Fault Classification in Power Transformers via Dissolved Gas Analysis and Machine Learning Algorithms: A Systematic Literature Review

Abstract

In electrical power systems, from generation power stations down to distribution substations, power transformers play a key role in ensuring reliable electricity transfer in the correct range from the generating source to the end-users. Over time, due to their operational demands and other various factors, transformers become susceptible to failures which threaten their reliability and life span. To address this issue, various transformer fault diagnosis methods are employed to detect and monitor the state of transformers, such as the dissolved gas analysis (DGA) method. In this paper, a systematic literature review (SLR) is conducted using the Preferred Reporting Items for Systematic Reviews (PRISMA) framework to record and screen current research work pertaining to the application of machine learning algorithms for DGA-based transformer fault classification. This study intends to assess and identify potential literature and methodology gaps that must be explored in this research field. In the assessment of the literature, a total of 124 screened papers published between 2014 and 2024 were surveyed using the developed PRISMA framework. The survey results show that the majority of the research conducted for transformer fault classification using DGA employs the support vector machine (32%), artificial neural network (17%), and k-Nearest Neighbor (12%) algorithms. The survey also reveals the countries at the forefront of transformer fault diagnosis and a classification based on DGA using machine learning algorithms. Furthermore, the survey shows that the majority of research conducted revolves around fault diagnosis with an emphasis on improving the accuracy of techniques such as SVM and ANN. At the same time, limited effort is put into other key metrics such as precision, Mean Squared Error, and R-Squared, and also, current works surveyed do not explore regularization techniques for preventing overfitting and underfitting of the proposed models.

1. Introduction

The reliability of every electrical reticulation system heavily depends on the health of its connected transformers as they play a key role in the transmission and distribution of electricity. However, over time, transformers tend to degrade and become susceptible to failures and breakdowns which may be caused by various elements such as lack of maintenance, environmental conditions, operation malfunctions, etc. One of the commonly adopted methods used to assess the condition of the transformers is the dissolved gas analysis (DGA) method. DGA is used to detect the presence of combustible gases such as hydrogen (H₂), methane (CH₄), ethylene (C₂H₄), ethane (C₂H₆), acetylene (C₂H₂), carbon monoxide (CO), and carbon dioxide (CO₂) in the transformer oil [1], and detect changes in their concentration [2]. The application of DGA has enabled researchers and professionals to conduct early diagnoses of transformers and to intervene with the necessary remedial methods to prolong the life of transformers [2,3,4]. The commonly used methods for assessing transformer oil concentration include Key Gas Analysis, Dornenberg and Rogers ratio methods, IEC Ratio, Duval triangle, and CIGRE method. However, various studies conducted on the application of the traditional interpretation methods have indicated their limitations and challenges associated with the uncertainty of the results due to manual handling of the test samples, the accuracy of the methods, inability to concurrently diagnose existing transformer faults, and identifying transformer incipient faults [5,6,7,8,9,10,11,12]. The combination of Ratio methods (IEC, Rogers, and Doernenburg) is limited in classifying the measured gas ratios into fault types in different interpretation techniques, often, inconsistent results are obtained when using different methods [5]. Also, it is noted that methods like the Rogers ratio are not capable of diagnosing borderline incipient faults.

1.1. Liquid Insulation

Liquid insulation plays a key role in the operation, safety, performance, and life span of transformers as it performs various functions such as providing electrical insulation between energized components and enhancing the dissipation of the heat generated by the transformer cores and windings. Liquid insulation also serves as a diagnosis agent that indicates the transformer’s health through testing of oil samples that are collected and analyzed to assess the chemical degradation of the oil caused by the energy dissipated through the oil [5]. A variety of insulating liquids are used in the industry, with mineral oil being the most commonly used insulation material. However, other insulating liquids are employed due to their superior tolerance in biodegradability and flash points when compared to the commonly used mineral oils [2,3,4]. Considering the various properties that each insulation liquid possesses which may be superior or inferior to other insulation liquids, carefully thought compromises are made when selecting a suitable liquid based on factors such as application, desired operating temperature, environmental impact, and associated costs.

1.2. Dissolved Gas Analysis

In most power system configurations under normal conditions, power transformers operate continuously providing the required power at the designed voltage level. Over time, the insulation of these transformers tends to gradually degrade due to various factors such as temperature, humidity, aging, loading stress, and oil quality. These elements make the transformers susceptible to insulation breakdowns that lead to transformer faults which interrupt the power system. To prevent interruptions, various power transformer monitoring techniques are employed to assess the condition of transformers, such as the dissolved gas analysis (DGA) technique. DGA is a routine test that monitors the presence and concentration of dissolved gases in the transformer oil as a result of electrical and thermal faults [5]. The dissolved gases present in the transformer oil that indicate incipient faults are Methane (CH₄), Ethane (C₂H₆), Ethylene (C₂H₄), Acetylene (C₂H₂), Hydrogen (H₂), Carbon Monoxide (CO), and Carbon Dioxide (CO₂). Generally, the procedure followed to conduct the analysis is as follows.

