Hindawi

Journal of Electrical and Computer Engineering

Volume 2023, Article ID 9948438, 10 pages
https://doi.org/10.1155/2023/9948438

Research Article
A GIS Partial Discharge Pattern Recognition Method Based on
Improved CBAM-ResNet

Di Hu ,1 Zhong Chen ,1 Wei Yang ,1 Taiyun Zhu ,1 Yanguo Ke ,1 and Kaiyang Yin 2

1
State Grid Anhui Electric Power Research Institute, Hefei 230601, Anhui, China
2
School of Electrical and Mechanical Engineering, Pingdingshan University, Pingdingshan 467000, China

Correspondence should be addressed to Kaiyang Yin; [email protected]

Received 24 February 2023; Revised 27 November 2023; Accepted 5 December 2023; Published 12 December 2023

Academic Editor: Gurvinder S. Virk

Copyright © 2023 Di Hu et al. Tis is an open access article distributed under the Creative Commons Attribution License, which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Diferent types of partial discharge (PD) cause diferent damages to gas-insulated substation (GIS), so it is very important to
correctly identify the type of PD for evaluating the GIS insulation condition. Te traditional PD pattern recognition algorithm has
the limitations of low recognition accuracy and slow recognition speed in engineering applications. To efectively diagnose the GIS
PD type and safeguard the safe and reliable operation of the distribution network, a GIS PD method based on improved CBAM-
ResNet was proposed in this paper. And the improved CBAM-ResNet takes advantage of the residual neural network and
attention mechanism. In particular, the channel attention module and the spatial attention module are connected in parallel in the
improved CBAM. Te experimental results showed that the GIS PD pattern recognition method proposed herein has a rec-
ognition rate of 93.58%, 95.00%, 93.55%, and 93.88% against the four PD types. Compared with the traditional PD pattern
recognition algorithm, the algorithm has the advantages of a lightweight model and more accurate recognition results, which carry
better engineering application values.

1. Introduction Te PD signal has obvious time-frequency-domain

characteristics, and its phase-resolved patterns of partial
GIS is widely used in power systems thanks to its advantages discharge (PRPD) can intuitively represent the relationship
of good insulation, high reliability, and small physical di- between the power frequency phase φ corresponding to the
mensions [1]. However, when the power system works in partial discharge pulse, the discharge amount q, and the
a complex environment of high temperature and high number of discharges n [6, 7]. Te PRPD pattern contains
pressure, or in the process of GIS manufacturing, some the intrinsic information of insulation defect types, which
safety hazards are inevitable, such as dust, metal tips, and air makes the identifcation of partial discharge PRPD pattern
gaps [2]. When there are insulation defects inside the GIS, an important method to diagnose defect types [8]. Li et al. [9]
PD will generate ultrahigh-frequency (UHF) signals [3]. PD frst extracted the statistical features in the PRPD pattern and
causes insulation deterioration of the GIS in the substation then used a probabilistic neural network to analyze the
and also leads to aggravated insulation deterioration. If not statistical features. Sha and Liang [10] established diferent
handled in time, it will eventually lead to insulation feature subsets based on the statistical features of the PRPD
breakdown of the GIS. Since diferent insulation fault defects pattern and eliminated redundant features through a prob-
require corresponding treatment methods, it is necessary to abilistic neural network algorithm, with a recognition ac-
accurately and efectively judge the type of equipment curacy of about 90% against partial discharge types. Yan
insulation faults and then take measures to avoid equipment et al. [11] extracted the time-frequency distribution image
insulation deterioration or even breakdown, which is an features of partial discharge and input them into SVM for
important technical means to safeguard the stable operation classifcation, but this method of image feature extraction
of substations and power systems [4, 5]. requires various artifcial transformations, which seriously
2 Journal of Electrical and Computer Engineering

