaws for NDE qualification – novel manufacturing method based on thermal fatigue. Dissertation for the degree of Doctor of Science in Technology. Espoo, Finland, 2006. (Available online from: http://lib.tkk.fi/Diss/2006/isbn9512282631/) 2) Mika Kemppainen, Iikka Virkkunen, Trueflaw Ltd. Raimo Paussu, Fortum NuclearServices Oy. Tapani Packalén, Juha Sillanpää, Inspecta Certification Oy. Importance of crack opening in U- inspection qualification

IMPORTANCE OF CRACK OPENING IN UT INSPECTION QUALIFICATION Mika Kemppainen, Iikka Virkkunen, Trueflaw Ltd., Finland Tapani Packalén, Juha Sillanpää, Inspecta Certification Oy, Finland Raimo Paussu, Fortum Nuclear Services Oy, Finland ABSTRACT Good, representative cracks are needed for reliable qualification. The qualification cracks should be representative of the expected in-service cracks in terms of essential characteristics that affect the inspection in question. For UT inspection, one such important characteristic is the crack opening. In order to find out the importance of crack opening to UT inspection qualification and to quantify and compare the available artificial flaw production technologies, a research project "Avauma" was carried out. The project included participants from the Finnish qualification body (Inspecta Certification), the Finnish utilities (TVO and Fortum), the Finnish NDT vendors (Polartest, VTT, Inspecta), and the Finnish flaw producers (Trueflaw, Fortum Nuclear Services). In the project, significance of crack opening on different ultrasonic techniques and inspection procedures is quantified, and conclusions drawn for further development of Finnish qualification practice. Cracks used in the project were produced by state of the art crack producing techniques capable of producing controlled, realistic cracks with known size and opening. Applied crack producing techniques are widely used in Finnish qualification processes, and similar techniques are considered for the qualification processes used for new plants, as well. 22 cracks in total were produced with different crack openings using three different crack production techniques. For some of the cracks, the opening was modified after the production with thermal loading to obtain wider range of crack openings. The samples were inspected by different ultrasonic techniques, procedures and by various inspectors. Finally, the samples were destructively studied to reveal the metallographic characteristics of the flaws. The flaw opening along the crack depth and other flaw characteristics were measured from the metallographic samples. In the present paper, the metallographic crack properties are documented and discussed. The paper presents the quantitative metallographic results together with selected NDE results. Results of the destructive testing are compared to the reported NDE results. The significance of the crack opening to the performance of different ultrasonic techniques and various inspection procedures is assessed. Metallographic results are also compared to the corresponding results from service-induced cracks, as reported in the open literature. The capabilities of different crack manufacturing techniques are analysed and discussed in terms of crack opening and its controllability, resulting crack characteristics, and the overall ability to achieve realistic artificial cracks with variable crack opening. INTRODUCTION Representative cracks are needed for reliable qualification. The qualification cracks should be representative of the expected in-service cracks in terms of essential characteristics affecting the inspection in question. Also, the artificial cracks should not contain any such features, that are not present in the expected in-service cracks and that would affect the performance or reliability of the inspection [1]. The previous studies on different flaw manufacturing techniques (refs. 2 and 3) underline the need for better understanding of the effects of non-similarities between artificially produced reflectors and service-induced cracks to NDE response in qualification. Furthermore, those studies expressed need for having better flaw manufacturing techniques not exhibiting weaknesses shown, e.g., in the flaws used during the first ENIQ pilot study [2]. After these studies, a new technology has been developed to meet this requirement [4]. For UT inspection, one important characteristic is expected to be the crack opening. Crack opening, in general, affects the UT response obtained from the crack. Greater opening is associated with stronger signal. It has been reported, that with very small opening, the crack may become at least partly transparent to ultrasound [5]. The crack tip signal, in particular, is sensitive to crack tip opening. Crack tip signal is crucial in many sizing techniques and hence it is important to faithfully reproduce the crack tip signal for sizing qualification. Though there have been several studies on the effect of crack mouth opening to NDE response (e.g., references 5, 6, and 7), much less studied is the effect of the opening distribution of the whole crack, and opening at the crack tip, to the NDE response. It is also essential to compare such studies to information published about the opening of real, service-induced flaws to have a connection to the qualification and its targets. Typical crack opening values in service Natural service-induced cracks exhibit a wide variety of different openings. This results from the wide variety of different loading conditions and mechanisms effective in operational conditions inducing cracks, as well as the material properties of the material in question. Wåle [8] has measured crack mouth and tip openings from numerous service-induced cracks. These values are summarized in Figures 1 and 2, respectively. It should be noted that bigger cracks with wider opening are more likely to be found. Consequently, the data is likely to be biased and the population shows too great opening values. Furthermore, these measurements were done from images in existing failure analyses. The primary objective of such analyses has been other than crack characterization for NDE and, consequently, all crack characteristics may not be readily observable from the images present. In particular, the crack tip radius is generally not well described in failure analyses. Crack opening values for different artificial flaw manufacturing techniques have been measured from images present in the open literature [9] by Virkkunen et al. [11]. These measurements represent values from numerous flaw manufacturers using crack simulation techniques and solidification cracks. This data is also summarized in Figures 1 and 2. When reading the data, it should be noted, that measurement of crack opening values from the images was rather difficult and, hence, the accuracy of the numbers is limited. Figure 1 Crack mouth opening (CMO) values of service-induced cracks and artificial cracks, in µm. The bars show the range between minimum and maximum values reported. The dots connected by line show the average value reported. Note, that the average value tends to be closer to minimum than maximum. Values for service-induced flaws were taken from [8]. Values for artificial flaws were measured from [9]. Figure 2 Crack tip opening radius (CTO) values of service-induced cracks and artificial cracks, in µm. The bars show the range between minimum and maximum values reported. The dots connected by line show the average value reported. Note, that the average value tends to be closer to minimum than maximum. Values for service-induced flaws were taken from [8]. Values for artificial flaws were measured from [9]. AIM OF PRESENT WORK The aim of present work was to compare available artificial flaw production technologies, produce solid metallographic data on the characteristics of different artificial flaws and to find out the importance of crack opening to UT inspection qualification. MATERIALS AND METHODS Three different flaw manufacturing techniques were used in this work for flaw production, two of them included welding. The aim was to produce artificial flaws with different techniques commonly used in today's qualification samples. The produced samples were inspected by popular inspection methods currently used in Finland. Finally, the samples were destructively examined and relevant crack characteristics measured. The results of the performed study are reported in three publications. The current work reports the metallographic results. Packalén et al. [10] analyse the NDE results. Paussu et al. [12] discuss the project from the utility perspective. Flaw production techniques Three flaw manufacturing techniques were used in this study. Two of these are welding- based: "solidification cracks" and "weld joined fatigue surfaces". Third technique "thermal fatigue crack" does not include welding. Solidification cracks are produced as follows: an excavation is ground in the place, where crack is to be introduced. The excavation is filled by welding with suitable filler material and welding parameters to create a solidification crack in the middle of each weld pass. Finally, the surface of the weld is ground to the final shape. The final depth is controlled by the depth of the opening ground on the surface, and control of the solidification process. The target depth is the depth of the opening. However, there is variance in the cracking of the first weld pass when filling the opening. Variance in the solidification process and possible machining give rise to the characteristic depth tolerance for the process. The tolerance given by the manufacturer is ±1.0 mm. The flaw type "weld joined fatigue surfaces" is produced in three steps with an assistance of an aid piece. First the aid piece is welded on the side wall of the prepared weld groove, and manually loaded to create a fatigue crack. The loading is continued until separation. The created fatigue surfaces of the aid piece are then manually ground to the desired shape. Shaped fatigue surfaces are welded tightly together and back to side wall, and welding of the actual joint is completed. Finally, the surface of the weld is ground to the final shape. The final depth is determined by the size of the shaped surfaces and the fixing welding procedure. The target depth is the depth of the shaped fatigue surface subtracted by the weld pass penetration of the fixing weld and machining thickness. Depth tolerance of this flaw type comes from the variations of the penetration of the joint weld pass in the edge of the fatigue surfaces and machining. The tolerance given by the manufacturer is ±1.0 mm. Flaw production by thermal fatigue is done in-situ with ready-made sample. The cyclic thermal fatigue loading is induced locally by alternating heating and water spray cooling, as described by Kemppainen [4]. The loading is based on pure thermal loading and there is no welding, machining, or any other mechanical treatment applied. The final depth is based on the applied process parameters (strength of the loading cycle, and amount of applied total cycles). Appropriate parameters are verified in advance by destructive validation, i.e. a crack is produced to similar material and destructively tested to reveal its true size. The best estimate depth is the statistical average depth from several validations. The tolerance comes from the statistical variance of validation results. The tolerance given by the manufacturer is ±1.0 mm. The opening of thermal fatigue cracks is based on variation of the applied process loading parameters. In this work, cracks with different openings were produced by applying different combinations of post-production loading sequences. The opening manipulation was done for six flaws with qualitative aim of seeing the effect of different combinations of loading parameters. Specimens and materials Solidification cracks and weld joined fatigue surfaces were produced to an austenitic stainless steel plate specimen with a X-groove butt-weld in the middle of the plate. Plate dimensions were 400 mm x 300 mm x 30 mm (length x width x thickness). In addition, two solidification cracks were produced to a ferritic base metal plate with dimensions of 400 mm x 300 mm x 30 mm (length x width x thickness). Thermal fatigue cracks were produced to two ready-made austenitic stainless steel base material plates with dimensions of 300 mm x 300 mm x 20 mm (length x width x thickness). Four cracks were produced to each plate, the other plate having cracks with 3 mm and the other one with 6 mm target depths. Cracks were produced in the centreline of the specimens. Destructive examination The produced flaws were destructively examined to reveal their characteristics. The measurements were conducted from high-resolution digital images using semi-automatic image analysis software "CrackMeasure" written by Trueflaw. The characteristics measured were similar to those measured by Wåle [8]. However, with automated measurement program and with metallographic samples made for this purpose in particular, much more detailed and accurate results were obtained. The measurements were done for the surface of the sample and for cross-section. Metallographic sample was prepared for each crack. The samples were manufactured with special care taken to ensure that crack opening was not altered in sample production. The cross-sectional sample was taken close to the deepest location of the flawed samples. Cross- sections were polished with normal metallographic procedure, and etched with appropriate etchant. Both the as-polished and etched surfaces were measured, because etching of the cracked area always rounds the corners of the fracture surface making the opening look bigger in the micrographs. Finally the cross-sectional samples were bent open to reveal the true depth and shape of the flaws and the opened surface was photographed. The destructive analysis was supervised by Inspecta Certification in order to ensure impartial and objective measurements. RESULTS Metallographic data The measured opening values as well as target and destructively revealed true depths for all the produced cracks are tabulated in Annex 1. Furthermore, the best estimate values as given by the manufacturer, are given for the thermal fatigue cracks. For opening, four values are reported: near mouth opening, middle opening, near tip opening and tip opening. The three first are reported as average measured opening from 0.5 mm length. Tip radius measurement was done by fitting a circle to high-magnification image of the crack tip and measuring its radius. The specified tolerance for all flaw types for length and depth were ±2 mm and ±1 mm, respectively. The actual correspondence between the specified, best estimate, and true depth values for the thermal fatigue cracks is presented in Figure 3. Figure 3 Comparison of specified, best estimate and true depth of the flaws, together with tolerance range given by the manufacturer. Opening manipulation of thermal fatigue cracks Post-production manipulation of the opening was performed for six thermal fatigue cracks. Different loadings were used to get different changes in the crack openings and the result of opening manipulation was measured on the surface opening. Surface opening values for all the produced thermal fatigue cracks are tabulated in Table 1. A graphical example of the comparison between the surface opening profiles for the as-produced and as-manipulated conditions of one crack is shown in Figure 4. Figure 4 Surface opening of a thermal fatigue crack (239AGB398) in as-produced condition and after opening manipulation. Table 1 Values for surface openings for thermal fatigue cracks in as-produced and as-manipulated conditions. Flaw ID As-produced (µm) Opening modified (µm) 186AGB364 60 78 193AGB367 67 92 211AGB385 70 - 218AGB389 46 65 235AGB395 123 127 237AGB396 51 - 239AGB398 82 139 250AGB405 130 103 Another aspect on the crack opening is the opening distribution in the depth direction of the crack. The difference in depth direction of an as produced and opening manipulated crack was revealed by comparing the measured cross-sectional openings of two different cracks, as the example shows in Figure 5. Both cracks were produced with the same process, the other one was left in as-produced condition while the other one was manipulated to have bigger opening. Figure 5 Opening profiles in cross-section of two cracks in as-produced and manipulated conditions, (237AGB396 and 239AGB398, respectively). Metallographic characteristics of flaws The metallographic images of the three different flaw types manufactured in the project reveal clear differences in their typical flaw appearances. Examples of cross-sectional images of solidification flaws, weld joined fatigue surfaces, and thermal fatigue cracks are shown in Figures 6, 7, 8 and 9, respectively. Two examples are shown for the weld joined fatigue surfaces. These images are selected to show certain features and they do not exhibit cross- sections typically produced by the welding-based techniques, as declared by the manufacturer. Figure 6 Cross-sectional image of a solidification crack (flaw 139AHS103). Figure 7 Cross-sectional image of a weld joined fatigue surface (flaw 139AHF103). Figure 8 Cross-sectional image of a weld joined fatigue surface (flaw 139AHF102). Figure 9 Typical cross-sectional image of a thermal fatigue crack with opening modified to larger value (crack 239AGB398). DISCUSSION Characteristic features of used flaw manufacturing techniques Three flaw manufacturing techniques were used in this study. Two of these are welding- based: "solidification cracks" and "weld joined fatigue surfaces". Third technique "thermal fatigue crack" does not include welding. The flaws were produced by two companies, Fortum Nuclear Services (solidification cracks and weld joined fatigue surfaces) and Trueflaw (thermal fatigue cracks). Thermal fatigue crack production technique is exclusively used by Trueflaw. The other two techniques are widely used by other flaw manufacturers, besides Fortum Nuclear Services. Consequently, it is of interest to compare the flaws analyzed here with those produced by other manufacturers using the same technique. Lemaitre et al. [9] and more recently Iacono et al. [3] have reported cross-section images from flaws manufactured by numerous artificial flaw production companies. Comparison with these publications reveals, that the features reported in this study represent the typical features of the used manufacturing techniques in general, and the results are applicable for all flaws manufactured by the studied techniques. In the cross-sectional images of the destructive testing results, it is seen that weld joined fatigue surfaces are located on one side of the joint weld and have weld metal on both sides resulting from the welding of the small aid piece to the wall of the original weld groove. Crack tips are melted by the welding and they exhibit relatively small radius. In the shallow surface layer these flaws have very small opening caused by the machining process. Solidification flaws are located in the middle of the weld passes. These flaws have relatively large openings through their whole depth. Furthermore, solidification flaws produced in austenitic stainless steel have major branches and unbroken ligaments causing the special aspects in their NDE response. In the shallow surface layer the solidification flaws have very small opening caused by the machining process. Thermal fatigue cracks are located in the base metal without introducing any other changes in the material. Especially, there is no welding done during the process. These cracks are tight, have natural propagation through the microstructure, and have small crack tip radiuses. The applied opening treatment of six of the cracks had changed their opening markedly. The opening profile in depth direction is natural exhibiting largest opening at the surface and smallest at the crack tip. Deviation between the best estimate and true depth values from the destructive testing reveal that thermal fatigue cracking process is well in line with the by manufacturer given characteristic tolerances of the technique. Comparison between different flaw types and service-induced cracks One of the aims of this study was to compare the characteristics of artificial flaws made with different techniques with the actual, service induced flaws as described in the Wåle report [8]. Figure 9 and 10 show the literature data together with data produced in the "Avauma" project. Figure 10 Crack mouth opening (CMO) values of service-induced cracks and artificial cracks, in µm. The bars show the range between minimum and maximum values reported. The dots connected by line show the average value reported. Note, that the average value tends to be closer to minimum than maximum. For Thermal fatigue cracks, average values are not reported. For this flaw type, the opening can be produced to specified value within the range given. Values for service-induced flaws were taken from [8]. Values for "ENIQ" flaws were taken from [11]. Figure 11 Crack tip opening radius (CTO) values of service-induced cracks and artificial cracks, in µm. The bars show the range between minimum and maximum values reported. The dots connected by line show the average value reported. Note, that the average value tends to be closer to minimum than maximum. For Thermal fatigue cracks, average values are not reported. For this flaw type, the opening can be produced to specified value within the range given. Values for service-induced flaws were taken from [8]. Values for "ENIQ" flaws were taken from [11]. Note also, that the crack-tip-radius values measured in this study were taken from polished (not etched) high-magnification images taken especially for this purpose. In contrast, the measurements from [8] and [11] are taken, in general, from etched lower magnification images. Consequently, the values taken from [8] and [11] may show greater values than would be correct. The above figures show the correspondence of the different artificial flaws to real, service-induced cracks. In addition, different flaw manufacturing techniques are divided to "ENIQ" flaws and flaws produced in this study. In the Figure 10 the range and average values for the "ENIQ" flaws are bigger than values for the most of the service-induced cracks. "ENIQ" flaws included both solidification flaws and weld joined fatigue surfaces. Solidification flaws produced in the current study in Figure 10 follow the typical range and average values of the "ENIQ" flaws. In the graphs the range and average values for the weld joined fatigue surfaces show roughly 50% smaller values than for the solidification flaws, but still exhibit bigger values than most of the service-induced flaws. The opening range of thermal fatigue cracks produced in this study is in the area of most of the service- induced cracks. There is no average value given for the thermal fatigue cracks because the openings can be manipulated to specified value. Figure 11 shows that the range and average values of the crack tip opening for the "ENIQ" flaws lay markedly in bigger values than for the service-induced cracks. However, for the flaws produced in current study the crack tip values for the solidification cracks show only a bit higher values than for the service-induced flaws. For the weld joined fatigue surfaces and thermal fatigue cracks the values are clearly smaller than those reported for the service-induced cracks. This difference can be attributed to the different measurement processes used by different authors, as described in the caption text of Figure 11. The effect of flaw characteristic on the reliability of sizing The crack characteristics may influence sizing reliability by affecting the likelihood of correct identification of crack tip signal. The stronger the tip signal is, the easier it is to correctly distinguish it from microstructural noise. The crack opening, and crack tip opening in particular, may affect the strength of the crack signal. Secondly, the artificial flaws may have "false tip candidates", i.e. features other than the deepest tip that give tip-like signal and thus may easily be misinterpreted as the crack tip. For example, strong twists or branches with secondary tips may act this way. As shown in Figures 6-9, weld joined fatigue surfaces and solidification cracks may show features that can act as "false tips". In contrast, thermal fatigue flaws do not exhibit such false tips. The effect of these features to ultrasonic inspection performed is analyzed in more detail in [10]. The used flaw manufacturing techniques caused characteristic opening profiles both in the surface and depth direction. As a result of the nature of the manufacturing process, the solidification flaws have the greatest opening values of all the flaws. Opening of solidification flaws cannot be manipulated and it is result of the solidification process. Also for weld joined fatigue surfaces, there is limited possibility to control the resulting flaw opening values. The production process for thermal fatigue cracks allows control of the opening profile both in the surface and depth direction. This capability was demonstrated in this study. One aim of the work was to reveal the effect of different openings of the manipulated thermal fatigue cracks to obtained NDE response. However, as detailed in the other publication [10], other features present in the samples proved to dominate the ultrasonic response and no distinct effect of the crack opening could be observed. In order to ensure relevance of training and reliability of qualification, the selection of flaws used should be based on good knowledge of flaw characteristics. Any potential "false tips" etc., which may lead to misinterpretation, should be noted and accounted for. Sizing based on "false tip" does not correspond to actual in service inspection. In particular, if there are false tips near the true tip (such as the examples show in Figures 6 and 8) erroneous interpretation based on the false tip may give sizing result close to the true crack size. In such a case, apparently accurate sizing of the flaw with false tip does not indicate accurate sizing in actual in-service inspection. Consequently, flaws with false tips may not be used reliably in qualification or performance demonstration, as they may give overly optimistic impression of the inspection performance. Future studies The results of this work revealed clear needs for future studies. Such studies should include the effect of weld noise to tip signal detection from thermal fatigue cracks, and crack detection reliability of different inspection techniques. CONCLUSIONS The following conclusions can be drawn from the study: 1) Thermal fatigue cracks can be produced, with the stated tolerances, without causing any additional disturbances. This was shown both by the metallographic images and NDE results. 2) All the flaw types included in this study produce tip signal that can be identified by at least some of the ultrasonic techniques used. None of the cracks were so tight as to become completely transparent to ultrasound. 3) The weld joined fatigue surfaces may have twists that give signal that was erroneously interpreted as the crack tip. This feature of the flaw type results in tendency to undersize the crack and decreased average error (when compared to cases when such "false tip" was not available and inspectors have instead interpreted microstructural noise as the crack tip). 4) The solidification crack may have branching and multiple crack tips. This feature of the flaw type results in tendency to undersize the crack and decreased average error (when compared to cases when such "alternate tip" was not available and inspectors have instead interpreted microstructural noise as the crack tip). 5) Thermal fatigue flaws showed no "false tips" to be identified. Consequently, mechanized and phased array ultrasonic inspections could locate the correct crack tip in the simple geometry, and low noise base material sample. In contrast, the manual inspections often failed to correctly identify the crack tip even in this simple case, but analysed signals from random microstructural noise. 6) The effect of the weld material was seen for the solidification flaws and weld joined fatigue surfaces. In this study the thermal fatigue cracks were located in the low noise base material, but if they had been in the weld material, it may be speculated that the mechanised and phased array techniques would have had difficulties in detecting signal from the crack tip. 7) Furthermore, opening manipulation of thermal fatigue cracks showed marked effect on the opening profiles as seen both in metallographic images and some NDE results. 8) The flaws included in this study give a representative sample of artificial flaw production techniques and artificial flaws in use today. 9) The performance of the manual inspection techniques is poorer than was expected, especially considering the simple, low noise, component in question. 10) The "false tip" features that may be present in solidification cracks and weld joined fatigue surfaces may lead to misleading results if used in qualification or performance demonstration. 11) Future studies could include studying the effect of weld noise to thermal fatigue crack tip signal detection and overall detection capability of different techniques. REFERENCES 1) Wüstenberg, H. and Erhard, A., "Problems with Artificial Test Reflectors at the Performance Demonstration of Ultrasonic Inspections" 6th European Conference on Non Destructive Testing, Nice, pp. 741-746, 1994 2) ENIQ, "Final Report of the First ENIQ Pilot Study", ENIQ Report nr 20, EUR 19026 EN, European Commission, JRC Petten, 38 p., 1999. 3) Iacono I, Eriksen B, Mendes J, Metten L, Lofaj F, Seldis T and Wallendorf M, NESC III Inspection Task Group, Destructive Examination Report, RRT Dissimilar Metal Weld Component, EC DG JRC Institute for Energy. EUR 22607 EN ISSN 1018-5593, 2006. 4) Kemppainen, M., Realistic artificial flaws for NDE qualification – novel manufacturing method based on thermal fatigue. Dissertation for the degree of Doctor of Science in Technology. Espoo, Finland, 2006. (Available online from: http://lib.tkk.fi/Diss/2006/isbn9512282631/) 5) Becker, F.L., Doctor, S.R., Heasler, P.G., Morris, C.J., Pitman, S.G., Selby, G.P. and Simonen, F.A., "Integration of NDE Reliability and Fracture Mechanics - Phase I Report", NUREG/CR-1696 PNL-3469, Vol 1, 170 p., 1981. 6) Yoneyama, H., Senoo, M., Miharada, H. and Uesugi, N., "Comparison of Echo Heights between Fatigue Crack and EDM Notch", 2nd International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurized Components, New Orleans, Louisiana, U.S.A., 8 p., 2000 7) Kemppainen, M., Virkkunen, I., Pitkänen, J., Hukkanen, K. and Hänninen, H., "Production of Realistic Flaw in Inconel 600", Conference on Vessel Penetration 13 Inspection, Crack Growth and Repair, USNRC and Argonne National Laboratory, Washington D.C., Gaithersburg, USA. NUREG/CP-0191, Vol. 1, pp. 51-60 and Vol. 2, pp. 181-196 (presentation slides), 2003. 8) Wåle, J. & Ekström, P., Crack Characterisation for In-Service Inspection Planning. SKI-project 14.4-940389, SAQ/FoU report 95/07, SAQ Kontroll Ab, Stockholm, Swerige, 1995. 9) Lemaitre, P., Iacono, I. & Vergucht, P. (eds.), Results of the Destructive Examination of the ENIQ Pilot Study: Defect Catalogue, ENIQ Report 19, EUR 19024 EN, JRC Petten, The Netherlands, 1999. 10) Packalen, T, Sillanpää J, Kemppainen M, Virkkunen I, Paussu R, "The influence of the crack opening in the UT inspection qualification", 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurized Components, Budapest, Hungary, 2007. To be published. 11) Virkkunen, I., Kemppainen, M. and Pitkänen, J., "Effect of Crack Opening on UT Response" The e-Journal of Nondestructive Testing & Ultrasonics, ISSN: 1435-4934. 11(11), Nov 2006. 9 p. (Available online from: http://www.ndt.net/article/ecndt2006/doc/Th.4.4.2.pdf) 12) Paussu R, Virkkunen I, Kemppainen M, "Utility aspect of applicability of different flaw types for qualification test pieces", 6th International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurized Components, Budapest, Hungary, 2007. To be published. Annex 1 Results for measured flaw characteristics; specified target size, best estimate values for the thermal fatigue cracks (as given by the manufacturer), and true values as the results of the destructive testing. Flaw ID Target Produced best estimate True Length (mm) Depth (mm) Surface opening (µm) Length (mm) Depth (mm) Surface opening (µm) Depth (mm) Mouth (µm) Middle (µm) Tip region (µm) Tip radius (µm) 186AGB364 6 3 80 7,7 - 80 4 96 73 33 1,5 193AGB367 6 3 160 11,2 - 100 4,3 88 66 50 2,2 211AGB385 6 3 40 6,9 3,2 60 3,2 101 30 20 0,4 218AGB389 6 3 120 8,1 3,2 80 3,3 54 44 12 0,4 235AGB395 12 6 120 21,7 6,5 140 6,2 104 54 20 0,8 237AGB396 12 6 40 20 6,5 75 5,8 67 40 14 0,1 239AGB398 12 6 160 20,9 6,5 150 6,5 95 60 43 5,5 250AGB405 12 6 80 19,4 6,5 105 5,7 101 62 23 4,8 139AHF101 3 - - 6,5 * - 5,7 68 53 21 3,9 139AHF102 6 - - 9,7 * - 8,1 149 102 12 0,5 139AHF103 3 - - 6,9 * - 5,4 133 68 53 0,3 139AHF104 6 - - 10,0 * - 7,8 105 92 64 0,7 139AHP101 10 - - 10,0 - 9,2 86 129 130 3,6 139AHP102 6 - - 6,5 - 6,5 126 110 31 10,2 139AHS101 3 - - 8,2 * - 6,9 191 233 44 9,1 139AHS102 3 - - 7,7 * - 6,4 227 149 38 12,5 139AHS103 6 - - 9,4 * - 8 184 255 87 14,9 139AHS104 6 - - 10,0 * - 9 158 217 151 5,7 * Depth values are before machining of root surface of the weld joint. Machining thickness is unknown.

