This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2020. Doi Number

Partial Discharge Localization in 3-D with a

Multi-DNN Model Based on a Virtual
Measurement Method
Atta Ul Munim Zaki1,2, Yue Hu1, Xiuchen Jiang1
1
Department of Electrical Engineering, Shanghai Jiao Tong University, No.800 Road Dongchuan, Shanghai, 200240, China
2
Department of Electrical Engineering, University of Azad Jammu and Kashmir, Muzaffarabad, 13100, Pakistan

Corresponding author: Yue Hu (e-mail: [email protected]).

This work was supported by the NSFC (51777122) and Natural Science Foundation of Shanghai, China (17ZR1414400).

ABSTRACT Time difference of Arrival (TDOA)-based localization method, although used widely, calls for
a fast and accurate solution owing to its time inefficiency and sensitivity to time delay estimation. In order to
speed up the solution for nonlinear TDOA equations, while guaranteeing the location accuracy, this paper
presents a hybrid approach namely multi-deep neural network model based on a virtual measurement method
(MDNNM-VMM). Data consisting of multiple time difference values, resulting from a virtual measurement
method (VMM), are fed to a pretrained multi-deep neural network model (MDNNM). These ‘n’ number of
virtually generated sequences of time delays are obtained from a single set of TDOA equations, while
conforming with a uniform distribution. The multi-DNN model using these data, outputs the required partial
discharge (PD) coordinates that help determine accurate PD location. While applying a measurement error
of 24 ns, the average error values Δr, Δθ, ΔФ, and Δd for the proposed method, compared to a multi-DNN
method, see a significant percentage decrease of 32 %, 24 %, 39 %, and 44 %, respectively. Additionally,
varying simulated error, different array designs, and certain other parameters are studied to make the PD
localization process more efficient and multifaceted.

INDEX TERMS Deep Neural Network (DNN), Partial Discharge (PD) localization, Time difference of
arrival (TDOA), Ultra-High Frequency (UHF), Virtual Measurement Method (VMM).

I. INTRODUCTION
Insulation condition monitoring in substations through the heart of these methods lies the concept of measuring
partial discharge (PD) detection has drawn great attention distance amongst sensors and PD source(s). Based on angle
since the last decade [1]-[2], as PD signals contain an ample and time, or strength information of UHF signals, UHF
amount of information regarding the insulation status of methods can either be based on Angle of Arrival (AOA),
electrical equipment [3]-[4]. PD is not merely an indicator to Time of Arrival (TOA), Time Difference of Arrival
degrading insulation level, rather it warns about an alarming (TDOA), or Received Signal Strength Indicator (RSSI) [8]-
condition that could henceforth lead to a complete [10]. TDOA-based method has already received more
breakdown. Therefore, detecting PDs is a must to carry out
attention than other methods due to its comparatively high
insulation diagnosis; and, localizing them could help avoid
accuracy. However, TDOA-based method has the curse of
complete insulation failures [5]-[6].
solving non-linear equations through iterative methods [11]-
One of the important problems with PD detection is its
[12], which in addition to being time consuming is
source localization, which is necessary for a fast and accurate
computationally inefficient at times (as the solution could be
fault detection [7]. Ultra-high frequency (UHF) sensor array-
non-optimal) [7].
based methods have been widely applied owing to the
To avoid these shortcomings, a shift toward data-driven
powerful characteristics of UHF electromagnetic waves,
methods was noticed. For this purpose, a number of methods
namely stable transmitting speed and anti-interference. At
were put forward that were based on shallow architectures,

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access
A. M. ZAKI: Partial Discharge Localization in 3D with a Multi-DNN Model Based on a Virtual Measurement Method (April 2020)

like Artificial Neural Networks (ANNs) and Support Vector detection problem. In addition to considering aspects related
Machines (SVMs). Mas’ud [1], in addition to encompassing to array size and shape etc., the proposed method presents a
the benefits of ANNs, details a number of well-known ANN- generalized methodology that combines the pros of previous
based architectures that have been put forward by scholars methods and avoids their corns, while comparing ‘N’ PD test
for the PD detection over the past years. These include points against ‘N’ PD predicted points. This method, unlike
Modular Neural Networks (MNNs), Ensemble Neural the conventional or contemporary methods, does not require
Networks (ENNs), Probabilistic Neural Networks (PNNs), massive datasets for predicting PD sources. Only a single
and Radial Basis Function Networks (RBFNs). Multiple TDOA equations’ set is required to predict each PD source
RBFNs has also been successfully applied by Nan Zhou [13]. efficiently and accurately. Neither repetitive measurements
Despite giving appreciable levels of accuracy, ANNs in fact nor computational burden is needed; therefore, this method
lack when selecting the feature vector(s) and during the is both time-efficient and cost-effective.
explanation of results. These shortcomings and limitations In order to validate this method, after getting a TDOA
associated with ANNs [1] called for some new and improved equations’ set for each randomly generated PD source, firstly
approaches. This created a room for Deep Neural Networks an ‘n’ number of similar sets of TDOA equations are
(DNNs), as they had already proven to be efficient in various obtained statistically, through a Virtual Measurement
domains, and were far better than ANNs [13]-[15]. Method (VMM). Secondly, these datasets of time difference
Reference [7] has presented a fast and efficient approach equations are fed to a multi-DNN model (MDNNM) that
based on single and multiple DNNs, including both 2-D and consequently predicts the PD location coordinates (r, θ, Ф)
3-D cases, that is by far the most promising and for all the PDs. Accuracy is ensured by VMM, while a fast
comprehensive attempt; as it includes various array designs, solution is guaranteed by MDNNM.
parameters related to array size, and different simulated error After the introduction, section 2 explains the theoretical
types added to the time delay sequences. The basis of this aspects of 2-D and 3-D models for the PD localization,
approach is to bypass the computational work required for including the details related to the VMM. Section 3 shows
solving nonlinear equations while considering massive data. how VMM helps fulfil the data demand of MDNNM.
PD sources are randomly generated initially as test cases and Subsequently, section 4 analyzes the obtained results and
distance between each sensor of the antenna array and every gives a comparative discussion with some previous works;
PD source is calculated by considering the time difference
and finally, the work ends with a conclusion section
amongst the electromagnetic signals reaching different
sensors, with the prior knowledge that speed of
II. LOCALIZATION METHOD
electromagnetic waves approximately equals the speed of Two major goals while working with PD localization
light. These time delay sequences are thereafter fed as input problems are their precision and speed [6], [9]. Therefore, to
to a multiple DNN-model, and the output of model for the ensure an accurate detection of PD source, a statistical
obtained PD source locations are the predicted coordinates simulation-based VMM is proposed in this section; whereas,
(r, theta, and phi). Average error between the test and for speedy detection of PD source a Multi DNN model is
predicted PD locations showed excellent performance, introduced in the next section. VMM-based mathematical
especially for the Y-shaped array. However, there is a model used in this paper has been introduced previously by
problem with this approach, that it depends on availability of [9], which itself is an improved version of extended 2-D
measured time delay data during its testing phase for its model presented by [10].
implementation. This demands repetitive measurements that The objective of proposing this Multi DNN Model based
are definitely time consuming. This weakness has been on Virtual Measurement Method (MDNNM-VMM) is to
overcome by [16], while extending and improving the ideas suffice the process of detecting PDs in the real-world
given by [17] and [18] that were limited to 2-D analysis. scenario. Obtaining a real-world dataset for PD
Reference [16] fulfilled the demand of massive data through measurements, especially with regards to those occurring in
a Data Enhancement Method (DEM) based on statistical the substations, is not always an easy task; that is why,
simulations. It also guaranteed a prompt solution for PD- publicly available standard datasets could hardly be found.
detection problem unlike [17] and [18], by considering a In addition to being time consuming, the process is costly
DNN model for both 2-D and 3-D cases. Furthermore, the as well. Repetitive measurements are generally needed to get
criticality of accurate time delay estimation, as highlighted the exact PD locations. These repetitive measurements
by [19], was efficiently dealt with. Thus, [16] was a hybrid generate bigdata, which requires costly arrangements for its
approach ensuring a fast and accurate solution. Despite storage and processing. Likewise, is the case for already
introducing a new idea, it was a specific case having available practical datasets, which require denoising and/or
limitations due to considering merely a single PD source; pre-processing prior to becoming useful for the PD detection.
therefore, it required generalization for its true applicability. Our approach chiefly provides an alternate solution in this
Most of the afore-mentioned PD detection approaches scenario, as it just requires a single set of TDOA equations;
either focus on accuracy or deal merely with speed. This that is, a single time measurement for the time difference
work, on the other hand, gives a wide-ranging and multi- values amongst array sensors for each PD. Furthermore, in
faceted solution for a fast and quick PD location and
2
VOLUME XX, 2020

