Multiple Sensors on Pulsed Eddy-Current Detection for 3-D ...

90

IEEE SENSORS JOURNAL, VOL. 5, NO. 1, FEBRUARY 2005

Multiple Sensors on Pulsed Eddy-Current Detection for 3-D Subsurface Crack Assessment Gui Yun Tian, Ali Sophian, David Taylor, and John Rudlin

Abstract—This paper proposes the use of multiple sensors in pulsed eddy-current detection for three-dimensional (3-D) subsurface flaw imaging. A normalization technique has been proposed to eliminate the characteristic variation among the Hall devices used in the probe and lift off effects. A principal component analysis-based feature extraction that provides orthogonal information for multiple sensor fusion has been introduced and investigated. Using the features of multiple projection coefficients, 3-D surface flaws can be measured and reconstructed. The experimental tests have illustrated that the proposed method has delivered more defect information than the conventional peak value and time for pulsed eddy-current sensors. Index Terms—Feature extraction, feature fusion, multiple sensors, normalization, principal component analysis (PCA), pulsed eddy-current sensors.

I. INTRODUCTION

E

DDY-CURRENT sensors typically have characteristics of robustness and small size [3]. Noncontact eddy-current sensors can be used in precision engineering for displacement and geometric measurement by using a novel sensor design and blind signal separation [1]–[6]. By applying advanced signal processing, the surface measuring resolution can be as small as 0.1 m [1], [4]. Eddy-current sensors have also been widely used for nondestructive evaluation (NDE) [7], in particular, surface crack detection. However, the measurement accuracy and resolution is much less than the surface geometric measurement because of the unknown features of the measurement [6]. Conventional eddy-current techniques use single-frequency sinusoidal excitation and detect flaws as impedance or voltage changes on an impedance plane display with inspectors interpreting the magnitude and phase changes. However, these techniques are sensitive to a variety of parameters that are inherent in the flaws [2]. To obtain the depth information of defects, multiple-frequency measurements have been combined to provide a more rigorous assessment of structural integrity by reducing signal anomalies that may otherwise mask the flaws [8]. Pulsed eddy-current (PEC) sensing is a new and emerging technique [9] that has been particularly developed and devised for subsurface flaw measurements. Some success has been reManuscript received March 28, 2002; revised July 1, 2004. This work was supported in part by the EPSRC research council, in part by the University of Huddersfield, and in part by the TWI, Ltd. The associate editor coordinating the review of this paper and approving it for publication was Dr. Mona Zaghloul. G. Y. Tian, A. Sophian, and D. Taylor are with The University of Huddersfield, Huddersfield HD1 3DH, U.K. (e-mail: [email protected]; [email protected]; [email protected]). J. Rudlin is with the Structural Integrity Department, TWI, Granta Park, Cambridge CB1 6AL, U.K. (e-mail: [email protected]). Digital Object Identifier 10.1109/JSEN.2004.839129

Fig. 1.

Layout of PEC probe.

ported by Iowa State University [10], [11], DERA [12], and the Royal Military College of Canada [13]. These techniques, particularly in the detection of corrosion, can work at some distance below the surface (up to 100 mm in aluminum). PEC techniques excite the probe’s excitation coil with a repetitive broadband pulse, usually a rectangular wave. The resulting transient current through the coil induces transient eddy currents in the test piece, which are associated with highly attenuated magnetic pulses propagating through the material. The probe provides a series of voltage-time data pairs as the induced field decays, and since the produced pulses consist of a broad frequency spectrum, the reflected signal contains important depth information. Physically, the field is broadened and delayed as it travels deeper into the highly dispersive material, and flaws or other anomalies close to the surface affect the eddy-current response earlier than deeper flaws. Peak values and peak times have been used for flaw detection and identification. However, these systems cannot currently be readily used because of difficulties in calibration and the lack of suitable response signal processing algorithms [14]. Typically, a single-sensor PEC probe consists of an excitation coil and a pickup coil or a magnetic field sensor, as shown in Fig. 1 [15]. A diagram block for a full PEC system developed by the authors in [16] is shown in Fig. 2. Briefly, the system works as follows; the waveform generator produces a rectangular waveform with variable frequency and duty cycle. The waveform is fed to a coil driver circuit, which excites the induction coil in the probe with pulsed current. The pickup sensor will measure the vertical resultant magnetic field, which is the sum of the one generated by the excitation coil and the opposing one generated by the induced eddy current in the sample. A voltage amplifier with variable gain then amplifies the signal so that the dynamic input range of the data acquisition card [or analog-to-digital (A/D) converter card] is used effectively. The A/D card will convert the input signal into digital data ready to be processed by software in the PC. The software performs communication with the data acquisition card, the control of the data transfer from DAQ card buffer to the PC RAM, signal preprocessing, feature extraction, defect categorization, and the presentation of the results on the PC monitor. To obtain more information about defect such as location and three-dimensional

