This article has been accepted for publication in IEEE Access.

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/ACCESS.2024.3366905

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2022.Doi Number

Deep Machine Learning-based Asset

Management Approach for Oil-Immersed Power
transformers using Dissolved Gas Analysis
Lan Jin1, Dowon Kim2, Kit Yan Chan3 and Ahmed Abu-Siada4, (Senior Member, IEEE)
1
Electrical and Computer Engineering Discipline, Curtin University, Perth 6102, Australia
2
Electrical and Computer Engineering Discipline, Curtin University, Perth 6102, Australia
3
Electrical and Computer Engineering Discipline, Curtin University, Perth 6102, Australia
4
Electrical and Computer Engineering Discipline, Curtin University, Perth 6102, Australia

Corresponding author: Ahmed Abu-Siada ([email protected]).

ABSTRACT Reliable operation of oil-immersed power transformers is crucial for electrical transmission
and distribution networks. However, the aging of high voltage assets including power transformers along
with the increasing of load demand have heightened the importance of adopting cost-effective asset
management strategies. Dissolved gas analysis (DGA) has been recognized as a valuable diagnostic tool for
detecting potential faults and monitoring the condition of oil-immersed power transformers. Traditional
offline DGA method involves periodic sampling and laboratory analysis, which often results in delayed
detection and response to emerging faults. To address these limitations, online DGA approach has been
emerged to provide real-time monitoring and continuous data acquisition. This paper presents a new asset
management approach for mineral oil-immersed power transformers by analysing the online DGA data using
convolutional neural networks. The proposed approach provides real time solutions to classify emerging fault
type and predict transformer health deterioration level with high accuracy. Results show that the accuracy of
fault diagnostics of the proposed approach is approximately 87%.

INDEX TERMS Power transformers, Dissolved gas analysis, Condition monitoring, Asset management,
Remnant life estimation.

I.INTRODUCTION The fundamental principle of DGA is to measure the levels

The reliable operation of a power system is largely of various dissolved gases in the transformer oil. These gases
dependent on the health condition and performance of its are generated as a result of overheating, arcing and partial
key equipment, particularly power transformers. Faults or discharge events [3]. By analysing the type and concentration
malfunctions in a power transformer can directly impact the of dissolved gases in transformer oil samples, potential faults
safety and reliability of the entire power grid. Therefore, it can be identified, and rate of insulation degradation can be
is essential to develop cost-effective asset management assessed; allowing timely maintenance and repair plans to
methods to assess its health condition and provide a timely prevent potential severe damages to the transformer. The
decision to rectify emerging faults and avoid any potential measured dissolved gases in transformer oil include Hydrogen
catastrophic consequences. Over decades, many condition (H2), Methane (CH4), Ethylene (C2H4), Ethane (C2H6),
monitoring methods have been evolved to detect various Acetylene (C2H2), Carbon Monoxide (CO) and Carbon
faults in power transformers [1]. In industry practice, one Dioxide (CO2). These gases have been considered as key
of the most widely used methods for analysing power indicators of potential transformer faults, and their
transformer oil to detect incipient faults is the Dissolved concentrations can provide valuable insights into the overall
Gas Analysis (DGA) method [2]. This technique has been health condition of the transformer.
proven to be effective in identifying potential issues in
power transformers and has become an essential asset
management tool.

VOLUME XX, 2023 1

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

Several conventional DGA interpretation methods, derived deterioration level of the solid insulation based on the amount of
from ANSI/IEEE standard and IEC publication 599, have been CO2 and CO gases obtained from online DGA measurement
extensively utilized in the power industry [3, 4]. These methods sensors. The fault diagnostic module encompasses “no fault”
include Key gas method, Rogers Ratios, Doernenburg Ratios, condition, in addition to identifying thermal fault, arc discharge
Duval Triangles and Pentagons graphical methods. However, and partial discharge as will be elaborated below.
each of these methods exhibits certain limitations such as out-of-
code ratios, distinct boundaries, and the exclusion of gas II.METHODOLOGIES UTILIZED IN DEVELOPING THE
evolution, which may lead to incorrect and inconsistent fault ASSET MANAGEMENT MODEL
diagnoses [5]. As a result, the accuracy of diagnostic results A. DATA PRE-PROCESSING – NORMALIZATION
relies more on the level of experience of the professionals The aim of data normalization in the data pre-processing stage is
conducting the test. To overcome such subjective interpretations, to bring features into a comparable scale to enhance the model
researchers have developed several artificial intelligence (AI)- performance and improve the training stability [27]. In the
based methods to improve the diagnostic accuracy. In [5], training datasets, the gas concentration exhibits a wide range
various AI-based methods for DGA interpretation have been from 0 to 100,000s ppm. To prevent the dominance of large
presented. One of the methods is Fuzzy Logic, which replaces values on training weights, which could potentially distort the
the precise values of input-output variables with a range of values training results [28], a commonly utilized normalization method
to handle the uncertainties and imprecisions in the DGA data [6, called Minimum-Maximum normalization as given by (1) is
7]. Other AI methods such as Decision Tree [8], Random Forest employed [29]. The Minimum-Maximum normalization
[9], k-Nearest Neighbours (KNN) [10, 11], Logistic Regression technique scales the data into a range between 0 and 1, based on
[12, 13], Support Vector Machine (SVM) [14, 15], Bayesian the minimum and maximum values in the datasets. Figure 1 (a)
Network [16], Artificial Neural Network (ANN) [17-19], depicts the raw data distribution of H2 gas measurements (in
Adaptive Neural Fuzzy Inference System (ANFIS) [20, 21] have ppm) collected from an online DGA sensor, while Figure 1 (b)
also shown promising results in enhancing the analysis of DGA illustrates the data distribution after normalization. Comparison
data. These methods are designed for various levels of data of the two figures shows that the normalization process does not
complexity, often requiring the use of multiple training models alter the essential features of the collected data.
to analyse DGA data comprehensively. For example, ANFIS 𝑋𝑋𝑛𝑛𝑛𝑛𝑛𝑛 =
𝑋𝑋−𝑋𝑋𝑚𝑚𝑚𝑚𝑚𝑚
(1)
𝑋𝑋𝑚𝑚𝑚𝑚𝑚𝑚 −𝑋𝑋𝑚𝑚𝑚𝑚𝑚𝑚
model encounters difficulties and may get stuck in the training
process. On the other hand, fuzzy logic requires ample number where 𝑋𝑋, is the original value before normalization. 𝑋𝑋𝑛𝑛𝑛𝑛𝑛𝑛 is the 𝑋𝑋
of fuzzy rules, which complicate and reduce the accuracy of the value after normalization. 𝑋𝑋𝑚𝑚𝑚𝑚𝑚𝑚 and 𝑋𝑋𝑚𝑚𝑚𝑚𝑚𝑚 represent the
DGA models of multiple input and output parameters. maximum and minimum of values of 𝑋𝑋 value; respectively.
In recent years, another powerful AI tool called Deep Machine
Learning algorithm has emerged to solve complex problems that
were previously challenging for traditional AI methods. Deep
learning is modelled to mimic the hierarchical structure of the
human brain and is designed to process data in a similar way,
starting with lower-level features and gradually building up to
higher-level concepts [22]. By doing so, deep learning models
are able to handle complex problems with large data sets. This
has made them highly effective for a wide range of applications,
including image and sound recognition.
As the concept of future smart grids continues to evolve, the
online monitoring of key assets including power transformers
has become more prevalent. This will lead to a substantial rise in
the amount of sophisticated data being collected and analysed.
Some of these raw data are presented in numerical form, while
others may be on the form of images or sound signals, such as
thermal and vibration analyses [23, 24]. This requires adaptable
diagnostic methods with enhanced learning and feature
extraction capabilities to effectively reflect meaningful insights
from the measured data. In [25, 26], a specific type of deep
machine learning called the probabilistic neural network with
optimizer was developed to provide fault diagnosis in power
transformers based on five DGA gas measurements. The output
of the method identifies four potential fault conditions, including
high/low temperature fault, partial discharge, and arc discharge.
This paper aims to provide a more comprehensive asset Figure 1. (a) Raw H2 Data (in ppm) collected from DGA sensor. (b) Data
management solution for mineral oil-immersed power processed using Minimum-Maximum normalization.

transformers. In addition to the fault diagnostic module, the

developed asset management system also comprises a life
management module to provide asset managers with the

VOLUME XX, 2023 3

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

B. PRE-PROCESSING – BALANCE DATASETS potentially have a significant impact on the outcomes of machine
The training datasets often exhibit uneven distribution, where learning algorithms.
minority classes are vastly outnumbered by majority classes. Another crucial issue that must be considered when
When the model is trained using such imbalanced dataset, utilizing machine learning algorithms is their ability to
machine learning algorithms tend to favour the majority classes, generalize beyond the training datasets. In the construction of
which may lead to potential misclassification of the minority the training datasets, the transformer fault types are typically
classes [30]. Based on the datasets presented in Table 1, it can be identified based on the expert judgements, which are often
observed that the various transformer conditions are not evenly derived from conventional interpretation methods such as
distributed. Specifically, the occurrence of partial discharge IEEE/IEC ratio and Duval Triangles/Pentagons. However, the
faults constitutes a relatively small percentage of all conditions heuristic nature of expert judgments, coupled with the fact
(accounting for only 8.7%). On the other hand, energy discharge that transformers may exhibit multiple faults simultaneously,
faults and thermal faults are much more prevalent, comprising can lead to varying degrees of inconsistency between datasets
approximately 35% and 41.3% of the total conditions, collected from different sources. This inconsistency in the
respectively. Interestingly, the normal condition makes up 15.7% training data can pose a significant challenge to the
of the conditions, which is noteworthy given that in real-world generalization capability of machine learning algorithms and
scenarios, normal conditions tend to be the most commonly may ultimately undermine their effectiveness in practical
observed results [31]. The imbalance nature of the datasets could applications.