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access
A. M. ZAKI: Partial Discharge Localization in 3D with a Multi-DNN Model Based on a Virtual Measurement Method (April 2020)

order to take into account, the complete range of possible d1 − d 4 = c(t140 +  14 +  '14 )
TDOA values, statistically generated datasets obey a
standard distribution (whether normal or uniform). d 2 − d3 = c(t230 +   23 +  ' 23 )
Moreover, different standard deviations, with different d1 − d 2 = c(t120 +  12 +  '12 ) (1)
ranges, are chosen so as to better replicate the system-related
errors and possible time delay errors. d 4 − d3 = c(t430 +   43 +  ' 43 )
Adding different values for the standard deviations also d 2 − d 4 = c(t240 +   24 +  ' 24 )
serves as an indirect way to knowing the strength of the
proposed method, as we can figure out that minimum value d1 − d3 = c(t130 +  13 +  '13 )
of Tm which can locate PD source(s) stably. Availability of
accurate data is not enough in getting a fast solution; and that Two additional parameters in (1), representing standard
is why, some recent researchers [17] working on statistical deviations 𝜎𝜀𝑖𝑗 and 𝜎′𝜀𝑖𝑗 respectively, denote the deviations
methods also leveraged iterative methods to locate PD caused by ‘error of TDOA measurement system’ and
locations. Iterative methods are well-known to be time- ‘supposed additional error for TDOA’. It is a known fact that
inefficient. At this stage, our approach uses a pretrained there is a difference between actual and measured values;
MDNNM that ensures a quick localization of PD(s). This thus, 𝜎𝜀𝑖𝑗 is inserted in the equations. As ɛi represents the
way the proposed approach ensures cost-effectiveness and measurement time error associated with P and Si, so
time-efficiency. measured time difference between time of arrival associated
with (P, Si) and (P, Sj) is given as ‘tij0+ij’, instead of simply
A. MATHEMATICAL MODELLING FOR PD
tij0. Moreover, here: ij=i-j. The measured value (that is,
LOCALIZATION
time difference between a pair of sensors) when multiplied
TDOA measurements are believed to be a simple and most
widely accepted means of evaluating the PD location by with the speed of light gives the distance between P and
straightforwardly solving certain simultaneous nonlinear relevant sensors (because UHF signals have approximately
equations. Four sensors are used to get the exact solution for same speed as that of light). The calculated distance
3-D case; that is, 6 TDOA measurements. Modelling the 3- difference equations for 2-D case, when solved through
D case involves complexity, as it considers the arrangement iterative methods, output the PD location angles. These are
types of antenna array [7]. plotted as angle histograms, with their highest peak regarded
Fig. 1 shows basic 2-D model for the PD localization. S1- as the PD source location. PD is located on the hyperbola
S4 denote four UHF sensors; while, P represents the PD with the corresponding coordinates of two sensors as focuses.
source with rectangular coordinates (x, y), and phasor Selecting different sensors results in different hyperbolas,
coordinates (r, θ). Proceeding in the anticlockwise direction, and their intersection point is theoretically the PD location.
the coordinates for S1, S2, S3, and S4 are (l, w), (-l, w), (-l, In hyperbolic equations (2), the major axial length and short
-w), and (l, -w), respectively. di is the distance between P and axial length of hyperbola are respectively 2aij and 2bij; while,
Si; whereas, the corresponding real time of arrival from P to the focal distance of hyperbola (decided by considering the
Si is ti. tij0 (as described in Fig. 2) equals ti-tj, and finally c shape and size of UHF array) is given as cij. The former
stands for the speed of light. system-related deviation σεij in (1), caused by ij, gets
For accurate PD detection in 3-D case, 6 TDOA affected by factors such as: measuring accuracy of arrival
measurements (obtained from 4 sensors) are considered. time, system’s sampling rate, and sensing ability of UHF
sensors. The latter deviation 𝜎′𝜀𝑖𝑗 caused by ′ ij, which is
related to difference in the arrival time, helps in finding out
its consequent effects on the PD location.

