Content uploaded by Hong Liu
Author content
All content in this area was uploaded by Hong Liu on Sep 19, 2022
Content may be subject to copyright.
1
> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
Partial Discharge Pattern Recognition based on a
Multifrequency F-P Sensing Array, AOK Time-
Frequency Representation, and Deep Learning
Zhixian Zhang, Student Member, IEEE, Weigen Chen, Kejie Wu, Haoyuan Tian, Ruimin Song,
Yuxuan Song, Hong Liu, Jianxin Wang
Abstract—Recognizing the patterns of partial discharge (PD) is
the key to assessing the hazard level of PD. This paper proposes an
oil-immersed transformer PD pattern recognition method based
on a multifrequency fiber-optic Fabry-Perot (F-P) ultrasound
sensing array and deep learning. Three F-P ultrasonic probes with
different resonance bands were used to form a sensing array for
collecting ultrasonic signals excited by PD. Five types of PD models
were prepared, including metal tip discharge in oil, partial
discharge in the air cavity, and surface discharge on the
pressboard, in addition to surface discharge after pressboard
blocking or after pressboard plus transformer winding blocking,
to study the influence of ultrasonic attenuation on recognition
accuracy. The collected PD ultrasound signal time-frequency
matrix was obtained using adaptive optimal-kernel time-
frequency representation to reinforce the differences in PD
patterns. Then, based on the features of the PD ultrasonic signals,
the signal time and frequency band for analysis were determined.
The intercepted time-frequency matrix of the signals from the
three F-P probes was formed into a (3, 300, 375) tensor. A modified
ResNet-18 net was used for PD pattern recognition, which
achieved a 98% recognition accuracy. The probabilities given by
the softmax function were used to study the confidence of the
model's predictions of signals belonging to known and unknown
types.
Index Terms—partial discharge; pattern recognition; optical
fiber sensors; acoustic signal processing; deep learning
I. INTRODUCTION
HE function and safety of a power transformer mainly
depend on its insulation state [1]. An inspection of the
insulation state helps to detect early defects in the power
transformer and can prevent serious accidents. Partial discharge
(PD) is an essential indicator of insulation aging and insulation
defects in power transformers [2][3]. Ultrasonic detection
[4][5] is the primary method used for PD detection, which is
especially suitable for online monitoring. The ultrasonic
sensors used for PD detection are mainly piezoelectric sensors
affixed to the transformer shell. However, the ultrasonic signal
will be considerably attenuated when propagating through the
This work was supported by the Science & Technology Project of State Grid
Corporation of China, grant/award number: 5500-202099279A-0-0-00.
(Corresponding author: Weigen Chen)
transformer shell; therefore, in recent years, many scholars have
carried out research on PD detection methods based on built-in
fiber optic F-P ultrasonic sensors, which have high sensitivity,
free from electromagnetic interference, and are also less likely
to be damaged by the internal harsh electromagnetic
environment [6][7].
The frequency response characteristics of the ultrasonic
sensor are the key to determining the sensitivity and anti-
interference ability of PD detection. The frequency band of PD
ultrasonic signal in oil is mainly concentrated in 20 kHz-180
kHz, while the noise in the field is generally below 60 kHz.
Therefore, IEEE recommends using resonant frequency in 120-
160 kHz piezoelectric sensors to improve the anti-interference
ability [8]. However, as for ultrasonic waves, higher frequency
means more attenuation, so some scholars and companies also
recommend using low-frequency (less than 60 kHz) resonant
piezoelectric ultrasonic sensors to improve the sensitivity of PD
detection [9]. For instance, Wojciech et al. proposed a
multifrequency piezoelectric sensor using multiple response
piezoelectric crystals with different resonant frequencies (68
kHz and 90 kHz) [9]. The resonant frequency bandwidth of one
F-P sensor is narrower, generally less than 50 kHz [10], and
with a certain diaphragm material, the higher the resonant
frequency, the lower the static pressure sensitivity. Therefore,
when building a PD detect system based on F-P sensors, it is
even more necessary to consider the acquisition requirements
of different frequencies, so the multifrequency F-P sensing
array (41 kHz, 75 kHz, and 112 kHz) was made and used to
detect PD in this paper. The 41 kHz F-P probe is used to
improve the system's detection sensitivity, the 112 kHz F-P
probe is used to improve the system's immunity to interference,
and the 75 kHz F-P probe is used to compensate for the low
amplitude region of the 41 kHz probe and the 112 kHz probe.
In addition to the need for sensitive and accurate detection
of the presence of PD, if information such as the pattern and
Zhixian Zhang, Weigen Chen, Kejie Wu, Haoyuan Tian, Ruimin Song,
Yuxuan Song, Hong Liu, and Jianxin Wang are with the State Key Laboratory
of Power Transmission Equipment & System Security and New Technology
(Chongqing University), Chongqing 400044, China (e-mail:
zhang.zhixian@cqu.edu.cn; weigench@cqu.edu.cn; wukejie@cqu.edu.cn;
hytian@cqu.edu.cn; ruiminsong@cqu.edu.cn; songyuxuan@cqu.edu.cn;
cqu_liuhong@outlook.com; wang.jianxin@cqu.edu.cn).
