This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

1
TIM-22-05197R1

Yizhao Wang, Jingbo Guo, Member, IEEE, Qihang Shi, and Tiehua Hu

Abstract A novel crack detection method using wavelet-based regions (scales and locations), and then identify inner-surface
decision-level data fusion is proposed and verified by simulations or outer-surface cracks of pipe walls.
and experiments. In this work, we established a crack signal model Until now, many methods to detect cracks from measured
using wavelet functions that fits all crack signal scales without
need of windowing, and showed that the signal-to-noise-ratio (SNR)
signals have been reported, which could be classified into two
difference between the independent signals to be fused has a types, i.e., machine learning (ML) methods and signal
significant effect on the overall detection performance. Based on processing (SP) methods. ML methods, such as convolutional
this observation, we designed four wavelet-based decision-level neural network (CNN) [14], support vector machine (SVM)
data fusion rules. We then presented a detection method where [15], random forest (RF) [16], etc., usually transform the
wavelet processing results of individual NDT input are accepted or detection problems into classification problems. Song et al. [14]
rejected based on said rules to produce optimized estimation
accuracy. To evaluate the proposed method, the first group of
proposed a data and decision level fusion-based crack detection
simulations were implemented to show the proposed method method using 1D CNN, and used a fixed sliding window of
identifies the inner and outer-surface cracks with good estimation 1024 points with 50% overlap rate to obtain raw samples for
accuracy; the second and third groups of experiments showed the CNN; this method reported better classification accuracy than
proposed method does improve upon individual detection methods comparable methods, but it remains difficult to produce a one-
alone, and has better detection performance than three state-of- size-fits-all fixed sliding window for all possible cracks and it
the-art methods; finally, the pull-rig experiments verified on our
own pipeline inspection gauge (PIG) that the proposed method
is difficult to eliminate the duplicate detection results in the
does improve the detection probability beyond that of individual overlap regions. Jing et al. [17] and Teng et al. [18] used serial-
detection methods, in an actual high-speed pipeline inline splicing and fixed sliding windows to acquire the detection
inspection (ILI). samples and reported the best classification results by CNN
among similar methods; the serial-splicing sliding window
Index Terms Crack detection, decision-level data fusion, high- avoids the problem of duplicate detection, but it introduces a
speed inspection, nondestructive testing (NDT), wavelets.
risk where one detection target becomes separated into two
segments, which increases the miss detection probability.
I. INTRODUCTION
One unavoidable caveat with ML methods is that they

C
RACKS often appear on the inner surface and outer require large quantity of training samples [19]. This limits how
surface of oil and gas pipelines due to transported large each sample can be, necessitating a maximum window
medium and harsh operating environment, which size which the detection algorithm has to look through.
threaten the safe operation of pipelines. At present, many Meanwhile, SP methods such as generalized likelihood ratio
mature nondestructive testing (NDT) technologies exist that test (GLRT) [20], Bayesian approach [21], locally most
detect cracks in pipelines, for instance, magnetic flux leakage powerful test (LMPT) [22], wavelet based singularity analysis
(MFL) testing [1]-[4], eddy current (EC) testing [5], pulsed [23], etc., have dealt with the problem of uncertain window
eddy current (PEC) testing [6]-[8], thermography testing (TT) sizes for a long time. For instance, wavelet functions can focus
[9], ultrasonic testing (UT) [10], etc. However, accurately on any moment and any frequency of signals, which allows the
detecting cracks from the low-SNR measured signals remains a wavelet transform an outstanding performance in time domain
crucial problem, because the crack sizes are usually smaller and frequency domain analyses [24], thereby fitting all possible
than corrosions (depths and openings at the micrometer scale) sizes of cracks without using any window. It also produces both
[11], [12], which means crack signals can easily vanish in the the scales and locations of the crack signals, which is necessary
noisy background, especially when PIGs run at high speed [13]. to provide in practical pipeline inspection operations and would
Thus, the motivation of this paper is to detect cracks of pipelines require extra work with a ML method that only provides
from the low-SNR measured signals, and estimate crack signal presence of cracks.
Detecting cracks with NDT methods is often not mutually
Manuscript received XX, 2022; revised XX, 202X; accepted XX, 202X. exclusive; far more often than not results from multiple
Date of publication XX, 202X; date of current version XX, 202X. This work
was supported by National Key Research and Development Program under technologies are available. For example, MFL PIGs often also
Grant 2016YFC0802100, and Regional Innovation Cooperation Project of incorporate EC or PEC sensors used to identify the inner-
Sichuan Province under Grant 20QYCX0009. The Editor coordinating the surface or outer-surface cracks of pipe walls [25]. This
review process was XX. (Corresponding author: Jingbo Guo.)
The authors are with Department of Electrical Engineering, Tsinghua inherently begs improvement with data fusion techniques.
University, Beijing 100084, China (e-mail: [email protected]). Generally, the fusion operations can be implemented at the

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

2
TIM-22-05197R1

signal, feature, and decision level [26]. The fusion at signal on dynamic search domain. Comprehensive simulations, Monte
level will combine measured signals in different methodologies Carlo experiments, method comparisons, and the tests with
to obtain better signal quality or image quality [27], [28]. For pull-rig data are introduced in Section IV. Finally, Section V
example, Heideklang and Shokouhi [12] in 2015 utilized concludes this paper.
discrete wavelet transform (DWT) to fuse and reconstruct the
images. However, the data fusion at signal level will involve II. WAVELET-BASED DECISION- LEVEL DATA FUSION
interpolation, which can introduce errors into location
A. Crack Signal Modeling Based on Wavelets
estimation [29] and identification of inner-surface or outer-
surface cracks. The feature-level fusion derives the features that In this paper, we take the data fusions in two kinds of
make the targets distinguishable in the feature space [26]. independent signals, based on different inspection technologies,
for examples to design the crack detection method for pipelines.
Extracting feature values from different signal sources is the
Specifically, the array signals of y-axis Hall-effect sensors in
first step of this approach, and then the appropriate mathematic
MFL testing, denoted as MFLY array signals, as well as the
methods, e.g., Hadamard product [30 array signals of novel pulse eddy current testing [25], denoted
evidence theory [31], will be used to fuse the feature values and as NPEC array signals, are selected to detect cracks of pipelines.
present defect information in the end. Nevertheless, the feature- Assume that for the same test area, it could be detected by the
level fusion may lose information under certain circumstances MFLY array signals and NPEC array signals simultaneously
[32], which leads to estimation errors of crack signal regions. and independently, and these two sorts of signals have the same
In contrast, decision-level fusion allows signals from different data dimensions. The problem of crack detection in pipelines
NDT technologies to be processed individually, and the final can be regarded as the detection of deterministic signals with
decision will be made with some fusion strategies, e.g., Kernel unknown parameters, such as signal amplitudes, signal scales,
density estimation (KDE) strategy utilized by Heideklang and and arrival time, in theory. Generally, the signal amplitudes and
Shokouhi [29] in 2016, to reduce uncertainty with redundant signal scales are related to the sizes of cracks, while the arrival
information, and does not modify original signals in whole time is correlated with locations of cracks.
fusion process. First and foremost, signal models should be established for
Not modifying original signals is desirable in our task, which MFLY and NPEC crack signals according to signal shapes. The
requires accurate computation of crack locations. Some SP shapes of MFLY crack signals are just like sine functions
methods can introduce phase error and impact this accuracy, constantly in the domain of [0, 2 ]. As to NPEC crack signals,
therefore not modifying the original signal at all makes it clear the shapes of them are different and a little complex between
that no phase error has been introduced. Thus, decision-level inner-surface and outer-surface cracks of pipe walls (as shown
data fusion strategy is most suitable for our purpose. in Fig. 19 of [25]): along the x axis, the maximal negative peak
appears first and the maximal positive peak appears last for
To the best of our knowledge, no existing strategies for
inner-surface cracks; while for outer-surface cracks, the
decision-level data fusion consider the SNR difference between
maximal positive peak appears first and the maximal negative
the independent signals to be fused; we shall show that this
peak appears last. According to the similarity of signal shapes,
difference illustrated in detail later in this paper has a significant the Gaussian wavelet with one vanishing moment, denoted as
effect on the fusion performance, and can be utilized to improve , serves as MFLY crack signal model, whereas the
the performance. Gaussian wavelet with five vanishing moments, denoted as
Based on the above considerations, we propose a novel crack , serves as NPEC crack signal model; specifically,
detection method for pipelines. The main contributions of this is used for outer-surface crack modeling, while is used
paper are summarized as follows. First, we systematically for inner-surface crack modeling. Note that and
derive the new wavelet-based decision-level data fusion rules, are both real as shown in Fig. 1.
by comparison of SNRs between the independent signals to be The Gaussian function, denoted as , can be represented
fused. Second, using the new data fusion rules, a crack detection by (1). The and are given by (2) and (3), which are
method for pipeline array signals is proposed, which can the first and fifth derivatives of the . The coefficient
determine whether there is a crack, what the crack signal region is defined such that the 2-norm of the
is, and whether it is the inner-surface or outer-surface crack. derivative of Gaussian function is equal to 1 [23]. Here,
Third, to rapidly estimate crack regions in the detection method, , , .
the golden section algorithm based on dynamic search domain (1)
is designed. Finally, the effectiveness and advantages of the (2)
proposed detection method are verified with three groups of
(3)
simulations and experiments, as well as the real world high-
speed pull-rig data. and can be uniformly described by mother
The rest of the paper is organized as follows. The crack signal wavelet , which gets scaled by a factor of a and translated
modeling, test statistic of crack detection, combined detection by a factor of b to generate the wavelet function given by [23]:
performance, and the new data fusion rules are presented in
(4)
Section II. Section III introduces the matrices of test statistic,
crack detection method, and the golden section algorithm based