• Extract a small oil portion from the transformer using a specialized syringe.
• Extract the gases in the oil collected from the transformer in the laboratory.
• Use chromatography to analyze and identify the individual concentrations of the extracted gases.
• Then, use the identified gas concentration to diagnose the incipient faults that may occur in the transformer using the DGA-based interpretation techniques.

Various DGA-based interpretation techniques are widely used for interpreting DGA data to identify incipient transformer faults governed by different international standards and to classify transformer faults such as thermal faults, electrical faults, or partial discharge faults [13,14,15,16]. These methods include the key gas method, Dornenburg ratio method, Rogers ratio method, Nomograph Method, IEC Ration Method, Duval triangle method, and the CIGRE method.

With the exponential deployment of artificial intelligence in various industries, the utilization of machine learning algorithms in electrical equipment has seen substantial growth over the past decade, especially in condition monitoring applications.

For power transformers, machine learning algorithms have been employed for analyzing key diagnosis data associated with DGA which are used for fault detection and prediction with the intent of improving the accuracy and reliability of conventional methods based on the transformer’s liquid insulation properties and oil concentration obtained from laboratory results [5,17,18,19].

The proposed systematic literature review investigates the current scholarly works on the application of machine learning algorithms for DGA-based fault classification in power transformers. This review assesses different machine learning algorithms and models that are currently employed for transformer fault diagnosis. Furthermore, this SLR searches for gaps in the methods used in comparison with the conventional DGA interpretation techniques to pave the way for further research and developments in the research area. To better ascertain the state of the art concerning the application of machine learning (ML) algorithms in classifying and interpreting incipient transformer faults based on dissolved gas analysis, a systematic literature review (SLR) was conducted. The primary goal of this SLR was to evaluate the extent to which researchers have investigated the application of ML using transformer oil characteristics obtained from DGA results to identify possible methodological or theoretical gaps in the existing literature. For a guarded search of the literature, a set of restrictive boundaries and criteria were applied, this included research period range, keywords, research domain, language, and publication type as shown in Section 2.1 and Section 2.3.

1.3. Research Problem

The significance of transformers in a power system can never be understated, they are one of the most essential equipment, and their reliability is just as significant. To ensure the reliability and stability of the power system, the condition of the power transformers must be carefully monitored. Currently, different traditional transformer condition monitoring methods are used to detect incipient transformer faults, one of such methods is the dissolved gas analysis method (DGA). The DGA has various fault classification and interpretation methods used such as the key gas method, Doernenburg ratio method, Duval’s pentagon and triangle methods, and the CIGRE method.

Although the mentioned methods have been somewhat effective in their application, these methods have indicated limitations and challenges associated with the uncertainty of the results due to the manual handling of the test samples, the accuracy of the methods, the inability to concurrently diagnose existing transformer faults and identify transformer incipient faults. These limitations and challenges prompted the adaptation of machine learning techniques to resolve these issues; however, despite the application of these intelligent approaches, the literature shows that they exhibit deficiencies related to accurately classifying faults based on DGA data.

To address this issue, this systematic review investigates the potential of utilizing DGA-based fault source classification pattern recognition hybrid algorithms with optimized functions to improve the reliability and accuracy of machine learning algorithms that are used to predict incipient transformer faults.

1.4. Research Questions

To guide and narrow the SLR, a set of research questions pertaining to the application of ML algorithms in transformer fault classifications based on the DGA results are presented. These research questions are developed using the Population, Intervention, Comparison, and Outcome (PICO) framework to help ensure that a structured approach is used for a thorough search and analysis of the existing literature. Applying the PICO framework to this study, different types of transformer faults from different regions over a set period are studied, different DGA-based fault interpretation methods are explored, a comparison of the fault interpretation methods is studied, and the performance accuracy and shortcomings of the interpretation methods are analyzed. In addition to the research questions, search keywords and inclusion and exclusion criteria were also applied to narrow the search and obtain more detailed search results. The research questions are as follows:

• What are the primary objectives and research aims that current research addresses on DGA-based transformer fault classification?
• Which research repositories and publication types are commonly used to study DGA-based transformer fault classification?
• Which global countries and regions are at the forefront of DGA-based transformer fault classification research?
• How has the research in the area of ML application in interpreting transformer faults using DGA evolved over the period 2014–2024?
• Which machine learning techniques are employed to investigate DGA-based transformer faults?
• Which software and engineering tools are frequently used to compute and analyze the DGA data?
• Which methods and industrial specifications are utilized for DGA data collection, analysis, and interpretation of the results?
• Which DGA gas components are analyzed?
• What shortcomings and limitations are identified in the literature relating to DGA-based transformer fault classification?
• What are the sources of DGA datasets used in the literature?