TABLE 1
DATASETS FOR FAULT DIAGNOSTIC MODULE TRAINING.

Energy
Reference No fault Partial Discharge Thermal Fault Total # of samples
Discharge
[11] 6 0 5 9 20
[21] 3 0 1 3 6
[7] 3 0 2 6 11
[32] 4 0 0 6 10
[33] 4 0 2 4 10
[34] 6 12 4 12 34
[35] 2 0 1 3 6
[36] 2 0 1 3 6
[37] 8 0 4 12 24
[38] 6 13 3 11 33
[39] 10 5 3 7 25
[40] 28 9 0 38 75
[41] 8 1 1 12 22
[19] 15 8 0 8 31
[42] 3 0 1 4 7
[43] 4 2 0 10 16
[16] 1 2 1 2 6
[44] 10 0 3 23 36
[45] 4 0 7 9 20
[46] 3 1 1 5 10
[47] 4 0 1 6 8
[48] 2 2 0 0 4
[49] 0 4 1 5 10
[50] 4 0 3 2 9
[4] 74 34 9 34 151
[51] 8 6 1 15 30
[2] 5 2 0 13 20
[52] 6 3 3 12 24
Total 233 104 58 274 664
Percentage 35% 15.7% 8.7% 41.3%
To overcome the above-mentioned issues of imbalanced data, through SMOTE processing. Figure 2 (a) provides an overview
Synthetic Minority Over-sampling Technique (SMOTE) has of the distribution of multi-class targets within the initial training
been implemented [53, 54]. The principle of SMOTE is to dataset before the application of SMOTE. Each bin within the
generate synthetic samples for minority classes. It begins by histogram corresponds to a distinct label. Specifically, the x-axis
randomly selecting a data point from the minority class and denotes the labels associated with the dataset. In this context,
identifying its k nearest neighbours. SMOTE then places a label “1” pertains to Thermal fault, label “2” signifies partial
synthetic point along the line connecting the chosen data point discharge (PD) fault, label “4” indicates No Fault, label “8”
and one of its nearest neighbours. These steps are repeated until represents Discharge fault, and label “9” corresponds to a
the dataset is balanced; thereby ensuring a more even combination of Discharge and Thermal faults. The y-axis
representation of all classes in the training data. Figure 2 presents represents the frequency (number of instances) that belongs to
a comparative histogram illustrating the effect of data balancing each label. On the other hand, Figure 2 (b) shows the distribution

VOLUME XX, 2023 3

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

of labels in the training data after applying SMOTE. As can be batch normalization layers normalize the outputs of the previous
observed, the frequency of labels “2”, “4”, “8”, and “9” have layers to enhance training stability and accelerate convergence.
been changed due to the introduction of synthetic samples. Finally, the flatten layers transform the multidimensional feature
maps into a one-dimensional vector. This process prepares the
extracted features for further processing in neural network. The
feature selection process in CNN provides the key advantage
over traditional machine learning algorithms, such as Support
Vector Machine (SVM). CNN can autonomously learn intricate
features and patterns directly from the raw input data that
eliminates the need for manual feature engineering. This
capability significantly reduces the burden of feature extraction
and enhances the overall efficiency of the model.
The neural network component comprises fully connected
layers, which integrate the extracted features and make
predictions based on the learned representations. In Figure 3, a
neural network with 2 hidden layers is depicted. The inputs 𝑥𝑥1
… 𝑥𝑥𝑖𝑖 , contain the features of the input data that are fed into the
network. Within the hidden layers, each neuron takes input from
the previous layer, applies a weight (i.e., 𝑤𝑤𝑖𝑖𝑖𝑖 or 𝑤𝑤𝑗𝑗𝑗𝑗 ) and a bias
(i.e., 𝑏𝑏𝑗𝑗 or 𝐵𝐵𝑘𝑘 ), and passes the results through an activation
function as presented by (2). During the training process, the
backpropagation algorithm determines the weights along with
biases of the neural network to minimize the error of difference
between the predicted output and the target or desired output.
This adjustment is performed using an optimization algorithm,
such as Adaptive Moment Estimation. The optimization
algorithm updates the weights based on the calculated error and
the network’s learning rate, which controls the step size of the
weights updates.
Figure 2. (a) Data distribution before SMOTE and (b) Data distribution after
SMOTE. (for reference to label numbers, see Table 3)

C. CONVOLUTION NEURAL NETWORK

Convolution Neural Network (CNN) has emerged as a popular
and effective deep learning technique, particularly for analysing
visual data [23]. In recent years, there has been growing interest
in leveraging CNN for condition monitoring applications. A
study published in [23] has utilized CNN to identify six types of
PD faults in power transformers by analysing Phase-Resolved
Partial Discharge (PRPD) signals captured by Ultra-High
Frequency (UFH) sensors. Another investigation in [24] focused
on using CNN to assess transformer winding conditions through
the analysis of vibration signatures. Furthermore, [55] explored
the application of CNN in assessing oil quality based on oil aging
images. Considering the future advancements and ongoing
developments of online condition monitoring methods, CNN Figure 3. General structure of neural networks with 2 hidden layers.
holds a great potential as a powerful tool that can provide more
comprehensive condition assessment of power transformers in Activation functions play a critical role in transforming the
real-time. This will provide a more accurate and holistic weighted sum of inputs and bias into an output value. For
understanding of transformer performance, facilitate timely example, the Rectified Linear Unit (ReLU) activation function is
maintenance interventions, and extend the operational lifespan of commonly used in neural networks, which results in the input
power transformers. value if it is positive, and zero otherwise. Another example is
The structure of a CNN comprises two primary components: Sigmoid function, which squashes the input value into a range
feature selection and conventional neural network (Figure 3) as between 0 and 1, in case output needs to be interpreted as
shown in Figure 4. The feature selection layers encompass probabilities.
convolutional layers, pooling layers, batch normalization layers, 𝑌𝑌𝑗𝑗 = 𝑓𝑓1 (∑𝑖𝑖𝑗𝑗=1(𝑤𝑤𝑖𝑖𝑖𝑖 × 𝑥𝑥𝑖𝑖 + 𝑏𝑏𝑗𝑗 )) (2)
and flatten layers [56]. The convolutional layers play a
significant role in feature extraction, applying filters or kernels where, 𝑌𝑌𝑗𝑗 is the output of the first hidden layer. 𝑥𝑥𝑖𝑖 is the
to the input data to capture patterns and spatial dependencies. input. 𝑓𝑓1 is the activation function. 𝑤𝑤𝑖𝑖𝑖𝑖 is the weight and 𝑏𝑏𝑗𝑗 is
The pooling layers reduce the spatial dimensions of the resulting the bias.
feature maps; effectively summarizing the learned features. The

VOLUME XX, 2023 4

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

Figure 4. The structure of Convolution Neural Network.

III.PROPOSED ASSET MANAGEMENT MODEL

The asset management approach proposed, as depicted in Figure In this section, a detailed explanation of the training process
5, utilizes gas measurements acquired from an online DGA for each module is presented, along with insights into the
sensor. This model consists of two distinct sub-modules: a Fault possible outputs derived from these modules.
Diagnostic module and a Life Management module.

Figure 5. Proposed workflow for the oil-immersed power transformer asset management model.