reduces the efciency of partial discharge pattern pattern recognition. Section 3 reports the experimental
recognition. results with discussions. Te paper is concluded in Section 4.
With the development of deep learning and computer
hardware, convolutional neural networks (CNN) have been 2. Proposed GIS PD Pattern
developed rapidly. Since CNN can automatically extract data Recognition Method
features and get rid of the dependence on manual feature
extraction and expert experience, it has the potential to im- Te common GIS partial discharge modes include four
prove the efciency of GIS partial discharge pattern recognition types: tip discharge, free particle discharge, creeping dis-
[12]. In recent years, more and more deep-learning methods charge, and suspended electrode discharge [17]. Te PRPD
have been applied in the feld of partial discharge pattern image of various types of partial discharges has distinctive
recognition. Zhang et al. [13] used a deep belief network for GIS characteristic diferences. Te proposed GIS PD pattern
partial discharge pattern recognition and compared the rec- recognition method is used for the PRPD image feature and
ognition results with SVM and BPNN, and the overall rec- the improved CNN to improve the PD-type recognition
ognition rate was signifcantly improved. Yin et al. [14] used the accuracy. In particular, based on the residual neural net-
PRPD image multifeature information fusion method for PD work, the attention module is added in this work, and then,
pattern recognition, and the online sequential-extreme learning an improved convolutional block attention module residual
machine (OS-ELM) algorithm for PD pattern recognition was network (improved CBAM-ResNet) is formed as the im-
proposed by Zhang et al. [15]. Tese methods have achieved proved CNN for GIS PD-type recognition. Following an
good recognition results, which prove the efectiveness of deep- overall overview of the proposed GIS PD pattern method,
learning algorithms in the feld of PD pattern recognition. the three key components of the recognition method, in-
However, the improvement of the performance of the CNN cluding the PRPD image, the improved CBAM-ResNet
model depends more on the deep network structure, but as the structure, and the GIS PD pattern recognition, are de-
number of network layers continues to increase, the perfor- tailed in this section.
mance decreases instead [16]. At this time, the network leads to
degradation problems such as information loss and gradient
2.1. Method Overview. Te overall architecture of the pro-
disappearance. In response to this problem, there have been
posed GIS PD pattern recognition method based on im-
techniques including improving the activation function and
proved CBAM-ResNet is illustrated in Figure 1, and the
using batch regularization to improve network performance,
network framework has three key components, including
but they have not been completely resolved. Te emergence of
the PRPD image input layer, feature extraction layer, and
the residual neural network has solved the degradation
classifcation output layer, which can realize the detection
problem caused by the increase in the number of network
from the PRPD image to the PD type. In particular, the
layers to a certain extent.
PRPD image is a source of information for the input layer,
Tis paper proposes a GIS PD pattern recognition
the two residual blocks, and the improved residual block as
method based on improved CBAM-ResNet, which takes the
the feature extraction layer, and the classifcation output
advantage of residual neural networks and convolutional
layer consists of the global average pooling operation and the
block attention modules. Te network framework has three
softmax classifer.
key components, including the PRPD image input layer,
Te feature extraction layer uses ResNet34 as the
feature extraction layer, and classifcation output layer. In
backbone network. ResNet34 contains 5 stages, that is, 5
particular, the two attention mechanisms of the channel
convolution stages with diferent parameters and features.
attention module and the spatial attention module are
Te residual network is used in the feature extraction layer
connected in parallel to solve the interference caused by the
to alleviate network performance degradation phenomena
cascade of the two attention mechanisms. Te network
such as gradient explosion or gradient dispersion caused
structure is simple and can obtain detailed information on
by the increase in network depth, to speed up model
the target, reducing the redundant feature information
convergence, and to improve model training efects.
generated in feature extraction. Each type of PD has a dif-
Furthermore, the CBAM attention mechanism is in-
ferent form of PRPD image which is selected as the input
troduced in the residual blocks of stage 3, stage 4, and
layer for PD-type recognition. To prove the efectiveness of
stage 5. It seeks to improve the feature extraction ability of
the algorithm, comparative experimentation was completed.
the proposed model, especially the extraction ability of
Te novelty and contribution of this work are threefold: (1)
low-contrast areas information, so as to improve the
proposing an improved convolutional block attention
detection accuracy of PD type. In the classifcation output
module residual network (improved CBAM-ResNet) for GIS
layer, the global mean pooling is used to fuse the image
PD-type recognition, (2) developing a “parallel connection”
features, which compresses the output feature dimension,
structure CBAM to improve the efciency of extracting
and the softmax function is used to calculate the classi-
efective information from PRPD images, and (3) applying
fcation result.
the proposed method to the GIS PD experimental platform
and engineering project for method verifcation and
demonstration. 2.2. PRPD Image. Te PRPD image is a commonly used
Te rest of the paper is organized as follows. Section 2 representation of PD information in the feld of GIS PD
details the proposed improved CBAM-ResNet for GIS PD pattern recognition, which mainly describes the relationship
Journal of Electrical and Computer Engineering 3

Channel Refined
Input Attention Feature
Feature Module

Spatial
Attention
-15 Module
Input
-27 Feature
PD type 1
-39

-51 PD type 2
Improved Improved Improved
-63 Residual Residual
Residual Residual Residual Global PD type 3
Block Block softmax
-75 Block Block Block Mean Pooling
0 90 180 270 360 PD type 4
Stage 1 Stage 2 Stage 3 Stage 4 Stage 5