Effect of Crack Opening on UT Response Iikka VIRKKUNEN, Mika KEMPPAINEN, Trueflaw OY, Espoo, Finland Jorma PITKÄNEN, Posiva, Finland Abstract. Crack opening is one of the key parameters affecting the UT response of the crack. Tight cracks with small opening tend to be more difficult to detect and characterize than cracks that have wider opening. In particular, the opening of crack tip has marked effect on the crack tip diffraction signal often used for crack sizing. Service-induced cracks found exhibit wide variety of different openings. The opening is affected by the service loads and crack growth mechanism. In general, cracks grown by high loads tend to have wider opening than cracks produced by small loads. Furthermore, residual stresses may alter the opening. In order to simulate the wide variety of openings of the service-induced cracks, a novel method for producing artificial flaws with controlled opening is presented. A set of similar realistic flaws was produced by controlled thermal fatigue loading. The as-produced "baseline" UT response of these cracks was recorded with phased array technique using shear waves. Some of the flaws were then subjected to different loading sequences to manipulate their opening. The UT response of the modified cracks was then recorded and compared to that of the baseline response. The crack tip signals were measured also with longitudinal waves before cutting the specimen. Finally, the sample was carefully sectioned to reveal the opening of the produced flaws and the effect of crack opening to the UT response is analyzed. 1 Introduction Trueflaw Ltd. manufactures artificial cracks for NDT training and qualification purposes. The cracks are produced directly (not weld-implanted) to ready-made samples using local thermal fatigue loading. The thermal fatigue loading is applied using successive high frequency induction heating and water-cooling cycles. No artificial initiators are used and no mechanical contact to the loaded sample is required. Consequently, the process does not damage or disturb the surface of the sample and the original surface roughness and features are retained. The used temperature cycles are chosen for each material to be such, that no microstructural alterations are caused by the process. For example, in case of the Inconel alloys (Inconel 600, 690, 182, 82, 152, 52 and the like), the highest temperature used during loading is 500°C. Also, the time spent around the highest temperature is short, typically below 10% of the total cycling time. In NDE qualification, the relevance and representativeness of used artificial flaws is very important. The artificial cracks or flaws used in qualification or training must give a relevant NDE-response. This is achieved by using flaws similar to the actual service- induced cracks or postulated cracks. More precisely, the used flaws must be similar in terms of some significant parameters, that control the obtained NDE response. One of those important parameters is the crack opening. Tight cracks with small opening tend to be more difficult to detect and characterize than cracks that have wider opening. In particular, the opening of crack tip has marked effect on the crack tip diffraction signal often used for crack sizing. ECNDT 2006 - Th.4.4.2 1 Natural cracks found during in-service inspections exhibit wide variety of different openings. The opening is affected by the service loads and crack growth mechanism. In general, cracks grown by high loads tend to have wider opening than cracks produced by small loads. Furthermore, residual stresses, if present during inspection, may alter the effective opening seen by the inspector. 1.1 Typical crack opening values in service Natural cracks exhibit a wide variety of different opening characteristics. This results from the wide variety of different loading conditions and mechanisms that have induced these cracks, as well as the material properties of the material in question. Wåle [1] has measured crack mouth and tip openings from numerous service- induced cracks. These values are summarized in Figures 1 and 2, respectively. It should be noted that bigger cracks with wider opening are more likely to be found. Consequently, the data is likely to be biased and the population shows too great opening values. Crack opening values for different artificial flaw manufacturing techniques are even more difficult to find from the literature than that of natural, service-induced flaws. Crack opening values can be measured from Lemaitre et al. [2]. These cracks represent values from numerous flaw manufactures using weld implanting techinques and solidification cracks. This data is summarized in Figures 1 and 2. When reading this data, it should be noted, that measurement of crack opening values from the images was rather difficult and the accuracy of these numbers is limited. Figure 1. Crack mouth opening (CMO) values of service-induced cracks and artificial cracks, in µm. The bars show the range between minimum and maximum values reported. The dots connected by line show the average value reported. Note, that the average value tends to be closer to minimum than maximum. Values for service-induced flaws were taken from [1]. Values for artificial flaws were measured from [2]. 2 Figure 2. Crack tip opening radius (CTO) values of service-induced cracks and artificial cracks, in µm. The bars show the range between minimum and maximum values reported. The dots connected by line show the average value reported. Note, that the average value tends to be closer to minimum than maximum. Values for service-induced flaws were taken from [1]. Values for artificial flaws were measured from [2]. 1.2 Aim of the study In order to simulate the wide variety of openings of service-induced cracks, a method for producing artificial flaws with controlled opening is desirable. Trueflaw cracks are produced directly to the specimen with applied thermal fatigue loading. Since the crack opening is affected by the loads used to produce the crack, this method also offers the possibility to alter the opening of produced flaws. The aim of present study was to study the effect of crack opening to ultrasonic inspection to confirm the importance of this parameter to inspection. The possibility of producing cracks with controlled opening was applied and the controllability of the crack opening studied. 2 Materials and methods To study the effect of crack opening to ultrasonic response and the performance of the Trueflaw crack manufacturing technology in controlling the crack opening, a test sample was manufactured with 5 identical cracks. The sample was prepared from 25 mm thick AISI 304 -type stainless steel and had dimensions of 150 by 150 mm. After manufacturing, a baseline NDT-responses of the cracks were recorded, as detailed in paragraph 2.1. After first inspection, some of the cracks were subjected to additional loading designed to increase the crack opening. Three different loadings were used to achieve different opening conditions. The loading was not sufficient to alter the size of the cracks; only the opening was altered. In succession, the NDT inspection was repeated in order to reveal the difference in UT-response caused by the changed opening. Finally, 3 the sample was destructively studied, as detailed in paragraph 2.2, to reveal the true crack opening of the cracks. The detailed information on the flaws manufactured are presented in Table 1. Unfortunately, the surface of the chosen sample had deep scratches, which initiated some secondary cracks during loading. Consequently, the NDT-response shows some disturbances caused by these cracks. Table 1. The flaws manufactured for the study. Number Length (l) Depth (a) Note 1 9.6 3.8 Opening increased ; loading 1 2 9.8 3.3 Opening increased ; loading 2 3 9.5 3.5 Opening increased ; loading 3 4 10.2 2.8 (deepest point destroyed during sample preparation) 5 10.4 3.2 2.1 NDT-inspection The UT-measurements were carried out with two phased array systems: Omniscan (RD Tech) with 16 elements 5MHz angle probe for shear waves and MultiX (M2M) with linear 64 element 5MHz probe with 1 mm pitch (Imasonic). Omniscan was calibrated for the measurement using notches of different sizes. The MultiX was calibrated using one notch to produce high focus to the area of interest. For the measurement 20 elements were used to avoid too high focusing. The passive aperture was 15 mm. The width of the notch was 1 mm in 20 mm depth. 2.2 Destructive analysis The specimen was first cut to 5 pieces each containing a single crack with an abrasive cutter. Then, each crack was sectioned by a precicion cutter (Struers Accutom-50). The location of sectioning was carefully selected to be near, but not at, the deepest location of the crack. Next, a metallographic sample was prepared from the piece still containing the deepest location of the crack. The sample was first grinded with SiC-papers and then polished with 1 µm diamond paste. Special care was taken in sample preparation to avoid any changes to crack opening. In particular, the time used in diamond polishing was minimized, in order to minimize increase of opening caused by rounding of the crack edges. Also, no etching was applied for the same reason. For each crack, the metallographic sample was prepared and photographed twice, to ensure that the measured values are not affected by variation in the sample preparation. After preparation, the samples were photographed by an inverted (metal) microscope. The crack tip opening was measured from digital micrographs by fitting a measurement circle to the tip location. Radius of this circle was reported. The greatest available magnification was used to obtain the best possible accuracy in tip measurements. However, the opening of some flaws was so small, that it could not be reliably measured with magnification available in optical microscopy. The crack opening at the crack mouth and along the crack depth was measured from the digital micrographs by custom made image analysis tool. 4 3 Results 3.1 Destructive analysis Figure 3 shows the crack opening measured from different samples. The measured crack tip opening and crack mouth opening values are summarized in Table 2. Table 2. Crack tip opening (CTO) and crack mouth opening (CMO) values measured (in µm). For cracks 1, 4 and 5 two CMO values are reported; these values correspond to opening before and after the opening treatment, respectively. Number CTO CMO 1 1.4 110 / 160 2 0.4 92 / 96 3 0.2 110 / 145 4 <0.05 77 5 0.1 100 3.2 NDT analysis Figures 4 shows characteristic NDT-response of each of the cracks before and after the altering treatment. The sample was also inspected with a phased array probe (linear array probe with 64 elements from which were used 20 elements, pitch was 1 mm) set-up after the alteration of the opening. The results of this inspection are shown in Figure 5. 5 Figure 3. Measured crack opening as a function of depth (in µm). 6 C ra ck 1 C ra ck 2 C ra ck 3 C ra ck 4 C ra ck 5 Figure 4.Phased array inspection results before (left side) and after (right side) modification of the flaw opening. Crack 5 did not experience any opening treatment. Thus, the differences in images for Crack 5 reflect the differences in inspection. 7 Figure 5. Phased array measurement results by focusing 5 MHz probe with 20 elements in the depth of 20 mm. 4 Discussion Table 2 shows, that the crack tip opening in Trueflaw cracks after production was rather small. The small crack opening typical to the cracks is a result of the crack growth process driven by the high crack tip stress concentration. This feature is shared by all natural cracks that grow due to the crack tip stress concentration, including thermal fatigue cracks, fatigue cracks and stress corrosion cracks. A blunt crack does not have the required stress concentration for crack growth and thus it would not grow. On the other hand, crack-like flaws that do not require the high crack tip stress concentration for growth, e.g. solidification cracks, do not necessarily show as small crack opening or as sharp tip. The crack opening increases with increasing loading due to crack tip blunting. The Trueflaw artificial crack production method is an accelerated process. Whereas in the actual power plants, when cracks appear, they typically form after years of service, the production of artificial thermal fatigue cracks only takes hours or days. This acceleration is achieved by using loads that exceed the service loads and higher frequency of the cyclic loading. Due to the greater loads, the crack opening in Trueflaw cracks is expected to be similar or a little greater than in actual service-induced cracks. However, most of the acceleration is gained during the initiation phase of the crack, which does not affect the crack opening. Consequently, it is expected that the difference in opening of Trueflaw cracks and real service-induced cracks is minimal. The values shown in Table 2 can be directly compared with the values shown in Figure 1 and 2. This comparison shows, that the initial condition of Trueflaw cracks is similar to many service-induced cracks. With the alteration in opening, the cracks can be adjusted to represent wider variety of service-induced cracks. However, at least corrosion fatigue cracks exhibit so great crack tip openings, that even in the altered condition, Trueflaw cracks remain much tighter than those cracks. The crack opening through the whole length of the crack can be adjusted by applying different thermal loads. Comparison of different cracks in Figure 3 and Table 2 8 show that the crack opening, and especially the important crack tip opening, can be changed markedly after flaw production. The crack tip radius shows increase by factor of 10. Figure 3 shows, that the crack opening is altered throughout the crack length. Also the NDT-results presented in Figure 4 show significant difference between the different cracks. In particular, the crack tip amplitudes show clear increase with increasing crack tip blunting. 5 Conclusions The crack opening of Trueflaw cracks is expected to be similar or slightly greater than opening of actual service-induced cracks. This view is also supported by the comparison with the data available in the literature for natural cracks. The typical crack opening values of artificial cracks produced by other means than Trueflaw technique is clearly bigger than that of most types of service-induced natural cracks. Thus, if crack opening is considered to be a significant factor for the inspection of NDE method in question, artificial flaws with big openings can not be recommended for these flaw types. The crack tip sharpness of Trueflaw cracks is considered similar to service-induced cracks. These features are expected to be shared by all types of cracks, that grow due to the crack tip stress concentration, including fatigue cracks and stress corrosion cracks. The opening of existing Trueflaw cracks can be altered to have bigger values, if so desired. The increase of opening changes the ultrasonic response of the cracks. 6 References [1] Wåle, J. & Ekström, P. 1995. Crack Characterisation for In-Service Inspection Planning. SKI-project 14.4-940389, SAQ/FoU report 95/07, SAQ Kontroll Ab, Stockholm, Swerige. [2] Lemaitre, P., Iacono, I. & Vergucht, P. (eds.) 1999. Results of the Destructive Examination of the ENIQ Pilot Study: Defect Catalogue. ENIQ Report 19, EUR 19024 EN, JRC Petten, The Netherlands. 9

Effect of Stress on Ultrasonic Response in Detection and Sizing of Cracks Jorma PITKÄNEN, Posiva Oy, Olkiluoto, Finland Anssi LAUKKANEN, VTT, Espoo, Finland Mika KEMPPAINEN, Iikka VIRKKUNEN, Trueflaw Ltd, Espoo, Finland Abstract In different NDT techniques huge developments have been achieved during the last few years with regard to crack detection and sizing. In a large range of materials, cracks are one of the most dangerous defect types. A crack is a planar reflector, which is sometimes extremely difficult to detect and to be sized. A crack can be totally open, partly closed or even totally closed because of compressive stresses. The effects of stresses cause problems, for instance, in in-service inspections of nuclear power plants in detection and sizing of closed cracks. This phenomenon causes similar effects in all kinds of plants and components. In this study some experimental inspections have been carried out as well some FEM calculations of stress field around the crack and compared to corresponding measurements in literature. The used method is dynamic loading, which gives during ultrasonic measurement clear evidence on the effect of the crack closure as well on the amplitude variation limits in ultrasonic testing. Materials used in the study are AISI 304, AISI 321 and ferritic piping steel. The load applied to the cracks was in form of different thermal cycles. The maximum temperature variations were from 20 °C to 600 °C depending on each dynamic loading cycle. Different types of ultrasonic methods were used in the measurements. The methods applied are normally used in the field in normal ultrasonic inspections like angle shear wave, creeping wave, TOFD and phased array. The results are presented and conclusions are drawn from the stress effect on the detection and sizing. Clearly the sizing methods have some differences when using manual sizing techniques, TOFD, phased array or SAFT technique. The effect of closure on the response of normal ultrasonic practical probes was recorded. Introduction Ultrasonic and radiographic testing are the typical volumetric test methods. They are normally used in many kinds of in-service inspections. One of the main tasks is the detection and sizing of cracks. The cracks are planar defects. According to fracture mechanics they are usually the most severe defect types, especially on the surface of the component. Cracks can be induced from various reasons during the manufacturing or in the use of the component. Typical in-service induced cracks are IGSCC, IGA, fatigue cracks and thermal fatigue cracks. Typically they are near the HAZ or in the weld toe, or in the geometrical discontinuities of the component. The orientation varies depending on the loading of the component. During in-service inspections detected indications have to be evaluated. Normally the size of the detected indication is the basic decision tool, if the component should be repaired or not. In order to detect cracks in the components there are several techniques ECNDT 2006 - Tu.3.3.4 1 applied depending mainly on the material and its geometry. The crack can have different loading history and this can affect drastically the defect detection and sizing. In the literature the cracks are normally studied under a static load situation. In these studies the corner echo and crack tip echo behaviour have been explained [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]. Applied load has normally been three- or four-point bending in these studies. Bending will cause extra stresses into the large volume of tested cracked specimen. Dynamic load onsite is normally caused by the shut-down, start-up or resonance of the piping during the use of the power plant. The thermal fatigue crack initiation is combined with the temperature variation of the environment of the cracked component. In these cases the effective loading caused by the thermal stresses is mainly local. If the crack is closed, the ultrasonic waves go partly through it and partly reflect. In this case the detection of the crack is minimised. The real information of the crack behaviour during dynamic loading, which can be similar loading as in the piping, is important for inspection. The dynamic behaviour of the crack affects the detection of cracks in the components. This study is concentrated to loading the crack dynamically with known interval of thermal stresses induced locally. This is different compared to 3- or 4-point bending, which exhibits a long range stress field. Under the dynamic loading the crack is studied with several ultrasonic techniques. The effect of the stress on the detection and sizing of cracks has been studied in many occasions, but the information leaves always some questions after each study. The ultrasonic behaviour of a closed crack has been studied in the following publications. Becker et al. [1] concentrated to inner-diameter transgranular cracks, which they considered as most critical crack type. In the study smooth and rough cracks were used, but no significant differences in ultrasonic response were obtained. Classification to rough or smooth cracks originated from the crack manufacturing. In the study mechanical fatigue cracks and thermal fatigue cracks were inspected. In four-point bending about 15 to 35 dB loss in corner echo response was measured, depending on the measured crack. Crack tip signals were not measured in the study. The stress cycle was about 105 - 125 % of the yield strength of the tested materials. Boström et al. [11] made calculations from transmission and reflection coefficient using spring model to describe behaviour of closed cracks. Persson [12] calculated on the basis of geometrical theory of diffraction model that the effect of 250 MPa compression stress reduced 5 - 20 dB the backscattered amplitude. Zetterwall et al. [13] reported partly own results and partly results from other investigations. The obtained decrease under compressive stresses in ultrasonic response was between 10 – 35 dB depending on the probe and wave type. In general, it was obtained that 100 MPa increase of the compressive stress caused a 6 dB drop in ultrasonic response. Fracture surface roughness affected about 3 - 10 dB change in the reflection, when the surface roughness varies between 5 – 30 μm. The theoretical curves were compared to Arakawa measurements [13], which showed 26 dB reflection differences between 1.5 to 40 μm of the surface roughness. According to these results transmission is higher with larger compressive stresses and lower fracture surface roughness. According to Zetterwal et al. [13] Whapman measured with TOFD 13 dB signal loss in diffraction signal response when the test specimen was loaded with 240 MPa compressive stresses. Exactly the same measurement setup as used by Whapman [13] was used by Mihara et al. [2] in their study. Mihara et al. [2] studied also the crack tip opening width effect on the measured diffraction signal and detected that between 0.01 – 0.1 μm crack tip opening during compression the ultrasonic signal disappeared. Saka et al. [3] measured the crack depth from the ultrasonic response of the crack face during different loads. Depending on the probe position curve and the amount of 2 compressive stresses, the crack depth can be estimated. In this study it is shown that the crack behaviour under compressive stresses is more complex than just linear decrease of the ultrasonic response. The effect on ultrasonic sizing in practice has not actually been discussed in these reports, it was just commented that there is an effect on the sizing and detection. In this study non-linearity effect of the crack on the ultrasonic response was not used, even though it is very useful to give information about crack closure [14, 15, 16, 17]. Effect of degradation mechanism Materials can be cracked in several ways. Typical cracking mechanisms are thermal fatigue cracking, mechanical fatigue or cracking induced by stress corrosion. The type of degradation mechanism is also important in order to understand the severity of the detected defects. It is not easy to be able to distinguish real crack types from each other and characterise them properly. In any case the degradation mechanism and the size of detected defects give us the need for repair or possibility to leave the defect to next in-service inspection. Figure 1 shows the crack growth of thermal fatigue crack, fatigue crack and IGSCC. As it can be seen from Figure 1, the mechanical fatigue leaves us more time for repair as compared to IGSCC or thermal fatigue cracking. In ASME XI the maximum allowed crack size for calculations is 75% of the wall thickness [18]. If the piping has 10 mm wall thickness, then a 1 mm deep crack grows with stress corrosion mechanism in 11-12 years over the ASME XI allowable size. Correspondingly the 5 mm deep IGSCC crack reaches the ASME allowable limit in less than 2 years, according to these calculations. So the proper sizing is as important as the degradation mechanism. The used cycles are typical for nuclear power plants in one year. The loading of the component can vary so that crack growth is very rapid or slow. The loading type and the amount of load can vary the dominating degradation mechanism. The used equations are based on ASME XI [18]. For calculation conservative half- elliptic cracks were used, with the length–depth relation of 1/6. Typical ultrasonic calibration defect for detection of cracks is 10% notch of the wall thickness. The recording level is normally 20% of the amplitude of the calibration notch. Alternative calibration reflectors are typically a flat bottom hole or side-drilled hole. According to Figure 1 case, the recording level would be -14 dB compared to 1 mm notch for this calibration type. The question arises: is a 1 mm crack detectable with this calibration type or not? According to Becker et al. [1] even large cracks (greater than 50% of pipe wall thickness) can be missed, when 100% recording level of calibration defect is used. This might be one reason, why -14 dB recording level was added to ASME XI. Figure 1 Crack growth values for different degradation mechanisms with starting crack sizes 1 and 5 mm 3 Thermal stress state effect on crack behaviour To estimate thermal stresses in the test specimens some FEM-analyses were carried out. In the analyses coupled thermal to continuum mechanical analysis were used. The analysis involved an initial crack essentially in a large test specimen of 2D nature, i.e. a through- wall crack. The constitutive formulation in the mechanical analysis was such that the material constitutive response was modelled as temperature dependent incremental isotropic elastic-plastic one (multi-linear stress-strain relationship), while the deformations were described as being finite. The finite element formulation was of p-type, and fourth order elements were used in the mechanical analysis and second order in the thermal analysis. As the crack face contact as a result of the thermal loading is imminent, the analyses were performed with penalty function controlled contact constraints imposed on the crack faces. The models had approximately 50000 degrees-of-freedom and 4000 triangular elements of the above specified degrees of interpolation. The mesh used in the calculations is shown in Figure 2. Figure 2 FEM mesh used in calculation with 8 mm deep crack in austenitic material Figure 3 Temperature and 1st principal stress distribution during one heating cycle and cooling cycle Heating and cooling cycles have been modelled. It can be seen from the calculations, that during the heating in the crack tip concentrated compression load is present and on the contrary in the cooling phase stress state changes to tensile load near the tip area, Figure 3. More accurate survey of the crack tip area shows that during the heating cycle the compression changes to tensile stresses on the left in Figure 4. Figure 4 Stress on the fracture surface near the crack tip on the left and derivative of 1st principal stress as function of cycling time on the right 4 This can be seen also on the right in Figure 4, where normalised amplitude is defined as an integral of time rate of change of tensile stress over the crack face is presented, i.e., derivative of 1st principal stress is shown as function of cycle time with different “amplitude” integration distances from the crack tip towards the crack mouth. On the right in Figure 4 the curves are calculated to 5%, 10%, 20% and 30% of the total crack length from the crack tip. The total crack length was 8 mm and this corresponds lengths of 0.4 mm, 0.8 mm, 1.6 mm and 2.4 mm from the crack tip. Definition of the crack for ultrasonic inspection In the crack three different areas can be distinguished: • Corner, which corresponds to CMOD (crack mouth opening displacement) • Crack face, which corresponds to fracture surface of the crack • Crack tip, which corresponds to effective area of crack tip opening (plastic zone), CTOD (crack tip opening displacement) and variable loading conditions All these areas have own impact to ultrasonic response. The size of the crack tip area is actually dependend on the load at the crack tip. Under tensile load the area can be estimated to be 3-4 x crack tip opening width (plastic zone) where crack tip affects. Under compressive load cyclic plastic zone corresponds to ¼ of maximum monotonic plastic zone. Cyclic plastic zone can be in maximum about the crack tip opening width. The load affects the crack tip radius. If we assume the crack tip radius to be 1 μm, we can have an effective zone of 4 μm under tensile stress and of 1 μm under compressive stress in maximum. In a typical crack growth situation the plastic zone can vary from 10 μm even to 1 mm. According to [2] the ultrasonic response is affected very strongly when the crack tip radius is less than about 0.1 μm. The effect of different zones of the crack on the ultrasonic response has been shown in principle in Figure 5 [19]. Other important factors are the roughness of the fracture surface, oxide films on the surfaces of the crack and orientation of the crack, because they affect the reflection behaviour of the ultrasonic waves and the crack behaviour under compressive stresses. In front of the crack tip there can be micro-voids in the plastic zone. These can also have an effect on the received ultrasonic response. Additionally, if the crack is not straightforward, the crack tip area can produce also reflection from the crack facets near the tip. Thus, the crack tip signal can be a sum of diffraction, scattering and reflection. Figure 5 Ultrasonic measurement of crack behaviour during dynamic loading with shear angle wave probes 5 Ultrasonic measurement techniques for detection and sizing Several techniques are available [20, 21, 22, 23] for ultrasonic detection and sizing. Ultrasonic detection of a crack in the materials is firstly based on corner echo detection. Typical way is the use of shear waves between angles of incidence 35° - 55°. The creeping wave technique as a part of mode conversion probe is extremely sensitive to detect crack-like defects in certain conditions. Creeping waves are not so sensitive to crack orientation compared to shear waves. Creeping wave (C2) conditions are that inner and outer surfaces have to be parallel. The outer and inner surface condition is also important factor. For instance, the counter bore is a factor which does not allow the use of creeping waves. In mode conversion probe there is also available other possibilities to characterise the crack such as 30T-70L-30T (C3) or 30T-70L-70L (C1). In TOFD technique diffraction from discontinuities is recorded and analysed based on the diffraction effects in cracks. TOFD (Time Of Flight Diffraction) technique uses longitudinal waves for detection and sizing. The main principle is to use crack tip signals in order to receive the dimensions of the crack. TRL-probes are often used for inspection of difficult materials such as austenitic weld, casting or dissimilar metal weld. Phased array technique is spreading fast into the industry applications. The applications of angular scanning and dynamic depth focusing in phased array analysis give good tools for defect characterisation and sizing. In this study typical manual and mechanised probes have been used in order to see the effect of stress state changes in normal inspection signals. Also some measurements with two phased array equipments have been carried out. The stress state in the material can also cause difficulties to interpret the results using TOFD as well other techniques for analysis. TOFD and phased array techniques have been utilised in measurements. For SAFT reconstruction normal angle shear wave probe has been used. Typical way to analyse defect size is to use crack tip signal if it is available. In order to analyse the defect sizing capability also crack tip signals have been recorded always, when it has been possible. Several ultrasonic equipment were used in this study: • Sumiad V3.34 multichannel equipment for mechanised inspection (Tecnatom) • Phased array equipment MultiX (M2M) • Phased array equipment Compas-128 (BAM) • PC-SAFT equipment (IzfP). The ultrasonic equipment Sumiad V3.34 was used to TOFD measurement and normal manual probe measurement during thermal cycling loading of the tested material. The equipment provides maximal 8 channel for data acquisition. During measurement each A- scan was digitised between 10 to 100 times in a second with 8 byte resolution. The amount of points was obtained to be too few in earlier work presented by Pitkänen et al. [19]. The collected data were either in form of RF-signals or logarithmic A-scans. Phased array measurements of thermal cyclic loading were carried out with 128 channel equipment called Compas-128. In the measurement an 8 element phased array probe was used. The loaded specimen was measured with a range of angles between 20° - 55°. The usable range of angles of incidence for that probe is between 20° - 48° according to the information given by the manufacturer of the probe. The frequency of the probe was 4 MHz. The measured data were stored with 14 bit resolution, which gave large dynamics for the measurement. The pitch of the used probe was 1.8 mm. The stored data was in rectified A-scan form. 6 SAFT measurements have been carried out with PC-SAFT equipment, in order to receive information from cracks before thermal loading and to visualise the effects seen in the SAFT reconstruction. The equipment is using one channel with 100 MHz sampling rate. The UT-measurements of closed cracks without thermal cycling were carried out with MultiX (M2M) using linear 64 element 5 MHz probe with 1 mm pitch (Imasonic). Omniscan was calibrated for the measurement with the help of notches of different sizes. Calibration concentrated to one notch to produce as high focus as possible for an interesting area. For the measurement, only 20 elements were used to avoid too high focusing. The passive aperture was 15 mm. The width of the notch was 1 mm with 20 mm depth in a similar austenitic specimen as the specimen with cracks. Figure 6 Used ultrasonic equipment for the measurement from the left: Sumiad V3.34, PC-SAFT, MultiX and Compas-128 The thermal loading was made by thermal fatigue equipment. The loading cycles varied from 2 s to 10 s. In each loading the cycles were repeated several times. The effects of different thermal loading and cooling cycles on the ultrasonic response were measured. The thermal cycling behaviour in combination with ultrasonic measurement has been reported by Pitkänen et al. [19]. Effect of the stress state of the cracks in some experimental data of ultrasonic inspections of cracks SAFT measurement with thermal fatigue cracks In austenitic steel piping material (AISI 321) several thermal fatigue cracks are detected, the origin of which is typical temperature differences between two connecting pipes. The low temperature flow in the smaller pipe had resulted a temperature difference on the inner surface of the larger pipe. This fluctuation of temperature caused the crack initiation and growth. In the measurement a 12 mm deep crack was measured with 4 MHz shear wave probe, the angle of incidence was 55° using SAFT. Reconstructions were carried out for each millimetre. The crack length was over 100 mm. The stress state of the crack itself has a strong effect on the detectability and sizing of the defect. In Figure 7 it can be seen, how the effect of the stress state of the crack changes the ultrasonic response. First picture, in the upper left, shows a weak corner echo when crack starts. In the middle area of the crack several different B-scans can be seen. Echoes are visible echoes from coming mode conversion, crack face (fracture surface) with direct sound path or through reflection and crack tip. Additionally, double corner echoes can be seen, which shows that the sound field sees crack jumping along the crack length in passive direction of the piezo crystal. The real crack indication depends on the local stress state. Three pictures in Figure 7 show that crack is open, because the crack face is clearly detectable. In two pictures in Figure 7 the crack tip echo can be seen together with the corner echo. It seems that in these situations the 7 crack is closed, but the crack tip is open. Thus, the loading history affects the response clearly. The sizing of this crack was easy, but in some parts the sizing could not be carried out properly. Mode conversion indication shows if the crack is in thickness direction or if it is tilted. But in the third picture of the lower row a clear crack face indication can be seen, but no mode conversion indication is present. In this picture the corner echo is low. From the carried out measurements can be obtained that indications received vary locally quite a lot, even though the height of the crack does not change remarkably. These obtained variations in the SAFT reconstructions are mainly due to the effect of local stress on the crack. Figure 7 Effect of stress state on the ultrasonic response from a real thermal fatigue crack (all data from the same crack at different positions, crack height is about 10 mm) Ultrasonic measurement of mechanical fatigue crack with phased array during thermal cycling An austenitic steel specimen (AISI 316) was loaded with dynamic thermal loading. The ultrasonic measurement was carried out with Compas-128 phased array equipment developed by BAM. In the test specimen a mechanical fatigue crack was produced by VTT. According to manual sizing the crack size was 4.2 mm in depth and 18 mm in length. The crack was loaded with 50 cycles. Each cycle has a heating and cooling phase. Heating took 6 s and in this cycle cooling was as long as the heating. Heating and cooling times can vary independently. During the heating cycle the temperature increased from 10 °C to 400 °C. The cooling was very rapid and reached in 1 s 10 °C. Figure 8 shows B-scan picture of Compas-128 equipment during cycling of the of the austenitic test specimen. Figure 8 Ultrasonic measurement of crack behaviour during dynamic loading with shear angle wave probes CR = Corner reflection, MCI = Mode conversion indication, CFI = Crack face indication, CTI = Crack tip indication, ISI = Inner surface indication, CFRI = Crack face reflection indication 8 By analysing Figure 9 data several results can be obtained. During cycling it can be seen that the corner echo decreases about 22 dB during the heating cycle. During the cooling small changes in the range of 2 dB can be seen in the corner echo. Small maximum can be obtained when cooling starts after 1 s in each cycle. Similar maximum can be seen in the crack tip signal, but the effect is larger than by the corner echo. After maximum crack tip closes partly and the amount of this effect is between 4 to 9 dB. The thermally loaded material stores some part of thermal energy by slowly increasing the temperature of the test object. The crack tip changes decrease because the cooling is not sufficient to cool to the start temperature. Crack tip amplitude shows small tendency to increase little bit towards to the end of heating. This means that crack tip width (CTOD) increases in small extent. The quantity of this effect varies between 2 to 9 dB. It seems that during the heating part of each cycle by closing the CMOD makes the crack tip to open partly, but not enough to produce a proper tip signal in this case. Figure 9 Ultrasonic measurement of crack behaviour during dynamic loading with shear angle wave probes. Corner echo and crack tip echo behaviour can be seen during dynamic thermal loading Ultrasonic measurement of mechanical fatigue crack in ferritic piping material The cracks were introduced by mechanical fatigue to a ferritic piping material. TOFD measurements were carried out with two 5 MHz longitudinal probes having angle of incidence 60°. The SAFT measurements were carried out with a 4 MHz 70° shear wave probe. Figure 10 shows similar results for both methods. In the TOFD picture in case of 8 mm crack height, the sizing is not clear but from the back-wall echo disturbances can be assumed, that it concerns a crack. The SAFT measurement was carried out from the same side as the crack is and TOFD measurement from the opposite side of the crack. In TOFD B-scans the lateral wave is on the bottom and the back-wall echo on the top. This has been done only for comparing and visualising the results of SAFT and TOFD methods. The crack-tip echo is very small in case of 8 mm crack. But it can be seen with SAFT method quite clearly. In TOFD the noise level is very near the crack-tip signal level in this case. It seems that the crack is closed. This can be estimated since the back-wall echo is seen and it does not disappear. Normally in case of open crack on the opposite side 9 of test specimen the back-wall echo is disturbed. Thus, after manufacturing the fatigue crack on the surface compressive residual stresses remain, which cause that the ultrasonic waves go through in the surface area of the inspected crack. This can be interpreted that there is no crack, which opens on the surface. Crack-tip echo can be assumed to come also from another discontinuity than from a crack in this case. Same effect can be seen for the larger crack (23 mm deep). The back-wall echo is seen through the specimen. High tip indications come from several depths in TOFD measurement. The SAFT measurement is similar. In the surface the crack is clearly closed and in the crack face there are several areas where ultrasonic waves do not reflect properly. Thus, the crack is partly under high compressive load in the areas of crack face and crack tip. Figure 10 Measurements of the specimen to analyse the cracks before thermal loading. On the left SAFT reconstructions are shown and on the right TOFD measurements of the corresponding test specimens, respectively Figure 11 Results of thermal cycling with TOFD measurement of 8 mm crack specimen. TOFD B-scan on the left for 2 cycles is shown and on the right TOFD signal behaviour during one cooling cycle Both ferritic specimens were also thermally loaded with different cycling. In Figure 11 changes in the TOFD signal during thermal cycling can be seen. In the left picture during the beginning of the heating, measurement shows that he crack tip opens and closes quickly after it. This happens in each similar cycle in the same way. At the beginning of 10 the cooling the crack moves. This is very clear effect. After that crack tip closes. This is a repeatable effect and it can be said that it is characteristic to the loading cycle. These effects are similar when thermal loading stays the same. The increase of temperature of test specimen changes the conditions and this can change a little the behaviour of the crack. On the right in Figure 11 the cooling phase from other type of cycles can be seen. It can be seen clearly, that an “opening wave” goes towards the crack tip and crack tip opens finally. Similar “opening wave”of the fracture surface of a crack fracture surface can be seen also in the mode conversion side of TOFD signals. Later in the cooling, the crack tip closes and cannot be seen clearly. Sizing is clearly difficult by every technique in this case when the crack tip is closed. Ultrasonic measurement of several mechanical and thermal fatigue cracks with ultrasonic probes used for manual inspection. By cycling measurement the sampling rate varied between 10 – 100 Hz. For instance the minimum heating cycle was 2 s. So for this heating time 200 A-Scan (samples) in 2 s were stored in maximum. From the measurement with 55° shear wave probe during cycling the effects in the dynamic crack behaviour can be recorded, Figure 12. Crack tip, indications from fracture surface and corner echo are visualised in the B-scan results. The movement of the fracture surface can be interpreted from the results. It is obvious that in the different areas of the crack the stress affects both effects: closure and opening. These closure and opening movements can be seen on the left of Figure 12. When cooling starts, the temperature decreases rapidly and crack starts to open