1.5. Research Contribution

This study assesses and introduces a detailed systematic literature survey of artificial intelligent-based DGA interpretation techniques for transformer incipient fault detection and classification by providing a thorough review of current developments and research carried out in the DGA interpretation using the machine learning domain. The literature review provides valuable insights into the landscape of the research area in terms of countries at the forefront of this research, the sources providing research datasets, the machine learning algorithms employed, the software used for analysis, the specifications and standards consulted, the gas components analyzed, and the trajectory of publications in the research area. Furthermore, this systematic literature survey indicates the limitations of the current research undertaken in transformer fault detection and classification based on dissolved gas analysis interpretation methods, thus providing a platform for further research and development opportunities in the techniques for incipient fault detection and classification using machine learning.

1.6. Research Novelty

The SLR revealed that various studies explored the application of machine learning algorithms for transformer incipient fault detection and identification; however, there is no evidence of studies that have systematically analyzed and implemented the application of hybrid pattern recognition algorithms for DGA-based fault source classification, especially in an African context with the SLR indicating that Africa has only contributed 10% of publications in this research area over the last decade, with South Africa contributing 2%. Addressing this identified gap is essential as transformers are critical equipment in power systems and being able to accurately detect and classify incipient transformer faults based on their oil characteristics will aid in timeous predictive maintenance, thus allowing for appropriate downtime planning to maintain transformers while saving costs associated with the unplanned and prolonged downtime of power systems.

1.7. Limitations of the Study

This systematic literature review was conducted within the study limits defined in the inclusion and exclusion criterion in Section 2.1. Also, the literature surveyed in this study was collected only from reputable repositories and within a set time frame of 2014–2024 to restrict the search with scalable boundaries in line with the methodology employed in this study. In light of the set search restrictions, there is a possibility that other scholarly works in line with this study may have been omitted due to exclusion criteria requirements, such as equipment type which only permits transformers and not tap-changers and bushings, and missing algorithm information in the considered literature.

1.8. Research Organization

Section 1 introduces and gives an overview of transformer liquid insulation and the conventional and intelligent methods used to interpret DGA samples. Section 2 outlines the materials and methods employed to develop and conduct the systematic literature review. Section 3 discusses the results of the conducted systematic review in detail. Section 4 concludes the findings of this study.

2. Materials and Methods

In this section, the materials, methods, and systematic review framework employed to develop and conduct the systematic literature review focused on machine learning-based DGA interpretation techniques for transformer incipient fault detection and classification from literature published from 2014 to 2024 are presented. Figure 1 shows the flow diagram of the research methodology employed for this study.

Figure 1. Systematic literature review methodology flow diagram.

2.1. Eligibility Criteria

To restrict the search of the publications obtained from various repositories using the predetermined keywords, a review protocol consisting of the inclusion and exclusion criteria tabulated in Table 1 was used to ensure that the search remains within the research scope. Based on these restrictions, a total of 132 publications studied and investigated from 2014 to 2024 satisfied the criteria and were eligible to be used for the SLR.

Table 1. Systematic literature review inclusion and exclusion criteria.

Criteria	Inclusion Criteria	Exclusion Criteria
Research Domain	The research is related to both machine learning and the classification of faults based on transformer oil DGA.	The research is not related to both machine learning and the classification of faults based on transformer oil DGA.
Publication period	The publications are within the range of 2014 and 2024	The publications are not within the range of 2014 and 2024.
Methodology	The publications include the methodology used (e.g., algorithms and dataset used).	The publications do include the methodology used (e.g., algorithms and dataset used).
Publication Type	The published paper (article or conference proceeding) is from a reputable source.	The published paper (article or conference proceeding) is not from a reputable source.
Language used	The communication language is English.	The communication language is not English.
Equipment Type	DGA application on transformers only, but excluding tap-changers and bushings.	DGA application on transformers including tap-changers and bushings.

2.2. Information Sources

For this systematic literature review, the collected literature as per the inclusion and exclusion criteria was sourced from 10 high-quality, reputable repositories that facilitate peer reviews for all their published research, namely, the Association for Computing Machinery Digital Library, Emerald Insight, IEEE Xplore, MDPI, ResearchGate, Science Direct, Scientific Reports, Springer Link, The Institute of Engineering and Technology, and Wiley Online Library.

2.3. Search Strategy

A combination of various search methods, such as the line-by-line search, block search, and single-line search approach, was used to search for research papers in various repositories. In the line-by-line approach, the defined keywords were used on separate lines to search and retrieve articles. With the block search approach, the research concepts were grouped into different parts search and retrieve articles. Lastly, with the single-line approach, the defined keywords were combined using Boolean operators such as “OR” and “AND” to search and retrieve articles. Table 2 shows the list of keywords used to search.

Table 2. Search keywords.

