ME FEATURE w

x
From Materials Evaluation, Vol. 76, No. 6, pp: 706-715.
Copyright © 2018 The American Society for Nondestructive Testing, Inc.

SteamGen by Jackson Daniel, A. Abudhahir,

K. Shankar, and Soundra Pandi

Editor’s Note: In this Feature Paper, statistics, wavelet

transform, and neural networks are used for the
detection and classification of material damages and
degradation based on images generated by MFL
testing of steam generator tubes in sodium-cooled
reactors of nuclear power plants. Although these tools

Detection and have been explored in many NDT research studies and
applications, as cited by the authors, their application
on MFL is relatively new.

Classification of
D
amage caused by the degradation of steam
generator tubes (SGTs), used in sodium-

Discontinuities cooled reactors, leads to leakage of water

or steam into the sodium coolant, which
may further lead to a high-temperature reaction. This

using Discrete paper explains how to detect and classify the disconti-
nuities that occur in the outer surface of SGTs using
the magnetic flux leakage (MFL) method. Generally,

Wavelet SGTs experience stress corrosion cracking, wear, and

mechanical damage, which can lead to catastrophic
failures. MFL is one of the nondestructive testing

Transform and
methods that can detect a localized discontinuity on
the outer surface of a SGT. The leakage flux detected
by the magnetic sensor is used to extract information
about the discontinuity. A multiphysics software

MFL Testing package is used to obtain simulated MFL images

showing discontinuities such as cracks, flat-bottom
holes, and rectangular notches. Statistical features are
extracted to detect the discontinuities in the images,
and discrete wavelet transform (DWT) features are
extracted to classify the discontinuities. Feed-forward
back propagation training and testing of neural
networks is used to classify the discontinuities. Using

706 MATERIALS EVALUATION • JUNE 2018

erator Tubes
these techniques, a 100% discontinuity detection rate aluminum plates by eddy current techniques. The
is achieved, and an average classification accuracy of processed signals from eddy current probes with giant
94.62% is obtained through the proposed algorithm. magnetoresistive (GMR) sensors were used as an
input to multilayered neural networks to classify
Introduction aluminum plate discontinuities such as holes and
These days, sodium-cooled nuclear reactors are the cracks.
most common type employed in nuclear power plants. Other researchers (Liu et al., 2013; Huda et al.,
If there are any ruptures in the SGTs, sodium will 2014) proposed a different way of detecting and char-
aggressively react with water, which leads to the struc- acterizing NDT signals using neural networks. Another
tural collapse of the heat-transfer components. So, work (Sambath et al., 2010) proposed an algorithm to
early detection of discontinuities is necessary to extract the wavelet transform features from ultrasonic
prevent the SGTs from failing. Different methods of signals and used it to classify the diverse types of
nondestructive testing for SGTs have been reported in discontinuities. Ultrasonic testing (UT) and ANNs were
the literature (Shafeek et al., 2004; de Mesquita et al., used in this algorithm to detect discontinuities in the
2004; Yoon et al., 2014). In one study, the authors materials. A classification rate of 85% was achieved.
(de Mesquita et al., 2002) developed a new classifica- Wavelet transform coefficient features play a signif-
tion technique to isolate the discontinuities in SGTs icant role in discontinuity detection and classification.
using eddy current testing, and the classification Other authors (Yang and Li, 2012) proposed an
accuracy was 85%. But, the eddy current technique automatic ultrasonic flaw signal classification model
fails to detect deeper discontinuities and is very using a wavelet transform–based feature extraction.
susceptible to magnetic permeability changes. Their proposed technique uses both discrete wavelet
Another work (Papaelias et al., 2016) compares the transforms and wavelet packet transform (WPT) as
pros and cons of different NDT techniques. They feature vectors. But the average classification results
concluded that MFL is best suited for discontinuity obtained for both SVM and ANN were less than 85%.
detection in SGTs. Different discontinuity classification Other researchers (Qiu et al., 2014) also used wavelet
algorithms have been developed by many researchers; transform features for classification and fault
however, there is very limited research available in the detection. The ANN was trained to differentiate the
literature concerning MFL-based classification of types of discontinuities found. Eddy current testing,
discontinuities with minimum dimensions. used to detect discontinuities and monitor the devel-
Artificial neural networks (ANN) have been exten- opment of preexisting defects during in-service inspec-
sively used for discontinuity detection and classifica- tions, was also successfully used to detect subsurface
tion. Other authors (Postolache et al., 2011) proposed discontinuities (Hur et al., 2014). Other authors
an algorithm to detect and classify discontinuities in (Zhang et al., 2012) presented a technique for

