SIGNAL PROCESSING ADVANCEMENTS FOR CASS UT EXAMINATIONS T. Seuaciuc-Osório, M. Dennis, D. MacDonald, EPRI, USA D. Braconnier, D.B. Ltd., Japan INTRODUCTION Regulatory requirements mandate that the welds of reactor coolant loop (RCL) piping in Westinghouse PWRs be examined to identify any potential defects throughout the component volume. Volumetric examination of cast stainless steel RCL piping offers several material and physical challenges to conventional nondestructive evaluation (NDE) methods. In addition, utilities have made license renewal commitments to examine these cast stainless steel components; however, a reliable NDE solution does not currently exist. Over the past two decades, several studies have been completed by various organizations in an effort to find a reliable volumetric examination solution for cast stainless steel piping. The anisotropic coarse grain structure of cast stainless steel results in significant changes in the acoustic material velocity, which has a detrimental impact on the ultrasonic responses. In the past, the use of standard ultrasonic techniques resulted in data images with very low signal-to-noise ratios, making flaw detection very challenging. Because of advancements in computer technology, it is now possible to create more complex signal processing routines to improve the ultrasonic data images. This paper presents some automated conventional ultrasonic data collected from a representative cast stainless steel pipe. These data were refined using signal processing routines including cross correlation, deconvolution by intercorrelation, L0- and L1-norm deconvolution methods, and decimation by threshold on average along the index axis. The processed data images using each of these routines are discussed. EXPERIMENTAL SETUP AND RESULTS In an effort to improve the ultrasonic responses obtained from the examination of cast austenitic stainless steel (CASS) materials, a pressurizer surge line mockup with three thermal fatigue cracks was inspected. Details of the specimen and experimental equipment and procedure utilized follow. Specimen The CASS pressurizer surge line courtesy of Pacific Northwest National Laboratories (PNNL) used in this study can be seen in Figure 1. This surge line was obtained from a cancelled power plant and used to build mockup 9C 002, which contains three base metal, circumferentially orientated thermal fatigue cracks manufactured by Truflaw. Only two of the flaws are considered in this investigation; the information regarding the dimensions of each analyzed thermal fatigue crack (099BAB1102 and 122BAB1110) can also be found in Table 1. Equipment and Procedure The Zetec DYNARAY ultrasonic phased array system with an automated pipe scanner was used to drive a 1.5 MHz, 45, 15 25 mm Applus RTD dual-conventional longitudinal wave transducer (see Figure 2). The ultrasonic data were collected using a raster scan approach with the ultrasonic beam oriented in the scan direction along the axis of the pipe (perpendicular to the thermal fatigue cracks). After completing one scan line, the transducer is indexed along the pipe circumference prior to executing the next scan line. At each probe location, the amplitude of the ultrasonic wave versus the time-of-flight (TOF) is collected 467
and produces an A-scan. These A-scans are used to generate Top (Scan-Index), Side/B-scan (Scan- Depth/TOF), and End/D-scan (Index-Depth/TOF). Figure 1: Pressurizer Surge Line Mockup 9C 002 utilized in this study, Table 1: Dimensions and Position of both analyzed thermal fatigue flaws in mockup 9C 002. Flaw 099BAB1102 122BAB1110 Length 0.76 in. (19.3 mm) 1.00 in. (25.3 mm) Height 0.13 in. (3.4 mm) 0.24 in. (6.0 mm) Position in depth 1.31 in. (33.3 mm) 1.31 in. (33.3 mm) Position in depth (µs with 45 ) 16.3 µs 16.3 µs Position along scan 4 in. (101.6 mm) 4 in. (101.6 mm) Position along index angle -45 40 Position along the circumference -127 mm 113 mm Position along the acquisition index (points) 20 ±5 118 ±5 Results Figure 3 through Figure 6 provide screen captures of angle-corrected data images for the two thermal fatigue cracks considered in this study with the ultrasonic beam oriented toward the pipe (termed looking up, LKUP) and toward the elbow (termed looking down, LKDN). As illustrated in these data images, the signal-to-noise ratio of each flaw depends on the scan direction, with the results with the probe oriented toward the elbow (LKDN) being the most challenging. 468
Figure 2: Zetec DYNARAY Ultrasonic Phased Array System (left) and 1.5 MHz 45 Dual Conventional Applus RTD Transducer (right). SIGNAL PROCESSING ALGORITHMS Several signal processing methods were executed on the raw data during this study: cross correlation, deconvolution by intercorrelation, L0- and L1-norm deconvolution, and decimation by threshold on average along the index axis; a brief description of each method follows. A combination of L0- or L1- norm deconvolution and decimation by threshold on average along the index axis was the method with the best performance; hence, these techniques are discussed in greater detail and are accompanied by sample processed B-scan images. Furthermore, greater emphasis will be given to data obtained with the transducer oriented toward the elbow (LKDN), as this is identified as the challenging direction. Cross Correlation This method involves simple cross correlation between the transducer wavelet (probe input signal) and the data files (A-scans). This intercorrelation represents the best