A. FAULT DIAGNOSTIC MODULE bubbles or voids. During PD activities, air or nitrogen in the gas
There are six primary types of faults that can be identified using phase undergoes ionization, forming a plasma of ionized oxygen
DGA method, as outlined in Table 2 based on the IEC 60599 and and nitrogen atoms.
IEEE57.104 [3]. Due to the limited information available in the TABLE 2
SIX FUNDAMENTAL TYPES OF FAULTS DIAGNOSED USING DGA.
datasets, fault types have been classified into 3 broader
Code Primary Faults
categories: Partial Discharge (PD), Energy Discharges and
Thermal Faults. T1 Thermal fault; T < 300°C
Partial Discharge occurs when a localized area of solid or T2 Thermal fault; 300°C < T < 700°C
fluid insulation material, exposed to high voltage stress, T3 Thermal fault; T > 700°C
undergoes a partial breakdown without fully bridging the gap D1 Low energy discharge
between two conductive materials [3, 57, 58]. In this context, PD D2 High energy discharge
specifically refers to corona-type PD that transpires within gas PD Corona Partial Discharge

VOLUME XX, 2023 4

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

This plasma interacts with the surrounding oil or cellulose, TABLE 3

POSSIBLE CONDITIONS REPRESENTED USING BINARY NUMBERS.
leading to the generation of hydrogen as the primary by-product.
Energy Discharges in oil-immersed power transformers occur
Case Discharge no Fault PD Thermal Diagnosis
when there is an energy discharge that creates a localized
conducting path or short circuit between conductive materials [3, 0 0 0 0 0 No Fault
57, 58]. This leads to sparking around loose connections within 1 0 0 0 1 Thermal Fault
the transformer. When low energy arcs, denoted as D1 in Table 2 0 0 1 0 PD Fault
2, occur in transformer oil, only a thin layer of the oil makes
3 0 0 1 1 PD, Thermal Fault
contact with the path of the arc. The high temperature of the arc,
exceeding 3000°C, causes decomposition of this small oil layer. 4 0 1 0 0 No Fault
The decomposition primarily yields acetylene, with traceable 5 0 1 0 1 Thermal Fault
amount of ethylene being produced. In contrast, high energy 6 0 1 1 0 PD Fault
arcs, referred to as D2 in Table 2, involve a greater current
7 0 1 1 1 PD, Thermal Fault
flowing through the arc path, resulting in a longer duration. The
extended duration allows for a larger volume of oil to be heated 8 1 0 0 0 Discharge Fault
by the arc. The convective flow of cooler surrounding oil 9 1 0 0 1 Discharge,
contributes to this process. Consequently, a significant Thermal Fault
10 1 0 1 0 Discharge, PD
temperature gradient is established in the oil surrounding the arc Fault
path, ranging from around 3000°C to 500°C. Interestingly, 11 1 0 1 1 Discharge, PD
despite the higher energy content of D2 arcs, their average oil &Thermal Fault
12 1 1 0 0 Discharge Fault
temperature is lower compared to D1 arcs. However, D2 arcs
generate a substantial amount of C2H4 in addition to C2H2 due to 13 1 1 0 1 Discharge,
the temperature gradient and longer duration of the arc. Thermal Fault
14 1 1 1 0 Discharge, PD
Thermal Faults in oil-immersed power transformers arise Fault
from the circulation of electric current within the insulating paper 15 1 1 1 1 Discharge, PD
due to excessive dielectric losses [3, 57, 58]. Thermal faults can &Thermal Fault
be classified into three categories: T1, T2 and T3. T1 fault occurs
TABLE 4
when there is an increase in the average winding temperature, FAULT DIAGNOSTIC MODULE OUTPUTS AND CORRESPONDING ASSET
typically caused by increased load or ambient temperature. T2 MANAGEMENT STATEMENT.
fault, on the other hand, involves localized hotspots within the
winding, resulting from electrical contact or excessive current Module Output Asset Management Statement – severity level
density. Localized hotspots generate higher temperatures, which
No Fault “No Fault detected, very low likelihood of failure.”
accelerate the insulation aging and compromise the
transformer’s overall performance. Lastly, T3 fault encompasses PD “Sign of Partial Discharge fault, low risk of failure.”
more severe conditions, such as arcing or short circuits, which
Thermal Fault “Sign of Thermal fault, moderate risk of failure”
cause significant temperature rise and pose an immediate risk to
the transformer’s operation and safety. PD & Thermal Fault “Sign of Thermal fault, moderate risk of failure”
The proposed Fault Diagnostic module shown in Figure 5
Discharge Fault “Sign of Discharge fault, very high risk of failure.”
employs CNN to analyse the measurements of five gases: H2,
CH4, C2H4, C2H6, and C2H2. As introduced in the previous Discharge & Thermal “Sign of Discharge fault, very high risk of failure.”
section, CNN has proven to be highly effective in processing
Discharge & PD “Sign of Discharge fault, very high risk of failure.”
complex data, making it ideal candidate for fault diagnosis in
power transformers. Based on the concentrations of these gases, Discharge, PD &
“Sign of Discharge fault, very high risk of failure.”
the module classifies the transformer’s condition into four Thermal
categories: “Discharge”, “no Fault”, “PD”, and “Thermal”.
Moreover, the module is capable of diagnosing combined fault
conditions, such as “Discharge and Thermal”, adding further B. LIFE MANAGEMENT MODULE
versatility to improve diagnostic accuracy. The life expectancy of power transformers predominantly relies
In the context of enabling the module for multi-label on their paper insulation [5]. The composition of transformer
classification tasks, a binary representation, as shown in Table 3 paper is primarily comprised of 90% cellulose by weight.
has been adopted. The proposed approach utilizes a four-digit Cellulose is an organic compound characterized by lengthy
binary representation based on 2n, where ‘n’ is set to four in this chains of glucose rings, typically ranging between 1000 to 1200
application, accommodating the representation of up to sixteen per chain for new paper [5]. The Degree of Polymerization (DP)
possible cases. refers to the average number of glucose rings present within each
Based on the output of the Fault Diagnostic module, an asset chain. Based on experience, it is commonly considered that
management decision will be provided to users with the severity transformers reach their end-of-life when the DP of paper
level as listed in Table 4. declines to 200, which means the tensile strength decreases to
approximately 40% of its initial value [59, 60]. However,
measuring DP requires a paper sample from the transformer,
which is not feasible for in-service transformers.

VOLUME XX, 2023 6

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

In the present industry practice, the estimation of DP value is adjustable parameters play crucial roles in shaping the model’s
often conducted through the analysis of furan compounds. performance. These parameters include the filter size, which
Furans are generated as by-products during the degradation of determines the width of the convolutional filters employed to
paper insulation impregnated with oil. Among the five furan extract features from the input data. Additionally, the kernel size
compounds, 2-furfural (2-FAL) is commonly utilized as a dictates the scope of each convolutional operation, influencing
predictor of DP due to its higher stability compared to other the receptive field of the network. The choice of padding,
compounds [59]. However, the development of online furan whether ‘valid’ or ‘same’, modifies the dimensions of the output
measuring methods is still an ongoing research area. feature maps. Lastly, the activation function governs the non-
Previous research has revealed a correlation coefficient of 0.87 linearity introduced within the network, contributing to its ability
between 2-FAL and DP [59]. Notably, the ratio of carbon dioxide to capture complex patterns and relationships within the data.
to carbon monoxide (CO2/CO) exhibits the highest correlation The module training process involves an exhaustive
coefficient of 0.97 [59], emphasizing its superior stability as an exploration of various configuration settings. Different options
indicator for assessing the condition of insulation paper. It is were tested at filter sizes: 32, 64, and 128, along with varying
important to highlight that CO2 and CO are also generated numbers of neural network layers. Throughout these
through the oxidation of oil [59]. The utilization of this ratio is experiments, layers have been systematically added to assess
justified by the fact that in situations of high thermal and arcing their impact on the model’s performance. Despite the array of
faults, CO shows a much more rapid increase compared to CO2 layer configurations tested, the final set of layers and parameters
[61]. Conversely, during significant heat generation in normal that yielded optimal results are as shown in Table 6.
operation conditions, CO2 increases at a faster rate than CO. TABLE 6
Unlike conventional machine learning approaches, deep OPTIMAL PARAMETERS OF THE DEVELOPED FAULT DIAGNOSTIC MODULE
learning has the capability to directly incorporate measurements Layer Parameter Setting
of CO2 and CO as input features, thereby eliminating the Filter size 64
requirement for extra features such as CO2/CO ratio. Kernel size 3
Convolutional layer
The proposed Life Management module, also depicted in Padding ‘same’
Activation function ‘ReLU’
Figure 5, utilizes a CNN regression approach, incorporating Number of neurons 16
simply two-gas measurements CO2 and CO to forecast the DP Dense layer
Activation function ‘ReLU’
value. Dense_1 layer Activation function ‘sigmoid’
Based on the DP value, the Life Management module delivers
an evaluation of the paper insulation condition and categorize it The ReLU activation function transforms negative input
into four levels: “Healthy insulation”, “Moderate deterioration”, values to zero while leaving positive values unchanged. The
“Extensive deterioration”, and “End of insulation life”. ReLU function finds extensive application in neural networks
TABLE 5 across diverse domains owing to its efficient computation and
DP VALUES AND THEIR ASSOCIATED SIGNIFICANCE [62] improved gradient propagation, thus facilitating the extraction of
DP Value Significance significant features from the input data.
1200-700 Healthy Insulation The sigmoid activation function possesses the ability to
700-450 Moderate Deterioration condense input values within a range between 0 and 1, as
450-250 Extensive Deterioration
< 250 End of Life
depicted by (4).
1
𝑓𝑓(𝑥𝑥) = −𝑥𝑥 (4)
1+𝑒𝑒
The inclusion of the estimated percentage of remaining life
The sigmoid function exhibits an S-shaped curve is capable of
within the asset management framework does not only enhance
transforming both positive and negative input values into
the precision of the assessments but also provide a dynamic
probabilities. Thus, binary classification will be provided. For
understanding of the insulation condition. An estimation for the
example, if both “Discharge” and “Thermal” faults are present,
percentage of remaining life based on the DP value can be
the predicted probabilities might be presented as [0.8, 0.2, 0.4,
conducted using (3) [63]. Upon review of Table 5, “Healthy
0.7] ([“Discharge”, “no Fault”, “PD”, “Thermal”]); with more
Insulation” category corresponds to DP values ranging from 700
probability assigned to the present faults. It’s noteworthy that
to 1200. In order to align the model with real-world applications
each output probability is determined independently, meaning
more effectively, a DP value of 800 is considered to be
the prediction for one condition does not influence the prediction
corresponding to 100% transformer remaining life as per (3).
of another.
% 𝑜𝑜𝑜𝑜 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 = 166.1 × 𝑙𝑙𝑙𝑙𝑙𝑙10 (𝐷𝐷𝐷𝐷) − 382.2 (3)
During the model compilation phase, the ‘Nadam’ optimizer,
IV.RESULTS AND DISCUSSION which is a combination of the Nesterov Accelerated Gradient
A. FAULT DIAGNOSTIC MODULE (NAG) and Adam optimizers has been applied. The selection of
The database comprises a collection of 1083 DGA samples ‘binary_crossentropy’ as the loss function, as given by (5), is a
incorporated from the 29 literatures listed in Table 1. 65% of common choice for binary classification problems. It measures
the samples were randomly selected for the training phase, with the dissimilarity between predicted probabilities and true labels
an additional 15% allocated for the validation phase while the (0 or 1), thus optimizing the model to achieve accurate binary
remaining 20% were designated for the testing phase. predictions.
The Fault Diagnostic module employs the capabilities of a
one-dimensional CNN, which has an architecture adept at
processing sequential data. Within this framework, several