PRPD Image Input Layer Feature Extraction Layer Classification output layer

Figure 1: Te overall architecture of the proposed GIS PD pattern recognition method based on improved CBAM-ResNet.

between the number of partial discharges n, the discharge formed for the GIS PD type recognition. ResNet34 is used as
quantity q, and the corresponding discharge phase φ [18]. In the backbone network in the proposed improved CBAM-
this paper, the PRPD image is a two-dimensional spectrum ResNet. Te fve convolutional stages with diferent pa-
constructed by continuously sampling 1 min PD signals, rameters are used to construct the feature extraction layer to
where the abscissa represents the discharge phase φ, the extract the PD image feature. Ten, the classifcation output
ordinate represents the discharge quantity q, and the layer is constructed using the global averaging pooling and
number of discharges n is mapped to the color space. Each softmax function to calculate the recognition results. In the
typical partial discharge has diferent discharge character- feature extraction layer, stages 1 and 2 consist of a convo-
istics, which are manifested as diferences in the initial lution layer and a maximum pooling layer, and stages 3–5
discharge phase, amplitude, and the number of partial are stacked by residuals with improved CBAM.
discharges on the PRPD image.
Te typical PRPD image of corona discharge is illus-
2.3.1. Residual Element. For normal CNN, expanding the
trated in Figure 2(a). It can be seen that the amplitude of the
depth of the neural network, that is, adding a new hidden
corona discharge is small and that the maximum discharge
layer, often through complete training, can efectively im-
amplitude appears at the peak voltage. Te discharge density
prove the training accuracy. However, in actual experiments,
of the negative half cycle is greater than that of the positive
blindly adding too many hidden layers and deepening the
half cycle, and the discharge is mainly distributed near the
depth of the neural network will easily cause the problem of
peak value of the negative half cycle. Te pulse phase of the
gradient disappearance and gradient explosion in the net-
surface discharge is mainly distributed at the rising edge of
work, However, in actual experiments, blindly adding too
the positive half cycle and the falling edge of the negative half
many hidden layers and deepening the depth of the neural
cycle, as shown in Figure 2(b). Te PD pulse phase distri-
network will easily cause the network to have gradient
bution of the suspended electrode discharge is symmetrical
disappearance and gradient explosion problems, and the
at the peak of the applied voltage, as demonstrated in
training accuracy cannot reach the expected goal. In re-
Figure 2(c). Te initial discharge point of free metal particles
sponse to the above problems, He et al. [20] introduced the
appeared earlier than the voltage peak, the distribution of PD
residual network structure into ordinary CNN. Te skip
was relatively compact, and the amplitude and density of PD
connection structure is adopted in the residual network as
were basically the same in the positive and negative half
shown in Figure 3.
cycles, as shown in Figure 2(d).
In particular, x and y represent the input and output of
the L-th residual network, respectively. F(x) is the residual
function, representing the residual learned by this residual
2.3. Improved CBAM-ResNet Structure. CNN is a kind of network. Te activation function ReLU is represented by
deep-learning network that simulates the hierarchical pro- σ(x). Te residual learned in the residual network can be
cessing of input data by brain neurons. It can extract the expressed as
features of the input layer, establish a structural model with
rich information, and obtain more essential abstract features F(x) � w2 σ w1 x􏼁, (1)
for better classifcation and recognition [19]. To improve the
efciency of type recognition, this paper introduces the where w1 and w2 represent the weight of the convolution
residual network structure based on the typical CNN and layer, H(x) stands for the identity map, that is, H(x) � x, and
adds the improved CBAM. Te improved CBAM-ResNet is the output of the L-th residual network can be defned as
4 Journal of Electrical and Computer Engineering

-15 -30

-27 -31.5 dbm -39

Amplitude (dbm)

Amplitude (dbm)
-39 -48 -49.73 dbm

-51 -57

-63 -66

-75 -75
0 90 180 270 360 0 90 180 270 360
(a) (b)
0 -30
-7.36 dbm
-38.86 dbm
-15 -39

Amplitude (dbm)
Amplitude (dbm)

-30 -48

-45 -57

-60 -66

-75 -75
0 90 180 270 360 0 90 180 270 360
(c) (d)

Figure 2: Te typical PRPD image for each PD type. (a) Corona discharge. (b) Surface discharge. (c) Suspended electrode discharge. (d) Free
metal particle discharge.