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

3
TIM-22-05197R1
(t) (t)
0.8
1
0.8
5
We consider the problem of detecting a deterministic signal
0.6 0.6 known except for A, a, b in white Gaussian noise (WGN).
0.4 0.4 Specifically,
0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4

-0.6 -0.6
(11)
-0.8 -0.8
-5 0 5 -5 0 5
Relative time Relative time

Fig. 1. Gaussian wavelet with one vanishing moment (left) and Gaussian where is WGN with variance and mean value 0. In the
wavelet with five vanishing moments (right). case of unknown parameters of deterministic signals, the GLRT
will decide if
where a>0 is the scaling factor denoted as the frequency of (12)
crack signals, while is the translation factor concerned
with the arrival time of crack signals. The factor of above
where , , and are the maximum likelihood estimations
can guarantee the has the same energy as , that is: (MLEs) of , , and , respectively; is the
(5) probability density function (PDF) of x under the condition that
we have got , , and based on hypothesis of ;
The continuous wavelet transform (CWT) of a time- is the PDF of x under hypothesis of . Here, can be
dependent real signal x(t) is defined as [23]: represented as:
(6) (13)

Mathematically, the center coordinate of , i.e., , can and can be given by:
be calculated by the center coordinate of , i.e., , as
. Besides, the half width of , i.e., , can be
represented by the half width of , i.e., , as .
Thus, the local information in the time domain of [
] is given by CWT. As shown in Fig. 1, the (14)
center coordinates of and are equal to 0, and the
half widths of them are approximately equal to 5, so the local where . Substituting (13) and (14)
into (12) produces:
information in the domain of [ ] is useful, and the
information out of this region could be treated as 0.
To employ the CWT into the digital signal processing, time (15)
domain should be sampled discretely; suppose
represents step size of the time domain, which means time
Taking logarithms we decide if
domain is discretized as
, that is also corresponded to (16)
= . Thus, the discrete-time
continuous wavelet transform (DCWT) can be derived by (4), It has been verified in Appendix I that the MLE of is given
(5) and (6) as: by:
(7)
(17)
(8)
(9)
Substituting (17) into (16) produces:
For DCWT, the scaling factor represents crack-signal (18)
frequency or period, while the translation factor
indicates the arrival time of crack signal.
Substituting (8) and (17) into (18) produces:
(19)
B. Test Statistic of Crack Detection
Considering the unknown parameters, i.e., signal amplitudes, It has been verified in Appendix II that, and are found by
signal scales, and arrival time, the signal model of cracks can maximizing the value of over all possible
be revised as:
and , where and are the MLEs of and . Thus, the test
(10) statistic of crack detection can be summarized as:
where [n] is the mother wavelet of DCWT and A is the signal (20)
amplitude.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

4
TIM-22-05197R1

It is noted that using the sign of absolute value in (19) and (20) (26)
makes this method suitable not only for inner-surface crack
detections, but also for outer-surface crack detections. Besides,
(17) inspires us an effective method to identify the inner-surface where is the deflection coefficient. (25) or (26)
and outer-surface cracks of pipe walls based on the sign of of describes the individual detection performance of one signal
NPEC crack signal: if of NPEC crack signal is more than 0, model as (10).
the result is identified as outer-surface crack; while if of If we consider the combined detection by two independent
NPEC crack signal is less than 0, the result is determined as signals, the combined false alarm probability, i.e., ,
inner-surface crack. becomes:

C. Analysis of Combined Detection Performance of Two

Independent Signals (27)
First, two definitions about decisions and detections are given
as follows:
1) Individual decision is defined that the final decision is (27) describes that if the decision of one MFLY signal or the
determined only by the MFLY signals or only by the NPEC decision of one NPEC signal determines false alarm, the
signals, and the decision is consistent with that of combined decision of them is false alarm. Assume that
corresponding independent signals. Besides, the detection , and then we have:
based on individual decision is defined as individual detection. (28)
2) Combined decision is defined that the final decision should
be considered together with the MFLY signals and the NPEC Thus, based on (26) and (28), the combined detection
signals. It means that if MFLY signals or NPEC signals probability, i.e., , becomes:
determine YES, the final decision is YES; only if all of MFLY
signals and NPEC signals determine NO, can the final decision
be NO. In addition, the detection based on combined decision
is defined as combined detection.
Then, starting from the analysis of individual detection
performance, we derive the combined detection performance of
two independent signals. The determination of the detection
performance of the GLRT given by (20) with unknown and
is difficult [33]. However, assume the MLEs of and could
be accurately estimated by some algorithm, and then we can (29)
equivalently analyze the detection performance given by (19)
approximately. (29) describes that if the decision of one MFLY signal or the
Assume that , and it has been decision of one NPEC signal determines true positive detection,
verified in Appendix III that: the combined decision of them is true positive. Substituting (28)
(21) into (29) produces relation between and , that is,
combined detection performance of two independent signals
based on combined decision, represented as .
Then, the individual false alarm probability and individual It is hard to verify the quantitative relation between combined
detection probability are: detection performance and individual detection performance
(22) mathematically, but we can plot receiver operator characteristic
(ROC) curves of them simultaneously, by MATLAB as Fig. 2
(23) based on (26), (28) and (29). Here, the individual detection
performance is given by solid lines, while the combined
where is the right tail probability of standard normal detection performance is plotted by dash-star lines. Note that
the performance comparisons between combined detections
distribution, that is, ; is and individual detections focus on the signals measured by
the inverse function of , and from (22) the detection sensors at the same time. For instance, the 0dB & -3dB in Fig.
threshold is: 2(b) represents the combined detection performance based on
(24) the signals with SNR of 0 dB and -3 dB; e.g., if the SNR of
MFLY crack signal is 0 dB, and the SNR of NPEC crack signal
so that: is -3 dB meanwhile, we could compare the combined detection
performance of 0dB & -3dB with the individual detection
(25)

or finally:

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

5
TIM-22-05197R1

(a) (b) (c)

Fig. 2. ROC curves given by theoretical calculations based on individual detections and combined detections: (a) the SNR of two independent signals is equal;
(b) the reference signal with SNR of -3 dB; (c) the reference signal with SNR of 10 dB.

performance of signal with SNR of 0 dB or -3 dB, rather than In next subsection, we will introduce the novel data fusion
with other SNRs, such as 10 dB or 5 dB, etc. rules, and expect them to improve the crack detection
As shown in Fig. 2(a), if two independent measured signals performance, no matter how different the SNRs between two
have the same SNR, the combined detection performance of independent signals.
them is clearly better than their individual detection
performance.
D. Wavelet-Based Decision-Level Data Fusion Rules
However, if the SNRs of two independent measured signals
are different, their combined detection performance is not Substituting the (24) into the (19) produces:
absolutely better than their individual detection performance. (30)
As shown in Fig. 2(b), if SNR of reference signal is just -3 dB
constantly, the combined detection performance of yellow line where is wavelet function of DCWT, and (30) is the
and part of green line is better than their individual detection criterion of crack detection affected by , and of WGN.
performance; while the combined detection performance of red,
Substituting the (17) into the (30) produces
blue and part of black dash-star lines is worse than the same
color of individual detection performance. Similarly, as shown ; then based on , we obtain:
in Fig. 2(c), if SNR of reference signal is 10 dB constantly, the
combined detection performance of red line and part of blue line (31)
is better than the individual detection performance of 10 dB. As
for other lines, their combined detection performance is worse We know that the power of pure signal is
than the individual detection performance of 10 dB. , and the power of noise is
Consequently, as for two independent signals, e.g., MFLY
signal and NPEC signal, measured simultaneously by different , where
sensors to the same test area, their combined detection ; thus, (31) derives the MLE of SNR as:
performance is better than corresponding individual
(32)
performance of higher-SNR signal, only when their SNR values
are relatively close, where the degree of closeness can be judged
by comparing combined detection probability, i.e., given Equation (32) is the equivalent transformation of (30), and it
by (29), with individual detection probability of higher-SNR illustrates that if (32) is used to detect cracks, the threshold of
signal, i.e., given by (26), under the same constant false (32) varies with under that and have been set as
alarm probability: if , it means the SNR values of constants in advance. Thus, when MFLY and NPEC signals
two independent signals are relatively close, and then we should work together to detect cracks for the same area, if one of them
trust combined decision to get higher detection performance; determines YES, i.e., , while
else if , it means the SNR values of two independent the other determines NO, i.e., ,
signals are relatively different, and then we should trust
individual decision of higher-SNR signal (MFLY signal or it is uncertain which is bigger between and . We
could summarize four fused decision rules, fused with the
NPEC signal), so that the overall detection performance will not
individual decisions and combined decisions according to
reduce for any . Based on above conclusions, the
fused decision and fused detection are defined as follows: and , as follows:
Fused decision is defined that the final decision is fused with 1) If and
the individual decision and combined decision, and the :
switchover between them depends on the wavelet-based
decision-level data fusion rules. Moreover, the detection based 1.1) If , we should trust combined decision,
on fused decision is defined as fused detection. i.e., YES;
1.2) If , we should trust individual decision of

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

6
TIM-22-05197R1

higher- signal, i.e., YES.

