DOI: 10.

2478/sbeef-2023-0014

MACHINE LEARNING ALGORITHMS FUSION BASED ON DGA DATA FOR IMPROVING

FAULT DIAGNOSIS OF ELECTRICAL POWER TRANSFORMER

Abdelmoumene Hechifa1, Abdelaziz Lakehal2,*, Arnaud Nanfak3, Lotfi Saidi4, Chouaib Labiod5
1
LGMM Laboratory, Faculty of Technology, University of 20 August 1955-Skikda, Skikda, Algeria.
2
Laboratory of Research on Electromechanical and Dependability, University of Souk Ahras, Souk-Ahras, Algeria
3
Laboratory of Energy, Materials, Modelling and Methods, National Higher Polytechnic School of Douala, University of
Douala ,Douala , Cameroon.
4
University of Tunis, ENSIT – Laboratory of Signal Image and Energy Mastery, Tunis, Tunisia
5
Electrical Engineering Department, Faculty of Technology, University of El Oued, El Oued, Algeria.
* Corresponding author: [email protected]

Abstract: Dissolved Gas Analysis (DGA) continues to be systems. Regular monitoring is necessary to maintain
widely recognized as a valuable method in recent times for the their availability and prevent any potential faults. In the
early identification of issues in oil-filled power transformers. event of a transformer fault, the entire power network can
It has gained extensive adoption as a primary approach for be affected, leading to catastrophic consequences in the
transmission of electricity [1]. Power transformers are
the early discovery of these issues, relying on the analysis of
susceptible to both thermal and electrical stresses, which
dissolved gases. This contributes to enhancing the
have the potential to induce decomposition of the
dependability of electrical systems. This paper proposes an
insulating oil, subsequently resulting in oxidation. This
efficient fusion method based on DGA data using the two best
process leads to the production of various gases,
Machine Learning algorithms , the neural network (MLP),
including Methane (CH4), carbon monoxide (CO),
the naïve Bayes (NB) throughdata input vector ppm, a Hydrogen (H2), Acetylene (C2H2), Ethylene (C2H4),
percentage input vector, and an Logarithmic input vector. The Ethane (C2H6), carbon dioxide (CO2) [2]. To analyze the
fusion method predictively combined the two classifiers and gases produced by transformer faults, a commonly used
obtained a statistical evaluation: accuracy, recall, precision, method is Dissolved Gas Analysis (DGA) [3].
and F-measure higher than both classifiers separately. The
DGA is a powerful diagnostic technique and widely used
proposed fusion method was evaluated for performance using
by the majority of power company for detecting thermal
a test database and compared with conventional and smart
and electrical faults in transformers. International
methods. Results showed that the proposed model
committees acknowledge its efficiency in diagnosing
outperformed both traditional and intelligent methods in issues resulting from oil or paper, including
terms of diagnostic accuracy when using percentage and Low/Medium/High thermal faults or electrical faults,
logarithmic input vectors. The Prediction Based Fusion including Partial discharge and Low/High energy
(PBF) vector Percentages achieved an accuracy rate of discharge [4]. To simplify the process of identifying
97.22%, while PBF vector Logarithmic achieved an accuracy potential faults in transformers, established rules and
rate of 95.83%. These rates were higher than those achieved methods rely on the analysis of dissolved gas
by traditional methods, such as the Modified RRM/CEGB concentrations [5].
method 91.67% and Modified RRM/IEC method 90.28%.
Accurate methods for diagnosing faults in power
Additionally, the proposed model surpassed the accuracy rates transformers are essential for conducting a thorough
of intelligent methods, such as CSUS ANN 88.89% and DGA analysis. At present, there are traditional methods
Conditional Probability 93.06%. represented in ratio methods: Dornenburg method [6],
Keywords: Dissolved gas analysis; fusion method, Multilayer modified Rogers four ratios, modified IEC ratios
Perceptron, Naïve Bayes, Percentages, Logarithmic. methods [7], HYOSUN Corporation gas ratio method
[8], three ratios technique [9], In addition, the graphic
1. INTRODUCTION methods are: Duval triangle method [10], Gouda triangle
method [11], pentagon methods [12] [13]. The traditional
The importance of power transformers lies in their methods are not accurate enough to know the type of
crucial role in ensuring the smooth operation of power fault, so they are weak and suffer from decision-making,