y 2 ( x − l )2 x 2 ( y − w) 2
2
− = 1; − =1
a14 b142 2
a12 b122
y 2 ( x + l )2 x 2 ( y + w) 2
2
− 2
= 1; 2
− 2
=1 (2)
a23 b23 a43 b43
FIGURE 1. Mathematical model of 2-D PD localization di − d j c(tij 0 +   ij +  ' ij )
aij = = ; bij = cij2 − aij2
3-D modelling involves a certain level of complexity [6], 2 2
as different array designs are used in addition to uniform
circular array (UCA). The rectangular coordinates for 3-D ij (causing a deviation in the measurements) has to have
case are given in Table I, and (1) corresponds to the an upper bound Ts, such that ij follows a specific gaussian
corresponding equations. distribution on [-Ts, Ts], as the naturally occurring
distribution is a gaussian distribution. Although unknown in

3
VOLUME XX, 2020

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access
A. M. ZAKI: Partial Discharge Localization in 3D with a Multi-DNN Model Based on a Virtual Measurement Method (April 2020)

an actual measurement, but as per assumption, the theoretical (evaluated time delay datasets) are not estimated properly,
value for time difference should be located in the range [-Ts, that would lead to false prediction of PDs.
Ts] (see Fig. 2). An appreciable effort regarding forecasting of stock price
It is noteworthy that a single set of time delay equations index using DNN has been given by [21], but it depends
obeys the normal distribution. On the other hand, the latter mainly on the past data. This creates a room for a statistical
deviation is considered to be a uniform distribution, as simulations-based method known as VMM, which fulfils the
supposed by [10], and proved to be stable (especially for θ quality data demand of DNN model(s) with merely a single
coordinate) by [9] also. To obtain a desired number of virtual set of TDOA equations’ set, subsequently making the
measurements for the VMM, in addition to studying the processing time efficient. Similar efforts based on statistical
impact of ‘error in time difference of arrival’ on the PD data evaluation could be seen in [22]-[24], where the authors
location, an error following a uniform distribution 'ij~U(- discuss multi-line outage identification, incipient building
Tm, Tm) is deliberately added to the measured tij0+ij. The faults, and short-circuit diagnosis, respectively.
upper bound of this purposely included simulated error is Tm. The data being fed to the DNN model is really critical for
The theoretical value for time difference must be located in the PD localization problem, because it directly affects the
the range (tij + ij -Tm, tij + ij +Tm), provided it is chosen output prediction (as illustrated by Fig. 4). Therefore, adding
properly. Following this pattern, for each amongst the ‘N’ varying amount of additional simulated error, and
PD sources generated randomly, an error-based time monitoring the prediction process beforehand, helps realize
sequences’ set is fed as input to the Multi-DNN model, the stability and robustness of the MDNNM-VMM in a real
which finally outputs the set of PD source coordinates. scenario. The detailed implementation of MDNNM-VMM
algorithm is shown in Fig. 3.
Another worth mentioning aspect in Fig. 3 is the
assignment of initial value, that is, choosing the values for
tij0, which being positive/ negative, decide about the quadrant
of PD source. For D1 (UCA), this is explained in Table II;
whereas, the other designs also possess similar
characteristics.

III. DEEP NEURAL NETWORK METHODS FOR THE PD

LOCALIZATION

FIGURE 2. Relationship among TDOA of theoretical value,measurement

A. DIFFERENCE BETWEEN SINGLE AND MULTI-DNN
value and simulation value of PD MODELS
Once, when a sequence consisting of six time-delays is
B. 3-D SENSOR ARRAY CONFIGURATIONS generated through VMM, it is fed as input to a pre-trained
Designs for 3-D array configurations have previously been DNN model that outputs three coordinates (r, θ, Ф); and this
detailed by [20], while skipping the uniform linear array happens with four antennas. The model presented in this
(ULA) due to its dramatically changing accuracy [7]. section is different from the single-DNN based model [16],
Table I lists the coordinates (x, y, z), for each amongst four as it is based on multi DNNs. Despite following [7], this
sensors, within a single array design. Simple formulas model helps overcome the shortcomings possessed by that
related to conversion can be used to switch from rectangular DNN-based approach. For satisfying the ‘immense-data-
to phasor coordinates or vice versa. All antennas are availability’ requirement, instead of averaging the repetitive
arranged on a spherical surface, which has a radius R. measurements to get minimum error-based sets of TDOA
Moreover, antennas are linearly and vertically polarized, and equations, this method makes the output of VMM method as
its input. This way, by overcoming the shortcomings and
omnidirectional with identical response.
benefitting from the strengths, the proposed model optimizes
The designs shown by Table I are: uniform circular array
the performance of previous methods [7], [16], and [17].
(D1), Y-shaped array (D2), and Novel Right Triangle
Moreover, the single-DNN model was used just with a
Pyramid-shaped array (D3) [7].
single PD source, where the theoretical coordinates were
initially known. Furthermore, after getting a single TDOA
C. VIRTUAL MEASUREMENT METHOD
set of time delays, and generating the desired number of
Owing to nonlinearities involved while solving TDOA
statistical simulations through a DEM, data were fed directly
equations, the solution process becomes time inefficient. To
overcome this situation, DNN-based techniques are as input to the single pre-trained DNN model. The output
incorporated that are better known for their capability of was plotted as histograms and kernel density estimation
solving nonlinearities efficiently. But the problem with (KDE) plots, separately, for all three coordinates; where, the
DNN-based algorithms during their testing phase is that they most frequently occurring peak was regarded as the PD
are sensitive to the quality of input data. If the data provided position.

4
VOLUME XX, 2020

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access
A. M. ZAKI: Partial Discharge Localization in 3D with a Multi-DNN Model Based on a Virtual Measurement Method (April 2020)