Color versions of one or more of the figures in this article are available
online at http://ieeexplore.ieee.org
T
This article has been accepted for publication in IEEE Transactions on Dielectrics and Electrical Insulation. 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/TDEI.2022.3199189
© 2022 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: CHONGQING UNIVERSITY. Downloaded on September 19,2022 at 11:21:04 UTC from IEEE Xplore. Restrictions apply.
2
> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
severity of PD can be obtained, this can provide more effective
guidance for maintenance, so many scholars have been studying
PD pattern recognition methods [11], which mainly involves
three aspects: (1) constructing a PD sample library, (2) studying
signal processing methods, and (3) studying pattern recognition
methods.
In terms of constructing a PD sample library, due to the
small number of PD signal samples collected in engineering
practice, the current research mainly uses specially processed
PD models to generate PD samples [9]. However, one problem
in studies that use ultrasonic signals to recognize PD is that the
influence of the ultrasonic signal propagation path has rarely
been considered. When the ultrasonic signal propagates inside
the transformer, the solid medium on the propagation path will
have a large attenuation effect on the signal. Therefore, the
same pattern of PD, such as surface discharge occurring on the
outside of the high-voltage winding or the inside of the high-
voltage winding, may have many differences in the signal
characteristics because the ultrasonic signal undergoes different
attenuation effects when propagating to the sensors, as shown
in Fig. 1. Therefore, in this paper, as for constructing the PD
sample library, in addition to conventional PD models, models
of surface discharge via pressboard blocking and pressboard
plus winding blocking were built and used to simulate the
surface discharge of the outside and inside of the high-voltage
winding, respectively.
Fig. 1. Schematic diagram of the propagation of the PD
ultrasonic signal in different positions.
In terms of signal processing methods, the main
approaches include short-time Fourier transform (STFT) [12],
wavelet transform [13], and empirical mode decomposition
(EMD)-based Hilbert–Huang transform (HHT) [14]. The core
concept of these methods is to decompose the signal and obtain
its time-frequency representation to highlight PD
characteristics. However, STFT cannot adjust resolution
adaptively for signals of different frequencies; wavelet
transform requires appropriate settings of wavelet basis
functions, which makes it challenging to achieve self-
adaptation of the signal [15], so the artificially set parameters
such as the time window length will significantly impact the
feature extraction results, which means the introduction of a
priori information in the process of signal processing and
restricts the universality of the method. EMD decomposition
can self-adaptation to signals, yet it has mode overlap problem.
Although many improved EMD algorithms can suppress mode
overlap, there are still problems such as slow calculation speed
and relatively sparse time-frequency expression. Considering
the adaptability to the signal and the calculation speed, in this
paper, adaptive optimal-kernel (AOK) time-frequency
representation [16] was used to obtain the time-frequency
representation of the PD ultrasonic signal, which is adaptive to
different frequencies and has a fast-computational speed.
In terms of pattern recognition, there are traditional
machine learning methods based on eigenvalue engineering
[17] [18] and methods based on deep learning (DL) [19]. DL-
based methods rely on learning machines to learn and extract
features automatically and have good generalizability and
recognition accuracy. However, a deep network often means a
lot of calculation operations and a much larger memory usage;
in addition, the PD sample library in this paper is small
compared to commonly used image recognition databases, and
a deep network may cause overfitting. Therefore, in this paper,
an optimized ResNet-18 was used to recognize PD patterns.
The core concept of the residual network is that every additional
layer should more easily contain the identity function as one of
its elements, which makes it easier to train a network, and can
achieve a higher image recognition accuracy; this design had a
profound influence on how to build deep neural networks [20]
[21]. The ResNet-18 means it belongs to the residual network
and only has 18 layers, so it's relatively small, and its
calculations can be done more quickly.
In addition, one issue has been controversial in the pattern
recognition of PD for a long time; in general, the pre-set PD
types in the laboratory are used as a scale to recognize PD, so
the validity of the pattern recognition may be unreliable if the
signals don't belong to the known types. To study the influence
of signals which belongs to unknown types, a noise database
including impulse noise with a center frequency of 20 kHz and
30 kHz has been made. A plate-plate discharge model has been
used to emit unknown PD ultrasound signals. The probabilities
given by the softmax function [22] were used to identify the
confidence of the model's predictions, and the confidence
difference between known and unknown types has been
discussed.
II. MULTIFREQUENCY F-P SENSING ARRAY
The multifrequency F-P sensing array contains three F-P
ultrasonic probes with different resonant bands. The laser light
emitted from the narrowband laser is fed into three circulators
via a 1×3 coupler and subsequently into three F-P ultrasound
probes. The laser light is reflected at the fiber end surfaces and
the diaphragm surfaces of the F-P probes and forms
interference. The output light from the F-P ultrasound probes is
input to three photodetectors, then to the DAQ System, as
shown in Fig. 2.
This article has been accepted for publication in IEEE Transactions on Dielectrics and Electrical Insulation. 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/TDEI.2022.3199189
© 2022 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: CHONGQING UNIVERSITY. Downloaded on September 19,2022 at 11:21:04 UTC from IEEE Xplore. Restrictions apply.
3
> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
Fig. 2. Schematic diagram of the multifrequency F-P ultrasound
sensing array.