VOLUME XX, 2023 6

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 − 𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 = −[𝑦𝑦𝑖𝑖 × 𝑙𝑙𝑙𝑙𝑙𝑙(𝑦𝑦�𝑖𝑖 ) + (1 − 𝑦𝑦𝑖𝑖 ) × within the “No Fault” condition as revealed by the confusion
𝑙𝑙𝑙𝑙𝑙𝑙(1 − 𝑦𝑦�𝑖𝑖 )] (5) matrix of the Fault Diagnostic module shown in Figure 7. This
where, 𝑦𝑦𝑖𝑖 is the actual target value (0 or 1) of the i-th data point. may be attributed to the fact that the majority of the collected
𝑦𝑦�𝑖𝑖 is the predicted value of the i-th data point generated by the datasets were predominantly geared towards the identification of
model. fault conditions.
Lastly, ‘accuracy’, a standard evaluation metric, is used for
classification tasks. It calculates the ratio of correctly predicted
instances to the total number of instances to provide insight into
the model’s overall performance.
During the last phase of model training, the model’s weights
are updated based on the provided training data. The training
process involves passing the training data through the network,
computing predictions, comparing them with the actual targets,
and then backpropagating the error to update the model’s
weights. ‘Epochs’ defines the number of times the model will
iterate over the entire training dataset, which is 1000 times in this
case. The batch size determines the number of training examples
the model processes in each update of the gradient. Smaller batch
sizes may lead to more frequent updates, while larger batch sizes
can speed up the training process. A batch size of 16 has been
chosen for the developed model based on running through many
simulations with different batch sizes.
The training process of the model randomly runs due to the
random initial weights, leading to varying results in accuracy and
loss. Following parameters adjustments, the training model has
been executed several times, and the run producing the highest
accuracy and lowest loss is selected. The generated plots shown
in Figure 6 provide valuable insights into the training process and
the performance of the developed CNN model. The alignment or
divergence of the curves reveals the overfitting or underfitting
phenomena and guides potential adjustments in the model
architecture or hyperparameters for optimal performance. In
Figure 6(a), the ‘Training loss’ curve, depicted in yellow, shows
how the model’s loss decreases as it learns to better fit the
training data. The ‘Validation loss’ curve, depicted in red,
demonstrates the model’s performance on unseen validation
data. A decreasing validation loss over epochs indicates
Figure 6. Fault Diagnostic module: (a) Training and validation loss plot. (b)
successful generalization of the model. In some runs, it was Training and validation accuracy plot.
noticed the validation loss started to rise after a certain number
of epochs while the training loss curve was stable. This
overfitting phenomenon indicates that the model has started to
memorize the training data instead of capturing underlying
patterns. In Figure 6(b), the ‘Training acc’ curve in yellow
illustrates the model’s accuracy using the training data, whereas
the ‘Validation acc’ curve in red showcases the model’s
performance using validation data. As epochs progress, observed
increasing validation accuracy that aligns with the training
accuracy reflects the model’s ability to generalize and predict
unseen data accurately.
The satisfied model performance was achieved with an
accuracy of 0.8479 and a corresponding loss of 0.2989 using the
test dataset. This model was subsequently saved for the Fault
Diagnostic module. To test the Fault Diagnostic module, new gas
measurements were fed into the module, enabling the prediction
of outputs through its learned capabilities.
In order to evaluate the performance of the Fault Diagnostic
module, a dataset comprising 151 samples from the IEC TC10
database was used [4]. Subsequently, a thorough assessment of Figure 7. Confusion matrix of the proposed Fault Diagnostic module.
misdiagnoses within each class was carried out. Notably, the
analysis revealed that misdiagnosis predominantly occurred

VOLUME XX, 2023 6

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

The samples presented in Table 7 originate from the IEC TC but also identified a potential discharge fault.
10 database (samples 1-21) and Korea Electric Power
6) In the overall assessment of diagnostic accuracy using IEC
Corporation (KEPCO) historical data [4, 64]. The second last
TC10 data, the developed module consistently demonstrates a
column in the table shows the actual condition of the
better accuracy when compared to traditional methods. It’s
transformers as determined through physical inspection while
worth noting that the Doernenburg ratio method appears to
the last column lists the diagnostic results generated from the
have higher accuracy when applied to KEPCO data. This
proposed Fault Diagnostic module. Certain discrepancies have
discrepancy can be attributed to a specific requirement outlined
been identified in the samples #4, #13, #16, #17, #20, and
in IEEE C57.104 [3], which mandates a minimum
KEPCO’s samples. In Table 8, traditional IEEE and IEC DGA
concentration limit for at least one key gas used in the ratios
interpretation methods are used to analyse all samples and
before the Doernenburg method can be applied. In the KEPCO
compare the results with those obtained from the module. The
dataset, the “no Fault” cases appear to have very low gas
following observations can be drawn out of these comparisons:
centration levels, therefore, identified as “no Fault”. On the
1) In the case of sample #4, the actual condition is described
other hand, in the IEC TC10 data, from samples #17 to #21, the
as “Tracking to the ground in glue of central beam”,
gas measurements exceed the concentration limit, resulting in
categorized as low energy discharge. Traditional methods such
inconclusive outcomes when utilizing the Doernenburg ratio
as Duval Triangle 1, IEEE and IEC ratios can identify this
method. This issue also manifests in the KEPCO data,
discharge fault. However, the developed module did not
producing inconclusive results for 12 or 24 months before the
capture this specific fault.
actual failures may really happen.
2) In the case of sample #17, where the inspection outcome
indicated no fault, the Roger ratio method indicated a discharge From the above analysis, it can be concluded that the
developed CNN-based fault diagnostic module comprises the
fault. However, both the Duval Triangle 1 and the developed
following unique features:
module reached the same conclusion, identifying a thermal
fault. • Ability to identify multi-label classification: Conventional
3) In the case of sample #20, the inspection result denoted no DGA interpretation techniques like Duval Triangle 1, IEEE and
fault. In contrast, the module and the Duval Triangle 1 method IEC ratios methods predominantly pinpointed singular faults.
yielded results indicating combined discharge and thermal Among them, Duval Triangle 1 method can only identify a
faults. The Roger ratio method, on the other hand, identified a combined discharge and thermal fault. In contrast, the CNN
discharge fault. model exhibits the ability to discern distinct features across all
four different conditions.
4) Observations reveal that the module occasionally indicated
the presence of multiple faults, as seen in samples #13 and #16. • Using a single training model: To distinguish three pivotal
In the case of sample #13, the module detected a discharge fault faults; Discharge, Partial Discharge, and Thermal faults, the
in addition to a thermal fault. Notably, if the traditional conventional machine learning methods require three distinct
methods were applied to analyse sample #13, the results would training models to collectively assess the transformer’s overall
align with the inspection result. For sample #16, the module condition. However, CNN method streamlines this process by
detected a partial discharge fault alongside a thermal fault. utilizing a single training model, which minimizes the training
Interestingly, when the IEC ratio and Duval Triangle methods duration substantially. This simplified approach not only
were applied to sample #16, both identified a partial discharge expedites the training phase but also produces a remarkable
fault, in contrast to the inspection result, which revealed a level of accuracy.
thermal fault. • Avoiding data manipulation: The CNN model possesses
5) Another assessment was conducted on four KEPCO inherent feature selection capabilities, which facilitates the use
transformers based on their annual DGA measurements. The of raw data directly and independently. This mechanism
results exhibited the capability of the developed module in circumvents the need for manually engineered features like gas
early fault detection for transformers 1, 2 and 3, which enables ratios or gas percentage. Additionally, this intrinsic capability
proactive maintenance measures to be taken. In the case of contributes to reduce execution time, making it particularly
transformer 4, the module not only detected the thermal fault suitable for real-time condition monitoring.