5
Scientific Bulletin of the Electrical Engineering Faculty – Year 23 No.2 (49) ISSN 2286-2455

but by turning to artificial intelligence, diagnosing faults procedure represents the pivotal phase for enhancing the
for transformers has become available and with high model's accuracy. In this study, three different input
precision. The methods used in the literature are: ANN vectors were employed, namely the original data input
vector (ppm), a percentage input vector, and a
[14], fuzzy logic [15], support vector machines [16], k-
Logarithmic input vector, as shown in Table 1.
nearest neighbor [17], Bayesian networks [18], ensemble
learning [19], and deep learning [20]. Table 1. Input features
Data Format
In this paper, both the neural network algorithm and the
naïve Bayes (NB) theory were presented using different X= [C2H6, C2H2, CH4, C2H4,
input vectors (ppm) H2]
input vectors and a proposed method for fusion them to
diagnose the six power transformer faults using the X= [%C2H6, %C2H2, %CH4,
KNIME analytics platform. input vectors (Percentage) %C2H4, %H2]

This paper is prepared as follows. In Section 2, the X= [LogH2, logCH4,

methodology utilized in this study is introduced, input vectors (Logarithmic) LogC2H2, LogC2H4, LogC2H6]
commencing with an elucidation of the employed dataset,
followed by a discussion of the various input vectors
considered, and an overview of the faults of transformers
targeted for detection. Section 3 provides an overview of 3. THE USED MACHINE LEARNING
the machine learning algorithms used in this study. In ALGORITHMS
Section 4, the model evaluation tools are shown. Section
3.1 Multilayer Perceptron (MLP)
5 evaluates the proposed fusion technique's performance
and effectiveness, comparing it with traditional and Artificial Neural Networks (ANNs) are symmetric
intelligent methods in the literature using test data. processors that exhibit high interconnectedness and
Finally, some conclusions of the presented study are represent an advanced and efficient algorithm for solving
given in the section 6. problems related to data scarcity and interference.
Neurons serve as the basic building blocks of neural
networks, drawing inspiration from the complex neurons
2. METHODOLOGY and neural connections found in the human brain [22].
The Multilayer Perceptron (MLP) is widely considered to
2.1 Data Description
be one of the most powerful and effective types of neural
Data collection is the first and most important step for networks for handling data. MLPs belong to the category
any study, as it is an integral part of the work that enables of supervised neural networks, which rely on desired
us to enhance the lifespan of machines by diagnosing and outputs for machine learning [23]. On the other hand, the
predicting their faults. The data for this study [21] was MLP algorithm comprises three primary layers. The
collected from 240 samples, comprising 129 electrical input layer establishes the connection between the data
fault samples and 111 thermal fault samples, which are input vector and the network, while the hidden layer
further categorized into six types of faults, as follows: varies from one algorithm to another, depending on the
Partial discharge (PD=27 cases), Low energy discharge sensory cells employed. Finally, the output layer
(D1=42 cases), High energy discharge (D2=55 cases), processes the data for the input vector via the sensory
Thermal fault below 300°C (T1=70 cases), Thermal fault cells and directs it through its activation function, as
between 300°C and 700°C (T2=18 cases), and Thermal represented by [24].
fault above 700°C (T3=28 cases). The training process
used 70 % of the data (168 samples), while the remaining 𝑦𝑦𝑖𝑖 = 𝑓𝑓�∑𝑛𝑛𝑗𝑗=1 𝜔𝜔𝑖𝑖𝑖𝑖 𝑥𝑥𝑗𝑗 + 𝜃𝜃𝑖𝑖 � (1)
30 % of the testing (72 samples) were used. 𝑥𝑥𝑗𝑗 : is the 𝑗𝑗th input of the 𝑖𝑖th neuron.
2.2 Input vector preparation 𝜃𝜃𝑖𝑖 : is called the bias of the 𝑖𝑖th neuron.
In order to ensure the quality of any data mining process,
𝜔𝜔𝑖𝑖𝑖𝑖 : is the weight from the 𝑗𝑗th input to the 𝑖𝑖th neuron.
the data must be processed by changing the input vectors
and data format to facilitate dealing with them. This 𝑦𝑦𝑖𝑖 : is the output of the 𝑖𝑖th neuron.