from the surface. The effect of this “opening wave” can be seen clearly. The yellow arrows show in Figure 12 the direction of the “opening wave”. The red arrows show in the contrary the “closure wave” direction. Crack tip opens also at the beginning of the cooling. But during cooling near the crack tip it can be seen that the crack tip divides in the two areas. On the right in Figure 12 the amplitude variation during one dynamic cycle in the crack tip area is shown. The opening effect at crack tip in the beginning of the cooling is similar to the opening, which happens in the beginning of the heating, when the crack is closed from the outer surface. This effect opens the crack tip for some moments until the temperature increases in the crack tip area and thermal compressive stresses close the crack. The behaviour of the crack changes according to the characteristics of the dynamic cycling. During the heating phase seen in Figure 12, the amplitude decrease of the corner echo was measured to be 27 dB. The changes of the crack tip echoes were about 10 dB. Figure 12 Ultrasonic measurement of crack behaviour during dynamic loading with shear angle wave probe. On the left B-scan as a function of time and on the right the crack tip variation during cooling (blue curve part) and heating (red curve part) is shown Figure 13 shows B-scan of thermal cycling measurements with 4 MHz 45° shear wave probe on the left and results of the measurements with creeping wave probe (WSY60- 11 2) on the right. Crack tip opening can be seen during the heating phase from the results of shear wave probe, Figure 13. Measurements using creeping wave probe show results which are comparable to previous results shown in Figure 14 on the right. Creeping wave has effective depth on the surface about one wavelength. Thus, it is very sensitive to crack mouth opening displacement (CMOD). Corner echo from other probe types comparing to creeping wave probe is reaching much deeper areas of the crack in the limits of the sound beam opening angle. The temperature increase of the tested material can also be obtained, which is seen in Figure 13 on the right from the creeping wave results. Creeping waves show tendency of decreasing of the amplitude after starting the cooling about 1.5 dB which is caused by defocusing because of the material temperature decrease. When the crack closes sharply the amplitude change was in the previous measurements not properly measured because of too small sampling rate. In the following measurement the amplitude decrease was much more, but the total amount of the decrease was not measured, but in each cycle it was more than 18 dB, Figure 13. Figure 14 shows that the mode conversion is more sensitive to the fracture surface behaviour and does not close totally during heating, even though creeping wave disappears totally during heating. This gives some hope in inspection cases, where the crack is closed from the surface (CMOD) but not from fracture surface. It can be clearly seen that behaviour of the crack depends on the stress state. If the cycles are similar, the behaviour of the crack follows the same pattern in each cycle and gives information about the characteristic crack behaviour if cycles are known. Figure 13 B-scans of thermal cycling measurements with 45° shear wave probe (on the left) and creeping wave probe (on the right) Figure 14 Thermal cycling ultrasonic measurements from creeping wave probe: on the left amplitude variation of mode conversion part during thermal cycling and on the right amplitude variation of creeping wave during thermal cycling Phased array measurement with closed thermal fatigue cracks Five thermal fatigue cracks in AISI 304 steel were produced by Trueflaw Co. The crack mouth opening on the surface varied between 77 µm and 160 µm. The crack tip signals 12 were very difficult to detect. Figure 15 shows results from 5 cracks, which were measured with phased array equipment. From each crack can be seen both A-scan and part of angular scanning in a form of B-scan. By moving the phased array probe, when it was doing angular scanning at the same time, it was possible to distinguish the movement of the crack face (fracture surface) from the noise level. To optimise the tip signal, which was carried out by tracking the maximum indication during movement of the crack face, the result is shown in Figure 15. Two of the cracks did not give a proper crack tip indication (crack 4 and 5, Figure 15). During the inspection the detection of the cracks was easy, but the sizing was not so easy. The corner echo was clearly detected with other probes like creeping wave probe or with 45° shear wave probe. But no proper tip was available. Figure 15 Ultrasonic measurement of crack behaviour during dynamic loading with shear angle wave probes For analyses of these cracks there were crack face indications available for each crack. The signal to noise ratio was less than 6 dB in case of crack tip signals of 2 cracks even using phased array probe which were focussed in the crack tip area in optimised inspection situation. The crack sizes were about 3.2 – 3.8 mm according to metallographic results. The similar evaluation of the closed crack was carried out by Dupond et al. [24]. According to their measurements the focussing of the phased array to the right depth gave better results. Also in our case ultrasonic field was focussed about 5 mm above the back wall. Summary and conclusions The stress has a large effect on the detectability of the cracks, and it seems to make sizing more difficult, too. This study has shown the variability of the ultrasonic indication as a function of stress. In the following the main items of this study are summarised: • It is possible to study the loading effects on the dynamic behaviour of the cracks. • Real loading on site can be introduced to the measurements. • Different probes have own characteristic information about stress state of the crack. • The movement of crack during the loading was recorded and visualised which shows that the loading history affects the ultrasonic response. • These effects were reproducible. • Dynamic loading analyses can be used to study crack behaviour from the stand point of 13 fracture mechanics. • Even though the crack is closed on the surface, the crack can still have areas which reflect properly. • If the loading cycles of the component are known the behaviour of the possible crack can be estimated by using similar loading as in power plant. The decrease of corner echo was measured to be between 22 – 27 dB, when the crack closes totally. This is comparable to other measurements in the literature. The crack tip indication in measured cases showed variation of about 18 dB in maximum. The creeping wave varied more than 18 dB in the measurements. The mode conversion decreases about 14 dB during the dynamic loading. The corner echo can be detected in most cases very clearly but the sizing, which is the criterion to repair or acceptance, will be affected in all techniques. When crack closes totally then the corner echo can bee difficult to detect. There are some non-linear techniques for measuring the opening of the crack tip in spite of this. As it can be seen the crack tip will be affected by the loading of the crack. The opening of the crack tip can be characterised, but more information is needed to have a clear picture of the opening effect. According to Mihara et al. [2] when the crack tip opening width is less than 0.1 μm the ultrasonic response decreases remarkable. The results of the measurements of this study can be interpreted so that the crack mouth opening and also the crack tip opening were less than 0.1 μm in closing phase of dynamic loading during heating. The changes for instance in the crack tip amplitude measured in this study in the beginning of the cooling let us to assume that the stress changes are about 150 MPa (9 dB in maximum) at the crack tip and 450 MPa (27 dB in maximum) at the surface of the crack according to [13]. The cyclic load affects the crack movement and can be characterised in the beginning of the heating in such away that closing affects crack mouth opening on the surface and introduces an “opening” wave towards the crack tip. Similar effect happens also when cooling starts. This information is important for qualification cases. How to size a closed crack? At which point it is not possible to do it. Are there some techniques which give more information from the closed cracks? The dynamic sizing is certainly one possibility. Information should be collected from detected and by the metallographic inspection of verified cracks. The crack mouth opening has been in some detected defects from dissimilar metal weld inspection very small [25]. How should this affect the qualification trials? Should there be closed cracks in the qualification tests? How should they be estimated? Suggestion is to use partly closed cracks for qualification, so that the inspectors can have some experience during the training. References [1] Becker, F.L., Doctor, S. R., Heasler, P. G., Morris, C. J., Pitman, S. G., Selby, G. P. and Simonen F.A., 1981, Integration of NDE reliability and fracture mechanics, Phase I report Nureg/CR-1696, PNL-3469, Vol. 1, R5, October, 170 p. + appendix 54 p. [2] Mihara, T, Nomura, S., Akino, M. & Yamanaka, K., 2004, Relationship between crack tip scattering and diffraction of longitudinal waves, Materials Evaluation, 62(2004)9, pp. 943-947. [3] Saka, M. & Salam Akanda, M. A., 2004, Ultrasonic measurement of the crack depth and crack opening stress intensity factor under a no load condition. Journal of Nondestructive Evaluation, 23(2004)2, pp. 49 – 63. [4] Denby, D., Duncumb, A.,C., 1985, The effect of stress on the ultrasonic detectability of defects, Nondestructive Testing in the Fitness-for-Purpose Assessment of Welded Constructions, London, UK, 20th – 22nd November 1984, pp. 73-81. 14 [5] Temple, J. A.G., 1985, The effect of stress and crack morphology on time-of-flight diffraction signals, Int. J. of Pressure Vessels & piping, 19(1985)3, pp. 185-211. [6] Ibraham, S. I. & Whittaker, V. N., 1981, The influence of crack topography and compressive stresses on the ultrasonic detection of fatigue cracks, British Journal of NDT, 23(1981)5, pp. 233 – 240. [7] Ahmed, S. R. & Saka M., 1998, A sensitive ultrasonic approach to NDE of tightly closed small cracks, Transactions of ASME, 120(1998)10, pp. 384 -392. [8] Saka, M. & Fukuda, Y., 1991, NDT of closed cracks by ultrasonic propagation along the crack surface. NDT International, 24(1991)4, pp. 191 – 194. [9] Brotherhood, C. J., Drinkwater, B. W. & Guild, F., J., 2002, The effect of compressive loading on the ultrasonic detecability of kissing bonds in adhesive joints, Journal of Non-destructive Evaluation, 21(2002)3, pp. 95 -104. [10] Mouchalin, J.-P., Ochai, M., Lévesque, D., Blouin, A., Talbot, R., & Fukumoto, A., 2003, Characterisation of surface breaking tight cracks using laser-ultrasonic shadowing, Review of progress in quantitative non-destructive evaluation 27th July – 1st August in Green Bay Wisconsin, Vol. 23, AIP, Melville, New York, pp. 1264 – 1271. [11] Böström, A. & Wickham, G., 1991, On the boundary conditions for ultrasonic transmission by partially closed cracks, Journal of Non-destructive Evaluation, 10(1991)4, pp. 139 – 149. [12] Persson G., 1991, Application of the geometrical theory of diffraction to closed cracks, Journal of Non-destructive Evaluation, 10(1991)3, pp. 97 – 109. [13] Zetterwall T. and Borgenstam C.J., 1993, Detektering och storleksbestämning av slutna sprickor, SKI rapport 94:1, November, 1993, 49 p. [14] Akino, M., Mihara, T. and Yamanaka, K., 2003, Fatigue crack closure analysis using non-linear ultrasound, Review of progress in quantitative non-destructive evaluation 27th July – 1st August in Green Bay Wisconsin, vol 23, AIP, Melville, New York, pp. 1256 – 1263. [15] Zagrai, A., Donskoy, D. & Lottiaux, J.-L., 2003, N-SCAN: New Vibro –modulation system for crack detection, monitoring and characterisation, Review of progress in quantitative non-destructive evaluation 27th July – 1st August in Green Bay Wisconsin, Vol. 23, AIP, Melville, New York, pp. 1414 – 1421. [16] Yamanaka, K., Mihara, T., Tsuji, T., 2003, Evaluation of nanoscale cracks by low-pass filter effect in non-linear ultrasound, Proceedings of IEEE Ultrasonics Symposium, Honolulu, 5th – 8th October, 6 p. [17] Yamanaka, K., Mihara, T. & Tsuji, T., 2004, Evaluation of closed cracks by analysis of subharmonic ultrasound, Insight, 46(2004)11, pp. 666 – 670. [18] ASME XI, 2001 Edition, 2001 Div 1; Article C-3300. [19] Pitkänen, J., Kemppainen, M. & Virkkunen, I, 2003, Ultrasonic study of crack under dynamic thermal load, Review of progress in quantitative non-destructive evaluation 27th July – 1st August in Green Bay Wisconsin, Vol. 23, AIP, Melville, New York, pp. 1582 – 1586. [20] Brekow, G., Schulz, E., Erhard, A. and Toffin, O., 1996, Ultraschallmethode zur Rißanalyse. DGzfP Jahrestagung, 13. -15. Mai 1996 in Lindau. pp. 577 – 588. [21] Gebhardt, W., Walte, F., 1989, Crack detection and defect classification using the LLT-technique. Review of progress in quantitative non-destructive evaluation, vol 8A, Plenum Publishing Corporation. 1989, pp. 591 – 598. 15 [22] Silk, M.G., 1987, Changes in ultrasonic defect location and sizing, NDT International, 20(1987)1, pp. 9 – 14. [23] Erhard, A., 1983, Untersuchungen zur Ausbreitung von Longitudinalwellen an Oberflächen bei der Materialprüfung mit Ultraschall, Forschungsbericht 88, BAM Berlin, 32 p. [24] Dupond, O., Bredif, P., Poidevin, C. & De Mathan, N, 2004, Advanced phased array transducer detection of closed crack tip diffraction, NDE in Relation to Structural Integrity for Nuclear and Pressurised Components, 6.-8. December, London UK, pp. 724 – 733. [25] Högberg, K., 2003, PostDas – Verifiering av ET/UT procedure för detektering och höjdbestämning av sprickor, IDSCC, i Alloy 182, Kärnteknik, 19.-20. November, 27 p. 16

PRODUCTION OF REALISTIC ARTIFICIAL FLAW IN INCONEL 600 SAFE-END Mika Kemppainen, Trueflaw Ltd., Espoo Finland Iikka Virkkunen, Trueflaw Ltd., Espoo Finland Jorma Pitkänen, VTT Industrial Systems, Espoo, Finland Kari Hukkanen, TVO Oy, Olkiluoto, Finland Hannu Hänninen, Helsinki University of Technology, Espoo, Finland The importance of NDT qualification has received significant attention during the recent years. Recent findings of cracks in Inconel 600 in different NPP components have also increased interest in the reliability of in-service inspections of this material. This, in turn, sets challenge for manufacturing of representative qualification specimens and flaws. A new, advanced flaw production technique has become available. The technique enables production of realistic cracks to ready-made mock-ups without implanting or welding. This paper describes the advanced crack production technique and its application to Inconel 600. A realistic, controlled crack was produced to a core spray nozzle safe-end mock-up. The technique produces true fatigue cracks, which are representative of most real, service-induced cracks. The technique is applicable to any shape or size of component and results only in an intended crack without unwanted disturbances. The technique allows production of a single or separate cracks as well as different combinations of them. In addition to the controlled crack production, the paper introduces studies of the effects of different thermal fatigue loading cycles on the ultrasonic response obtained from the crack in Inconel 600. Results of the study show the effect of different thermal fatigue loading cycles on the obtained ultrasonic response during dynamic loading of the artificially produced crack. Control of crack growth and relationship between loading parameters and ultrasonic response are discussed. Introduction The last decade has brought new challenges for the nondestructive testing in the nuclear power field. Several through-the-wall leakages in components and structures that have not been covered by in-service inspection programs have gathered attention of the whole nuclear community. One of current concerns is the primary water stress corrosion cracking of Inconel 600 alloy and its weld metals in the pressure vessel head and bottom penetration nozzles. This type of degradation and crack growth was not originally considered in components in question. The NDE qualification procedures are still under development all over the world. This includes development of better flaw production techniques producing representative flaws. There are certain factors that have to be taken into account when a flaw is used as a reflector for ultrasonic inspection. The ultrasonic response is affected by different crack characteristics, among others, location, orientation and size of a crack1, the opening of a crack and crack tip2,3,4, the remaining residual stresses in the material5,6, fracture surface roughness7,4, crack tip plastic zone8 and filling of the crack with water9. These characteristics of cracks affect propagation, reflection, diffraction, transmission, attenuation and diffusion of ultrasonic energy9,10. Wüstenberg et al.11 mentioned, that if the main interaction of a flaw used in qualification is based on the crack tip diffraction, the only possibility would be use of service-induced flaws as cut outs from real components and weld implant them to qualification mock-ups. This was based on the fact that there was no flaw production technique capable of producing realistic cracks or flaws which represent sufficient weak crack tip diffraction. Hence, there is a need to develop a flaw manufacturing technique that is capable of producing realistic flaws representative from all typical characteristics point of view. A novel artificial flaw production technique and its applicability for Inconel 600 is introduced in this paper. The technique is used to produce a realistic crack in a core spray nozzle mock-up component of a BWR-type nuclear power plant. Furthermore, the ultrasonic response of the crack under dynamical thermal loading was studied in order to understand the relationship between ultrasonic response and different crack opening conditions. Materials and Methods The flaw production technique is based on thermal fatigue loading. Loading is applied by high frequency induction heating and water or air spray cooling. Produced flaws are representative of real, service-induced fatigue flaws in metallographic sense and hence they are supposed to be representative also in terms of NDE response. The technique allows production of realistic flaws with controlled location, orientation and size. Characteristics of flaws produced with the technique are introduced in more detail in references 12,13,14. The technique is applicable to different materials and virtually any shape or size of a sample. The only requirement for the crack production is that the intended location must be accessible. Sample This paper introduces flaw production to a full-size core spray nozzle, safe-end mock-up (BWR-type nuclear power plant). Figure 1 shows the nozzle consisting of three different materials: A508 carbon steel, Inconel 600 and AISI 316 type austenitic stainless steel. There is a buttering and a joint weld between the carbon steel (with cladding on the inner surface) and Inconel 600 safe-end, and a butt weld between Inconel 600 safe-end and AISI 316 austenitic stainless steel pipe. Both welds were made with Inconel 182 filler material with Inconel 82 root pass. After welding the working allowances were machined away. The finishing machining removed the root pass so, that the welds of the ready-made mock-up are Inconel 182. Figure 1 Core spray nozzle mock-up with Inconel 600 safe-end. Figure 2 shows the drawing of the nozzle mock-up and the intended location of the flaw production. The intended location is in Inconel 600 in the HAZ of the buttering weld. The wall thickness of the Inconel 600 safe-end in the intended location is 23 mm. Nozzle was received as ready-made and no machining or welding was allowed. Flaw was to be produced to the inner surface in as-received condition of the nozzle. The specimen was nondestructively tested after flaw production and no destructive tests were performed. Figure 2 Drawing of the core spray nozzle and location of the produced crack in Inconel 600 safe-end in the heat affected zone of the buttering weld. NDT set-up A pulse-echo shear wave probe (41°, 1.5 MHz) was used when performing the inspection of the nozzle after crack production. The same probe was used during the studies of the relationship between ultrasonic response and crack loading. These studies were performed with a ready-made crack. The probe was attached on the outside surface of the mock-up and the surface breaking crack in the inner surface was monitored through the wall, in front of the weld. Ultrasonic signals were gathered in-situ during continued thermal fatigue cycling of the crack. Details about the NDT measurement system are given in reference15. Applied loads In order to study the effect of different loadings, two different thermal fatigue loading cycles were applied. Temperature curves of applied cycles are shown in Figure 3 as measured from the sample surface. The first cycle (B1) had high heating rate and short cooling time with heating and cooling times of 10 and 15 s, respectively. The second cycle (B2) had lower heating rate and longer cooling time with heating and cooling times 20 and 25 s, respectively. Water spray cooling was applied for both cycles. The first cycle reached higher temperature than the second cycle. In order to see the effect of the stabilised cycles, B1 loading was applied as 20 and B2 as 16 successive cycles. Figure 3 Two different temperature loading cycles used in the studies. FEM-analysis Applied cycles were analysed by finite element modeling (FEM) giving results of temperature and strain distributions through the material thickness during dynamical loading. Used finite element model is presented in more detail in reference16. Results A realistic crack was produced in the inner surface of the nozzle. Figure 4 shows the dye penetrant indication of the produced single crack in Inconel 600 safe-end in the heat affected zone of the buttering weld. The weld is located in the upper part and Inconel 600 base material in the lower part of the figure. The length of the crack is 14.2 mm and the depth is 5 mm, thus being about 22% through the wall. The maximum surface opening of the crack varies locally between 30 – 45 µm. In the figure, there is also a very small (less than 1 mm deep) secondary indication in the corner of the shoulder visible in the lower part of the figure. The initiation of the secondary crack was caused by the stress rising effect of the shoulder. Without vicinity ofsuch a stress riser, there would have been no secondary cracking. The secondary indication does not affect the performance of ultrasonic testing as it is located about 7 mm away from the actual crack. Figure 4 Dye penetrant indication of the produced realistic crack in Inconel 600 safe-end in the heat affected zone of the buttering weld. The size of the crack was controlled by process control during the production and confirmed by ultrasonic testing. The obtained signal from the crack at room temperature is shown in Figure 5. The reflections from crack opening corner and subsurface parts of the crack are visible in the figure. The ultrasonic inspection sized the crack to be 18 mm long and 6 mm deep. The measured length by ultrasonic testing is clearly bigger than the actual value as seen from Figure 4. Also the measured depth differs from the given process value, but it lies inside the production tolerances (±1 mm). Figure 5 A-scan obtained from the crack at room temperature (41°, 1.5 MHz, shear wave probe). The studies of ultrasonic response versus dynamical thermal loading resulted in a large amount of ultrasonic data. Figure 6 shows the ultrasonic signal obtained from the crack in the end of cooling and heating phases of cycle B2. The figure clearly shows the differences between different crack opening states. Results shown in the figure have been obtained in the turning points of surface temperature cycles. Figure 6 A-scans from the crack in the end of cooling and in the end of heating of thermal fatigue loading cycle. Differences in ultrasonic response are related to the crack opening and closing behaviour. Results of finite element modeling gave temperature and strain distributions through the wall thickness. Figures 7 and 8 show solved strain distributions for analysed cycles B1 and B2, respectively. Nozzle ID is in the left side and OD on the right side of both figures. The results clearly show the difference between the faster and slower loading rates. Figure 7 Strain distribution for loading cycle B1. Nozzle ID on the left and OD on the right side of the figure. Figure 8 Strain distribution for loading cycle B2. Nozzle ID on the left and OD on the right side of the figure. The obtained ultrasonic signal amplitudes from corner reflection and crack tip varied during the loading. These variations are related to the opening behaviour of the different parts of the crack. Figures 9 and 10 show the combined results of strain variations from modeling and measured changes of ultrasonic amplitudes from corner reflection and crack tip for cycles B1 and B2, respectively. Figure 9 Combined results of strain and ultrasonic amplitude variations from crack corner and tip caused by loading cycle B1. Figure 10 Combined results of strain and ultrasonic amplitude variations from crack corner and tip caused by loading cycle B2. Discussion The results show that a realistic crack was produced in the heat affected zone of the buttering weld in Inconel 600 safe-end, as intended. The flaw location and size were accurately controlled. The dye penetrant indication shows a single, tortuous crack, which has a natural propagation in the heat affected zone of the buttering weld. The crack is narrow and its opening varies in different parts of the crack. The ultrasonic response is determined to be a crack-like indication. Similarly, the amplitudes from corner, face and crack tip are representative and set realistic challenge for the inspection. Ultrasonically the produced crack represents a difficult reflector caused by its realistic characteristics. The realistic crack causes unhomogeneous reflections affecting the detection. The tight crack tip and small crack tip radius make the sizing of the crack challenging. It was shown that the technique is applicable to ready-made mock-up without causing any alterations to the component. The results show, that the technique fulfills the important factors to be taken into account when performance demonstration is designed and an artificial flaw is used as a reflector. These factors include correspondence of reflector dimensions and dynamic range of echo amplitude, representativeness of position, orientation, fracture surface roughness and reproducibility of the artificial reflector both metallografically and echodynamically1,11. The results of ultrasonic response versus thermal fatigue loading show how different parts of the crack are opening and closing at different time moments. For example, the corner amplitude decreased during heating and increased during cooling. While the crack tip amplitude increased during heating and decreased during cooling. That is, crack tip amplitude changes were opposite to the corner amplitude. Amplitude decrease is caused by crack closure and increase by opening of the crack. It is known that the surface breaking part of the crack is closed during heating and opened during cooling as described, e.g., in reference17. However, the ultrasonic results of the crack tip amplitude show, that the tip is openend during heating and closed during cooling. This is caused by temperature cycling inducing stress gradients in the specimen. During heating the surface layer of the material is heated up and experiencing increased compressive stresses. At the same time, subsurface parts of the crack are at lower temperature and may be under tensile stress. The increase of crack tip amplitude during heating clearly indicate that the crack tip is opened, i.e. under tensile stress. The finite element modeling, however, shows different results for the strain variations in the depth of the crack. For both analysed cycles the model shows decreasing strains during heating and increasing strains during cooling at the crack tip. This is explained by the fact that the model was made for solid material and does not take into account the flaw in the material. Conclusions The novel artificial flaw production technique is available for different materials including Inconel 600. The technique is applicable to full size mock-ups with challenging multi-material structures. Flaw production does not cause any unwanted alterations and is applied to ready- made, finished surfaces. The produced flaws are realistic thermal fatigue cracks. Cracks are tortuous, tight, narrow and have a small crack tip radius. Hence, the reflection properties of produced cracks are realistic. Flaws produced with the new technique can be used in NDE training and qualification purposes. The accurate positioning, control of crack size and reproducibility offer an opportunity to have realistic reflectors in testing, training or qualification specimens. The production process does not set any requirements for the specimen and, hence, also specimens with existing flaws can be used. Acknowledgements This work was performed in a research and development project funded by Technology Agency Finland (Tekes), Trueflaw Ltd., Pacific Northwest National Laboratory (PNNL, USA), TVO Oy and Fortum Nuclear Services Ltd. The participants are acknowledged for giving the funding, delivery of test materials and technical support. References 1. G. Waites, C. and Whittle, J., 1998. The Status of Performance Demonstration and Evaluation Developments. Insight, 40 (12), December, pp. 810-813. 2. Ahmed, S.R. and Saka, M., 1998. A Sensitive Ultrasonic Approach to NDE of Tightly Closed Small Cracks. Journal of Pressure Vessel Technology, Transactions of the ASME, 120, November, pp. 384-392 . 4. Wirdelius, H. and Österberg, E., 2000. Study of Defect Characteristics Essential for NDT Testing Methods ET, UT and RT. SKI Project Number 98267, SKI Report 00:42, October, Sweden. 50 p. 3. Yoneyama, H., Senoo, M., Miharada, H. and Uesugi, N., 2000. Comparison of Echo Heights between Fatigue Crack and EDM Notch. Proceedings of 2nd International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurized Components, 24-26 May, New Orleans, Louisiana, U.S.A. 8 p. 5. Gauthier, V., 1998. Thermal Fatigue Cracking of Safety Injection System Pipes Non Destructive Testing Inspections Feedback. Proceedings of NEA/CSNI Specialists' Meeting on: Experiences with Thermal Fatigue in LWR Piping Caused by Mixing and Stratification, 8-10 June, Paris, France. pp. 436-453. 6. Iida, K., Takumi, K. and Naruse, A., 1988. Influence of Stress Condition on Flaw Detectability and Sizing Accuracy by Ultrasonic Inspection. The Ninth International Conference on Nondestructive Evaluation in the Nuclear Industry, 25-28 April, Tokyo, Japan. pp. 563-567. 7. Ogilvy, J.A., 1989. Model for the Ultrasonic Inspection of Rough Defects. Ultrasonics, 27, pp. 69-79. 8. Saka, M., Fukuda, Y., 1991. NDT of Closed Cracks by Ultrasonic Propagation along the Crack Surface. NDT&E International, 24 (4), pp. 191-194. 9. Becker, F.L., Doctor, S.R., Heasler, P.G., Morris, C.J., Pitman, S.G., Selby, G.P. and Simonen, F.A., 1981. Integration of NDE Reliability and Fracture Mechanics - Phase I Report. NUREG/CR-1696 PNL-3469, 1. 170 p. 10. Ibrahim, S.I. and Whittaker, V.N., 1981. The Influence of Crack Topography and Compressive Stresses on the Ultrasonic Detection of Fatigue Cracks in Submerged Arc Welds. British Journal of NDT, September, pp. 233-240. 11. Wüstenberg, H. and Erhard, A., 1994. Problems with Artificial Test Reflectors at the Performance Demonstration of Ultrasonic Inspections. Proceedings of 6th European Conference on Non Destructive Testing, Nice, pp. 741-746. 12. Kemppainen, M., Virkkunen, I, Pitkänen, J., Paussu, R. and Hänninen, H., 2002. Realistic Cracks for In-Service Inspection Qualification Mock-ups. Proceedings of the 8th European Conference on Non-destructive Testing, Barcelona, Spain. 13. Kemppainen, M., Virkkunen, I, Pitkänen, J., Paussu, R. and Hänninen, H., 2002. Comparison of Realistic Artificial Cracks and In-service Cracks. Proceedings of the 8th European Conference on Non-destructive Testing, Barcelona, Spain. 14. Kemppainen, M., Virkkunen, I, Pitkänen, J., Paussu, R. and Hänninen, H., 2003. Advanced Flaw Production Method for In-service Inspection Qualification Mock-ups. Journal of Nuclear Engineering and Design, 224, pp. 105-117. 15. Pitkänen, J., Kemppainen, M., Virkkunen, I., and Hänninen, H., 2003. Ultrasonic Study of Crack under a Dynamic Thermal Load. Proceedings of Review of Progress in Quantitative Nondestructive Evaluation QNDE, Melville, New York. To be published. 16. Virkkunen, I., Kemppainen, M., Pitkänen, J. and Hänninen, H., 2003. Effect of Thermal Stresses along Crack Surface on Ultrasonic Response. Proceedings of Review of Progress in Quantitative Nondestructive Evaluation QNDE, Melville, New York. To be published. 17. Kemppainen, M., Virkkunen, I., Pitkänen, J. and Hänninen, H., 2003. Advanced Flaw Manufacturing and Crack Growth Control. Proceedings of Review of Progress in Quantitative Nondestructive Evaluation QNDE, Melville, New York. To be published.

Nuclear Engineering and Design 224 (2003) 105–117 Advanced flaw production method for in-service inspection qualification mock-ups Mika Kemppainena,∗, Iikka Virkkunena, Jorma Pitkänenb, Raimo Paussuc, Hannu Hänninend a Trueflaw Ltd., P.O. Box 540, FIN 02151 Espoo, Finland b VTT Industrial Systems, Espoo, Finland c Fortum Nuclear Services Ltd., Espoo, Finland d Helsinki University of Technology, Espoo, Finland Received 21 October 2002; received in revised form 27 February 2003; accepted 8 March 2003 Abstract One of the key issues in in-service inspection qualification is the representativeness of the defects used in qualification specimens. The best representativeness is achieved with realistic defects. However, present specimen production techniques have some significant weaknesses, such as unrealistic defects or additional alterations induced in the surrounding material. Specimens manufactured, for example, by weld implantation or with weld solidification defects always result in one or more extra weld interfaces. These interfaces can be detected by NDT. To overcome problems with the current specimens, a new defect manufacturing technique was developed. The new technique produces natural, representative defects without introducing additional weld metal or other unwanted alterations to the specimen. The new method enables artificial production of single, separate fatigue cracks by thermal loading. The method is based on a natural thermal fatigue damage mechanism and enables production of real cracks directly into the samples. Cracks are produced without welding or machining and without any preliminary surface treatment or artificial initiator such as a notch or a precrack. Single crack or a network of cracks can be induced into the base material, welded areas, HAZ, weld claddings, threaded areas, T-joints, etc. The location, orientation and size of produced cracks can be accurately controlled. Produced cracks can be used to simulate different types of service-induced cracks such as thermal fatigue, mechanical fatigue and stress corrosion cracks. It is shown that artificially produced thermal fatigue cracks correspond well with the real, service-induced cracks and overcome the problems of traditional qualification specimen manufacturing techniques. © 2003 Elsevier Science B.V. All rights reserved. ∗ Corresponding author. Tel.:+358-45-6354414; fax: +358-9-455-3117. E-mail addresses: mika.kemppainen@trueflaw.com (M. Kemppainen), iikka.virkkunen@trueflaw.com (I. Virkkunen), jorma.pitkanen@vtt.fi (J. Pitkänen), raimo.paussu@fortum.com (R. Paussu), hannu.hanninen@hut.fi (H. Hänninen). 1. Introduction Thermal fatigue, that is, material degradation due to successive temperature changes, is one of the life-limiting mechanisms in nuclear power plant con- ditions. During the operation of a power plant ther- mal fatigue cracks can initiate and grow in various components. Causes for this are mixing, striping or stratification of hot and cold water (Hytönen, 1998). 0029-5493/$ – see front matter © 2003 Elsevier Science B.V. All rights reserved. doi:10.1016/S0029-5493(03)00078-5 106 M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 A typical component, where thermal fatigue cracking occurs, is a T-joint where hot and cold fluids meet and mix. The turbulent mixing of fluids with different temperatures induces rapid temperature changes to the pipe wall. The resulting uneven temperature distribu- tion prevents thermal expansion and gives rise to ther- mal stresses. The successive thermal transients cause varying, cyclic thermal stresses. These cyclic ther- mal stresses cause fatigue crack initiation and growth similar to cyclic mechanical stresses. (Virkkunen, 2001). Cracks occur in nuclear power plant components in different locations such as in straight pipe sections, valve bodies, pipe elbows, collector head screw holes, etc., as well as in base material and in weld joints (ASME, 1990). Crack growth direction depends on the component and the location, based on the local shape effect and the loading conditions. Pipe cracks can grow both axially and circumferentially, in weld joints both parallel to the weld in the heat-affected zone (HAZ) and transverse to the weld, in threads both vertically and horizontally. Thermal stresses are typically equi-biaxial and they are highest at the loaded surface. The loading is strain controlled and very high local stresses can arise. If the stresses locally exceed the yield strength of the material, thermally induced residual stresses arise (Virkkunen et al., 2000). Due to the high sur- face stresses, the thermal fatigue cracks often form a mosaic-like crack pattern of shallow cracks. There have been several incidents showing that some of the shallow cracks extend deeper into the material and can grow through the wall thickness (ASME, 1990; Hänninen and Hakala, 1981). During the in-service inspections thermal fatigue cracks create a challenge for ultrasonic inspection, both for detection and sizing. The difficulty of the in- spection is caused by the characteristics of the cracks, which cause the ultrasonic energy to diffuse, attenu- ate, scatter, diffract, etc. (Becker et al., 1981). Typical characteristics affecting the ultrasonic inspection are crack opening (Yoneyama et al., 2000), fracture sur- face roughness (Ogilvy, 1989), branching (Wirdelius and Österberg, 2000), residual stress (Iida et al., 1988), plastic zone (Saka and Fukuda, 1991), etc. Each of these factors affects the performance of ul- trasonic testing. Therefore, it is crucial to identify the key characteristics of service-induced and artificially induced flaws and to understand how they affect the ultrasonic response obtained during inspection. The most effective in-service inspection qualifica- tion is achieved with qualification specimens, which are as representative as possible of all service con- ditions. In order to have similar physical reflector to that of service-induced crack, it is commonly under- stood that real or realistic cracks must be used in the in-service inspection qualification mock-ups. Prob- lems with the artificial defect incorporation methods have, so far, restricted the use of realistic cracks in qualification mock-ups. The developed method that is described in this paper allows controlled production of realistic cracks that have NDE responses similar to service-induced cracks. In this paper, the performance of the new method for producing artificial flaws with thermal fatigue is demonstrated and validated using the results of de- structive and non-destructive tests. Destructive tests re- veal typical metallographic characteristics of produced cracks. Metallographic results are compared to corre- sponding service-induced cracks documented in the open literature. Non-destructive results are compared to corresponding results from EDM-notches and real, service-induced cracks. The feasibility of the method is demonstrated by three different cases where artifi- cial defects have been produced in real components. 2. Experimental Different test samples were produced in order to verify the new crack production method. Samples were studied both non-destructively and destructively. Ther- mal fatigue loading was applied with high frequency induction heating and water or air cooling in order to achieve high heating and cooling rates. Controlled ini- tiation and growth of cracks were followed by replica assisted light optical microscopy. Ultrasonic examina- tions were performed to determine the non-destructive response of the cracks produced. Destructive testing and scanning electron microscopy were carried out to study the microstructural propagation and fracture surface morphology. Ultrasonic response as well as metallographic properties of artificial cracks were compared to properties of real, service-induced cracks. The manufacturing process employed does not re- quire any crack starter or artificial initiator and it is M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 107 Fig. 1. Two surface breaking thermal fatigue cracks: (a) surface length approximately 8 mm (30,021 thermal fatigue cycles) and (b) surface length approximately 20 mm (166,971 thermal fatigue cycles) in AISI 304 type austenitic stainless steel plate. applied to the actual surface of a component. Con- sequently, the process does not leave any unwanted alterations to the material, which may be detected in ultrasonic or eddy current examination. Crack properties affecting the non-destructive ex- amination include among others: location and ori- entation, microstructural propagation, branching, crack opening/closure and surface roughness (Rz) of the crack and the shape of the crack tip. These parameters affect both detection and sizing of service-induced cracks. After characterising the prop- erties of artificially produced thermal fatigue cracks they were compared to corresponding properties of real, service-induced thermal fatigue cracks. The new defect production method utilises the same, natural thermal fatigue damage mechanism, which is also present in real components. The method enables controlled defect production in similar locations and directions as those where real cracks occur. For the artificial crack production, there are no limitations for component shape or size, or for crack orientation or location. 2.1. Microscopy In austenitic stainless steels thermal fatigue cracks initiate from slip bands. From the multitude of initial microcracks one single crack can be grown by con- trolled thermal loading. InFig. 1 there are two exam- ples of artificially produced surface breaking cracks. Cracks follow the crystallographic path through the microstructure. The crack path is tortuous, but the macroscopic surface crack growth follows the prede- termined direction. InFig. 2, cross-sections of two different cracks are shown, where tortuous and trans- granular crack propagation can be seen. The propaga- tion is different for each crack because these cracks were produced with different thermal fatigue parame- ters. The cracks show minor branching, they are nar- row and the crack tip radii are small. The crack depth is controlled cycle-by-cycle as the produced crack depth depends on the magnitude of the applied thermal cycle and the total number of cycles. Artificially produced cracks have rough fracture surfaces where the per-cycle crack extension Fig. 2. Cross-sections of two artificially produced thermal fatigue cracks in AISI 304 type austenitic stainless steel with different fracture surface roughness resulting from different production pa- rameters. Cracks after (a) 30,000 thermal fatigue cycles and (b) 2770 thermal fatigue cycles. 108 M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 Fig. 3. Artificially produced thermal fatigue crack (a) has fatigue striations on the fracture surface and (b) follows elliptical shape in its growth (82% of the wall thickness penetrating crack). White arrows show crack growth directions. Material is AISI 304 type austenitic stainless steel. can be followed from the fatigue striations. Each striation represents an incremental advance of the crack front during one load cycle. The magnitude of the increment, that is, striation spacing, depends on the stress/strain range. Typically the striation spacing varies in different areas of the crack. In a small crack the spacing is small, but it increases as the crack gets bigger. The crack growth follows an elliptical shape, which in the limit approaches a circular configura- tion. Typical fracture surface and shape of a big crack are shown inFig. 3. The white arrows in the picture show the crack growth direction. InFig. 3b half of the fracture surface of an 82% through wall thickness (7 mm deep) crack is shown. The initiation surface, that is, pipe inner surface, is in the upper part of the figure and the outer surface of the sample pipe is shown in the lower part of the figure. 2.2. Metallographic comparison of artificially produced and service-induced thermal fatigue cracks The non-destructive response reveals the micro- structural characteristics of the service-induced crack. In order to have similar non-destructive response, artificially produced cracks should simulate the mi- crostructural properties of service-induced cracks. In Fig. 4, a service-induced and an artificially produced thermal fatigue crack are presented side by side. The macroscopic propagation of real and artificially produced thermal fatigue cracks is similarly tortuous. As shown inFig. 4, both cracks are narrow, propa- gate transgranularly in the microstructure and show minor branching. The real, service-induced crack has the larger opening or width near the surface and it is tight in the vicinity of the crack tip. The typical crack surface width of a service-induced thermal fatigue crack in austenitic stainless steels varies between 5 Fig. 4. Comparison of cross-sections of (a) a service-induced (Hänninen and Hakala, 1981) and (b) an artificially produced thermal fatigue crack (6500 thermal fatigue cycles). M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 109 Fig. 5. Comparison of crack tips of (a) a service-induced (Wåle and Ekström, 1995) and (b) three artificially produced thermal fatigue cracks (after 30,000 thermal fatigue cycles). and 380