Finally, all three obtained coordinates were compared 1) Randomly generate a training set ‘ytrain’, containing m
against the set theoretical coordinates, and error was found training samples and three coordinates (r, θ, Ф), such
for different values of simulated Tm. That method worked that the size of the set is [m,3]. Similarly, generate
well, but was confined to just a single PD. This motivated another set for testing ‘ytest’, containing n testing
the authors of the present work to bring an improvement by samples and three coordinates (r, θ, Ф), such that the size
generating ‘N’ PD points randomly, instead of merely a of this set is [n,3]. Declare both these sets as an output
single point. The scheme for training and testing these ‘N’ for the training and testing of multi-DNN model,
randomly generated PDs (with three coordinates each, as the respectively.
case is 3-D) is given in the forthcoming subsection. 2) Transform both the above sets from phasor (r, θ, Ф) to
rectangular coordinates (x, y, z) with the help of simple
3-D conversion formulas.
TABLE I 3) Calculate the distance ‘di’ between ‘rectangular
RECTANGULAR COORDINATES FOR 3-D ARRAY DESIGNS
coordinates of PD source (x, y, z)’ and ‘rectangular
Type Axis S1 S2 S3 S4
coordinates of each amongst the four antennas S i (xi, yi,
D1 X 1 1 1 1 zi)’; where: i=1, 2, 3, and 4. Coordinates for the four
R − R R − R sensors are taken in compliance with Table I for any of
2 2 2 2
Y 1 1 1 1 the chosen antenna array designs.
R R − R − R
2 2 2 2 4) While using the calculated ‘di’, its corresponding time
Z 0 0 0 0 ‘ti’ is calculated by the relation ti=di/c; where, ‘c’ is for
D2 X 0 3 3 0 the speed of light that approximates the UHF signals (as
R − R
2 2 UHF signals have got approximately same speed as that
Y 1
R
1
− R
1
− R
0 of light).
2 2 2 5) Then, calculate the corresponding time delay sequence
Z 0 0 0 0
‘tij’ between four antennas using the relation tij=ti-tj. This
D3 X 0 2 2 2 2 results in two sequences, separately for the train and test
− R 3
R
3
R
data sets- that are regarded as the input for the training
3
Y 0 0 6 6
and testing of DNN-based model, respectively. The
R − R training set has ‘m’ number of samples and six time-
3 3
Z R
difference values, such that x-train input has the size
1 1 1
R R R [m,6], that is, x-train (t12, t43, t14, t23, t24, t13). Similarly,
3 3 3 the testing set has ‘n’ number of samples and six time-
difference values, such that x-test input has the size [n,6],
that is, x-test (t12, t43, t14, t23, t24, t13). Generally, ratio for
training samples versus testing samples is 70/30 %.
6) Add random simulated error to the TDOA sequence, and
generate the desired dataset in accordance with the
TABLE II
procedure described for the VMM in section II.
DATA CHARACTERISTICS OF TIME DELAY SEQUENCE
C. CLASSIFIER
Case t120 t140 θ range Quadrant The classifier, shown in Fig. 4, has been added between the
error carrying time delay sequences and multi-DNNs in
1 -ve -ve θ ∈ (0°, 90°) 1st order to reduce the error between test (randomly generated)
2 -ve +ve θ ∈ (90°, 180°) 2nd
coordinates and predicted coordinates. It is done by
3 +ve +ve θ ∈ (180°, 270°) 3rd
4 +ve -ve θ ∈ (270°, 360°) 4th
following the data characteristics of time delay sequence as
shown by Table II; the basis of which relies on the idea that
tij being positive or negative decides the quadrant of PD
source. Therefore, while taking into account estimated θ
range initially, the sequences for D1, and D2-D3 can
B. TRAINING AND TESTING PHASE STRATEGY respectively be classified as 4 and 3. θ range is realized on
This sophisticated scheme has been borrowed form [7]. A basis of a trained classifier. The number of DNNs for multi-
careful approach is needed for all this process because, as per DNN based model is same as the output size of the classifier.
Fig. 4, the input to Multi-DNN model directly affects the Firstly, each neural network (NN) gets trained with sample
output, which is actually the prediction result of PD location. datasets having specific θ range; and afterward, for the
Thus, the more properly the training and testing of model is prediction purpose, the most suitable DNN amongst multi-
carried out, the more efficiently this pre-trained-pre- NNs is used. In this manner, the optimal result obtained for
classified-Multi-DNN model would predict the PD location. PD source has less error as compared to non-classified
The scheme follows through as: single-DNN model used by [16].

5
VOLUME XX, 2020

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access
A. M. ZAKI: Partial Discharge Localization in 3D with a Multi-DNN Model Based on a Virtual Measurement Method (April 2020)

IV. RESULTS AND ANALYSIS

This section shows the performance of the MDNNM-VMM
by carrying out various simulations and comparing the
obtained results with some of the latest research works [7]
discussing MDNN. Furthermore, it also compares the
performance of different array designs by inserting varying
values for the simulated error. It is to be noted that the
coordinates of PD sources for all simulations were randomly
generated with some limits: 𝑟 ∈ (0𝑚, 60𝑚) , 𝜃 ∈ (0°, 360°) , 𝜑 ∈
(0°, 90°) . Additionally, R (that is the radius of sphere
introduced in Table I) for all the array designs is set as 18 m
in order to get a smooth comparison.
For the comparison of coordinates, average results based
on (3) are used. The set of coordinates for each randomly
generated PD source is (rtest, θtest, Фtest); whereas, the set for
the predicted coordinates by the pre-classified pretrained
multi-DNN model is (rpredicted, θpredicted, Фpredicted).
The deviation between these two sets, with respect to the
number of test samples, is called the mean absolute error
(MAE). This error, which is based on (3), determines how
efficient the proposed model works for different array
designs and varying simulated errors.

r = rpred − rtest
 =  pred − test
 =  pred −  test (3)
d = d ( S pred − Stest ) OR
d = ( x pred − xtest ) 2 + ( y pred − ytest ) 2 + ( z pred − ztest ) 2

FIGURE 3. Algorithm flowchart for the PD location with MDNNM-VMM

A. EFFECT OF MEASUREMENT ERROR ON LOCATION
ACCURACY
In order to monitor the robustness and efficiency of the
MDNNM-VMM, this section fixes the system error Ts at 2
ns, and varies the simulated error up to a large value of 24
ns. Afterward, the results are compared against the results of
[7]. The test case is set at N=200 for the simulation studies.
Error is applied to two time delays, that are, t12 and t43.
1) EFFECT OF VARYING SIMULATED ERROR Tm ON
INDIVIDUAL PD COORDINATES
Different values of simulated error are considered here. Prior
to deeply pondering on the individual PD coordinate
behavior (see Figs. 5(a)-5(c)), while comparing the test and
predicted PD points, we observe that coordinate θ behaves
quite linearly compared to r and Ф, even for large error
values. Less variations being noted, from a small error value
to large error values, depict the robustness of coordinate θ.
On the other hand, other two coordinates go on spreading
apart as the error values increase; and, certain outliers are
also seen for the Ф coordinate. For an error value of 2 ns,
both these coordinates show a matching linear trend quite
similar to that of θ; but for increasing error values like 10 ns
FIGURE 4. Comparison between pre-trained single-DNN and multi-DNN and 24 ns, there is a significant mismatch amongst the test
models and predicted points. Therefore, it can be concluded that θ