6
Scientific Bulletin of the Electrical Engineering Faculty – Year 23 No.2 (49) ISSN 2286-2455

The corresponding Figure 1 represents the classified

MLP architecture, which consists of three layers, input
layer, hidden layer, and output layer.

Figure 2. Simplified diagram of the NB algorithm

4. MODEL EVALUATION
Figure 1. The structure of the (MLP) based classifier
3.2 Naive Bayes (NB) In the field of data mining, there exists a dependable
method to assess the accuracy of data, which is crucial in
The Naive Bayes (NB) algorithm is technique inspired by supporting documentation systems. The effectiveness of
Bayes' theorem. It operates on the principle of probability the model is evaluated using statistical measures,
theory and statistical methods. Bayes theorem establish including Accuracy, recall, precision, and F-measure
independent values based on what precedes them [25]. [28].
The NB algorithm is considered "naive" because it
𝑇𝑇𝑇𝑇+𝑇𝑇𝑇𝑇
assumption of independence among all variables 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 = (4)
𝑇𝑇𝑇𝑇+𝑇𝑇𝑇𝑇+𝐹𝐹𝐹𝐹+𝐹𝐹𝐹𝐹
concerning class values, which may not hold in real- 𝑇𝑇𝑇𝑇
world scenarios. Nonetheless, this approach is favored 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 = (5)
𝑇𝑇𝑇𝑇+𝐹𝐹𝐹𝐹
for its fast-learning capabilities [26]. The Bayes theorem 𝑇𝑇𝑇𝑇
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 = (6)
algorithm offers the advantage of classifying various 𝑇𝑇𝑇𝑇+𝐹𝐹𝐹𝐹
objects, with the posterior probability equation being 2(𝑃𝑃𝑃𝑃 𝑒𝑒𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐×𝑅𝑅𝑅𝑅 𝑐𝑐𝑎𝑎𝑎𝑎𝑎𝑎)
𝐹𝐹−𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 = (7)
described as follows [27]: 𝑃𝑃𝑃𝑃 𝑒𝑒𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐+𝑅𝑅𝑅𝑅 𝑐𝑐𝑎𝑎𝑎𝑎𝑎𝑎

𝑃𝑃�𝐵𝐵 �𝐴𝐴�𝑃𝑃(𝐴𝐴)
𝑃𝑃(𝐴𝐴|𝐵𝐵) = (2)
𝑃𝑃(𝐵𝐵)
5. RESULTS AND DISCUSSION
𝑃𝑃(𝐴𝐴|𝐵𝐵) is the posterior probability of A under condition
To assess the diagnostic efficacy of the proposed model
B.
and to predict transformer faults that could endanger the
𝑃𝑃(𝐴𝐴) i prior probability A. condition of the electrical system, a dataset consisting of
240 samples was used which is as follows: 168 training
𝑃𝑃(𝐵𝐵|𝐴𝐴) posterior probability under condition A.
samples and 72 testing samples. The used data are in
𝑃𝑃(𝐵𝐵) is the prior probability B. ppm, percentage, and arithmetic. Where this paper
The posterior probability: represents one of the most powerful and accurate
algorithms, namely, the naive rule and neural networks in
𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑜𝑜𝑜𝑜𝑜𝑜×𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝
𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 = (3) diagnosing transformer faults, which in turn, their
𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒
prediction explaining the first stage, which is the
The principle of the NB algorithm can be illustrated preparation and processing of data involves collecting
through a simplified diagram and how to classify things; data and extracting features for input vectors. In the
Figure 2 illustrates the basic structure of a NB. second stage, classifiers are trained and tested. The third
stage, known as the Prediction Based Fusion (PBF)
method, is a machine learning and data science technique