JUNE 2018 • MATERIALS EVALUATION 707

ME BACK TO BASICS w
x detecting damage in steam generator tubes

This technique uses 3D modeling and

simulation of the MFL imaging system to
detect the different shapes of discontinuities
combining time and frequency features from pulsed discontinuities. The proposed technique provides
eddy current signals to enhance the accuracy of satisfactory performance for both discontinuity
discontinuity classification. Even though it works detection and discontinuity classification.
better for surface discontinuities, it produced an 86%
success rate for subsurface classification. Another Materials and Techniques
work (Khodayari-Rostamabad et al., 2009) demon- This technique uses 3D modeling and simulation of
strated a technique for the detection of discontinuities the MFL imaging system to detect the different shapes
and irregularities in an oil pipe wall and the evaluation of discontinuities in ferromagnetic SGTs. Tubes with
of the severity of those discontinuities. It also an outer diameter 17.2 mm and a wall thickness of
included a kernelized technique for determining the 2.3 mm are considered in this work. The model-
depth and severity of each discontinuity. The average predicted MFL signals were obtained and the MFL
detection performance in recognizing major metal images of surface discontinuities were constructed,
discontinuities was reported as 96%. The MFL which included rectangular notches, flat-bottom holes
technique was used to automate testing on coiled (FBH), and cracks. Discontinuities ranging from 1 mm
tubing quickly and accurately in other research (Kang in length and a minimum depth of 10% wall
et al., 2012). Other authors (Zhang et al., 2013) intro- thickness, meaning 0.23 mm in depth and a width of
duced techniques that are presently used extensively 4.5 mm, were considered.
for detecting and characterizing wire rope discontinu- The constructed images were then subjected to
ities. An image processing technique was utilized to the detection and classification algorithm as per the
preprocess the MFL images obtained by hall sensors. flowchart shown in Figure 1.
The texture analysis approach for MFL images is also
suitable for feature extraction and discontinuity classi- Multiphysics Software Modeling
fication. Other authors (Carvalho et al., 2006) A finite element model of a SGT with a bobbin coil
evaluated the use of ANNs for pattern recognition of arrangement and a simulated rectangular notch was
MFL signals in weld joints of pipelines obtained by developed using multiphysics software. The model is
intelligent pigging. MFL signals were applied as inputs shown in Figure 2a.
for a neural network. An average success rate of The bobbin coil is used to magnetize the SGT. The
71.7% for the validation set was obtained. The simula- simulated defect in the tube wall results in leakage of
tion and modeling of an MFL-based discontinuity magnetic flux, which is detected by a hall sensor array.
detection system is commonly used in the literature MFL modeling provides high-resolution data of the
(Ji et al., 2009; Dutta et al., 2009; Xu et al., 2015) to entire circumference of the SGT, allowing for inclusive
test and validate the algorithm in detecting and structural evaluations of the tube wall. The simulated
classifying the discontinuities. hall sensor points are circumferentially placed near
The technique proposed in this paper demon- the inner surface of the tube wall and in proximity to
strates a detection and classification algorithm for the coil for the best sensitivity to flux variation. Figure 2b
MFL-generated images of steam generator tubes. shows the orientation of the hall sensors placed over
Statistical features and discrete wavelet transform the bobbin coil. Hall sensors are used to obtain the
features are extracted, and the back-propagation magnetic flux density in the scanned portion of the
artificial neural network is used to classify the SGT. The defective regions produce leakage flux.

