You are on page 1of 8

J. Shanghai Jiaotong Univ. (Sci.

), 2015
DOI: 10.1007/s12204-015-1691-y

Pulsed Eddy Current Signal Denoising Based on


Singular Value Decomposition

ZHU Hong-yun (), WANG Chang-long∗ (), CHEN Hai-long (), WANG Jian-bin ()
(Department of Unmanned Aerial Vehicles Engineering, Ordnance Engineering College, Shijiazhuang 050003, China)

© Shanghai Jiaotong University and Springer-Verlag Berlin Heidelberg 2015

Abstract: The noise as an undesired phenomenon often appears in the pulsed eddy current testing (PECT)
signal, and it is difficult to recognize the character of the testing signal. One of the most common noises presented
in the PECT signal is the Gaussian noise, since it is caused by the testing environment. A new denoising approach
based on singular value decomposition (SVD) is proposed in this paper to reduce the Gaussian noise of PECT
signal. The approach first discusses the relationship between signal to noise ratio (SNR) and negentropy of PECT
signal. Then the Hankel matrix of PECT signal is constructed for noise reduction, and the matrix is divided into
noise subspace and signal subspace by a singular valve threshold. Based on the theory of negentropy, the optimal
matrix dimension and threshold are chosen to improve the performance of denoising. The denoised signal Hankel
matrix is reconstructed by the singular values of signal subspace, and the denoised signal is finally extracted from
this matrix. Experiment is performed to verify the feasibility of the proposed approach, and the results indicate
that the proposed approach can reduce the Gaussian noise of PECT signal more effectively compared with other
existing approaches.
Key words: pulsed eddy current testing (PECT), singular value decomposition (SVD), negentropy, denoising
CLC number: TN 911.4 Document code: A

0 Introduction filtering in double logarithmic domain, which could in-


crease the SNR of PECT signal effectively. Unfortu-
Pulsed eddy current testing (PECT) technology has nately, there is little literature about noise reduction
been proved to be an effective nondestructive method, of PECT signal. Though there are many methods for
which can be used in conductive material defect detec- noise reduction, it should be discussed whether they are
tion, thickness measurement and so on[1-4] . The PECT suitable for PECT signal denoising.
is excited by a pulse or a square-wave rather than a Recently, singular value decomposition (SVD) has
sinusoidal waveform. The broad band of pulse eddy been widely used in noise reduction[12-14] . It can di-
current makes the response signal contain more infor- vide the original signal matrix into noise subspace and
mation about the defect and thickness[5-6] . Besides, signal subspace by a singular valve threshold. The de-
the PECT has shown some advantages, including more noised signal can be extracted by reconstructing the
extended detection depth, higher robustness and con- matrix with the singular values of signal subspace. One
venient operation[7-10] . of the most common noise presented in the PECT sig-
However, the noise as an undesired phenomenon of- nal is the Gaussian noise, which is caused by the test-
ten occurs in the PECT signal because of the interfer- ing environment[15] . To reduce the Gaussian noise, we
ence of environment electromagnetism and other fac- present an optimal denoising method based on SVD
tors, and the useful signal character is covered by the in this paper. First, the Hankel matrix of the PECT
noise when the signal to noise ratio (SNR) is very signal is constructed. The relationship between SNR
low[11] . Therefore, noise reduction in the pre-processing and negentropy of the PECT signal is studied to de-
stage is necessary in PECT signal processing. In order termine the optimal dimension of Hankel matrix and
to reduce the noise, Huang et al.[11] presented a de- the singular value threshold. Then, the noise subspace
noising method based on signal averaging and median and signal subspace are obtained based on the optimal
Hankel matrix and threshold, the denoised matrix is re-
Received date: 2014-09-01
constructed with the singular values of signal subspace,
Foundation item: the Twelve-Five Pre-Research Project and the denoised PECT signal can be extracted from
(No. 51325010602) the reconstructed matrix. Experimental results show
∗E-mail: wang-oec@126.com that the negentropy can be treated as a criterion of
J. Shanghai Jiaotong Univ. (Sci.), 2015

the noise suppression performance evaluating, and the where ΣX and Σh are the singular values of signal sub-
proposed approach is able to achieve a better perfor- space and noise subspace, respectively.
mance in reducing the Gaussian noise of PECT signal The Hankel matrix of noise-free signal and noise can
compared with other denoising methods. be deduced from Eqs. (5) and (7):