6
VOLUME XX, 2020

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access
A. M. ZAKI: Partial Discharge Localization in 3D with a Multi-DNN Model Based on a Virtual Measurement Method (April 2020)

coordinate has better prediction accuracy compared to the compared to that of Δr, depicts that Δθ coordinate is less
other coordinates. affected by the increase in time delay error. Less number of
Fig. 6 gives a deeper insight of individual PD coordinates by spikes means there are a few cases for which the predictions
illustrating the averaged error results between the test and have gone wrong/ less accurate, unlike the case for Δr.
predicted coordinate values. X-axis represents the number of Figs. 6(g)-6(i) show how coordinate ΔФ changes w.r.t
PD sources that have been predicted by the pre-trained changing Tm values. Its behavior can also be monitored
MDNNM; whereas, Y-axis shows the averaged error corresponding to the values provided by Table III. Its
between the test and predicted PD coordinates. Therefore, performance remains in-between that of the other two
the first value for the error would be against the first test PD, coordinates.
and so on. The spike(s) in these illustrations would show a Once all coordinates are discussed individually, now Fig.
positive error value, provided the difference between the test
7 shows their respective variation, along with the
and predicted coordinates is positive; otherwise, the spike(s)
corresponding distance coordinate, when Tm is varied up to
would be on the negative side.
24 ns. It is seen that for Tm = 24 ns and N = 200 PDs, the
It can be seen in Figs. 6(a)-6(c) that the error value Δr for
r coordinate, when Tm = 2 ns, fluctuates between ± 2.5 m absolute error values Δr, Δθ, ΔФ, and Δd are, respectively,
mostly with some peaks going beyond ± 5 m; whereas, when below 8 m, 8°, 9°, and 7 m. There is a significant percentage
Tm is increased to 10 ns, Δr varies in between ± 5 m for most decrease in error compared to some recent researches [7] that
cases with some peaks surpassing ± 10 m. Likewise, for Tm would be further discussed in the upcoming subsection. It
= 24 ns, Δr has nearly same fluctuation pattern as compared can be seen in Fig. 7 that Δr shows less variations, for
to the 10 ns case, but with a greater number of spikes going increasing Tm values, as compared to Δθ. This can be
beyond ± 10 m. This behavior of Δr is comprehensible when understood by monitoring the trend of ΔФ, which shows
we see the absolute error values for Δr in Table III, that are abnormal behavior for Tm values surpassing 10 ns. Owing
2.53, 4.54, and 7.49, respectively, for Tm values 2 ns, 10 ns, to this behavior, Δθ is affected much more than Δr; which is
and 24 ns. It is noteworthy that the greater number of spikes quite obvious, as strangely changing ΔФ would verily affect
for Tm = 24 ns case has brought a significant variation to the Δθ directly. On the other hand, radial distance would not be
Δr value owing to averaging the results with respect to the affected directly. However, the positive aspect being noticed
number of test cases. If the Tm value is small, it indicates that is that the finally computed factor Δd remains stable no
we are simulating a condition where time delay error is matter how Δr, Δθ, and ΔФ alter. This figure has been
small; whereas, when this value is increased, it means time explored more in the upcoming section, where for each given
delay error is large. Less time delay error means lesser Tm value, the corresponding error value has been compared
interference with other sources and/or lesser noise added to with the error values given by [7].
the UHF signal approaching the array sensors. Now, when
2) COMPARATIVE ANALYSIS OF MDNNM-VMM AND
the number of spikes increases, it specifies that the pre-
MDNN
trained MDNNM is not performing very well. It is because
For further analysis of the proposed method, results obtained
of the fact that large time delay error has affected the quality
are compared against those of [7]. For the comprehensibility
of input data that have to be fed to the MDNNM. This can be
of figures, our method would be regarded as VMM; while,
understood the other way around, that is, for a certain time
that of reference [7] would be called as DNN.
delay error, what is the strength of the MDNNM. Here, once
While applying a measurement error of 24 ns, the average
again, our proposed model excels any practical PD
error values Δr, Δθ, ΔФ and Δd for the proposed method,
measurement scheme, because it takes much time and
compared to the multi-DNN method, respectively see an
excessive measurements for a practical methodology to
approximate percentage decrease of 32 %, 24 %, 39 %, and
know the strength of any scheme; whereas, testing a few
44 %. Percentage change is obtained by (4); while, absolute
random values for Tm can be enough to know the strength of
error values for both methods being compared can be found
our method.
in Table III. The corresponding graphical comparison for all
Figs. 6(d)-6(f) show the variation trend for Δθ. It could be the coordinates can be visualized from Fig. 8 to Fig. 11,
seen in Fig. 6(d) that for Tm = 2 ns, Δθ alters within ± 5° where Tm values range from 0-24 ns.
range, while some spikes go even beyond ± 10°; while, when Fig. 8 shows that for a value of T m = 0 (supposedly no
Tm = 10 ns, Δθ remains approximately within ± 10° range simulated error), the absolute error Δr between the test and
alongside rarely occurring spikes approaching ± 20°. The predicted cases for the proposed method is 2.47 m. This error
trend for Tm = 24 ns is similar to that of previous case, but a value, without considering the simulated error, is because of
few spikes now touch the limits of ± 30°. The absolute error the system-related error Ts that has been taken as 2 ns
values from Table III make this behavior clear, as for Tm throughout. This has already been discussed in section II
values 2 ns, 10 ns, and 24 ns, the respective error values are (under subsection A). Further, as we go on increasing the Tm,
obtained as 4.78°, 6.74°, and 7.52°. A very small difference Δr value slowly increases compared to that of [7]; but this
between absolute error values for Tm= 10 ns and Tm = 24 ns, difference approximately doubles when the Tm approaches
unlike that for Δr again, shows that θ coordinate has a stable 14 ns. This shows that how the performance of proposed
nature compared to that of r coordinate. The behavior of Δθ, method is far less affected by the increasing time delay
7
VOLUME XX, 2020