VOLUME XX, 2023 6

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

TABLE 7
COMPARISON BETWEEN THE DIAGNOSTIC RESULT FROM THE FAULT DIAGNOSTIC MODULE AND ACTUAL CONDITION FROM INSPECTION.
(ALL GASES ARE MEASURED IN PPM)
Actual
condition Module’s
IEC TC10 Samples H2 CH4 C2H4 C2H6 C2H2
from Result
inspection
1 543 120 411 41 1880 Discharge Discharge
2 1230 163 233 27 692 Discharge Discharge
3 645 86 110 13 317 Discharge Discharge
4 60 10 4 4 4 Discharge No Fault
5 95 10 11 0 39 Discharge Discharge
6 32930 2397 0 157 0 PD PD
7 37800 1740 8 249 8 PD PD
8 92600 10200 0 0 0 PD PD
9 8266 1061 0 22 0 PD PD
10 9340 995 6 60 7 PD PD
11 1270 3450 1390 520 8 Thermal Thermal
12 3420 7870 6990 1500 33 Thermal Thermal
Discharge;
13 360 610 260 259 9 Thermal
Thermal
14 0 18900 540 410 330 Thermal Thermal
15 960 4000 1560 1290 6 Thermal Thermal
16 2031 149 3 20 0 Thermal PD; Thermal
17 125 100 150 100 20 No Fault Thermal
18 175 0 375 100 3 No Fault No Fault
19 50 30 0 0 5 No Fault No Fault
Discharge;
20 250 150 250 150 150 No Fault
Thermal
21 134 224 224 550 154 No Fault No Fault
KEPCO Module’s
KEPCO Samples H2 CH4 C2H4 C2H6 C2H2
Health State result
1999 0 6 2 2 0 No Fault No Fault
2000 0 25 13 9 0 No Fault No Fault
Transformer 1 2001 0 35 37 31 0 No Fault Thermal
2002 0 44 28 85 0 No Fault Thermal
2003 251 139 256 123 1064 Electrical Fault Discharge
2011 10 7 2 5 0 No Fault No Fault
2012 13 11 3 26 0 No Fault No Fault
Transformer 2
2013 48 24 12 63 14 No Fault Discharge
2015 335 246 1324 150 1123 Electrical Fault Discharge
2000 0 1 5 0 0 No Fault No Fault
2002 0 7 11 14 0 No Fault No Fault
Transformer 3
2003 0 64 150 99 0 No Fault Thermal
2004 218 744 1743 264 7 Thermal Fault Thermal
2000 5 44 4 9 0 No Fault No Fault
2001 6 42 10 9 0 No Fault No Fault
2002 6 44 12 10 0 No Fault No Fault
Transformer 4
2003 7 56 12 10 0 No Fault No Fault
Discharge;
2004 628 1381 1873 351 2.8 Thermal Fault
Thermal

VOLUME XX, 2023 2

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

TABLE 8
COMPARISON OF DIAGNOSTIC RESULTS: PROPOSED MODULE VS. THE TRADITIONAL METHODS.
(ALL GASES ARE MEASURED IN PPM)
Actual
Roger
condition Doernenburg IEC Ratio Duval Module’s
IEC TC10 Samples Ratio
from Ratio Method Method Triangle result
Method
inspection
1 Discharge N/A Discharge Discharge Discharge Discharge
2 Discharge Discharge Discharge Discharge Discharge Discharge
3 Discharge Discharge Discharge Discharge Discharge Discharge
N/A No Fault
4 Discharge Discharge Discharge Discharge
5 Discharge N/A Discharge N/A Discharge Discharge
6 PD N/A N/A N/A PD PD
7 PD N/A PD PD PD PD
8 PD N/A N/A N/A PD PD
9 PD N/A N/A N/A PD PD
10 PD N/A N/A N/A PD PD
Thermal
11 Thermal Thermal Thermal Thermal Thermal
Thermal
12 Thermal Thermal Thermal Thermal Thermal
Discharge;
13 Thermal Thermal Thermal Thermal Thermal
Thermal
14 Thermal N/A N/A N/A Thermal Thermal
15 Thermal Thermal Thermal Thermal Thermal Thermal
16 Thermal N/A N/A PD PD PD; Thermal
17 No Fault Discharge N/A N/A Thermal Thermal
18 No Fault N/A N/A N/A Thermal No Fault
19 No Fault N/A N/A N/A Discharge No Fault
Discharge; Discharge;
20 No Fault Discharge N/A N/A
Thermal Thermal
21 No Fault N/A N/A Thermal Discharge No Fault
Roger
KEPCO Doernenburg IEC Ratio Duval Module’s
KEPCO Samples Ratio
Health State Ratio Method Method Triangle result
Method
1999 No Fault N/A No Fault N/A Thermal No Fault
2000 No Fault N/A No Fault N/A Thermal No Fault
Transformer 1 2001 No Fault N/A No Fault N/A Thermal Thermal
2002 No Fault N/A N/A N/A Thermal Thermal
Electrical
2003 N/A Discharge N/A Discharge Discharge
Fault
2011 No Fault Thermal No Fault N/A Thermal No Fault
2012 No Fault N/A No Fault N/A Thermal No Fault
Transformer 2
2013 No Fault N/A N/A N/A Discharge Discharge
2015 Electrical
Discharge Discharge Discharge Discharge Discharge
Fault
2000 No Fault N/A No Fault N/A Thermal No Fault
2002 No Fault N/A No Fault N/A Thermal No Fault
Transformer 3
2003 No Fault N/A N/A N/A Thermal Thermal
Thermal
2004 Thermal Thermal Thermal Thermal Thermal
fault
2000 No Fault Thermal No Fault N/A Thermal No Fault
2001 No Fault Thermal No Fault Thermal Thermal No Fault
Transformer 4 2002 No Fault Thermal No Fault Thermal Thermal No Fault
2003 No Fault Thermal No Fault Thermal Thermal No Fault
Thermal Discharge;
2004 Thermal Thermal Thermal Thermal
Fault Thermal

VOLUME XX, 2023 2

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

B. LIFE MANAGEMENT MODULE

The training of the Life Management module employed
47 datasets sourced from three distinct literatures. These
datasets were divided into a training set comprising 80%
of the data and a testing set constituting the remaining
20%.
The training model in this context also incorporated a
one-dimensional CNN architecture. The specific parameters
utilized in configuring the CNN architecture are detailed in
Table 9.
TABLE 9
OPTIMAL PARAMETERS OF THE DEVELOPED LIFE MANAGEMENT
MODULE.

Layer Parameter Setting

Filter size 128 Figure 8. The training and validation loss plot for the Life management
Convolutional layer Kernel size 1 module.

Activation function ‘ReLU’

In the past, there was lack of records connecting CO2 and
Number of neurons 64 CO measurements with the degree of deterioration of
Dense layer
Activation function ‘ReLU’ insulating paper. To pursue further module testing, an
Dense_1 layer Number of neurons 1 additional 131 CO2 and CO measurements have been
collected from diverse sources of the literature. The CO2 and
In the phase of compiling the model, the Adaptive CO measurements are fed into the life management module
Moment Estimation ‘adam’ optimizer and ‘mse’ loss for processing based on the ranges and corresponding
function are adopted. As stated above, ‘adam’ enhances conditions given in Table 10. Upon comparing the module’s
optimization by independently adjusting learning rates for outcomes with the estimated results, a variance was
each parameter based on the historical gradients. The Mean observed in 20 out of 131 cases, demonstrating an
Squared Error ‘mse’ loss function as given by (6), quantifies approximate 85% concurrence rate.
the average squared difference between the predicted values
generated by a model and the target values in the dataset. It C. CONTRIBUTION AND SIGNIFICANCE
is often used as a loss function for regression models to Results show that the developed asset management
guide the optimization process. module offers a generalized approach to evaluate power
1
transformer condition. Constructed using diverse datasets
𝑀𝑀𝑀𝑀𝑀𝑀 = × ∑(𝑦𝑦𝑖𝑖 − 𝑦𝑦�𝑖𝑖 )2 (6) from reputable literature sources, this model emerges as a
𝑛𝑛
where, n is the total number of samples in the dataset used comprehensive tool for users at all stages. The model
to evaluate the model performance. 𝑦𝑦𝑖𝑖 is the actual observed features a continuous learning capacity, progressively
value while 𝑦𝑦�𝑖𝑖 is the corresponding predicted value by the enhancing its performance as it encounters new data, thus
model of the i-th data point. ensuring its adaptability to specific requirements, including
power transformers operating within distinct conditions and
By setting the number of epochs to 5000 and the batch environments.
size to 16, an optimal model performance has been The model’s foundation lies in the employment of CNN
observed. The progression of the training process is visually deep machine learning algorithms, empowering it to
depicted in Figure 8. A substantial reduction in both the assimilate fresh information and fine-tune its predictions
training and validation losses over the course of training can accordingly. By harnessing this innovative model, users can
be observed from the figure. access invaluable insights into the well-being of their power
To assess the accuracy of the developed module, a transformers, facilitating well-informed decisions regarding
database that was not included in the training process was maintenance and potential replacements.
analysed as outlined in Table 11 [65]. As can be seen in the While the model was developed mainly for mineral oil
Table, only two samples (#5 and #7) out of the total 15 immersed-power transformers due to the availability of
results deviate from the expected target values. This reflects required data, same concepts can be used to modify the
an accuracy level of 86.7% for the developed model. model to other transformer types once sufficient data are
available to train the model.