1 Noise Reduction Based on SVD AX = UX ΣX VXT , (8)


Ah = Uh Σh VhT . (9)
The noisy PECT signal can be mathematically rep-
resented as We can know that the Σ is divided into two parts
y(n) = X(n) + h(n), (1) by the threshold for singular value; the singular values
where, y(n), X(n) and h(n) denote the noisy PECT of noise subspace which are lower than the threshold
signal, noise-free signal and noise, respectively; n is the should be set to be zero, and the denoised matrix can
sample index. For a discrete PECT signal y(n), n = be reconstructed with the singular values of signal sub-
1, 2, · · · , N , the Hankel matrix can be constructed by space. So it is very important to select the threshold.
the following signal: If the threshold is too low, there will be a residual noise
⎡ ⎤ in the denoised signal, or when threshold is larger, the
y(1) y(2) ··· y(k) denoised signal will distort seriously.
⎢ ⎥ Generally, there are break points in the curve of sin-
⎢y(2) y(3) · · · y(k + 1)⎥
⎢ ⎥ gular values of Σ, and the break point where the curve
A=⎢ . .. .. .. ⎥, (2)
⎢ .. . ⎥ slope changes drastically can be treated as the thresh-
⎣ . . ⎦
old point[16] . Then the singular values of noise subspace
y(l) y(l + 1) · · · y(N )
which are lower than the threshold are set to be zero
where 1 < k < N , A ∈ Rl×k , l + k = N + 1. for space division[17] , and the denoised signal can be ex-
Generally, for any matrix A ∈ Rl×k , the singular tracted from the denoised matrix which is reconstructed
value decomposition of A is of the form: with the singular values of signal subspace.

A = U ΣV T , (3) 2 The Optimal Denoising Method


Based on SVD
where U ∈ Rl×m and V ∈ Rk×m are orthogonal ma-
trices, and Σ is an m × m diagonal matrix of singular It is known that the denoising effect will be influenced
values. The diagonal matrix Σ can be expressed as by the dimension of Hankel matrix, normally, the num-
 ber of rows and columns of Hankel matrix is equal or
P 0 close, but this situation may not be appropriate for the
Σ= , (4)
0 0 PECT signal noise reduction. A negentropy is intro-
duced to determine the optimal dimension of Hankel
where P = diag(σ1 , σ2 , · · · ), σ1  σ2  · · · > 0. matrix.
The fundamental concept of noise reduction based on The negentropy can reflect the Gaussian characteris-
SVD is to select a threshold for singular value to divide tics of the noisy signal; the lower the negentropy of the
the original signal matrix into noise subspace and sig- signal is, the more the Gaussian noise is[18] . So, in order
nal subspace, then to reconstruct the denoised matrix to reduce the Gaussian noise, we take the negentropy as
with the singular values of signal subspace, and finally a criterion of the noise suppression performance. The
to extract the denoised signal from the reconstructed negentropy can be mathematically expressed as
matrix.
The Hankel matrix of the noisy PECT signal can be J(x) ≈
written as k1 {E[F1 (x)]}2 + k2 {E[F2 (x)] − E[F2 (μ)]}2 , (10)
Ay = AX + Ah , (5)
where J(x) is the negentropy, k1 and k2 are positive in-
where Ay , AX and Ah are the Hankel matrices of noisy tegers, x is the standardized random variable, μ is the
PECT signal, noise-free signal and noise, respectively. standardized Gaussian variable, F1 is an odd function
The SVD of Ay can be mathematically expressed as measuring the antisymmetric character of random vari-
  able, and F2 is an even function measuring the bimodal
T ΣX 0 VXT peak value of random variable at the origin. Functions
Ay = U ΣV = [UX Uh ] , (6)
0 Σh VhT F1 and F2 are defined as

F1 (x) = log cosh(ax) ⎬
1
and
a , (11)
2
F (x) = −e−x /2 ⎭
Ay = UX ΣX VXT + Uh Σh VhT , (7) 2
J. Shanghai Jiaotong Univ. (Sci.), 2015

where a = 1. the increase of SNR, the increase rate of negentropy


The simulated PECT signal is extracted by finite el- is higher when the SNR is lower, and the rate reduces
ement modeling (FEM) to study on the negentropy of with the increase of SNR. The negentropy and the SNR
noisy PECT signal. Then the signal is added by Gaus- are into a non-linear proportional relationship, and the
sian noise, the SNR of the noisy signal is set to be 20, negentropy can represent the character of SNR quali-
22, · · · , 50 dB, and the negentropy of these noisy sig- tatively. Therefore, it can be treated as an evaluating
nals is calculated. We plot the negentropy of different parameter of the noise suppression.
noisy signals with respect to the SNR in Fig. 1. 2.1 Determination of Hankel Matrix Optimal
As shown in Fig. 1, the negentropy increases with Dimension
The denoising result based on SVD is different with
different dimensions of Hankel matrices. When the
25.0 dimension of Hankel matrix is appropriate, the SNR
and the distortion of the denoised signal are lower.
24.5 Four simulated signals with different SNR levels are
discussed to determine the best dimension of Hankel
Negentropy

