Instrumentation & Measurement Magazine 25-2 - 74

is fusion-bonded epoxy
powder anti-corrosion,
rigid polyurethane foam
insulation, and a high-density
plastic protective layer.
The designed throughput
is 290×104 t/a. The delivery
oil is 90 octane gasoline.
The magnetic flux leakage
detector is shown in Fig. 5.
The experiment collects
four working conditions:
normal, weld, weld defect,
and defect. The actual
working conditions inTable
1 - Network structure of the CNN model
Number of layers
Input layer
Convolution layer 1
Pooling layer 1
Convolution layer 2
Pooling layer 2
Fully connected layer
SoftMax layer
clude weld data, weld defect data, defect data and normal
data. In each working condition, 2000 sample points are continuously
collected as a sample, and the samples are repeated
2000 times. The final data set contains 20,000 samples, and each
sample contains 10,000 data points. Then, the time-frequency
graph is converted. Each working condition contains 2,000 picture
samples with a size of 36×36. Time-frequency pictures of
different defect types are shown in Fig. 6.
Because of different pipeline materials, pipeline weld size
and defect size will affect the effect of magnetic flux leakage
signal and the corresponding calculated time-frequency diagram.
However, this method can distinguish between pipeline
weld defect, weld, defect and a normal condition.
Seventy percent of the pictures are selected as the training
set, and the remaining 30% of the samples are selected as the
test set to verify the performance of the model. Then, 14,000
picture samples with a size of 36×36 are used as the training
set, and the remaining 6000 picture samples with a size of
36×36 are used as the test set.
According to the test of input field pictures with a size of
36×36, in which the convolution kernel of the pooling layer
and convolution layer in the CNN model is in the form of matrix,
the output of each layer is also matrix. The output of CNN
model is divided into four types of working conditions, such
as weld defect, weld, defect and normal working condition.
The CNN model structure established with Table 1 can obtain
a good recognition effect.
The changing trend of the loss function of the MFL diagnosis
model without data augmentation is as shown in Fig. 7
where the loss function of the training set finally converges to
0, while the loss function of the test set oscillates at approximately
0.25, indicating that the model has overfitted due to
insufficient training data.
The changing trend of the loss function of the fault diagnosis
model after the proposed data augmentation based on
ACGANs is shown in Fig. 8 where the loss function value converges
to approximately 0 on both the training set and the test
set, showing that the overfitting phenomenon disappears after
data augmentation.
The accuracy rate of the recognition model in the case of
different numbers of datasets is shown in Table 2. The model
74
3×3×1
2×2
3×3×16
2×2
4
16
16
32
32
1
1
2
1
2
Convolution
kernel size
Number of
convolution
kernels
Convolution
step
Output size
36×36×1
36×36×16
18×18×16
18×18×32
9×9×32
2592
4
Fig. 7. Loss function values before data augmentation.
Fig. 8. Loss function values after data augmentation.
IEEE Instrumentation & Measurement Magazine
April 2022
Instrumentation & Measurement Magazine 25-2

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 25-2