708 MATERIALS EVALUATION • JUNE 2018

The hall sensor has been placed with a liftoff of
0.1 mm. A total of 64 sensors are placed circumferen-
tially and 51 movements are provided to inspect the
portion (50 mm) of the tube’s 100 mm length as Magnetic flux leakage (MFL) image data
shown in Figure 2a. The magnetic flux leakage dataset acquisition using multiphysics software
of 64 sensors × 51 movements (each movement of
1 mm) was obtained. Statistical features are calcu- 51 × 64 MFL data set acquisition
lated from the 64 × 51 dataset to find whether a from 100 mm SGT
discontinuity is present or not. The dataset from the
64 × 51 data matrix represents magnetic flux value in
Noise removal using daubechies 4
gauss units. Then, the modality transformation is wavelet filter
carried out to convert the raw MFL data from the hall
sensors into images. These values are transferred
linearly into a grayscale color map, in which each pixel Statistical features Construct MFL
represents a gray value proportional to the amplitude extraction images
of the magnetic flux density sensed by the hall effect
sensor, and the brighter regions correspond to
Convert grayscale
leakage flux arising from the discontinuities. The to binary image
minimum depth of 10% of wall thickness has been Discontinuity Yes
considered. The images obtained using MFL may detection
contain noise due to permeability variations in the 2D discrete wavelet
transform
material. These gaussian noises are reduced using a No
wavelet filter. A daubechies 4 wavelet filter will
successfully filter out the noises present in the signal No discontinuity DWT-based feature
with good signal-to-noise ratio. extraction

Discontinuity Detection Feed-forward back propagation

Discontinuities were detected in the filtered MFL data neural network training and testing
by extracting statistical features. This takes care of for discontinuity classification
discontinuity variations to some extent. It can be used
for reducing the size of the feature set and providing
high speed and low complexity. The major statistical Flat-bottom hole Crack Rectangular notch
features used for discontinuity detection are
mentioned in Equations 1 through 5. In this work,
four statistical features are extracted from the MFL
data, which are obtained from the four classes of
Figure 1. Flow chart for the proposed discontinuity classification technique.
discontinuities.

Sensors
Modeled
SGT 10 mm Mild steel
Coil
100 mm
Modeled
bobbin
and coil
Artificial arrangement
rectangular
2.3 mm notch
30 mm
50 mm
(a) 12.6 mm (b)

Figure 2. Model of steam generator tube: (a) 3D geometry of tube and bobbin coil arrangement; (b) orientation of hall sensors on bobbin coil.

JUNE 2018 • MATERIALS EVALUATION 709

ME BACK TO BASICS w
x detecting damage in steam generator tubes

The given MFL data are converted into

grayscale images. The binary image reveals the
shape rather than the grayscale images.
Mean Standard Deviation
Mean is the average value of the gauss values present The standard deviation is a measure of how spread
in the MFL data, out the numbers are. It is the square root of the
1  n variance:
(1) Mean, µ =   ∑ Xi
 n  i =1 Standard deviation, σ
1 n
∑ ( X − µ)
2
(5) =
where n − 1 i =1 i
Xi is the gauss value of ith data,
n is the total number of data points in the MFL The output in the discontinuity detection
data matrix. algorithm, output = discontinuity, results in
“discontinuity” based on the expressions provided
The probability of the leakage flux value i is in Equation 6,
denoted by  If mean value ≥ Tm , 
n K −1  
(2) Pi = i , Pi ≥ 0, and ∑ Pi = 1  If maximum MFL ≥ Tmax , 
N i= 0 (6)  
 If standard deviation ≥ Tsd ,
where  If variance ≥ T 
v
N = n1 + n2 + … + nK,
the flux data is represented by K gauss, meaning and “discontinuity free” based on the expression in
0,1, … K-1. K is assumed to be 50 gauss. Equation 7,
 If mean value ≤ Tm , 
 
 If maximum MFL ≤ Tmax 
Variance (7)  
The variance is defined as the sum of square  If standard deviation ≤ Tsd ,
 If variance ≤ T 
distances of each term in the distribution from v

the mean, divided by the number of terms in the

distribution: where
 1  n Tm, Tmax, Tsd, and Tv are the threshold values for
 ∑ (Xi − µ)
2
(3) Variance, σ 2 =  mean, maximum value, standard deviation,
 n − 1  i=1
and variance, respectively.

Maximum Amplitude The threshold values have been fixed just above
The maximum amplitude is the maximum gauss value the values obtained for a “discontinuity free” MFL
of the MFL data matrix, image output.