1530-437X/$20.00 © 2005 IEEE

TIAN et al.: MULTIPLE SENSORS ON PULSED EDDY-CURRENT DETECTION

Fig. 2.

91

Block diagram of the PEC system.

(3-D) defects, the paper extends the work to multiple sensors for PEC systems. With the increase of computing power, advanced signal processing has been used for computer science and engineering, pattern recognition, and artificial intelligence, particularly [16]. In this paper, advanced signal processing has been applied for a novel multisensor system to extract robust features, which are extended from our other work in computer vision, for crack detection and identification. A feature fusion based on the multiple features from the sensors is introduced for 3-D assessment. The rest of the paper is organized as follows. Section II will discuss signal normalization and feature extraction for pulsed eddy-current sensors by using principal component analysis (PCA); Section III will introduce multiple sensor design and their signal fusion for 3-D flaw assessment; Sections IV and V will present the experimental tests and derived conclusions.

II. NORMALIZED PRINCIPAL COMPONENT ANALYSIS FOR FEATURE EXTRACTION In NDE, it is desired to be able to detect, classify, and quantify defects that may occur in metal structures, such as aircraft lap joints. In such structures, defects mainly take place as surface cracks, subsurface cracks, and hidden corrosion. To answer this requirement, we have developed a PEC system based on Hall-effect devices and a new feature extraction technique. PEC is believed to be able to give more information than sinusoidal eddy currents as they are rich in frequency contents. They are also most likely faster than multifrequency eddy currents as PEC exercises a string of different frequencies of field simultaneously [7]. As opposed to conventional eddy currents, PEC data analysis is mainly carried out in the time domain. The base response signal where the probe situated on a defect-free area is taken as the reference. The defect free area could be a standard sample or pretested sample, which was inspected by radiography nondestructive testing (NDT) instruments. A differential signal is obtained by subtracting the testing sample response signal from

the reference signal. Then the differential signal is analyzed for gaining information about the testing conditions. Fig. 3(a) shows an example of a response using our pulsed eddy-current sensor C4 as structured in Fig. 1, where the internal diameter and external diameters of the excitation coil of C4 are 18 and 23 mm. Features commonly used are the magnitude and the arrival time of the positive going peak of the differential signal [14]. The disadvantage of the technique is that the peak value and the arrival time are prone to noise. Besides, two parameters are not good enough to identify the locations and the characters of flaws [15]–[18]. Also, the same flaw responses under different magnetic sensors such as Hall devices have different peak values and peak times, as illustrated in Fig. 3(b). The different sensors C3 and C4 have quite different responses, such as peak values and peak times for testing the same Aluminum samples, where C3 and C4 are similar Hall devices with different coil diameters. The sensors’ layout and the samples will be illustrated in Sections III and IV. The internal and external diameters of the excitation coil C3 are 25 and 30 mm, respectively. The variation in responses due to different devices and manufacturing errors will be reduced by normalization for effective multiple sensor integration. Fig. 3(c) illustrates examples of normalized signals, where their peak values are normalized. In order to efficiently describe the clusters of response signals and identify the defects and their sizes, we have to choose the set of directions in the signalspace along which the variance of the cluster is maximum. This is achieved through the standard procedure of PCA, or the Karhunen–Loeve transform [19], [20]. Transforming coordinates amounts to projection onto new coordinates and expressing a response signal as a linear combination of base signals. The notion of direction of variance in a high-dimensional space can be extracted from the covariance matrix of the data points. The eigenvectors of the covariance matrix define a space in which the covariance among dimensions is zero, so the matrix takes on a diagonal form. Thus, essentially, the PCA is solved by computing the eigenvectors of the covariance matrix and projecting the data points into this new space. We would like to apply the projection coefficients as the new features for flaw detection and identification.