VOLUME XX, 2023 15

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

TABLE 10
CO2 AND CO CONCENTRATION RANGES AND THE CORRESPONDING V.CONCLUSION
DETERIORATION LEVEL [66].
This paper presents a comprehensive approach for
CO2 CO Significance transformer asset management through the integration of
0 - 2500 0 - 350 Health Insulation two modules: Fault Diagnostics and Life Management.
Both modules have undergone training utilizing deep CNN
2500 - 4000 350 - 570 Moderate machine learning technique. This technique empowers the
Deterioration model to harness the potential of online DGA
Extensive
measurements, providing asset managers with a streamlined
4000 - 10000 570 - 1400
Deterioration means to obtain highly accurate insights into the health
condition of power transformers. Furthermore, the model
≥ 10000 ≥ 1400 End of Life offers indication of paper insulation deterioration in real
time using the measurements of CO and CO2 that can be
obtained using online DGA sensors. This feature is crucial
Overall, key advancements and contributions highlighted information for effective real time asset management
in this study include: schemes. The adoption of CNN not only simplifies the
• Developed transformer asset management solely relying process, but it utilizes a single training model and requires
on measurements obtained from online DGA sensors. minimal input features. The proposed approach enhances
• Empowerment of the model to interpret online DGA the precision of predictions, hence facilitating informed
measurements accurately and diagnose multiple faults to decision-making for asset managers. This comprehensive
provide more insights into transformer health condition. strategy, encompassing fault diagnostics and life
• Estimating the DP value based on CO and CO2 assessment, demonstrates the integration of cutting-edge
measurements and hence eliminating the need to measure technology into asset management practices, and
furan compounds offline. This feature facilitates the contributes to the enhanced reliability and longevity of
online implementation of the developed asset power transformers.
management model.
• Employing CNN, which utilizes a single training model
and requires minimal input features.

TABLE 11
COMPARISON BETWEEN THE PREDICTED RESULT FROM THE LIFE MANAGEMENT MODULE AND ACTUAL CONDITION FROM DATASET.
Condition
Condition based on
Targeted Predicted
Sample # CO2 CO based on predicted DP
DP DP
targeted DP (using CO,
CO2)
1 812 62 1304 Healthy 1020 Healthy
2 2628 370 490 Moderate 540 Moderate
3 1498 132 963 Healthy 984 Healthy
4 1878 164 1189 Healthy 727 Healthy
5 2298 214 1149 Healthy 596 Moderate
6 4400 594 266 Extensive 338 Extensive
7 2562 146 846 Healthy 565 Moderate
8 2587 387 500 Moderate 544 Moderate
9 2502 353 654 Moderate 556 Moderate
10 4348 576 362 Extensive 339 Extensive
11 4206 586 268 Extensive 341 Extensive
12 2984 503 600 Moderate 489 Moderate
13 1815 211 785 Healthy 748 Healthy
14 4217 566 276 Extensive 342 Extensive
15 2421 372 652 Moderate 564 Moderate

VOLUME XX, 2023 15

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

power transformers," Electronics (Basel), vol. 10, no. 10, p.

1206, 2021, doi: 10.3390/electronics10101206.
ACKNOWLEDGMENT [13] M. Božić, M. Stojanović, Z. Stajić, and Đ. Vukić, "Power
The authors sincerely acknowledge the support from BHP transformer fault diagnosis based on dissolved gas analysis
for providing a scholarship to the first author to conduct her with logistic regression," Przegląd Elektrotechniczny, vol. 6,
pp. 83-87, 2013.
research on high voltage assets condition monitoring [14] H. A. Illias and W. Z. Liang, "Identification of transformer
methodologies. The authors also acknowledge the support fault based on dissolved gas analysis using hybrid support
from Curtin University, Australia, for facilitating research vector machine-modified evolutionary particle swarm
optimisation," PLoS One, vol. 13, no. 1, pp. e0191366-
resources. e0191366, 2018, doi: 10.1371/journal.pone.0191366.
[15] K. Bacha, S. Souahlia, and M. Gossa, "Power transformer
REFERENCES fault diagnosis based on dissolved gas analysis by support
[1] L. Jin, D. Kim, A. Abu-Siada, and S. Kumar, "Oil-Immersed vector machine," Electric power systems research, vol. 83,
Power Transformer Condition Monitoring Methodologies: A no. 1, pp. 73-79, 2012, doi: 10.1016/j.epsr.2011.09.012.
Review," Energies (Basel), vol. 15, no. 9, p. 3379, 2022, doi: [16] J. I. Aizpurua et al., "Power transformer dissolved gas
10.3390/en15093379. analysis through Bayesian networks and hypothesis testing,"
[2] G. Odongo, R. Musabe, and D. Hanyurwimfura, "A IEEE Transactions on Dielectrics and Electrical Insulation,
Multinomial DGA Classifier for Incipient Fault Detection in vol. 25, no. 2, pp. 494-506, 2018, doi:
Oil-Impregnated Power Transformers," Algorithms, vol. 14, 10.1109/TDEI.2018.006766.
no. 4, doi: 10.3390/a14040128. [17] C. Lefeng and Y. Tao, "Dissolved Gas Analysis Principle-
[3] "IEEE Guide for the Interpretation of Gases Generated in Based Intelligent Approaches to Fault Diagnosis and
Mineral Oil-Immersed Transformers," IEEE Std C57.104- Decision Making for Large Oil-Immersed Power
2019 (Revision of IEEE Std C57.104-2008), pp. 1-98, 2019, Transformers: A Survey," Energies (Basel), vol. 11, no. 4, p.
doi: 10.1109/IEEESTD.2019.8890040. 913, 2018.
[4] M. Duval and A. dePabla, "Interpretation of gas-in-oil [18] J. Faiz and M. Soleimani, "Assessment of computational
analysis using new IEC publication 60599 and IEC TC 10 intelligence and conventional dissolved gas analysis methods
databases," IEEE Electrical Insulation Magazine, vol. 17, for transformer fault diagnosis," IEEE Transactions on
no. 2, pp. 31-41, 2001, doi: 10.1109/57.917529. Dielectrics and Electrical Insulation, vol. 25, no. 5, pp.
[5] L. Jin, D. Kim, and A. Abu-Siada, "State-of-the-art review 1798-1806, 2018, doi: 10.1109/TDEI.2018.007191.
on asset management methodologies for oil-immersed power [19] H. A. Illias, X. R. Chai, and A. H. Abu Bakar, "Hybrid
transformers," Electric Power Systems Research, vol. 218, p. modified evolutionary particle swarm optimisation-time
109194, 2023/05/01/ 2023, doi: varying acceleration coefficient-artificial neural network for
https://doi.org/10.1016/j.epsr.2023.109194. power transformer fault diagnosis," Measurement, vol. 90,
[6] A. Abu-Siada, S. Hmood, and S. Islam, "A new fuzzy logic pp. 94-102, 2016/08/01/ 2016, doi:
approach for consistent interpretation of dissolved gas-in-oil https://doi.org/10.1016/j.measurement.2016.04.052.
analysis," IEEE Transactions on Dielectrics and Electrical [20] T. Kari et al., "An integrated method of ANFIS and
Insulation, vol. 20, no. 6, pp. 2343-2349, 2013, doi: Dempster-Shafer theory for fault diagnosis of power
10.1109/TDEI.2013.6678888. transformer," IEEE Transactions on Dielectrics and
[7] M. Žarković and Z. Stojković, "Analysis of artificial Electrical Insulation, vol. 25, no. 1, pp. 360-371, 2018, doi:
intelligence expert systems for power transformer condition 10.1109/TDEI.2018.006746.
monitoring and diagnostics," Electric Power Systems [21] R. A. Hooshmand, M. Parastegari, and Z. Forghani,
Research, vol. 149, pp. 125-136, 2017/08/01/ 2017, doi: "Adaptive neuro-fuzzy inference system approach for
https://doi.org/10.1016/j.epsr.2017.04.025. simultaneous diagnosis of the type and location of faults in
[8] A. G. C. Menezes, M. M. Araujo, O. M. Almeida, F. R. power transformers," IEEE Electrical Insulation Magazine,
Barbosa, and A. P. S. Braga, "Induction of Decision Trees to vol. 28, no. 5, pp. 32-42, 2012, doi:
Diagnose Incipient Faults in Power Transformers," IEEE 10.1109/MEI.2012.6268440.
Transactions on Dielectrics and Electrical Insulation, vol. [22] Y. Wang and L. Zhang, "A Combined Fault Diagnosis
29, no. 1, pp. 279-286, 2022, doi: Method for Power Transformer in Big Data Environment,"
10.1109/TDEI.2022.3148453. Mathematical Problems in Engineering, vol. 2017, p.
[9] N. Haque, A. Jamshed, K. Chatterjee, and S. Chatterjee, 9670290, 2017/05/18 2017, doi: 10.1155/2017/9670290.
"Accurate Sensing of Power Transformer Faults From [23] T.-D. Do, V.-N. Tuyet-Doan, Y.-S. Cho, J.-H. Sun, and Y.-
Dissolved Gas Data Using Random Forest Classifier Aided H. Kim, "Convolutional-Neural-Network-Based Partial
by Data Clustering Method," IEEE Sensors Journal, vol. 22, Discharge Diagnosis for Power Transformer Using UHF
no. 6, pp. 5902-5910, 2022, doi: Sensor," IEEE access, vol. 8, pp. 207377-207388, 2020, doi:
10.1109/JSEN.2022.3149409. 10.1109/ACCESS.2020.3038386.
[10] Y. Kim, T. Park, S. Kim, N. Kwak, and D. Kweon, [24] K. Hong, M. Jin, and H. Huang, "Transformer Winding Fault
"Artificial Intelligent Fault Diagnostic Method for Power Diagnosis Using Vibration Image and Deep Learning," IEEE
Transformers using a New Classification System of Faults," transactions on power delivery, vol. 36, no. 2, pp. 676-685,
Journal of electrical engineering & technology, vol. 14, no. 2021, doi: 10.1109/TPWRD.2020.2988820.
2, pp. 825-831, 2019, doi: 10.1007/s42835-019-00105-0. [25] L. Tao, X. Yang, Y. Zhou, and L. Yang, "A Novel
[11] O. Kherif, Y. Benmahamed, M. Teguar, A. Boubakeur, and Transformers Fault Diagnosis Method Based on
S. S. M. Ghoneim, "Accuracy Improvement of Power Probabilistic Neural Network and Bio-Inspired Optimizer,"
Transformer Faults Diagnostic Using KNN Classifier With Sensors, vol. 21, no. 11, p. 3623, 2021. [Online]. Available:
Decision Tree Principle," IEEE Access, vol. 9, pp. 81693- https://www.mdpi.com/1424-8220/21/11/3623.
81701, 2021, doi: 10.1109/ACCESS.2021.3086135. [26] Y. Zhou, L. Tao, X. Yang, and L. Yang, "Novel Probabilistic
[12] Y. D. Almoallem, I. B. M. Taha, M. I. Mosaad, L. Nahma, Neural Network Models Combined with Dissolved Gas
and A. Abu-Siada, "Application of logistic regression Analysis for Fault Diagnosis of Oil-Immersed Power
algorithm in the interpretation of dissolved gas analysis for Transformers," ACS Omega, vol. 6, no. 28, pp. 18084-
18098, 2021/07/20 2021, doi: 10.1021/acsomega.1c01878.