The acoustic response of the F-P ultrasound probe is
determined by the material, structure, and fixation of its
diaphragm and the mechanical properties of its surrounding
medium, and the (0,0) inherent frequency of a clamped
diaphragm in contact with the liquid on one side can be
estimated by Eq. (1) [7]
( )
2
2
020
0
3.2 1
431 1 0.669
hE
faa
h
−
+
(1)
where h is the thickness of the diaphragm, a is the effective
vibration radius of the diaphragm, ρ is the density of the
diaphragm, μ is the Poisson's ratio of the diaphragm, and ρ0 is
the density of the liquid where the F-P probe is located. In this
paper, the liquid is the #10 transformer insulating oil, and the
density is taken as 895 kg/m3, the breakdown voltage is 42 kV,
and the kinematic viscosity is 8.931 mm2/s (40℃).
The Corning glass (Young's modulus 73.6 GPa, density 2380
kg/m3, Poisson's ratio 0.23) coated with a dielectric reflective
film (reflectivity 99%) was used to make the sound-sensitive
diaphragm. Three sound-sensitive diaphragms with different
resonant frequencies were designed, and the structural
parameters are as follows: F-P 1: a=1.7 mm, h=115 μm; F-P 2:
a=1.7 mm, h=215 μm; F-P 3: a=1.4 mm, h=190 μm. To obtain
the acoustic response of each F-P probe, the F-P probes were
placed in transformer oil, and the ultrasonic waves were excited
by a REF-VL piezoelectric sensor at 10 cm directly opposite the
F-P probe. A signal generator was used to input a sinusoidal
signal with a peak-to-peak value of 15 V to REF-VL, and the
frequency of the sinusoidal signal was changed in turn to obtain
the frequency response of each F-P probe, as shown in Fig. 3.
(a) F-P 1
(b) F-P 2
(c) F-P 3
Fig. 3. The frequency responses of F-P 1, F-P 2, and F-P 3.
As seen in Fig. 3, F-P1, F-P2, and F-P3 have different
resonant frequencies. To improve the signal-to-noise ratio of PD
detection, the passband frequencies of all three photodetectors
are set to 10 kHz–300 kHz, and the amplifying factor is set to
100. The structure of the F-P probe is shown in Fig. 4.
(a) F-P probe structure schematic
(b) F-P probe picture
Fig. 4. The structure of the F-P probe.
Ⅲ. TIME-FREQUENCY REPRESENTATION OF PARTIAL
DISCHARGE ULTRASONIC SIGNAL
Three PD models were designed to simulate the metal tip
discharge in oil, partial discharge in the air cavity, and surface
discharge on the pressboard, which are common discharges
inside a power transformer. In addition, to consider the
attenuation effect of solid insulation, a surface discharge model
with pressboard blocking and with pressboard plus winding
blocking was added, as shown in Fig. 5.
This article has been accepted for publication in IEEE Transactions on Dielectrics and Electrical Insulation. 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/TDEI.2022.3199189
© 2022 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: CHONGQING UNIVERSITY. Downloaded on September 19,2022 at 11:21:04 UTC from IEEE Xplore. Restrictions apply.
4
> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
(a) MT
(b) AC
(c) SD
(d) SD-p
(e) SD-p-w
Fig. 5. Partial discharge models.
In Fig. 5, the discharge electrodes were made of brass. The
winding model was made of iron plates, each with a thickness
of 3 mm, a width of 35 mm, and a spacing of 3 mm between
each other, which were covered with insulating paper and fixed
with wooden boards. The oil tank was filled with 10#
transformer insulating oil at a height of 20 cm (2/3 of the tank
height).
The PD test platform is shown in Fig. 6.
Fig. 6. The schematic of the partial discharge test platform: T=
transformer (60 kVA/60 kV); R= resistor (100 kV, 10 kΩ); C=
coupling capacitor; CD= coupling device; PDG= electrode
system for partial discharge generation; OT= mineral oil filled
resin tank; PT= piezoelectric sensor (R15α); SA= signal
amplifier; O= oscilloscope (640Zi).
The picture of the experimental oil tank and sensor
arrangement is shown in Fig. 7.
Fig. 7. The picture of the experimental oil tank and sensor
arrangement.
The voltage was increased at a rate of 1 kV/s until it reached
5 kV; then, every 1 minute, the voltage was increased by 0.5 kV
until F-P 1 detected the PD ultrasonic signal. At this time, the
voltages applied to the MT, AC, SD, SD-p, SD-p-w models
were around 15.3 kV, 18.2 kV, 14.4 kV, 15.7 kV, and 16.7 kV,
respectively; then stop increasing the voltage, set the
oscilloscope to be triggered by the channel 1, which is
corresponding to F-P 1, and continuously collect the ultrasonic
signals excited by PD. When the collected signal sets reach 50,
set the voltage to 0 V, replace the pressboard and repeat the
above process.
The PD ultrasonic signals collected by the F-P sensing array
and R15α sensor of the MT model are shown in Fig. 8. The
R15α sensor was connected to a bandpass amplifier with a
bandwidth of 10 kHz to 2 MHz and an amplification factor of
100 times.