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access
A. M. ZAKI: Partial Discharge Localization in 3D with a Multi-DNN Model Based on a Virtual Measurement Method (April 2020)

errors. Moreover, no rapid increase in the error values depict the array. Moreover, it is also observed that as the T m goes
that the proposed method is also stable. on increasing, the circular spread of the error vector plot also
Fig. 9, while comparing the Δθ coordinate, shows that the increases. However, this increase in circular spread is not
proposed method performs quite better as compared to [7]; found to be very significant, which again endorses the fact
however, the difference is not so significant as compared to that the proposed method ensures stability and efficiency
that of Δr (see Fig. 8) and ΔФ (see Fig. 9) coordinates. It while making predictions, even for those cases where time
means that MDNN [7] was also good at predicting, provided delays are inclusive of large-scale errors.
θ coordinate is considered only.
C. EFFECT OF DIFFERENT ARRAY DESIGNS ON THE
Fig. 10, unlike Fig. 9, demonstrates that the performance
LOCATION ACCURACY
of [7] compared to our approach was poor for Ф coordinate.
A method cannot solely meet the challenges of accurate PD
As the value of Tm reached 16 ns, the error value ΔФ doubled
detection and localization unless the best array design for its
as compared to our approach. This large error value resulted
true practical applicability is determined. For this purpose,
in a bad overall performance for the MDNN [7] method,
we studied the comparative behavior of sensor array designs
which is quite apparent from Fig.11.
in 3-D (as illustrated by Fig. 13), while applying varying
While comparing all the coordinates individually for
amount of simulated error Tm. In order to compare the
varying simulated error values, our proposed method shows
designs more meaningfully, we have plotted error
greater prediction accuracy, which results in a little
distributions for all the coordinates. The mean and standard
mismatch between the test and predicted PD points. Thus,
deviation values are demonstrated too; and this is done while
Fig. 11 demonstrates that how the proposed method brings
the time delays are subjected to varying error values, that is,
the mismatch half way down from 12 m to 6 m
for Tm ranging from 0-14 ns.
approximately for even a large simulated error value of 24
ns. It is also observed that, unlike MDNN [7], the proposed TABLE III
method does not rapidly undergo a large deviation (after Tm MEAN ABSOLUTE ERROR FOR THE PD COORDINATES
value 10 ns) between the test and predicted cases, rather it
alters slowly; and this fact also endorses the stability and Tm rVMM rDNN θVMM θDNN ФVMM ФDNN dVMM dDNN
efficiency of this method. 0 2.47 2.89 4.77 4.87 4.55 4.41 2.28 2.46
2 2.53 3.12 4.71 5.00 4.57 4.54 2.36 2.65
AverageErrorMultiDNN − AverageErrorMDNNM −VMM 4 3.30 3.93 4.82 5.70 4.10 4.85 2.80 3.29
Percentage change = (4)
AverageErrorMultiDNN 6 3.39 4.83 5.51 6.11 4.41 5.52 2.91 4.05
8 4.09 5.92 6.11 6.55 5.70 6.95 3.68 5.06
The discussion until now has been confined to polar 10 4.54 6.73 6.74 7.35 6.00 7.85 3.99 5.91
coordinate system. Now, Fig. 12 collectively analyzes the 12 4.81 7.67 6.14 7.78 5.76 8.84 4.12 7.02
behavior of r and θ coordinates in cartesian coordinate 14 4.89 8.43 7.30 7.74 5.69 9.00 4.47 7.71
system, while the time delays are subjected to different 16 5.68 9.18 6.55 8.73 5.03 10.34 4.90 8.63
values of Tm. This figure shows an error vector plot between 18 6.24 9.78 7.67 8.70 6.35 10.48 5.51 9.52
the test and predicted cases, with vectors representing the 20 6.86 10.19 8.67 9.25 6.55 11.36 5.79 9.99
errors in-between them. As the test PDs have been randomly 22 7.25 10.48 7.21 9.32 7.49 12.55 6.43 11.01
generated in space, so error vector plot can better visualize 24 7.45 11.01 7.52 9.86 8.04 13.08 6.75 12.01
how far are the predicted PDs situating. This plot gives a
deep insight of the proposed method, while demonstrating TABLE IV
the strength and stability collectively. For Tm = 2 ns, we see COMPARATIVE ANALYSIS AMONGST ARRAY DESIGNS
that error vectors (blue arrows) look like small dots in the
central region, which means there is a small difference Tm 2 4 6 8 10 12 14
between the test and predicted PDs. As we move away from r1 4.57 4.89 5.84 6.33 6.55 6.71 6.81
the center, the error vectors start getting significant. r2 2.53 3.30 3.39 4.09 4.54 4.81 4.89
Similarly, when Tm = 10 ns, the error vectors (now red r3 6.64 6.58 6.79 7.07 7.16 6.98 8.03
arrows) follow the same trend, but this time the circular θ1_ 4.89 4.94 4.42 7.94 5.56 5.35 7.63
spread has increased. This is due to the fact that for this case θ2_ 4.71 4.82 5.51 6.11 6.74 6.14 7.30
simulated error value is large, and as a consequence, some θ3_ 6.03 5.12 5.76 6.18 6.16 6.72 8.00
predictions go wrong/ far away from the test points. Ф1_ 10.00 10.25 8.26 9.07 11.44 11.25 12.61
Likewise, is the scenario for Tm = 24 ns, which is Ф2 4.57 4.10 4.41 5.70 6.00 5.76 5.69
demonstrated with black arrows as error vectors. Ф3_ 1.93 2.88 2.53 4.19 2.89 3.55 3.90
It is apparent that as we move outwards from the center, d1_ 7.33 7.56 8.43 9.73 10.25 10.89 10.98
we observe error vectors becoming significant. This shows d2 2.36 2.80 2.91 3.68 3.99 4.12 4.47
that those PD sources which are far from the antenna array
d3 4.50 4.68 4.67 4.91 5.07 5.13 5.69
would not be predicted as accurate as those in the vicinity of
8
VOLUME XX, 2020

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access
A. M. ZAKI: Partial Discharge Localization in 3D with a Multi-DNN Model Based on a Virtual Measurement Method (April 2020)

(a) (b) (c)

FIGURE 5. Prediction Versus Test plot for Ts = 2 ns (a) r when Tm = 2 ns, 10 ns, and 24 ns (b) θ when Tm = 2 ns, 10 ns, and 24 ns (c) Ф when Tm = 2 ns,
10 ns, and 24 ns

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

FIGURE 6. Error plot for Ts = 2 ns (a) r when Tm = 2 ns (b) θ when Tm = 2 ns (c) Ф when Tm = 2 ns (d) r when Tm = 10 ns (e) θ when Tm = 10 ns

(f) Ф when Tm = 10 ns (g) r when Tm = 24 ns (h) θ when Tm = 24 ns (i) Ф when Tm = 24 ns