conventional filtering that can be expected, and results in a slight improvement from the raw data. However, this technique is inefficient when the signal has a considerable amount of structural noise (scattering noise from the grain structure), as is the case with CASS, and the results show little differentiation between flawed and unflawed regions. Deconvolution by Intercorrelation This method involves performing conventional de-convolution of the data files by the ultrasonic transducer wavelet. The results thus obtained are better when compared to simple cross correlation for much of the structural noise and virtually all the floor noise can be removed from the B-scan images, while the flaw response is enhanced. However, the structural noise is still amplified in some regions of the B-scans, making it hard to differentiate between flawed and flawless regions in the processed B-scans. 469
Figure 3: Unprocessed UltraVision Data Images of Flaw 099BAB1102 with Transducer Oriented Toward the Pipe (LKUP). Figure 4: Unprocessed UltraVision Data Images of Flaw 099BAB1102 with Transducer Oriented Toward the Elbow (LKDN). 470
Figure 5: Unprocessed UltraVision Data Images of Flaw 122BAB1110 with Transducer Oriented Toward the Pipe (LKUP). Figure 6: Unprocessed UltraVision Data Images of Flaw 122BAB1110 with Transducer Oriented Toward the Elbow (LKDN). 471
L0/L1-Norm Deconvolution The L0/L1-norm deconvolution method offers two advantages: echoes resulting from reflectors in close proximity to each other can be resolved, and the detectability and noise filtering levels can be adjusted with a set of parameters. The convolution model [1, 2] is where y is the response signal (A-scan), h is the wavelet (transducer input signal), x* represent the pure echo signals (desired answer) and ε is an error term. The star (*) represents the convolution operator. The deconvolution method then is a constrained least-squares approach to minimize the error (noise) between the measured response and the pure echo signals according to the convolution model above, that is, subject to where ϕ is a norm and λ is the regularization parameter. A norm is defined to apply a sparsity constraint on the solution: there exist many signals which, when convoluted with the transducer signal h, will yield responses close to the measured A-scans y, but which are very different from the true solution. This sparsity constraint on x leads to more stable solutions which are not so sensitive to the noise embedded in the signals. Different norms can be defined; in the present context, two norms are used, L0 and L1. L0-norm In essence, this norm uses the regularization parameter λ as a limit to the number of spikes allowed in the signal x. It is defined as [1, 2] This is a suboptimal method as they only explore a subset of all solutions, and are not guaranteed to find the best solution [1]. L1-norm Instead of using the regularization parameter λ to limit the number of spikes in the solution x, the L1-norm limits the sum of its amplitudes. It is simply defined as [1] The L1-norm is more computationally expensive than the L0-norm, and does not always yield considerably better results. This method yields the most promising results, as it makes it easier to differentiate between flawed and unflawed areas. Decimation by Threshold on Average along the Index Axis This process involves averaging n adjacent B-scans (that is, along the index axis of the C-scan) and applying a threshold on the result. Since the structure noise has a low consistency in comparison to a flaw response, the probability to obtain a signal at the same scan position and time-of-flight in adjacent B-scans is much lower for structural noise than for flaw responses. As a result, much of the signal noise is removed. The challenge in this method is to determine a suitable selection of the two parameters, namely the averaging distance and the threshold value, which effectively rejects only structural noise while maintaining and emphasizing any relevant response signals. The challenge in choosing the detection threshold in such noisy data was observed in the results: while most of the noise was eliminated, the flaw response was not emphasized and could be mistaken for structural noise. Again, there was little difference in the processed B-scans with and without flaws, making data analysis challenging. 472
Combination of L0/L1-Norm Deconvolution and Decimation by Threshold on Average along the Index Axis Among the tried algorithms, the best observed performance was achieved by the L0/L1-Norm Deconvolution method. However, the signal processing can be further enhanced by combining this method with the Decimation by Threshold on Average along the Index Axis method. Since much of the noise is removed from the raw signal by the deconvolution, procedure, the choice of a proper threshold value is not as challenging as when the Decimation by Threshold method is applied by itself. The results of applying the Decimation by Threshold on Average along the Index Axis method on the data processed by the L0/L1-Norm Deconvolution method can be seen in Figure 7 for a flawed B-scan, and in Figure 8 for a flawless B-scan. It is clear from these images that this combined method yields remarkable signal processing results: the flaws are greatly emphasized while the structural noise is eliminated. Figure 7: Raw (Left) and Processed (Right) B-scans Containing Flaw 122BAB1110 Using a Combination of L1-Norm Deconvolution and Decimation by Threshold on Average along the Index Axis Methods (LKDN) Figure 8: Raw (Left) and Processed (Right) B-scans Containing No Flaw Using a Combination of L1-Norm Deconvolution and Decimation by Threshold on Average along the Index Axis Methods (LKDN) 473