(4) max (Xi ) Discontinuity Segmentation

If the discontinuity present in the MFL data is
where confirmed through the above process, then the nature
i =1,2……. n of the discontinuity is recognized by classifying its
shape. Otsu’s algorithm is a better threshold-based
segmentation method for the sensor images (Zhang et
al., 2012). The given MFL data are converted into

710 MATERIALS EVALUATION • JUNE 2018

grayscale images. The binary image reveals the shape and effective representation of low-frequency image
rather than the grayscale images. Hence, the grayscale backgrounds as well as high-frequency edges of the
images are segmented by Otsu’s segmentation discontinuity image. As most of the image information
technique, which chooses the threshold (t) to is stored in the approximation coefficients, it offers a
minimize the weighted within-class variance as better estimate of the discontinuity’s properties. The
depicted in Equations 8 through 11. 2D discrete wavelet transform (as shown in Figure 3)
Within class variance, leads to a decomposition of approximation coeffi-
cients at level j in four components. It computes the
(8) σ w 2 (t ) = q1 (t) σ 12 (t) + q 2 (t) σ 22 (t ) approximation coefficients matrix cA and details coeffi-
cients matrices such as horizontal (cDh), vertical (cDv),
probabilities of the classes are estimated as, and diagonal (cDd), respectively.
t
The approximations coefficients are used to extract
q1 (t) = ∑ p ( i) , and the wavelet features such as mean, variance, mean of
i= 0
(9) K −1
energy, maximum amplitude, minimum amplitude, and
q 2 (t ) = ∑ p ( i ) maximum energy. The six extracted feature sets for the
i= t+1
three types of discontinuity images are applied to the
means of the classes are, input of the feed-forward back propagation neural
network. The schematic representation of the designed
t i ( pi )
µ 1 (t ) = ∑ , and neural network is shown in Figure 4. The features are
i =0 q1 (t) separated into two groups: training and testing groups.
(10) K −1 i ( pi )
µ 2 (t) = ∑
i =t +1 q2 (t )
Rows Columns
and individual class variances are,
cAj L 2 1 L 1 2 cAj+1
t P ( i)
σ 1 (t) = ∑ ( i − µ1 t)
2 2
and
i =1 q1 (t) H 1 2 cDj+1h
(11)
N P ( i)
σ 2 2 (t) = ∑ ( i −µ 2 t)
2

i =t +1 q 2 (t)
H 2 1 L 1 2 cDj+1v

where
H 1 2 cDj+1d
t is the optimal threshold value,
P( i ) is the probability of occurrence of the ith
Down sample Down sample
image pixel. columns rows

For images, K represents the gray level of the Figure 3. Wavelet decomposition for two-dimensional images, where L is low-pass
filter; H is high-pass filter; cA and cD are the approximation and detailed
image, meaning 0 to 255. q1(t) are the probabilities of coefficients, respectively; cDh, cDv, and cDd are the horizontal, vertical, and
the foreground and background classes, respectively. diagonal coefficients, respectively; and j is the level of decomposition.
Based on the threshold (t) value, the images have
been segmented as foreground (discontinuity) and
background (discontinuity free) regions. The mean, Input layer Hidden layer Output layer
probability, and variances are computed for all t
values from 1 to 256, and the best threshold t value
that produces minimum sw2(t) is chosen to segment
the image. Based on the threshold t value, the images
X1
are segmented as foreground (discontinuity) and back-
ground (discontinuity free) regions.
X2
Discontinuity Classification with 2D Discrete O
Wavelet Transform X3
The wavelet transform coefficients are now computed
for the segmented image. The main feature of the X6
wavelet transform is its ability to condense most of the
image’s energy into a few transformation coefficients Figure 4. Schematic representation of feed-forward neural network.

JUNE 2018 • MATERIALS EVALUATION 711

ME BACK TO BASICS w
x detecting damage in steam generator tubes

Forty images from each discontinuity (meaning 120 total Results and Discussion
images) are used to train the network. A total number of The input data are obtained using the MFL method by
42 images are used to test the algorithm. modeling the MFL system using the multiphysics
software for different discontinuity dimensions and

(a) (b) (c) (d)

Figure 5. Grayscale images of input MFL data: (a) discontinuity free; (b) crack; (c) flat-bottom hole; (d) rectangular notch.