9
VOLUME XX, 2020

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access
A. M. ZAKI: Partial Discharge Localization in 3D with a Multi-DNN Model Based on a Virtual Measurement Method (April 2020)

Starting from coordinate r, we see that D2 not only has D1, which showed a considerable behavior for Δθ,
small absolute error as compared to the other two designs, performs poorly for the ΔФ, with error values distributed
but it also deviates less from its mean value. It means this between 10-12°. Here, D3 performs better than D2 as well,
model is efficient and stable both. On the other hand, D 1 has and D2 lies in between both the other models (as is generally
better properties compared to D3; and this could easily be observed.
understood by consulting Table IV. Δr for D1 has an The overall behavior can be monitored while observing
approximate range 5-7 m, unlike that of D3, which is Δd. It is obvious that D2 tops the other two designs, as its
approximately 7-8 m. Right from the start, where Tm is too standard error distribution and mean value have small
small, D1 and D3 as compared to D2 have twofold and deviation. However, D3 has better performance compared to
threefold value for error. This means that these models are D1; and this is chiefly owing to the ΔФ values, where D3
not efficient because a small deviation (or error in the time performs way better than D1.
delay sequences) should not have this much compromising All studies till now have been carried out on the basis of
impact on the PD accuracy. MAE values, and the results are found to be far better
Moving on to the next coordinate θ, it is observed that all compared to some of the latest methods. Nevertheless, we
designs perform quite well, and the behavior of D 1 and D2 is also obtained the standard deviation-based results, and found
nearly similar. For certain Tm values, D1 has smaller value of that they provide comparatively improved numbers. Table
Δθ compared to D2, but the error deviation from its mean IV gives the comparison between Δds obtained both ways.
depicts that D2 experiences less deviations, which is again a Correspondingly, Fig. 14 shows the comparison through
sign for its robust nature. their respective distribution plots alongside mean values.

FIGURE 7. Effect of varying Tm on PD coordinates FIGURE 9. MDNNM-VMM Vs MDNN w.r.t. θ

FIGURE 8. MDNNM-VMM Vs MDNN w.r.t. r FIGURE 10. MDNNM-VMM Vs MDNN w.r.t. Ф

10
VOLUME XX, 2020

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access
A. M. ZAKI: Partial Discharge Localization in 3D with a Multi-DNN Model Based on a Virtual Measurement Method (April 2020)

V. CONCLUSION
This paper proposes a multi-DNN model (MDNNM) based
on a virtual measurement method (VMM) for the accurate
and prompt detection of ‘N’ PD sources in 3-D. ‘N’ test PD
sources are compared against ‘N’ predicted PD sources; and
averaged error for all coordinates are obtained. Important
conclusions that have been drawn are as follows:

1) The obtained results depict that the proposed approach

is robust and efficient, even for large values of
measurement error Tm. While applying a measurement
error of 24 ns, the average error values Δr, Δθ, ΔФ, and
Δd for the proposed method, compared to the simple
multi-DNN method, respectively see a significant
FIGURE 11. MDNNM-VMM Vs MDNN w.r.t. d approximate percentage decrease of 32 %, 24 %, 39 %,
and 44 %.
2) While relying on single-valued system measurement
error Ts (reflected by a Gaussian distribution), the
varying Tm (reflected by a uniform distribution) fulfils
the requirement of a large dataset for the multi-DNN
model.
3) This helps realize the real PD signals’ measurement
scenario with merely a single value; which is both time
efficient and cost-effective. In addition, changing values
of simulated error helps figure out those error values for
which the proposed approach produces stable results.
4) After incorporating different sensor array designs and
different test sizes, it was revealed that the overall best
FIGURE 12. Error vector plot for r and θ in cartesian coordinate system sensor array design for 3-D is Y-shaped array.
5) If standard deviation-based error is used instead of mean
absolute error, comparatively better results are achieved.
6) Error vector plots both in polar and cartesian
coordinate systems exhibit overall robustness of the
proposed methodology. After a multidimensional study,
MDNNM-VMM was found to outperform multi-DNN
model for the PD detection and localization.

Future works focusing on PD localization, while aiming

to simultaneously improve speed and accuracy, can improve
this VMM-based approach by incorporating VMM during
the training phase instead of testing phase.

REFERENCES
FIGURE 13. Standard deviation w.r.t. different arrays [1] Mas’ud, A., Albarracín, R., Ardila-Rey, J., et al.: 'Artificial Neural
Network Application for Partial Discharge Recognition: Survey and
Future Directions', Energies, 2016, 9, (8), p. 574
[2] Barrios, Buldain, Comech, Gilbert, Orue: 'Partial Discharge
Classification Using Deep Learning Methods-Survey of Recent
Progress', Energies, 2019, 12, (13), p. 2485
[3] Catterson, V.M., Sheng, B.: 'Deep neural networks for understanding
and diagnosing partial discharge data'. Proc. IEEE Electrical
Insulation Conference (EIC), Seattle, WA, June. 2015, pp. 218-221
[4] Li, Z., Luo, L., Zhou, N., Sheng, G., Jiang, X.: 'A Novel Partial
Discharge Localization Method in Substation Based on a Wireless
UHF Sensor Array', Sensors, 2017, 17, (8), p. 1909
[5] Ren, L., Jia, T., Sun, A., Zheng, N., Wu, F., Luo, L.: 'Partial Discharge
Angle of Arrival Estimation Using UHF Wireless Sensor Array'. IOP
Conference Series: Materials Science and Engineering, 2019, 486, (5),
p. 012031
[6] Chai, H., Phung, B.T., Mitchell, S.: 'Application of UHF Sensors in
FIGURE 14. Standard deviation-based error Vs Mean-based error

11
VOLUME XX, 2020

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2993190, IEEE Access
A. M. ZAKI: Partial Discharge Localization in 3D with a Multi-DNN Model Based on a Virtual Measurement Method (April 2020)