VOLUME XX, 2023 15

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

[27] L. Zhang, G. Sheng, H. Hou, and X. Jiang, "A Fault [40] L. V. Ganyun, C. Haozhong, Z. Haibao, and D. Lixin, "Fault
Diagnosis Method of Power Transformer Based on Cost diagnosis of power transformer based on multi-layer SVM
Sensitive One-Dimensional Convolution Neural Network," classifier," Electric Power Systems Research, vol. 74, no. 1,
in 2020 5th Asia Conference on Power and Electrical pp. 1-7, 2005/04/01/ 2005, doi:
Engineering (ACPEE), 4-7 June 2020 2020, pp. 1824-1828, https://doi.org/10.1016/j.epsr.2004.07.008.
doi: 10.1109/ACPEE48638.2020.9136223. [41] W. Mang-Hui, "A novel extension method for transformer
[28] C. Sun, Y. Chen, and N. Tang, "Fault Diagnosis of Power fault diagnosis," IEEE Transactions on Power Delivery, vol.
Transformer Based on DGA and Information Fusion," in 18, no. 1, pp. 164-169, 2003, doi:
2022 IEEE/IAS Industrial and Commercial Power System 10.1109/TPWRD.2002.803838.
Asia (I&CPS Asia), 8-11 July 2022 2022, pp. 247-251, doi: [42] J. Fan, F. Wang, Q. Sun, F. Bin, F. Liang, and X. Xiao,
10.1109/ICPSAsia55496.2022.9949927. "Hybrid RVM–ANFIS algorithm for transformer fault
[29] J. Li, C. Hai, Z. Feng, and G. Li, "A Transformer Fault diagnosis," IET generation, transmission & distribution, vol.
Diagnosis Method Based on Parameters Optimization of 11, no. 14, pp. 3637-3643, 2017, doi: 10.1049/iet-
Hybrid Kernel Extreme Learning Machine," IEEE Access, gtd.2017.0547.
vol. 9, pp. 126891-126902, 2021, doi: [43] H. Hongsheng, Q. Suxiang, and C. Jian, "Monitoring and
10.1109/ACCESS.2021.3112478. fault diagnosing system design for power transformer based
[30] Y. Cui, H. Ma, and T. Saha, "Improvement of power on temperature field model and DGA feature extraction," in
transformer insulation diagnosis using oil characteristics data 2008 7th World Congress on Intelligent Control and
preprocessed by SMOTEBoost technique," IEEE Automation, 25-27 June 2008 2008, pp. 1800-1805, doi:
Transactions on Dielectrics and Electrical Insulation, vol. 10.1109/WCICA.2008.4593195.
21, no. 5, pp. 2363-2373, 2014, doi: [44] M. Badawi et al., "Reliable Estimation for Health Index of
10.1109/TDEI.2014.004547. Transformer Oil Based on Novel Combined Predictive
[31] S. Kim et al., "A Semi-Supervised Autoencoder With an Maintenance Techniques," IEEE Access, vol. 10, pp. 25954-
Auxiliary Task (SAAT) for Power Transformer Fault 25972, 2022, doi: 10.1109/ACCESS.2022.3156102.
Diagnosis Using Dissolved Gas Analysis," IEEE Access, vol. [45] B. Zeng, J. Guo, W. Zhu, Z. Xiao, F. Yuan, and S. Huang,
8, pp. 178295-178310, 2020, doi: "A transformer fault diagnosis model based on hybrid grey
10.1109/ACCESS.2020.3027830. wolf optimizer and LS-SVM," Energies (Basel), vol. 12, no.
[32] S. A. Khan, M. D. Equbal, and T. Islam, "ANFIS based 21, p. 4170, 2019, doi: 10.3390/en12214170.
identification and location of paper insulation faults of an oil [46] Y. Zhang, X. Ding, Y. Liu, and P. J. Griffin, "An artificial
immersed transformer," in 2014 6th IEEE Power India neural network approach to transformer fault diagnosis,"
International Conference (PIICON), 5-7 Dec. 2014 2014, IEEE Transactions on Power Delivery, vol. 11, no. 4, pp.
pp. 1-6, doi: 10.1109/POWERI.2014.7117715. 1836-1841, 1996, doi: 10.1109/61.544265.
[33] Y. Benmahamed, M. Teguar, and A. Boubakeur, [47] X. Hao and S. Cai-xin, "Artificial Immune Network
"Application of SVM and KNN to Duval Pentagon 1 for Classification Algorithm for Fault Diagnosis of Power
transformer oil diagnosis," IEEE transactions on dielectrics Transformer," IEEE Transactions on Power Delivery, vol.
and electrical insulation, vol. 24, no. 6, pp. 3443-3451, 22, no. 2, pp. 930-935, 2007, doi:
2017, doi: 10.1109/TDEI.2017.006841. 10.1109/TPWRD.2007.893182.
[34] M. S. Katooli and A. Koochaki, "Detection and [48] V. Duraisamy, N. Devarajan, D. Somasundareswari, A. A.
Classification of Incipient Faults in Three-Phase Power M. Vasanth, and S. N. Sivanandam, "Neuro fuzzy schemes
Transformer Using DGA Information and Rule-based for fault detection in power transformer," Applied Soft
Machine Learning Method," Journal of control, automation Computing, vol. 7, no. 2, pp. 534-539, 2007/03/01/ 2007,
& electrical systems, vol. 31, no. 5, pp. 1251-1266, 2020, doi: https://doi.org/10.1016/j.asoc.2006.10.001.
doi: 10.1007/s40313-020-00625-5. [49] N. Yadaiah and N. Ravi, "Internal fault detection techniques
[35] J. Faiz and M. Soleimani, "Dissolved gas analysis evaluation for power transformers," Applied Soft Computing, vol. 11,
in electric power transformers using conventional methods a no. 8, pp. 5259-5269, 2011/12/01/ 2011, doi:
review," IEEE Transactions on Dielectrics and Electrical https://doi.org/10.1016/j.asoc.2011.05.034.
Insulation, vol. 24, no. 2, pp. 1239-1248, 2017, doi: [50] A. R. G. Castro and V. Miranda, "Knowledge discovery in
10.1109/TDEI.2017.005959. neural networks with application to transformer failure
[36] M. Elsisi, M. Q. Tran, K. Mahmoud, D.-E. A. Mansour, M. diagnosis," IEEE Transactions on Power Systems, vol. 20,
Lehtonen, and M. M. F. Darwish, "Effective IoT-based deep no. 2, pp. 717-724, 2005, doi:
learning platform for online fault diagnosis of power 10.1109/TPWRS.2005.846074.
transformers against cyberattacks and data uncertainties," [51] M. M. Ibrahim, M. M. Sayed, and E. E. A. El-Zahab,
Measurement : journal of the International Measurement "Diagnosis of power transformer incipient faults using Fuzzy
Confederation, vol. 190, p. 1, 2022, doi: Logic-IEC Based Approach," in 2014 IEEE International
10.1016/j.measurement.2021.110686. Energy Conference (ENERGYCON), 13-16 May 2014 2014,
[37] I. B. M. Taha, S. S. M. Ghoneim, and H. G. Zaini, pp. 242-245, doi: 10.1109/ENERGYCON.2014.6850435.
"Improvement of Rogers four ratios and IEC Code methods [52] H. Malik and S. Mishra, "Selection of Most Relevant Input
for transformer fault diagnosis based on Dissolved Gas Parameters Using Principle Component Analysis for
Analysis," in 2015 North American Power Symposium Extreme Learning Machine Based Power Transformer Fault
(NAPS), 4-6 Oct. 2015 2015, pp. 1-5, doi: Diagnosis Model," Electric power components and systems,
10.1109/NAPS.2015.7335098. vol. 45, no. 12, pp. 1339-1352, 2017, doi:
[38] A. Peimankar, S. J. Weddell, T. Jalal, and A. C. Lapthorn, 10.1080/15325008.2017.1338794.
"Evolutionary multi-objective fault diagnosis of power [53] H. He and E. A. Garcia, "Learning from Imbalanced Data,"
transformers," Swarm and Evolutionary Computation, vol. IEEE Transactions on Knowledge and Data Engineering,
36, pp. 62-75, 2017/10/01/ 2017, doi: vol. 21, no. 9, pp. 1263-1284, 2009, doi:
https://doi.org/10.1016/j.swevo.2017.03.005. 10.1109/TKDE.2008.239.
[39] A. Kirkbas, A. Demircali, S. Koroglu, and A. Kizilkaya, [54] K. N. V. P. S. Rajesh, U. M. Rao, I. Fofana, P. Rozga, and
"Fault diagnosis of oil-immersed power transformers using A. Paramane, "Influence of Data Balancing on Transformer
common vector approach," Electric Power Systems DGA Fault Classification With Machine Learning
Research, vol. 184, p. 106346, 2020/07/01/ 2020, doi: Algorithms," IEEE Transactions on Dielectrics and
https://doi.org/10.1016/j.epsr.2020.106346. Electrical Insulation, vol. 30, no. 1, pp. 385-392, 2023, doi:
10.1109/TDEI.2022.3230377.