m, at one half the through wall depth of the crack between 2 and 190
m and at the crack tip between 1 and 18
m (Wåle and Ekström, 1995). The opening of a crack is affected by many factors Fig. 6. Comparison of fracture surfaces of (a) a service-induced (Pirson and Roussel, 1998) and (b) an artificially produced thermal fatigue crack. including the condition of the residual stress along the fracture surface. Fracture surface of a service-induced thermal fatigue crack is rough. Typical surface roughness (Rz) values in austenitic stainless steels 110 M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 vary between 6 and 140
m (Wåle and Ekström, 1995). In Fig. 4 the artificially produced thermal fatigue crack has a similar opening behaviour to that of the service-induced crack. The approximate width of the artificially produced thermal fatigue crack is 70
m at surface, in the middle-section 40
m and at the crack tip 2
m. Typical values of surface roughness (Rz) for artificially produced thermal fatigue crack varies between 35 and 125
m. Surface roughness was cal- culated digitally from cross-sectional images. The difference of height between ten highest peaks and ten lowest valleys on the fracture surface were mea- sured for each image and the average of measurement results was reported asRz. The tortuous crack path has formed when the crack propagation direction changes crystallographically at grain boundaries in the microstructure.Fig. 5shows a comparison of crack tips of service-induced and arti- ficially produced thermal fatigue cracks. The crack tip of a service-induced thermal fatigue crack has typically a very small radius and crack surfaces near the tip are close to each other, which is shown inFig. 5. The tip of artificially produced ther- mal fatigue cracks is similarly tight and has a very small radius. Fatigue crack growth can be followed from the fracture surface from fatigue striations. In Fig. 6, a comparison of striations on fracture surfaces of a service-induced and an artificially produced thermal fatigue crack is shown. Striations on the fracture surfaces of service-induced and artificially produced thermal fatigue cracks are similarly visible, as shown inFig. 6. The cycle- dependent incremental crack growth makes it possible to control the artificial crack production accurately. Fig. 7. Ultrasonic A-scans (time–amplitude) from an artificially produced circumferential thermal fatigue crack obtained with 70◦T probe (MWK 70-2E). 2.3. Non-destructive testing Artificially produced thermal fatigue cracks in austenitic stainless steel pipe base material and austenitic stainless steel strip welded cladding were characterised by ultrasonic and eddy current meth- ods. The objective of ultrasonic measurements was to study the detectability of artificially produced thermal fatigue cracks and to characterise the cracks from the ultrasonic signal point of view. Characterisation included studying the ultrasonic signal obtained from crack corner reflection effect, fracture surface and crack tip. Additionally, the variation of ultrasonic signal from different areas of the crack was studied. The objective of using the eddy current method was to study the detectability and sizing capability in austenitic stainless steel cladding. 2.3.1. Base material Two cracks were produced (one in axial and the other in circumferential directions) in austenitic stain- less steel pipe base material (Fig. 17). The diameter of the pipe was about 360 mm and wall thickness 28 mm. Both cracks were placed in the inner surface of the pipe. The surface length of the circumferential crack was 20 mm and depth approximately 6 mm. The length of the axial crack was 20.5 mm and depth approximately 3 mm. Cracks in the base material were characterised ultrasonically with longitudinal, transverse and sec- ondary creeping waves. Two techniques with longitu- dinal waves, 0◦L (longitudinal) 5 MHz single element probe and ADEPT60◦L 3 MHz dual element special probe, were used. With transverse waves two differ- ent 2 MHz composite probes (MWK 45-2 and MWK M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 111 Fig. 8. Ultrasonic A-scan from an artificially produced circum- ferential thermal fatigue crack obtained with 45◦T probe (MWK 45-2). Fig. 9. Ultrasonic corner echo dynamics from an artificially pro- duced circumferential thermal fatigue crack obtained with sec- ondary creeping wave probe (WSY 70-2). 70–2E) and 2 MHz mode conversion probes (WSY 70-2) were used. The inspection results obtained us- ing these different probes from artificially produced circumferential crack are shown inFigs. 7–10. In the austenitic stainless steel base material the tip of circumferential crack was detected with both types of ultrasonic longitudinal wave probes as shown in Fig. 10. Ultrasonic A-scan from an artificially produced circumfer- ential thermal fatigue crack obtained with mode conversion probe (ADEPT60◦L). Fig. 11. Eddy current signal from (a) an artificially produced thermal fatigue crack in austenitic stainless steel strip welded cladding and (b) three EDM-notches (depths 2.0, 1.0 and 0.5 mm) in austenitic stainless steel base material. Fig. 15 (0◦L) and Fig. 10 (ADEPT60◦L). However, with 0◦L no proper echo from the crack face was detected. With transverse waves, echoes from crack tips were detected both with 70◦T (S/N approximately 10 dB) and 45◦T (S/N approximately 15 dB) 2 MHz probes (Figs. 7 and 8, respectively). The obtained signal-to-noise ratio results show that with transverse waves the crack tip was more clearly seen with the 45◦ probe than with 70◦ probe. During ultrasonic testing with the mode conversion probe (Fig. 9) the received secondary creeping wave echo from the transverse crack disappeared locally. This is caused most likely by a discontinuous crack extension or local compressive residual stress, which presses the crack surfaces closely together, thus en- abling ultrasonic waves to penetrate through the crack. 2.3.2. Cladding One crack was produced in a strip welded AISI 316 type austenitic stainless steel cladding of a fer- ritic steel test block (465 mm× 150 mm× 100 mm, length× width × height). The cladding was 10 mm thick and the produced crack was about 10 mm long and 3 mm deep. Three different size (depths 0.5, 1.0 and 2.0 mm) of rectangular (70 mm long and 0.3 mm wide) EDM-grooves used for eddy current calibration were produced in a separate austenitic stainless steel plate. The eddy current inspection was performed with a 100 kHz probe (Zetec 195-801P02 Fe). The crack was detected in the cladding, although the impedance level varied depending on the-ferrite content and the probe location on the strip welded cladding (middle of strip, HAZ, fusion line). The crack was estimated to be about 3 mm deep.Fig. 11shows eddy current signals from the artificially produced crack and from three different EDM-grooves used for calibration. Also the 112 M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 Fig. 12. Ultrasonic testing results of an artificially produced thermal fatigue crack in austenitic stainless steel cladding: (a) from outside surface with long sound path with 41◦T (1 MHz) wave probe and (b) from crack opening surface with 70 TRL (2 MHz) wave probe. sizing results of eddy current testing may be affected by the possible compressive residual stress affecting the crack tip. Ultrasonic inspection of the cladding sample was performed from the crack opening surface (with 70◦TRL, 2 MHz probe) and with a long sound path (about 200 mm) from the outside surface of the test block (with 41◦T, 1 MHz probe). The results inFig. 12 show two ultrasonic B-scans in the bottom line and one C-scan above them. The orientations of scan di- rections are visualised in the figure. The crack was detected from both sides of the test block, but no crack tip echo was detected in either case. This is explained by the limited resolution of the probing frequencies used and a possible compressive residual stress closing the crack tip. 2.4. Non-destructive comparison of artificially produced and service-induced thermal fatigue cracks The non-destructive testing results obtained from artificially produced thermal fatigue cracks were compared to the results obtained with an EDM-notch and a real service-induced thermal fatigue crack. The service-induced thermal fatigue crack was caused by mixing of fluids at different temperatures in a T-joint process pipe. The cracked area was found based on pipe leakage. By closer inspection, numerous transversal cracks were found, one of them penetrat- ing through the wall of the pipe. A 13 mm deep crack was selected to be used in this study. The objective of the comparison of different reflec- tors was to show the difference between ultrasonic re- sponses obtained. The inspection of all three different types of reflectors was performed during the same ses- sion with the same equipment and personnel. The arti- ficial thermal fatigue crack was produced in the same pipe containing the service-induced crack. The pipe material was AISI 321 type austenitic stainless steel (Ti-stabilised). The third reflector, a semi-elliptical EDM-notch, was in a separate plate (thickness 30 mm) of AISI 316 type austenitic stainless steel. Crack tip echoes obtained with 0◦L 5 MHz sin- gle probe from the EDM-notch, the service-induced thermal fatigue crack and the artificially produced thermal fatigue crack are presented inFigs. 13–15, respectively. The crack tip echo is clearly seen from EDM-notch but from the real and artificially produced thermal fatigue cracks the S/N ratio is low, being about 6–10 dB. However, the S/N ratios of crack tip signals were about the detection limit value (6 dB). With artificially produced thermal fatigue crack the crack tip signal was not equally clear over the whole length of the crack. The signal shown inFig. 15 is from an area where the crack tip signal was optimal. Fig. 13. Ultrasonic A-scan of a 15 mm deep semi-elliptical EDM-notch in AISI 316 type austenitic stainless steel (V110-0◦L). M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 113 Fig. 14. Ultrasonic A-scan of a 13 mm deep real, service-induced thermal fatigue crack in AISI 321 type austenitic stainless steel (V110-0◦L). 2.5. Residual stress Residual stress state affects the opening of the produced crack, which in turn affects the reflection and transmission behaviour of the ultrasonic energy. Residual stress can be measured on the surface of the sample with relative ease, while measurement through the thickness of the sample near a crack is difficult. Fig. 16 shows an example of cumulative change of surface residual stress caused by continued thermal fatigue cycling. Residual stresses were mea- sured in two directions during test interruptions when the sample had cooled down to room temperature. During the test, an individual, 8 mm long crack was produced with a total of 35,000 cycles. Residual stress measurements were performed in the mid-section of the crack in the solid material, 1 mm away from the crack opening. Measurements were done with X-ray diffraction, which reveals stresses from a thin surface layer (thickness 15–30
m). Measurements were per- formed with Cr radiation, 3 mm (diameter) collimator Fig. 15. Ultrasonic A-scan of an artificially produced circumfer- ential 20 mm long and approximately 6 mm deep thermal fatigue crack in AISI 321 type austenitic stainless steel (V110-0◦L). Fig. 16. Change of surface residual stresses as a function of con- tinued thermal fatigue cycling in the vicinity of crack produced in AISI 304 type austenitic stainless steel plate (T: stresses transverse to the crack, P: stresses parallel to the crack). and 60 s exposure time. Residual stress measure- ments through the thickness of the sample were not performed in this study. The change of surface residual stresses inFig. 16 demonstrates that the applied thermal fatigue cycle has a strong effect on the final residual stress distribution. Residual stress transverse (marked as “T”) and parallel (marked as “P”) to the crack show opposite behaviour as transverse stress moves very quickly in tension and parallel stress in compression. In this example ther- mal cycling was controlled so that the crack opening is pronounced, that is, transverse stress was set in ten- sion. Thus, by controlling the thermal cycling also the residual stresses can be controlled. 2.6. Real components The developed defect production method works well both from the metallographical and non-destruc- tive response point of view. A further interest was to verify the applicability of the method to real com- ponents. Verification has been done with different types of components: pipe section, butt-welded pipe, T-joint of two pipes and a collector head threaded screw hole. Materials have been different types of austenitic stainless steels. Results of these tests are presented in the following. 2.6.1. Pipes Three different sizes of pipes were examined dur- ing the experiments. Cracks were produced on the in- ner surface of the pipes. InFig. 17is an example of a pipe (diameter approximately 360 mm, wall thickness 114 M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 Fig. 17. Austenitic stainless steel pipe: (1) circumferential 20 mm long and (2) axial 20.5 mm long artificially produced thermal fa- tigue crack (89,200 and 112,820 thermal fatigue cycles, respec- tively) (pipe diameter 360 mm, wall thickness 28 mm). 28 mm) on the inner surface of which one circumferen- tial and one axial thermal fatigue crack were produced. The cracks were produced in the base material with lengths of about 20 mm. In another case, a crack was produced in the inner surface of a 159 mm× 8.5 mm (diameter× wall thickness) pipe. The crack was 82% of the wall thickness being about 7 mm deep (Fig. 3). Additionally, cracks have been produced in the heat affected zones of butt-welded pipes. 2.6.2. T-joint of pipes Artificially produced thermal fatigue cracks were produced in the corner of a T-joint of pipes of different sizes (110 mm× 20 mm and 570 mm× 35 mm). Nei- ther surface nor any other pre treatment was performed on this area. Two individual cracks were produced in the desired locations with location accuracy of Fig. 19. Artificial thermal fatigue crack at the bottom of a thread: (a) cross-section and (b) crack opening on the surface of the thread bottom. Fig. 18. Production of artificial thermal fatigue cracks inside of austenitic stainless steel collector head threaded screw hole. ±1◦ and length sizing accuracy of±0.3 mm (surface length). 2.6.3. Collector head The method was applied to a piece of a VVER steam generator collector head that had previously been in use in a nuclear power plant. The piece included three threaded holes (diameter 48 mm) with a bottom cup in the end of the holes. Artificial cracks were produced in the threaded area and the bottom cup area. Cracks produced in the threaded area were used as initiating flaws for subsequently growing transgranular stress corrosion cracks. A general picture of the collector head mock-up is shown inFig. 18. The arm of the induction heater is shown in the left side hole of the figure. An example of a horizontal crack produced at the bottom of a thread is shown inFig. 19. The crack has first initiated by thermal fatigue mechanism and then grown further by stress corrosion. During the production of the horizontal crack shown, a vertical crack developed. M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 115 3. Discussion Production of realistic thermal fatigue cracks is based on controlled, cyclic thermal loading. During the manufacturing process, varying heating and cool- ing periods are applied in order to control the thermal stresses induced. Single thermal fatigue cracks first initiate and then grow. There are a wide variety of thermal fatigue cracks, which can be produced by this method. These are single cracks with, for ex- ample, different fracture surface roughness, different tortuous crack paths, many or few branches, differ- ent growth directions (e.g. when compared to rolling direction of a plate), different aspect ratios (depth to length ratios), different states of residual stresses, etc. Furthermore, it is possible to grow combinations of single cracks, as parallel cracks, intersecting cracks or several cracks forming a network. The crack path of the artificially produced thermal fatigue crack is tortuous. Tortuous path is a result of the natural crack growth mechanism. The produced thermal fatigue crack has rough fracture surfaces ex- hibiting clear fatigue striation formation and typically crack has a semielliptical shape. Thermal loading exceeding the yield strength of the material causes residual stresses. The applied cyclic thermal loading and the number of cycles determine the magnitude of the residual stress. Thus, the residual stress near the crack, along the crack length and path, can be controlled by controlling the applied thermal loading. It is commonly known, that the ultrasonic signal obtained from a crack is affected by the stresses present in the vicinity of the produced crack (Iida et al., 1988; Yoneyama et al., 2000). According toWirdelius and Österberg (2000), for example, 200 MPa increase in the closing pressure of a flaw causes 10 dB drop in ultrasonic signal amplitude. Changes in the residual stresses along the crack path (at the crack tip, in the middle of the crack and near the crack mouth) affect the ultrasonic energy penetration in different locations. These changes increase the uncertainty of the signal analysis and affect flaw detection and the difficulty of sizing flaws accurately. The artificially produced cracks realistically simu- late the real, service-induced cracks from the metal- lurgical and ultrasonic point of view. Metallurgically the produced cracks are narrow, have a rough fracture surface, show minor branching, propagate transgranu- larly and have a tight crack tip with a small radius. Ar- tificially produced thermal fatigue cracks do not have an oxide layer on their fracture surface, as is the case with the service-induced cracks. However, an oxide layer can be grown by a subsequent heat treatment. Artificially produced thermal fatigue cracks gave similar ultrasonic responses compared to real service-induced cracks (corner echo, crack tip, frac- ture surface). Studied cracks in austenitic stainless steel base material pipe were reliably detected and sized. No additional cracks were detected near the produced cracks. Crack in the austenitic stainless steel cladding was detected easily, but no crack tip echo was detected which makes the sizing of the crack difficult. This may be caused by compressive stresses affecting the crack tip. The possible compres- sive stress at the crack tip may also affect the use of eddy current-based crack sizing. Further, the inhomo- geneous welded cladding caused some uncertainties in eddy current inspection. The ultrasonic signal amplitudes from a service- induced and artificially produced thermal fatigue cracks are not as high as those obtained from an EDM-notch. A reason is that the ultrasonic reflectiv- ity is affected by the rougher reflection surface and the narrow opening (width) of the cracks, while the EDM-notch has a smoother reflection surface and is relatively wide. The tip of the EDM-notch gave a clear and sharp signal, while the signals from crack tips were formed by multiple signals produced by the irregularities of the crack tips. Furthermore, the possible stresses present at the crack tips may affect Fig. 20. Comparison of (a) a service-induced thermal fatigue crack (Hänninen and Hakala, 1981), (b) an artificially produced thermal fatigue crack and (c) an EDM-notch (modified PISC type A). 116 M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 Fig. 21. Comparison of defects produced with different techniques: (a) a weld solidification crack (W tson and Edwards, 1996), (b) a surface breaking crack in a weld (Uddcomb, 1999) and (c) an artificially produced surface breaking thermal fatigue crack. the ultrasonic signals, while the tip of an EDM-notch is stress free. In this study, the crack tip of an arti- ficially produced thermal fatigue crack in austenitic stainless steel base material was detected by several probes and the crack tip echo had similar S/N ratio as obtained from real service-induced thermal fatigue cracks.Fig. 20 illustrates the difference between the service-induced crack, artificially produced crack and modified PISC type A EDM-notch. Controlled thermal fatigue cracks are produced without causing any additional alterations in the sur- rounding material. As an exampleFig. 21 compares defects produced by two other qualification defect production techniques and the new artificial thermal fatigue crack production method. Fig. 21 shows that the other techniques introduce additional weld metal volume, while controlled pro- duction of thermal fatigue cracks results precisely in the desired crack without any alterations in the sur- roundings. During the production of artificial thermal fatigue cracks only a small area of interest is loaded. Consequently the size of the specimen is irrelevant to the production method and cracks can be produced in very large components. The artificial crack production method has been applied to different nuclear power plant components: pipes, T-joints of pipes and collector head samples. Materials have been austenitic stainless steels com- monly used in the nuclear power plants. So far, practically no limitations have been met for the ap- plicability of the method. The only requirement for the applicability is that the location where the crack is to be produced must be accessible for the induction heating coil. 4. Conclusions A new artificial flaw production method for NDT-qualification defect production purposes has been developed. The method is based on controlled thermal fatigue loading. The flaw production can be controlled so that the location, orientation and size of the cracks are accurately adjusted. Single cracks are grown without any additional alterations caused in the material. Cracks can be produced in base material or welded areas, in simple plate samples or full scale mock-ups. It is shown that artificially produced cracks correspond well with the service-induced cracks both non-destructively and destructively. Artificial, representative cracks can be produced in samples and mock-ups of different sizes and shapes. The experiments performed proved that artificial sur- face breaking thermal fatigue cracks can be induced in any component with practically no limitations in lo- cation, orientation or size. With this method realistic, controlled cracks can be produced in new mock-ups, in existing mock-ups containing other defects or the method can be used to grow further existing defects of different types. The method is also applicable to change the opening and the residual stress state of an already existing real crack. M. Kemppainen et al. / Nuclear Engineering and Design 224 (2003) 105–117 117 Acknowledgements The main part of the work was carried out at Helsinki University of Technology as a post-graduate study financed by Foundation of Walter Ahlström, Foundation of Runar Bäckström and Technology De- velopment Foundation (TES). The rest of the work has been carried out in Technical Research Centre of Finland (VTT) and Fortum Nuclear Services Ltd. References ASME, 1990. Metal Fatigue in Operating Nuclear Power Plants— A Review of Design and Monitoring Requirements, Field Failure Experience, and Recommendations for ASME Section XI Action. Prepared by ASME Section XI Task Group on Fatigue in Operating Plants. Revision 0 (Draft), January 1990. Becker, F.L., Doctor, S.R., Heasler, P.G., Morris, C.J., Pitman, S.G., Selby, G.P., Simonen, F.A., 1981. Integration of NDE Reliability and Fracture Mechanics—Phase I Report, NUREG/CR-1696 PNL-3469, vol. 1. 170 pp. Hytönen, Y., 1998. Two leakages induced by thermal stratification at the Loviisa power plant. In: Proceedings of NEA/CSNI/R(98)8 Specialists’ Meeting, Paris, France, 8–10 June, pp. 115–160. Hänninen, H., Hakala, J., 1981. Pipe failure caused by thermal loading in BWR water conditions. Int. J. Pressure Vessel Piping 9, 445–455. Iida, K., Takumi, K., Naruse, A., 1988. Influence of stress condition on flaw detectability and sizing accuracy by ultrasonic inspection. In: Proceedings of the 9th International Conference on Nondestructive Evaluation in the Nuclear Industry, Tokyo, Japan, 25–26 April, pp. 563–567. Ogilvy, J.A., 1989. Model for the ultrasonic inspection of rough defects. Ultrasonics 27, 69–79. Pirson, J., Roussel, G., 1998. Emergency core cooling system pipe crack incident at the Tihange unit 1 plant. In: Proceedings of NEA/CSNI/R(98)8 Specialists’ Meeting, Paris, France, 8–10 June, pp. 103–114. Saka, M., Fukuda, Y., 1991. NDT of closed cracks by ultrasonic propagation along the crack surface. NDT&E Int. 24 (4), 191– 194. Uddcomb Engineering AB, 1999. Testblock med Implanterade Defekter. Brochure from Uddcomb Engineering. Virkkunen, I., 2001. Thermal Fatigue of Austenitic and Duplex Stainless Steels, Acta Polytechnica Scandinavica. Mechanical Engineering Series No. 154, Espoo, 115 pp. Available online at: (http://lib.hut.fi/Diss/2001/isbn9512256878/). Virkkunen, I., Kemppainen, M., Hänninen, H., 2000. Residual stresses induced by cyclic thermal loads. In: Proceedings of the Sixth International Conference on Residual Stresses ICRS-6, Oxford, UK, 10–12 July, pp. 529–536. Wåle, J., Ekström, P., 1995. Crack Characterisation for In-service Inspection Planning, SKI Projekt 14.4-940389, 94164 SAQ/FoU-Rapport 95/70, SAQ Kontroll AB, Stockholm, Sweden, 84 pp. Watson, P., Edwards, R.L., 1996. Fabrication of test specimens simulating IGSCC for demonstration and inspection technology evaluation. In: Proceedings of the 14th International Conference on NDE in the Nuclear and Pressure Vessel Industries, Stockholm, Sweden, 24–26 September, pp. 165–168. Wirdelius, H., Österberg, E., 2000. Study of Defect Characteristics Essential for NDT Testing Methods UT, EC and RT. SKI Project Number 98267, SKI Report 00/42, Sweden, 50 pp. Yoneyama, H., Senoo, M., Miharada, H., Uesugi, N., 2000. Comparison of echo heights between fatigue crack and EDM-notch. In: Proceedings of the 2nd International Conference on NDE in Relation to Structural Integrity for Nuclear and Pressurized Components, New Orleans, Louisiana, USA, 24–26 May, 8 pp.

(12) United States Patent Elfving et al. USOO6723185B1 (10) Patent No.: US 6,723,185 B1 (45) Date of Patent: Apr. 20, 2004 (54) METHOD FOR PRODUCING DEFECTS AND TENSILE RESIDUAL STRESSES (75) Inventors: Kai Elfving, Kirjala (FI); Hannu Hanninen, TKK (FI); Mika Kemppainen, Kirkommuni (FI); Pekka Saarinen, Helsinki (FI); Iikka Virkkunen, Helsinki (FI) (73) Assignee: Trueflaw Oy (FI) (*) Notice: Subject to any disclaimer, the term of this patent is extended or adjusted under 35 U.S.C. 154(b) by 0 days. (21) Appl. No.: 09/856,008 (22) PCT Filed: Nov. 16, 1999 (86) PCT No.: PCT/FI99/00949 S371 (c)(1), (2), (4) Date: Jul. 19, 2001 (87) PCT Pub. No.: WO00/29841 PCT Pub. Date: May 25, 2000 (30) Foreign Application Priority Data Nov. 16, 1998 (FI) ................................................. 982471 (51) Int. Cl. ............................ C21D 11/00; C21D 9/00 (52) U.S. Cl. ....................... 148/510; 148/508; 148/511; 148/639; 148/644 (58) Field of Search ................................. 148/508, 511, 148/510, 639, 644 (56) References Cited U.S. PATENT DOCUMENTS 4,108,647 A 8/1978 Shaw ......................... 42O/448 5/1980 Vieu et al. 10/1980 Matsuda et al. 11/1987 Tarnai 3/1988 Podlech 3/1989 Shinozaki et al. 5/1991 Diaz 4.203,315 A 4,229.235 A 4,704892. A 4,729,235 A 4,810,400 A 5,013,370 A FOREIGN PATENT DOCUMENTS JP 410170421. A * 6/1998 WO WO 98/19 155 5/1998 - - - - - - - - - - - - GO1 N/3/60 * cited by examiner Primary Examiner Deborah Yee (74) Attorney, Agent, or Firm-Skinner and Associates (57) ABSTRACT A method, which is used to make controlled defects corre sponding to natural flaws and residual Stresses in various kinds of test pieces. Defects identical to natural flaws are required to qualify non-destructive testing (NDT) proce dures. In the method, Sequential, repeated, heating-cooling cycles are used to create defects and residual Stresses. The shape of the heating and cooling pattern, the duration of the heating and cooling, and the number of thermal cycles are used to control the size of the defects and residual Stresses obtained. The defect is grown without initial flaw or other nucleator. The defects correspond to natural flaws in terms of morphology and also of the signals obtained with NDT methods, and are suitable for use in, for example, NDT. qualification blockS. 6 Claims, 1 Drawing Sheet US 6,723,185 B1 Apr. 20, 2004 U.S. Patent US 6,723,185 B1 1 METHOD FOR PRODUCING DEFECTS AND TENSILE RESIDUAL STRESSES The present invention relates to a method for producing controlled defects and residual Stresses in test pieces. Defects Such as thermal fatigue cracks may appear in Various components, e.g., in nuclear power plants, during operation. Non-destructive testing (NDT) procedures are used to examine pre defined valves, pipes, pipe connections, etc., during inspections carried out at regular intervals. In inspections carried out under field conditions, inaccuracies due to both the method and the perSonnel carrying them out always arise. Test pieces incorporating artificially created cracks similar to real ones are used to qualify NDT proce dures and inspectors. The international PISC I, II, and III studies of the reliability of the non-destructive testing of materials, which have revealed obvious deficiencies in the detection and definition of defects, have demonstrated the need for qualification. For example, qualification is required in the inspection of nuclear power plants. Instructions have been issued for the qualification of the NDT inspection of nuclear power plants (in the USA, ASME Code Section XI, in Europe, NRWG and ENIQ), according to which equipment, test pieces, procedures, and inspectors should be qualified. According to the European qualification procedures, the dimensions (diameter, wall-thickness, etc.) of test pieces should correspond to the real objects being inspected. Similarly, the test pieces other properties, Such as material, shape, Surface quality, method of manufacture, and location of welds must also correspond to those in actual plant components. The type, Shape, size, location, orientation, and opening of the defects that occur must correspond closely enough to natural flaws. The characteristics of the defects in qualification test pieces are highly significant to the entire qualification process. The use of test pieces with defects as Similar as possible to natural defects will ensure that the inspection procedure in question can detect and define Such defects with the required precision. Personnel-qualification tests determine if the inspector can detect and define the relevant defects with Sufficient accuracy. At present, methods are known for producing various kinds of cracks and defects in test pieces. In the known methods, an artificial crack is implanted to pieces usually by welding, a welding fault is made, or a notch is machined in the test piece. Japanese patent publication JP 57-034439 discloses a method that can be used to produce a crack in the Surfacing of a plate-like piece. Another, also Japanese, patent publication JP 58-055752 discloses a method for making an artificial crack, in which a hole is machined in the Surface of a joint between two pieces, after which the pieces are joined together. A third Japanese patent publication, JP 8-219953, discloses a method for manufacturing defects by machining grooves in a piece and filling them with a material with different acoustic properties to those of the parent material. No method is known for producing a defect in a piece of any shape at all, at any place at all, and of any desired orientation and shape. Producing cracks Similar to natural ones is one of the central problems in qualification. A method for producing a State of tensile residual StreSS is also known from U.S. Pat. No. 5,013,370, in which a tensile residual StreSS is induced in a test piece by cooling one Surface of the piece and locally heating the opposite surface. The invention disclosed in the above publication requires the object or test piece being dealt with to have at least one StreSS-free portion, where a State of tensile residual StreSS can be induced. Heating is directed to this area and, if 15 25 35 40 45 50 55 60 65 2 the tensile residual StreSS is to be produced in more than one area, the initial State of each of them must be StreSS-free. In the publication referred to, opposite Sides of the piece are heated and cooled. This arrangement induces an uncon trolled State of tensile StreSS in the area being treated. According to the patent in, question, a crack can be created in the induced field of tensile Stress, e.g., by means of an aggressive crack-promoting environment, Such as a boiling magnesium chloride Solution, oxygenated water at a high temperature, etc. Unlike the invention according to U.S. Pat. No. 5,013, 370, the present invention can be used to create a wanted and controlled crack or a controlled residual stress state (either tensile or compressive), with no environmental or stress State demands. AS, in the present invention, the heating/ cooling is not directed to different Sides of the piece, there are no requirements concerning the size and shape of the piece. In addition, there are no requirements concerning the residual StreSS State in the test piece in its initial State. The above and other advantages and benefits of the invention are achieved by means of a method, the charac teristic features of which are described in the accompanying Claims. The basic idea of the invention is to repeatedly alter nately heat and cool the piece being treated, i.e. to fatigue it thermally, which will result in a crack identical to a natural flaws, or in a desired controlled State of residual StreSS. In general, it can be Stated that cracks induced by means of the methods disclosed in the above publications do not correspond to natural flaws, which is an obvious deficiency when manufacturing Such test pieces for qualification. At present, cracks are produced in qualification test pieces either by Welding Separate pieces containing cracks into a test piece, or by welding a hot crack into a test piece, or Simply by machining a notch in a test piece. Cracks implanted by welding may be natural and taken from actual pieces that have cracks in operation, or be artificially pro duced in Separate test pieces. No matter how the crack being implanted has originated, the Weld from the implanting will remain in the material. A hot crack is made by machining a narrow groove in the test piece and then by welding it shut using parameters that will cause the Weld to crack in the desired direction. The welded joints in test pieces resulting from these methods can be easily detected by NDT inspec tion methods. Such aspects of machined notches as width and progression do not correspond to those of natural flaws. The method now developed can be used to flexibly manufacture cracks Similar to natural ones in any location in the test piece, irrespective of the shape or dimensions. The cracks are nucleated directly in the Surface of the test piece. No crack initiator (machined notch etc.) is required. The cracks are grown in the Surface of the test piece without the material experiencing micro-structural or other changes detectable by NDT methods. This is a significant advantage, because when transplanted or welded cracks are used, the inspector may notice the welded Seams in the test piece and be alerted to make a more thorough Search for cracks in the Same area. Cracks created by the method that has been developed also correspond well to natural cracks, in terms Such as the propagation and branching of the crack and the radius of the crack tip, all of which affect, for instance, the Signals received in ultrasonic inspection and their interpre tation. The method now developed is based on the phenomenon of thermal fatigue and a new application of it. The phenom enon of thermal fatigue as Such has been known for a long time, particularly for materials used at high temperatures. US 6,723,185 B1 3 The rapid cycling of heating and cooling in thermal fatigue causes Steep temperature gradients to fluctuate in the test piece, resulting in StreSS and Strain cycling that depend on the coefficient of thermal expansion of the material, and finally leading to fatigue damage. Cracks manufactured using the method now developed are suitable for use in test pieces used to qualify NDT procedures. The cracks can be manufactured according to the following requirements: the morphology of the crack corresponds Sufficiently to a natural crack for the response to it, when inspected by an NDT procedure, to be similar to that to a genuine crack the method can be used to manufacture individual cracks and networks of cracks the orientations of individual cracks can be varied cracks can be made in different sizes and shapes, unaf fected by the thickness of the material of the piece if it is not intended to destruct the test pieces after the qualification test, the size of the cracks to be manufac tured can be evaluated either during the fatigue cycling or on the basis of the fatigue parameters the heating and cooling patterns can be altered to make the cracks grow in desired directions. BRIEF DESCRIPTION OF DRAWINGS In the following, the operation of the invention is described with reference to the accompanying drawing, which show the heating and cooling arrangements. In the FIGURE, forced cooling (either liquid or gas cooling) is marked with the number 1, heating with the number 2, and the heating pattern (isotherms) with the number 3. The arrangement according to the figure operates So that heater 2 is used to heat the Surface to the desired temperature, after which the Surface is cooled 1 to a lower temperature. During heating, a heating pattern 3, which is dependent on the heating output and time, arises, the shape of the pattern influencing the orientation of the defects created. The operation of the invention is described with the aid of the following examples. EXAMPLE 1. A Single crack, or Several individual parallel cracks, was to be created in a pipe that corresponded to those used in a nuclear power plant. The heating pattern was shaped with its longer dimension circumferential to the pipe. This caused the cracks to grow in the pipe's axial direction. The heating and cooling cycles both lasted for 30s. During the test, the maximum temperature was approximately 700° C. and the minimum temperature approximately 10 C. Forced cooling was used to prevent the heating pattern from Spreading by thermal conduction. In forced cooling, cooling water is continuously directed to both sides of the heated area, also during the heating cycle. The total number of cycles was 6500, both heating and cooling being included in a Single cycle. At this number of cycles, three axially-oriented cracks grew from micro-cracks nucleated in the test piece. 15 25 35 40 45 50 55 60 4 The method according to the invention was used to form a single crack or multiple cracks in the parent material. EXAMPLE 2. The method was used to manufacture a crack or multiple cracks in a welded Seam in a pipe. The cracks were manu factured in the inner Surface of the pipe. The heating pattern was shaped in Such a way that it extended over the root pass of the Weld, with its centre point close to the root pass. During calibration, a temperature of 700 C. was achieved when a heating time of 6 s was used. The cooling time was 10 s. A total of 17803 cycles were made during the test. Penetrant testing was performed, demonstrating that cracks had formed on both Sides of the weld. Besides the aforementioned cracks, a Single crack, not extending over the root pass, had grown transversely to the weld. The method can be used to manufacture cracks with properties corresponding to those of natural flaws, and which thus can be used in qualification test pieces. The method can be used to grow both a network of cracks Specific to thermal fatigue and Single cracks. Ultrasonic inspections of the manufactured cracks showed that the cracks challenge to inspection procedures corresponded to reality. The method has been used in the manner described in the above examples. In addition, the method can also be applied to produce residual Stresses in a test piece. Residual Stresses are also produced by thermally cycling on a desired area in a test piece, when permanent residual StreSS State will arise in the test piece. The changes of the residual StreSS-State during the cycling depend on the heating and cooling parameters used. It should be noted that the invention is in no way restricted to the above disclosure or examples, but can be varied within the Scope of the following claims and the Stated inventive idea. What is claimed is: 1. A method for producing artificial defects and/or residual Stresses in test pieces, characterized by the Steps of alternately heating and cooling the test object, on the same Side of the test object, to create the defects and/or residual StreSSeS. 2. A method according to claim 1, characterized by forming a shape of heating and/or cooling patterns whereby the defects and/or residual Stresses are created. 3. A method according to claim 2, characterized in that the shape of the heating pattern is controlled by cooling the test piece outside the area of the desired pattern. 4. A method according to claim 1, characterized by Sequentially repeating, a Sufficient number of times, alter nate heating and cooling cycles to achieve thermal fatigue damage. 5. A method according to claim 1, characterized by growing the defect without an initial flaw or other nucleator. 6. A method according to claim 1, characterized by controlling, by means of the heating and cooling output, the duration of the heating and cooling, and the number of thermal cycles, whereby the Size of the manufactured defects and residual Stresses is controlled. k k k k k