x information, use more attention resources for target areas

that need to be focused on, and improve the efciency and
accuracy of visual information processing.
Convolution Layer
Te CBAM attention mechanism consists of a channel
attention module and a spatial attention module and im-
proves the ability of the network to extract implicit im-
F (x) ReLu x portant features. It collaboratively learns the key
information in the image, assigns a higher weight to the PD
information area in the PRPD image, and assigns a lower
Convolution Layer weight to the background information to improve the neural
network. Attention to PD information can improve the
feature learning and expression ability of the network.
F (x)+x In the “serial connection” structure, the channel atten-
ReLu tion module and the spatial attention module are linked in
y series. Whether the channel attention module is enabled frst
Figure 3: Basic structure of the residual element. and then the spatial attention module is enabled, or the
spatial attention module is enabled frst and then the channel
attention module is enabled, the weight in the back is
y � σ(F(x) + H(x)) + σ w2 σ w1 x􏼁 + x􏼁. (2) generated by the feature image in front of it [21]. Te output
of the attention module in front focuses more on valid
information. However, the local information degree of the
PRPD image is very high. It is considered that the attention
2.3.2. Improved CBAM. Te PD information only occupies module in front “modifes” the original input feature image,
a small part of the PRPD image, and most of the content is which to a certain extent afects the characteristics learned by
the natural background. In order to eliminate the redundant the attention module at the end of the row. In other words,
characteristic information irrelevant to the partial discharge the attention module in front focuses on the valid in-
information and improve the detection accuracy of the GIS formation under its own module mechanism and ignores the
PD, the residual block is reconstructed in ResNet34, and an information required by the other module. It will be det-
improved CBAM attention mechanism is designed and rimental to the improvement of recognition accuracy.
added in this work. CBAM can imitate human visual per- In order to solve this problem, this work improves the
ception ability, automatically flter out unimportant serial attention module of CBAM and adopts a “parallel
Journal of Electrical and Computer Engineering 5

connection” structure, so that both attention modules can Te channel attention mechanism treats the feature
directly learn the original input feature image and that there information of diferent channels diferently by constructing
is no need to pay attention to the order of the spatial at- the relationship between the channels of diferent feature
tention and channel attention, then obtaining an improved maps. Te input feature map is marked as F1 ∈ RC×H×W ,
CBAM. Te overall process of the improved CBAM at- where H, W, and C are the height, width, and channel
tention mechanism is shown in Figure 4. number of the input feature maps, respectively.
Te proposed improved CBAM frst obtains the corre- Te global spatial information of the feature graph F1 is
sponding weights from the input feature image F1 through compressed by global maximum pooling and global average
channel attention and spatial attention and then performs an pooling, and then, two feature maps S1 and S2 with size 1 ×
inner product operation on the weights and the original 1 × C are generated. Te two feature maps are input into the
input feature image F1 to obtain the feature image F2 . Te shared multilayer perceptron network to get two one-
process can be expressed as dimensional feature maps. Te two one-dimensional fea-
ture maps are summed according to channels and then
F2 � MC (F) ⊗ Ms (F) ⊗ F1 , (3)
normalized using the sigmoid function to obtain the weight
where MC (F) stands for the weights of the channel attention Mc with the size of each channel being 1 × 1 × C. Te
channel and Ms (F) is the weights of the spatial attention channel attention weight calculation process can be
channel. expressed as

Mc (F) � σ 􏼈MLP[AvgPool(F)] + MLP[MaxPool(F)]􏼉

⎪
⎧
⎨ H W (4)
⎡ 1
� σ ⎪MLP⎢
⎣ 􏽘 􏽘 fx0 i0 , j0 􏼁⎥⎦⎤ + MLP􏼢 max fx0 i0 , j0 􏼁􏼣􏼩,
⎩ H × W i �1 j �1 i∈H,j∈W
0 0