24.0 matrix, and the number of signal sample points is 800.


After the noise is reduced based on SVD with different
23.5 dimensions of Hankel matrices, the SNR and the ne-
gentropy of the denoised signals are calculated. In the
23.0 denoising process, the break point where the slope of
the singular value curve changes drastically is treated
22.5 as the threshold. The SNR and the negentropy of the
20 25
35 30
40 45 50
SNR/dB denoised signals with respect to different dimensions of
Fig. 1 The negentropy of noisy signals with different SNR Hankel matrices are plotted in Fig. 2, where SNRb is
levels the SNR of noisy signal before filtering.

31.50 24.40 34 24.46

29.25 24.15 32 24.32

Negentropy
Negentropy
SNR/dB

SNR/dB

27.00 23.90 30 24.18

24.75 23.65 28 24.04

22.50 23.40 26 23.90


0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Number of Hankel matrix rows Number of Hankel matrix rows
(a) SNRb = 20 dB (b) SNRb = 24 dB
39.00 24.60 42.60 24.620

36.75 24.53 40.35 24.585


Negentropy
Negentropy
SNR/dB

SNR/dB

34.50 24.46 38.10 24.550

32.25 24.39 35.85 24.515

30.00 24.32 33.60 24.480


0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Number of Hankel matrix rows Number of Hankel matrix rows
(c) SNRb = 28 dB (d) SNRb = 32 dB

Fig. 2 SNR and negentropy of the denoised signals with different dimensions of Hankel matrix
J. Shanghai Jiaotong Univ. (Sci.), 2015

As illustrated in Fig. 2, the distribution of the ne- rience sometimes. The points from 2 to 15 are chosen
gentropy is the same as SNR with different dimensions as the threshold for signal denoising based on SVD to
of Hankel matrices. We can deduce from Fig. 2(a) that reduce the interference and study the denoising per-
when the number of Hankel matrix rows is 270, the SNR formance. Then, the SNR and the negentropy of the
and the negentropy reach their maximum values simul- denoised signals are calculated, respectively. In this
taneously, When the number of Hankel matrix rows is section, four simulated signals with different SNR lev-
less than 270, the SNR and the negentropy increase els are discussed again. The SNR and the negentropy of
with the increase of rows number. On the other hand, the denoised signals with respect to the threshold point
when the number of rows is more than 270, the SNR and are plotted in Figs. 4 and 5.
the negentropy will decrease slowly with the increase of As shown in Figs. 4 and 5, for all noisy signals, when
rows number. Besides, Fig. 2 illustrates that when the the third singular value is treated as the threshold, the
SNR is different, the optimal dimension of the noisy sig- negentropy and the SNR of the denoised signal reach
nal Hankel matrix is also different. For the experiment their maximum values simultaneously, and the maxi-
testing, there is no such “original” signal for calculating mum value point can be seen clearly in every curve. It
the SNR, so the negentropy can be used to evaluate the can be deduced that the former three singular values
result of noise reduction. And the negentropy can be are obtained from the noise-free signal, and the others
used for determining the optimal dimension of Hankel are obtained from the noise signal. Namely, it indicates
matrix. When the negentropy reaches the maximum that the selected threshold can distinguish the singular
value, we treat the dimension of the Hankel matrix as values of signal subspace and the singular values of noise
the optimal dimension. subspace accurately. As pointed out earlier, when the
2.2 Determination of the Singular Value negentropy reaches the maximum value, the result of
Threshold signal denoising is the best, so the method proposed in
As mentioned above, when the optimal dimension of this section can be used to select the threshold exactly.
Hankel matrix is determined, the noisy signal matrix
can be divided into noise subspace and signal subspace 1.0
by a singular valve threshold based on SVD, and the
selection of the threshold can influence the noise re- 0.8
duction result. Normally, there are break points in Point 1
Amplitude

