PIERS ONLINE, VOL. 3, NO. 4, 27 57 Advances in Signal Processing to Reduce Lift-off Noise in Eddy Current Tests M. Cacciola, A. Gasparics 2, F. C. Morabito, M. Versaci, and V. Barrile Universitá Mediterranea degli Studi di Reggio Calabria, Reggio Calabria, Italy 2 Research Institute for Technical Physics and Materials Science, Budapest, Hungary Abstract Nowadays, Non Destructive Testing and Evaluation (NDT/NDE) are frequently used to evaluate integrity of manufactured articles in civil or industrial applications. In this framework, Eddy Current Test (ECT) technique has a primary role for inspection of conducting materials. Solution of many related inverse problems requires an accurate comprehension of measured data: it is one of the most challenging problem in this context. That s why a suitable signal pre-processing is necessary. In this paper, advanced signal processing techniques are evaluated in order to reduce the impact of lift-off effect on the eddy current data. Some hybrid approaches are also depicted, with encouraging results. DOI:.2529/PIERS6725. INTRODUCTION Improving manufacturing quality and ensuring public safety are two of the most important targets in various applicative fields. Thus, components and structures are commonly inspected for early detection of defects or faults which may reduce their structural integrity. In experimental NDT/NDE, the available measurement data are explored, so that some clues may emerge in the inspection signal that are possibly representative of structural modification of the specimen, like cracks, flaws and corrosion. Practically solving of these inverse problems needs a suitable signal processing procedure, in order to mitigate the difficulties introduced by presence of noise into the measured data. In fact, all the measured experimental data is a composition of the informative pattern modified by non-informative variations, can be considered as the error or noise. The different noise sources can be classified as systematic (error due to the some know effects, like the lift-off effect) and non-systematic (uncorrelated or content independent sources, like the not-position but the time dependent noise during the data-acquisition period). In order to obtain suitable input data for the inverse problem solution, it is necessary to emphasize the informative pattern in the data set by filtering the noise. However, it is also critical to keep the information unchanged or unperturbed by the applied noise filtering techniques. Otherwise, the inverse problem solution will produce fraudulent result. Basically, the enhancement of the noise filtering can be characterized by a single parameter: the signal-to-noise ratio (SNR). However, it can reliable only if it is possible to assume that the informative content of the data set is untouched. In ECTs, one of the noise source is the so-called lift-off effect. The lift-off is defined as the distance between the specimen and the probe used for the inspection. During the scanning over the specimen surface, sensor lift-off is not constant. So, it non-linearly influences the measured data by introducing a typical low frequency and variable spatial distribution noise. Scientific literature proposes a lot of signal processing approaches exploited to reduce lift-off effect on ECTs (e.g., see [ 3] and references within); the main idea proposed in this work consists on an advanced signal processing approach based on the joint use of Independent Component Analysis (ICA) and Wavelet filtering techniques. In this way, it has been possible to improve the informative content of eddy current data, emphasizing useful signal and noteworthily increasing the SNR of measurements. 2. EXPERIMENTAL SET-UP AND COLLECTED DATA In ECT, the probe allows for complex, time-varying voltage measurements and the presence of a flaw affects the formation of eddy currents. In our experimentations, probe is a single pancake exciting coil with Fluxset sensor R [4], provided by the Research Institute for Technical Physics and Materials Science (Fig. ). Inspected specimen was a square plate 8 8.25 cm made of INCONEL6 (σ = 5, µ = µ ) material. It has a rectangular thin crack (9 mm in length and.2 mm in width) in the central region; depth of crack is 2% of the plate thickness. Defect is