(a) F-P 1
(b) F-P 2
(c) F-P 3
(d) R15α
Fig. 8. PD ultrasonic signal detected by F-P sensing array and
piezoelectric sensor.
This article has been accepted for publication in IEEE Transactions on Dielectrics and Electrical Insulation. 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/TDEI.2022.3199189
© 2022 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: CHONGQING UNIVERSITY. Downloaded on September 19,2022 at 11:21:04 UTC from IEEE Xplore. Restrictions apply.
5
> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
As seen in Fig. 8, the F-P sensing array can pick up the PD
ultrasound signals at different frequency bands with a high
signal-to-noise ratio.
The time-frequency representation can provide the PD
ultrasonic signal's energy distribution in the time-frequency
domain, providing the learning machine with richer information
to realize higher recognition accuracy. The adaptive optimal
kernel (AOK) method provides an effective approach to
analyzing nonstationary signals in fine time-frequency
resolution and to suppressing cross-terms [23] [24]. A
comparative analysis of the AOK algorithm with other time-
frequency analysis algorithms can be found in reference [24].
Most of the time-frequency representations with a bilinear
form that have been proposed can be regarded as members of
the Cohen class, and their general representation is
*
()
1
( , : ) ( / 2) ( / 2) ( , )
2
x
j t u
C t g x u x u g
e dud d
− + −
= + −
(2)
In Eq. (2), t is the time, Ω is the frequency, τ is the time shift,
θ is the frequency shift, u is the integral variable, and g (θ, τ) is
the kernel function. Eq. (2) can also be expressed with the
ambiguity function
()
( , ) ( , ) ( , ) jt
xx
C t A g e d d
− +
=
(3)
where
( , )
x
A
is the symmetrical ambiguity function of the
signal x(t).
Different choices for the kernel function yield widely
different time-frequency representations (TFR's). However, for
a determined kernel function, the performance of the TFR's
when acting on different signals is not identical. The reference
[25] gives an optimal kernel function design method. The radial
Gaussian functions
( )
2 2 2
( )/ 2
( , )ge
−+
=
(4)
is tuned to design the best kernel function for a specific signal
x(t). Using polar coordinates for Eq. (4),
22
/2 ( )
( , ) r
g r e
−
=
(5)
where
22
r
=+
is a radial variable. The optimization
problem of solving the optimal kernel function of the signal can
be expressed as
2
2
00 ( , ) ( , )
max x
g
A r g r rdrd
(6)
22
/2 ( )
22
200
( , )
s.t. 1( , ) , 0
4
r
g r e
g r rdrd
−
=
(7)
However, the method in reference [25] can only be used on
block data and is not adaptable to subsequent signal changes.
Therefore, [16] proposed an adaptive kernel function design
method, which defines a short-time ambiguity function (STAF)
*
1
( ; , ) ( , ) ( / 2) ( / 2)
2
ju
xx
A t r u w u t w u t e du
= − + − −
(8)
where
( , )
x
ru
is the instantaneous autocorrelation function of
the signal
( )
xt
,
( )
wu
is the window function, and t is the
center position of the applied window function. The purpose of
using an STAF is to divide the signal into small segments and
transform it. Therefore, the AOK step brings the
( ; , )
x
At
obtained at time t into the objective function of Eq. (6) and
solves the linear programming problem under the constraints of
Eq. (7) to obtain the optimal kernel
( ; , )
opt
gt
at time t and
then continues to change t to find the optimal kernel at different
moments.
From the tank structure and sensor position shown in Fig. 6,
and the MT model PD ultrasonic signals shown in Fig. 8, and
considering the speed of sound in oil, the second time for the
sound wave to reach the sensor position after reflection is
approximately 0.5 ms; therefore, the collected PD ultrasonic
signal can be roughly divided into 3 phases, as shown in Fig. 9.
Fig. 9. Partial discharge ultrasound signal measured by three
F-P sensors.
The echo signal contains information about the oil tank
structure and the location of the sensor array and PD point,
which can be used for PD localization. However, when
choosing a characteristic signal for PD pattern recognition, it
should be omitted to emphasize the information about the PD
itself. Besides, there is a segment of noise before the PD signal,
and the characteristic signal contains part of this noise to
represent the amplitude of the PD signal after normalization.
The selected characteristic signal is shown in Fig. 10.
Fig. 10. Partial discharge characteristic signal extraction.
As shown in Fig. 10, to extract the characteristic signal, the
Teager energy operator combined with the double threshold
This article has been accepted for publication in IEEE Transactions on Dielectrics and Electrical Insulation. 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/TDEI.2022.3199189
© 2022 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: CHONGQING UNIVERSITY. Downloaded on September 19,2022 at 11:21:04 UTC from IEEE Xplore. Restrictions apply.
6
> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
method (In Fig. 10, Th 1 = first threshold, Th 2 = second
threshold) [26] was used to extract the signal start-point position
of F-P 1. At the start-point position, 0.05 ms is taken forward
and 0.1 ms is taken backward to intercept the characteristic
signal used for analysis, and the signals collected by F-P 2 and
F-P 3 are also intercepted according to the same parameters.