7
Scientific Bulletin of the Electrical Engineering Faculty – Year 23 No.2 (49) ISSN 2286-2455

used to combine multiple predictions from MLP and NB The KNIME analytics platform environment allows
algorithms to improve the overall accuracy and engineers to develop and implement algorithms in a short
robustness of predictions. This is achieved by averaging time through a group of interconnected nodes, each of
the predictions of different models. Finally, in stage four, which performs a specific function and enables the expert
the proposed model is evaluated. They were combined to enter output and modify data [29]. Figure 4 represents
through ensemble learning to increase the accuracy of the the proposed model and PBF with a simplified
model, using the KNIME analytics platform. Figure 3 explanation of the steps.
represents the proposed model structure.
Data pre-processing: The input data is processed by the
Reader node, which scans the input file to ascertain the
quantity and categories of columns present. The
Duplicate Rows node then removes all duplicate rows
within the input table, followed by processing any
instances of missing data.
Model training and testing: First, the data is partitioned
into training and testing segments via the partition node.
Then, the model is trained and evaluated using a node
learner and predictor.
Combined prediction: The predictions are collected first
by the joiner node, which keeps the class and prediction
columns, and the PBF node is collected using the mean
predictions of both NB and MLP algorithms.
Evaluation: The proposed model was evaluated through
accuracy, recall, precision, and F-measure.
Figure 3. Flowchart of the proposed model

Figure 4. The proposed model using PBF to combine predictions of different classifiers
5.1 Performance evaluation results The best results of the proposed method for PBF at input
vector ratio of 97.22% accuracy, 98.33% recall, 97.62%
When comparing Figure 5 with Table 2, it can be seen
precision, and F-measure of 97.83% were superior to
that the PBF of assembling the algorithm of NB with the
both input vector ppm with 91.67% accuracy, 86.08
algorithm of MLP over both algorithms separately when
recall, 92.73% precision, and F-measure. 94.75, as well
the input vectors are ppm, percentages, and logarithmic.
as a logarithmic input vector with 95.83% precision,
The proposed method of PBF is characterized by its
96.14% recall, 94.72% precision and 95.32% F-measure.
reliability in fault detection in terms of overall accuracy,
and macro average : precision, recall, and F-measure.

8
Scientific Bulletin of the Electrical Engineering Faculty – Year 23 No.2 (49) ISSN 2286-2455

While for the algorithms separately, the high efficiency percentages and ppm, and the algorithm of the NB
of the MLP algorithm in terms of the input vector was in prevailed when the input vector was logarithmic.
Table 2. Performance results of the proposed model
Input features
Proposed ppm Percentages Logarithmic
classification Acc% Rec% Pre% F-m% Acc% Rec% Pre% F-m% Acc% Rec% Pre% F-m%
techniques
NB 77.78 70.94 72.30 81.91 91.67 94.56 90.88 92.45 93.06 94.56 91.81 92.61
MLP 84.72 81.53 86.70 81.6 94.44 96.67 95.83 95.77 88.89 84.97 85.63 84.37
Prediction
Based Fusion of 91.67 86.08 92.73 94.75 97.22 98.33 97.62 97.84 95.83 96.14 94.72 95.32
NB With MLP

100
90
80
70
60
50
40
30
20
10
0
Acc% Rec% Pre% F-m% Acc% Rec% Pre% F-m% Acc% Rec% Pre% F-m%
ppm Percentages Logarithmic