Search Keywords

“Dissolved Gas Analysis” or “DGA” or “Transformer faults” or “Transformer anomalies” or “Transformer incipient faults” or “Transformer Fault Classification” or “Transformer fault detection” or “Transformer Maintenance” or “Transformer Condition monitoring” or “Transformer Health Index” or “Transformer Fault Diagnostics” or “Transformer Fault Prognosis” or “Transformer Reliability” or “Transformer Insulation Liquids” or “Transformer Oil” or “DGA Gas components” or “feature selection and extraction” or “Pattern Recognition Algorithms” or “Classification Algorithms”

2.4. Selection Process

To effectively conduct this systematic literature review, a sequential procedure consisted of 7 key steps, as illustrated in Figure 2. In step 1, key and concise research questions related to the review scope are formulated. In steps 2 and 4, the literature selection inclusion and exclusion criteria were defined to determine the literature to be used in the study, and the qualifying studies were selected based on the inclusion criteria. In step 3, a thorough search of the relevant literature in various databases was conducted. Once the relevant literature was selected, key information to be used for the analysis of the systematic literature review was extracted in step 5. In step 6, a data quality assessment of the selected literature was conducted, and lastly, for step 7, the results of the conducted systematic literature review have been presented in Section 3.

Figure 2. Systematic literature review procedure flow diagram.

2.5. Data Collection Process

The systematic literature review consisted of a comprehensive data collection process from 10 repositories, which was conducted by two reviewers. To restrict the search of the publications obtained from various repositories, a set of predetermined keywords was used as detailed in Section 2.3 and Table 2. The studied and investigated data were extracted from the literature published from 2014 to 2024, which satisfied the inclusion criteria defined in Table 1. In addition, a set of data quality assessment questions in support of the PRISMA flow diagram in Section 3.1, A proposed PRISMA flow diagram and an adapted Critical Appraisal Skills Programme (CASP) systematic review checklist was used to assess the validity of the results obtained from studies, ascertain the study results, and evaluate their impact and applicability.

• Q1. Is the research related to both machine learning and the classification of transformer faults based on DGA data?
• Q2. Does the research stipulate the analysis methodology used in the study?
• Q3. Does the research specify the performance accuracy of the methods used for DGA interpretation?
• Q4. Does the research stipulate DGA data sources and data size?
• Q5. Does the research consider DGA applications for transformers only and not tap changers and bushings?

The extracted data were then organized in an Excel spreadsheet in preparation for further analysis of trends and patterns that could be identified from the data to be presented in the form of graphs, charts, and tables for interpretation.

2.6. Data Items

This study intends to identify the extent to which machine learning has been used concerning transformer fault detection and classification; to achieve this and organize the data to be analyzed and interpreted into meaningful information, data collection and organization criteria were used based on the items listed in Table 3.

Table 3. Key data items from literature.

Criteria	Description
Paper Title	The title of the paper studied
Publication Year	The year in which the paper was published
Publication Type	The type of publication of the studied paper: journal or conference paper.
Publishing Repository	Which repository was used to collect the papers
Publication Country	The country in which the paper was published
Primary Condition Monitored	The type of transformer fault that was studied
Classification Method	The type of machine learning classification method: binary or multi-classification
Algorithm Used	The type of machine learning algorithm used
Algorithm Performance	The accuracy performance of the algorithm used
Dataset Used and Their Source	The source of the DGA dataset used for analysis
Algorithm Development Software Used	The software used to develop the machine-learning model
Industrial Standard/Specification Used	The industrial specification referenced or adhered to

2.7. Study Risk of Bias Assessment

In support of the PRISMA flow diagram, an adapted Critical Appraisal Skills Programme (CASP) systematic review checklist was used to assess the validity of the results obtained from studies, ascertain the study results, and evaluate their impact and applicability.

• Q1. Is the research related to both machine learning and the classification of transformer faults based on DGA data?
• Q2. Does the research stipulate the analysis methodology used in the study?
• Q3. Does the research specify the performance accuracy of the methods used for DGA interpretation?
• Q4. Does the research stipulate DGA data sources and data size?
• Q5. Does the research consider DGA applications for transformers only and not tap changers and bushings?

To address the data quality and risk of bias, a set of assessment questions for the literature papers analyzed was used with a scaling matrix used to assess the quality of the papers primarily consisting of three criteria, “yes” as a score of “1”, “partially” as a score of “0.5”, and “no” as a score of “0”. A summary of the results of the studied papers is tabulated in Table 4.

Table 4. Data quality and risk of bias assessment.

Paper ID	Q1	Q2	Q3	Q4	Q5	Total	%
1, 2, 4, 5, 6, 7, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 29, 30, 31, 32, 34, 35, 36, 37, 38, 39, 40, 41, 42, 44, 45, 46, 47, 48, 49, 50, 52, 53, 54, 55, 60, 62, 66, 67, 68, 70, 71, 72, 73, 76, 78, 79, 81, 82, 83, 85, 86, 88, 90, 91, 92, 93, 94, 95, 97, 98, 100, 102, 104, 105, 106, 107, 108, 110, 111, 112, 115, 116, 119, 121	1	1	1	1	1	5	100
3, 8, 26, 27, 28, 43, 56, 58, 61, 63, 69, 74, 77, 78, 84, 89, 96, 99, 101, 109, 113, 114, 117, 118, 120, 122, 124	1	1	1	0.5	1	4.5	90
33, 51, 57, 59, 64, 65, 80, 87	1	1	1	0	1	4	80
75, 123	1	1	0	0.5	1	3.5	70

2.8. Effect Measures

Effect measures were used to present the results obtained from the analysis of the systematic review data pertaining to the application of machine learning algorithms in transformer fault detection and classification. The effect measure process commences with defining the research questions which guides the systematic review data collection; this is followed by appropriate measures that align with the desired search outcome, and this is reported in the form of tables, graphs, and charts to ensure that consistency and clarity are maintained throughout the search results. After that, a meta-analysis of the data is conducted to distinguish the heterogeneity of the data studied.