VOLUME XX, 2023 15

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
 This article has been accepted for publication in IEEE Access. 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/ACCESS.2024.3366905

[55] M. M. Alam et al., "Assessing Transformer Oil Quality

using Deep Convolutional Networks," in 2019 29th Dowon Kim worked as a transmission and
Australasian Universities Power Engineering Conference substation engineer at Korea Electric Power
(AUPEC), 26-29 Nov. 2019 2019, pp. 1-6, doi: Corporation from 1998 to 2011, and worked in the
10.1109/AUPEC48547.2019.211896.
fields of power system study, condition
[56] J. Lin, J. Ma, J. G. Zhu, and Y. Cui, "A Transfer Ensemble
Learning Method for Evaluating Power Transformer Health monitoring and protection in Australia since
Conditions With Limited Measurement Data," IEEE 2012. He completed his Ph.D. in Electrical
Transactions on Instrumentation and Measurement, vol. 71, Engineering from Curtin University, Perth
pp. 1-10, 2022, doi: 10.1109/TIM.2022.3175268. Australia, in 2021 and he has been working as a Lecturer with the School
[57] G. Odongo, R. Musabe, and D. Hanyurwimfura, "A of Electrical Engineering, Computing, and Mathematical Sciences,
multinomial dga classifier for incipient fault detection in oil- Curtin University, Perth, WA, Australia since August 2020. His research
impregnated power transformers," Algorithms, vol. 14, no. 4,
interests include power system protection, high voltage condition
p. 128, 2021, doi: 10.3390/a14040128.
[58] "D1/A2 Technical Brochure - Advances in DGA monitoring and diagnosis, and wireless power transfer systems.
interpretation," vol. 771, July 2019. CIGRE.
[59] A. Teymouri and B. Vahidi, "CO2/CO concentration ratio: A Kit Yan Chan received the Ph.D. degree in
complementary method for determining the degree of computing from London South Bank University,
polymerization of power transformer paper insulation," IEEE London, U.K., in 2006. He is a Senior Lecturer in
Electrical Insulation Magazine, vol. 33, no. 1, pp. 24-30, the School of Electrical Engineering, Computing
2017, doi: 10.1109/MEI.2017.7804313.
and Mathematical Science, Curtin University,
[60] M. Duval, A. D. Pablo, I. Atanasova-Hoehlein, and M.
Grisaru, "Significance and detection of very low degree of Perth, WA, Australia. He was a Full Time
polymerization of paper in transformers," IEEE Electrical Researcher in Hong Kong Polytechnic University
Insulation Magazine, vol. 33, no. 1, pp. 31-38, 2017, doi: (2004–2009) and Curtin University (2009–2013).
10.1109/MEI.2017.7804314. His research interests include artificial intelligence, machine learning and
[61] F. R. Barbosa, O. M. Almeida, A. P. S. Braga, M. A. B. their applications to power flow optimization, underwater acoustic
Amora, and S. J. M. Cartaxo, "Application of an artificial
communications, load consumption forecasting etc. He has published
neural network in the use of physicochemical properties as a
low cost proxy of power transformers DGA data," IEEE more than 100 journal papers and several books and serves as an associate
Transactions on Dielectrics and Electrical Insulation, vol. editor for several reputed journals.
19, no. 1, pp. 239-246, 2012, doi:
10.1109/TDEI.2012.6148524.
[62] S. Forouhari and A. Abu-Siada, "Application of adaptive
neuro fuzzy inference system to support power transformer Ahmed Abu-Siada, received his B.Sc. and M.Sc.
life estimation and asset management decision," IEEE degrees from Ain Shams University, Egypt and PhD
Transactions on Dielectrics and Electrical Insulation, vol. degree from Curtin University, Australia, all in Electrical
25, no. 3, pp. 845-852, 2018, doi: Engineering. He is a full professor and the head of the
10.1109/TDEI.2018.006392. Electrical and Computer Engineering Discipline, Curtin
[63] S. S. M. Ghoneim, "The Degree of Polymerization in a University. He has published more than 300 Journal and
Prediction Model of Insulating Paper and the Remaining conference papers and holds several patents on condition
Life of Power Transformers," Energies, vol. 14, no. 3, p. monitoring, fault diagnosis and remnant life estimation of high voltage
670, 2021. [Online]. Available: assets. He has delivered several keynote speeches in various IEEE and
https://www.mdpi.com/1996-1073/14/3/670. international conferences and has been a general chair for the IEEE
[64] S. Kim et al., "A Semi-Supervised Autoencoder With an iSPEC 2022, CMD 2018 conferences. Prof Abu-Siada is the founding
Auxiliary Task (SAAT) for Power Transformer Fault editor-in-chief of 2 international journals and an editor for several
Diagnosis Using Dissolved Gas Analysis," IEEE access, vol. journals such as IET GTD.
8, pp. 178295-178310, 2020, doi:
10.1109/ACCESS.2020.3027830.
[65] M. M. Nezami, M. D. Equbal, S. A. Khan, S. Sohail, and S.
S. M. Ghoneim, "Classification of Cellulosic Insulation State
Based on Smart Life Prediction Approach (SLPA),"
Processes, vol. 9, no. 6, p. 981, 2021. [Online]. Available:
https://www.mdpi.com/2227-9717/9/6/981.
[66] M. Nezami, M. Equbal, S. A. Khan, and S. Sohail, "An
ANFIS based comprehensive correlation between diagnostic
and destructive parameters of transformer’s paper
insulation," Arabian Journal for Science and Engineering,
vol. 46, no. 2, pp. 1541-1547, 2021.

Lan Jin received her B.Eng degree in Electrical and

Electronic Engineering from the University of
Melbourne, Australia, in 2007. Currently, she is
working toward a master’s degree at the School of
Electrical and Computer Engineering, Computing
and Mathematical Science at Curtin University,
Perth, Australia. Lan Jin’s primary areas of research
focus on the application of artificial intelligence
algorithms in power transformer condition
monitoring and fault diagnosis.

VOLUME XX, 2023 15

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4