For the five types of PD, the characteristic signals collected
by F-P 1, 2, and 3 are analyzed by the AOK, the sampling rate
of the oscilloscope is set to 2500 kS/s, the width of the analysis
window is set to 32, the moving step of the analysis window is
1, and the length of the FFT is 2048, so we obtain a time-
frequency matrix with a frequency interval of 0 Hz–1250 kHz
and a time length of 0.15 ms, and the size of the matrix is
1025×375. To improve the proportion of effective information
in the time–frequency matrix and reduce the computational
complexity, the number of rows in the matrix is reduced from
1025 to 300 so that the frequency interval is approximately 0
Hz–365 kHz, and the matrix size is 300×375. The time-
frequency representation of the PD ultrasonic signals detected
by F-P 1, 2, and 3 are shown in Fig. 11, and the coordinates of
the signal starting points are set to 0.
(a) MT, F-P 1
(b) MT, F-P 2
(c) MT, F-P 3
(d) AC, F-P 1
(e) AC, F-P 2
(f) AC, F-P 3
(g) SD, F-P 1
(h) SD, F-P 2
(i) SD, F-P 3
(j) SD-p, F-P 1
(k) SD-p, F-P 2
(l) SD-p, F-P 3
(m) SD-p-w, F-P 1
(n) SD-p-w, F-P 2
(o) SD-p-w, F-P 3
Fig. 11. The AOK time-frequency representation of PD ultrasonic signals.
This article has been accepted for publication in IEEE Transactions on Dielectrics and Electrical Insulation. 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/TDEI.2022.3199189
© 2022 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: CHONGQING UNIVERSITY. Downloaded on September 19,2022 at 11:21:04 UTC from IEEE Xplore. Restrictions apply.
1
> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
As shown in Fig. 11, (1) The PD emitted by MT and SD
model have stronger signal amplitudes and can be sensitively
detected by all three F-P probes, because they occur at the oil
and oil-paper interfaces, respectively, which means the
ultrasonic signal is not blocked by the insulating medium during
propagation to the sensing array. (2) The PD emitted by AC
model occurs in the air cavity inside the solid insulation material
and is blocked by the pressboard in the propagation process, so
the signal amplitude of F-P 3 is weaker. (3) The PDs emitted by
SD-p and SD-p-w model are blocked by the pressboard and the
pressboard plus winding, respectively, so the signal amplitude
of F-P 3 is significantly attenuated compared to the SD model;
In addition, the attenuation of SD-p-w model is much higher,
and only F-P 1 can clearly show the difference between the PD
phase and the noise phase. (4) The high-frequency component
of the signal is mainly concentrated in the head of the signal,
and the attenuation of the high-frequency signal is more obvious
due to the blocking of the solid material.
IV. PARTIAL DISCHARGE PATTERN RECOGNITION
The convolutional neural network (CNN) [20] is designed
for computer vision applications, which can make the learning
machine automatically extract features from 2D data. For image
recognition, through the reasonable design of model structures
and hyperparameters, CNN-based methods can achieve much
higher accuracy than traditional machine learning, especially
when using deep networks.
In this paper, the ResNet-18 [21] was used as a basis to
construct a pattern recognition neural network for PD, as shown
in Fig 12.
Fig. 12 The modified ResNet-18 network for PD pattern
recognition.
In Fig.12, the convolution operation can learn and build
feature extraction methods for various types of PD signals
based on the input labeled data samples. For each input sample
of dimension (3,300,375), the convolution operation processes
it to a tensor of dimension (512,10,12) and, by global averaging,
to a tensor of dimension (512,1,1). Then, the output vector (1,5)
is obtained by a fully connected layer, and then based on the
softmax function, the predicted class and probability can be
obtained [20] [22].
In this paper, the database has 1100 samples, in other words,
220 samples for each type of PD, of which 180 samples are used
for training, 20 samples are used for tuning, and another 20
samples are used for testing. The PD ultrasonic signals
measured by the three probes, each normalized to [0,1], were
input into the network as three channels of one tensor. The mini-
batch stochastic gradient descent was used for model
optimization. The batch size was set to 5, the epoch was set to
10, the initial value of the learning rate was set to 0.001, and
every 2 epochs, the learning rate decayed by 20%, and the
weight decay was set to 1e-5. One RTX 3060 was used for
training, and the loss and accuracy variation during the training
process is shown in Fig. 13, and the test accuracy reached 98%
in the 10th epoch.
Fig. 13. Loss and accuracy variation.
The confusion matrix for the recognition is shown in Fig. 14.
Fig. 14. The confusion matrix.
Using the same signal analysis method and training
parameters, only the signal start points, and the network's
channel number were changed, the PD time-frequency matrix
(1,300,375) collected by R15α was input into the network, and
the recognition accuracy was 92.9% in the 10th epoch.
An MT sample, as shown in Fig. 11 (a)-(c) was used to
analyze the feature extraction process of the network. The
visualization of this sample based on RGB color (R = F-P 1, G
= F-P 2, B = F-P 3) is shown in Fig. 15 (a), and its first 8 feature
maps output in the first 3 convolutional layers are shown in Fig.
15 (b)-(d).
This article has been accepted for publication in IEEE Transactions on Dielectrics and Electrical Insulation. 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/TDEI.2022.3199189
© 2022 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: CHONGQING UNIVERSITY. Downloaded on September 19,2022 at 11:21:04 UTC from IEEE Xplore. Restrictions apply.