92

IEEE SENSORS JOURNAL, VOL. 5, NO. 1, FEBRUARY 2005

Fig. 3.

(a) Differential response signal. (b) Response variation with different coils. (c) Examples of normalized signals.

Consider a response signal among a collection of signals, which can be obtained from various samples, define as the average signal

(1) Every signal differs from the mean by a vector The covariance matrix of the data is, thus, defined as

.

Fig. 4.

(2) is a column-wise concatenation of all the s. note that has dimension from the sample data in each response signal. We note that if are the eigenvec, then tors of where

where are the eigenvalues, then are the eigenvectors or as we see by multiplying on the left eigensignal of by , the previous equation (4) So, defining

(3)

Multiple sensor layout (top view).

, the eigenvectors of

, we have (5)

TIAN et al.: MULTIPLE SENSORS ON PULSED EDDY-CURRENT DETECTION

93

Fig. 5. Specimen 1: thickness variation.

From here on, we assume that the order of is such that the eigenvalues are decreasing (these values are the variances along the new coordinate space). We decide to project a response dimensions by computing signal onto only

image dependency on lighting geometry and illumination color, as a preprocess for the computation of eigensignal [21]. The preprocess will normalize the peak value to one unit by

(6)

(7)

where and . is the th coordinate of in the new system. The eigensignals that retain the most significant amount information are those that correspond to the largest eigenvalues. To carry out the analysis, each data set that is made of elements from each sample testing condition or any object to recognize/classify is lined into a column in a table, after the mean of each data set is subtracted from each data set. If there sample-testing conditions, then the table will consist of are columns. Then, the covariance matrix is worked out from the table. The covariance matrix will have dimension of , which is a square matrix. From this covariance matrix, are deeigenvalues and their corresponding eigenvectors rived. When an eigenvector is multiplied by a data set, a value will be produced that represent the data when it is mapped into the axis corresponding to the related eigenvalue. These values are the new features that can be used for classification and recognition purposes, and, in our case, they might correlate with quantities to be measured. Only two or three eigenvectors usually chosen based on the values of their eigenvalues. The ones with highest values are selected. The initial estimates of the subspaces are determined by PCA. This gives eigensignals and eigenvalues, the variance on each eigensignal. Each signal is then coded across the whole set of subspaces, giving a set of weights for that signal. Since the different subspaces overlap, these weights are not unique. Thus, we further require the power consumption of each signal’s weights be minimized, using ratio of weight to eigenvalue. This reduces the weights given on low-variance eigensignals, which code similar variations. If these projections on the subspaces generate new subspaces, they will be progressively more orthogonal and functional. Our approach has two processes of creating eigensignals and extracting key parameters by mapping a response signal to the eigensignal space. As indicated in Fig. 3(c), the raw response signals show large variation in the peak time of the signals. We apply the color normalization algorithm, a comprehensive image normalization which removes