NB MLP Prediction Fusion

Figure 5. Performance chart algorithms for different evaluation metric

5.2 Comparisons with previous studies format. The PBF (vector Percentages) achieved an
accuracy rate of 97.22%, while PBF (vector Logarithmic)
For the effectiveness of the proposed fusion prediction
achieved an accuracy rate of 95.83%. These rates were
model and to make comparisons with previous studies,
higher than those achieved by traditional methods, such
the test data set consisting of 72 samples was used to find
as the Modified RRM/CEGB method 91.67% and
out the reliability of the proposed method in predicting
Modified RRM/IEC method 90.28%, as well as
transformer faults. Table 3 shows a comparison of the
intelligent methods, such as CSUS ANN 88.89% and
results obtained with other techniques, whether
Conditional Probability 93.06%.
traditional or Intelligent.
The comparison results demonstrated that the proposed
model outperformed both traditional and intelligent
methods in terms of diagnostic accuracy when using
percentage and logarithmic input vectors. This
improvement in accuracy can be attributed to the
preprocessing of the data, which involved converting the
original ppm data to either percentage or logarithmic

9
Scientific Bulletin of the Electrical Engineering Faculty – Year 23 No.2 (49) ISSN 2286-2455

Table 3. Comparison between the proposed method with traditional and Intelligent methods

Number of correct cases Total accuracy

Diagnostic methods PD D1 D2 T1 T2 T3 (%)

Traditional Three ratios technique [9] 4 6 1 12 0 8 43.06

methods
Modified RRM/CEGB [7] 7 10 15 20 6 8 91.67
Modified RRM/IEC [7] 6 10 15 20 6 8 90.28
Hyosun Corporation gas ratio method [8] 2 10 1 8 2 8 43.06
Duval triangle method [10] 4 8 2 10 0 5 43.06
Gouda triangle method [11] 3 10 3 9 0 8 45.83
Intelligent methods CSUS ANN [30] 7 10 14 19 6 8 88.89
Conditional probability [31] 7 11 16 19 6 8 93.06
Proposed method Prediction Based Fusion (vector ppm) 7 16 16 18 1 8 91.67
Prediction Based Fusion (vector 9 12 18 17 5 9 97.22
Percentages)
Prediction Based Fusion (vector 7 16 14 20 7 5 95.83
Logarithmic)
Aided by Data Clustering Method." IEEE Sensors Journal 22, no.
6 (2022): 5902 – 5910.
[2] Taha, Ibrahim B.M, and DiaaEldin, A. Mansour. "Novel power
6. CONCLUSION transformer fault diagnosis using optimized machine learning
methods." Intelligent Automation & Soft Computing 28, no. 3
This article proposes an efficient fusion method based on (2021): 739-752.
DGA data, utilizing the two best Machine Learning
[3] Passara, Chanchotisatien, and Vong Chanvichet. "Feature
algorithms, MLP and NB. By using input vectors of ppm, engineering and feature selection for fault type classification from
percentage, and logarithmic data, the proposed fusion dissolved gas values in transformer oil." International Computer
Science and Engineering Conference (ICSEC), 2021: 75-80.
method combines the classifiers and achieves higher
accuracy rates, recall, precision, and F-measure than the [4] Rao, U. Mohan, I. Fofana, K. N. V. P. S. Rajesh, and P. Picher.
"Identification and Application of Machine Learning Algorithms
classifiers alone. The performance of the proposed model for Transformer Dissolved Gas Analysis." IEEE Trans Dielectr
was evaluated using a test database and compared to Electr Insul. 28, no. 5 (2021): 1828–1835.
traditional and intelligent methods. Results showed that [5] Taha, Ibrahim B. M., Saleh Ibrahim, and Diaa-Eldin A. Mansour.
"Power Transformer Fault Diagnosis Based on DGA Using a
the proposed model outperformed all other methods, Convolutional Neural Network with Noise in Measurements."
achieving an accuracy rate of 97.22% for the PBF vector IEEE Access 9 (2021): 111162 –111170.
Percentages and 95.83% for PBF vector Logarithmic. [6] Wannapring, Ekkarach, Cattareeya Suwanasri, and Thanapong
These rates were higher than those achieved by Suwanasri. "Dissolved Gas Analysis methods for distribution
transformers." International Conference on Electrical
traditional methods such as the Modified RRM/CEGB Engineering/Electronics, Computer, Telecommunications and
91.67% and Modified RRM/IEC 90.28% methods, as Information Technology (ECTI CON), 2016: 1-6.
well as intelligent methods such as CSUS ANN 88.89% [7] Taha, Ibrahim B.M, Ayman Hoballah, and Sherif SM Ghoneim.
and Conditional Probability 93.06%. Overall, this study "Optimal ratio limits of rogers' four-ratios and IEC 60599 code
methods using particle swarm optimization fuzzy-logic approach."
demonstrates that the proposed fusion method is a IEEE Transactions on Dielectrics and Electrical Insulation 27, no.
promising approach for DGA analysis and can 1 (2020): 222-230.
significantly improve the accuracy of DGA-based [8] Sung-wook, Kim, Sung-jik, Kim, Hwang-dong, Seo, Jae-ryong,
diagnosis. Jung, Hang-jun, Yang, Michel, Duval, et al. "New methods of
DGA diagnosis using IEC TC 10 and related databases Part 1:
application of gas-ratio combinations." IEEE Transactions on
7. REFRENCES Dielectrics and Electr Insula 20, (2013): 685-690.