Thermal Fatigue of Austenitic and Duplex Stainless Steels IIKKA VIRKKUNEN Helsinki University of Technology Department of Mechanical Engineering Puumiehenkuja 3 A P.O.Box 4200 FIN-02015 HUT Finland Dissertation for the degree of Doctor of Science in Technology to be presented with due permission for public examination and debate in Council room at Helsinki University of Technology (Espoo, Finland) on the 23rd of November, 2001, at 12 o'clock noon. Espoo 2001 2 Virkkunen, Iikka. Thermal fatigue of austenitic and duplex stainless steels. Acta Polytechnica Scandinavica, Mechanical Engineering Series No. 154, Espoo 2001, 115 pp. Published by the Finnish Academies of Technology Keywords: thermal fatigue, austenitic stainless steel, duplex stainless steel, residual stress, Barkhausen noise, fatigue crack growth ABSTRACT Thermal fatigue behavior of AISI 304L, AISI 316, AISI 321, and AISI 347 austenitic stainless steels as well as 3RE60 and ACX-100 duplex stainless steels was studied. Test samples were subjected to cyclic thermal transients in the temperature range 20 – 600°C. The resulting thermal strains were analyzed with measurements and numerical calculations. The evolution of thermal fatigue damage was monitored with periodic residual stress measurements and replica-assisted microscopy. The elastic strains of the ferrite phase in duplex stainless steels were studied using Barkhausen noise. Finally, destructive analyses including fractographic scanning electron microscopy (SEM) studies and transmission electron microscopy (TEM) analyses were performed. The surface residual stresses changed markedly during the first load cycles. In the austenitic stainless steels yielding during the rapid cooling resulted in compressive residual stresses from -200 MPa (20 – 300°C temperature cycle) to -600 MPa (20 – 600°C temperature cycle). After 10 cycles the residual stresses stabilized and then started to relax due to crack formation. Cracks were seen to initiate from persistent slip bands (PSBs) and in 3RE60 from MnS inclusions. In duplex stainless steels the phase boundaries retarded crack growth markedly. In the austenitic stainless steels, the fracture surfaces of thermal fatigue cracks showed extensive striation formation, i.e. they were similar to mechanical fatigue. The dislocation density was lower than expected based on mechanical fatigue data. Dislocation tangles and occasional cell tendency was observed. In duplex stainless steels the plastic deformation concentrated to the austenite phase. The obtained thermal fatigue data were compared with mechanical fatigue data from literature and with the ASME design curve. The ASME design curve was found to give safe design life, although the remaining safety factor on strain is decreased to 1.5. The total strain (elastic+plastic) caused by thermal loading was solved with linear-elastic FE-analysis. Thermal fatigue crack growth was predicted successively using total strain solution of an uncracked component and a strain- based growth model:       da dN C atot m= 7 7∆ε , where C7=1.6 and m7=1.3 for the studied austenitic stainless steels. The model is applicable to small fatigue cracks (0.05 - 4 mm) growing in varying temperature and strain fields and is temperature-independent in the studied range. © All rights reserved. No part of the publication may be reproduced, stored in a retrieval system, or transmitted, in any form of by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior written permission of the author. 3 PREFACE The research work presented in this thesis was mainly accomplished as a part of the research project "Thermal Fatigue Mechanisms" financed by TEKES, Fortum Nuclear Services Ltd., Fortum Oil & Gas Oy, Metso Paper Oy, Outokumpu Poricopper Oy, and Teollisuuden Voima Oy. The steering committee of the project also gave valuable comments and suggestions during my work. The support is gratefully acknowledged. The financial support of Imatran Voiman Säätiö and KAUTE (Laura & Aarne Karjalainen fund) is also acknowledged. I am deeply grateful to professor Hannu Hänninen for the advice and unlimited expertise he provided for my guidance. Numerous co-workers and friends also offered their help and support. I wish to express my gratitude to Mika Kemppainen for the fruitful collaboration, for Sergiy Smuk, Jaakko Virkkunen, and Ville Virkkunen for their encouragement and perceptive comments on my thesis, to Wade Karlsen for advice on the grammar of the thesis, to Laura Taivalaho for TEM work and help with the grammar of the thesis, to Pertti Nenonen for the exquisite TEM work, and for the workers of the Laboratory of Engineering Materials for creating a fruitful and always enjoyable working atmosphere. I also wish to extend my gratitude to the pre-examiners of this work, Prof. Gary Marquis and Dr. Arttu Laitinen for their very valuable and constructive criticism. Finally, I wish to express my deepest gratitude to my wife Päivi and to my newborn daughter, who made my life outside my work exceedingly rewarding. Otaniemi, June 2001 Iikka Virkkunen 4 CONTENTS Abstract 2 Preface 3 Contents 4 Original features 6 Nomenclature 7 1 Introduction – current understanding of thermal fatigue 10 1.1 Thermal stresses ............................................................................................11 1.2 Residual stresses............................................................................................13 1.3 Fatigue..........................................................................................................14 1.4 Dislocation structures produced by cyclic loading...........................................15 1.5 Fatigue crack initiation ..................................................................................16 1.6 Fatigue crack growth .....................................................................................16 1.7 Fracture mechanics .......................................................................................20 1.8 Crack growth models for fatigue design..........................................................22 1.9 Fatigue properties of austenitic and duplex stainless steels ..............................29 1.10Multiaxial fatigue ..........................................................................................32 1.11Environmental effects ....................................................................................32 1.12Dynamic strain ageing...................................................................................36 1.13Thermal fatigue .............................................................................................36 2 Aims of the current work 43 3 Experimental procedure 44 3.1 Materials.......................................................................................................44 3.2 Thermal fatigue tests......................................................................................46 3.3 Thermal load determination ..........................................................................49 3.4 Residual stress profile measurements by contour method................................49 3.5 Cyclic stress-strain tests..................................................................................50 3.6 Numerical simulations...................................................................................52 4 Results 54 4.1 Cyclic stress-strain curves...............................................................................54 4.2 Barkhausen noise measurements....................................................................55 4.3 Cyclic strain caused by thermal loading..........................................................57 4.4 X-ray measurements ......................................................................................61 4.5 Contour method............................................................................................69 4.6 Build-up of surface damage............................................................................69 4.7 Fractography.................................................................................................72 4.8 Dislocation structures ....................................................................................77 5 Discussion 81 5.1 Cyclic thermal loads and residual stresses.......................................................81 5.2 Micromechanism of crack growth ..................................................................82 5.3 Thermal fatigue prediction.............................................................................84 5 5.4 Comparison of thermal and mechanical fatigue ..............................................88 6 Conclusions 91 7 References 92 Appendix 1. Two-bar model for thermal stresses 109 Appendix 2. Derivation of equation 41 111 Appendix 3. Used FE-models 112 Appendix 4. Stress solutions 114 6 ORIGINAL FEATURES The experimental data and analyses of this thesis are directed towards modeling and predicting thermal fatigue damage and especially thermal fatigue crack growth in stainless steels. The following features are believed to be original: 1. Detailed study of the strains, residual stresses, dislocation structures and fatigue crack growth caused by cyclic thermal loading was performed. 2. An original method was developed to measure the elastic strains of the ferrite phase in a two-phase ferritic-austenitic duplex stainless steel using Barkhausen noise. 3. Linear-elastic FEM solution was found to give good approximation of the total (elastic + plastic) strains in strain-controlled, thermal load case. 4. The acquired thermal fatigue data was analyzed according to the ASME (1995) pressure vessel code. The code was found to give safe design life; the remaining safety factor on strain is 1.5. 5. It was shown that thermal fatigue crack growth in the crack depth range of 0.1 – 5 mm can be predicted using a total strain based crack growth model:       da dN C atot m= 7 7∆ε , whereC7=1.6 and m7=1.3 for the studied austenitic stainless steels. 6. It was shown that the model used is independent of test temperature in the studied range. 7 . Verified range of the applicability for the crack growth model was extended from mechanical fatigue to thermal fatigue and for cracks growing in varying strain and temperature fields. 7 NOMENCLATURE A Area (m2) a Crack depth a0 Initial crack depth aeff Effective crack length for the El Haddad model b Burgers vector C1 Parameter for the Paris law C2 Parameter for the Paris equation used with Kε C3 Parameter for the Tomkins model C4 Parameter for the Solomon model C5 Parameter for the Ahmad-Yates model C6 Parameter for the Ahmad-Yates model C7 Parameter for the Taira model C8 Parameter for the Nisitani model D Deformation zone size (µm) Plastic zone size (µm) d Grain size of the material (µm) E Young's modulus (GPa) Fen Fatigue life correction factor for environmentally assisted fatigue G Energy release rate Iβ Material constant for the Skelton model J J integral Je Elastic part of J Jp Plastic part of J K Stress intensity factor (MPa m0.5) KI Stress intensity factor for mode I loading (MPa m0.5) KISCC Threshold stress intensity factor for stress corrosion cracking (MPa m0.5) Kε Strain intensity factor (m0.5) M Material constant for the Paris law m1 Parameter for the Paris law m2 Parameter for the Paris equation used with Kε m4 Parameter for the Solomon model m5 Parameter for the Ahmad-Yates model m6 Parameter for the Ahmad-Yates model m7 Parameter for equation 40 m8 Parameter for the Taira model N Cycle number O' Parameter describing the effect of dissolved oxygen on corrosion fatigue q Parameter for the Haigh-Skelton model R Load ratio (εmin/εmax) r Radius (mm) rI Inner radius (mm) rO Outer radius (mm) 8 S Striation spacing (µm) s1 Strain hardening exponent s2 Work hardening exponent T Temperature (°C) T' Temperature dependent parameter for estimating Fen Te Equivalent temperature (°C) Tmax Maximum temperature during a cycle (°C) Tmin Minimum temperature during a cycle (°C) Tt Traction vector for the J integral t Parameter for the Gabetta model u Displacement (mm) Vα Volume fraction of α-phase (%) Vγ Volume fraction of γ-phase (%) W Strain energy density wc Hysteresis energy per cycle We Elastic strain energy density Wf Energy required for failure Wp Plastic strain energy density Y Geometry factor for K and Kε ∆Kth Threshold stress intensity factor range below which crack growth stops ∆σl Fatigue limit Γ Arbitrary path surrounding the crack tip α Linear coefficient of thermal expansion (K-1) Ferrite phase in duplex stainless steels β Material constant for the Skelton model ε Strain (mm/mm) ε̇ Strain rate (s-1) ε̇' Parameter describing the effect of strain rate on corrosion fatigue εa Axial strain (mm/mm) εe Elastic strain (mm/mm) εeb Elastic biaxial strain (mm/mm) εeq Equivalent strain for the Haigh-Skelton model (mm/mm) εeu Elastic uniaxial strain (mm/mm) εm Measured apparent strain (mm/mm) εmax Maximum strain, e.g., during a cycle (mm/mm) εmin Minimum strain, e.g., during a cycle (mm/mm) εp Plastic strain (mm/mm) εpb Plastic biaxial strain (mm/mm) εpu Plastic uniaxial strain (mm/mm) εr Radial strain (mm/mm) εt Tangential strain (mm/mm) εtot Total (elastic+plastic) strain (mm/mm) εΤ Thermal expansion (mm/mm) εσ Constrained thermal expansion (mm/mm) γ Austenite phase in duplex stainless steels ν Poisson's ratio νeff Effective Poisson's ratio 9 σ Stress (MPa) σµ Microscopic (average) stress (MPa) σa Axial stress (MPa) σf Flow stress (MPa) σM Macroscopic stress (MPa) σmax Maximum stress during a cycle (MPa) σphase Total average phase stress (MPa) σt Tangential stress (MPa) σy Yield stress (MPa) σyc Cyclic yield stress (MPa) σµα Microscopic (average) stress in phase α (MPa) σµγ Microscopic (average) stress in phase γ (MPa) σr Radial stress (MPa) 10 1 INTRODUCTION – CURRENT UNDERSTANDING OF THERMAL FATIGUE Components subjected to alternating heating and cooling have been found to crack and eventually fail. This phenomenon is termed thermal fatigue. The reason for cracking is believed to be twofold. Firstly, temperature change in the material induces thermal expansion (or contraction). If surrounding material or external constraints hinder this expansion, thermal stresses arise. These cyclic thermal stresses cause fatigue similar to that of mechanical stresses. Typical components where this kind of problems arise are, e.g., power and process industry pipings where turbulent mixing of fluids causes quick thermal transients, or rotating components in the paper industry where frictional heating causes rapid temperature cycles (e.g., Dean, 2000; Hänninen & Hakala, 1981). Secondly, at high temperatures, materials are found to exhibit time dependent deformation, i.e. creep. Failure is generated by a process combining creep and fatigue (e.g., Taira & Fujino, 1979; Zamrik, 1990; Goswami, 1999; Gandossi, 2000). Creep becomes more significant with increasing temperature and prolonged exposure. Typical components, where creep-fatigue interaction is significant include gas turbine blades and soldered joints of electric circuitry. The motivation for this study of thermal fatigue is to enhance the reliability of affected components. The economical effects of unexpected failures can be very high, e.g., in power, process and paper industries. Failures may threaten property and even life. The existing design methods do not give adequate tools for design against thermal fatigue. In the case of thermal fatigue, the increase of safety by using heavier constructions is not only uneconomical but also often impossible. In some cases a heavy design increases the constraint to thermal expansion and thus has a negative effect on component safety. This study is focused on the thermal fatigue in power, process and paper industries, where time-dependent effects can be presumed to be negligible (Zauter et al., 1994). The theoretical background on which the current understanding of the phenomenon is based is that of mechanical fatigue (or just fatigue) and thermal stresses. 11 1.1 Thermal stresses Change in temperature induces thermal expansion in a material: ε αT T= ∆ , (1) whereεT is the thermal expansion (mm/mm), α is the linear coefficient of thermal expansion (K-1) and ∆T is the temperature change causing the expansion (K). If this expansion is hindered, thermal stresses arise. The magnitude of the arising stresses is such that if applied as an external load they would result in strain equal to the hindered expansion. Thermal stresses can arise because of either external or internal constraints. Internal constraints can be caused by nonuniform temperature distribution or nonuniform material properties. If the thermal stresses arise due to internal constraints, the loading is called pure thermal loading, or just thermal loading. If the stresses are due to external constraints, term thermomechanical loading is used. The elementary theory of thermal stresses is quite well established, and several textbooks have been written on the subject (e.g., Boley & Weiner, 1960). Analytical solutions to elastic and elastic-perfectly plastic thermal stress problems in various geometries have been presented in literature (e.g., Fritz & Schenectady, 1954). For example, the thermal stresses in an elastic cylinder or tube with radial temperature distribution are given by (Boley & Weiner, 1960): σ α νr I O I r r r r r E r r r r r Trdr Trdr II O( ) = −( ) − − −        ∫∫ 1 2 2 2 2 2 (2) σ α νt I O I r r r r r E r r r r r Trdr Trdr Tr II o( ) = −( ) + − + −        ∫∫ 1 2 2 2 2 2 2 , (3) whereσr is the radial stress (MPa), σt is the tangential stress (MPa), α is the linear coefficient of thermal expansion (K-1), E is Young's modulus (GPa), ν is Poisson's ratio, T(r) is temperature (K), r is the radius (mm), rI is the inner radius of the tube (mm) and rO is the outer radius of the tube (mm). If the axial deformation of the cylinder is unconstrained, the axial stress is given by: 12 σ α ν σ σa O I r r r tr E r r Trdr T I O( ) = −( ) − −         = +∫ 1 2 2 2 , (4) where σa is the axial stress (MPa) and the other symbols are as in equations 2 and 3. In addition to full solutions, simplified parameters for the severity of thermal transients have been proposed. Yoshimoto et al. (1999) showed that the maximum thermal stress, which develops during a quench test, can be estimated as follows: σ αmax . .= ( ) −0 705 43 80E T∆ , (5) where σmax is the maximum thermal stress during a quench test, E is Young's modulus (GPa), ∆T is the temperature difference (K) and α is the linear coefficient of thermal expansion (K-1). In a two-phase material, such as duplex stainless steel, where the two phases have different thermal expansion behavior, change in temperature gives rise to strain mismatch and thermal stresses between the two phases. The resulting microscopic thermal stresses, i.e. microstresses, can be very complex. The microstresses have to be superimposed on the macroscopic stresses to give the total phase-specific stresses. Stevens (1999) studied the effect of thermal treatment simulating welding on the corrosion fatigue life of duplex stainless steels, and found that the applied temperature cycle reduced the corrosion fatigue life of the samples. He attributed this effect to both secondary phase precipitation and residual stresses. Behnken & Hauk (1992a; 1992b) studied the effect of microstresses on fatigue behavior, and concluded rather surprisingly that, since the residual microstresses are typically much smaller than the macroscopic residual stresses, their effect can be neglected. Other sources (Pohl & Bracke, 1997; 1999) however, concluded that thermal cycling up to 900°C causes microscopic thermal strains that result in stresses exceeding the yield strength in every cycle. Because in duplex stainless steels both constituent phases are capable of plastic deformation, pure thermal cycling can accumulate macroscopic strain to the specimens (Pohl & Bracke, 1999; Fischer et al., 1990; Siegmund et al., 1993; 1995). As temperature rises, stresses in the austenite phase are shifted towards compression, and stresses in the ferrite phase towards tension, due to the greater thermal expansion of the austenite phase. During cooling the stresses are reversed. This has been found by both X-ray residual stress measurements (Kamachi et al., 1982) and numerical simulations (Siegmund et al., 1993). The microscopic thermal stresses are highly dependent on the local microstructure, i.e. local phase morphology. In order to model the phenomenon, a set of quantitative parameters are needed to describe the investigated microstructure (Siegmund et al., 1995; Silberschmidt & Werner, 1999). 13 1.2 Residual stresses In addition to external loads, thermal or mechanical residual stresses often exist in a component. These stresses result from uneven yielding during, e.g. manufacturing or thermal treatments. Residual stresses are difficult to model or measure, and they often form an unknown factor in fatigue. The residual stresses tend to fade during fatigue loading (Almer et al., 2000). Stabilized residual stresses correlate with fatigue crack initiation and fatigue life. In a two-phase material, such as duplex stainless steels, microscopic residual stresses in the constituent phases can develop. These microstresses can result from uneven yielding of the phases during plastic deformation (e.g., Inal et al., 1999) or from uneven thermal expansion of the two phases. Irrespective of their origin, the residual stresses have strong impact on load sharing between the phases and on mechanical behavior, especially on the fatigue properties of the material. The residual stresses in a two-phase material affect the apparent thermal expansion behavior (Wang et al., 1999). The uneven yielding also causes a surface relief to develop (Polák et al., 1999). The uneven yielding during plastic deformation gives rise to residual microstresses. The total stress in a phase is then the sum of the micro- and macrostresses: σ σ σµphase M= + , (6) whereσphase is the total (average) stress in a phase, σM is the macroscopic stress and σµ is the microscopic (average) stress in a phase. In a randomly oriented microstructure the microscopic stresses of the constituent phases are related through the equilibrium condition: V Vα µα γ µγσ σ+ =0 , (7) whereVα and Vγ are the volume fractions of phases α and γ and σµα and σµγ are the (average) microstresses in phases α and γ, respectively. Johanson & Odén (1999; 2000a; 2000b) studied the load sharing between the austenite and ferrite phase in nitrogen containing duplex stainless steels. They concluded that the hardness and the yield strength of the austenite phase are higher than those of the ferrite phase. Despite its greater strength, more plastic deformation occurs in austenite than in ferrite because of the compressive residual stress present in the ferrite phase after cooling (Johanson & Oden, 1999; 2000a; 2000b). A higher load, therefore, is transferred through the ferritic matrix. The ratio of plastic strain in austenite to that in the ferrite increases with decreasing 14 nitrogen content and increasing total strain amplitude (Degallaix et al., 1995). Plastic deformation changes the residual microstresses due to uneven yielding in the constituent phases. Moverare (2001) studied the evolution of residual stresses in uniaxial tension. The experiment was repeated in several loading directions on duplex stainless steels. He concluded that the load partitioning between the two phases is dependent on the loading direction: for loading in the rolling direction, both phases deform plastically to the same degree, while more plastic deformation occurs in the austenitic phase during loading in the transverse direction. The difference in load partitioning was attributed to crystallographic texture in the ferritic phase. 1.3 Fatigue Fatigue is defined as the degradation or failure of a material caused by cyclic loads. The study of fatigue started with the so-called phenomenological models, which ignore the mechanisms by which the fatigue failure takes place, but aim to establish a tool with which the fatigue endurance of a component can be predicted. The Wöhler and Coffin & Manson plots have been the basis for fatigue design for many decades and are still used in many design codes such as the ASME pressure vessel code (1995). During the years, several improvements have been proposed to the plots (e.g., Muralidharan & Manson, 1988; Hatanaka, 1990). At first the material properties were assumed to remain constant during the fatigue life. Especially in low cycle fatigue, this is often not the case. Cyclic loading hardens or softens the material. The change in material properties starts from the first cycles and continues to final fracture. Consequently, the monotonic material properties do not describe the material adequately, and fatigue life predictions can be improved by considering the cyclic properties (Mughrabi, 1993, 1996; Mughrabi et al., 1997). Several corrections for the Coffin-Manson model have been proposed which take into account the cyclic properties (e.g., Kliman, 1983; Mughrabi, 1996; Shi & Pluvinage, 1994; Hatanaka, 1990). Mughrabi (1996) studied the cyclic stress-strain behavior of f.c.c and b.c.c materials. He compared the cyclic properties of these materials with several life prediction models and concluded that the microstructural interpretations of fatigue behavior are still far from being comprehensive and consistent. The cyclic stress-strain curve is affected by metallurgical factors, such as mobility of screw dislocations, strain-induced martensite and dynamic strain ageing (Hatanaka, 1990). For austenitic stainless steels several constitutive models for predicting cyclic properties have been proposed (Ohno & Kachi, 1986; McDowell, 1987). 15 The main drawback of the phenomenological approach is that because it lacks a direct relation to the cumulating fatigue damage, it provides inadequate tools to assess the remaining safe life of a cracked component. Also, the effect of various defects from manufacturing etc. is difficult to take into account. In addition, the cyclic stresses or strains in the component may not be uniform throughout the component, which leads to retardation or acceleration of crack growth. This cannot be adequately modeled with the phenomenological approach. Despite their inadequacies, the phenomenological models still remain in many cases the only practical design method for components with small or no initial defects. Chopra & Shack (1999) note that, although the S-N curves are conventionally used to give the number of cycles to produce gross failure for given stress or strain amplitude, they would be better interpreted as specifying the number of cycles required for the formation of an "engineering" crack for example 3 mm deep. 1.4 Dislocation structures produced by cyclic loading Numerous studies on the dislocation structures generated by mechanical fatigue in austenitic and duplex stainless steels can be found in literature. Gerland et al. (1989; 1997) fatigued AISI 316L isothermally at different temperatures ranging from 20 to 600°C. In the beginning of fatigue at a small plastic strain amplitude (εp=2x10-3), dislocation tangles and irregular cells were found. After failure, or with greater strain amplitudes, high dislocation densities and tightly arranged dislocation structures, such as walls-and-channels, cells or corduroy, were observed. Kruml et al. (1997) found ladder-like dislocation structures in AISI 316L fatigue-tested at room temperature. This structure was explained to be responsible for the strain localization. Li & Laird (1994) also found cells and walls-and- channels dislocation structures after room temperature fatigue in an AISI 316L single crystal. Orbtlík et al. (1994) (also for AISI 316L) explained that the different dislocation structures found in austenitic stainless steels are due to their intermediate stacking fault energy (SFE). Hence the dislocation structure depends on the applied strain range. For small strain amplitudes (<10-4), the planar dislocation structures typical for low SFE materials prevail, i.e. corduroy, dislocation alignments and planar slip bands. For high strain amplitudes (>10-3) walls-and-channels and cells typical for high SFE materials prevail. Xia et al. (1992) studied a manganese-alloyed austenitic stainless steels and found similar dislocation arrangements. Similar dislocation arrangements are also observed in the austenite phase of duplex stainless steels. In the ferrite phase of duplex stainless steels, veins-and-channels and ladder structures are observed (Degallaix et al., 1995). In the ferrite phase, the observed dislocation structures do not change markedly with the strain amplitude. 16 1.5 Fatigue crack initiation Traditionally the failure is divided into three phases: crack initiation, crack growth or propagation, and final fracture. During the initiation period, a visible crack or cracks form. The initiated cracks then grow and at some point the growing cracks have weakened the structure so much that the load bearing capacity is exceeded and final fracture occurs. At very low load amplitudes, near the fatigue threshold, most of the life is spent on the crack initiation phase. As the load amplitudes increase, the proportion of the life spent in the propagation phase increases rapidly. The distinction between the initiation and the propagation phases is somewhat ambiguous. With the development of better methods of detecting cracks, smaller cracks have become detectable and the initiation period has shortened. Also considerable scatter now exists in the definition of initiation: it can mean cracks visible to the naked eye (in the range of 1 – 2 mm) or it can mean smallest cracks visible in the electron microscope (in the range of a few µm). It has been suggested that initiation is absent altogether and that first undetectable cracks initiate during the first load cycle (Miller, 1999). Crack initiation typically takes place at material inhomogenities. These can be for example persistent slip bands caused by cyclic loading, grain boundaries or nonmetallic inclusions. The fatigue strength depends on the "weakest link", e.g., from the largest inclusion present (Murakami & Endo, 1993; Murakami et al., 1994). 1.6 Fatigue crack growth After crack initiation, the fatigue damage is concentrated to the highly strained area of the crack tip. In small crack lengths the crack growth rate is highly dependent on the local microstructure and coalescence of cracks (Miller 1999). After the crack length is more than about 3 times the grain size, the crack growth can be described by a single mechanical parameter (Kawagoishi et al. 2000). Ductile metals deform plastically at the highly stressed area around the crack tip. In the area where stresses exceed the yield strength of the material, the monotonic plastic zone is formed. The plastic zone alters the stress and strain distribution at the crack tip. In the cyclic case, unloading reverses the stress direction. The changes in the stresses and strains are given by a solution identical to that of monotonic loading but with yield strain and stress replaced by twice their values (Rice, 1967). Consequently, a much smaller cyclic plastic zone forms inside the monotonic plastic zone (Figure 1). The formation of these plastic zones was also experimentally verified (e.g., Davidson & Lankford, 1976; Williams et al., 1980). Lankford et al. (1984) studied the crack tip deformation of small fatigue cracks and found that the measured crack tip opening (CTOD) and crack tip plastic zones were much larger than those predicted based on long 17 crack models. They concluded that the large plastic zone compared to the crack depth at least partly explained the anomalous crack growth rate of small cracks. Figure 1. Schematic illustration of monotonic and cyclic plastic zones and stress distributions (after Rice, 1967). The crack propagation phase is subdivided based on the plane of the crack advance. The crack can grow by two processes and corresponding planes: shear decohesion on one plane directly ahead of the crack tip, or shear decohesion of two mutually perpendicular planes inclined at about 45° to the crack plane (Miller, 1999). The former is termed stage I and the latter stage II crack growth. Stage I crack growth is mainly observed at small applied load ranges, whereas stage II is typical for high loads. Cheng & Laird (1983) suggested that the transition occurs due to work hardening on the primary slip system. The work hardening causes strain to be transferred partly to the secondary slip system. This causes persistent slip band (PSB) formation on the secondary slip system and transition from stage I to stage II crack growth. The orientation of the crystal relative to loading affects the stresses on the secondary slip system and consequently the transition behavior. The higher the stresses on the secondary slip system are, the earlier the transition to stage II crack growth occurs. Lukáš  & Kunz (1984) and Lukáš  et al. (1985) studied the dislocation structures ahead of a fatigue crack tip in copper (Figure 2). They found that the high strains directly ahead of the crack tip are extended by the formation of slip bands. The crack was found to follow the path of the highest local deformation. At low amplitude loading, the zone of the cell structure ahead of the crack tip is small and the crack follows PSBs. With higher load amplitudes, the crack tip is surrounded by an extensive zone of dislocation cells which does not contain preferred paths, and the crack propagates non-crystallographically. It was also found that the crack growth rate was proportional to the local crack tip plastic strain (Lukáš  et al., 1985). 18 Figure 2. Dislocation structures ahead of fatigue crack tip in copper (after Lukáš Aet al., 1985). The portions of fatigue life spent in initiation, stage I and stage II of crack growth depend on the environment, material and loading (Socie & Bannantine, 1988). The fracture surface left by a stage I crack appears faceted. The facets form as the crack changes the plane of advance. By contrast, stage II cracks leave ripple-like markings, striations, which describe the crack advance by cycle. Several models have been proposed to explain the formation of striations. The blunting model, proposed by Laird & Smith (1962), is presented in Figure 3. Tomkins (1968) presented an alternative model (Figure 4). His model was criticized by Laird & de la Veaux (1977). The model was further developed in later publications (Tomkins, 1980; 1981; 1983a). The main difference between these two models is that slip is condensed to either one or several sets of PSBs ahead the crack tip. Neumann (1969), Pelloux (1969) and Vehoff & Neumann (1978) developed an alternative model for striation formation (Figure 5). In this model, the crack forms a V-shaped tip and advances by alternating slip on two slip bands. This model was supported by recent fracture surface studies and was developed further by Bichler & Pippan (1999). It should be noted, however that great variation in the appearance of fatigue striations is perceived in different materials, environments and loading conditions. Hence, it may be that a single model for their formation is not conceivable. 19 Figure 3. The blunting model of striation formation (after Laird & Smith, 1962). Figure 4. The Tomkins model of striation formation (after Tomkins, 1968). 20 Figure 5. The Neumann-Pelloux-Bichler model of striation formation (after Bichler & Pippan, 1999). Striation spacing is generally considered to be an indicator of the crack advance per cycle. However, in-situ observations have revealed that, in the near-threshold regime where crack growth rate is low, this is not necessarily true. As the crack growth rate diminishes, more than one cycle is needed to form each striation. Finally, very close to the threshold the striation spacing reaches a minimum of 0.1 – 0.4 µm and is independent of the crack growth rate (Davidson & Lankford, 1992). Recently, Uchida et al. (1999) showed that the stress ratio (R) could also be determined from the fracture surface by studying the height of the striations. At greater crack growth rates the striations give a direct indication of the crack growth rate. Since fatigue crack growth does not occur continuously, but the advancing crack (at least locally) slows or stops at microstructural barriers (Tanaka, 1989), the growth rate calculated from striation spacing is a conservative estimate. Tanaka et al. (1981) proposed a correlation between the striation spacing and ∆K-level (see sections 1.7 and 1.8): S b K E b J Eb = ∆    = ∆ 10 10 103 2 , (8) whereS is the striation spacing, b is the atomic distance or Burgers vector, ∆K is the stress intensity factor range (see section 1.7) and ∆J is the J integral range (see section 1.7). 1.7 Fracture mechanics In 1920 Griffith (1920) published his classic paper on the phenomena of rupture and flow of solids. His paper concludes that the strength of brittle glass samples was determined by the size of flaws present. He proposed that the amount of surface energy required by the growing crack 21 determined the stress needed to break the sample. He then concluded that the strength of a cracked plate was inversely proportional to the square root of the length of the crack present. Later Irwin (1947) generalized this idea to include energy dissipated by plastic flow and named the defined parameter energy release rate, G (Irwin, 1957). With the help of a stress solution of a cracked plate given by Westergaard (1939), it was then shown that the crack tip stresses are proportional to a quantity called the stress intensity factor, given for plane stress by: K EG=    π 1 2 , (9) where K is the stress intensity factor (MPa m0.5) and G is the energy release rate (J mm-2). In a force-controlled load case the solution of G usually gives the a dependence noticed by Griffith (1920). In a displacement-controlled load case, the dependence of K on the crack length a can be quite different or even nonexistent, as for example in the case of displacement-controlled loading of a very long crack in an infinite plate. The linear elastic fracture mechanics (LEFM) parameter K has been a very successful tool in describing many failure mechanisms that feature crack growth. However, with high stresses or small crack lengths the small scale yielding condition of K is violated, and K loses its validity. Rice & Rosengren (1968), Rice (1968) and Hutchinson (1968) proposed that the crack tip could be characterized by a line integral (J-integral or J) around the crack tip even in the case of plasticity exceeding the small scale yielding condition of K: J Wdy T du dx dst= − ⋅    ∫ Γ , (10) where W is the strain energy density, Tt is the traction vector and Γ is a path surrounding the crack tip. The elastic-plastic fracture mechanics (EPFM) parameter J is directly related to K in the case of small scale yielding (Rice & Rosengren, 1968): J K E I= −( )1 2 2ν . (11) J is extensively used for crack tip characterization in inelastic load cases. In the case of thermal gradients, the conventional definition of J loses its path independence. Wilson & Yu (1979) showed that J can be calculated also in thermal stress crack problems, but the solution involves a computationally difficult area integral: 22 J E T x x T dAii ii A = − ( ) − ( )      ∫α ν ∂ ∂ ε ∂ ∂ ε 1 2 1 21 10 , (12) where T is temperature. Following the analysis of Wilson and Yu (1979) the J equation for an arbitrary eigenstrain distribution can be solved and used to solve J in a residual stress field, which again involves an area integral. 1.8 Crack growth models for fatigue design Paris et al. (1961) and Paris & Erdogan (1963) showed that the fatigue crack growth rate could be characterized with the LEFM-based parameter, ∆K: da dN K M = ∆( )4 , (13) whereM is a material constant and ∆K is the stress intensity range. The empirical Paris-Erdogan equation is now by far the most used crack growth model (see Tomkins (1984) for a review of classical crack growth based fatigue analysis). Usually the exponent is also considered as a material constant or a fitting parameter and the equation is presented in the following, more general form: da dN C K m= ∆1 1 , (14) where C1 and m1 are material constants. With very high loading the small scale yielding condition of LEFM is no longer valid and ductile tearing fracture and other fracture mechanisms take part in the process. Hence the applicability of Paris-Erdogan law is limited at high ∆K-values. With low ∆K-values the crack growth rate is observed to deviate from the Paris law (equation 14) and to decrease rapidly. This point is termed the threshold stress intensity range, ∆Kth. Consequently, the fatigue crack growth is divided into three regimes. A schematic crack growth plot is shown in Figure 6. 23 Figure 6. Schematic crack growth plot with Paris parameters (after Tanaka, 1989). At small crack lengths, the applied stress to produce a certain K-value is very high. With high stresses the plastic zone of the crack is large compared to the length of the crack and the small scale yielding condition of K is violated. Consequently, the observed crack growth rates are higher than predicted by the Paris law. Because the fatigue life spent during the early crack growth is very long, the inability of the Paris law to correctly predict the small crack growth is a serious handicap. The limits of the linear elastic fracture mechanics were extended by the introduction of elastic-plastic fracture mechanics (see Paragraph 1.7). Similarly the limitations of the Paris law that are caused by the limitations of linear elastic fracture mechanics could be extended by the use of EPFM parameters such as J. Dowling & Begley (1976) and Dowling (1976; 1977) were among the first to test the applicability of the elastic-plastic J parameter to fatigue crack growth characterization. They concluded that for ductile materials and long cracks J characterized the crack growth for high load state and ligament plasticity. Hoshide et al. (1985) tried to correlate small crack growth with ∆J and found that with crack lengths greater than three times the grain size of the material correlation could be found. For smaller crack lengths, the coalescence of cracks caused a higher than predicted crack growth rate. Observed fatigue crack growth depends on numerous other factors in addition to ∆K or ∆ J. For example, the load ratio R, environment, microstructure, temperature, overloads, and short cracks change the crack growth rate. Elber (1970) proposed that these effects could be explained by premature crack closure, which reduces the effective ∆K. The concept has 24 been used widely since then. Dowling & Begley (1976) and Dowling (1976) corrected their ∆J values for crack closure, which appeared at nonzero (compressive) load in their analysis. Chen et al. (1997) concluded that both ∆J and closure-corrected ∆Jeff were successful in predicting high temperature crack growth in AISI 304 steel. Sadananda & Vasudevan (1995) showed that crack closure is often overestimated and is not sufficient to explain the observed anomalies in crack growth. They proposed instead that fatigue crack growth is dependent upon two parameters: Kmax and ∆K (or the equivalent Jmax and ∆J) (Vasudevan et al., 1992; 1993; Vasudevan & Sadananda, 1993; 1995a; 1995b; Sadananda & Vasudevan, 1995; 1997a; 1997b). The main disadvantage of the K- and J-based models is the problem with predicting small crack growth and crack growth from notches (Pippan et al., 1986). In addition, the calculations required for the assessment of loading at the crack tip are often very difficult. Many analytical solutions for K are available (e.g., Murakami, 1987). For J the solutions are not that readily available. Dowling (1977) presented the following approximate J solution for small half-circular surface cracks: J J J W a W a E se p e p p= + = + = + + 3 2 5 0 3 2 2 5 0 1 2 1 . . . . σ σε , (15) where Je is the elastic part of J, Jp is the plastic part of J, We and Wp are the elastic and plastic strain density and s1 is the strain hardening exponent and a is the crack depth. Nowadays J-based design is usually based on J calculated from finite element (FE) analysis. The calculations are rather demanding due to the elastic-plastic modeling required. If the temperature is not uniform throughout the specimen, additional difficulties arise due to the area integrals present. Several extensions and modifications have been proposed to the above mentioned crack growth laws (e.g., review by Hussain, 1997). Many of them are developed to extend the applicability of the traditional LEFM and J approaches to small crack growth and crack growth from notches. El Haddad et al. (1980) proposed that an effective crack length aeff should be added to the crack length a when calculating ∆K of ∆J to account for the fatigue limit. The effective crack length aeff can be obtained if the threshold stress intensity range and fatigue limit of the material are known: a K eff th e = ∆ ∆      σ π 2 1 , (16) where∆σe is the fatigue limit. 