22 22 22

22
60 22 22 22 22 22
22

22

22
Circumferential angle (degree) 22 22

22
36

22
22 22 22

22
22
50
Hall output voltage (V)

22
22
22
34 22

22
22
22

22

22

22
32
30 40 22

22
22
22 22
22

28 22
22
22

26

24 262333803284
30 22

24
24

22
22

22
20
22
22

20
22

22 22
22

80

22
22

22
22
22 22
60 60 22

22
50 22 22 22
22
Cir 40 10 22 22
c 40 22 2 22
ang umfer 20 30 ) 22 22 2

22
20 (mm 22
le ( ent nsor
22

22

22
10 22
22
e
22

deg ial 0 0 of s 22 22 22 22 22
ree tion
Posi
5 10 15 20 25 30 35 40 45 50
(a) ) (b) Position of the sensor (mm)

Figure 6. MFL image data: (a) raw signals of hall sensor array; (b) contour plot of the noised MFL data.

22 22
60
Circumferential angle (degree)

36 22
Hall output voltage (V)

50 22
34
32
22

30 40
28
22

26
30 33642
24 3308
24

2
264
22 2
22

20 20
80 22
60 60
Cir 50
c 40 40 10
ang umfer 20 30 mm)
le ( ent 10
20
s ensor ( 22
deg ial 0 0 of
tion
22 22
ree
) Posi 5 10 15 20 25 30 35 40 45 50
(a) (b) Position of the sensor (mm)

Figure 7. MFL image data: (a) denoised MFL data; (b) contour plot of the denoised MFL data.

712 MATERIALS EVALUATION • JUNE 2018

diverse types of discontinuities. The modeled MFL MFL values of the flat-bottom hole discontinuity for the
image dataset will be of 64 × 51 size. The data 16 samples and threshold value. The variation in mean
obtained through the multiphysics software are and maximum values is due to the diverse discontinuity
converted into grayscale images, as shown in Figure 5. dimensions (length, width, and depth). Similarly, the
The gaussian noise due to permeability variations in other two statistical features are analyzed, and the
the material is removed using the daubechies 4 threshold has been fixed for the four features. The total
wavelet filter. The noisy and denoised images with 120 images are tested with the threshold, and the
their contour plots are shown in Figure 6 and Figure 7, values of the features of all the images with discontinu-
respectively. ities are above the defined threshold values. Therefore,
Figure 8 and Figure 9 show the maximum and mean 100% discontinuity detection accuracy has been

47.7788

47.653
60

44.6903
43.7626
Maximum MFL value (Gauss)

37.2549
36.2939
50

35.0144
35.1358

37.069
29.6059
28.0267
27.2518

27.7854

28.1853
25.8221

40
24.7368

30

20

10 Discontinuity Threshold value

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
MFL image data (No.)

Figure 8. Maximum MFL data plot for the sample discontinuity images and threshold value.

22.1106
22.0756

22.15
22.0466

22.0403
22.0358
22.0221

22.1
21.9954
21.9869
Mean MFL value (Gauss)

21.9783
21.9557

22.05
21.9449

21.9299
21.9339

21.9243

21.9099

22
21.8973

21.95
21.9
21.85
21.8
21.75 Discontinuity Threshold value
21.7
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
MFL image data (No.)

Figure 9. Mean MFL data plot for the sample discontinuity images and threshold value.

TABLE 1
Success and error rates of the proposed algorithm
Discontinuity Training Testing
Success rate Error rate Success rate Error rate
Discontinuity free 100% 0% 100% 0%
Crack 100% 0% 95.23% 4.76%
Flat-bottom hole 100% 0% 90.40% 9.60%
Rectangular notch 100% 0% 92.85% 7.15%

JUNE 2018 • MATERIALS EVALUATION 713

ME FEATURE w
x detecting damage in steam generator tubes

(a) (b) (c) (d)

Figure 10. Denoised MFL images: (a) discontinuity free; (b) crack; (c) flat-bottom hole; (d) rectangular notch.

(a) (b) (c)

Figure 11. Binary images of input MFL data: (a) crack; (b) flat-bottom hole; (c) rectangular notch.