2) If and
:
2.1) If , we should trust combined decision,
i.e., YES;
2.2) If , we should trust individual decision of
higher- signal: if , output YES; else if
, output NO. Fig. 3. Flow chart to get the matrix of test statistic for MFLY/NPEC array
3) If and signal.

:
3.1) If , we should trust combined decision, (33)
i.e., YES;
3.2) If , we should trust individual decision of
higher- signal: if , output YES; else if which is the -dimensional matrix, where is
, output NO.
the wavelet function of MFLY or NPEC crack signal given by
4) If and scaling factor , and translation factor ; assume that
: output NO in any case. and ; and .
In summary, the above fused decision rules derive the Note that (33) is equivalent to calculate the moduli of DCWT
wavelet-based decision-level data fusion rules as follows: for in the total domain of and .
1) If and Assume that the channel count of MFLY and NPEC signals
both are m, and if we consider the inspection data of all channels
: output YES and locate the cracks together, the three-dimensional matrix of test statistic for
based on the signal with higher . MFLY or NPEC array signals is defined as:
2) If and (34)

: which is the -dimensional matrix. The physical

2.1) If and : output NO; meaning of is that there are channels of
2.2) Else: output YES and locate the cracks based on the signals, numbers of sampling points in the inspection data,
signal with . and the domain of scaling factor is . Note that (34)
3) If and consists of moduli of DCWT for every signal channel, in the
total domain of and . The flow chart to get the
: matrix of test statistic for MFLY/NPEC array signal is
3.1) If and : output NO; summarized in Fig. 3.
3.2) Else: output YES and locate the cracks based on the The target of crack detection is to find every peak that is more
signal with . than detection threshold in (34) for MFLY and NPEC signals,
or equivalently, to find every peak that is more than detection
4) If and
threshold in (33), when and {MFLY, NPEC}.
: output NO. In next subsection, a new crack detection method is designed to
In next section, we will design the crack detection method by achieve above target.
virtue of above data fusion rules.
B. Crack Detection Method for Array Signals
III. CRACK DETECTION METHOD USING WAVELET-BASED
DECISION- LEVEL DATA FUSION The crack detection method consists of 6 steps in total as
shown in Fig. 4. From Step1 to Step3, the processing procedures
A. Matrices of Test Statistic for MFLY and NPEC Array of MFLY signals are listed in the left side, while the right side
Signals is for the NPEC signal processing.
Generally, the inspection data of pipelines consist of array Step1: Acquire two-dimensional MFLY and NPEC array
signals with 2 dimensions, i.e., dimension and signals. For the same test area, the two-dimensional MFLY
clock dimension, as two pictures of Step1 in Fig. 4, where the array signals and corresponding NPEC array signals can be
inspection data are sampled along the direction, obtained from the hard disk of the PIG; the MFLY and NPEC
while the different data channels, i.e., different blue lines, are array signals should have the same data dimensions, that is,
distributed along the clock direction. As to the MFLY or NPEC . Assume that the domains of MFLY and NPEC signals
signal of channel i, i.e., , where {MFLY, NPEC}, it both are [0, 4095], and the domain of scaling factor is
can be transformed by (30) to two-dimensional submatrix of here. Besides, one of -dimensional 0-1
test statistic as: crack-decision matrix should be initialized to 0 for every
element; at the end of Step5, we will see the corresponding

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

7
TIM-22-05197R1

level data fusion rules. The optimal estimated set, supposing

that is {( ), ( ), ( )}, is output based
on data fusion rules, which has been introduced in the Section
II-D in detail.
Step5: Determine the crack signal regions. Based on Step4,
the crack signal regions are calculated as ,
, and , respectively,
and then set the decision-matrix elements corresponding to
crack regions to 1.
Step6: Mark the crack regions. Switch to another channel and
repeat Step2-Step5 until all channels of the MFLY and NPEC
signals have been detected. At last, the areas where the elements
of 0-1 crack-decision matrix are equal to 1 represent the crack
regions, and then mark these regions by red minimum bounding
boxes. Note that the 6 red bounding boxes in the Step6 of Fig.
4, actually describe 3 cracks, i.e., No. 1, No. 2, No.3, given by
0-1 crack-decision matrix; they are just shown respectively in
MFLY and NPEC array signals.

C. Golden Section Algorithm Based on Dynamic Search

Domain
In the Step3 of Fig. 4, the target of crack detection is to find
every peak that is more than detection threshold in (33), when
and {MFLY, NPEC}. Note that (33) is
-dimensional submatrix, and using traversal method to find
every peak will consume vast computing resources, because the
matrix elements are obtained by numerical integration.
Unfortunately, we could only find the peaks step by step along
the direction, because the numbers and locations of
Fig. 4. Schematic diagram of crack detection method. cracks are stochastic and unknown beforehand. However, along
the direction, it is unnecessary to check every point to
elements of decision matrix are set to 1 according to the crack find the optimal scaling factor, because we have proved in
signal regions. Appendix IV that, under the non-noise condition, in the crack
Step2: Detect cracks in every signal channel. Starting from the regions, the test-statistic distributions of cracks along the
first channel, detection is performed simultaneously on MFLY direction are unimodal; so we could use golden section
and NPEC signals channel by channel. As Step2 in Fig. 4, one algorithm to find peaks along the direction. As for other
channel of MFLY and NPEC signal is presented. regions without cracks, the test statistics are zero under the non-
Step3: Find all of the peaks from test-statistic submatrix. Each noise condition, which are easier for optimization.
channel of MFLY signal corresponds to a submatrix of test To rapidly find every peak that is more than detection
statistic given by (33), and its spatial distribution of values is threshold in (33), the golden section algorithm based on
shown as the left figure of Step3 in Fig. 4, where the blue region dynamic search domain is proposed. Assume the search domain
represents the test statistic is more than detection threshold. The of is , and the golden section points are defined
optimal estimated set of scaling factors, translation factors, and as [34]:
the test statistics for MFLY signal, given by three peaks, i.e.,
black dots in the figure, is {( ), ( ), (35)
( )}. Each channel of NPEC signal also corresponds to
a test-statistic submatrix defined by (33), and its spatial The initial value of , and ; but the
distribution of values is shown as the right figure of Step3 in and will be adjusted dynamically according to the crack
Fig. 4, where the yellow region describes the test statistic is regions, which is the characteristic of this algorithm. Besides,
more than detection threshold. The optimal estimated set of to reduce the false alarm probability further, a priori threshold,
scaling factors, translation factors, and the test statistics for i.e., , is involved given by the minimum amplitude of
NPEC signal is {( ), ( ), ( )}. crack signals we have known, or given by M times of the
Specifically, the method to calculate peaks will be illustrated in detection threshold, where M is an empirical value based on
the Section III-C. simulations. To sum up, the steps of algorithm for the MFLY or
Step4: Output optimal estimated sets of crack parameters. NPEC signal of channel are listed as follows, and its flow
Along the direction, put the sets of {( ), chart is shown in Fig. 5.
( )}, {( ), ( )}, and {( ), Step1: Set , , , and clear the
( )}, successively, into the wavelet-based decision- ROM of temporary parameters and clear ROMp of peaks;