[1] Nasirul, Haque, Jamshed Aadil, Chatterjee Kingshuk, and [9] Gouda, Osama E., Salah H. El Hoshy, and Hassan H. EL Tamaly.
Chatterjee Soumya. "Accurate Sensing of Power Transformer "Proposed three ratios technique for the interpretation of mineral
Faults From Dissolved Gas Data Using Random Forest Classifier oil transformers based dissolved gas analysis." IET Generation,
Transmiss Distribu 12, (2018): 2650-2661.

10
Scientific Bulletin of the Electrical Engineering Faculty – Year 23 No.2 (49) ISSN 2286-2455

[10] Khiar, Mohd Shahril Ahmad, Sharin Ab Ghani, Imran Sutan [24] Xiang, Weiming, Hoang-Dung Tran, and Taylor T. Johnson.
Chairul, Yasmin Hanum Md Thayoob, and Young Zaidey Yang "Output reachable set estimation and verification for multilayer
Ghazali. "On-site OLTC monitoring using Duval triangle and neural networks." IEEE Transactions on Neural Networks and
DWRM." In Proceedings of the 2nd International Conference on Learning Systems 29, no. 11 (2018): 5777–5783.
Technology, Informatics, Management, Engineering and
Environment, TIME E 2014, Bandung, Indonesia 1 (2014): 216– [25] Ketjie, Viny Christanti Mawardi, and Novario Jaya Perdana.
221. "Prediction of credit card using the Naïve Bayes method and C4.5
algorithm. IOP Conference Series." Materials Science and
[11] Gouda, Osama E., Salah H. El‐Hoshy, and Hassan H. EL‐Tamaly. Engineering 1007, no. 1 (2020): 012161.
"Condition assessment of power transformers based on dissolved
gas analysis." IET Generation, Transmission & Distribution 13, [26] Mahamdi, Yassine, Ahmed Boubakeur, Abdelouahab Mekhaldi,
no.12 (2019): 2299-2310. and Youcef Benmahamed. "Power Transformer Fault Prediction
using Naive Bayes and Decision tree based on Dissolved Gas
[12] Duval, Michel, and Laurent Lamarre. "The duval pentagon-a new Analysis." ENP Engineering Science Journal 2, no. 1 (2022): 1-5.
complementary tool for the interpretation of dissolved gas analysis
in transformers." IEEE Electrical Insulation Magazine 30, no. 6 [27] Balaji, V.R., S.T. Suganthi, R. Rajadevi, V. Krishna Kumar, B.
(2014): 9–12. Saravana Balaji, and Sanjeevi Pandiyan. "Skin disease detection
and segmentation using dynamic graph cut algorithm and
[13] Mansour, Diaa-Eldin A. "Development of a new graphical classification through Naive Bayes classifier." Measurement:
technique for dissolved gas analysis in power transformers based Journal of the International Measurement Confederation 163
on the five combustible gases." IEEE Transactions on Dielectrics (2020): 107922.
and Electrical Insulation 22, no. 5 (2015): 2507–2512.
[28] Rahmad, F, Y Suryanto, and K Ramli. "Performance Comparison
[14] Li, Anyi, Xiaohui Yang, Zihao Xie, and Chunsheng Yang. "An of Anti-Spam Technology Using Confusion Matrix
optimized GRNN-enabled approach for power transformer fault Classification." IOP Conference Series: Materials Science and