EFFECT OF PARAMETERS In the previous section, a combination of L0/L1-Norm Deconvolution and Decimation by Threshold on Average along the Index Axis was shown to yield the best performance among the attempted signal processing routines. This combined method, however, requires the initially arbitrary selection of some parameters, such as the regularization parameter (along with the choice of norm), the size of the average window and the threshold value. The choice of these parameters can greatly affect the performance of the signal processing routine, and a study needs to be performed to determine their effect and how to make a suitable choice for the data set being analyzed. As a first step in such a study, two different set of parameters were tested on the data collected on mockup 9C 002, and the results obtained with each set is shown here. The parameter values for each set is listed in Table 2, while the resulting top (C-scan) and end (D-scan) views are shown in Figure 9 and Figure 10, respectively. It is seen that Parameter Set II yields better results, as it is capable of eliminating more of the structural noise while maintaining good emphasis on the flaws. The difference in the sets which leads to this discrimination is the higher threshold, indicating that the threshold values in Parameter Set I are too low. Further analysis is required to determine whether or not the threshold values in Parameter Set II can be increased without loss of significant indications. There are other parameters that can be adjusted, such as the size of the average window (the number of points averaged along the index axis) and the maximum number of peaks of the regularization parameter, and further parameter studies need to be performed. Ultimately, the goal is to remove the arbitrary nature of these parameters, perhaps by relating them to physical parameters of the equipment used. One example of such relation would be determining the size of the average window based on the width of the probe and the index resolution of the scan. In addition, more data sets need to be evaluated with this signal processing routine to help determine ideal parameters which are not adequate only for a specific data set, but which can be applied, with the adequate adjustments, to any data set. As an example, it may be that one single threshold level cannot be efficiently determined for all data sets, but it needs to be adjusted according to the observed noise level in each specific data set. Once more, further studies are necessary to remove the arbitrary nature of these parameters. Table 2: Parameter Sets Utilized Parameter Set I Set II Deconvolution Norm L0 Regularization Parameter 5000 1000 Number of Maximum Peaks 5 5 Number of Point Averaged Along Index Axis 4 4 Threshold on point at index -2 4% 5% Threshold on point at index -1 8% 10% Threshold on point at index +1 8% 10% Threshold on point at index +2 4% 5% 474
Figure 9: Top View (C-Scan) Obtained with Combined Method Using Parameter Set I (Left) and Parameter Set II (Right) Figure 10: End View (D-Scan) Obtained with Combined Method Using Parameter Set I (Left) and Parameter Set II (Right) CONCLUSIONS Currently a reliable volumetric inspection approach for cast stainless steel piping does not exist due to the anisotropic coarse grain structure of this material which typically results in poor signal-to-noise flaw responses. Advancements in computing power have enabled the use of sophisticated signal processing algorithms which may be able to refine the raw ultrasonic data to provide enhanced images for analysis. Several signal processing routines, namely cross correlation, deconvolution by intercorrelation, L0/L1-norm deconvolution and decimation by threshold on average along the index axis, have been applied in an attempt to enhance data quality resulting from the ultrasonic inspection of CASS. The routine which yielded the best results is a two-step combination of two of these methods: first, a L0/L1- norm deconvolution is applied on the raw, original data; then, the decimation by threshold on average along the index axis method is applied on the deconvoluted data. While the results obtained with this combined process are remarkable and encouraging, a study needs to be performed to fully determine the effect of the few arbitrary parameters that it requires, as well as to provide guidelines on how to appropriately select them; ideally, such a study would link these parameters to physical properties of the inspection, such as the width of the probe, inspection index 475
resolution, base noise level, etc. In addition, more data files need to be evaluated to determine the limitations of this approach. Lastly, a performance measure needs to be defined in order to quantitatively and objectively compare the results of different parameter sets on different data sets. REFERENCES 1) Charles Soussen, Jérôme Idier, Ewen Carcreff, Laurent Simon, Catherine Potel. "Ultrasonic nondestructive testing based on sparse deconvolution". Journal of Physics : Conference series, 2011, vol. 353 012018. 2) Charles Soussen, Jérôme Idier, David Brie, Junbo Duan. "From Bernoulli Gaussian Deconvolution to Sparse Signal Restoration". IEEE Transactions on Signal Processing, October 2011, 59(10) 4572-4584. 476