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

8
TIM-22-05197R1

Step2: For , find the peaks step by step

along direction; if searching for is over, exit the
algorithm and check the peaks in ROMp; else go to Step3;
Step3: Calculate ,
, , and
;
Step4: Repeat Step5 until , where ; then,
go to Step6;
Step5: If , set ; ; ;
; ;
Else, set ; ; ;
; ;
Step6: Set ; ;
Step7: If , save to ROM;
; , where
is the adjusted value of search domain; here assume
; if , go to Step8; else go to
Step2;
Else: ; ; then go to Step8;
Step8: If ROM is empty, go to Step2; Fig. 5. Flow chart of golden section algorithm based on dynamic search
Else: Calculate and corresponding and domain.
from ROM; then go to Step9;
Step9: If , determine A. Evaluation for Crack Detection Method
is one of peaks we find in (33), and save this peak to ROMp; We simulate crack detections in one channel of MFLY and
then clear the ROM and go to Step2; NPEC signals for instances to evaluate the detection method.
Else, reject and clear the ROM; We design three simulating MFLY crack signals given by (36),
then go to Step2. where , , for ; ,
The golden section algorithm based on dynamic search , for ; and , ,
domain will find optimal scaling factor from , for , respectively. Similarly, three simulating
when the search footprint is far from the peaks; but when the NPEC crack signals given by (37) are prepared, where
search footprint is close to the peak, as the blue region and , , for ; , ,
yellow region in Step3 of Fig. 4, the search domain will be for ; and , ,
adjusted to , so as to for , respectively. Note that ,
accelerate the algorithm based on smaller search domain. Note , and represent three cracks
that the process of using to filter peaks in above Step9 in substance, and each crack is detected by MFLY signal and
also could be put between the Step4 and Step5 in Fig.4, because NPEC signal simultaneously; besides, the regions of them are
this process is just for eliminating the false alarming peaks, and given by as shown in Table III.
the final results are unrelated to the elimination time.
In fact, the actual measured signals contain noises, which
probably make the signals discontinuous or heteromorphic; this (36)
situation may reduce the performance of golden section
algorithm. However, by virtue of the data fusion, simulations
demonstrate the proposed crack detection method actually has
the good performance even if the SNR is 0 dB or -3 dB. (37)

IV. SIMULATIONS AND EXPERIMENTS As for other parameters, the domain of , is [1, 100],
This section consists of 4 parts. Part A performs three with step size ; so we get
simulations to evaluate the crack detection method based on for definitional domain; while the
four data fusion rules. Part B implements six groups of Monte domain of is [0.4, 2.0] with step size of 0.1. The WGNs
Carlo experiments to evaluate the detection performance of the with constant are added on the MFLY and NPEC
method. Part C is for method comparisons. Part D evaluates the signals, so as to get different SNRs for signals, as shown in
proposed method in actual pull-rig experiments. Table II. Besides, the false alarm probability is set to be
0.01 for all of the experiments; according to (24), the detection
threshold, i.e., , determined by , and , approaches to

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

9
TIM-22-05197R1

(a) (b)

(c) (d)
Fig. 6. Simulations in the first experiment.
TABLE I
ERRORS OF AMPLITUDES, SCALING FACTORS, AND TRANSLATION FACTORS IN THE FIRST EXPERIMENT

/ / /
A a b A a b A a b
Design value 2.000 1.500 10.000 3.000 1.000 50.000 1.000 0.800 90.000
MFLY Estimate 1.863 1.400 9.800 3.059 1.000 50.100 1.017 0.700 90.100
Error -0.137 -0.100 -0.200 0.059 0.000 0.100 0.017 -0.100 0.100
Design value -2.000 1.500 10.000 3.000 1.000 50.000 -1.000 0.800 90.000
NPEC Estimate -2.036 1.500 10.000 2.939 1.000 50.100 -0.765 0.900 90.100
Error -0.036 0.000 0.000 -0.061 0.000 0.100 0.235 0.100 0.100

TABLE II TABLE III

SNR AND PD IN THE FIRST EXPERIMENT ERRORS OF CRACK REGIONS IN THE FIRST EXPERIMENT

Crack 1 Crack 2 Crack 3 Crack 1 Crack 2 Crack 3

LB RB LB RB LB RB
Designed SNR (dB) 0.398 -0.023 5.922 5.095 -1.775 -3.216 Target region 2.5 17.5 45.0 55.0 86.0 94.0
Estimated SNR (dB) -0.066 0.405 5.690 5.342 -2.342 -5.898 Estimated region 2.5 17.5 45.1 55.1 86.6 93.6
SNR Error (dB) -0.464 0.428 -0.232 0.247 -0.568 -2.681 Error 0.0 0.0 0.1 0.1 0.6 -0.4
1.000 1.000 1.000 1.000 0.999 0.988 Midpoint error 0.0 0.1 0.1
1.000 1.000 1.000 Region-length error 0.0 % 0.0 % -12.5 %

0.407; while a
priori threshold is are shown in Fig. 6.
. The 3D test statistics of one channel of MFLY and NPEC
Note that the parameter values of , , , i=1,2,3, are not signal are shown in Fig. 6(a), and we find three pinnacles
very important, but the SNR affected by these parameters is the corresponding to three crack signals in each figure; besides,
consideration for the detection performance, as shown in Fig. 2. many red dots are the optimal positions of scaling factors step
Thus, if SNRs meet the design requirements, as listed in Table by step along the direction, given by the golden
II, the parameters of , , , i=1,2,3, are consistent with the section algorithm based on dynamic search domain.
actual application scenarios. Besides, the parameters of The original signals including pure signals and signals with
, and are affected by the sampling WGN are shown in Fig. 6(b). The parameters of different crack
precision and spatial resolution, while is decided by signals and their estimated results are shown in Table I; the
engineering requirements, so they all could be adjusted estimated SNRs, and are listed in Table II. In particular,
according to actual situations. the amplitudes of NPEC crack signals are designed as
To evaluate the crack detection method based on four data , respectively, where the negative
fusion rules in Section II-D, three simulations are performed. numbers represent inner-surface cracks, while the positive
The first experiment is prepared to evaluate the crack detection number is for outer-surface crack. It can be found that the errors
method based on the first and the fourth data fusion rules. At of amplitudes, scaling factors, and translation factors are small
first, suppose all of the MFLY sensors and NPEC sensors work for MFLY and NPEC crack signals; especially the signs of
normally so that three MFLY crack signals and three NPEC amplitudes for NPEC crack signals are all estimated correctly,
crack signals are added in the experiment, and their computed although the designed SNR of is only -3.216 dB; it
results are listed from Tables I to III, and the simulation results means that the proposed crack detection method can identify

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

10
TIM-22-05197R1

the inner and outer-surface cracks accurately even if the SNR is

lower.
Fig. 6(c) shows distributions of estimated test statistics given
by optimal along the direction, where the blue
lines are detection thresholds, and the red dash lines are a priori
thresholds. The dual thresholds both act important roles for the
detection method: the PFA will increase due to noises if only
detection thresholds work in the experiments, while the
repetitive results will be reported for one crack if only a priori
(a)
thresholds work in the experiments. Take the right figures of
Fig. 6(c) for example, the test statistics of the first NPEC crack
signal have multiple crests exceeding a priori threshold, but the
valleys between the crests do not fall below the detection
threshold; thus, these crests will be processed together to
generate only the highest peak corresponding to the first crack.
In Fig. 6(d), the green curves denote estimated crack signals
whose test statistics are more than detection threshold, while the
vertical boundaries of blue dash lines represent the estimated
target regions. By comparing the boundaries of green curves (b)
and blue dash lines whether coincide, we could know which
estimated signals are used to locate the cracks. In Fig. 6(d), the
NPEC signal is selected to locate the first crack, because
and ; while the MFLY signals
are selected to locate the second and third cracks, because
and for the second crack,
and and for the third
crack; the results illustrate that crack signal regions are
determined by the signals with higher when both of the (c)
test statistics of MFLY and NPEC crack signals are more than
detection threshold; these results are consistent with the first
data fusion rule and verify the effectiveness of detection method.
Furthermore, at the domains where there are no crack signals in
Fig. 6(b), all of the estimated SNRs are no more than a priori
thresholds, and the majority of those are no more than detection
thresholds, so the estimated results are determined no cracks,
which verifies the effectiveness of detection method based on
the fourth data fusion rule.
(d)
Table III, where LB is the left boundary, and RB is the right Fig. 7. Simulations in the second experiment.
boundary, shows that the region-length errors (%) of the first
crack and the second crack are both 0.0%, while the region- search domain is also effective in this experiment.
length error (%) of the third crack is -12.5%, because the SNRs The original MFLY signals including pure signal and signal
of the first two cracks are higher than that of the third one. As with WGN are shown in the left figure of Fig. 7(b), while the
to midpoint errors, their values are all small, at most up to one right figure only has WGN. The parameters of different crack
step size, i.e., 0.1, for the second and the third cracks. signals, the estimated SNRs, and are listed in Tables IV
The second experiment is prepared to evaluate the crack
and V, respectively. It can be found that the errors of amplitudes,
detection method based on the second data fusion rule. Suppose
scaling factors, and translation factors are small for MFLY
only the MFLY sensors work normally in the experiment, but
crack signals.
the NPEC sensors are all broken. Hence, the three MFLY crack
Fig. 7(c) shows distribution of test statistics in the left figure
signals are added in the experiment, and their test statistics are
are normal, and the test statistics of three cracks are more than
more than dual thresholds; however the NPEC signal is only
detection threshold; however, the results of right figure are
filled with WGN. The computed results are listed from Tables
caused by WGN, and the test statistics of three crack positions,
IV to VI, and the simulation results are shown in Fig. 7.
where , , and , are no more than
In Fig. 7(a), we find three pinnacles corresponding to three
detection threshold, so the NPEC estimated crack signals are
crack signals in the left figure, but chaotic distribution in the
not saved to present in the right figure of Fig. 7(d). Moreover,
right figure; the red dots are located in the positions of optimal
as shown in Figs. 7(c), 7(d) and Table V, only the MFLY
scaling factors step by step along the direction,
signals are used to locate all of the cracks, because
which means the golden section algorithm based on dynamic
and the of MFLY crack signals are more than those of