III. MULTIPLE PULSED EDDY-CURRENT SENSOR AND FEATURE FUSION To be able to process 3-D flaw assessment and improve the detecting resolution and overcome the problem of lift off, we have designed and developed a multiple eddy-current sensing system. In the new sensor, an excitation coil is used to generate a transient magnetic field and three Hall devices are incorporated to detect the distribution of eddy-current magnetic field. The sensor layout is illustrated in Fig. 4, where sensors 1, 2, and 3 are Hall devices. The spatial resolution of Hall devices can be reduced down to 0.5 mm. The diameter of Hall devices used is 2 mm in our design. It is expected to reconstruct the 3-D flaw from the three magnetic sensors by their correlation and correspondence, which will depend on the feature extraction. It is our intention to find the depth information in subsurface from the above feature extraction. Then, we can apply orthogonal information to obtain the other spatial information. The eigenvectors are orthogonal each other. Interestingly, the coordinates, which the signal maps on three eigensignals from above normalized PCA have reflected the flaw coordinate in subsurface. More details can be found in Section IV. IV. EXPERIMENT TESTS AND DISCUSSION For testing purposes, we have made two aluminum samples. One is for thickness variation testing, and can be used for corrosion or metal loss simulation, and the other for surface and subsurface defects detection and quantification. The samples can be seen in Figs. 5 and 6. For subsurface crack simulation, we probe the sample in Fig. 6 over its top side, and for surface crack, we do it over the bottom side. In our investigation, we have proposed two different approaches that differ in the type of data sets to be used. We use time domain data of the differential response signals and

94

IEEE SENSORS JOURNAL, VOL. 5, NO. 1, FEBRUARY 2005

Fig. 6. Specimen 2: slot quantification.

Fig. 7. First two eigensignals (eigenvectors) for pulsed eddy-current sensor.

Fig. 8. Comparison of flaw identification. (a) Using peak values/times. (b) Uing the proposed PCA approach.

frequency domain data by wavelet analysis for feature extraction and 3-D flaw assessment. Both of these data processing approaches have similar results. Therefore, we only present the time domain data analysis in the paper. The differential signals from different testing conditions were arranged into a matrix , where and are the length of with dimension of each signal and the number of signals, respectively. were investigated and it was The resulting eigenvectors found that the first eigenvector seemed to try to amplify the peak value differences. However, as there is variation in time scale, it can only manage to exploit the values at points in the proximity of the highest peak. The resolution given by the scores using

the first eigenvector is slightly higher than what the peak values give. The second eigenvector yields a good discrimination of surface slots from other conditions. By applying all the response signals, the eigensignals can be computed by using (1)–(5) and (7). Fig. 7 displays the first two eigensignals, where the horizontal axis is the time in and the vertical axis is the amplitude in millivolts. Then, we apply the eigensignals to identify various flaws by using (6). Fig. 8 illustrates the comparison of flaw identification using classical peak values/peak time and our normalized PCA, where two eigensignals have been used. As indicated from Fig. 8, normalized PCA for feature extraction has improved the

TIAN et al.: MULTIPLE SENSORS ON PULSED EDDY-CURRENT DETECTION

95

Fig. 9. Three-dimensional defect distribution obtained using time domain PCA.

flaw classification significantly. The test results based on normalized PCA are robust to different sensors and lift off, which is important for commercial applications. By applying three eigenvectors (eigensignals), we can map the sensor response signal to the three principal component vectors as coordinates. Based on the three coordinates, a 3-D distribution of flaws of our test samples can be visualised and grouped reflects the in Fig. 9. As indicated in the color plot in Fig. 9, depth information of the flaw in subsurface. Then, we can apply the eigenvector orthogonal property, their correlation of project coefficients, and the magnetic sensor geometric configuration to derive the flaw’s location and shape. Based on the normalized PCA mapping, the 3-D assessment can be obtained by simple geometric computation in the multiple sensor system. Our test results have reflected the above analysis.

microsensors for integrated diagnostics. Our work offers tremendous opportunities to extend the functionality of a single sensor by incorporating spatial information gained from arrays of these sensors, e.g., to localize and separate multiple signal sources impinging on the arrays. Our approach has illustrated its advantage in comparison with physical models and FE computing [22], [23], which are time consuming and sensitive to the target surface geometry. Integrating the signal processing with the sensing can have substantial advantages especially when the size of the sensors becomes smaller, e.g., using the MEMS technique. The 3-D assessment resolution can be improved significantly and the measurement efficiency can be advanced [24]–[26].