25 Lantaigne & Baïlon (1981) suggested an improved model for near- threshold crack growth. In their model fatigue damage was modeled to cumulate in a certain damage zone ahead of the crack tip. Lal (1994) proposed that the applicability of the Paris approach could be extended to small and near-unstable crack growth by dividing the crack growth into three regimes. The crack growth in each regime was described by a Paris type equation (equation 14) with different material constants C2 and m2. Crack growth from notches has also been successfully predicted by replacing the stress intensity factor K with a strain intensity factor Kε (Leis, 1985; Hatanaka et al., 1989): K K E aYε ε π= = , (17) where Y is a geometry factor. When the Paris law (equation 14) is used together with the strain intensity factor, the corresponding material parameters C and m have the following relationships: m m1 2= (18) C C E m1 2 2 = , (19) where the subscripts 1 and 2 denote parameters to be used with stress and strain intensity factors, respectively. Liu & Kobayashi (1980) proposed the following relation based on the unzipping model for crack growth: da dN K E J yc yc = −( ) ∆( ) =0 02 1 0 022 2 . .ν σ σ ∆ , (20) where σyc is the cyclic yield stress. Tomkins (1968) proposed a crack growth model based on the micromechanism of fatigue crack growth (see section 1.4). The basic assumption in the model is that the crack growth rate can be calculated from the size of the crack tip deformation zone: ∆ =a Dpε , (21) where∆a is the crack advance, εp is the applied plastic strain and D is the deformation zone size. To calculate D, Tomkins (1968) used the following solution: 26 D a ≈    sec π 2 . (22) This gives the following cyclic crack growth law (Tomkins 1968): da dN C ap= 3∆ε , (23) whereC3 is a material constant. In later publications, the equation was modified to (Tomkins 1980; 1981; 1983a): da dN C ap t f = 3 2 2∆ε σ σ , (24) whereσt is the stress amplitude and σf is the material flow stress. Solomon (1972) proposed the following, more generalized crack growth law for high strain fatigue: da dN C ap m= 4 4∆ε , (25) whereC4 and m4 are material constants and ∆εp is the bulk plastic strain range. Equation 25 cannot be used for high cycle fatigue, where εp<<εe. To bridge the gap between high strain fatigue and the Paris law, Haigh & Skelton (1978) and Skelton (1982) proposed an equivalent strain range: ε ε εeq p eq= + , (26) whereεeq is the equivalent strain, εp is the plastic strain, εe is the elastic strain and q is a parameter describing the part of the cycle that the   crack stays open; a value of 0.5 was chosen based on crack   closure measurements by electric conduction across the crack   faces. With q=0.5 the εeq is equal to arithmetic mean of   total and plastic strain. High strain fatigue and linear elastic crack growth can then be unified by defining the strain intensity factor in terms of equivalent strain. This parameter is roughly equal to ∆J (Starkey & Skelton, 1982): ∆ ∆K q ap eε ε ε π= +( )( ) 1 2 . (27) 27 Ahmad & Yates (1994) also proposed a fatigue crack growth model based on the Tomkins model. In their model, the Tomkins term was combined with a term for long crack growth to obtain a unified model for cracks growing from notches: da dN C K T C am p m= − + +( ) 5 6 1 25 6∆ ∆ε , (28) whereT is a threshold value for crack growth and C5, C6, m5 and m6 are material constants. More recently Skelton et al. (1998) proposed a model for correlating tested low cycle fatigue (LCF) properties to LEFM fatigue crack growth via an energy-based parameter. In this model the crack growth rate is calculated from the energy required to fail the specimen and the hysteresis energy per cycle, i.e.: da dN w W J I W fc f f = = +     ( )1 1 β β β ∆ , (29) wherewc is the hysteresis energy per cycle, Wf is the energy required for failure, β is a material constant, Iβ is a constant depending on the value of β and f(β) is 1 for plane stress. Taira et al. (1976) studied isothermal high temperature low cycle fatigue of 0.16% carbon steel and AISI 304 austenitic stainless steel. They proposed the following crack growth law: da dN C am= 7 7∆ε , (30) where C7 and m7 are fitting parameters depending on the material and test setup. Equation 30 was found to hold for both plastic and total strain range. The exponent m7 was seen to be almost constant over a wide temperature range (Figure 7). 28 Figure 7. The temperature dependence of exponent α in equation 30 (after Taira et al., 1976). Nisitani et al. (1992) proposed the following small crack growth model for stress ranges below the yield stress: da dN C am= 8 8∆σ , (31) where C8 and m8 are material parameters. According to Nisitani et al. (1992) the apparent contradiction between equation 31 and the Paris law (equation 14) can be explained by their different range of applicability. With long cracks and small stresses the crack tip cyclic plastic zone size is proportional to ∆Km1 and the Paris law describes crack growth. With small cracks and large stresses the small scale yielding condition of LEFM is violated. In this regime, the crack tip plastic zone is given by equation 32 and crack growth rate can be predicted with equation 31. D a y m ∝       σ σ 8 , (32) where D is the plastic zone size. Nisitani et al. (1992) noted that equation 31 is not suitable for comparing different materials due to differences in the values of the parameters C8 and m8. To overcome this, he proposed the following equation in which the material properties are partly included: da dN C a y m =      9 9 ∆σ σ , (33) where C9 and m9 are again material parameters. 29 Kawagoishi et al. (2000) recently studied the range of applicability of this equation with 0.42% carbon steel. It was shown that the lower limit of applicability of equation 31 is around a crack length of 3 times the grain size of the material (d). Below this value, the strong influence of the microstructure prohibits the use of a mechanical parameter. Microstructural influence continues to be significant until about 8d, but below this value the mean crack growth can be estimated with a mechanical parameter. When cracks grow long enough to satisfy the small scale yielding condition of LEFM, the crack growth becomes ∆K- controlled, and equation 31 is no longer valid. With higher stress amplitudes, equation 31 is applicable to greater crack lengths. Kawagoishi et al. (2000) described these different regimes of crack growth with the following map (Figure 8). Figure 8. The applicability of the Paris law (equation 14) and equation 31 as a function of crack length and stress amplitude (after Kawagoishi et al., 2000). It should be noted that equations 30 and 31 are linearly dependent in the linear-elastic region studied by Nisitani et al. (1992) and Kawagoishi et al. (2000). The equations are equal if: m m7 8= and (34) C C E m7 8 8 = , (35) where E is the Young's modulus. 1.9 Fatigue properties of austenitic and duplex stainless steels The fatigue properties of different stainless steels have been studied widely at different temperatures and environments. Both S-N plots and crack 30 growth data are widely available in various environments. In Figure 9, the S-N plots for AISI 304 and AISI 316NG stainless steels in air and water are presented. In Figure 10 the crack growth rates of various austenitic stainless steels in different environments are presented. From the presented crack growth data, the material constants C1 and m1 of the Paris law (equation 14) can be extracted. The extracted values are presented in Table 1. Figure 11 shows the crack growth rate for a duplex stainless steel in air.    Figure 9. S-N plots for AISI 304 and AISI 316NG austenitic stainless steels in air and water (after Chopra & Gavenda, 1998). Table 1. The material constants C1 and m1 of the Paris law from Figures 10 and 11. Material C m AISI 304, 316, 321, 348 at 25°C 4.6x10-9 3.1 AISI 304 at 593°C 4.7x10-8 2.8 AISI 316 at 593°C 7.3x10-7 2.1 AISI 321 at 593°C 1.2x10-7 2.6 AISI 348 at 593°c 1.2x10-8 3.1 20 Cr, 5.5Ni Duplex at RT 4.6x10-11 3.1 31 Figure 10. Crack growth data for austenitic stainless steels (after Boyer, 1986). Figure 11. Crack growth data for A790-grade duplex stainless steels fatigued in air at RT (after Rajanna et al., 1997). 32 1.10 Multiaxial fatigue Multiaxial loading shortens expected fatigue life as compared to uniaxial loading. The fatigue life in biaxial loading can be reduced by more than a factor of 2 as compared to uniaxial loading (Itoh et al., 1995; 2000). Numerous models have been proposed to predict the effect of various proportional or nonproportional load paths (e.g., Brown & Miller, 1982; You & Lee, 1996; Socie & Marquis, 2000). The proposed multiaxial fatigue models can be broadly classified into the following approaches (Socie & Marquis, 2000): stress-based models, strain-based models, energy- based models, critical plane models and fracture mechanics models. Stress and strain based models aim to extend the use of the known phenomenological Wöhler or Coffin-Manson models by providing an equivalent damage parameter to be used for multiaxial loading. In energy- based models, the fatigue life is related to the plastic strain energy dissipated during loading. In critical plane models, it is noted that fatigue crack growth can occur predominantly along planes of maximum shear or tensile strain and no single damage model can be expected to be successful for both modes (e.g., Socie, 1987; Socie & Bannantine, 1988). The fracture mechanics models aim to extend crack growth models, such as the Paris law (Equation 14), for mixed mode loading. Marsh (1981) used the following formulas to convert equi-biaxial thermal fatigue loading to equivalent uniaxial loading: ε εpb pu= 1 2 , (36) ε ε νeb eu= −( )1 , (37) where subscripts e and p refer to elastic and plastic, and b and u refer to biaxial and uniaxial loading, respectively. These simple formulas assume, that biaxial loading can be correlated with uniaxial loading by subtracting strain caused by the Poisson effect of the transverse loading (νεeu for elastic and νeffεpu for the plastic case). 1.11 Environmental effects Fatigue crack growth tends to accelerate in aggressive environments. Even fatigue tests conducted in air show significantly higher crack growth rates than tests conducted in vacuum (Wadsworth & Hutchings, 1958). In more aggressive environments, time dependent stress corrosion cracking (SCC) accelerates the crack growth rate. According to McEvily & Wei (1971) and Gabetta et al. (1990) the environmental effects on fatigue crack growth can be divided into true corrosion fatigue (TCF) and stress corrosion fatigue (SCF) (Figure 12). 33 Figure 12. Separation of corrosion fatigue to true corrosion fatigue and stress corrosion fatigue (after McEvily & Wei, 1971; Gabetta et al., 1990). In true corrosion fatigue, the crack growth rate follows the conventional Paris law (equation 14), and the environment merely has an effect on changing the material constants C1 and m1 to reflect the accelerated crack growth rate. The fracture surfaces show ductile features, such as ductile striations, similar to the fracture surfaces produced, e.g., in air. In stress corrosion fatigue, a time-dependent crack growth component accelerates the crack growth. The fracture surfaces typically show brittle striations or intergranular fracture. The effect of test frequency is decisive. The stress corrosion fatigue crack growth can be correlated to the stress corrosion crack growth under similar environmental and loading conditions. The maximum expected combined crack growth rate can be expressed as a combination of the crack growth rates caused by true corrosion fatigue and stress corrosion (Gabetta et al., 1990): da dN da dN t da dNTCF SCC     =     +     , (38) where the true corrosion fatigue crack growth (subscript TCF) follows the Paris law (equation 14), and stress corrosion crack growth (subscript SCC) is almost constant and can be estimated from slow strain rate test (SSRT) results (Gabetta et al., 1990). The stress corrosion component is present only during the time when the applied ∆K is large enough to exceed the limiting KISCC. This time is a function of the test frequency and is represented by the term t in the equation. 34 Two mechanisms have been suggested to account for the environmental effects on fatigue crack growth: slip oxidation/dissolution (Ford, 1993) and hydrogen-induced cracking (Hänninen et al., 1986). In the former mechanism the crack growth rate is related to the oxidation rate at the strained crack tip and periodicity of oxide rupture at the crack tip. In the latter mechanism crack growth is related to localized hydrogen induced cracking resulting from hydrogen generation due to the corrosion reaction. In stainless steels the fact that fatigue life is longer in water containing low dissolved oxygen than in high dissolved oxygen water suggests that the contribution of slip oxidation/dissolution mechanism is limited (Chopra & Muscara, 2000). Studies on crack initiation in water environments reveal that the observed decrease in fatigue life is caused primarily by the effects of environment on the growth of small cracks less than 500 µm deep (Chopra & Muscara, 2000). The environmental effects depend on strain, dissolved oxygen (DO), strain rate and temperature. The main effects and threshold values of the parameters are summarized in Table 2. Table 2. The effect of various factors on the decrease of the fatigue life of austenitic stainless steels in aqueous environment (Chopra & Muscara, 2000). Factor Effect Dissolved oxygen (DO) The environmental effects are more pronounced in low-DO (<0.01 ppm) than in high-DO water. In high-DO water the environmental effects are moderate when coductivity is low and electrochemical potential (ECP) has reached a stable value. In low-DO water conductivity or additions of lithium and boron have no effect on the fatigue life. Applied strain A minimum threshold strain is required for the environmental decrease of fatigue life. During each cycle, the environmental effects are limited to the part of the cycle where the applied strain is above this threshold. Strain rate In high-DO water with low conductivity and stable ECP, fatigue life is insensitive to strain rate. In low-DO water fatigue life decreases logarithmically with decreasing strain rate below about 0.004 s-1. Temperature High temperature increases the effect of environment. Environmental effects are minimal below 200°C and become significant above 250°C. Fatigue life is insensitive to changes in temperature in the range of 250 – 330°C. Higuchi & Iida (1991) have suggested that the effect of environment for austenitic stainless steels can be expressed in terms of fatigue life 35 correction factor Fen, which is defined as the ratio of life in air at room temperature to that in water at service temperature. For stainless steels, Fen can be estimated as follows (Chopra & Muscara, 2000): F een T O= − ′ ′ ′0 935. ε̇ , (39) whereT'=0 for T<180°C, T'=(T-180)/40 for 180220°C, ε̇'=0 for ε̇>0.004 s-1, ε̇'=ln( ε̇/0.004) for 0.000004< ε̇<0.004 s-1, ε̇'=ln(0.000004/0.004) for ε̇<0.000004 s-1, O'=0.260 for DO<0.05 ppm and O'=0 for DO>0.05ppm. Corrosion fatigue crack growth in austenitic stainless steels in aqueous environments has been studied extensively (e.g., Kawakubo et al., 1980; Endo et al., 1983; Hishida et al., 1986). Kawakubo et al. (1980) measured crack growth rates of AISI 304 type stainless steel in oxygenated water at 290°C in the solution treated and sensitized conditions (Figure 13). Figure 13. Crack growth rates of AISI 304 stainless steel in oxygenated water at 290°C (sinusoidal wave loading, frequency 10Hz, stress ratio 0.1, O2 concentration 8 ppm, electrical conductivity <10-4 S/m) (after Kawakubo et al., 1980). 36 1.12 Dynamic strain ageing In certain ranges of temperatures and strain rates, many solid solutions exhibit serrated yielding behavior (Portevin–Le Chatelier effect). The phenomenon is explained by interaction between moving dislocations and diffusing solute atoms, i.e. dynamic strain ageing (DSA). According to the model proposed by McCormick (1972), the movement of dislocations during plastic deformation is discontinuous. Dislocations are temporarily arrested at obstacles. They break free and advance with high velocity until they are arrested again at the next obstacle. The average dislocation velocity is determined mainly by the average arrest time spent at the obstacles. If the arrest time of the dislocation at an obstacle is sufficiently long, the dislocation may be locked by solute atoms, and stress increase is needed for the dislocations to be released from the locking atmosphere formed by the solute atoms. The ageing time required to form the locking atmosphere depends on the diffusion rate of the solute atoms. During plastic deformation, the ageing time decreases due to vacancy production. Dislocation multiplication caused by straining increases the waiting time. At some critical strain the waiting time becomes equal to the required ageing time, and serrated yielding begins (McCormick, 1972). The strength of the material is increased by dynamic strain ageing. The fracture toughness, ductility and low cycle fatigue resistance are reduced (e.g., Srinivasan et al. 1997). For a given alloy, DSA occurs typically for very limited temperature and strain rate regimes. Dynamic strain ageing is observed in different austenitic stainless steels at temperatures 200–600°C (Srinivasan et al., 1997; Nilsson & Thorvaldsson, 1985; Jenkins & Smith, 1969; Zauter et al., 1993). Due to the limited temperature range of occurence, DSA can cause inverse dependence of strength on temperature, i.e. strength increases with increasing temperature. Dynamic strain ageing is related to interstitial atoms or interstitial atom-vacancy pairs at temperatures below 500°C and to substititional atoms at higher temperatures. It can be associated with pipe diffusion below 600°C and also with bulk diffusion above 800°C (Ilola, 1999). 1.13 Thermal fatigue Cyclic thermal loads cause fatigue damage and crack growth. If the fatigue loading originates from cyclic pure thermal loads, i.e. thermal loads resulting from internal constraints, then loading is termed thermal fatigue loading. If the loads arise due to external constraints or forces, then the term thermomechanical fatigue is used. The loading is said to be in-phase when the highest temperature coincides with the highest stress and out-of- 37 phase when the highest temperature coincides with lowest stress (Figure 14). Figure 14. Schematic illustration of in-phase and out-of-phase loading. Typically cyclic thermal loads arise from rapid heating or cooling of a component surface, e.g., due to turbulent mixing of fluids of different temperatures. It is characteristic to thermal loads that they are highest at the surface and attenuate towards the inside of the component. The loading is strain-controlled. Because the stresses are self-equilibrating, there exists a surface of zero stress in the component at any given time. This causes the characteristic ∆ε distribution, described schematically in Figure 15, to develop. Figure 15. Typical distribution of ∆εσ as a function of depth (x) caused by thermal fatigue loading. At any given time the strain distribution shows a point of zero strain (the dashed lines). The corresponding ∆εσ distribution over the whole cycle shows a minimum at a certain distance (the solid line). 38 The strain range at the surface is often high and exceeds the yield strength of the material. Due to the equilibrating nature of the stresses, residual stresses arise as soon as plastic deformation takes place. Typically, the loading is applied to a surface equally in all directions, and the thermal strains are biaxial. The characteristics of thermal fatigue loading are reflected in the characteristics of thermal fatigue cracking. The high strains at the surface give rise to rapid crack initiation on a multitude of initiation sites. Because of the biaxiality, cracks initiate and grow equally in all directions and soon form a mosaic-like network. The cracks relax the stresses perpendicular to the crack plane in their vicinity. Hence new cracks turn to meet older cracks at 90° angles. A typical thermal fatigue crack pattern is shown in Figure 16. Figure 16. Typical mosaic-like crack pattern formed by thermal fatigue. Due to the rapidly decaying ∆ε seen by a small surface crack, crack growth slows down after the initial rapid growth. Cracks typically slow down at some depth close to the minimum in ∆ε. A smaller number of cracks extend over this minimum and continue to grow with an increasing rate. Thermomechanical fatigue is closely related to thermal fatigue, although the characteristic features of the load pattern are lost in the case of external constraints or loads. More studies are available on thermomechanical fatigue than pure thermal fatigue probably due to easier comparison and integration to mechanical fatigue testing. Thermal fatigue studies started, as did mechanical fatigue studies, with empirical tests aimed at providing a simulative verification of a design problem at hand (e.g., Coffin et al., 1954; Coffin & Schenectady, 1954; Tidball & Shrut, 1954). These kinds of studies have continued and S-N curves for different materials and environments have been measured (Hirano et al., 1994; Hayashi 1994; 1998a; 1998b; Hayashi et al., 1998). 39 Petersen & Rubiolo (1991) compared isothermal, thermal and thermo- mechanical fatigue data from different sources (Figure 17). The biaxial thermal fatigue data fits well with the uniaxial isothermal fatigue data, which lead Peterson & Rubiolo (1991) to conclude that the effect of biaxial stress condition on fatigue life is negligible. Also studies on multi-material components, such as coated composite tubes, have been made (e.g., Keiser et al., 1996; 1997; Revel et al., 2000; Saarinen et al., 2000), and thermal fatigue caused by a traveling temperature distribution has been investigated (Ogawa et al., 1991; Igari et al., 2000). Figure 17. Fatigue data for isothermal, thermal and thermo-mechanical fatigue in air at different temperatures (after Petersen & Rubiolo, 1991) The established design methods for mechanical fatigue were soon applied to thermal fatigue. Wareing et al. (1973) studied the effects of temperature and strain rate on the fatigue behavior of two austenitic stainless steels and concluded that the difference in fatigue behavior was explained by the effect of temperature and strain rate on cyclic behavior. Models have been proposed to predict the cyclic stress-strain curves in the case of thermomechanical loading based on isothermal data (Shi & Pluvinage, 1994; Maier & Christ, 1996; 1997). The stress-strain behavior in thermomechanical fatigue saturates after fewer cycles than in mechanical fatigue (Spindler, 1998). Zauter et al. (1994) were among the first to study dislocation arrangements caused by the thermomechanical fatigue of AISI 304L. They found cell structures at lower temperatures (below 625°C), where creep was not significant. Above 625°C transition into directional sub-grain structure was observed. This transition was attributed to creep. Maier & Christ (1997) found that under thermomechanical fatigue conditions cell structures form. Armas et al. (1992) studied AISI 316 type steel with 40 thermal expansion fully restricted by external constraints. Walls-channels and cell structures were observed similar to those seen in isothermal loading. Studies on dislocation structures caused by pure thermal fatigue are rare to find. Recently Keiser et al. (1996; 1997) observed low dislocation density and dislocation tangles when studying possible thermal fatigue of recovery boiler tubes. Recent studies by Saarinen et al. (2000) on thermal fatigue of recovery boiler tubes confirm these results. Taira (1973) studied the difference between thermal fatigue and high temperature isothermal fatigue. He proposed that a Coffin-Manson curve for thermal fatigue could be determined by simpler isothermal tests if the temperature for the test was chosen correctly. He found that at low temperatures (below 400°C for the 0.16% carbon steel he studied) this equivalent temperature (Te) was equal to the arithmetic mean of the Tmax and Tmin. In this temperature regime the diffusion-controlled effects are negligible and thermal and mechanical fatigue are expected to be similar (Taira, 1973). At higher temperatures, the damage accumulated close to the maximum temperature dominates and Te≈Tmax. Crack growth studies on thermal and thermomechanical fatigue are rare. Notable exceptions are the measurements by Shi et al. (1996) presented in Figure 18, and Marsh (1981), presented in Figure 19. Figure 18. Striation spacing measured from thermomechanical fatigue fracture surface of AISI 316L austenitic stainless steel with in- phase (IP) and out-of-phase (OP) loading of two different temperature regimes with total strain range of 1.6% (after Shi et al., 1996). 41 Figure 19. Crack growth rate of AISI 304 and 316 austenitic stainless steels subjected to thermal fatigue loading of effective ∆T of 208 – 288°C. Loading was applied with cyclic resistance heating and water quenching. (After Marsh, 1981) The design codes developed for mechanical fatigue, such as the ASME Boiler and Pressure Vessel code (1995), are currently used to design against thermal fatigue although experimental verification of their applicability is limited (Merola, 1995; Fissolo et al., 1996; Kerezsi et al., 2000). The K- or J-solutions for cracks growing in thermal or residual stress fields have been studied, e.g., by Hirano et al., (1979); Zhang & Burger (1986); Kuo & John, (1992); John et al. (1992); Jin & Noda (1994); Fischer et al. (1996) and Tanigawa & Komatsubara (1997). The K- solutions are limited to small ∆T due to the small scale yielding condition of LEFM. The crack growth models for mechanical fatigue have been applied to thermal fatigue in several studies. The used models include the Paris law (dell'Erba & Aliabadi, 2000; Marsh et al., 1986; Green et al., 1987; Bressers et al., 1994), strain-based models (Tomkins, 1983b; Green & Munz, 1996) or both (Marsh, 1981). In many cases, the complexity of the thermal fatigue loading imposed serious restrictions or approximations on the analysis. The Haigh-Skelton model (equation 27) was recently applied to thermal fatigue crack growth analysis by Green & Munz (1996). The results were promising, although the thermal stress analysis was limited and the actual crack growth rates could not be measured or compared to calculations. 42 Marsh (1981) studied the thermal fatigue behavior of AISI 304 and 316 steels. He concluded that the thermal fatigue initiation life correlated with the initiation life observed in uniaxial isothermal fatigue if the biaxial thermal fatigue data was converted to an equivalent uniaxial load case with equations 36 and 37 (Figure 20). Figure 20. Comparison of initiation life of thermal and isothermal fatigue (after Marsh, 1981). Marsh (1981) obtained good correlation between the thermal and isothermal fatigue crack growth rates by describing the crack growth in the highly strained near surface area with a strain-based model (equation 40) and the crack growth in the elastic bulk with the conventional Paris law (equation 14): da dN a sf p=       +         π σ σ ε2 2 28 2 1 2 ∆ , (40) whereσ is the maximum tensile stress in the cycle, σf is the flow stress, s2 is the work hardening exponent and ∆εp is the plastic strain range. 43 2 AIMS OF THE CURRENT WORK The tools for fatigue design offered in the present open literature can be divided into three categories: phenomenological models (1), crack growth models based on fracture mechanics (2), and crack growth models not based on fracture mechanics (3). In thermal fatigue, the loading typically causes a varying strain and temperature field in the affected component, as presented in Figure 15. This makes the application of the phenomenological models to the thermal fatigue load case difficult, since a single representative load parameter is not readily available. The application of LEFM-based crack growth models is limited by the small scale yielding condition. This prohibits the use of LEFM-based models for small cracks and high loads. The EPFM parameters are valid for greater amounts of plastic deformation, but are difficult to determine for a crack advancing in a varying strain and temperature field. Consequently, the usefulness of EPFM-based models is limited by the reliability and feasibility of the stress analysis required. For the crack growth models not based on the fracture mechanics, such as the strain-based models, the main problem is that their range of applicability is unknown. Numerous crack growth parameters have been proposed in the open literature, but for each parameter the applicability is demonstrated for a rather limited range of materials and load cases. Applicability of these models to cracks growing in varying strain and temperature fields caused by cyclic thermal loads is unclear. The existing tools for fatigue design are, thus, seen to be inadequate for predicting fatigue damage and crack growth in the thermal fatigue load cases. The aim of the present work is to study the micromechanisms of thermal fatigue crack growth in austenitic and duplex stainless steels, to study the loads and residual stresses caused by cyclic temperature changes, and finally, to find a crack growth model applicable to small thermal fatigue cracks. 44 3 EXPERIMENTAL PROCEDURE Thermal fatigue test samples were subjected to controlled thermal loads. The thermal loads used were selected to be similar to those found in actual industry applications. Applied thermal loads were characterized by thermal load measurements and numerical calculations. The evolution of thermal fatigue damage in the samples was monitored by periodic residual stress measurements and replica-assisted microscopy. To aid the numerical calculations and to compare thermal fatigue and high temperature mechanical fatigue, cyclic stress-strain curve measurements were conducted on some of the materials. Finally, a destructive analysis including fractographic scanning electron microscopy (SEM) studies and transmission electron microscopy (TEM) analyses was performed on the fatigued samples. 3.1 Materials Altogether 9 different materials or thermal treatments were studied. Three test materials common in nuclear power plants were chosen for testing: AISI 304L, 316 and 321. AISI 321 and AISI 347 were studied in solution annealed and stabilized conditions, as used in the oil refinery industry. (Stabilization heat treatment is used in the oil refinery industry to prevent sensitization at the high working temperature. The used materials were held at 920 °C for 2 h and then cooled in air.) Two duplex stainless steels used in pulp & paper industry were chosen: ACX-100 – a cast duplex stainless steel made by Kubota and 3RE60 – a wrought duplex stainless steel manufactured by Avesta. The chemical compositions and the material codes (used hereafter to identify the materials) of test materials are shown in Table 3. Available mechanical properties of the test materials are shown in Table 4. The test materials were characterized by standard metallographic analyses. 45 Table 3. Chemical compositions of the materials studied. Material code Type Cr Ni Mo Mn Si Other A AISI 304L 17.4 10.2 1.6 0.8 B AISI 316 17.9 12.7 2.7 1.33 0.3 C AISI 321 17.8 10.7 0.4 1.6 0.3 0.1 Cu, 0.5 Ti, 0.2 Co D AISI 321 19.6 10.0 0.9 0.7 0.4 Ti E AISI 321 " " " " " " F AISI 347 17.3 9.2 0.4 1.5 0.4 0.4 Cu, 0.5 Cb, 0.1 Co G AISI 347 " " " " " " H 3RE60 18.5 4.9 2.9 1.5 1.5 0.1 N I ACX-100 23.4 5.3 2.2 0.6 0.6 0.2 N Table 4. Room temperature mechanical and physical properties of the materials studied. (Properties reported based on manufacturers information or minimum required standard values for similar grades (Wegst, 1995)). Material code Type RP0.2 MPa Rm MPa E GPa k W/mK α x10-6 K-1 A AISI 304L 180 460 200 15 16,0 Solution annealed B AISI 316 205 510 200 15 16,5 Solution annealed C AISI 321 180 460 205 15 16,0 Solution annealed D AISI 321 205 500 200 15 16,0 Solution annealed E AISI 321 205 500 200 15 16,0 Stabilized F AISI 347 205 510 200 15 16,0 Solution annealed G AISI 347 205 510 200 15 16,0 Stabilized H 3RE60 450 730 200 15 14 42% ferrite I ACX-100 460 660 200 15 13 56% ferrite The microstructures of the studied materials were examined in three different orientations. In Figures 21 – 24 the microstructures of the studied materials are shown. Figure 21. Microstructures of studied A, B and C materials. 46         Figure 22. Microstructures of studied D and E materials. The stabilization treatment has refined the grain size considerably and caused an increase in the ferrite content (material E).         Figure 23. Microstructure of studied F and G materials. The stabilization treatment has left the microstructure virtually unchanged (material G).       Figure 24. Microstructure of studied H and I materials. The H material shows anisotropic phase structure due to rolling. The cast I material has considerably coarser, but isotropic, phase structure. Samples were manufactured so that the axial direction of the sample is perpendicular to face 3. 3.2 Thermal fatigue tests Thermal loads were applied using a prototype thermal fatigue testing machine built in the Laboratory of Engineering Materials (Kemppainen, 1997). Two kinds of test specimens were used (Figure 25). 47 Figure 25. Specimen geometries used. (Dimensions in mm.) Specimens were heated rapidly by 156 kHz high frequency induction and cooled by water spray. The test apparatus (Figure 26) consisted of an induction heating unit, a cooling ring and a pneumatic positioning system that moved the specimen between heating and cooling. Specimens were rotated to achieve uniform heating and cooling. Figure 26. Thermal fatigue test apparatus. 48 Rapid heating and cooling caused a steep, cyclic temperature gradient in the specimen. Uneven temperature distribution in the radial direction constrained the thermal expansion of the sample and gave rise to thermal loads. A more detailed description and verification of the test apparatus is given by Kemppainen (1997). The thermal cycles were chosen to simulate the thermal loads expected in actual industrial components. For the nuclear industry materials (codes A, B and C) the maximum expected temperature in use was about 300°C. A temperature cycle of 20 – 300°C was chosen. For the oil refinery industry materials (codes D, E, F and G) a higher temperature cycle of 20 – 600°C was utilized. For the pulp & paper industry materials (codes H and I) a temperature cycle of 20 – 280°C was chosen. The temperature cycles used, as measured at the sample surface, are presented in Figure 27. (The mentioned temperature ranges are used hereafter to identify the cycles, although some deviation is present in the actual cycles, as can be seen from Figure 27.) Figure 27. The temperature cycles used. Temperature cycles were measured at the sample surface using a thermocouple. Thermal cycling was stopped at predefined intervals, and the residual stresses on the sample surface were measured with X-ray diffraction. Measurements were conducted with a Stresstech X3000 commercially available residual stress measurement unit using Cr-Kα radiation. Residual stresses were calculated using the sin2Ψ -method (Cullity, 1978). The surface of the sample was replicated with cellulose acetate. With successive replicas taken from the sample surface, it was possible to monitor the initiation and growth of individual thermal fatigue cracks. Finally, a destructive examination was conducted on the samples. The samples were cut and studied with optical and SE-microscopy to reveal the 49 radial crack growth. When possible, one of the cracks in the sample was opened in liquid nitrogen and the fracture surface was studied with SEM. A TEM study of the fatigued material was also conducted. TEM samples were taken from a depth of about 2 mm. At this depth the calculated ∆ε was much lower than at the surface (see Figures 39 and 40). 3.3 Thermal load determination Thermal loads resulting from the applied thermal loading were determined at the sample surface. For the determination, a sample was instrumented with an air-cooled high temperature strain gage and thermocouples. The instrumented sample was then subjected to thermal cycling in the temperature range of interest. During the thermal cycling, the axial thermal expansion and the temperature of the specimen surface were measured. When the temperature of the sample is known, the expected thermal expansion can be calculated from equation 1. By comparing the expected thermal expansion to the actual measured thermal expansion, the constrained expansion can be calculated taking into account that at the surface the radial stresses are zero and the axial and tangential strains are equal (equation 41, see Appendix 2 for derivation). ε ε ε νσ = − − m T 1 , (41) where εσ is the constrained strain that causes stresses, εT is the calculated thermal expansion and εm is the strain measured from the sample surface. In the linear elastic region, the resultant thermal stresses can be calculated from the Hooke’s law (Boley & Weiner, 1960): σ ν ε ε ε ν ν ε νa a t r aE E= + +( ) +( ) −( ) + +1 1 2 1 , (42) where the subscripts a, t and r denote axial, tangential and radial stresses,    respectively. 3.4 Residual stress profile measurements by contour method To find out the residual stresses inside the sample, the destructive contour method was used. This method was recently introduced by Prime & Gonzales (2000) and it is not a standardized method for residual stress measurements. The method is based on measuring displacement caused by relaxation of residual stresses on a cut surface. As with all relaxation techniques for residual stress measurement, the method assumes, that the relaxation of residual stresses occurs elastically and that the cutting does not induce stresses. In addition, because the displacements are measured directly from the cut surface, in the contour method it is assumed, that the cut occurs along a plane that was flat when the cutting started. This 50 assumption requires, that the sample is constrained from both sides, when the cut is made. The main error sources for this method are inadequate constraint during cutting and the stresses caused by the cutting (Prime, 2001). For a small cylindrical sample, conventional methods such as the hole drilling or the crack compliance methods are not usable. The neutron diffraction method has a spatial resolution of about 1 mm, which is not sufficient to capture the residual stress profile resulting from thermal loads. Hence, the contour method is the only feasible way of measuring the residual stresses inside the samples. For the measurement, a sample was subjected to 10 thermal cycles of interest. 10 cycles were chosen to ensure that a stabilized residual stress configuration was reached, but without crack formation altering the residual stresses. Samples were cut in two by electric discharge machining (EDM), and the shape of the cut surface was measured with an inductive surface finish and form measurement unit (Taylor Hobson Form Talysurf Series 2). Deviations from planeness develop due to relaxation of residual stresses perpendicular to the cut surface. The measured deviations were then inserted into a linear-elastic FE-model, and the pre-cut residual stresses were solved as proposed by Prime & Gonzales (2000) and Prime (2001). The FE-analysis was conducted with a commercially available ANSYS FE-code (ANSYS, release 5.5.1, 1998). An axisymmetric model was used. 3.5 Cyclic stress-strain tests In order to study the cyclic behavior of the test materials as well as to obtain needed material data for the numerical simulations, cyclic stress- strain measurements were conducted on F, G, H and I materials. Measurements were conducted in strain-control with repeating load blocks containing varying amplitudes (Figure 28). The specimen geometry is presented in Figure 29. Cyclic stress-strain curves were determined at room temperature (materials F, G, H and I) and at 100, 200, 300, 400, 500 and 600°C (materials F and G). 51 Figure 28. Schematic image of the loading pattern used in the cyclic stress- strain curve measurements for materials F and G. For materials H and I the load pattern is similar but with εmin=0. Figure 29. The specimen geometry used for cyclic stress-strain measurements. The dimensions are as follows: L=30 mm, D1=7.5 mm, D2=15 mm, R=75 mm. In the tests made with duplex stainless steels (materials H and I) at room temperature, Barkhausen noise was measured. The Barkhausen noise signal originates from the turning of the magnetic domains in ferromagnetic materials. Consequently, in duplex stainless steels the signal comes from the ferrite phase alone. Using the Barkhausen noise data, it was possible to measure the load carried by the ferrite phase. However, the local phase stresses depend on the microstucture, i.e. phase morphology, and the austenite phase stresses cannot be resolved. Previously the load sharing between austenite and ferrite phases has been measured using in-situ X-ray diffraction measurements during tensile testing (Johanson & Odén, 2000a; 2000b; Inal et al., 1999). This is very laborious and time consuming: a single X-ray measurement can take hours to complete. Measurement of repeated cycles is therefore not practical. The Barkhausen noise signal is very sensitive to the magnetic properties of the material, i.e., to the ferrite content of a duplex stainless steel. As the ferrite content varies markedly even within a single sample, each time the Barkhausen sensor was attached to a specimen, a new calibration had to be made. In practice this meant that, before each test the sensor was firmly attached to the sample, and the small, elastic cycles in 52 the beginning of the test were then used for calibration. It was assumed that no microscopic residual stresses exist at the beginning of the test. The Barkhausen noise signal is primarily dependent upon the elastic strain (Stefanita et al., 2000). However, the measurement is also sensitive to hardness, microstructure and prior plastic deformation (Lindgren & Lepistö, 2000a; 2000b). 3.6 Numerical simulations The test sample (Figure 25) was modeled with an axisymmetric, infinite height model (see Appendix 3 for description of the used models). An uncoupled analysis was conducted with 8 node rectangular elements. In thermal analysis, the temperature cycles measured in the thermal load measurements were used as loading. The outer surface of the model was forced to follow the measured temperature cycle. The temperature cycle was divided to about 30 steps, and linear temperature solution was used to solve each step. Solved temperature distributions were then read to the mechanical solution as load steps. A non-linear structural analysis was conducted for the F material for which the cyclic stress-strain curves had been measured. A hybrid multilinear isotropic-kinematic material model was used, and the resulting cycle stabilized within 3 cycles. In addition, all materials and used cycles were analyzed with a linear elastic structural analysis. In the strain-controlled thermal loading, the linear-elastic and elastic- plastic total strain solutions give very similar results. In fact, in many geometries, such as the infinite axisymmetric model considered here, the total strain solution of pure thermal loading is independent of the value of E. Consequently, in the uniaxial case the strain field is not affected by plastic deformation, and a linear model can correctly solve the total strain. In a multiaxial case, the resulting strain field depends also on the effective Poisson's ratio νeff, which changes from elastic value of ν to the plastic value of 0.5 with increasing amount of plastic deformation. This induces some error in the solved strain field in the multiaxial case. If the thermal strains change the residual stresses of the material, the initial strain field is changed for the next load cycle. For cyclic loading the developing residual stresses cause additional error. This affects mainly the mean strain during a cycle and not so much the total strain range solved. Consequently the total strain ranges solved from the elastic and elastic- plastic analyses are very similar, even in the 20 – 600°C temperature cycle where high plastic strains were calculated (Figure 30). The elastic-plastic analysis gives additional information about the stresses and the amount of plastic strain, which cannot be modeled by the elastic analysis. The elastic- plastic solution also gives predictions about the residual stresses caused by 53 the cycle. The FE-analysis was conducted with a commercially available ANSYS 5.5 FE-code (ANSYS Release 5.5.1, 1998). Figure 30. Comparison of the ∆ε distributions solved from elastic-plastic (solid and dashed lines) and linear elastic (dots) models. Dimensions in (mm). 