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

11
TIM-22-05197R1

TABLE IV
ERRORS OF AMPLITUDES, SCALING FACTORS, AND TRANSLATION FACTORS IN THE SECOND EXPERIMENT

/ / /
A a b A a b A a b
Designed value 2.000 1.500 10.000 3.000 1.000 50.000 1.000 0.800 90.000
MFLY Estimate 1.739 1.400 10.100 2.484 1.000 50.100 1.023 0.900 90.000
Error -0.261 -0.100 0.100 -0.516 0.000 0.100 0.023 0.100 0.000
NPEC Designed value 0.000 0.000 0.000

TABLE V
SNR AND PD IN THE SECOND EXPERIMENT

Crack 1 Crack 2 Crack 3

Designed SNR (dB) 1.109 - 5.186 - -1.666 -

Estimated SNR (dB) -0.664 -17.388 3.879 -22.908 -3.375 -16.294
SNR Error (dB) -1.773 -1.307 -1.709
1.000 0.166 1.000 0.032 1.000 0.133
1.000 1.000 0.999

TABLE VI (a)
ERRORS OF CRACK REGIONS IN THE SECOND EXPERIMENT

Crack 1 Crack 2 Crack 3

LB RB LB RB LB RB
Target region 2.5 17.5 45.0 55.0 86.0 94.0
Estimated region 3.0 17.0 45.0 55.0 85.5 94.5
Error 0.5 -0.5 0.0 0.0 -0.5 0.5
Midpoint error 0.0 0.0 0.0
Region-length error -6.7 % 0.0 % 12.5 %

NPEC crack signals. Above results verify the effectiveness of (b)

detection method based on the second data fusion rule.
Table VI shows that the region-length errors (%) of three
cracks are 6.7%, 0.0%, and 12.5%, respectively, because the
SNR of the second crack is highest, but lowest for the third
crack. Note that the midpoint errors of three estimated regions
are all 0.0, which means there are no location deviations of
estimated results in the second experiment.
The third experiment is implemented to evaluate the crack
detection method based on the third data fusion rule. Suppose
only the NPEC sensors work normally in the experiment, but (c)
the MFLY sensors are all broken. Thus, the three NPEC crack
signals are added in the experiment, and their test statistics are
more than dual thresholds; however the MFLY signal is only
filled with WGN. The computed results are listed from Tables
VII to IX, and the simulation results are shown in Fig. 8.
Similarly, in the left figures from Figs. 8(a) to 8(c), the results
of them affected by WGN look like chaotic and abnormal;
however, the right figures about NPEC crack signals are normal
and correct as expected.
(d)
The errors of amplitudes, scaling factors, and translation Fig. 8. Simulations in the third experiment.
factors are small for NPEC crack signals as shown in Table VII;
besides, as shown in Figs. 8(c), 8(d) and Table VIII, only the the first and the third NPEC crack signals are negative, but
NPEC signals are used to locate all of the cracks, because positive for that of the second NPEC crack signal, which
and the of NPEC crack signals are more than coincides with the designed values and verifies again that the
those of MFLY crack signals. Furthermore, the three test crack detection method can accurately identify the inner and
statistics of MFLY crack signals, in the positions of , outer-surface cracks even if the SNR is lower.
, and , are no more than detection threshold, Table IX shows that the region-length errors (%) of three
so the MFLY estimated crack signals are not saved to present cracks are 6.7%, 0.0%, and 12.5%, respectively. Meanwhile,
in the left figure of Fig. 8(d). Above results verifies the the midpoint error of the second estimated region is lowest, but
effectiveness of detection method based on the third data fusion highest for the first estimated region.
rule.
Moreover, Fig. 8(d) shows that the estimated amplitudes of

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

12
TIM-22-05197R1

TABLE VII
ERRORS OF AMPLITUDES, SCALING FACTORS, AND TRANSLATION FACTORS IN THE THIRD EXPERIMENT

/ / /
A a b A a b A a b
MFLY Designed value 0.000 0.000 0.000
Designed value -2.000 1.500 10.000 3.000 1.000 50.000 -1.000 0.800 90.000
NPEC Estimate -1.681 1.400 10.100 2.972 1.000 50.100 -1.163 0.900 90.000
Error 0.319 -0.100 0.100 -0.028 0.000 0.100 -0.163 0.100 0.000

TABLE VIII TABLE IX

SNR AND PD IN THE THIRD EXPERIMENT ERRORS OF CRACK REGIONS IN THE THIRD EXPERIMENT

Crack 1 Crack 2 Crack 3 Crack 1 Crack 2 Crack 3

LB RB LB RB LB RB
Designed SNR (dB) - 0.571 - 6.346 - -2.725 Target region 2.5 17.5 45.0 55.0 86.0 94.0
Estimated SNR (dB) -18.982 -0.963 -17.862 5.439 -18.863 -2.255 Estimated region 3.0 19.0 45.1 55.1 85.0 94.0
SNR Error (dB) -1.533 -0.907 0.470 Error 0.5 1.5 0.1 0.1 -1.0 0.0
0.107 1.000 0.099 1.000 0.068 1.000 Midpoint error 1.0 0.1 -0.5
1.000 1.000 1.000 Region-length error 6.7 % 0.0 % 12.5 %

TABLE X
AUCs GIVEN BY SIMULATIONS BASED ON INDIVIDUAL DETECTIONS AND FUSED DETECTIONS

Individual detections (a) (b) (c) (d) (e) (f)

SNR (dB) AUC SNR (dB) AUC SNR (dB) AUC SNR (dB) AUC SNR (dB) AUC SNR (dB) AUC SNR (dB) AUC
10 0.974 10 & 10 0.986 10 & -3 0.974 10 & 0 0.974 10 & 3 0.974 10 & 5 0.979 10 & 10 0.986
5 0.813 5&5 0.964 5 & -3 0.849 5&0 0.894 5&3 0.958 5&5 0.965 5 & 10 0.981
3 0.733 3&3 0.913 3 & -3 0.779 3&0 0.880 3&3 0.912 3&5 0.951 3 & 10 0.975
0 0.635 0&0 0.820 0 & -3 0.730 0&0 0.818 0&3 0.879 0&5 0.895 0 & 10 0.975
-3 0.573 -3 & -3 0.658 -3 & -3 0.656 -3 & 0 0.730 -3 & 3 0.781 -3 & 5 0.850 -3 & 10 0.974

Note that in Table X, the first data set on the far left is the
B. Evaluation for Detection Performance
AUC of individual detections, while the other data sets are the
To estimate the given by and SNR for ROC curves, AUCs of fused detections corresponding to Fig.9 (a) - (f), where
based on the detection method proposed in this paper, Monte A & B represents the fusion of two signals with SNR of A and
Carlo experiments are implemented. We design a simulating B. Similar to Fig.2, the performance comparisons between
MFLY crack signal given by (36) and a NPEC crack signal fused detections and individual detections focus on the signals
given by (37), where , , for both of measured by sensors at the same time. For instance, the 0dB &
them; thus, a priori threshold is . The time -3dB in Fig. 9(b) and Table X represents the fused detection
domain is [1, 20], with step size ; so we get performance based on the signals with SNR of 0 dB and -3 dB;
for definitional e.g., if the SNR of MFLY crack signal is 0 dB, and the SNR of
domain; while the variation region of is [0.4, 2.0] with NPEC crack signal is -3 dB meanwhile, we could compare the
step size of 0.1. The SNR sets of MFLY and NPEC crack fused detection performance of 0dB & -3dB with the individual
signals are all {-3 dB, 0 dB, 3 dB, 5 dB, 10 dB}; specifically, detection performance of signal with SNR of 0 dB or -3 dB,
the NPEC crack signal serves as reference signal. We do 2000 rather than with other SNRs, such as 10 dB or 5 dB, etc. It is
detection experiments for one result of given by and obvious that the AUC of 0dB & -3dB, i.e., 0.730, is greater than
SNR. At last, six experiment results are shown in Fig. 9. It is that of 0dB, i.e., 0.635, and that of -3dB, i.e., 0.573.
noted that the individual detection performance, given by (26) In summary, as shown in Fig. 9(a) and the first two data sets
and plotted by solid lines, are derived from the detection of in Table X, when the SNR of two independent signals is equal,
deterministic signals with known parameters; however, the the fused detection performance is all superior to their
real-world individual detection performance is probably individual detection performance; this result is consistent with
inferior to that of (26), because the parameters of are the Fig. 2(a). As for other figures and data sets shown in Fig. 9
unknown beforehand. Thus, the fused detection performance, and Table X, we could summarize two results as follows:
plotted by dash-star lines in Fig. 9, is actually compared with 1) The performance of fused detection is not weaker than
the ideal performances of individual detection, so it is more that of any corresponding individual detection, no matter
convincing to verify the proposed crack detection method. how different the SNRs between two independent signals;
Besides, the areas under the curve (AUCs) are calculated by 2) The performance of individual detection of the signal
ROC curves given by (38), and the results are listed in Table X. with higher SNR is the lower bound of that of the
(38) corresponding fused detection.
Take the Fig. 9(e) for example, the fused detection
performance of -3dB & 5dB is not weaker than the individual
where is the ith value of vector, and is the ith value
performance of 5 dB when ; as for other fusions,
of vector; N is the length of the vector and vector.
such as 0dB & 5dB, the performance of fused detection is better