2
> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
(a)
(b)
(c)
(d)
Fig. 15. Tensor visualization and feature maps.
In Fig. 15 (b)-(d), we use a jet colormap to visualize the
feature maps, and the high-amplitude regions in each map are
the regions that the neural network focuses on. Compared to Fig.
11 (a)-(c), it can be seen that the energy distribution of the PD
in each channel's time-frequency representation has a decisive
influence on the feature maps.
The recognition accuracy and computational complexity of
the network in this paper are compared to 3 commonly used
networks, and each network is trained for 10 epochs, as shown
in Table I. TABLE I
ACCURACY COMPARISON USING DIFFERENT NETWORKS
Net
Recognition
accuracy
Computational
complexity
Resnet-18
98%
4.17 GMac
Resnet-50 [21]
58%
9.55 GMac
Mobilenet_v3_s [27]
77%
0.14 GMac
VGG-11 [20]
85%
16.68 GMac
To analyze the influence of unknown classes, a plate-plate
electrode (upper plate diameter 80 mm, lower plate diameter
100 mm, plate spacing 2 mm, noted as PP) was used to emit the
PD signal of unknown class. Besides, the piezoelectric crystal
combined with signal amplifier and signal generator was used
to generate impulse ultrasound with center frequencies at 20
kHz (noted as N20) and 30 kHz (noted as N30) to simulate the
noise signal, as shown in Fig 16.
(a) F-P 1
(b) F-P 2
(c) F-P 3
Fig. 16. The impulse ultrasound with center frequency at 30
kHz.
Take 10 samples of each type from the above unknown class,
and take 10 samples of each type from the test database as for
the known class, and input them into the trained network to
obtain the confidence probability based on the softmax function
[22] as shown in Fig.17.
Fig. 17. The prediction confidence of the network model for
different samples.
This article has been accepted for publication in IEEE Transactions on Dielectrics and Electrical Insulation. 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/TDEI.2022.3199189
© 2022 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: CHONGQING UNIVERSITY. Downloaded on September 19,2022 at 11:21:04 UTC from IEEE Xplore. Restrictions apply.
3
> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
As shown in Fig. 17, as for the most known type samples,
the network model has very high confidence with a probability
higher than 0.99. However, for SD-p, and SD-p-w there are also
some samples with low diagnostic confidence, which may be
due to their same PD source and a low signal-to-noise ratio. For
the unknown noise signals, the model gives predictions with
low confidence. For the unknown type of PD, a greater
dispersion of confidence can be seen, therefore further research
is needed for the analysis and processing of these samples cause
some samples of this type can also have very high confidence.
V. CONCLUSIONS
This paper proposed a PD pattern recognition method based
on a multifrequency fiber-optic F-P ultrasound sensing array,
AOK time-frequency representation, and ResNet-18 network,
and obtained the following conclusions:
(1) Combining several F-P ultrasonic probes with different
resonant bands helps improve the sensing system's detection
band and can achieve a higher recognition accuracy than the
piezoelectric sensor, which was installed on the shell outside.
(2) The high-frequency components of the PD ultrasound
signal are mainly concentrated in its head, and there are
apparent differences in the time-frequency distribution of
different PD types. The solid insulation materials produce a
noticeable attenuation effect on the ultrasound signal, and the
high-frequency signal attenuation is more prominent.
(3) The selection and tuning of the neural network model
significantly influence the final recognition accuracy. This
paper finally achieves 98% recognition accuracy for five PD
types based on ResNet-18. As for unknown types, the network
model has low confidence in predicting unknown noise types
yet more dispersed confidence of unknown PD types, which is
worth further research.
REFERENCES
[1] P. Rozga, T. Piotrowski, and R. Kozak, "Experiences with detection of
transformer failure based on insulating oil tests," in 2017 IEEE 19th
International Conference on Dielectric Liquids (ICDL), 25-29 June 2017
2017, pp. 1-4, doi: 10.1109/ICDL.2017.8124628.
[2] J. Du, W. Chen, L. Cui, Z. Zhang, and S. Tenbohlen, "Investigation on
the Propagation Characteristics of PD-Induced Electromagnetic Waves
in an Actual 110 kV Power Transformer and Its Simulation Results,"
IEEE Trns. Dielectr. Electr. Insul., vol. 25, no. 5, pp. 1941-1948, Oct
2018, doi: 10.1109/tdei.2018.007336.
[3] G. Ma, Y. Wang, W. Qin, et al., "Optical sensors for power transformer
monitoring: A review," High Voltage, vol. 6, no. 3, pp. 367-386, Jun
2021, doi: 10.1049/hve2.12021.
[4] H. D. Ilkhechi and M. H. Samimi, "Applications of the Acoustic Method
in Partial Discharge Measurement: A Review," IEEE Trns. Dielectr.
Electr. Insul., vol. 28, no. 1, pp. 42-51, 2021, doi:
10.1109/TDEI.2020.008985.