PIERS ONLINE, VOL. 3, NO. 4, 27 58.645.64.635.63.625-2 -5 - -5 5 5 2 Figure : Layout of the Fluxset probe (left). V out measured at Y =, 2 X 2 in the ID analysis (right): it is possible to denote the lift-off noise into the signal trend. superficial (inner defect, ID) or hidden (outer defect, OD) according to the inspection side. The specimen was inspected by both ID and OD analysis, using a.7 ma (rms) exciting current at frequency 2 KHz. The amount of exciting coil turns is 93. The lift-off of the sensor core is about.75 mm. The measured quantity is the magnitude of output pick-up voltage ( V out ). The offset signal was balanced out at the position: X = 2 mm, Y = mm, i.e., the origin of used reference system. Scanned area was a 4 4 mm region, referred to the plate center, with.5 mm spacing along x and y direction ( 2 mm X 2 mm, 2 mm Y 2 mm) [5]. A strong discontinuity in the homogeneity of the V out profile in a spatial location should evidence the presence of a defect in that zone of the volume. But, due to lift-off effect, irregularities caused by defects could not be clearly observable (Fig. ). Therefore, a suitable signal pre-processing is necessary in order to remark the effective variation on V out caused by the defect. Figure 2: V out measured on the whole scanned area: ID analysis (at left), OD analysis (at right). Due to relatively small depth of defect, it is not easily visible by an OD analysis. 3. WT-BASED APPROACH FOR THE CASE-OF-STUDY Wavelet approach measuring average fluctuations at different scales might prove less sensitive to noise than Fourier Transform. A suitable bank of discrete Wavelet filters can split low- and highfrequency contribute for a signal by a multi-resolution analysis in a Wavelet domain. In this way, global dynamics of signal f(x) can be condensed in a series of Wavelet coefficients related to the low frequencies, i. e., the approximation coefficients at the higher multi-resolution level, A M. On the other hand, local oscillations of f(x) are depicted in a set of Wavelet details coefficients at different scales (D j, j =, 2,..., M), related to the high frequencies. In this way f(x) = A M (x) + M j= D j(x). This wavelet decomposition enables the noise in the signal to be separated from the useful components. For the considered case-of-study, th-order Daubechies Wavelet [6] (Db) has been used, decomposing the signal into five frequency levels by the Discrete WT (DWT).
PIERS ONLINE, VOL. 3, NO. 4, 27 59 The high regularity of Db assures to restrict the lift-off noise into A 5. Thus, usage of the Inverse DWT (IDWT) excluding A 5 can reduce the lift-off effect on eddy current signal []: Fig. 3 shows both the WT-based denoising for Y =, 2 mm X 2 mm of ID scanning, and the iteration of the whole procedure for the ID and OD scannings. Due to small visibility of crack into OD data, also the related WT reconstruction is not recommendable for an useful experimentation and will be excluded in the prosecution of the work..645.64.635.63.625 Observed signal Wavelet low pass filtering signal x -4 3 2 2 5 5 5 5 2 x 4 D D2 D3 2 2 5 5 5 5 2 x 4 2 5 5 5 5 2 D4 D5 [mv] - -2-3 -4-5 -6-2 -5 - -5 5 5 2 - Figure 3: At top: Db decomposition of V out for the ID analysis (left); V out reconstruction without considering A 5 coefficients (right). At bottom: results of WT-based denoising on V out for ID (left) and OD (right) analysis. 4. ICA: THEORETICAL FUNDAMENTALS AND EXPLOITATIONS FOR THE CASE-OF-STUDY Let us consider a signal s(t) in time-domain, which is registered by j receivers and is the result of mixing of j sources, such that s(t) = {s (t), s 2 (t),..., s j (t)}, where s k (t) = j h= a khx h (t). Under particular assumption, it is possible to recover the set of j sources by calculating a suitable mixing matrix A, i. e., the matrix with elements a kh. It is the aim of ICA, where each mixture s k as well as each Independent Component (IC) x h is a random variable. Mixtures must have zero mean and unitary variation (whitening process). Once A is calculated, it is possible to obtain its inverse A and retrieve ICs having non-gaussian distributions [7]. In the case-of-study, signal s is considered as the mixing of V out for both ID and OD analysis. The FastICA software based on fixed-point algorithm [8] has been used, by exploiting the function g = tanh(a u); a =.7 to estimate the differential entropy (i. e., negentropy). Results of ICA-based pre-processing is shown in Fig. 4: it is possible to discriminate the component related to lift-off noise to the one representing the useful denoised V out signal. Compared with WT-based approach, the ICA method shows less micro-oscillations but higher lift-off at the top and bottom side of inspected area.