REFERENCES V. CONCLUSION A PCA-based normalized feature extraction algorithm has been developed and tested. The new feature extraction is robust to magnetic sensing devices’ variation and lift off effects. The approach provides a flexible framework to extract more features by using more eigensignal or eigenvectors. Both the time-domain and frequency domain-based PCA produce potential features that give a good discrimination of surface defects and reflect the spatial information, particularly the depth information in subsurface. Multiple sensors can be applied to construct 3-D cracks below the surface. It provides an effective approach on location and sizing of the cracks without using scanning. Further investigation on the accuracy of the 3-D flaw assessment will be studied in the future. More features can be obtained by using more eigensignals. Although normalized PCA in the frequency domain has similar results as the normalized PCA in the time domain, the frequency domain also has good repeatability that is better than the peak height and occurrence time performance. This shows that it is less sensitive to noise in the signals. More investigation on frequency domain-based PCA will be required. The normalized PCA-based feature extraction and 3-D flaw identification are important technique for eddy-current

[1] G. Y. Tian, “Frequency output eddy current sensors for precision engineering,” Insight, vol. 43, no. 5, pp. 315–318, May 2001. [2] G. Y. Tian, Z. X. Zhao, and R. W. Baines, “Computational algorithms for linear variable differential transformers (LVDTs),” Proc. Inst. Elect. Eng., vol. 144, no. 4, pp. 189–193, Apr. 1997. [3] , “The research of inhomogeniety in eddy-current sensors,” Sens. Actuators A, vol. 58, pp. 153–156, 1998. [4] , “A miniaturised displacement sensor for deep hole measurement,” J. Precision Eng. Amer. Soc., vol. 23, pp. 236–242, 1999. [5] , “Precision measurement using an eddy-current sensor device,” in Proc. 12th Nat. Conf. Manufacturing Research, Bath, U.K., 1996, pp. 88–94. [6] G. Y. Tian, Z. X. Zhao, R. W. Baines, and P. Corcoran, “Blind sensing,” Inst. Elect. Eng. Manufact. Eng., vol. 76, no. 4, pp. 179–183, Apr. 1997. [7] A. Sophian, G. Y. Tian, D. Taylor, and J. Rudlin, “Electromagnetic and eddy-current NDT: a review,” Insight, vol. 43, no. 5, pp. 308–313, May 2001. [8] T. Clauzon, F. Thollon, and A. Nicolas, “Flaws characterization with pulsed eddy currents,” IEEE Trans. Magn., pt. 1, vol. 35, no. 3, pp. 1873–1876, May 1999. [9] B. C. Yoseph, “Emerging NDE technologies and challenges at the beginning of the 3rd millennium—Part I, part II,” NDT.net, vol. 5, no. 1, Jan. 2000. [10] [Online] Available: http://www.cnde.iastate.edu/cnde.html [11] J. Bowler, “Pulsed eddy-current inversion for the determination of crack shape,” Electromagn. Nondestruct. Eval., pp. 263–269, 1997. [12] D. J. Harrison, “Eddy-current inspection using Hall sensors and transient excitation,” DRA, Famborough, U.K., Defence Research Agency, Tech. Rep. DRA/SMC/TR941 008, 1994.