diagnosis." IEEJ Trans. Electr. Electron. Eng 14, no. 8 (2019): Engineering 879, no. 1 (2020): 012076.
1181–1188.
[29] Nicola, George, Michael R. Berthold, Michael P. Hedrick, and
[15] Taha, Ibrahim B. M., Sherif S. M. Ghoneim, and Hatim G. Zaini. Michael K. Gilson. "Connecting proteins with drug-like
"A Fuzzy Diagnostic System for Incipient Transformer Faults compounds: Open source drug discovery workflows with
Based on DGA of the Insulating Transformer Oils." International BindingDB and KNIME." Database-Oxford, 2015: 1–22.
Review of Electrical Engineering (I.R.E.E.) 11, no. 3 (2016): 305-
313. [30] Ghoneim, Sherif SM, Ibrahim BM Taha, and Nagy I. Elkalashy.
"Integrated ANN-based proactive fault diagnostic scheme for
[16] Zhang, Yiyi, et al. "A fault diagnosis model of power transformers power transformers using dissolved gas analysis." IEEE
based on dissolved gas analysis features selection and improved Transactions on Dielectrics and Electrical Insulation 23, no. 3
Krill Herd algorithm optimized support vector machine." IEEE (2016): 1838-1845.
Access 7 (2019): 102803–102811.
[31] Taha, Ibrahim B.M, Diaa-Eldin A. Mansour, S. S. Ghoneim, et
[17] Benmahamed, Y, Y Kemari, M Teguar, and A Boubakeur. Nagy I. Elkalashy, "Conditional probability based interpretation
"Diagnosis of power transformer oil using KNN and nave bayes of dissolved gas analysis for transformer incipient faults." IET
classifiers." IEEE 2nd International Conference on Dielectrics Generation, Transmission & Distribution 11.4 (2017): 943-951.
(ICD), 2018: 1–4.
[18] Lakehal, Abdelaziz, and Fouad Tachi. "Bayesian duval triangle
method for fault prediction and assessment of oil immersed
transformers." Measurement and control 50, no. 4 (2017): 103-
109.
[19] Nanfak, Arnaud, Charles Hubert Kom, and Samuel Eke. "Hybrid
Method for Power Transformers Faults Diagnosis Based on
Ensemble Bagged Tree Classification and Training Subsets Using
Rogers and Gouda Ratios." Int. J. Intell.Eng. Syst 15, no. 5
(2022): 12‑24.
[20] Wu, Xiaoxin, Yigang He, and Jiajun Duan. "A deep parallel
diagnostic method for transformer dissolved gas
analysis." Applied Sciences 10, no. 4 (2020): 1329.
[21] Ibrahim, Saleh I., Sherif S.M. Ghoneim, and Ibrahim B.M Taha.
"DGALab: an extensible software implementation for DGA." IET
Gener. Transm. Distrib 12, no. 18 (2018): 4117-4124.
[22] Patekar, Kalinda D., and Bhoopesh Chaudhry. "DGA analysis of
transformer using artificial neutral network to improve reliability
in Power Transformers." IEEE 4th International Conference on
Condition Assessment Techniques in Electrical Systems
(CATCON), 2019: 1–5.
[23] Ramkumar, M., C. Ganesh Babu, K Vinoth Kumar, D Hepsiba4,
A. Manjunathan, and R. Sarath Kumar. "ECG Cardiac arrhythmias
Classification using DWT, ICA and MLP Neural Networks."
Journal of Physics: Conference Series 1831, no. 1 (2021): 012015.

11