54 4 RESULTS The yield behavior of the studied materials was examined with cyclic stress-strain measurements (materials F and G) and Barkhausen noise measurements (materials H and I). The strains caused by the applied thermal load cycles were characterized with strain measurements on the surface and with numerical modeling. The damage caused by the thermal cycling was studied using replica-assisted microscopy, fracture surface studies (SEM) and dislocation structure analysis (TEM). 4.1 Cyclic stress-strain curves Cyclic stress-strain curves were measured for the solution annealed material F and stabilized material G. Material F showed higher strength than material G during the first load block. Subsequent 10 load blocks caused cyclic softening in material F. Material G showed slight cyclic hardening during the first load blocks and subsequent small cyclic softening. The difference in the cyclic stress-strain behavior decreases the difference of strength between the different heat treatments. The behavior during the 10 first load blocks at room temperature is shown in Figure 31. The stabilized curves at different temperatures are shown in Figure 32.    Figure 31. Measured stresses during the first 10 load blocks measured at room temperature for materials F (a) and G (b). Material F shows cyclic softening after the first load block. Material G shows initial cyclic hardening followed by cyclic softening. 55     Figure 32. Stabilized stress-strain curves measured for materials F and G at different temperatures. Stabilized curves were taken after about 80 load blocks. For material F the cyclic stress-strain curve at 400°C indicates higher strength than at 200°C and 300°C. This hardening can be attributed to dynamic strain ageing. The strength of the materials decreases with increasing temperature. However, for the solution annealed material F the strength at 400°C is higher than at 200°C or 300°C. This effect was not observed in the stabilized material G. The strengthening can be attributed to dynamic strain ageing (DSA). DSA is observed in austenitic stainless steels in the temperature range of 250 – 600°C, depending upon the strain rate (Ilola, 1999; Zauter et al., 1993). Evidently the stabilization heat treatment has decreased the amount of free interstitial solute atoms. Consequently DSA is not observed in material G. The measured isothermal stress-strain curves can be considered representative or conservative as compared to thermomechanical curves (Spindler, 1998; Mallet et al., 1995; Skelton & Webster, 1996). 4.2 Barkhausen noise measurements With Barkhausen noise measurements it was possible to measure the elastic strains of the ferrite phase and the load sharing between the austenite and ferrite phases in duplex stainless steel (materials H and I) during cyclic loading. At small strain amplitudes the load in ferrite is linearly proportional to the macroscopic elastic strain. With increasing load uneven plastic deformation gives rise to interphase residual stresses. The residual stresses form during the first several load-blocks and then stabilize. The effect of residual stresses is that the load sharing between the phases is equalized. 56 The greater yielding in tension in the austenite phase results in tensile residual stress in the ferrite phase, as expected. In material I the microstructure is anisotropic and the phases impose little constraint on each other. At small load amplitudes, both phases deform elastically. However, when the load amplitude increases, the austenite phase starts to yield. This causes the load raise in ferrite to slow down, and tensile residual stress to develop in ferrite. As the load amplitude increases, austenite hardens, and greater residual stresses develop (Figure 33). After about 5 load blocks a saturated state is achieved (Figure 34). Figure 33. The macroscopic stress and phase stresses during first load cycles in material I. Sample was strained as shown in Figure 28 at room temperature. Figure 34. After about 5 load blocks a saturated state is achieved. The macroscopic stress and phase stresses of a few cycles from a stabilized load block are presented for material I. Sample strained at room temperature. In material H the elongated microstructure imposed greater constraint to the phase deformation. During the first load block the behavior is 57 similar to that of material I (Figures 33 and 34). However, with larger stress amplitude the phase stresses cannot develop freely, and the stress in the ferrite increases even though the austenite deforms plastically. The resulting stress cycle is presented in Figure 35. Figure 35. After about 5 load blocks a saturated load block is achieved. The macroscopic stress and phase stresses of a few cycles from a stabilized load block are presented for material H. Sample strained at room temperature 4.3 Cyclic strain caused by thermal loading The axial strains applied to the sample surface could be determined from the measurement. The numerical analysis gave information about the strains inside the specimen. In the austenitic materials (materials A – G) the largest strains were observed in the beginning of the cooling phase, where the temperature change is very quick. In the duplex materials (H and I), where greater heating power was used, the temperature change and, thus, the thermal strains were highest during the heating phase. The applied axial strains were determined at the sample surface. Figure 36 shows a strain plot where the temperature range was increased after each cycle in the austenitic material G. The largest compressive strain depends mainly on the heating power, which was kept constant during the cycles. The largest tensile strain depends mainly on the cooling power, i.e. the temperature difference between the sample and the cooling medium. Hence, the increase in the temperature range increases the largest tensile strain. The strains determined from all the samples are presented in Table 5. 58 Figure 36. Constrained thermal strain plot from material G. During the first 5 cycles the temperature range was increased by 100°C after each cycle by increasing the heating time. Temperature cycles of 20 – 200°C, 20A–A300°C, 20 – 400°C, 20 – 500°C and 20 – 600°C were applied. The temperature cycling was then continued at 20 – 600°C. Table 5. Thermal strain measurements. The applied strains were calculated from the measured temperature and strain using equation 41. Material Thermal cycle (approximate) Minimum thermal strain (µm/m) Maximum thermal strain (µm/m) Thermal strain amplitude (µm/m) A 20 – 300°C -2000 4400 3200 B 20 – 300°C -2000 800 1400 C 20 – 300°C -2000 2000 2000 D 20 – 600°C -1300 6800 4050 E 20 – 600°C -1000 7500 4250 F 20 – 600°C -1700 6700 4200 G 20 – 600°C -1600 6700 4150 H 20 – 280°C -2460 850 1660 I 20 – 280°C -2340 1050 1700 The monotonic yield strain of the studied austenitic stainless steels (materials A – G) is about 1000 µm/m. The determined strain amplitudes exceed the yield strain, and the cycles give rise to plastic deformation both in tension and compression during each cycle. The duplex stainless steels are considerably harder: their yield strain is about 2000 µm/m. The yield 59 strain is exceeded in compression but not cyclically, during every cycle, as the strain amplitude is below this value. Thermal strains were calculated with the finite element method (FEM) from the temperature curves measured on the sample surface. A typical time-varying temperature solution is presented in Figure 37. The corresponding strain solution is presented in Figure 38. The ∆ε distributions calculated from the linear elastic FE-analysis are shown in Figures 39 and 40. The axial strain distribution shows a minimum at about 2.5 mm for all temperature cycles. For the tangential strain solution the minimum is deeper - at about 5 mm. Consequently, the loading is equi- biaxial at the surface, but the tangential strains dominate at a depth of about 2.5 mm. Axial strains then increases again, and finally at depths greater than about 4 mm the axial strains dominate. Figure 37. A typical temperature solution obtained from FEM-calculations. The outer surface of the sample (OD) is forced to follow the measured temperature curve. The calculated temperature response from inner surface (ID) and from about r=7.5 mm (R7.5) are presented. Note, that during the long cooling phase the temperature of the inner surface decreases to room temperature. 60 Figure 38. Axial strain solution corresponding to Figure 37. The ∆ε over the whole cycle is greatest at the outer surface (OD) and smallest at about r=7.5 mm. Figure 39. Axial ∆ε distributions calculated by FEM for different cycles. Sample surface is at r=10 mm. See Appendix 4 for more detailed presentation of this data. 61 Figure 40. Tangential ∆ε distributions calculated by FEM for different cycles. Sample surface is at r=10 mm. See Appendix 4 for more detailed presentation of this data. 4.4 X-ray measurements Residual stresses were measured before the fatigue testing and during the inspections. The tests were stopped for inspection after the cooling phase. Consequently, the residual stresses were measured after the high tensile strains caused by the cooling. After surface yielding in tension, compressive residual stresses are expected. The residual stresses measured are shown in Figures 41 – 45. In materials H and I the measured phase stresses (Figure 45) were converted to macroscopic stresses and microscopic stresses (Figure 47). In all the studied cycles the first large temperature cycles had the largest effect on residual stresses. Then cycles saturated, and, finally, as cracks formed in the sample, the surface residual stresses were relaxed. 62 Figure 41. Residual stresses measured in materials A, B and C. Samples were loaded with the 20 – 300°C thermal cycle. Yielding caused by the large tensile strain at the beginning of the cooling phase causes compressive residual stresses at the surface. The residual stresses saturate after about 10 cycles. Some fluctuation is present in the measured residual stresses due to measurement uncertainty. For the 20 – 300°C cycle (materials A, B and C) compressive residual stresses were measured. Before loading, each sample had a different residual stress state due to the thermal and manufacturing history of the particular sample. For the tangential direction, where the initial residual stresses were low (Figure 41), thermal cycling shifts residual stresses towards compression. After about 10 cycles the residual stresses saturate, and changes caused by subsequent thermal cycling are modest. For materials F – G (Figures 42 and 43) the cycling started with lower temperature amplitude cycles and the first cycle shifts residual stresses towards tension. With increasing temperature amplitude the tensile strain applied increases and the measured residual stresses shift towards compression. A saturated residual stress state is reached after about 10 cycles, where measured residual stresses correspond to the cyclic flow stress of the material. After about 1000 cycles the measured residual stresses decrease due to crack formation. 63 Figure 42. Residual stresses measured in materials D and E. Yielding during the large tensile strain at the beginning of the cooling phase induces compressive residual stresses to the surface. The residual stresses saturate after about 10 cycles. With high number of cycles, crack growth shifts stresses towards tension again. Figure 43. Residual stresses measured from materials F and G. Yielding during the large tensile strain at the beginning of the cooling phase induces compressive residual stresses to the surface. The residual stresses saturate after about 10 cycles. With high number of cycles, crack growth shifts stresses towards tension again. Some fluctuation is present in the measured residual stresses due to measurement uncertainty. 64 Figure 44. Residual stresses measured from material H. Numbers refer to the sample orientation. Sample was loaded with the 20 – 280°C temperature cycle. Figure 45. Residual stresses measured from material I. Sample was loaded with the 20 – 280°C temperature cycle. Measuring residual stresses from the duplex stainless steels proved to be demanding due to the two-phase structure. The varying microstresses between the phases cause fluctuation in the test results. Microstresses also increase the uncertainty of the used sin2Ψ method due to a strain gradient normal to the measured surface. The large grain size of the cast material I significantly weakened the reliability of the measurements. In Figures 44 and 45, the measured stresses are presented. Large fluctuations are evident in measurements from both materials. The calculated micro- and macrostresses are presented in Figures 46 and 47. For the cast material I the fluctuations are still very high and evidently reflects the large measurement uncertainties. For the material H the axial macrostress is 65 shifted towards tension, which is consistent with the calculated strain cycle. Large fluctuations still exist in the calculated microstresses. These fluctuations are due to differences in the phase morphology, which results in different microstresses in different grains. Figure 46. Residual stresses measured from material H (Figure 44) converted to macro- and microstresses. The axial macrostress is initially highly compressive in material H. The compressive stress during thermal cycling causes yielding in compression and shifts stresses towards tension. Figure 47. Residual stresses measured from material I (Figure 45) converted to macro- and microstresses. The width of the X-ray diffraction peak was also monitored. In Figures 48 – 52 the measured peak's full width half maximum (FWHM) values are presented. Initially samples showed FWHM values varying from about 1.5 to 2.5 degrees. This variance is caused by the different manufacturing 66 histories of the samples. Thermal fatigue loading decreased the FWHM values in the samples with high initial values and increased FWHM in the samples with low initial values. The FWHM values of the samples stabilize to a value of about 2 in all of the materials and temperature cycles. Figure 48. FWHM values measured from materials A, B and C. Samples were loaded with the 20 – 300°C temperature cycle. Changes in the FWHM values are modest. Slight decrease in the initially higher FWHM value of material C is observed and stabilization to a level of about 2 occurs. For materials A, B and C the observed changes of FWHM due to thermal cycling are modest. The initial FWHM values around 2 are not altered by thermal cycling. Samples D, E and F were machined after heat treatment. Deformation during manufacturing has evidently increased the dislocation density in the sample, resulting in high initial FWHM values. Thermal cycling is observed to reduce the FWHM values. Sample G was stabilization heat treated after machining. Consequently the dislocation density is expected to be low. Also, the FWHM values measured in the sample before the cycling were very low. In this material, thermal cycling increased the FWHM values. After 2000 cycles the FWHM values reached a value of about 1.8. 67 Figure 49. FWHM values measured in materials D and E. After deformation caused by machining the dislocation density of the sample surfaces can be assumed to be quite high. Consequently also the FWHM value measured is quite high. Thermal cycling decreases FWHM until it stabilizes to about 2. Figure 50. FWHM values measured in the materials F and G. In material F the initial dislocation density after machining is high and a decrease of FWHM similar to that of materials D and E is observed (Figure 49). The material G samples were solution heat treated after machining. Consequently the initial FWHM value (and dislocation density) is low and thermal cycling increased the FWHM value. FWHM values of both materials stabilize to about 2. 68 For materials H and I, large fluctuations are observed in the FWHM values. In the duplex stainless steels this is due to the scatter in microstresses between grains. The fluctuations in ferrite phase are smaller than in the austenite phase. This can be attributed to the microstructure; ferrite forms a continuous matrix, in which separate austenite grains are embedded. Figure 51. FWHM values measured in material H. Sample was loaded with the 20 – 280°C temperature cycle. In the austenite phase a stabilization of FWHM values to about 2 is seen. The FWHM values of the ferrite phase are unaffected by the thermal cycling. Figure 52. FWHM values measured in material I. Sample was loaded with the 20 – 280°C temperature cycle. The FWHM values of the ferrite phase are unaffected by the thermal cycling. 69 4.5 Contour method The residual stress profiles of 8 samples of materials B to I loaded with 10 thermal cycles were measured with the contour method (Prime & Gonzales, 2000; Prime, 2001). The measurements presented here are not directly comparable to residual stress measurements presented in paragraph 4.4 due to possible stress relaxation before the residual stress measurements. The residual stress profiles measured are presented in Figure 53 together with FEM calculations and X-ray data. Figure 53. Residual stresses measured in the axial direction with the contour method presented with the FEM results and X-ray data. The residual stress profiles measured show a point of zero stress at a depth of about 2.5 mm. The strain range solution from the FE analysis also showed a minimum at a depth of about 2.5 mm (Figure 53). However, the elastic-plastic FE analysis predicts plastic deformation only at the very surface of the sample thus deviating considerably from the measured profiles. The austenitic samples (materials B – G) show compressive residual stress on the surface due to yielding in tension during the cooling phase. In the duplex samples (materials H – I) stresses were highest during the heating phase and yielding in the surface occurred in compression. Consequently, tensile residual stresses were measured. 4.6 Build-up of surface damage In the austenitic materials (A – G), fatigue damage started with the formation of persistent slip bands (PSBs). As the fatigue continued, more PSBs formed and existing PSBs grew to form microcracks. Microcracks 70 formed in all directions and no preferred orientation was seen. This indicates an equi-biaxial stress state. Soon, the cracks formed a mosaic-like network typical for thermal fatigue. In such a network, the length and growth of individual cracks could not be monitored from the surface. In Figures 54 and 55 the crack growth of different materials as seen in the replicas is presented. Cracks with surface length of 100 µm or more were observed after about 500 – 1500 cycles. Figure 54. Crack growth as seen in the replicas. On the left material D is shown after 100 (a), 500 (b) and 1500 (c) cycles of 20 – 600°C. On the right material E is shown after 100 (d), 500 (e) and 1500 (f) cycles of 20 – 600°C. 71 Figure 55. Crack growth as seen in the replicas. On the left material F is shown after 200 (a), 1500 (b), 2000 (c) and 5000 (d) cycles of 20 – 600°C. On the right material G is shown after 200 (e), 1500 (f), 3000 (g) and 5000 (h) cycles of 20 – 600°C. In the duplex stainless steel samples (H and I) cracks were seen to initiate from PSBs in austenite grains and from MnS inclusions. The primary initiation site depended upon the size of the inclusions; if the inclusions were bigger than the available austenite grains, then dominating 72 cracks initiated from the inclusions. This behavior resulted in a significant anisotropy in the fatigue properties of the wrought material H. In material H, where the nitrogen content in the austenite is lower, PSBs often occurred in multiple directions. In the material I with the higher austenite nitrogen content, PSBs formed predominantly in a single direction within the grains. Phase boundaries were seen to retard the crack growth significantly. Often microcracks initiated in two adjacent austenite grains, and the phase boundary was overcome only when the ferrite ligament between them finally broke. In the duplex stainless steels, the growth of individual cracks could be measured in the replicas. The measured crack growth data for materials H and I and temperature cycle 20 – 280°C are presented in Figure 56. Figure 56. Crack growth in materials H and I, temperature cycle 20 – 280°C. Material H showed significant anisotropy in crack growth. Crack growth in orientation 1 (see Figure 24 for the definition of orientations) was so small that it could not be measured. 4.7 Fractography In all samples, where possible, an axial crack was opened for fractographic examination. In the austenitic materials (A, D, E, F and G) the fracture surfaces showed extensive striation formation (Figure 57). The striation spacing was measured at several locations on the fracture surface. The measured striation spacings are shown in Figure 58. 73 Figure 57. Typical fracture surface of austenitic materials showed extensive striation formation. Image from material A after 15625 cycles of 20 – 300°C. Figure 58. Striation spacings measured from materials C, D, E, F and G. In the measured range the striation spacing can be considered as an indication of the crack advance per cycle. The scatter in local crack growth rate increased with increasing crack length. The observed decrease in the macroscopic crack growth rate scatter is due to averaging of local crack growth rates in the crack front. By taking the striation spacing as a direct indication of the instantaneous crack growth rate and assuming that the crack growth rate changes linearly between the measured points, the crack growth as a function of cycle number could be calculated. The obtained growth curves are presented in Figure 59. 74 At small crack depths the striation spacing is small. With increasing crack depth the striation spacing increases sharply although the nominal ∆ε at the crack tip location diminishes. After the crack depth reaches about 1.5 mm, the effect of decreasing strain amplitude surpasses the effect of increasing crack depth and the crack growth slows down. Figure 59. Crack growth calculated from the measured striation spacings. In the duplex stainless steels, the two-phase microstructure significantly affected the crack growth. Fracture surfaces showed signs of many different crack growth mechanisms. Occasionally in the austenite phase, ductile striations similar to those in austenitic materials were seen (Figure 60). In addition, crack growth by alternating fracture plane (Figure 61) and in ferrite, in one location, brittle striations were observed (Figure 62). 75 Figure 60. Ductile striations similar to those in austenitic materials in the austenite phase of material I after 20000 cycles of 20 – 280°C. Figure 61. Crack growth by alternating fracture plane. Austenite phase of material H after 20000 cycles of 20 – 280°C. 76 Figure 62. Brittle striations in the ferrite phase of material I after 20000 cycles of 20 – 280°C. The edge of the fracture surface and the sample surface was studied to reveal the crack initiation and growth in the two-phase materials. Figure 63 shows that the crack growth mechanism is different in the austenite and ferrite phases. In austenite a pattern typical for stage I fatigue crack growth is seen, while in ferrite no deformation outside the crack wake is visible. Also, secondary cracks in PSBs are observed in stage II fatigue crack growth (Figure 64). Crack initiation was seen to be caused by extrusions and intrusions in the austenite grains (Figure 65). Figure 63. Crack grows from ferrite (right) to austenite (left) in material I. See text for details. 77 Figure 64. Secondary cracks along PSBs in stage II crack growth (material H) (schematic graphics after Suresh, 1992). Figure 65. Extrusions and intrusions on sample surface of material I (schematic graphics after Suresh, 1992). 4.8 Dislocation structures The TEM studies showed that dislocation density in all studied materials was lower than expected based on mechanical fatigue data. Dislocation 78 tangles and occasional cell tendency were observed. Typical dislocation structures observed are shown in Figures 66 – 67. Figure 66. Dislocation structure of material B after 15625 cycles of 20 – 300°C. Dislocation tangles and low dislocation density was observed. Figure 67. Dislocation structure of material F after 10000 cycles of 20 – 600°C. Dislocation tangles and low dislocation density was observed. Numerous Nb-rich precipitates are visible. 79 In the duplex stainless steels the dislocation density of the austenite phase was greater than that of the ferrite phase. This indicates that plastic deformation has been localized in the austenite phase. Typical dislocation structures are shown in Figures 68 – 69. More stacking faults are observed in the austenite phase of material H than of material I or in the studied austenitic materials (A – G). This indicates that the stacking fault energy (SFE) of the austenite phase of material H is lower than that of other studied materials. In material H, small amounts of χ-phase were observed in phase boundaries (Figure 70). Figure 68. Dislocation structure of material H after 15625 cycles of 20 – 280°C. Planar slip bands are observed in the austenite phase. The dislocation density of austenite is higher than that of ferrite. Numerous stacking faults are observed in the austenite phase. 80 Figure 69. Dislocation structure of material I after 15625 cycles of 20 – 280°C. Dislocation tangles are observed in the austenite phase. The ferrite phase shows parallel dislocation alignments. The dislocation density of austenite is higher than that of ferrite. Figure 70. The secondary phase observed at the phase boundaries of material H was identified as χ -phase. 81 5 DISCUSSION 5.1 Cyclic thermal loads and residual stresses The surface stresses during all of the used thermal cycles exceeded the monotonic yield limit of the material. Consequently, the thermal cycling altered the residual stresses of the samples. A schematic stress-strain curve of the load cycle on the sample surface of the austenitic stainless steel samples is presented in Figure 71 (see also Appendix 1). Figure 71. A schematic stress-strain curve for the surface of an austenitic sample during thermal cycling. In the beginning of the cycle heating causes compressive stress at the specimen surface (0 – 1). When the cooling phase begins, high tensile stresses and yielding are induced to the sample surface (1 – 2). When cooling is continued, the temperature distribution in the sample evens out (2 – 3). Because of uneven yielding during the cooling phase, a compressive residual stress is produced at the sample surface (3). The surface residual stresses changed markedly during the first load cycles (Figures 41 – 47). After about 10 cycles the stresses stabilized. In cycles with a high strain amplitude, the stabilized value corresponded to the cyclic flow stress of the cyclic stress-strain curve (materials F and G). In these samples the residual stress state prior to thermal cycling had no effect on the final stress state. In samples where the thermal strains were lower, the prior residual stresses also had an influence on the stabilized residual stresses. After continued fatigue, thermal fatigue cracks started to alter the residual stresses and the residual stresses shifted towards tension. The 20 – 600°C cycle (material F) was modeled with an elastic-plastic numerical model, which predicted the residual stress state. In Figure 53 the calculated residual stresses are compared with the stresses measured by the contour method and X-ray diffraction. The agreement of different 82 measurement techniques is good, while the agreement with the numerical simulation is poor. The FE-model predicts plastic strains only on the very surface, whereas the contour measurements show that plastic deformation is more evenly distributed in the sample. The difference is apparently because the yield strength in the stabilized stress-strain curve is larger than the yield strength observed in the monotonic tensile test or during the first fatigue cycle. The used material model and data describe heavily deformed material, such as that in the surface layer. Therefore, the surface residual stress is correctly predicted by the FE-model. TEM images show, however that this does not describe the whole sample (Figures 66 – 67). To correctly model the loading and to correctly predict the residual stresses, a history- dependent material model should be used, which would update the stress- strain curve on each load reversal based on prior deformation. Such a material model is not available, and hence the plastic strains in the thermal fatigue test could not be adequately modeled. In duplex stainless steels, the plastic deformation in uniaxial mechanical loading was seen to concentrate in the austenite phase. The concentration was greater in the cast material I, where the phase morphology caused little interphase constraint. Uneven plastic deformation in tension resulted in compressive residual microstress in the austenite phase and tensile microstress in the ferrite phase, as already noted by several authors (Kamachi et al., 1982; Siegmund et al., 1993). The concentration of plastic deformation in the austenite phase was also seen from the TEM-studies of thermally fatigued samples. 5.2 Micromechanism of crack growth In the studied austenitic materials, cracks initiated from PSBs. In materials with larger grain size, the PSBs were broader and the initiation occurred more easily. In duplex stainless steels the initiation was observed at PSBs in austenite grains (both materials H and I) and at MnS inclusions (material H, orientation 2). In material H, where the inclusions were bigger than the austenite grains, the dominating cracks initiated from these inclusions. In material I, where round inclusions were smaller than the austenite grain size, the inclusions had no notable effect on the fatigue behavior. The fatigue life of the specimens was seen to depend on the largest initiation site available. The biggest cracks were found from the orientation 2 of the material H, where elongated MnS inclusions were available as crack initiators. In the orientation 1 of material H the MnS inclusions were not available and the austenite phase size was very small. Consequently, the crack initiation occurred later than in other cases, and crack growth was considerably slower. In material I the crack growth was between these two extremes. The cracks initiated from PSBs in the austenite grains, but due to the bigger grain size of material I, crack growth was faster than in orientation 1 of material H. 83 The fracture surfaces of the austenitic test materials showed extensive striation formation (Figure 57). This is typical for austenitic stainless steels also in mechanical fatigue. Similar striations were seen in the cyclic stress- strain samples, which were isothermally fatigued at different temperatures. The fracture surfaces of the austenitic materials are therefore very similar to those of mechanical fatigue. The fracture surfaces showed no evidence of stress corrosion fatigue, and the possible environmental effects of the water quenching used are a result of true corrosion fatigue. In the duplex stainless steels the two-phase microstructure allowed many different fracture mechanisms to be active. Ductile striations, similar to those in austenitic materials, were seen in the austenite phase of the duplex samples. In addition, fracture by altering fracture plane and brittle- like fracture, both transgranular and intergranular, were seen. These complex fracture surfaces are also typical in mechanical fatigue of duplex stainless steels. Apparently the interphase stresses and the different crack growth mechanims in the two phases cause high local fluctuations in the strains at the crack tip, and thus the complex fracture surfaces observed. While the micrometer-scale fatigue mechanism of thermal fatigue seems to be similar to that of mechanical fatigue, the dislocation structures observed are quite different. In mechanical and thermomechanical fatigue with externally applied loads or constraints, tightly arranged dislocation structures, such as walls-and-channels, cells or corduroy, are typical (see paragraph 1.4). In the present study, dislocation tangles and low dislocation density were observed (Figures 66 – 67). However, similar dislocation structures have recently been found in AISI 304 cracked cladding of recovery boiler composite tubes (Keiser et al., 1996). At first this dislocation arrangement was considered as evidence of the absence of thermal fatigue. Further study (Keiser et al., 1997) with a thermally fatigued reference samples showed that the observed dislocation structure is typical for thermal fatigue. Other, more recent, studies on recovery boiler composite tubes confirmed those results (Saarinen et al., 2000). The low dislocation densities observed in the present study can be attributed to the fact that the TEM sample was taken from a depth of about 2 mm. The strain amplitude seen by the material at this depth is much smaller than at the surface. The axial and tangential strain amplitudes are about 14% and 40% of the maximum amplitude, respectively. It is generally found that cold work and fatigue tend to increase the width of X-ray diffraction peaks. From Figures 48 – 52 it is seen that this is not necessarily the case in thermal fatigue. The diffraction peaks of some samples do show a small increase, while for other samples a decrease in FWHM is observed. The increase of FWHM is observed predominantly in stabilized samples, where initial dislocation density can be assumed to be low. The decrease of FWHM is seen in materials, where initial dislocation 84 density is higher due to sample manufacturing. The FWHM values of different materials tend towards the value of about 2. This indicates that the low dislocation density observed in the TEM studies is characteristic to thermal fatigue loading. 5.3 Thermal fatigue prediction Based on the replica studies and fractographic studies, the fatigue crack growth rate in the samples could be estimated. Following the interpretation of Chopra & Shack (1999), an S-N plot for the number of cycles required to grow a 3 mm deep "engineering crack" can be drawn (Figure 72). The load is plotted according to the calculated total surface strain range. The strain range decreased rapidly towards the inside of the specimen. Consequently the plot should be considered as a rough estimate only. Figure 72. S-N plot for cycles required to grow a 3 mm deep surface crack. Total strain range is plotted according to calculated surface strain range. Open dots are run-outs. The strains were obtained from FEM-calculations (Figure 40) and the cycle numbers from striation spacing measurements (Figure 59) and replica studies (Figure 56). Numerous crack growth models have been proposed in literature (see paragraph 1.8). The application of most of these models to the current thermal fatigue load case is, however, not trivial. The linear-elastic K- based models are not applicable to current thermal fatigue crack growth due to plasticity in the specimen surface layer. To numerically determine the J-values, the crack tip stresses and strains (elastic and plastic) must be known. The stresses and plastic strains of even an uncracked geometry could not be reliably solved using the current nonlinear material models (paragraph 5.1). In addition, the temperature gradients present make the numerical evaluation of J more difficult. Hence, the ∆J-distribution of the 85 current thermal fatigue load case could not be determined numerically. Similarly, lack of a reliable cyclic, non-linear stress solution for the uncracked geometry prevents the direct use of some strain-based parameters, which require the plastic strain amplitude to be known. Equations 17 and 27 can be used with q set to 1, i.e. εeq=εtot, together with the Paris law (a similar approach was used earlier by Solomon, 1972) (equation 43): da dN C atot m = [ ]2 2∆ε π . (43) The correlation with experimental data is, however, poor, as seen in Figure 73. Figure 73. Crack growth rates calculated from equation 43 and measured striation spacings. A much better correlation is achieved using equation 30 with total strain, i.e.: da dN C atot m= 7 7∆ε . (44) A similar approach has been used by Solomon (1972), Skelton et al. (1996), Taira et al. (1976), Nisitani et al. (1992) and Kawagoishi et al. (2000). Crack growth rate curves based on equation 44 are presented in Figure 74. Because of the inherent scatter in the striation spacing measurements, correlation with experimental data points can be considered good. The thermal fatigue crack growth rates for several materials and temperature cycles can, thus, be predicted based on the linear-elastic total strain solution of an uncracked component together with equation 44, as long as the required parameters C7 and m7 are 86 experimentally determined. In Figure 75 experimental thermal fatigue data from Marsh (1981) and the present study is presented. The agreement with the present results is excellent. Figure 74. Crack growth rates calculated using equation 44 and measured striation spacings. The open symbols represent data points measured at very low crack growth rates. These points were not included in the curve fitting. Striation spacing divided by the crack length (S/a) is plotted against the strain range (∆ε) on a log-log scale to give linear fit for equation 44. The crack growth predicted by Figure 74 is presented in Figure 76. In order to calculate the crack growth from equation 44, an initial crack length a0 is assumed. A value of a0=50 µm was used based on the upper limit grain size of the materials. Due to fast crack initiation, the estimation is not sensitive to the chosen value of a0. The assumed initial crack size adds conservativism to the prediction. In the case of the 20 – 600°C temperature cycle this effect can be estimated based on the replica studies (Figures 54 and 55) to be about 500 – 1500 cycles. With decreasing strain range the number of cycles required to form an initial crack of 50 µm increases. 87 Figure 75. Data from Marsh (1981) and from the present study. Striation spacing divided by the crack length (S/a) is plotted against the strain range (∆ε) on a log-log scale to give linear fit for equation 44. Figure 76. Crack growth measured from striation spacing (dots) and calculated from equation 44 (solid lines) with initial crack size of 50 µm. The main difference between equations 43 and 44 is the exponent imposed on the crack depth a. In equation 43 the exponent is m2/2, while in equation 44 it is 1. In the case of m2=m7=2, the two equations are equal. 88 The a-dependence has also been noticed for high strain mechanical fatigue for different steels (Solomon, 1972; Tomkins, 1981; Skelton, 1993; Skelton et al., 1996; Nisitani et al., 1992). There is some theoretical (Liu & Kobayashi, 1980) and experimental studies (Bates & Clark, 1969), which indicate that m2=2 in many materials. In published literature (e.g., Boyer, 1986) m2 values ranging from 2 to 4 are presented. Values quoted for the materials currently studied are typically about 3. Crack growth in the duplex stainless steels was calculated based on the replica measurements shown in Figure 56. The calculated crack growth data are presented in Figure 77. No correlation with equation 44 can be observed. This is apparently due to the large effect of the phase boundaries on the crack growth rate, which causes the crack growth rate to vary greatly. Figure 77. Crack growth rates measured at the sample surface (materials H and I). The crack growth rate estimated from the replica studies divided by the crack length (S/a) is plotted against strain range (∆ε) on a log-log scale to give linear fit for equation 44. 5.4 Comparison of thermal and mechanical fatigue The observed fatigue lives shown in Figure 72 are shorter than those observed for mechanical fatigue (see Figure 17), even when the strains are presented according to the maximum value at the surface. This difference can be attributed to the biaxial stress state and environmental effects, which promote crack initiation and small crack growth at the surface. To compare the present result with S-N plots for mechanical fatigue, the data points shown in Figure 72 were converted to equivalent uniaxial strains 89 with equations 36 and 37, following the analysis by Marsh (1981) (elastic and plastic strains were calculated according to monotonic yield stress). In Figure 78 the converted points are superimposed on Figure 17. The points from the present study agree with the isothermal 527 – 600°C fatigue and thermal fatigue data for AISI 316L steel. Figure 78. Data points from Figure 72 converted to equivalent uniaxial strains with equations 36 and 37 and superimposed on S-N plot from Petersen & Rubiolo (1991) (Figure 17). Open dots are run-outs. In the ASME pressure vessel code (1995) the effects of biaxial loading and environment are not explicitly taken into account. To allow for these effects, the ASME design curves include a safety factor of about 2 on strain and 20 on cycles. To assess the sufficiency of these margins in the studied load case, the data is also presented in the ASME curve (Figure 79). It is seen that the data points are correctly predicted failures and that the ASME design curve gives a safe design life, although the remaining safety factor is decreased to about 1.5 on strain. 90 Figure 79. Data points converted from Figure 72 and plotted on the ASME design curve for austenitic stainless steels (the results on duplex stainless steels are not plotted). Material color codes as in figure 78. Open dots are run-outs. In paragraph 5.3, it was shown that thermal fatigue crack growth can be predicted from the total strain solution of an uncracked specimen and the crack depth using equation 44. In open literature similar models have been shown to be successful in predicting crack growth for mechanical fatigue (Solomon, 1972), high strain fatigue (Skelton et al., 1996), high temperature isothermal fatigue (Taira et al., 1976), and small crack growth (Nisitani et al., 1992; Kawagoishi et al., 2000). As suggested by Nisitani et al. (1992), an equation similar to 44 describes crack growth when the small scale yielding condition of LEFM is violated by high stresses. The similarity in the crack growth models for mechanical and thermal fatigue crack growth indicates that the thermal fatigue crack growth mechanism in the studied regime is similar to the mechanical fatigue crack growth mechanism. If the complex thermal loading is modeled as suggested in paragraph 5.3, the same equation, i.e. equation 44, can be used to predict both thermal and mechanical fatigue crack growth. The similarity seen in crack growth laws for mechanical and thermal fatigue also indicates that the strain-based crack growth model (equation 44) is valid for mechanical fatigue cracks growing in a varying strain and temperature field. Hence, the present results extend the verified range of validity of equation 44. 91 6 CONCLUSIONS The following conclusions can be drawn from the studies: 1. The fracture surfaces of thermal fatigue cracks are similar to those of mechanical fatigue. 2 . The dislocation structures caused by thermal fatigue differ considerably from those of mechanical fatigue. The width of the X- ray diffraction peaks do not necessarily increase in thermal fatigue loading. 3. The load sharing between the phases in duplex stainless steels is more even in the wrought material than in the cast material due to the different phase morphology. 