the curve of singular values of the noisy signal Han- 0.6


Point 2
kel matrix, and the break point where the slope of the
0.4
curve changes drastically can be treated as the thresh-
old point. The singular value curve of the Hankel ma-
0.2
trix for a PECT signal is plotted in Fig. 3.
As illustrated in Fig. 3, there are many break points
in the curve. It can be seen that the changes of the 0 20 10 30 40 50
curve slope are drastic and similar at Point 1 and Point Index number
2. Namely, there are interference points that interfere Fig. 3 Singular values of the Hankel matrix constructed
from a PECT signal
the determination of threshold in the curve, and the
threshold might only be determined upon human expe-

50 25 50 25

24 24
40 40
Negentropy

Negentropy
SNR/dB
SNR/dB

23 23
30 30
22 22

20 20
21 21

10 20 10 20
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16
Index number Index number
(a) SNRb = 20 dB (b) SNRb = 24 dB

Fig. 4 SNR and negentropy of the denoised signals with different singular value threshold (SNRb = 20, 24 dB)
J. Shanghai Jiaotong Univ. (Sci.), 2015

50 25 60 25

24 50 24
40

Negentropy

Negentropy
SNR/dB
SNR/dB

23 40 23
30
22 30 22

20
21 20 21

10 20 10 20
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16
Index number Index number
(a) SNRb = 28 dB (b) SNRb = 32 dB

Fig. 5 SNR and negentropy of the denoised signals with different singular value threshold (SNRb = 28, 32 dB)

3 Performance Evaluation ciency of the algorithm and the noise suppressing per-
formance simultaneously, the order of the median filter
Two simulated PECT signals are extracted from the is set to be 31. Then the noisy PECT signals of flat
flat and crack zones of the model to prove the advan- and crack are filtered by the three different approaches,
tage of the proposed approach in this paper. And the and the denoising results are shown in Table 1.
crack is 2 mm in depth and 1 mm in width. The sig- The results in Table 1 indicate that both the SNR
nals are added Gaussian noise with 15 dB. Performance and the negentropy of the noisy PECT signals are im-
of the proposed approach is compared with the tradi- proved by the three approaches respectively, and for
tional SVD method which is introduced in Noise reduc- the same noisy signal, the SNR and the negentropy
tion based on SVD section and the approach proposed of denoised signal based on the approach proposed in
in Ref. [11]. The dimensions of Hankel matrix for the this paper are maximum. Besides, it can be inferred
proposed approach and the traditional SVD method are that the relationships between the SNR and the negen-
230 × 571 and 400 × 401 respectively, the third singular tropy are different for the two kinds of signals, but for
value is treated as the threshold for the proposed ap- a specific kind of signal, the negentropy increases with
proach, and the tenth singular value where the slope of the increase of SNR. We can deduce that the denoising
the singular value curve changes drastically is treated performance of the approach of Ref. [11] outperforms
as the threshold for the traditional SVD method. For the traditional SVD method, and the approach pro-
the approach proposed in Ref. [11], considering the effi- posed in this paper is the best one.

Table 1 Comparing the performance of three different denoising approaches on the PECT signals of flat
and crack
SNR/dB Negentropy
Signal
Approach of Approach of Traditional Approach of Approach of Traditional
this paper Ref. [11] SVD method this paper Ref. [11] SVD method

Flat 38.132 6 35.281 3 30.012 5 24.369 2 24.351 1 24.321 2


Crack 38.863 2 35.436 2 30.435 6 24.536 8 24.501 5 24.448 2

4 Experimental Results and Discussion driver coil. The Hall sensor is used to extract the PECT
signal. The output voltage of the Hall element is ampli-
Experimental study is also performed to verify the fied, digitized, and then sampled by a data acquisition
performance of the approach proposed in this paper. card; finally, the data are transmitted to a computer
The schematic block diagram of experimental setup is for further analysis. An aluminum plate with a crack
illustrated in Fig. 6. is applied as the specimen, the crack is 2 mm in depth
A square wave voltage signal generated by a function and 1 mm in width, and the parameters of the driver
generator is converted to the current signal. The square coil are listed in Table 2.
wave current signal is amplified, and it is applied to the The signals are extracted from the flat and crack
J. Shanghai Jiaotong Univ. (Sci.), 2015