achieved, as shown in Table 1. Once the discontinuity Conclusion

is detected, the shape classification algorithm will find In this paper, a technique for detecting discontinuities
its exact shape. Since image-based classification in SGTs using statistical features and thresholding is
produces better accuracy, the given MFL data are explained. The classification of discontinuities was
converted into grayscale images. Figure 10 shows the done using wavelet features and neural networks.
grayscale images of the denoised images of several Features were extracted from the input MFL data after
types of discontinuities. suitable preprocessing. The statistical features and
The grayscale images are converted into binary wavelet features were used as a feature set. The
images by thresholding. The constructed binary feature vector was trained using a feed-forward neural
images are shown in Figure 11. network. Back propagation learning was used for this
The 2D discrete wavelet transform is performed network. A total of 42 testing images were used to
over the obtained images to reduce the data without evaluate the performance of the trained neural
losses. From the transformed images, the features network. The neural network was designed and
are computed for all the discontinuity images. trained with extracted wavelet features. The data
Similarly, the wavelet features are extracted from all features were given as input for the neural networks,
120 images. These transformed wavelet features act and the overall classification accuracy is computed.
as the input feature vector for the back propagation As the statistical features for discontinuity and
neural network. The performance of the algorithm is discontinuity-free images are clearly distinguished, a
very good for the detection and classification of 100% discontinuity detection rate is achieved. The
discontinuities with very small dimensions. As the classification results for crack, flat-bottom hole, and
MFL value clearly differentiates the discontinuity rectangular-notch type discontinuities are obtained
region from the discontinuity-free region, a proper with minimum error percentage. This can be further
threshold value clearly predicts discontinuities with reduced by utilizing more MFL data. The average
100% detection accuracy. classification accuracy obtained is 94.62%. w x