1 n c
where σ stands for the sigmoid function and fx0 (i0 , j0 ) yc � 􏽘x , (6)
n i�1 i
represents the pixel value of the coordinate (x, y) in the x0
channel of the input feature graph F1 . where yc represents the global average of the c-th channel of
Te improved CBAM proposed in this paper adopts the the input feature maps, n indicates the number of features of
“parallel connection” structure of two attention mecha- each channel, and xci is the i-th eigenvalue of the c-th
nisms, so the spatial attention mechanism also takes the channel.
original input feature map F1 as input, performs global Te softmax classifer is an extension of logistic re-
maximum pooling and average pooling operations on the gression on multiple classifcation problems. Its principle is
channel dimension, respectively, and generates two feature simple, and it can solve multiple classifcation problems.
maps P1 and P2 with a size of 1 × H × W. Te two feature Terefore, this work uses the softmax classifer to realize the
maps are merged in the channel-based dimension and then classifcation of GIS PD types. We suppose that the set of
normalized by the sigmoid function after the convolution input feature is X � 􏼈x1 , x2 , ..., xi , ..., xT 􏼉, and the PD type of
operation, and the spatial weighted information Ms is ob- the input feature is one of the sets of C � 􏼈c1 , c2 , ..., ck , ..., cT 􏼉.
tained. Te process can be expressed as follows: Ten, the probability of classifying the input sample xi to be
ck which is one of the sets of C can be expressed as
Ms (F) � σ fconv 􏼂AvgPool F1 􏼁; MaxPool F1 􏼁􏼃􏼁, (5)
􏼌􏼌 T 1
conv P􏼐ck 􏼌􏼌xi 􏼑 � exi ck xTk ck
, (7)
where f stands for the convolution operation. 􏽐K
k�1 e
T
where exi ck is the correlation between category ck and the
xTk ck
2.4. Classifcation Output. Te classifcation output layer entire classifcation type xi and 1/􏽐Kk�1 e is the normal-
consists of the global average pooling (GAP) operation and ization function.
softmax classifer in this work. GAP operation refers to the
spatial averaging operation of the output feature maps of the 3. Experimentation
last convolution layer or pooling layer [22]. Compared with
the fully connected operation, the GAP operation can en- To verify the efectiveness of the proposed improved CBAM-
hance the correlation between the feature and the target ResNet for GIS partial discharge pattern recognition in this
type, and the GAP operation has no parameters to be paper, the GIS PD experimental platform was set up in the
trained, so it can greatly reduce the network parameters and laboratory with a PD detector as the acquisition equipment,
the amount of calculation, efectively prevent overftting, as shown in Figure 5. In the GIS PD experimental platform,
and make the learned features more robust. Te mathe- corona discharge, surface discharge, suspended electrode
matical expression can be written as discharge, and free metal particle discharge defect are set
6 Journal of Electrical and Computer Engineering

Channel Attention Module

Max Pooling Shared MLP

Sigmoid
Output Feature
Input Feature
Average Pooling F2
F1

Spatial Attention Module Average

Pooling
Sigmoid

Max
Pooling

Sum by channel operation

Sigmoid operation
Inner product operation

Figure 4: Improved CBAM structure diagram.

rate of the recognition method, the last column in the gray

box represents the recall rate, and the fnal dark gray box
represents the average recognition precision rate.
From the experimental results, it can be clearly seen that
the average recognition precision rate is the highest,
reaching 94%, when the training set and the test set are
randomly allocated at a ratio of 0.8 : 0.2. At the same time,
the precision rate and the recall rate of the four PD types are
more than 93.55% and 92.93%, respectively. When the
training set and the test set are randomly allocated at a ratio
of 0.4 : 0.6, the average recognition precision rate of the
Figure 5: Te diagram of the GIS PD experimental platform. recognition method is the lowest, which is 82.25%. Tis
shows that increasing the number of training sets can im-
prove the recognition performance of the proposed im-
artifcially to form four typical PD PRPD image acquisition. proved CBAM-ResNet for GIS PD-type recognition.
500 sets of 200 × 120 × 3 PRPD maps were collected of each
typical PD type for a total of 2000 PRPD maps. Te rec-
3.2. Recognition Performance for Diferent Recognition
ognition performance analysis experiment based on dif-
Methods. To further verify the recognition performance of
ferent recognition methods and diferent size training sets
the proposed improved CBAM-ResNet for GIS PD-type
were completed, and then, the proposed improved CBAM-
recognition, the normal CNN and SVM recognition
ResNet for GIS partial discharge pattern recognition was
methods were used to perform the comparative experiment
tested during the engineering project application.
in this work. In particular, the SVM uses the statistical
feature values of the PRPD image as the PD-type recognition
3.1. Recognition Performance for Diferent Training Set Size. and chooses the radial basis function as the kernel function.
In order to verify the infuence of training set size on the Four related standard normal distribution statistical oper-
recognition performance of the proposed improved CBAM- ators are used as statistical feature values of the PRPD image,
ResNet recognition method, the collected PRPD images are which are the skewness Sk , the steepness Ku , the phase
randomly divided into training and test sets in the ratios of asymmetry Φ, and the phase correlation coefcient Cc .
0.8 : 0.2, 0.6 : 0.4, and 0.4 : 0.6 in this work. Ten, the rec- According to the above experiment, increasing the
ognition performance of the proposed method under dif- number of training sets can improve the recognition per-
ferent training datasets is obtained in three groups of formance. Terefore, the collected PRPD images are divided
experiment. Te experimental results are shown in Figure 6. into training and test sets in the ratios of 0.8 : 0.2 in this
Te diagonal black box represents the number of samples experiment. Tat is, among the 2000 collected PD PRPD
whose recognition type is consistent with the actual PD type, images, 400 PRPD images of each PD type were randomly
the white box is the number of incorrectly recognized selected as the training set, a total of 1600 PRPD images were
samples, the last row in the gray box indicates the precision used to train the three PD recognition methods, and the
Journal of Electrical and Computer Engineering 7