2.9. Synthesis Analysis

2.9.1. Criteria for Studies Eligible for Synthesis

The criteria used for the analysis of the systematic literature review in Section 3.2 were extracted from all the eligible research papers; this information included publication period (2014–2024), research type, the repository used, the country of publication, the transformer fault type monitored, the machine learning algorithm, classification technique, data-processing technique, feature extraction and selection techniques used, the accuracy of the machine learning algorithms, the dataset source and size, the gas components analyzed, and the analysis software used, as well as the industrial standards referenced in the analysis. Additionally, all the papers used had to adhere to the inclusion and exclusion criteria defined in Section 2.1, Table 1, as well as the risk of bias assessment requirements as stated in Section 2.7.

2.9.2. Data Preparation

In preparing the data for analysis, a well-managed Excel spreadsheet was used for data management to categorize all key elements to be extracted from the literature and also note any missing data from the studied literature which could affect the accuracy of the systematic review. In handling the missing data, it was excluded from the final quantification analysis and only noted as “Not Specified” to indicate a gap in the studied literature that qualified for analysis as per the inclusion criteria.

2.9.3. Results Presentation

Various methods were utilized to visually present the analysis results, these included tables, pie-charts, bar graphs, and line graphs for listing results, comparing results, assessing the distribution of the results, studying patterns and trends and identifying gaps in the literature. The first step was to tabulate all contents of all the studied papers in an Excel spreadsheet; this was then followed by the application of the Excel Data Analysis Toolbox to visually present the results using the methods described above.

2.9.4. Methods for Data Synthesis and Meta-Analysis

A categorical approach was employed in Excel to systematically classify the studies based on the criteria described in Section 2.6, Table 3, which assessed factors such as the publication year, publication type, publishing country, transformer fault classification methods, algorithm used, dataset source, etc. This approach was pivotal in assessing the state of the art and identifying patterns and trends in the research area.

2.9.5. Exploring Causes of Heterogeneity

Due to the uncertainties observed in the analyzed results, it was imperative to understand the root of these uncertainties for the study to reach a sound and reliable conclusion. Additionally, understanding these uncertainties would further aid in gauging the extent to which machine learning is utilized in transformer fault detection and classification and identifying gaps within the research field. In exploring the possible causes of the uncertainties, subgroup analyses of various data attributes were conducted to assess the variations among these attributes such as the type of algorithm, its accuracy performance, dataset size, its source, and the country in which the study was conducted. By analyzing the type of algorithm used and its accuracy performance, the study was able to link the general performance of the algorithms used in the literature against the type of methodology required for transformer fault detection and classification (multi-classification problem type). Similarly, the sample size has a substantial impact on the performance of models.

2.9.6. Sensitivity Analysis Approach

To better understand the relationship between machine learning applications in transformer fault detection and classification, as part of the systematic review, a sensitivity analysis was conducted to assess the robustness of the synthesized results where various transformer fault classification methods, algorithms, datasets, sources, and performance are covered to aid in ensuring scrutinized results that are reliable. Figure 3 shows the sensitivity analysis process used for this study.

Figure 3. Sensitivity analysis process.

2.10. Reporting Bias Assessment

To ensure the validity of the systematic review on the application of machine learning in transformer fault detection and classification based on dissolved gas analysis, it was imperative to assess the risk of bias in the gathered literature. The areas that were identified as having potential biases included the selection of repositories for extracting the literature, the country in which the literature was published, the limitations presented by publication language (only English, non-translated literature was considered for this study), and the inclusion of only certain attributes deemed appropriate for this study on the selected literature. For the bias assessment, data analysis was carried out manually to prevent any errors and missing out on key findings; this varies from data collection from different repositories, categorizing the data in Excel, and drawing up conclusions.

2.11. Certainty Assessment

In support of the PRISMA flow diagram in Figure 4, an adapted Critical Appraisal Skills Programme (CASP) systematic review checklist was used to assess the validity of the results obtained from studies, ascertain the study results, and evaluate their impact and applicability.

Figure 4. Proposed PRISMA flow diagram.

• Q1. Is the research related to both machine learning and the classification of transformer faults based on DGA data?
• Q2. Does the research stipulate the analysis methodology used in the study?
• Q3. Does the research specify the performance accuracy of the methods used for DGA interpretation?
• Q4. Does the research stipulate DGA data sources and data size?
• Q5. Does the research consider DGA applications for transformers only and not tap changers and bushings?