96

[13] B. A. Lepine, J. S. R. Giguere, D. S. Forsyth, A. Chahbaz, and J. M. S. Dubois, “Interpretation of pulsed eddy-current signals for locating and quantifying metal loss in thin skin lap splices,” Rev. Quant. Nondestruct. Eval., vol. 21, pp. 415–422, 2002. [14] J. R. Bowler and D. J. Harrison, “Measurement and calculation of transient eddy-currents in layered structures,” Rev. Progr. Quant. Nondestruct. Eval., pt. 1, vol. 11, pp. 241–248, 1996. [15] A. Sophian, G. Y. Tian, D. Taylor, and J. Rudlin, “Feature extraction techniques for pulsed eddy-current NDT,” NDT & E Int., vol. 36, no. 1, pp. 37–41, Jan. 2003. [16] J. R. Rudlin, “A beginner’s guide to eddy-current testing,” Brit. J. NDT, vol. 31, no. 6, pp. 314–320, Jun. 1989. [17] R. A. Smith and G. R. Hugo, “Transient eddy-current NDE for ageing aircraft—capabilities and limitations,” Insight, vol. 43, no. 1, Jan. 2001. [18] B. Lebrun, Y. Jayet, and J. C. Baboux, “Pulsed eddy-current signal analysis: application to the experimental detection and characterization of deep flaws in highly conductive materials,” NDT&E Int., vol. 30, no. 3, pp. 163–170, 1997. [19] V. N. Reddy, W. M. Miller, and M. L. Mavrovouniotis, “Inverse-signal analysis with PCA,” Chemometr. Intell. Lab. Syst., vol. 36, pp. 17–30, 1997. [20] A. M. Martinez and A. C. Kak, “PCA versus LDA,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 23, no. 2, pp. 228–233, Feb. 2000. [21] G. D. Finlayson and G. Y. Tian, “Color normalization for color object recognition,” Int. J. Pattern Recognit. Artif. Intell., vol. 13, no. 8, pp. 1271–1285, 1999. [22] K. Miya, M. Uesaka, and Y. Yoshia, “Applied electromagnetics research and application,” Progr. Nuclear Energy, vol. 32, no. 1/2, pp. 174–194, 1998. [23] U. Patel and D. Rodger, “Finite element modeling of pulsed eddy currents for nondestructive testing,” IEEE Trans. Magn., vol. 32, no. 3, pp. 1593–1596, May 1996. [24] G. Y. Tian, “Design and implementation of distributed measurement systems by fieldbus-based intelligent sensors,” IEEE Trans. Instrum. Meas., vol. 50, no. 5, pp. 1197–1202, Oct. 2001. [25] J. Xue, S. Ramalingam, and Z. Shi, “Eddy-current flaw imaging using micro-sensor arrays,” in Proc. Int. Mechanical Engineering Congr., Dallas, TX, Nov. 1997. [26] G. Y. Tian, A. Sophian, D. Taylor, and J. R. Rudlin, “Pulsed eddy-current system for dynamic inspection of defects,” Insight, vol. 46, no. 5, pp. 256–260, May 2004.

Gui Yun Tian received the B.Sc. degree in metrology and instrumentation and the M.Sc. degree in precision engineering from the University of Sichuan, Chengdu, China, in 1985 and 1988, respectively. He is currently a Senior Lecturer at the School of Computing and Engineering, University of Huddersfield, Huddersfield, U.K., and a Visiting Professor at the University of Sichuan. He maintains diverse and active research in the area of sensors, intelligent instrumentation, nondestructive testing, digital signal processing, computer vision, and microelectromechanical systems (MEMS), which have been funded by the EPSRC, the Royal Society, the Royal Academy of Engineering, and industry. He has published over 100 books and papers in English and Chinese in the above areas. He is a regular reviewer for international journals and conferences.

IEEE SENSORS JOURNAL, VOL. 5, NO. 1, FEBRUARY 2005

Ali Sophian was born in 1974 in Indonesia. He received the B.Eng. (Hons.) and the Ph.D. degrees in electronics engineering from the University of Huddersfield, Huddersfield, U.K., in 1998 and 2004, respectively. He is currently a Research Fellow at the School of Computing and Engineering, University of Huddersfield. His research interests include eddy-current NDT system design and signal processing.

David Taylor works extensively in both mixed analog/digital ASIC testing and the design and synthesis of high-performance DSP and error-correction systems. In particular, he has been responsible for developing the transient response testing technique for analog components in mixed-signal systems. Currently, he is the head of the department of Multimedia and Information Systems, University of Huddersfield, Huddersfield, U.K.

John Rudlin has worked in nondestructive testing research for over 25 years. This has included employment at Rolls-Royce and Associates, UKAEA Risley, John Laing R&D, Hocking NDT, and the University College London NDE Centre, London, U.K. He joined TWI, Cambridge, U.K., in 1999 as a Principal Consultant in NDT. His recent research interests include inspection reliability of high-speed ultrasonic techniques for pipeline and riser inspection, robotic inspection of aircraft, corrosion inspection methods and strategies, and electromagnetic and thermography techniques.