Suspended Suspended
Sufface Electrode Free Metal Recall Sufface Electrode Free Recall
Corona PD PD PD PD Rate Corona PD PD PD Metal PD Rate

Corona Corona
102 0 1 4 95.33% 188 3 6 11 89.52%
PD PD

Sufface PD 1 95 4 1 94.06% Sufface PD 4 178 11 7 89.00%

Suspended Suspended
Electrode 2 3 87 1 93.55% Electrode 10 6 166 6 88.30%
PD PD

Free Metal 4 2 1 92 92.93% Free Metal 14 8 6 176 86.27%

PD PD

Precision Precision
93.58% 95.00% 93.55% 93.88% 94.00% 87.04% 91.28% 87.83% 88.00% 88.50%
Rate Rate

(a) (b)
Suspended
Sufface Electrode Free Metal Recall
Corona PD PD PD PD Rate

Corona
261 14 18 17 84.19%
PD

Sufface PD 15 242 17 15 83.74%

Suspended
Electrode 26 19 238 11 80.95%
PD
Free Metal
24 22 15 246 80.13%
PD

Precision
80.06% 81.21% 82.64% 85.12% 82.25%
Rate

(c)

Figure 6: Te recognition performance for three diferent recognition methods. (a) Te ratio of training and test sets is 0.8 : 0.2; (b) the ratio
of training and test sets is 0.4 : 0.6; (c) the ratio of training and test sets is 0.6 : 0.4.

remaining 400 PRPD maps were used as the test set to test the recognition methods have poorer accuracy for PD-type
above three recognition methods. Te recognition results are recognition. Te normal CNN realizes the PD-type recog-
shown in Figure 7. Te diagonal black box represents the nition by automatically obtaining PRPD map features. Te
number of samples whose recognition type is consistent with shapes of diferent PD type PRPD images are relatively close,
the actual PD type, the white box is the number of incorrectly which seriously afects the accuracy of the local placement
recognized samples, the last row in the gray box indicates the pattern recognition [12]. SVM uses the statistical charac-
precision rate of the recognition method, the last column in teristic parameter for the PD-type recognition. For very
the gray box represents the recall rate, and the fnal dark gray close statistical characteristic parameter values, it is difcult
box represents the average recognition precision rate. to accurately identify PD types by a single analysis [23].
About the normal CNN, the average recognition accu- Ten, CNN and SVM recognition methods are naturally
racy is 86.5%, and the recognition precision rate is 85.32%, dependent on large-scale training sample data. To sum up,
88.00%, 86.02%, and 86.73% for corona discharge, surface the proposed improved CBAM-ResNet method is more
discharge, suspended electrode discharge, and free metal suitable for PD-type recognition.
particle discharge, respectively. For the SVM, the average
recognition accuracy is 85.00%, and the recognition pre-
cision rate is 81.65%, 84.00%, 86.02%, and 88.77% for corona 3.3. Engineering Project Application. Te improved CBAM-
discharge, surface discharge, suspended electrode discharge, ResNet model trained with experimental data was applied to
and free metal particle discharge, respectively. In contrast, diagnose PD types at a GIS in Hefei. In the engineering
the improved CBAM-ResNet has the highest average rec- application, the directional antenna was used to obtain the
ognition accuracy which is 94.00%; the SVM has the lowest PD ultrahigh-frequency pulse (UHF) signals, the LeCroy
average recognition accuracy which is 845.00%. Te same WR640Zi oscilloscope was used to acquire the UHF signals
conclusion holds for the recall rate. Tat is, compared with with a bandwidth of 4 GHz, and then, the PD UHF signals
the improved CBAM-ResNet PD-type recognition method are used to generate the PRPD images. Te UHF PD signal
proposed in this work, both normal CNN and SVM acquisition site is shown in Figure 8.
8 Journal of Electrical and Computer Engineering

Suspended Suspended
Sufface Electrode Free Recall Sufface Electrode Free Recall
Corona PD PD PD Metal PD Rate Corona PD PD PD Metal PD Rate

