Steam Gen: Detection and Classification of Discontinuities Using Discrete Wavelet Transform and MFL Testing
Steam Gen: Detection and Classification of Discontinuities Using Discrete Wavelet Transform and MFL Testing
x
From Materials Evaluation, Vol. 76, No. 6, pp: 706-715.
Copyright © 2018 The American Society for Nondestructive Testing, Inc.
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-
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,
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
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.
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.
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.
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
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.
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
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%
Figure 10. Denoised MFL images: (a) discontinuity free; (b) crack; (c) flat-bottom hole; (d) rectangular notch.
Figure 11. Binary images of input MFL data: (a) crack; (b) flat-bottom hole; (c) rectangular notch.