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

13
TIM-22-05197R1

(a) (b) (c)

(d) (e) (f)

Fig. 9. ROC curves given by simulations based on individual detections and fused detections: (a) the SNR of two independent signals is equal; (b) the reference
signal with SNR of -3 dB; (c) the reference signal with SNR of 0 dB; (d) the reference signal with SNR of 3 dB; (e) the reference signal with SNR of 5 dB; (f)
the reference signal with SNR of 10 dB.

than that of individual detection of 0 dB or 5dB when dB and -3 dB respectively for the first scenario, with SNR of 5
. dB and -3 dB respectively for the second scenario, and so on
To sum up, the simulations above illustrate the proposed fused until with SNR of -3 dB and -3 dB for the fifth scenario, as
detection method has the better detection performance than shown in Table XI; while the noise testing samples consist of
corresponding individual detection method in any case; this noise pairs given by the signals with WGN in every scenario
feature is superior to the traditional combined detection method, subtracting the corresponding pure crack signals. After model
whose performance may be inferior to that of corresponding testing, we could calculate AUC results given by (38) for three
individual detection with higher-SNR signals. comparable methods in five scenarios, and summarize them in
Table XI.
C. Comparison The individual detection results and the proposed fused
To further illustrate the advantages of the proposed method detection results are cited from Table X, and the AUCs of fused
for crack detection in pipelines, other three state-of-the-art detection in every scenario are better than or equal to those of
detection methods, called the comparable methods here, corresponding individual detection with higher-SNR signals.
including the data and decision level fusion method [14], However, the AUCs of comparable methods are less than
adaptive data fusion based on DCNN [17], and the decision those of the proposed fused detection method in five scenarios.
level fusion using CNN [18], were used for reference and Compared with the individual detection method, the AUCs of
comparison. To this end, we temporarily transformed the comparable methods, with SNR combinations of 0 dB and -3
detection problem in this paper into binary classification dB as well as -3 dB and -3 dB, are more than those of
problem. First, we designed 6000 pairs of training samples to corresponding individual detection method with higher-SNR
train classification models, where 3000 pairs of samples of signals, because the SNRs of the MFLY and NPEC crack
them, made up of MFLY crack signals, NPEC crack signals, signals in these two scenarios are relatively close. On the
and WGN like those in Section IV-B, represent crack samples contrary, the AUCs of comparable methods in other three
with different SNR combinations, given by the sets of {-3 dB, scenarios are less than those of corresponding individual
0 dB, 3 dB, 5 dB, 10 dB}; while the other 3000 pairs of samples detection method with higher-SNR signals, because the
denoted no cracks only consist of WGN with different comparable methods do not consider the SNR differences of
variances. Second, we tested the classification models in 5 sample pairs used for data fusion, so they could be classified
scenarios, and there were 200 crack testing samples and 200 into combined detection methods. For the combined detection
noise testing samples in every scenario. Crack testing samples methods, as shown in Fig. 2(b), the signals with SNR as low as
consist of MFLY and NPEC crack signal pairs with SNR of 10 -3 dB weaken the detection performance of signals with higher

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

14
TIM-22-05197R1

TABLE XI Figs. 10(g) and 10(h), where the origins of coordinates are set
AUC RESULTS OF DIFFERENT METHODS on the downstream flange. The sampling points on the right side
Individual SNR (dB) 10 5 3 0 -3 of downstream flange are 557, with sampling interval of 2 mm;
Individual detection method 0.974 0.813 0.733 0.635 0.573 [1,557], with step size ; while
Combined SNR (dB) 10&-3 5&-3 3&-3 0&-3 -3&-3 [0.4,7.0], with step size of 0.1; is set to be 0.01;
The proposed fused detection method 0.974 0.849 0.779 0.730 0.656
Data and decision level fusion method 0.972 0.811 0.732 0.714 0.643
according to the simulation experiences in Section IV-A, a
Adaptive data fusion based on DCNN 0.969 0.806 0.730 0.699 0.637 priori threshold, i.e., , is set to 2.457, i.e., 1 / 0.407,
Decision level fusion using CNN 0.965 0.796 0.729 0.684 0.629 times the detection threshold here.
Based on the proposed detection method, 14 artificial linear
SNRs, i.e., 3 dB, 5 dB, and 10 dB; in other words, the signals cracks are all detected, as No. 1 to No. 14 shown in Figs. 10(g)
with SNR of -3 dB in the combined detections make the and 10(h), marked by red minimum bounding boxes, according
increase of PFA stronger than the increase of PD in above three to the binary image given by 0-1 crack-decision matrix in Fig.
scenarios. Consequently, the AUCs of three comparable 10(i). By calculating the locations of geometric centers of these
methods in Table XI, with SNR combinations of 10 dB and -3 bounding boxes, we could estimate the and and get
dB, 5 dB and -3 dB, as well as 3 dB and -3 dB, are inferior to estimated relative errors for these cracks, as shown in Table XII,
those of corresponding individual detection method with where the relative error of [-0.50%, 5.00%], with mean
higher-SNR signals. relative error (MRE) of 1.87%, and the relative error of
In summary, owing to the wavelet-based decision-level data [-1.04%, 0.93%], with MRE of 0.33%.
fusion, the proposed detection method has the better detection Moreover, considering the detection method has been
performance than individual detection methods and three performed on MFLY and NPEC array signals channel by
comparable methods even though the SNRs of two independent channel, we take 3 channels of 8:00, 9:00, and 10:00,
signals to be fused are quite different. corresponding to different linear-crack angles, for instances to
present the detection results, as shown in Figs. 10(g) and 10(h),
which run through the geometric centers of 14 linear cracks
D. Tests with Pull-Rig Data exactly. The detection results of these cracks are shown in Table
The ILI PIG developed by our research group is shown in Fig. XIII, and Figs. 11 to 13, where the origins of coordinates are on
10(a), and a novel differential probe circuit was designed as the right side of downstream flange with the offset of 50
shown in Fig. 10(b) [25], [35]. The probe circuit consists of 2- sampling points (0.1 m), to focus on showing the crack signals.
channel differential coils to transmit NPEC signals, 4-channel First, the results of actual oblique cracks No. 1 to No. 4, for
tri-axial Hall-effect sensors to transmit MFL signals, signal- the channel of 8:00, are shown in Fig. 11. The 3D test statistics
flow part, and data-flow part, respectively. The gap between of MFLY and NPEC signals are shown in Fig. 11(a), and there
center point of differential coils and y-axis Hall-effect sensors are 4 pinnacles corresponding to 4 crack signals in each figure.
is 11 mm. Before data analysis, two preprocessing steps should The original signals with noise are shown in Fig. 11(b);
be done. First, the measured MFLY signals are shifted in specifically, the MFLY original signal is measured by Hall-
effect sensors, with the unit of Gs, while the NPEC original
parallel by 11 mm in the positive direction to be set
signal is measured by differential coils, with the unit of mV; it
to the same measuring point as the NPEC signals; second, the
is noted that the shapes of actual linear crack signals with noise
NPEC signals should be interpolated to the same number of are similar to the wavelets we have selected in Fig. 1. The 2D
channels as the MFLY signals along the clock direction. Note test statistics of original signals are shown in Fig. 11(c): the test
that our probe circuit can collect MFLY and NPEC signals statistics of oblique cracks No. 2 to No. 4 exceed a priori
simultaneously and independently, which makes it possible to thresholds for MFLY and NPEC signals, so these oblique
realize fused detection. cracks are detected by both MFLY and NPEC signals; however,
To test the proposed crack detection method in engineering, as for the oblique crack No. 1, the NPEC test statistic is lower
the pull-rig experiments were implemented in one actual crack than a priori threshold but the MFLY test statistic is more than
pipe with outer diameter of 1016 mm. The material of the pipe a priori threshold, so the oblique crack No. 1 could only be
is X80 steel with wall thickness of 15.30 mm, and 14 artificial detected by the MFLY signal according to the crack detection
linear cracks with different angles, lengths, widths, and depths method. Fortunately, the four test statistics of NPEC crack
were machined by wire cut in the pipe wall, as shown in Figs. signals in Fig. 11(c) are all more than detection threshold, so
10(c) to 10(f) and Table XII, where DR is the relative distance the crack identifications of inner surface or outer surface by
from cracks to the downstream flange, and the clock describes NPEC crack signals are believable: as the right figure shown in
the center position of linear cracks from upstream perspective. Fig. 11(d), the crack No. 1 is the inner-surface crack, while the
Note that there are 3 inner linear cracks, i.e., No. 1, No. 5, and others are the outer-surface cracks, which coincides with the
No. 10, machined in the inner-surface of pipe wall near the designed results. After estimations, as shown in Table XIII, the
downstream flange, while the other linear cracks were of MFLY crack signals for above 4 oblique cracks are all
machined in the outer-surface of pipe wall. more than those of NPEC crack signals when , so
The running speed of the PIG was set to 8 m/s to get relative the regions of them are all determined by the estimated MFLY
lower SNRs. After a pull-rig experiment, the experimental data crack signals, as shown in Fig. 11(d), where the estimated target
were got from the disk of the PIG, and the original MFLY and regions, given by blue dash lines, are the bases to generate 0-1
NPEC array signals around the artificial cracks are shown in crack-decision matrix.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