Corona Corona
93 2 3 6 89.43% 89 3 4 6 87.25%
PD PD

Sufface PD 3 88 6 4 87.13% Sufface PD 6 84 5 3 85.71%

Suspended Suspended
Electrode 6 4 80 3 86.02% Electrode 8 6 80 2 83.33%
PD PD

Free Metal Free Metal

7 6 4 85 83.33% 6 7 4 87 83.65%
PD PD

Precision Precision
85.32% 88.00% 86.02% 86.73% 86.50% 81.65% 84.00% 86.02% 88.77% 85.00%
Rate Rate

(a) (b)

Figure 7: Te recognition performance for two diferent recognition methods. (a) Te normal CNN recognition performance; (b) the SVM
recognition performance.

Figure 8: Te UHF PD signal acquisition site.

Table 1: Recognition results of PRPD images in the feld test.

PD type Corona PD Surface PD Suspended electrode PD Free metal PD
Number 3 17 24 5
Recognition accuracy 93.88%

-30
-36.73 dbm
-39
Amplitude (dbm)

-48

-57

-66

-75
0 90 180 270 360
Figure 9: Te randomly selected PRPD image.

A large number of PRPD images were collected in the discharge, surface discharge, suspended electrode dis-
feld tests, and then, the trained CBAM-ResNet model is charge, and free metal particle discharge was 3, 17, 24, and
used to identify the PD type. Te recognition results are 5, respectively. Te average recognition accuracy is
shown in Table 1. Among them, the number of corona 93.88%.
Journal of Electrical and Computer Engineering 9