4. The total strains caused by thermal loading can be approximated with a simple linear-elastic FE-analysis for many geometries and load cases. 5. The ASME (1995) pressure vessel code gives a safe design life for thermal fatigue loading; the remaining safety factor is about 1.5 on strain. 6. The total strain solution for an uncracked geometry can be used to predict the crack growth rate. The crack growth is observed to follow the strain-based crack growth equation:       da dN C atot m= 7 7∆ε , whereC7=1.6 and m7=1.3 for the studied austenitic materials. 7. Similar crack growth models have been shown to be successful in predicting mechanical fatigue crack growth under various load cases and temperatures. In the present work the model is shown to be successful also for small cracks growing in varying temperature and strain fields. 8. The similitude in the crack growth models for thermal and mechanical fatigue indicates that the crack growth mechanism is similar. 92 7 REFERENCES Ahmad, H. Y. & Yates, J. R. 1994. An elastic-plastic model for fatigue crack growth at notches. Fatigue and Fracture of Engineering Materials and Structures, 17 (6), pp. 651 – 660. Almer, J. D., Cohen, J. B. & Moran, B. 2000. The effects of residual macrostresses and microstresses on fatigue crack initiation. Materials Science and Engineering, A284, pp. 268 – 279. ANSYS Release 5.5.1.1998.UP19981001.SAP IP Inc. Armas, A. F., Alvarez-Armas, I. & Petersen, C. 1992. Thermal fatigue behavior and dislocation substructures of 316-type austenitic stainless steels. Jounal of Nuclear Materials, 191 – 194, pp. 672 – 675. ASME Boiler and Pressure Vessel Code. 1995. Section III, Division 1, Appendix 1. Design stress intensity values, allowable stresses, material properties, and design fatigue curves Bates, R. C., & Clark, W., G., Jr. 1969. Fractography and fracture mechanics. Transactions of the Americal Society for Metals, 69, pp. 380 – 388. Behnken, H. & Hauk, V. 1992a. On the influence of microresidual stresses during cyclic loading. Proc. 3rd European Conference on Residual Stresses. Frankfurt-am-Main. Germany. 4 – 6 Nov. DGM Informationsgesellschaft mbH. pp. 733 – 742. Behnken, H. & Hauk, V. 1992b. X-ray studies on a friction welded duplex steel. Proc. 3rd European Conference on Residual Stresses. Frankfurt-am- Main. Germany. 4 – 6 Nov. DGM Informationsgesellschaft mbH. pp. 165 – 170. Bichler, Ch. & Pippan, R. 1999. Direct observation of fatigue crack tip deformation in the midsection of a specimen. The seventh international fatigue congress, Fatigue '99, Beijing, P. R. China, June 8 – 12. Beijing institute of aeronautical materials, Institute of metal research, Fatigue society, pp. 2789 – 2794. Boley, B. & Weiner, J. 1960. Theory of thermal stresses. John Wiley & Sons. 586 p. ISBN 0-486-69579-4 Boyer, H. E. (ed.) 1986. Atlas of Fatigue Curves. American Society for Metals. Metals Park. Ohio, p. 214 93 Bressers, J., Hurst, R. C., Kerr, D. C., Lamain, L., Sordon, G. & Tartaglia, G.P. 1994. Thermal fatigue crack growth: modelling and experimental verification. Journal of Nuclear Materials, 212 – 215, pp. 448 – 452. Brown, M. W. & Miller, K. J. 1982. Two decades of progress in the assessment of multiaxial low-cycle fatigue life. In: Low-Cycle Fatigue and Life Prediction, Amzallag, C., Leis, B. N. & Rabbe, P., Eds., ASTM-STP 770, Americal Society for Testing and Materials, pp. 482 – 499. Chen, J., Takezono, S., Tao, K. & Hazawa, T. 1997. Application of fracture mechanics to the surface crack propagation in strainless steel at elevated temperatures. Acta Materialia, 45 (6), pp. 2495 – 2500. Cheng, A. S. & Laird, C. 1983. The transition from stage I to stage II fatigue crack propagation in copper single crystals cycled at constant strain amplitudes. Materials Science and Engineering, 60, pp. 177 – 183. Chopra, O. K. & Gavenda, D. J. 1998. Effects of LWR coolant environments on fatigue lives of austenitic stainless steels. Transactions of the ASME, 120, pp. 116 – 121. Chopra, O. K. & Muscara, J. 2000. Effects of light water reactor coolant environments on fatigue crack initiation in piping and pressure vessel steels. Proceedings of ICONE8 - 8th International Conference on Nuclear Engineering, April 2-6, Baltimore. ASME. pp. 1 – 12. Chopra, O. K. & Shack, W. J. 1999. Overview of fatigue crack initiation in carbon and low-alloy steels in light water reactor environments. Journal of Pressure Vessel Technology, American Society for Testing and Materials. 121, pp. 49 – 60. Coffin, L. F., Wesley, R. P. & Schenectady, N. Y. 1954. Apparatus for study of effects of cyclic thermal stresses on ductile metals. Transactions of the ASME, 76, pp. 923 – 930. Coffin, L. F. & Schenectady, N. Y. 1954. A study of the effects of cyclic thermal stresses on a ductile metal. Transactions of the ASME, 76, pp. 931 – 950. Cullity, B. D. 1978. Elements of X-ray Diffraction. 2nd ed. Addison Wesley Publishing Company. 555 p. ISBN 0-201-01174-3. Davidson, D. L. & Lankford, J. Jr. 1976. Plastic strain distribution at the tips of propagating fatigue cracks. Transactions of the ASME, 98(1), pp. 24 – 29. 94 Davidson, D. L. & Lankford, J. 1992. Fatigue crack growth in metals and alloys: mechanisms and micromechanics. International Materials Reviews, 37 (2), pp. 45 – 76. Dean, S. 2000. Chloride SCC of stainless steel? No – cyclic strain cracking. Materials Performance, Sept., pp. 78 – 82. dell'Erba, D. N., Aliabadi, M. H. 2000. Three-dimensional thermo- mechanical fatigue crack growth using BEM. International Journal of Fatigue, 22, p. 261 – 273. Degallaix, S., Seddouki, A., Degallaix, G., Kruml, T., Polák, J. 1995. Fatigue damage in austenitic-ferritic duplex stainless steels. Fatigue and Fracture of Engineering Materials and Structures, 18 (1), pp. 65 – 77. Dowling, N. E. 1976. Geometry effects and the J-integral approach to elastic-plastic fatigue crack growth. In: Cracks and Fracture, ASTM-STP 601, American Society for Testing and Materials, pp. 19 – 32. Dowling, N. E. 1977. Crack growth during low-cycle fatigue of smooth axial specimens. In: Cyclic Stress-Strain and Plastic Deformation Aspects of Fatigue Crack Growth, ASTM-STP 637, American Society for Testing and Materials, pp. 97 – 121. Dowling, N. E. & Begley, J. A. 1976. Fatigue crack growth during gross plasticity and the J-integral. In: Mechanics of crack growth, ASTM STP 590, American Society for Testing and Materials, pp. 82 – 103. Endo, K, Komai, K, Murayama, S. 1983. Influence of Cl- concentration on corrosion fatigue crack growth of an austenitic stainless steel. Bulletin of the JSME, 26(218), pp. 1281 – 1287. El Haddad, M. H., Dowling, N. E., Topper, T. H. & Smith, K. N. 1980. J integral applications for short fatigue cracks at notches. International Journal of Fracture, 16 (1), p. 15 – 30. Elber, W. 1970. The significance of fatigue crack closure. In: Damage Tolerance in Aircraft Structures, ASTM-STP 486, American Society for Testing and Materials, pp. 230 – 242. Fissolo, A., Marini, B., Nais, G. & Wident, P. 1996. Thermal fatigue behaviour for a 316L type steel. Journal of Nuclear Materials, 233 – 237, pp. 156 – 161. Fischer, F. D., Rammerstorfer, F. G. & Bauer, F. J. 1990. Fatigue and fracture of high-alloyed steel specimens subjected to purely thermal cycling. Metallurgical Transactions, 21A (4), pp. 935 – 948. 95 Fischer, F. D., Mayrhofer, K. & Parteder, E. 1996. Elliptical subsurface cracks under a normal stress and a residual stress field. Fatigue and Fracture of Engineering Materials & Structures, 19 (1), pp. 129 – 139. Ford, F. Prediction of corrosion-fatigue initiation in low-alloy steel and carbon-steel/water systems at 288°C. In: Proceedinngs of the Sixth International Symposium on Environmental Degradation of Materials in Nuclear Power Systems-Water Reactors. Gold, R. E., Simonen, E. P. (Eds.). August 1-5, San Diego, California. The Minerals, Metalss & Materials Society. pp. 9 – 18. Fritz, R. J. & Schenectady, N. Y. 1954. Evaluation of transient temperatures and stresses. Transactions of the ASME, 76, pp. 913 – 920. Gabetta, G., Rinaldi, C. & Pozzi, D. 1990. A model for environmetally assisted crack growth rate. In: Environmentally Assisted Cracking: Science and Engineering, Lisagor, W. B., Crooker, T. W. & Leis, B. N. (Eds.), ASTM-STP 1049, American Society for Testing and Materials, pp. 266 – 282. Gandossi, L. 2000. Crack growth behaviour in austenitic stailess steel components under combined thermal fatigue and creep loading. Doctor of Philosophy thesis. University of Wales, Swansea, 419 p. Gerland, M, Mendez, J., Violan, P. & Ait Saadi, B. 1989. Evolution of dislocation structures and cyclic behaviour of a 316L-type austenitic stainless steel cycled in vacuo at room temperature. Materials Science and Engineering, A 118, pp. 83 – 95. Gerland, M., Alain, R., Ait Saadi, B. & Mendez, J. 1997. Low cycle fatigue behaviour in vacuum of a 316L-type austenitic stainless steel between 20 and 600°C – Part II: dislocation structure evolution and correlation with cyclic behaviour. Materials Science and Engineering, A229, pp. 68 – 86. Goswami, T. 1999. High temperature fatigue – dwell sensitivity and life prediction. Doctor of Technology thesis. Helsinki University of Technology. Laboratory of Engineering Materials. TKK-MTR-5/99. ISBN 951-22-4794-1. Green, D., Parker, R. & Marsh, D. 1987. Comparison of theoretical estimates and experimental measurements of fatigue crack growth under severe thermal shock conditions – Part II: theoretical assessment and comparison with experiment. Journal of Pressure Vessel Technology, 109, pp. 421 – 427. Green, D. & Munz, D. 1996. Thermal fatigue crack growth experiments on austenitic steel plates. International Journals of Pressure Vessels & Piping, 65, p. 369 – 378. 96 Griffith, A. A. 1920. The phenomena of rupture and flow in solids. Philosophical Transactions of the Royal Society of London, 221, pp. 163 – 198. Haigh, J. R. & Skelton, P. 1978. A strain intensity approach to high temperature fatigue crack growth and failure. Materials Science and Engineering, 36, pp. 133 – 137. Hatanaka, K., Fujimitsu, T. & Shiraishi, S. 1989. An analysis of surface crack growth in circumferentially grooved components under low-cycle fatigue. JSME International Journal, 32(2), pp. 245 – 255. Hatanaka, K. 1990. Cyclic stress-strain response and low-cycle fatigue life in metallic materials. JSME International Journal, 33 (1), pp. 13 – 25. Hayashi, M. 1994. High cycle thermal fatigue crack initiation and growth behavior in the semi-infinite plate model. ASME, PVP, Vol. 287, pp. 49 – 54. Hayashi, M. 1998a. Thermal fatigue strength of type 304 stainless steel in simulated BWR environment. Nuclear Engineering and Design, 184, pp. 135 – 144. Hayashi, M. 1998b. Thermal fatigue behavior of thin walled cylindrical carbon steel specimens in simulted BWR environment. Nuclear Engineering and Design, 184, pp. 123 – 133. Hayashi, M., Enomoto, K., Saito, T. & Miyagawa, T. 1998. Development of thermal fatigue testing apparatus with BWR water environment and thermal fatigue strength of austenitic stainless steels. Nuclear Engineering and Design, 184, pp. 113 – 122. Higuchi, M. & Iida, K. 1991. Fatigue strength correction factors for carbon and low-alloy steels in oxygen-containing high-temperature water. Nuclear Engineering and Design, 129, pp. 293 – 306. Hirano, A., Hayashi, M., Sagawa, W., Takehara, H., Tanaka, M. & Iikura, T. 1994. High cycle thermal fatigue crack initiation behavior of type 304 stainless steel in pure water. ASME, PVP, Vol. 287, pp. 19 – 25. Hirano, K., Kobayashi, H. & Nakazawa, H. 1979. Elastic-plastic fracture mechanics study of thermal shock cracking. ICM 3 – Mechanical Behavior of Materials, Vol. 3, Cambridge, England, 20 – 24 Aug., Pergamon Press Ltd., pp. 457 – 466. Hishida, M., Saito, M., Hasegawa, K., Enomoto, K., Matsuo, Y. 1986. Experimental study on crack growth behavior for austenitic stainless steel in high temperature pure water. Transactions of the ASME, 108, pp. 226 – 233. 97 Hoshide, T., Yamada, T., Fujimura, S. & Hayashi, T. 1985. Short crack growth and life prediction in low-cycle fatigue of smooth specimens. Engineering Fracture Mechanics, 21 (1), pp. 85 – 101. Hussain, K. 1997. Short fatigue crack behaviour and analytical models: a review. Engineering Fracture Mechanics, 58 (4), pp. 327 – 354. Hutchinson, J. W. 1968. Singular behavior at the end of a tensile crack in a hardening material. Journal of the Mechanics and Physics of Solids, 16, pp. 13 – 31. Hänninen, H. & Hakala, J. 1981. Pipe failure caused by thermal loading in BWR water conditions. International Journal of Pressure Vessels & Piping, 9, pp. 445 – 455. Hänninen, H., Törrönen, K. & Cullen, W. H. 1986. Comparison of proposed cyclic crack growth mechanisms of low alloy steels in LWR environments. Poceedings of the 2nd International Atomic Energy Agency Specialists' Meeting on Subcritical Crack Growth, NUREG/CP-0067, MEA-2090, 2, pp. 73 – 97. Igari, T., Wada, H. & Ueta, M. 2000. Mechanism-based evaluation of thermal ratchetting due to traveling temperature distribution. Transactions of the ASME, 122, pp. 130 – 138. Ilola, R. 1999. Effect of temperature on mechanical properties of austenitic high nitrogen steels. Doctor of Technology thesis. Acta Polytechnica Scandinavica, Mechanical Engineering Series, 136. 100 p. Inal, K. Gergaud, P., François, M. & Lebrun, J. L. 1999. Stress analysis in a duplex steel. Scandinavian Journal of Metallurgy, 28, pp. 139 – 150. Irwin, G. R. 1947. Fracture dynamics. In: Fracturing of Metals. A Seminar on the Fracturing of Metals. American Society for Metals, Ohio. pp. 147 – 166. Irwin, G. R. 1957. Analysis of stresses and strains near the end of a crack traversing a plate. Journal of Applied Mechanics, 24, pp. 361 – 364. Itoh, T., Chen, X., Nakagawa, T. & Sakane, M. 2000. A simple model for stable cyclic stress-strain relationship of type 304 stainless steel under nonproportional loading. Journal of Engineering Materials and Technology, 122, pp. 1 – 9. Itoh, T., Sakane, M., Ohnami, M. & Socie, D. 1995. Nonproportional low cycle fatigue criterion for type 304 stainless steel. Journal of Engineering Materials and Technology, 117, pp. 285 – 292. 98 Jenkins, C. F. & Smith, G. V. 1969. Serrated plastic flow in austenitic stainless steel. Transactions of the Metallurgical Society of AIME, 245, 2149 – 2156. Jin, Z.-H. & Noda, N. 1994. Transient thermal stress intensity factors for a crack in a semi-infinite plate of a functionally gradient material. International Journal of Solids & Structures, 3(2), pp. 203 – 218. Johansson, J., Odén, M. & Zeng, X.-H. 1999. Evolution of the residual stress state in a duplex stainless steel during loading. Acta Materialia, 47(9), pp. 2669 – 2684. Johansson, J. & Odén, M. 2000a. Load sharing between austenite and ferrite in duplex stainless steel during cyclic loading. Metallurgical and Materials Transactions, 31A (6), pp. 1557 – 1570. Johansson, J, & Oden, M. 2000b. Influence of texture and anisotropy on microstresses and flow behavior in a duplex stainless steel during loading. The 49th annual Denver X-ray Conference, 31 July- 4 August 2000, Denver, Colorado, USA. John, R., Hartman, G. A. & Gallagher, J. P. 1992. Crack growth induced by thermal-mechanical loading. Experimental Mechanics, June, pp. 102 – 108. Kamachi, K., Okada, T., Kawano, M., Namba, S. Ishida, T., Tani, N. & Kubohori, T. 1982. Thermal fatigue by impact heating and stresses of two phase stainless steel at elevated temperature. Proc. 4th Int. Conf. on Composite Materials: Progress in Science and Engineering of Composites, Tokyo, pp. 1383 – 1389. Kawagoishi, N., Chen, Q. & Nisitani, H. 2000. Significance of small crack growth law and its practical application. Metallurgical and Materials Transactions, 31A, pp. 2005 – 2013. Kawakubo, T., Hishid, M., Amano, K., Katsuta, M. 1980, Crack growth behavior of type 304 stainless steel in oxygenated 290°C pure water under low frequency cyclic loading. Corrosion-NACE, 36(11), pp. 638 – 647. Keiser, J., Taljat, B., Wang, X.-L., Maziasz, P., Hubbard, C. & Swindeman, R. 1996. Analysis of composite tube cracking in recovery boiler floors. 1996 Engineering Conference, TAPPI Proceedings, pp. 693 – 705. Keiser, J., Taljat, B., Wang, X.-L., Swindeman, R., Maziasz, P., Thomas, R. & Payzant, E. 1997. Analysis of cracking of co-extruded recovery boiler floor tubes. 1997 Engineering & Papermakers conference. TAPPI Proceedings, pp. 1025 – 1041. 99 Kemppainen, M. 1997. Termisen väsymisen testausjärjestelmän suunnittelu ja toteutus. Diplomityö. Espoo, Teknillinen korkeakoulu, Konetekniikan osasto, Materiaalitekniikan laboratorio. 110 p. In Finnish. Kerezsi, B. B., Kotousov, A. G. & Price, J. W. H. 2000. Experimental apparatus for thermal shock fatigue investigations. International Journal of Pressure Vessels and Piping, 77, pp. 425 – 434. Kliman, V. 1983. Fatigue life prediction for a material under programmable loading using the cyclic stress-strain properties. Materials Science and Engineering, 68, pp. 1 – 10. Kruml, T., Polák, J., Obrtlík, K. & Degallaix, S. 1997. Dislocation structures in the bands of localised cyclic plastic strain in austenitic 316L and austenitic-ferritic duplex stainless steels. Acta Materialia, 45 (12), pp. 5145 – 5151. Kuo, A.-Y. & John, R. 1992. Analytical and experimental treatment of a single-edge-crack plate subjected to arbitrary point heat sources. In: Fracture Mechanics: Twenty Second Symposium. Eds.: Ernst, H. A., Saxena, A. & McDowell, D. L. ASTM-STP 1131. American Society for Testing and Materials. Philadelphia, pp. 342 – 366. Laird, C. & de la Veaux, R. 1977. Additional evidence for the plastic blunting process of fatigue crack propagation. Metallurgical Transactions, 8A (4), pp. 657 – 664. Laird, C. & Smith, G. C. 1962. Crack propagation in high stress fatigue. Philosophical Magazine, 7(77), pp. 847 – 857. Lal, D. 1994. A new mechanistic approach to analyzing LEFM fatigue crack growth behaviour of metals and alloys. Engineering Fracture Mechanics, 47 (3), pp. 379 – 401. Lankford, J., Davidson, D. L. & Chan, K. S. 1984. The influence of crack tip plasticity in the growth of small fatigue cracks. Metallurgical Transactions, 15A (8), pp. 1579 – 1588. Lanteigne, J. & Baïlon, J-P. 1981. Theoretical model for FCGR near the threshold. Metallurgical Transactions, 12A (3), pp. 459 – 466. Leis, B. N. 1985. Displacement controlled fatigue crack growth in inelastic notch fields: implications for short cracks. Engineering Fracture Mechanics, 22 (2), pp. 279 – 293. Li, Y. & Laird, C. 1994. Cyclic response and dislocation structures of AISI 316L stainless steel. Part 2: polycrystals fatigued at intermediate strain amplitude. Materials Science and Engineering, A186, pp. 87 – 103. 100 Lindgren, M. & Lepistö, T. 2000a. Effect of prior plastic deformation on the relation between residual stresses and Barkhausen noise. The sixth international conference on residual stresses ICRS-6, Oxford, UK, July 10 – 12 2000. pp. 853 – 860. ISBN-I-86125-123-8 Lindgren, M. & Lepistö, T. 2000b. Application of a novel type Barkhausen noise sensor to continuous fatigue monitoring. NDT&E International, 33, pp. 423 – 428. Liu, H. W. & Kobayashi, H. 1980. Stretch zone width and striation spacings – the comparison of theories and experiments. Scripta M e t a l l u r g i c a , 1 4 , p p . 525 – 530. Lukáš , P. & Kunz, L. 1984. Threshold stress intensity and dislocation structures surrounding fatigue cracks in polycrystalline copper. Materials Science and Engineering, 62, pp. 149 – 157. Lukáš , P., Kunz, L. & Knésl, Z. 1985. Fatigue crack propagation rate and the crack tip plastic strain amplitude in polycrystalline copper. Materials Science and Engineering, 70, pp. 91 – 100. Maier, H. J. & Christ, H.-J. 1996. Modelling of cyclic stress-strain behavior under thermomechanical fatigue conditions – a new approach based upon a multi-component model. Scripta Materialia, 34 (4), pp. 609 – 615. Maier, H. J. & Christ, H.-J. 1997. Modelling of cyclic stress-strain behavior and damage mechanisms under thermomechanical fatigue conditions. International Journal of Fatigue. 19 (1), pp. S267 – S274. Mallet, O., Engler-Pinto, C. C.Jr, Rézaï-Aria, F., Skelton, R. P., Nikbin, K. & Webster, G. A. 1995. Influence of material stress-strain characteristics on thermomechanical fatigue analysis of IN100 superalloy. Materials at High Temperatures, 13 (1), pp. 47 – 54. Marsh, D. J. 1981. A thermal shock fatigue study of type 304 and 316 stainless steels. Fatigue of Engineering Materials and Structures, 4 (2), pp. 179 – 195. Marsh, D., Green, D. & Parker, R. 1986. Comparison of theoretical estimates and experimental measurements of fatigue crack growth under severe thermal shock conditions – Part I: experimental observations. Journal of Pressure Vessel Technology, 108, pp. 501 – 506. McCormick, P. G. 1972. A model for the Portevin – Le Chatelier effect in substitutional alloys. Acta Metallurgica, 20, pp. 351 – 354. 101 McDowell, D. 1987. Simple experimentally motivated cyclic plasticity model. Journal of Engineering Mechanics, 113 (3). McEvily, A. J. & Wei, R. P. 1971. Fracture mechanics and corrosion fatigue. NACE-2, Corrosion Fatigue: Chemistry, Mechanics and Microstructure, Devereux, O., McEvily, A. J. & Staehle, R. W. (Eds.) University of Connecticut, June 14 – 18, pp. 381 – 395. ISBN 915567-58-X Merola, M. 1995. Normative issue in thermal fatigue design of nuclear components. Nuclear Engineering and Design, 158, pp. 351 – 361. Miller, K. J. 1999. A historical perspective of the important parameters of metal fatigue; and problems for the next century. Proceedings of the seventh international fatigue congress, Fatigue '99, Beijing, P. R. China, June 8 – 12 1999. Beijing Institute of Aeronautical Materials, Institute of Metal Research, Fatigue Society, pp. 15 – 39. Moverare, J. J. 2001. Microstresses and anisotropic mechanical behaviour of duplex stainless steels. Linköpings Universitet, Institute of Technology, Department of Mechanical Engineering, Division of Engineering Materials. Dissertations No. 699. 135 p. ISBN 91-7373-043-2. Mughrabi, H. 1993. Cyclic plasticity and fatigue of metals. Journal de Physique IV, 3, pp. 659 – 668. Mughrabi, H. 1996. Cyclic stress-strain behaviour, microstructure and fatigue life. Fatigue '96: Proceedings of the sixth international fatigue congress, 6 – 10 May, Berlin, Germany. pp. 57 – 68. Mughrabi, H. & Christ, H.-J. 1997. Cyclic deformation and fatigue of selected ferritic and austenitic steels: specific aspects. ISIJ International, 37 (129), pp. 1154 – 1169. Murakami, Y. (ed.). 1987. Stress intensity factors handbook. Pergamon Press. ISBN 0-08-034809-2. Murakami, Y., Endo, M. 1993. Effects of defects, inclusions and inhomogeneities on fatigue strength. International Journal of Fatigue, 16 (4), pp. 163 – 182. Murakami, Y., Toriyama, T & Coudert, E. M. 1994. Instructions for a new method of inclusion rating and correlations with the fatigue limit. Journal of Testing and Evaluation, 22 (4), pp. 318 – 326. Muralidharan, U. & Manson, S. S. 1988. A modified universal slopes equation for estimation of fatigue characteristics of metals. Journal of Engineering Materials and Technology, 110, pp. 55 – 62. 102 Navarro, A. & de los Rios, E. R. 1988. Short and long fatigue crack growth: a unified model. Philosophical Magazine A, 57 (1), pp. 15 – 36. Neumann, P. 1969. Coarse slip model of fatigue. Acta Metallurgica, 17, pp. 1219 – 1225. Nilsson, J.-O. & Thorvaldsson, T. 1985. The influence of nitrogen on microstructure and strength of a high-alloy austenitic stainless steel. Scandinavian Journal of Metallurgy, 15, pp. 83 – 89. Nisitani, H., Goto, M. & Kawagoishi, N. 1992. A small-crack growth law and its related phenomena. Engineering Fracture Mechanics, 41 (4), pp. 499 – 513. Ohno, N. & Kachi, Y. 1986. A constutive model of cyclic plasticity for nonlinear hardening materials. Journal of Applied Mechanics, 53, pp. 395 – 402. Ogava, K. Ueta, M. Kanaoka, T. Kitade, S, Ueno, T. & Takashi, S. 1991 Applicability of simplified estimation method to thermal ratchetting of FBR component. Transaction of the 11th international conference on structural mechanics in reactor technology. H Shibata (ed.) Atomic Energy Society of Japan, Tokyo. Vol. E, pp. 209 – 214. Orbtlík, K., Kruml, T. & Polák, J. 1994. Dislocation structures in 316L stainless steel cycled with plastic strain amplitudes over a wide interval. Materials Science and Engineering, A187, pp. 1 – 9. Paris, P., Gomez, M. & Anderson, W. 1961. A rational analytic theory of fatigue. The Trend in Engineering, 13 (1), pp. 9 – 14. Paris, P. & Erdogan, F. 1963. A critical analysis of crack propagation laws. Transacions of the ASME. Journal of Basic Engineering, 85(4), pp. 528 – 534. Pelloux, R. M. N. 1969. Mechanisms of formation of ductile fatigue striations. Transactions of American Society for Metals, 62, pp. 281 – 285. Petersen, C. & Rubiolo, G. H. 1991. High-temperature thermal fatigue of AISI 316L steel. Journal of Nuclear Materials, 179 – 181, pp. 488 – 491. Pippan, R., Berger, M. & Stüwe, H. P. 1986. The influence of crack length on fatigue crack growth in deep sharp notches. Metallurgical Transactions, 18A (3), pp. 429 – 435. Pohl, M. & Bracke, A. 1997. The influence of changing the microstructure on the plastic deformation of duplex stainless steels after thermal cycling. Duplex Stainless Steels 97 – 5th World Conference. The Netherlands, 21 – 103 23 October. Stainless Steel World. KCI Publishing, pp. 891 – 896. ISBN 90-73168-02-3. Pohl, M. & Bracke, A. 1999. Thermisch induzierte Veränderungen der Mikrostruktur ferritisch-austenitischer Duplex-Stähle und deren Einfluß auf den Eigenspannungszustand. Z. Metallkd., 90 (8), pp. 551 – 556 Polák, J., Kruml, T., Degallaix, S., Nilsson, J.-O. 1999. Fatigue damage in austenitic-ferritic duplex stainless steel. ICM8 – Eigth International Conference on the Mechanical Behaviour of Materials, Victoria, British Columbia, Canada, 16 – 21 May, Vol. I.2, pp. 47 – 52. ISBN 1-55058-163- 5 Prime, M. 2001. Cross-sectional mapping of residual stresses by measuring the surface contour after a cut. Submitted for publication in Journal of Engineering Materials and Technology. Prime, M. B. & Gonzales, R. 2000. The contour method: simple 2-D mapping of residual stresses. The sixth international conference on residual stresses ICRS-6, Oxford, UK, July 10 – 12, 2000. pp. 617 – 624. ISBN-I-86125-123-8. Rajanna, K., Pathiraj, B. & Kolster, B.H. 1997. Duplex stainless steel fracture surface analysis using X-ray fractography. Journal of Materials Engineering and Performance, 6 (1), pp. 35 – 40. Revel, P., Kircher, D. & Bogard, V. 2000. Experimental and numerical simulation of a stainless steel coating subjected to thermal fatigue. Materials Science and Engineering, A290, pp. 25 – 32. Rice, J. R. 1967. Mechanics of crack tip deformation and extension by fatigue. In: Fatigue Crack Propagation, ASTM-STP 415, American Society for Testing and Materials, pp. 247 – 311. Rice, J. R. 1968. A path independent integral and the approximate analysis of strain concentration by notches and cracks. Journal of Applied Mechanics, 35, pp. 379 – 386. Rice, J. R. & Rosengren, G. F. 1968. Plane strain deformation near a crack tip in a power-law hardening material. Journal of the Mechanics and Physics of Solids, 16, pp. 1 – 12. Saarinen, P., Nenonen, P. & Hänninen, H. 2000. Thermal fatigue of new composite tube materials. Proceedings of TAPPI 2000 Engineering Conference, Atlanta Hilton & Towers, Atlanta, Sept. 17 – 21.(CD-ROM) Sadananda, K. & Vasudevan, A. K. 1995. Analysis of fatigue crack closure and thresholds. In: Fracture Mechanics: 25th Volume, ASTM-STP 1220. 104 Ed.: Erdogan, F. American Society for Testing and Materials, pp. 484 – 501. Sadananda, K. & Vasudevan, A. K. 1997a. Analysis of high temperature fatigue crack growth behavior. International Journal of Fatigue, 19 (1), pp. S183 – S189. Sadananda, K. & Vasudevan, A. K. 1997b. Short crack growth behavior. In: Fatigue and Fracture Mechanics: 27th volume. Eds.: Piascik, R. S., Newman, J. C. & Dowling, N. E. ASTM-STP 1296. American Society for Testing and Materials, pp. 301 – 316. Shi, H. J. & Pluvinge, G. 1994. Cyclic stress-strain response during isothermal and thermomechanical fatigue. Fatigue, 16, pp. 549 – 557. Shi, H. J., Wang, Z., G. & Su, H. H. 1996. Thermomechanical fatigue of a 316L austenitic steel at two different temperature intervals. Scripta Materialia, 35 (9), pp. 1107 – 1113. Siegmund, T., Werner, E. & Fischer, F. D. 1993. The irreversible deformation of a duplex stainless steel under thermal cycling. Materials Science and Engineering, A169, pp. 125 – 134. Siegmund, T., Werner, E. & Fischer, F. D. 1995. On the thermomechanical deformation behavior of duplex-type materials. Journal of the Mechanics and Physics of Solids, 43 (4), pp. 495 – 532. Silberschmidt, V. V., Werner, E. 1999. Analysis of thermal residual stresses in duplex-type materials. Computational Materials Science, 16, pp. 39 – 52. Skelton, R. P. 1982. Growth of short cracks during high strain fatigue and thermal cycling. In: Low Cycle Fatigue and Life Prediction. ASTM-STP 770. Philadelphia, pp. 337 – 381. Skelton, R. P. 1993. Cyclic hardening, softening, and crack growth during high temperature fatigue. Materials Science and Technology, 9, pp. 1001 – 1008. Skelton, R. P., Vilhelmsen, T. & Webster, G. A. 1998. Energy criteria and cumulative damage during fatigue crack growth. International Journal of Fatigue, 20 (9), p. 641. Skelton, R. P., Webster, G. A. 1996. History effects on the cyclic stress- strain response of a polycrystalline and single crystal nickel-base superalloy. Materials Science and Engineering, A216, pp. 139 – 154. Skelton, R. P., Rees, C. J. & Webster, G. A. 1996. Energy damage summation methods for crack initiation and growth during block loading 105 in high temperature low-cycle fatigue. Fatigue and Fracture of Engineering Materials and Structures, 19 (23), pp. 287 – 297. Socie, D. 1987. Multiaxial fatigue damage models. Journal of Engineering Materials and Technology, 109, pp. 293 – 298. Socie, D. & Bannantine, J. 1988. Bulk deformation fatigue damage models. Materials Science and Engineering, A103, pp. 3 – 13. Socie, D. F. & Marquis, G. B. 2000. Multiaxial fatigue. Society of Automotive Engineers. Warrendale, U.S.A. 484 p. ISBN 0-7680-0453-5 Solomon, H. D. 1972. Low cycle fatigue crack propagation in 1081 steel. Journal of Materials, JMLSA, 7 (3), pp. 299 – 306 Spindler, M. W. 1998. Thermal mechanical cyclic stress-strain testing of a type 316 steel. In: Allison, A. (ed.). Experimental Mechanics: Advances in Design, Testing and Analysis. Balkema. Rotterdam. ISBN:90-5809-016-7, pp. 1127 – 1132. Srinivasan, V. S., Sandhya, R., Valsan, M., Bhanu Sankara Rao, K., Mannan, S. L. & Sastry, D. H. 1997. The influence of dynamic strain ageing on stress response and strain-life relationship in low cycle fatigue of 316L(N) stainless steel. Scripta Materialia, 37 (10), pp. 1593 – 1598. Starkey, M. S. & Skelton, R. P. 1982. Strain intensity and cyclic J approaches to crack growth. Fatigue of Engineering Materials and Structures, 5 (4), pp. 329 – 341. Stefanita, C.-G., Atherton, D. L. & Clapham, L. 2000. Plastic versus elastic deformation effects on magnetic Barkhausen noise in steel. Acta Materialia, 48, pp. 3545 – 3551. Stevens, K. J. 1999. Fatigue performance and microanalysis of heat treated 2205 duplex stainless steel. Materials Science and Technology, 15, pp. 903 – 908. Suresh, S. 1992. Fatigue of materials. Cambridge Solid State Science Series. Cambridge University Press. 617 p. ISBN 0-521-437-63-6. Taira, S. 1973. Relationship between thermal fatigue and low-cycle fatigue at elevated temperature. In: Fatigue at Elevated Temperatures, ASTM- STP 520, American Society for Testing and Materials. pp. 80-101. Taira, S. & Fujino, M. 1979. A damage analysis in high temperature thermal fatigue. Transactions of the Iron and Steel Institute of Japan, 19 (3), pp. 185 – 190 106 Taira, S., Maruyama, S. & Fujino, M. 1976. Effect of temperature on the rate of fatigue crack propagation in steels during low cycle fatigue. Transactions of the Iron and Steel Institute of Japan, 16 (3), pp. 146 – 152 Tanaka, K., Masuda, C. & Nishijima, S. 1981. The generalized relationship between the parameters C and m of Paris' law for fatigue crack growth. Scripta Metallurgica, 15, pp. 259 – 264. Tanaka, K. 1989. Mechanics and micromechanics of fatigue crack propagation. In: Fracture Mechanics: Perspectives and Directions, Twentieth Symposium. Eds.: Wei, R. P. & Gangloff, R. P. ASTM-STP 1020. American Society for Testing and Mateirals. Philadelphia, pp. 151 – 183. Tanigawa, Y. & Komatsubara, Y. 1997. Thermal stress analysis of a rectangular plate and its thermal stress intensity factor for compressive stress field. Journal of Thermal Stresses, 20, pp. 517 – 542. Tidball, R. A. & Shrut, M. M. 1954. Thermal-shocking austenitic stainless steel with molten metals. Transactions of the ASME, 76, pp. 639 – 643. Tomkins, B. 1968. Fatigue crack propagation – an analysis. Philosophical Magazine, 18, pp. 1041 – 1066. Tomkins, B. 1980. Micromechanisms of fatigue crack growth at high stress. Metal Science, 14 (8-9), pp. 408 – 417. Tomkins, B. 1981. High strain fatigue. In: Subcritical Crack Growth due to Fatigue, Corrosion and Creep, Ispra, Italy, 19 – 23 Oct., Elsevier Science Publishing Co., Inc., pp. 239 – 266. Tomkins, B, 1983a. Fatigue: mechanisms. In: Creep and Fatigue in High- Temperature Alloys, Applied Science Publishers, Ltd., pp. 111 – 143. Tomkins, B. 1983b. Life prediction at elevated temperature. Journal of Pressure Vessel Technology, 105, pp. 269 – 272. Tomkins, B. 1984. Crack growth – present status, future direction. Fatigue Crack Growth: 30 Years of Progress, Cambridge, UK, 20 Sept, pp. 133 – 146. Uchida, Y., Shimojo, M. & Higo, Y. 1999. Relationship between fatigue striation height and stress ratio. Journal of Materials Science, 24, pp. 2411 – 2419. Vasudevan, A. K., Sadananda, K. & Louat, N. 1992. Reconsideration of fatigue crack closure. Scripta Metallurgica et Materialia, 27, pp. 1673 – 1678. 107 Vasudevan, A. K., Sadananda, K. & Louat, N. 1993. Two critical stress intensities for threshold fatigue crack propagation. Scripta Metallurgica et Materialia, 28, pp. 65 – 70. Vasudevan, A. K. & Sadananda, K. 1993. Fatigue crack growth in metal matrix composites. Scripta Metallurgica et Materialia, 28, pp. 837 – 842. Vasudevan, A. K. & Sadananda, K. 1995a Classification of fatigue crack growth behavior. Metallurgical and Materials Transactions, 26A (5), pp. 1221 – 1234. Vasudevan, A. K. & Sadananda, K. 1995b Fatigue crack growth behavior of composites. Metallurgical and Materials Transactions, 26A (12), pp. 3199 – 3210. Vehoff, H. & Neumann, P. 1978. In situ SEM experiments concerning the mechanism of ductile crack growth. Acta Metallurgica, 27, pp. 915 – 920. Wadsworth, N. J., Hutchings, J. 1958. The effect of atmospheric corrosion on metal fatigue. Philosophical Magazine, 3, pp. 1154 – 1166. Wang, X.-L., Hoffman, C. M., Hsueh, C. H., Sarma, G. & Hubbard, C. R. 1999. Influence of residual stress on thermal expansion behavior. Applied Physics Letters, 75(21), pp. 3294 – 3296. Wareing, J., Tomkins, B. & Sumner, G. 1973. Extent to which material properties control fatigue failure at elevated temperatures. In: Fatigue at Elevated Temperatures, ASTM-STP 520, American Society for Testing and Materials, pp. 123 – 137 Wegst, C. W. 1995. Stahlschlüssel. Verlag Stahlschlüssel Wegst GmbH. Germany. Westergaard, H. M. 1939. Bearing pressure and cracks. Journal of Applied Mechanics, 6, pp. 49 – 53. Williams, D. R., Davidson, D. L. & Lankford, J. 1980. Fatigue-crack-tip plastic strains by the stereoimaging technique. Experimental Mechanics, 22, pp. 134 – 139. Wilson, W. K. & Yu, I.-W. 1979. The use of the J-integral in thermal stress crack problems. International Journal of Fracture, 15 (4), pp. 377 – 387. Xia, Y. & Wang, Z. 1992. Low cycle fatigue behaviour of a new type of stainless steel. Materials Science and Engineering, A151, pp. 29 – 35. Yoshimoto, T., Ishihara, S., Goshima, T., McEvily, A. J. & Ishizaki, T. 1999. An improved method for the determination of the maximum 108 thermal stress induced during a quench test. Scripta Materialia, 41 (5), pp. 553 – 559. You, B.-R. & Lee, S.-B. 1996. A critical review on multiaxial fatigue assessment of metals. International Journal of Fatigue, 18(4), pp. 235 – 244. Zamrik, S. Y. 1990. An interpretation of axial creep-fatigue damage interaction in type 316 stainless steel. Transactions of the ASME, 2, pp. 4 – 19. Zauter, R., Christ, H. J., Mughrabi, H. & Petry, F. 1993. Thermomechanical fatigue of the austenitic stainless steel AISI 304L. Thermomechanical Fatigue Behavior of Materials, San Diego, USA, 14-16 Oct. American Society for Testing and Materials. ASTM-STP 1186. pp. 70-90. Zauter, R., Christ, H. J. & Mughrabi, H. 1994. Some aspects of thermomechanical fatigue of AISI 304 L stainless steel: Part II. Dislocation arrangements. Metallurgical and Materials Transactions, 25A (2), pp. 401 – 406. Zhang, P. & Burger, C. P. 1986. Transient thermal stress-intensity factors for short edge cracks with equal depth of crack tips. Engineering Fracture Mechanics, 24 (4), pp. 589 – 599. 109 APPENDIX 1. TWO-BAR MODEL FOR THERMAL STRESSES The thermal strains and stresses caused by rapid temperature changes and internal constraint can be very complicated. A simple two-bar model can be used to illustrate qualitatively the effect of a thermal cycle. Consider a situation where two bars are constrained so, that their strains are forced equal (Figure A1.1). The thinner Bar 1 simulates the sample surface, whereas the thicker Bar 2 simulates the bulk of the material. Both bars are from the same linear elastic – ideally plastic material. Figure A1.1 Two bars are forced to equal strain. Figure A1.2 The temperature in Bar 1 simulates the quick transient of the specimen surface, whereas the temperature curve of Bar 2 simulates the temperature deeper in the specimen. Consider now the system going through a thermal cycle each bar following the corresponding temperature curve in Figure A1.2. The temperature curve of Bar 1 simulates the rapid transients experienced by the surface, whereas the curve of Bar 2 simulates the slower heating and cooling, that takes place deeper in the material. A qualitative stress-strain plot can now be drawn for the two bars (Figure A1.3). 110 Figure A1.3. Schematic stress-strain plot for Bar 1 during thermal cycle of the system. See text for details. The development of the stress-strain plot can be divided into 6 distinct phases, as follows: 0) The heating starts. The temperature of Bar 1 increases faster than temperature of Bar 2. Consequently the thermal expansion causes compressive stress to develop into Bar 1 and a balancing tensile stress into Bar 2. Because of the greater thickness of Bar 2 the stresses in Bar 2 are much smaller than those in Bar 1 1) The yield point is exceeded in Bar 1. Bar 1 starts to deform plastically, whereas Bar 2 continues to deform elastically. The uneven plastic deformation induces a mismatch between the two bars. 2) The cooling starts. Rapid cooling changes the stress in Bar 1 from compressive to highly tensile. 3) Yielding starts in tension in Bar 2. Again the uneven plastic deformation induces a mismatch between the two bars. 4) The temperature difference between the bars begins to decrease again as Bar 2 slowly cools down. 5) Finally, the temperature in Bar 1 and 2 is equal again. Due to the yielding in tension in phase 3 in Bar 1 there is a compressive residual stress present in Bar 1 and a balancing tensile residual stress in Bar 2. 111 APPENDIX 2. DERIVATION OF EQUATION 41 Equation 41 was used to calculate the applied strains from the measured temperature and measured axial strain data (paragraph 3.3). The equation can be derived as follows: in an unconstrained case an increase in temperature gives rise to thermal expansion proportional to the temperature change: , (A2.1) where εT is free, uniaxial thermal expansion, α is the linear coefficient of thermal expansion and ∆T is the temperature change. In the test sample the thermal expansion is partially constrained by the uneven temperature distribution. The uniaxial constrained thermal expansion is given by the difference between measured and calculated (from Equation A2.1) thermal expansion: , (A2.2) where εT is calculated thermal expansion, εm is measured expansion and εσ is the constrained thermal expansion. Due to the Poisson effect an applied strain causes a corresponding strain in perpendicular direction: ε νεy x= − (A2.3) where εy and εx are strains in perpendicular directions and ν is the Poisson's ratio. Because the thermal expansion in the test specimen is constrained in the axial and tangential directions, Equation A2.2 must be corrected to take into account the Poisson effect: ε νε ε εσ σa t m T− = − (A2.4) where εσa is the constrained thermal expansion in axial direction and εσt is the constrained thermal expansion in tangential direction. On the surface, the radial stresses are zero and axial and tangential strains are equal. Hence we get Equation 41: ε ε ε νσ = − − m T ( )1 (A2.5) 112 APPENDIX 3. USED FE-MODELS The test sample was modeled with an axisymmetric, infinite height model (Figure A3.1). The specimen radius was divided to 100 elements. The model height was adjusted to give rectangular elements for best accuracy. Figure A3.1. The axisymmetric element mesh used to model the test sample. The symmetry axis equals the y-axis in figure. The bottom nodes are fixed and the top nodes are forced to have an equal displacement, i.e. the model represents an object of infinite height. In elastic-plastic modeling a multilinear isotropic-kinematic material model was used. The measured cyclic stress-stain curves (Paragraph 4.1) were inserted to the model as a temperature-dependent multilinear approximations. The material model then applied kinematic hardening rule to the material, while limiting the kinematic hardening based on the given multilinear curves to produce more stable behavior of the numerical model (Ansys Release 5.5.1, 1998). In the contour method, the residual stresses are solved using a linear- elastic, axisymmetric FE-model describing the cut half of the specimen. The used element mesh is presented in Figure A3.2. An example of the deformed shape and the stress solution is presented in Figure A3.3. 113 Figure A3.2. The axisymmetric element mesh used in the contour method. The symmetry axis equals the y-axis in the figure. Figure A3.3. An example of the deformed shape and the stress solution for the contour method. D e f o r m e d s h a p e i s exaggerated for better illustration. 114 APPENDIX 4. STRESS SOLUTIONS ∆ε distributions for all the used cycles were calculated with linear-elastic FE-method. The results for each material are presented in Figures A4.1 – A4.9. 0.0000 0.0005 0.0010 0.0015 0.0020 0.0025 0.0030 0 2 4 6 8 10 r (mm) ε ( m m /m m ) Figure A4.1. Axial (sol id l ine) and tangential (dotted line) ∆ε solutions for material A (20 – 300 °C). 0.0000 0.0005 0.0010 0.0015 0.0020 0.0025 0.0030 0 2 4 6 8 10 r (mm) ε ( m m /m m ) Figure A4.2. Axial (sol id l ine) and tangential (dotted line) ∆ε solutions for material B (20 – 300 °C). 0.0000 0.0005 0.0010 0.0015 0.0020 0.0025 0 2 4 6 8 10 r (mm) ε ( m m /m m ) Figure A4.3. Axial (sol id l ine) and tangential (dotted line) ∆ε solutions for material C (20 – 300 °C). 0.0000 0.0010 0.0020 0.0030 0.0040 0.0050 0.0060 0.0070 0.0080 0 2 4 6 8 10 r (mm) ε ( m m /m m ) Figure A4.4. Axial (sol id l ine) and tangential (dotted line) ∆ε solutions for material D (20 – 600 °C). 115 0.0000 0.0010 0.0020 0.0030 0.0040 0.0050 0.0060 0.0070 0.0080 0 2 4 6 8 10 r (mm) ε ( m m /m m ) Figure A4.5. Axial (solid line) and tangential (dotted line) ∆ε solutions for material E (20 – 600 °C). 0.0000 0.0010 0.0020 0.0030 0.0040 0.0050 0.0060 0.0070 0 2 4 6 8 10 r (mm) ε ( m m /m m ) Figure A4.6. Axial (solid line) and tangential (dotted line) ∆ε solutions for material F (20 – 600 °C). 0.0000 0.0010 0.0020 0.0030 0.0040 0.0050 0.0060 0.0070 0 2 4 6 8 10 r (mm) ε ( m m /m m ) Figure A4.7. Axial (solid line) and tangential (dotted line) ∆ε solutions for material G (20 – 600 °C). 0.0000 0.0005 0.0010 0.0015 0.0020 0.0025 0.0030 0.0035 0 2 4 6 8 10 r (mm) ε ( m m /m m ) Figure A4.8. Axial (solid line) and tangential (dotted line) ∆ε solutions for material H (20 – 280 °C). 0.0000 0.0005 0.0010 0.0015 0.0020 0.0025 0.0030 0 2 4 6 8 10 r (mm) ε ( m m /m m ) Figure A4.9. Axial (solid line) and tangential (dotted line) ∆ε solutions for material I (20 – 280 °C).