15
TIM-22-05197R1

(a) (b)

(c) (d) (e) (f)

(g) (h) (i)

Fig. 10. Hardware platform and the pull-rig experiment: (a) the ILI PIG; (b) schematic diagram of probe circuit; (c) actual pipeline with artificial cracks; (d) the
oblique crack No. 3; (e) the transverse crack No. 8; (f) the longitudinal crack No. 12; (g) detection results of original MFLY array signals; (h) detection results of
original NPEC array signals; (i) binary image given by 0-1 crack-decision matrix.

Second, the results of actual transverse cracks No. 5 to No. 9, 13(b): the shapes of MFLY and NPEC crack signals with noise
for the channel of 9:00, are shown in Fig. 12. Benefited from are both normal. The 2D test statistics of original signals are
very tight arrangement of the PIG probes, our developed PIG is shown in Fig. 13(c): the test statistics of the longitudinal cracks
also sensitive to the transverse cracks, although the amplitudes No. 11 to No. 14 exceed a priori thresholds for MFLY and
of MFLY signals and NPEC signals for transverse cracks are NPEC signals, so these cracks are detected by both MFLY and
lower than those of oblique cracks and longitudinal cracks for NPEC signals; but the crack No. 10 could only be detected by
the same sizes. The 3D test statistics of MFLY and NPEC NPEC signal because only its test statistic exceeds a priori
signals are shown in Fig. 12(a), and there are 5 pinnacles threshold. After estimations, as shown in Table XIII, the
corresponding to 5 crack signals in each figure. The original of NPEC crack signals for No. 10 and No. 11 are more than
signals with noise are shown in Fig. 12(b): the shapes of MFLY those of MFLY crack signals when , so the regions
crack signals with noise are a little deformed, but basically of these two cracks are determined by the estimated NPEC
normal for the NPEC crack signals. The 2D test statistics of crack signals; while the situations of other cracks in Fig. 13(d)
original signals are shown in Fig. 12(c): all of the test statistics are opposite, so their regions are determined by the estimated
of the transverse cracks exceed a priori thresholds for MFLY MFLY crack signals. In addition, according to the NPEC
and NPEC signals, so all of the transverse cracks are detected estimated results in Fig. 13(d), the crack No. 10 is the inner-
by both MFLY and NPEC signals. After estimations, as shown surface crack, while the others are the outer-surface cracks,
in Table XIII, the of MFLY crack signals for above 5 which coincides with the designed results.
transverse cracks are all more than those of NPEC crack signals Furthermore, as shown in Table XIII, the results for cracks
when , so the regions of them are all determined by No. 1 to No. 14 are consistent with the results in Figs. 11 to 13.
the estimated MFLY crack signals. Furthermore, according to Besides, the NPEC test statistic of crack No. 1 and the MFLY
the NPEC estimated results in Fig. 12(d), the crack No. 5 is the test statistic of crack No. 10 are inferior to a priori thresholds,
inner-surface crack, while the others are outer-surface cracks, so the individual decisions via these independent signals are NO;
which coincides with the designed results. fortunately, the MFLY test statistic of crack No. 1 and the
Third, the results of actual longitudinal cracks No. 10 to No. NPEC test statistic of crack No. 10 are more than a priori
14, for the channel of 10:00, are shown in Fig. 13. The 3D test thresholds, and these two cracks are determined YES by fused
statistics of MFLY and NPEC signals are shown in Fig. 13(a), decisions in the end. As for other cracks in Table XIII, their
and there are 5 pinnacles corresponding to 5 crack signals in individual decisions of MFLY and NPEC signals are all YES,
each figure. The original signals with noise are shown in Fig. and the final decisions are absolutely YES.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

16
TIM-22-05197R1

(a) (b)

(c) (d)
Fig. 11. Processing results for actual oblique cracks No. 1 - No. 4.

(a) (b)

(c) (d)
Fig. 12. Processing results for actual transverse cracks No. 5 - No. 9.

(a) (b)

(c) (d)
Fig. 13. Processing results for actual longitudinal cracks No. 10 - No. 14.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

17
TIM-22-05197R1

TABLE XII
ARTIFICIAL LINEAR-CRACK INFORMATION

Crack number 1 2 3 4 5 6 7 8 9 10 11 12 13 14
DR (m) 0.200 0.400 0.600 0.800 0.200 0.400 0.600 0.800 1.000 0.200 0.400 0.600 0.800 1.000
Clock 8:00 8:00 8:00 8:00 9:00 9:00 9:00 9:00 9:00 10:00 10:00 10:00 10:00 10:00
Inner/Outer Inner Outer Outer Outer Inner Outer Outer Outer Outer Inner Outer Outer Outer Outer
Angle (°) 75 65 55 45 0 0 0 0 0 90 90 90 90 90
Length (mm) 45.0 51.9 88.7 91.0 46.0 57.8 75.1 84.4 88.0 31.1 53.8 75.5 93.9 128.8
Width (mm) 1.5 1.7 1.6 2.0 1.4 1.5 1.8 1.8 2.1 1.4 1.3 2.0 2.1 1.7
Depth (%) 26.32 43.55 53.86 100.00 26.38 45.53 57.71 79.61 100.00 12.40 17.04 48.00 58.74 100.00
(m) 0.208 0.410 0.608 0.808 0.210 0.408 0.598 0.796 1.006 0.210 0.410 0.598 0.798 1.008
7:55 7:58 8:00 8:00 9:05 9:03 9:01 8:58 8:57 10:02 10:01 10:00 10:00 10:00
error (%) 4.00 2.50 1.33 1.00 5.00 2.00 -0.33 -0.50 0.60 5.00 2.50 -0.33 -0.25 0.80
error (%) -1.04 -0.42 0.00 0.00 0.93 0.56 0.19 -0.37 -0.56 0.33 0.17 0.00 0.00 0.00

TABLE XIII
ACTUAL LINEAR-CRACK DETECTION RESULTS OF MFLY AND NPEC SIGNALS

Crack number 1 2 3 4 5 6 7 8 9 10 11 12 13 14
(dB) 1.61 11.72 15.90 25.25 11.20 16.26 16.51 20.86 29.64 -4.06 5.97 19.65 21.76 27.74
PD 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
MFLY T(x) 76.40 247.70 439.60 1372.90 19.21 34.40 35.40 58.39 160.48 34.30 85.40 422.00 544.00 1081.80
61.28 61.28 61.28 61.28 4.00 4.00 4.00 4.00 4.00 41.43 41.43 41.43 41.43 41.43
Decision YES YES YES YES YES YES YES YES YES NO YES YES YES YES
(dB) -2.56 11.17 15.77 18.40 -0.68 7.27 14.88 15.64 16.52 2.35 11.93 17.77 19.29 21.34
PD 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
NPEC T(x) 12.85 60.95 119.11 158.22 4.89 11.44 24.28 26.82 30.77 15.74 46.34 93.94 100.36 149.20
16.86 16.86 16.86 16.86 4.43 4.43 4.43 4.43 4.43 11.33 11.33 11.33 11.33 11.33
Decision NO YES YES YES YES YES YES YES YES YES YES YES YES YES
PD,C 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
Fused decision YES YES YES YES YES YES YES YES YES YES YES YES YES YES

The experiments above illustrate that for the actual high- complex shapes can also be studied to improve its applicability
speed pipeline ILI, compared with the individual detections, the in practical implementations once more real world samples
proposed crack detection method is effective for inner-surface become available.
cracks and outer-surface cracks with different sizes and angles;
moreover, the method has the good location precision, and can APPENDIX I
further improve the detection probability, in an environment
Proof: We first write the PDF from (14) as:
that constant false alarm probability is low.