To address the data quality and the risk of bias, a set of assessment questions for the literature papers analyzed was used with a scaling matrix used to assess the quality of the papers primarily consisting of three criteria, “yes” as a score of “1”, “partially” as a score of “0.5”, and “no” as a score of “0”. A summary of the results of the studied papers is tabulated in Table 4.

3. Results

This section presents the results of the studied and analyzed literature to address the research questions defined in Section 1.4.

3.1. Study Selection

Using the keywords listed in Table 2 in Section 2.3, the combined literature search yielded a total of 26,573 research papers across 10 repositories. From these results, a substantial amount of 26,292 papers were excluded from the study based on their eligibility, leaving a total of 281 for assessment. Upon further assessment and screening of the research papers using the proposed inclusion and exclusion criteria, only a total of 124 papers were deemed eligible for inclusion in the systematic literature review study. Table 5 shows the initial search results from each considered repository before the screening using all the elements of the inclusion and exclusion criteria.

Table 5. Search results for the number of publications per repository.

Research Repositories	Number of Results Obtained from the Initial Search
Association for Computing Machinery Digital Library	852
Emerald Insight	786
IEEE Xplore	3531
MDPI	5324
ResearchGate	1493
Science Direct	4120
Scientific Reports	1265
Springer Link	4836
The Institute of Engineering and Technology	2353
Wiley Online Library	2013
Total	26,573

Furthermore, to screen the eligibility of the search results, a Preferred Reporting Items for Systematic Reviews (PRISMA) flow diagram was generated to record the search results and screen the collected papers, employing a data and statistics-driven quantitative approach. Figure 4 shows the proposed PRISMA-adapted flow diagram for identifying, screening, and selecting the literature for further analysis.

3.2. Study Characteristics

To obtain insights into the research questions defined in Section 1.4, the 124 resulting papers from the screening conducted using the PRISMA flow diagram were examined concerning transformer fault interpretation methods based on DGA data. The studied literature shows that the existing literature seeks to improve on the existing traditional fault interpretation methods which include SVM, ANN, kNN, etc. The systematic literature review further shows that even though research has been conducted in this research area, there are still shortcomings and limitations. Out of the 10 repositories which were consulted for the literature, MDPI had the highest output at 28 publications, closely followed by Science Direct, Springer Link, and IEEE Xplore at 22, 20, and 19, respectively. The rest of the papers were published by the other six repositories. Figure 5 shows the distribution of publishers.

Figure 5. The distribution of publishers.

Furthermore, from these repositories, it was investigated which type of research output was reported and it was found that journal articles are the most published papers at 105 outputs compared to the 18 outputs of conference proceedings shown in Figure 6.

Figure 6. Type of research outputs.

The study also reveals compelling evidence that China is at the forefront of this research with a dominant 53% of the total publications on electrical transformers monitoring based on DGA using machine learning algorithms. China’s interest in and work on machine learning algorithms and transformer fault monitoring using DGA reflect their involvement in the adaptation of artificial intelligence in electrical power systems. From China, the other countries following from a distance are India at 5%, Saudi Arabia at 4%, and the UK, Algeria, Australia, and Indonesia all at 3%. The rest of the publications spread over 20 other countries including South Africa at 2%. The significant gap between China and the rest of the world in terms of participation in this research area could result from the educational and industrial investments China have made in advancing their infrastructures and being intentional in leading the fourth industrial revolution compared to the other countries which are following the trend. Figure 7 shows the distribution of countries publishing works on transformer DGA and machine learning.

Figure 7. Countries of presented publication sources.

Furthermore, with the growing development of machine learning integration into electrical power systems, there has been a substantial growth in research publications of machine learning algorithms application in electrical transformers monitoring techniques to interpret faults using DGA data. Figure 8 shows research activities on machine learning and transformers with the presented data ranging from January 2014 to April 2024. From the figure, it is noticeable that there has been a steady incline in the publications, with years 2021 and 2023 topping the charts in research outputs, while 2024, at the time of generating this graph (April 2024), only had seven reported outputs. The trends exhibited by this graph suggest that artificial intelligence in electrical transformers is attracting researchers to explore its potential in monitoring transformers. With the growing interest in artificial intelligence, there will likely be linear or exponential growth in the research outputs in this study area.

Figure 8. Ten-year publication trends in transformer DGA and machine learning.

Table 6 tabulates a cited comprehensive overview of the current state-of-the-art DGA interpretation methods for transformer faults based on artificial intelligence techniques of scholarly works published from 2014 to 2024, with a key focus on the machine learning methods used for fault diagnosis and their performance against the conventional DGA interpretation methods.

Table 6. Transformer DGA-based fault classification literature survey.