PIERS ONLINE, VOL. 3, NO. 4, 27 52 Figure 4: ICs of measured data: it is possible to denote how the component at right represents V out denoised by ICA. Figure 5: ICs of measured data: it is possible to denote how the component at left represents V out denoised by proposed WAVICA approach. 5. PROPOSED HYBRID WAVICA AND ICAWAV APPROACHES The WT- and the ICA-based approaches have been subsequently mixed, in order to obtain an hybrid approach able to improve performances of exploited techniques avoiding the disadvantages of the two methods. In a first attempt, a 5th-level Db based Wavelet analysis has been initially taken into account to improve the SNR of both ID and OD ECTs; subsequently such denoised signals have been mixed in order to extract the lift-off component by using the ICA: it is the WAVICA approach. Results are shown in Fig. 5. In this way, it has been possible to improve the quality of ECT s measurement by reducing the amount of whole noise. Some peaks on V out are still present at the bottom-left corner of inspected area. Therefore, a further approach has been considered. In this case, the ICA has been firstly carried out on the mixture of ID and OD ECTs, and subsequently the useful component has been processed by means of Db WT at five multi-resolution scales (ICAWAV approach). So, it has been possible to obtain the best results (see Fig. 6), emphasizing the influence of defect in output pick-up voltage and strongly lowering the lift-off noise even at the borders (and corners) of examined area. Table : SNR obtained by different exploited pre-processing techniques. SNR ID SNR W T SNR ICA SNR W AV ICA SNR ICAW AV 4.623 9.29 6.954 9.27 9.92
PIERS ONLINE, VOL. 3, NO. 4, 27 52 Figure 6: Results of ICAWAV-based denoising on V out for the ID analysis. The OD analysis has not been considered due to the previously presented results of the WT-based approach. 6. DISCUSSIONS ABOUT RESULTS AND CONCLUSIONS In this paper, some advanced techniques have been evaluated to reduce the lift-off noise in eddy current NDT/NDE. Besides WT- and ICA-based procedures, hybrid WAVICA and ICAWAV approaches have been proposed, obtaining a noteworthily decrement of lift-off effect. In terms of SNR, the gain retrieved by proposed hybrid approaches is between 4.54 db and 4.569 db, if compared with SNR of the original ID signal (SNR ID ). Moreover, ICAWAV technique assures a raising on SNR of.63 db and 2.238 db compared to WT and ICA respectively. Table resumes all obtained SNR values. ACKNOWLEDGMENT Authors are very pleased to thank all of the participants of the MANODET project, for providing on Web the dataset and information about Fluxset #6/6 used in this work [5]. REFERENCES. Chen, G., A. Yamaguchi, and K. Miya, A novel signal processing technique for eddy-current testing of steam generator tubes, IEEE Trans. on Magnetics, Vol. 34, No. 3, 642 648, 998. 2. Morabito, F. C., Independent component analysis and feature extraction techniques for NDT/NDE data, Materials Evaluation, Vol. 58, No., 85 92, 2. 3. Fiori, S. and P. Burrascano, ECT-data fusion by the independent component analysis for non-destructuve evaluation of metallic slabs, Third International Conference on Independent Component Analysis and Signal Separation, 323 327, San Diego, California, 2. 4. Gasparics, A., C. S. Daroczi, G. Vertesy, and J. Pavo, Improvement of ECT probes based on Fluxset-type magnetic field sensor, Electromagnetic Nondestructive Evaluation, No. 2, 46 5, 998. 5. Gasparics, A., Probe 98/6 Surface scan results within MANODET Project, available at http://alag3.mfa.kfki.hu/ect/manodet/results/fs6b/index.htm, October 3, 26. 6. Daubechies, I., Ten lectures on wavelets, CBMS-NSF Series in Applied Mathematics, 992. 7. Hyvarinen, A. and E. Oja, Independent component analysis: Algorithms and applications, Neural Networks, Vol. 3, No. 4 5, 4 43, 2. 8. Hyvarinen, A., Fast and robust fixed-point algorithms for independent component analysis, IEEE Trans. on Neural Networks, Vol., No. 3, 626 634, 999.