Power System Equipment for Partial Discharge Detection: A Review', A.M. ZAKI was born in Rawalakot, Azad
Sensors, 2019, 19, (5), p. 1029 Kashmir, Pakistan, in 1989. He received the B.S.
[7] Liu, J., HU, Y., Peng, H., Zaki, A. ul M.: 'A Data-Driven Method degree in electrical engineering from the
Using DNN for PD Location in Substations', IET Science, University of Engineering and Technology,
Measurement & Technology, 2019 Lahore, Pakistan, in 2011 and the M.S. degree also
[8] Li, Z., Luo, L., Sheng, G., Liu, Y., Jiang, X.: 'UHF partial discharge in electrical engineering from the University of
localization method in substation based on dimension-reduced RSSI Engineering and Technology, Taxila, Pakistan, in
fingerprint', IET Generation, Transmission & Distribution, 2018, 12, 2015. Currently, he is pursuing the Ph.D. degree
(2), pp. 398–405 in electrical engineering at Shanghai Jiao Tong
[9] Zhou, N., Luo, L., Sheng, G., Jiang, X.: 'Direction of arrival estimation University, Shanghai, China.
method for multiple UHF partial discharge sources based on virtual He is on study leave nowadays from the Department of Electrical
array extension', IEEE Transactions on Dielectrics and Electrical Engineering, University of Azad Jammu and Kashmir, Muzaffarabad,
Insulation, 2018, 25, (4), pp. 1526–1534 Pakistan; where he had served as Laboratory Engineer from 2014-2016. His
[10] Zhu, M.-X., Wang, Y.-B., Chang, D.-G., Zhang, G.-J., Tong, X., current research interests include the partial discharge detection and location
Ruan, L.: 'Quantitative comparison of partial discharge localization using virtual measurement methods alongside data-driven approaches.
algorithms using time difference of arrival measurement in substation',
International Journal of Electrical Power & Energy Systems, 2019, YUE HU (M’15) was born in Yueyang, Hunan,
104, pp. 10–20 China, in 1978. He received the B.S. and M.S.
[11] Ataei, M., Osanloo, M.: 'Using a Combination of Genetic Algorithm degrees from Shanghai Jiao Tong University,
and the Grid Search Method to Determine Optimum Cutoff Grades of Shanghai, China, in 2001 and 2007, respectively,
Multiple Metal Deposits', International Journal of Surface Mining, and the Ph.D. degree from the Politecnico di
Reclamation and Environment, 2004, 18, (1), pp. 60–78 Turin, Turin, Italy, in 2010. In 2010, he joined the
[12] Bing Hwa Cheng, Hudson, R.E., Lorenzelli, F., Vandenherghe, L., Department of Applied Electromagnetism,
Kung Foo: 'Distributed Gauss-Newton method for node loclaization in Istituto Nazionale di Ricerca Metrologica, Turin,
wireless sensor networks'. Proc. IEEE 6th Workshop on Signal related to the characteristic of partial discharge
Processing Advances in Wireless Communications, 2005., 2019 (PD) instruments.
[13] Zhou, N., Luo, L., Chen, J., Sheng, G., Jiang, X.: 'Error correction Since 2011, he has been with Shanghai Jiao Tong University. His current
method based on multiple neural networks for UHF partial discharge research interests include the PD detection and location, and nanogrid
localization', IEEE Transactions on Dielectrics and Electrical energy system.
Insulation, 2017, 24, (6), pp. 3730–3738
[14] Karimi, M., Majidi, M., Mirsaeedi, H., Arefi, M.M., Oskuoee, M.: 'A XIUCHEN JIANG was born in Shandong,
Novel Application of Deep Belief Networks in Learning Partial China, in January 1965. He received the B.E.
Discharge Patterns for Classifying Corona, Surface and Internal degree in high voltage and insulation technology
Discharges', IEEE Transactions on Industrial Electronics, 2019, pp. 1– from Shanghai Jiaotong University, Shanghai,
1 China, in 1987, the M.S. degree in high voltage
[15] Bengio, Y., Courville, A., Vincent, P.: 'Representation Learning: A and insulation technology from Tsinghua
Review and New Perspectives', IEEE Transactions on Pattern University, Beijing, China, in 1992, and the Ph.D.
Analysis and Machine Intelligence, 2013, 35, (8), pp. 1798–1828 degree in electric power system and automation
[16] Zaki, A.., Hu, Y.: '2D and 3D Partial Discharge Localization with a from Shanghai Jiaotong University in 2001.
Deep Neural Network Based on a Data Enhancement Method', Currently, he is a Professor in the Department
Submitted to CSEE JPES of Electrical Engineering, Shanghai Jiaotong University. His research
[17] Hu, Y., Zeng, Z., Liu, J., Wang, J., Zhang, W.: 'Design of a Distributed interests are electrical equipment online monitoring, condition-based
UHF Sensor Array System for PD Detection and Location in maintenance, and automation.
Substation', IEEE Transactions on Instrumentation and Measurement,
2019, 68, (6), pp. 1844–1851
[18] Yue, H., Jianwen, W., Xiangyang, P.: 'An improved partial discharge
location method in substations: Error reduction based on Monte-Carlo
simulation'. Proc. Conference on Precision Electromagnetic
Measurements (CPEM), Ottawa, ON, Jul. 2016, pp.1-2
[19] Li, P., Dai, K., Zhang, T., Jin, Y., Liu, Y., Liao, Y.: 'An Accurate and
Efficient Time Delay Estimation Method of Ultra-High Frequency
Signals for Partial Discharge Localization in Substations', Sensors,
2018, 18, (10), p. 3439
[20] Liu, Q., Zhu, M.-X., Wang, Y.-B., et al.: 'UHF antenna array
arrangement optimization for partial discharge direction finding in air-
insulted substation based on phased array theory', IEEE Transactions
on Dielectrics and Electrical Insulation, 2017, 24, (6), pp. 3657–3668
[21] S. Lee and T. Jeong, “Forecasting Purpose Data Analysis and
Methodology Comparison of Neural Model Perspective,” Symmetry,
vol. 9, no. 7, p. 108, Jul. 2017.
[22] Y. Zhao, J. Chen and H. V. Poor, "A Learning-to-Infer Method for
Real-Time Power Grid Multi-Line Outage Identification," in IEEE
Transactions on Smart Grid, vol. 11, no. 1, pp. 555-564, Jan. 2020.
[23] B. Jin, D. Li, S. Srinivasan, S. Ng, K. Poolla and A. Sangiovanni-
Vincentelli, "Detecting and Diagnosing Incipient Building Faults
Using Uncertainty Information from Deep Neural Networks," 2019
IEEE International Conference on Prognostics and Health
Management (ICPHM), San Francisco, CA, USA, 2019, pp. 1-8.
[24] J. Tong, H. Wu, Y. Lin, Y. He and J. Liu, "Fog-Computing-Based
Short-Circuit Diagnosis Scheme," in IEEE Transactions on Smart
Grid.

12
VOLUME XX, 2020

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.