714 MATERIALS EVALUATION • JUNE 2018

AUTHORS Khodayari-Rostamabad, Ahmad, James P. Reilly, Natalia K.
Nikolova, James R. Hare, and Sabir Pasha, 2009, “Machine
Jackson Daniel: Department of Electronics and Instrumenta- Learning Techniques for the Analysis of Magnetic Flux
tion Engineering, National Engineering College, Kovilpatti, Leakage Images in Pipeline Inspection,” IEEE Transaction
Tamil Nadu, PIN code: 628503, India; email on Magnetics, Vol. 45, No. 8, pp. 3073–3084.
[email protected]
Liu, Shudong, Chunling Du, Jianqiang Mou, Landong Martu,
A. Abudhahir: Department of Electrical and Electronics Engi- Jingliang Zhang, and F.L. Lewis, 2013, “Diagnosis of Struc-
neering, Vel Tech Dr.RR &Dr.SR University, Avadi, Chennai, tural Cracks using Wavelet Transform and Neural Networks,”
Tamil Nadu, PIN code: 600062, India NDT&E International, Vol. 54, pp. 9–18.
K. Shankar: Department of Electronics and Instrumentation Papaelias, Mayorkinos, Liang Cheng, Maria Kogia, Abbas
Engineering, National Engineering College, Kovilpatti, Tamil Mohimi, Vassilios Kappatos, Cem Selcuk, Louis Constan-
Nadu, PIN code: 628503, India tinou, Carlos Quiterio Gómez Muñoz, Fausto Pedro Garcia
Soundra Pandi: Department of Electronics and Instrumenta- Marquez, and Tat-Hean Gan, 2016, “Inspection and Struc-
tion Engineering, National Engineering College, Kovilpatti, tural Health Monitoring Techniques for Concentrated Solar
Tamil Nadu, PIN code: 628503, India Power Plants,” Renewable Energy, Vol. 85, pp. 1178–1191.
Postolache, O., H. Geirinhas Ramos, and A. Lopes Ribeiro,
REFERENCES 2011, “Detection and Characterization of Defects using
Carvalho, A.A., J.M.A. Rebello, L.V.S. Sagrilo, C.S. Camerini, GMR Probes and Artificial Neural Networks,” Computer
and I.V.J. Miranda, 2006, “MFL Signals and Artificial Neural Standards & Interfaces, Vol. 33, No. 2, pp. 191–200.
Networks Applied to Detection and Classification of Pipe Qiu, Xuan Bing, Lu Lu Liu, Chuan Liang Li, Ji Lin Wei, Ying Fa
Weld Defects,” NDT&E International, Vol. 39, No. 8, Wu, and Xiao Chao Cui, 2014, “Defect Classification by
pp. 661–667. Pulsed Eddy-Current Technique Based on Power Spectral
de Mesquita, Roberto N., Daniel K.S. Ting, Eduardo L.L. Density Analysis Combined with Wavelet Transform,” IEEE
Cabral, and Belle R. Upadhyaya, 2004, “Classification of Transactions on Magnetics, Vol. 50, No. 9, pp. 1–8.
Steam Generator Tube Defects for Real-Time Applications Sambath, S., P. Nagaraj, and N. Selvakumar, 2010, “Auto-
Using Eddy Current Test Data and Self-Organizing Maps,” matic Defect Classification in Ultrasonic NDT Using Artificial
Real-Time System, Vol. 27, No. 1, pp. 49–70. Intelligence,” Journal of Nondestructive Evaluation, Vol. 30,
de Mesquita, Roberto Navarro, Daniel Kao Sun Ting, No. 1, pp. 10–28.
Eduardo Lobo L. Cabral, Luis A. Negro M. Lopez, and Belle Shafeek, H.I., E.S. Gadelmawla, A.A. Abdel-Shafy, and I.M.
R. Upadhyaya, 2002, “Development of a System for Moni- Elewa, 2004, “Automatic Inspection of Gas Pipeline
toring and Diagnosis of Steam Generator Tubes using Artifi- Welding Defects using an Expert Vision System,” NDT&E
cial Intelligence Techniques on Eddy Current Test Signals,” International, Vol. 37, No. 4, pp. 301–307.
INAC 2002: International Nuclear Atlantic Conference; 13. Xu, Z.H., X.M. Zha, H.G. Chen, Y.H. Sun, and M.Z. Long,
Brazilian National Meeting on Reactor Physics and Thermal 2015, “A Simulation Study for Locating Defects in Tubes
Hydraulics; 6. Brazilian National Meeting on Nuclear Appli- using the Weak MFL Signal based on the Multi-channel
cations, Rio de Janeiro, Brazil. Correlation Technique,” Insight, Vol. 57, No. 9,
Dutta, Sushant M., Fathi H. Ghorbel, and Roderic K. Stanley, pp. 518–527.
2009, “Simulation and Analysis of 3-D Magnetic Flux Yang, Peng, and Qiufeng Li, 2012, “Wavelet Transform-
Leakage,” IEEE Transactions on Magnetics, Vol. 45, No. 4, based Feature Extraction for Ultrasonic Flaw Signal Classifi-
pp. 1966–1972. cation,” Neural Computing Applications, Vol. 24, Nos. 3–4,
Huda, A.S.N., S. Taib, K.H. Ghazali, and M.S. Jadin, 2014, pp. 817–826.
“A New Thermographic NDT for Condition Monitoring of Yoon, Byungsik, Yongsik Kim, and Seunghan Yang, 2014,
Electrical Components using ANN with Confidence Level “An Experimental Result on Flaw Sizing of Steam Generator
Analysis,” ISA Transactions, Vol. 53, No. 3, pp. 717–724. Tube at Nuclear Power Plants using Ultrasonic Testing,”
Hur, Do Haeng, Myung Sik Choi, Hee-Sang Shim, Deok Hyun Journal of Nuclear Science and Technology, Vol. 49, No. 7,
Lee, and One Yoo, 2014, “Influence of Signal-to-Noise Ratio pp. 760–767.
on Eddy Current Signals of Cracks in Steam Generator Zhang, Donglai, Min Zhao, Zhihui Zhou, and Shimin Pan,
Tubes,” Nuclear Engineering and Technology, Vol. 46, 2013, “Characterization of Wire Rope Defects with Gray
No. 6, pp. 883–888. Level Co-Occurrence Matrix of Magnetic Flux Leakage
Ji, Fengzhu, Changlong Wang, Shiyu Sun, and Weiguo Wang, Images,” Journal of Nondestructive Evaluation, Vol. 32,
2009, “Application of 3-D FEM in the Simulation Analysis No. 1, pp. 37–43.
for MFL Signals,” Insight, Vol. 51, No. 1, pp. 32–35. Zhang, Qi, Tian-lu Chen, Guang Yang, and Li Liu, 2012,
Kang, Yihua, Jianbo Wu, and Yanhua Sun, 2012, “The Use “Time and Frequency Domain Feature Fusion for Defect
of Magnetic Flux Leakage Testing Method and Apparatus Classification Based on Pulsed Eddy Current NDT,” Research
for Steel Pipe,” Materials Evaluation, Vol. 70, No. 7, in Nondestructive Evaluation, Vol. 23, No. 3, pp. 171–182.
pp. 821–827.

JUNE 2018 • MATERIALS EVALUATION 715