[5] Z. Zhou, G. Ma, M. Zhang, H. Zhang and C. Li, "A High Sensitivity
Optical Fiber Interferometer Sensor for Acoustic Emission Detection of
Partial Discharge in Power Transformer," IEEE Sens J, vol. 21, no. 1, pp.
24-32, 1 Jan.1, 2021, doi: 10.1109/JSEN.2019.2951613.
[6] C. Gao, L. Yu, Y. Xu, W. Wang, S. Wang, and P. Wang, "Partial
Discharge Localization Inside Transformer Windings via Fiber-Optic
Acoustic Sensor Array," IEEE Trns. Dielectr. Electr. Insul., vol. 34, no.
4, pp. 1251-1260, Aug 2019, doi: 10.1109/tpwrd.2018.2880230.
[7] Z. Zhang, J. Lei, W. Chen, et al., "Oil-paper insulation partial discharge
ultrasonic multifrequency sensing array based on fibre-optic Fabry-Perot
sensor," High Voltage, 2021, doi: 10.1049/hve2.12123.
[8] IEEE Guide for the Detection, Location and Interpretation of Sources of
Acoustic Emissions from Electrical Discharges in Power Transformers
and Power Reactors, IEEE Standard C57.127,2018.
[9] W. Sikorski, "Development of Acoustic Emission Sensor Optimized for
Partial Discharge Monitoring in Power Transformers," Sensors, vol. 19,
no. 8, Apr 2 2019, Art no. 1865, doi: 10.3390/s19081865.
[10] Q. C. Chen, W. C. Zhang, and H. Zhao, "Response Bandwidth Design
of Fabry-Perot Sensors for Partial Discharge Detection Based on
Frequency Analysis," Journal of Sensors, vol. 2019, Nov 2019, Art no.
1026934, doi: 10.1155/2019/1026934.
[11] S. Lu, H. Chai, A. Sahoo, and B. T. Phung, "Condition Monitoring Based
on Partial Discharge Diagnostics Using Machine Learning Methods: A
Comprehensive State-of-the-Art Review," IEEE Trns. Dielectr. Electr.
Insul., vol. 27, no. 6, pp. 1861-1888, Dec 2020, doi:
10.1109/tdei.2020.009070.
[12] T. Boczar, S. Borucki, A. Cichon, and D. Zmarzly, "Application
Possibilities of Artificial Neural Networks for Recognizing Partial
Discharges Measured by the Acoustic Emission Method," IEEE Trns.
Dielectr. Electr. Insul., vol. 16, no. 1, pp. 214-223, Feb 2009, doi:
10.1109/tdei.2009.4784570.
[13] C. Boya, M. Ruiz-Llata, J. Posada and J. A. Garcia-Souto, "Identification
of multiple partial discharge sources using acoustic emission technique
and blind source separation," IEEE Trns. Dielectr. Electr. Insul., vol. 22,
no. 3, pp. 1663-1673, June 2015, doi: 10.1109/TDEI.2015.7116363.
[14] F. Gu, H. Chen, and M. Chao, "Application of Improved Hilbert-Huang
Transform to Partial Discharge Signal Analysis," IEEE Trns. Dielectr.
Electr. Insul., vol. 25, no. 2, pp. 668-677, Apr 2018, doi:
10.1109/tdei.2017.006922.
[15] J. Gilles, "Empirical Wavelet Transform," IEEE Transactions on Signal
Processing, vol. 61, no. 16, pp. 3999-4010, 2013, doi:
10.1109/TSP.2013.2265222.
[16] D. L. Jones and R. G. Baraniuk, "An adaptive optimal-kernel time-
frequency representation," IEEE Trans Signal Process, vol. 43, no. 10,
pp. 2361-2371, 1995, doi: 10.1109/78.469854.
[17] M. Harbaji, K. Shaban, and A. El-Hag, "Classification of Common
Partial Discharge Types in Oil-paper Insulation System Using Acoustic
Signals," IEEE Trns. Dielectr. Electr. Insul., vol. 22, no. 3, pp. 1674-
1683, Jun 2015, doi: 10.1109/tdei.2015.7116364.
[18] J. Tang, D. Wang, L. Fan, R. Zhuo and X. Zhang, "Feature parameters
extraction of GIS partial discharge signal with multifractal detrended
fluctuation analysis," IEEE Trns. Dielectr. Electr. Insul., vol. 22, no. 5,
pp. 3037-3045, October 2015, doi: 10.1109/TDEI.2015.004556.
[19] Y. Wang, J. Yan, Z. Yang, Y. Zhao, and T. Liu, "Optimizing GIS partial
discharge pattern recognition in the ubiquitous power internet of things
context: A MixNet deep learning model," Int J Electr Power Energy Syst,
vol. 125, Feb 2021, Art no. 106484, doi: 10.1016/j.ijepes.2020.106484.
[20] Zhang A, Lipton Z C, Li M, "Dive into deep learning," arXiv preprint
arXiv:2106.11342, 2021.
[21] K. He, X. Zhang, S. Ren and J. Sun, "Deep Residual Learning for Image
Recognition," 2016 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2016, pp. 770-778, doi: 10.1109/CVPR.2016.90.
[22] D. Hendrycks, K. Gimpel, "A baseline for detecting misclassified and
out-of-distribution examples in neural networks," arXiv preprint
arXiv:1610.02136, 2016.