Ref.	Cites	Year	ML Algorithm	ML Dataset Source	Contribution
[20]	145	2014	Adaptive Neuro-Fuzzy Inference System	IEC TC10	It develops an Adaptive Neuro-Fuzzy Inference System model to identify transformer incipient faults using DGA. The model achieved an accuracy of 98% when applied to the IEC standard, 97% on the Rogers ratio method, and 99% on the Doernenburg ratio method.
[21]	73	2019	Improved Krill Herd SVM	IEC TC10 and Private Company	It proposes an Improved Krill Herd optimized SVM algorithm based on cross-validation principles to improve fault diagnosis accuracy. The accuracy of the proposed model reaches 85.71%.
[22]	2	2024	Deep CNN	Literature	It presents an asset management technique for classifying incipient faults and predicting the health index of transformers using deep CNN; the model achieves 87% accuracy.
[23]	4	2022	kNN, DT	New York Power Authority	It presents a fault detection model using kNN. It applies the key gas method to detect outliers, and then the Basic Gas Ratio is used to detect faults, the accuracy of this model reaches 88%.
[24]	178	2017	Deep believe network	State Grid Corporation of China and Literature	It presents a fault diagnosis method based on the deep believe networks algorithm which assesses the transformer health index including the impact of discharge and overheating. This model reaches an accuracy of 90%.
[25]	90	2019	Hidden Markov model	State Grid Zhejiang Electric Power Co. Ltd.	It develops a dynamic early prediction and warning model which combines the Gaussian mixture model with the Hidden Markov model. The proposed model achieved an accuracy of 95%.
[26]	34	2020	Lasso regression	State Grid Zhejiang Electric Power Co. Ltd.	It uses the Pearson correlation coefficient to filter and optimize the feature set to reduce the redundancy of the selected features and applies Lasso regression to build a multi-dimensional linear model of the selected features. The proposed model achieved an accuracy of 87.5%.
[27]	66	2021	Teaching-learning based optimization	Egyptian Electricity Holding Company	It develops an efficient teaching–learning-based optimization (TLBO) model to enhance the diagnostic accuracy of the transformer faults based on gas concentration percentages limits and gas ratios. The developed model achieved an accuracy of 86.27%.
[28]	19	2021	Gaussian process multi-classification	General Electric	It proposes a Gaussian process multi-classification model to enhance and produce a probabilistic interpretation of the analysis results. The proposed model achieved an accuracy of 95%.
[29]	5	2023	k-means clustering algorithm	IEC TC10 and Literature	It proposes a two-step hybrid method consisting of a k-Means Clustering algorithm based on the generic algorithm for subset analysis. The proposed model achieved a diagnostic accuracy of 98.29%.
[30]	38	2020	Hybrid gray wolf optimization algorithm	Egyptian Electricity Utility and Literature	It develops a hybrid gray wolf optimization algorithm to produce an efficient diagnostic code matrix based on dissolved gas percentages for accurate fault diagnosis using the fuzzy system to map the limits of gas ratios. The developed model achieved an accuracy of 97.45%.
[31]	39	2021	Auto-WEKA	IEC TC10 and Literature	It adopts the Auto-Weka framework for fault detection by choosing the best configuration and comparing its performance against other existing methods to assess its efficiency. The proposed model achieved an accuracy of 93.08%.
[32]	134	2016	ANN	Egyptian Power System Network	It develops an ANN model which is used to feed inputs to a Smart Fault Diagnostic Approach (SFDA) based on the integration of the conventional DGA interpretation techniques. The developed model achieved an accuracy of 97.3%.
[33]	86	2015	ANN, SVM, ELM, SaE-Elm	Electrical Utilities and Literature	It presents an incipient fault diagnosis method developed using the combination of a multi-classification Self-adaptive Evolutionary Extreme Learning Machine (SaE-ELM) algorithm and arctangent transform (AT). The presented model achieved an accuracy of 93.57%.
[34]	173	2015	SVM	IEC TC10	It proposes the application of SVM and the optimum dissolved gas ratios obtained using the generic algorithm which also optimizes the SVM parameters. The proposed model achieved a diagnosis accuracy of 87.18%.
[35]	55	2017	Bayesian networks	IEC TC10	It presents a multi-classification probabilistic diagnosis framework based on Bayesian networks and hypothesis testing for dissolved gas analysis. The presented model achieved a diagnosis accuracy of 88.9%.
[36]	61	2021	CNN	Electrical Utilities and Literature	It proposes a convolutional neural network (CNN) model for predicting transformer fault types under different noise levels in measurements, the proposed model is then applied to conventional ratios, new ratios, and hybrid ratios. The proposed model achieved a diagnosis accuracy of 98.5%.
[37]	68	2017	SVM and kNN	IEC TC10	It assesses the combination of Particle swarm-optimized SVM and kNN algorithms to Duval’s pentagon 1 method for transformer fault diagnosis. The developed model is then tested against the five faults detectable in transformer oil. The assessed model achieved a diagnosis accuracy of 88%.
[38]	53	2022	ANN	Egyptian Ministry of Electricity and Literature	It aims to detect incipient faults in power transformers using DGA using an integrated method that implements conditional probability, clustering, Duval triangle, Roger’s four ratios refined, and an artificial neural network. The proposed model achieved a diagnosis accuracy of 93.6%.