V. CONCLUSION
This paper proposes a crack detection method for pipelines
using wavelet-based decision-level data fusion. The scales and (39)
locations of crack signals can be accurately determined by
DCWT; the decision-level data fusion helps to identify inner-
surface or outer-surface cracks of pipe walls and improve the Considering this as a function of , it becomes the likelihood
crack detection performance. function. Differentiating the log-likelihood function, we have:
The first group of simulations show the proposed method has
the good precision at estimations of crack signal regions and
(40)
can identify the inner and outer-surface cracks accurately with
SNR as low as -3.216 dB. The second and third groups of
experiments show the proposed method does improve upon Based on (8) and (9), setting (40) to zero produces:
individual detection methods alone, and has better detection
(17)
performance than three state-of-the-art methods. Finally, the
real world pipeline experiments verify that, in actual high-speed
pipeline ILI, the proposed method can further improve the Thus, the MLE of is given by (17).
detection probability, in an environment that constant false
alarm probability is low; besides, the crack location accuracy is APPENDIX II
very high: the MREs of crack relative distances and crack center
Proof: We first write the PDF as (39), and consider it as a
clocks are only 1.87% and 0.33% respectively.
function of and , so the MLEs of and are found by
In this paper, two independent signals were implemented in maximizing:
designing the fusion rules. Signals from more sources can be
incorporated in future to further investigate extent of possible
improvement from these rules. Cracks with more topologically

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

18
TIM-22-05197R1

(41)
(22)
over all possible , and ; maximizing (41) is equivalent to
minimize:
(42)

But according to (8), and (23)

it is not a function of and . Hence, the MLEs of and are
found by maximizing: APPENDIX IV
(43) Proof by contradiction:
At first, outside the crack regions, the value of crack test
Theoretically, (17) could be modified as follow: statistic, i.e., is zero under the non-
(44) noise condition, which is not within the scope of this question.
Substitute (9) and (44) into above (43) and we verify that the In the crack regions, if is constant, must has peak
MLEs of and are found by maximizing: values in the domain of ; here, assume that there are two
(45) scaling factors, and , corresponding to two peak values
of , when ;
or, equivalently, by maximizing over all It means that when , the measured signal is similar to
possible and . the crack signal model, i.e., or ; and when ,
the measured signal is also similar to the crack signal model.
APPENDIX III The scaling factor represents the signal period; if the
Proof: Under either hypothesis is Gaussian and since assumption above is right, then the measured crack signal must
is a linear combination of Gaussian random variables, be discontinuous or heteromorphic in one signal period when
is also Gaussian. Let and ; however, this result contradicts with the fact that crack
denote the expected value and variance of under . Then: signals are continuous and regular, and the shapes of them are
(46) similar to the shapes of crack signal models.
Thus, the assumption is wrong, and the proposition is correct.

REFERENCES
[1]
approximation approach to defect opening profile recognition in magnetic
IEEE Trans. Instrum. Meas., vol. 70, pp. 1-12, Mar.
2021.
(47) [2] -based

IEEE Trans. Instrum. Meas., vol. 70, pp. 1-10, Mar. 2021.
[3]
measurements using low-rank recovery in pipeline transportation
IEEE Trans. Instrum. Meas., vol. 69, no. 9, pp. 6776-6786,
Sept. 2020.
[4] n iterative stacking method for
(48) IEEE Trans. Instrum.
Meas., vol. 69, no. 6, pp. 3780-3788, June 2020.
[5]
characterization with eddy current testing using nonlinear-regression
IEEE Trans. Instrum.
Meas., vol. 62, no. 5, pp. 1207-1214, May 2013.
[6]
testing for wall thickness of IEEE Trans. Instrum.
Meas., vol. 69, no. 12, pp. 9766-9773, Dec. 2020.
(49) [7] V. Arjun, B. Sasi, B. P. C. Rao, C. K. Mukhopadhyay, and T. Jayakumar,
-surface
defects in stainless Sens. Actuators A, Phys., vol. 226, pp. 69-
Thus,
75, Feb. 2015.
[8]
(21) liftoff point of intersection in pulsed eddy current responses for elimination
Sens. Actuators A, Phys., vol. 251, pp. 66-74, Oct. 2016.
[9]
microwave thermography for nondestructive evaluation of surface cracks
IEEE Trans. Instrum. Meas., vol. 68, no. 2, pp. 576-
585, Feb. 2019.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in IEEE Transactions on Instrumentation and Measurement. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TIM.2023.3244211

19
TIM-22-05197R1

[10] -scan [34]

IEEE Trans. Instrum. Meas., Eng.
vol. 54, no. 1, pp. 192-199, Feb. 2005. Appl. Artif. Intell., vol. 50, pp. 201-214, Feb. 2016.
[11]R. S. Edwards [35] -
Sens. Actuators A, Phys., D defect profile from MFL signals using key physics-based parameters and
vol. 144, no. 1, pp. 222-228, Jan. 2008. NDT & E Int., vol. 103, pp. 26-38, Jan. 2019.
[12] -sensor image fusion at signal level
for improved near- NDT & E Int., vol. 71, pp. 16-
22, Jan. 2015.
[13]
dynamic magnetization in
NDT & E Int., vol. 64, pp. 7-12, Feb. 2014.
[14] D. Song, T. Ma, Y. Li and F. Xu, "Data and Decision Level Fusion-Based
Crack Detection for Compressor Blade Using Acoustic and Vibration
Signal," IEEE Sensors Journal, vol. 22, no. 12, pp. 12209-12218, 15
June15, 2022.
[15] Y. Yang et al. -image-feature-based hierarchical concrete crack
identification framework using optimized SVM multi-classifiers and D S
Remote Sens., vol. 13, no. 2, p. 240,
Jun. 2021.
[16] M. Cerrada, G. Zurita, D. Cabrera, R.-V. Sánchez, M. Artés, and C. Li,
random
Mech. Syst. Signal Process., vols. 70 71, pp. 87 103, Mar. 2016.
[17] L. Jing, -sensor data
fusion method based on deep convolutional neural networks for fault
Sensors, vol. 17, no. 2, p. 414, Feb. 2017.
[18] S. Teng, G. Chen, Z. Liu, L. Cheng, and X. Sun, -sensor and
decision-level fusion-based structural damage detection using a one-
Sensors, vol. 21, no. 12, p.
3950, Jun. 2021.
[19] Z. Zhao et al.
intelli IEEE Trans.
Instrum. Meas., vol. 70, 2021, Art. no. 3525828.
[20] Y. -k. Fu, G. -H. Yang, H. -
Diagnosis for Quality- IEEE
Trans. Ind. Electron, vol. 70, no. 1, pp. 898-906, Jan. 2023.
[21]
IEEE Trans. Ind.
Informat., vol. 17, no. 4, pp. 2675-2685, April 2021.
[22] A

IEEE Trans. Signal Inf. Process. Netw., vol. 8, pp. 528-542, 2022.
[23] oring of eddy current
NDT & E
Int., vol. 43, no. 9, pp. 687-694, July 2010.
[24]
Wavelets and Machine Learning for Unmanned Surface IEEE
Trans. Ind. Electron., vol. 68, no. 8, pp. 7464-7473, Aug. 2021.
[25]
current method for high- Sens. Actuators
A, Phys., vol. 295, pp. 244-258, May 2019.
[26] -flux
IEEE Trans. Magn.,
vol. 56, no. 6, pp. 1-15, June 2020.
[27]
nondestructive evaluation of defects in steel elements under various
Sensors, vol. 18, no. 7, pp. 2091, June 2018.
[28] perimenting with
pixel- IEEE Trans. Instrum. Meas., vol.
49, no. 5, pp. 1083-1090, Oct. 2000.
[29] -level fusion of spatially
scattered multi-modal data for nondestructive inspection of
Sensors, vol. 16, no. 1, pp. 105, Jan. 2016.
[30]
Springer Science Business Media, 2005.
[31]
Univ. Press, 1976.
[32]
NDT & E Int., vol. 71, pp.
54-60, Jan. 2015.
[33]
Englewood Cliffs, NJ: Prentice Hall, Inc., 1998.

© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: China University of Petroleum. Downloaded on February 20,2023 at 04:20:53 UTC from IEEE Xplore. Restrictions apply.