[23] Z. Gao, Q. Cai, Y. Yang, N. Dong, and S. Zhang, "Visibility Graph from
Adaptive Optimal Kernel Time-Frequency Representation for
Classification of Epileptiform EEG," Int J Neural Syst, vol. 27, no. 4, Jun
2017, Art no. 1750005, doi: 10.1142/s0129065717500058.
[24] Z. Feng, M. Liang, and F. Chu, "Recent advances in time-frequency
analysis methods for machinery fault diagnosis: A review with
application examples," Mech Syst Signal Proc, vol. 38, no. 1, pp. 165-
205, Jul 5 2013, doi: 10.1016/j.ymssp.2013.01.017.
[25] R. G. Baraniuk and D. L. Jones, " Signal-dependent time-frequency
analysis using a radially Gaussian kernel," Signal Processing, vol. 32, no.
3, pp. 263-284, Jun 1993, doi: 10.1016/0165-1684(93)90001-q.
This article has been accepted for publication in IEEE Transactions on Dielectrics and Electrical Insulation. 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/TDEI.2022.3199189
© 2022 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: CHONGQING UNIVERSITY. Downloaded on September 19,2022 at 11:21:04 UTC from IEEE Xplore. Restrictions apply.
4
> REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <
[26] M. Ghislieri, G. L. Cerone, M. Knaflitz, and V. Agostini, "Long short-
term memory (LSTM) recurrent neural network for muscle activity
detection," J NeuroEng Rehabil, vol. 18, no. 1, Oct 2021, Art no. 153,
doi: 10.1186/s12984-021-00945-w.
[27] A. Howard et al., "Searching for MobileNetV3," 2019 IEEE/CVF
International Conference on Computer Vision (ICCV), 2019, pp. 1314-
1324, doi: 10.1109/ICCV.2019.00140.
Zhixian Zhang was born in Chongqing,
China, in December 1993. He received his
B.Sc. degree from Chongqing Jiaotong
University in 2016 and his M.Sc. degree
from Chongqing University, Chongqing,
China in 2019. Currently, he is pursuing his
Ph.D. degree at the School of Electrical
Engineering at Chongqing University. His
primary research interests include partial
discharge detection, fiber-optic sensing, battery condition
assessment, signal processing, and machine learning.
Weigen Chen was born in Zhejiang, China
in August 1967. He received his B.Sc.,
M.Sc., and Ph.D. degrees in Electrical
Engineering from Chongqing University.
Currently, he is a Professor at the School of
Electrical Engineering at Chongqing
University. His primary research interests
include online monitoring and fault
diagnosis of power equipment, condition-
based maintenance, and the internal insulation and thermal
properties of power transformers.
Kejie Wu was born in Sichuan, China, in
January 1998. He received his B.Sc. degree
from North China Electric Power
University, Beijing, China, in 2020.
Currently, he is pursuing his M.Sc. degree
at the School of Electrical Engineering at
Chongqing University. His main research
interests include online monitoring and
fault diagnosis of power equipment and
partial discharge fiber optic sensing technology.
Haoyuan Tian was born in Shandong,
China in November 1999. He received the
B.Sc. degree from Harbin University of
Science and Technology, Heilongjiang,
China in 2021. He is currently pursuing his
M.Sc. degree at the School of Electrical
Engineering at Chongqing University. His
primary research interests include
detecting the temperature, vibration, and
partial discharge of electrical equipment through fiber optic
sensors.
Ruimin Song was born in Chongqing,
China in April 1997. He received his B.Sc.
degree from Chongqing University,
Chongqing, China in 2019. He is currently
pursuing his Ph.D. degree at the School of
Electrical Engineering at Chongqing
University. His primary research interests
include detecting the aging product of
internal insulation, and the use of Raman
spectroscopy to diagnose the aging stage of the oil-paper
insulation.
Yuxuan Song was born in Heilongjiang,
China, in 1998. He received his B.Sc.
degree from Chongqing University,
Chongqing, China, in 2020. Currently, he
is pursuing his Ph.D. degree at the School
of Electrical Engineering at Chongqing
University. His primary research interests
include detecting the temperature,
vibration, and partial discharge of
electrical equipment through fiber optic sensors.
Hong Liu was born in Chongqing, China in
1999. He received his B.Sc. degree in
Electrical Engineering from Southwest
University. He is currently studying for a
M.Sc. degree in Electrical Engineering at
Chongqing University. His primary
research interests include gas detection by
nanosensor, online monitoring and fault
diagnosis of power equipment.
Jianxin Wang was born in Jiangsu, China,
in 1994. He received his B.Sc. degree from
Chongqing University, Chongqing, China,
in 2017. Currently, he is pursuing his Ph.D.
degree at the School of Electrical
Engineering at Chongqing University. His
primary research interests include the
detection of trace features with enhanced
Raman spectroscopy and fault diagnosis of
power equipment based on deep learning.
This article has been accepted for publication in IEEE Transactions on Dielectrics and Electrical Insulation. 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/TDEI.2022.3199189
© 2022 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: CHONGQING UNIVERSITY. Downloaded on September 19,2022 at 11:21:04 UTC from IEEE Xplore. Restrictions apply.