Function generator
structed firstly to reduce the Gaussian noise in the orig-
inal signals, and the dimension of the matrices is dis-
cussed based on the theory of determination of Hankel
Power amplifier matrix optimal dimension section. The negentropy of
the denoised signals with respect to different dimensions
of Hankel matrices is plotted in Fig. 8.
PECT probe Figure 8 shows that for the flat signal, the negen-
tropy is maximum when the number of Hankel matrix
rows is 180, and for the crack signal, the negentropy
Preamplifier is maximum when the number of Hankel matrix rows
is 160. Considering the sample point number of the
PECT signals, the optimal dimensions of Hankel ma-
Data acquisition
trix are 180 × 1 821 and 160 × 1 841 for the two signals,
respectively. The selection of threshold is discussed to
Computer further reduce the noise. Different singular values of the
optimal Hankel matrix are treated as threshold. Then
Fig. 6 Block diagram of pulsed eddy current experimental the noise is reduced based on SVD. The negentropy of
setup the denoised signals with respect to the singular value
threshold is plotted in Fig. 9.
Table 2 Parameters of the driver coil According to Fig. 9, the negentropy achieves the
Coil parameter Value maximum when treating the third singular value as
threshold for both signals. Therefore, when the third
Inner radius/mm 15
Outer radius/mm 32 0.30
Height/mm 55
0.25
Turns 600
0.20
Amplitude/V

0.15
zones of the specimen respectively, and the number of
Signal of flat
sample points is 2 000. The original PECT signals are 0.10
Signal of crack
shown in Fig. 7. It is well-known that the peak ampli- 0.05 Differential signal
tude and the peak reaching time of the differential sig-
nal are the important features of the PECT signal[19] , 0
but Fig. 7 shows that the differential signal is inter-
−0.05
fered by the Gaussian noise. The peak amplitude and 0 0.5 1.0 1.5 2.0
the peak reaching time of the differential signal cannot Index number×10−3
be seen clearly, so it is necessary to reduce the noise.
Fig. 7 The original PECT signals
The Hankel matrices of the PECT signals are con-

24.880 25.44

24.848 25.40
Negentropy

Negentropy

24.816 25.36

24.784 25.32

24.752 25.28

24.720 25.24
0 200 400 600 800 1 000 0 200 400 600 800 1 000
Number of Hankel matrix rows Number of Hankel matrix rows
(a) The PECT signal of flat (b) The PECT signal of crack
Fig. 8 The negentropy of the denoised PECT signals with different dimensions of Hankel matrix
J. Shanghai Jiaotong Univ. (Sci.), 2015

25.0 25.5

24.5 25.0

Negentropy
Negentropy
24.0 24.5

23.5 24.0

23.0 23.5

22.5 23.0
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16
Index number Index number
(a) The PECT signal of flat (b) The PECT signal of crack
Fig. 9 The negentropy of the denoised PECT signals with different singular value thresholds

singular value of the optimal Hankel matrix is treated reaching time of the differential signal can be found
as threshold, the denoising approach based on SVD is easily, and these three approaches can reduce the noise
optimal. The denoised signals filtered by the optimal effectively. However, Table 3 shows that for the same
method based on SVD are shown in Fig. 10 and the PECT signal, the negentropy of denoised signal which
negentropy of the denoised signals filtered by different is filtered by the approach proposed in this paper is
approaches is listed in Table 3. For the traditional SVD maximum. Therefore, the denoising performance of the
method, the dimension of Hankel matrix is 1 000×1 001, approach proposed in this paper is the best one, and it
the sixth singular value is treated as the threshold, and can reduce the noise more effectively.
the order of the median filter is set to be 31 in the
approach proposed in Ref. [11]. 5 Conclusion
As shown in Fig. 10, the peak amplitude and the peak
A new approach based on SVD is proposed to re-
duce the Gaussian noise of PECT signal. In this paper,
0.30 the negentropy is treated as an evaluating parameter
of noise suppression, and it is used to determine the
0.25 optimal dimension of Hankel matrix and the singular
value threshold. Experimental results show that the
0.20 proposed approach can reduce the Gaussian noise of
Amplitude/V