A randomly selected PRPD image is shown in Figure 9. [6] H. Song, J. Dai, G. Sheng, and X. Jiang, “GIS partial discharge
Te pulse phase of this PRPD image is mainly distributed at pattern recognition via deep convolutional neural network
the positive half-cycle rising edge and the negative half-cycle under complex data source,” IEEE Transactions on Dielectrics
falling edge; that is, the discharge phenomenon occurs be- and Electrical Insulation, vol. 25, no. 2, pp. 678–685, 2018.
tween 0° and 90° and 180° and 270°. Te recognition result of [7] P. Romano, G. Presti, A. Imburgia, and R. Candela, “A new
approach to partial discharge detection under DC voltage,”
this PRPD image is that surface discharge agrees with the
IEEE Electrical Insulation Magazine, vol. 34, no. 4, pp. 32–41,
real PD type. 2018.
[8] H. O. Mota, F. H. Vasconcelos, and C. L. de Castro, “A
4. Conclusion comparison of cycle spinning versus stationary wavelet
transform for the extraction of features of partial discharge
PD-type recognition is very important for evaluating the GIS signals,” IEEE Transactions on Dielectrics and Electrical
insulation condition. To solve the limits, the traditional PD- Insulation, vol. 23, no. 2, pp. 1106–1118, 2016.
type recognition method has low recognition accuracy and [9] Z. Li, L. Qian, and J. Li, “Identifcation of transformer partial
limits its engineering application. Tis paper proposed the discharge types based on statistical features and probabilistic
improved CBAM-ResNet for GIS PD-type recognition. And neural networks,” Power System Protection and Control,
the improved CBAM-ResNet takes advantage of the residual vol. 46, no. 13, pp. 55–60, 2018.
neural noetwork and attention mechanism. In particular, the [10] Z. Sha and J. Liang, “Pattern recognition of partial discharge
channel attention module and the spatial attention module based on moment features and probabilistic neural network,”
are connected in parallel in the improved CBAM. Te Power System Protection and Control, vol. 44, no. 3, pp. 98–
comparison test results show that the improved CBAM- 102, 2016.
[11] T. Yan, L. Yushun, X. Jun, and L. Pengfei, “Feature parameters
ResNet for the GIS PD-type recognition method proposed in
extraction method of partial discharge UHF signal based on
this work has a great improvement in recognition accuracy textural features in time-frequency representation image,”
compared with normal CNN and SVM and has a good vol. 53, no. 07, pp. 30–37, 2018.
application prospect in engineering practice. [12] Y. Wang, J. Yan, Z. Yang, Q. Jing, J. Wang, and Y. Geng,
“GAN and CNN for imbalanced partial discharge pattern
Data Availability recognition in GIS,” High Voltage, vol. 7, no. 3, pp. 452–460,
2022.
Te data used to support the fndings of this study are in- [13] X. B. Zhang, J. Tang, C. Pan et al., “Research of partial dis-
cluded within the article. charge recognition based on deep belief nets,” Power System
Technology, vol. 40, no. 10, pp. 3283–3289, 2016.
Disclosure [14] K. Yin, Y. Wang, S. Liu et al., “GIS partial discharge pattern
recognition based on multi-feature information fusion of
Tis work was based on the research of the conference PRPD image,” Symmetry, vol. 14, no. 11, p. 2464, 2022.
paper [24]. [15] Q. Zhang, H. Song, Y. Jiang, Y. Chen, G. Sheng, and X. Jiang,
“Partial discharge pattern recognition of transformer based on
Conflicts of Interest OS-ELM,” High Voltage Engineering, vol. 44, no. 4,
pp. 1122–1130, 2018.
Te authors declare that they have no conficts of interest. [16] M. Ghasemzadeh, M. Samragh, and F. Koushanfar, “ReBNet:
residual binarized neural network,” in Proceedings of the 2018
References IEEE 26th annual international symposium on feld-
programmable custom computing machines, pp. 57–64,
[1] V. N. Tuyet-Doan, H. A. Pho, B. Lee, and Y. H. Kim, “Deep Boulder, CO, USA, April 2018.
ensemble model for unknown partial discharge diagnosis in [17] B. Guo, S. Li, Q. Tang, L. Li, and P. Li, “A fuzzy interpolation
gas-insulated switchgears using convolutional neural net- cross-correlation method for time-delay estimation of partial
works,” IEEE Access, vol. 9, pp. 80524–80534, 2021. discharge UHF pulse signals,” Journal of Electrical and
[2] M. T. Nguyen, V. H. Nguyen, S. J. Yun, and Y. H. Kim, Computer Engineering, vol. 2021, Article ID 853271, 10 pages,
“Recurrent neural network for partial discharge diagnosis in 2021.
gas-insulated switchgear,” Energies, vol. 11, no. 5, pp. 1202– [18] W. J. K. Raymond, C. W. Xin, L. W. Kin, and H. A. Illias,
1213, 2018. “Noise invariant partial discharge classifcation based on
[3] X. Han, X. Zhang, R. Guo et al., “Partial discharge detection in convolutional neural network,” Measurement, vol. 177, Article
gas-insulated switchgears using sensors integrated with UHF ID 109220, 2021.
and optical sensing methods,” IEEE Transactions on Di- [19] L. Alzubaidi, J. Zhang, A. J. Humaidi et al., “Review of deep
electrics and Electrical Insulation, vol. 29, no. 5, pp. 2026– learning: concepts, CNN architectures, challenges, applica-
2033, 2022. tions, future directions,” Journal of big Data, vol. 8, pp. 53–74,
[4] X. Wang, B. Niu, and F. Ma, “Research on partial discharge 2021.
fusion diagnosis and intelligent early warning system of [20] K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into
electrical equipment,” High Voltage Apparatus, vol. 57, no. 11, rectifers: Surpassing human-level performance on imagenet
pp. 93–100, 2021. classifcation,” in Proceedings of the IEEE International
[5] J. Yang, M. Deng, Y. Ma, A. Chun, Y. Li, and C. Liu, “Feature Conference on Computer Vision, pp. 1026–1034, Santiago,
extraction method of PRPD data based on deep learning,” Chile, December 2015.
Electrical Measurement & Instrumentation, vol. 57, no. 3, [21] H. Su, X. Wang, T. Han, Z. Wang, Z. Zhao, and P. Zhang,
pp. 99–104, 2020. “Research on a U-net bridge crack identifcation and feature-
10 Journal of Electrical and Computer Engineering

calculation methods based on a CBAM attention mecha-

nism,” Buildings, vol. 12, no. 10, p. 1561, 2022.
[22] T. Y. Hsiao, Y. C. Chang, H. H. Chou, and C. T. Chiu, “Filter-
based deep-compression with global average pooling for
convolutional networks,” Journal of Systems Architecture,
vol. 95, pp. 9–18, 2019.
[23] L. Yin, D. Bai, M. Wang, X. Gong, and C. Gu, “Svm-based
partial discharge pattern classifcation for gis Journal of
Physics: conference Series,” IOP Publishing, vol. 960, no. 1,
Article ID 012051, 2018.
[24] D. Hu, Z. Chen, W. Yang, T. Zhu, and Y. Ke, “Partial dis-
charge pattern recognition of GIS based on CBAM-ResNet,”
in Proceedings of the 2022 4th International Conference on
Robotics, Intelligent Control and Artifcial Intelligence,
pp. 548–552, Dongguan, China, April 2022.