PECT signal more effectively compared with other ex-


0.15 isting approaches. Though the proposed approach can
reduce the Gaussian noise of PECT signal effectively,
0.10 Signal of flat
there may be other kinds of noise in the PECT signal,
Signal of crack such as probe wobble noise. Our future work may focus
0.05 Differential signal on the performance of the proposed denoising approach
when it comes to the other kinds of noise.
0
0 0.5 1.0 1.5 2.0 References
Index number×10−3
Fig. 10 The denoised PECT signals [1] Xu Z Y, Wu X J, Li J, et al. Assessment of wall thin-
ning in insulated ferromagnetic pipes using the time-
to-peak of differential pulsed eddy current testing sig-
Table 3 The negentropy of the denoised signals nals [J]. NDT & E International, 2012, 51: 24-29.
filtered by different approaches
[2] Alamin M, Tian G Y, Andrews A, et al. Principal
Negentropy component analysis of pulsed eddy current response
Signal from corrosion in mild steel [J]. IEEE Sensors Journal,
Approach of Approach of Traditional
2012, 12(8): 2548-2553.
this paper Ref. [11] SVD method
[3] Angani C S, Park D G, Kim C G, et al. The pulsed
Flat 24.869 7 24.851 9 24.815 6 eddy current differential probe to detect a thickness
Crack 25.421 8 25.392 2 25.336 2 variation in an insulated stainless steel [J]. Journal of
Nondestructive Evaluation, 2010, 29: 248-252.
J. Shanghai Jiaotong Univ. (Sci.), 2015

[4] Cheng W Y. Pulsed eddy current testing of car- borne time domain electromagnetic data [J]. Journal
bon steel pipes wall-thinning through insulation and of Applied Geophysics, 2011, 75: 264-276.
cladding [J]. Journal of Nondestructive Evaluation, [13] Rajwade A, Rangarajan A, Banerjee A. Image
2012, 31: 215-224. denoising using the higher order singular value decom-
[5] Mukriz I, Tian G Y, Li Y. 3D transient mag- position [J]. IEEE Transactions on Pattern Analysis
netic field mapping for angular slots in aluminium [J]. and Machine Intelligence, 2013, 35(4): 849-862.
Insight, 2009, 51(1): 21-24.
[14] Zhao X Z, Ye B Y. Similarity of signal processing
[6] Kiwa T, Hayashi T, Kawasaki Y, et al. Magnetic
effect between Hankel matrix based SVD and wavelet
thickness gauge using a Fourier transformed eddy cur-
transform and its mechanism analysis [J]. Mechanical
rent technique [J]. NDT & E International, 2009, 42:
Systems and Signal Processing, 2009, 23: 1062-1075.
606-609.
[7] He Y Z, Luo F L, Pan M C. Defect characterization [15] Zhang Li-zhi. Surface defects inspection for con-
based on pulsed eddy current imaging technique [J]. tinuous casting slab by pulsed eddy current [D].
Sensors and Actuators A, 2010, 164: 1-7. Chongqing, China: College of Material Science and
Engineering, Chongqing University, 2011 (in Chinese).
[8] Hosseini S, Lakis A A. Application of time frequency
analysis for automatic hidden corrosion detection in a [16] Hassanpour H, Zehtabian A, Sadati S J. Time do-
multilayer aluminum structure using pulsed eddy cur- main signal enhancement based on an optimized singu-
rent [J]. NDT & E International, 2012, 47: 70-79. lar vector denoising algorithm [J]. Digital Signal Pro-
[9] Theodoulidis T, Wang H T, Tian G Y. Extension cessing, 2012, 22: 786-794.
of a model for eddy current inspection of cracks to [17] Shih Y T, Chien C S, Chuang C Y. An adaptive pa-
pulsed excitations [J]. NDT & E International, 2012, rameterized block based singular value decomposition
47: 144-149. for image denoising and compression [J]. Applied Math-
[10] Fan M B, Huang P J, Ye B, et al. Analytical model- ematics and Computation, 2012, 218: 10370-10385.
ing for transient probe response in pulsed eddy current [18] Lago-Fernández L F, Sánchez-Montañés M,
testing [J]. NDT & E International, 2009, 42: 376-383. Corbacho F. The effect of low number of points in
[11] Huang C, Wu X J, Xu Z Y, et al. Pulsed eddy clustering validation via the negentropy increment [J].
current signal processing method for signal denoising Neurocomputing, 2011, 74: 2657-2664.
in ferromagnetic plate testing [J]. NDT & E Interna- [19] Tian G Y, He Y Z, Abewale I, et al. Research on
tional, 2010, 43: 648-653. spectral response of pulsed eddy current and NDE ap-
[12] Reninger P A, Martelet G, Deparis J, et al. Sin- plications [J]. Sensors and Actuators A, 2013, 189:
gular value decomposition as a denoising tool for air- 313-320.

You might also like