[39]	35	2020	SVM	Electrical Utilities	It proposes a support vector machine-based Duval pentagon Method to determine the severity of transformer faults by assessing the graphical Duval pentagon method using incorporating gas level, gas rate, and DGA interpretation results into one interpretation method. The proposed model achieved a diagnosis accuracy of 88%.
[40]	1	2024	TimesNet	Literature	It proposes a multi-model fusion transformer fault diagnosis method based on TimesNet and Informer with the adoption of an MCA module to reduce complexity and computational burden. The proposed model achieved a diagnosis accuracy of 88.46%.
[41]	19	2021	KosaNet	Kenya Power Ltd.	It proposes a multinomial KosaNet classification model which is based on the decision tree algorithm, and investigates the application of various machine learning techniques to detect incipient faults and interpret DGA results. The proposed model achieved a diagnosis accuracy of 99.98%.
[42]	33	2020	Spatial Clustering	Huzhou Power Supply Company and IEC TC10	It proposes a fault diagnosis method based on the correlation coefficient density clustering which extrapolates the correlation coefficient of DGA data using density clustering. The proposed model achieved a diagnosis accuracy of 90%.
[43]	4	2023	TPE-XGBoost	Literature	It proposes an Extreme Gradient Boosting algorithm for fault diagnosis; it utilizes a Random Forest Regression algorithm to determine missing DGA data and hyperparameter optimization of data using a Tree-structured probability density estimator algorithm. The proposed model achieved a diagnosis accuracy of 89.5%.
[44]	9	2022	BCSVM	South African transformer manufacturer	It proposes a binary classification support vector machine with the application of the kernel function for transformer fault prediction; the model accuracy is then compared against the IEC gas ratio and ANN methods and yields an accuracy of 92%.
[45]	16	2021	Logistic Regression	Not Specified	It presents the application of a regularized Logistic Regression Algorithm for improving DGA interpretation accuracy. The model accuracy is then compared against the conventional and intelligent methods and yields an accuracy of 85.6%.
[46]	18	2023	ISCSO	Not Specified	It proposes a power transformer fault diagnosis method based on landmark isometric mapping (L-Isomap) and Improved Sand Cat Swarm Optimization (ISCSO) to optimize the BiGRU (ISCSO-BiGRU). The proposed model achieved an accuracy of 94.8%.
[47]	11	2023	Improved BP neural network	Not Specified	It proposes an improved transformer fault diagnosis method based on a residual BP neural network which uses mapping to extract gas feature information with more significant feature differences after fusion expansion.
[48]	2	2023	Deep believe network	Literature	It applies a non-coding ratio method to generate 35 fault gas ratios based on the 5 fault gases, subsequently refined through correlation analysis to eliminate redundant feature variables, resulting in 15 significantly representative fault gas ratios. The model achieved an accuracy of 90.2%.
[49]	35	2018	GRNN	Shandong Substation	It proposes a kernel principal component analysis and a generalized regression neural network (GRNN) using an improved fruit fly optimization algorithm (FFOA) to select the smooth factor. The proposed model achieved an accuracy of 95%.
[50]	40	2018	ISGOSVM	IEC TC10	It proposes a fault diagnostic method integrating feature selection and diagnostic model optimization using genetic algorithm SVM feature screen and SVM classifier optimized by improved social group optimization. The proposed model achieved an accuracy of 92.86%.
[51]	35	2018	Improved fuzzy c-means clustering	UCI Iris dataset	It improves the fuzzy c-means clustering algorithm and the resultant dissolved gas data clustering by utilizing an exponential form of the membership function to obtain distance monotony. The proposed model achieved an accuracy of 94.67%.
[52]	30	2019	Gene Expression Programming	Literature	It introduces an improved approach to overcome the limitations of conventional DGA interpretation techniques based on Gene Expression Programming; this approach further automates and standardizes the DGA interpretation process.
[53]	32	2019	TWSVM	State Grid Corporation of China	It presents a fault diagnosis model based on twin support vector, Restricted Boltzmann Machines, and chemical reaction optimization for data preprocessing and improving efficiency and accuracy. The presented model achieved an accuracy of 94.64%.
[54]	27	2019	HMSVM	Hebei North Electric Power Company and Literature	It proposes a fault diagnosis method for solving misdiagnosis problems of conventional methods using hypersphere multiclass support vector machine (HMSVM) and Dempster–Shafer (D–S) Evidence Theory (DET). The proposed model achieved an accuracy of 90%.
[55]	45	2019	Hybrid gray wolf optimizer and LS-SVM	Literature	It proposes and applies a hybrid gray wolf optimized least square support vector machine (HGWO-LSSVM) fault diagnosis method. The proposed model achieved an accuracy of 97.45%.
[56]	10	2020	GAPSVM	State Grid Corporation of China and IEC TC10	It proposes a genetic algorithm-optimized probabilistic SVM integrated with the fuzzy three-ratio method which judges the decision boundary according to its output probabilities. The proposed model achieved an accuracy of 86.8%.
[57]	26	2021	NN	Laboratories	It presents a method based on the analysis of the gases dissolved and the analysis of furan compounds for identifying different faults using neural networks by extending the diagnosis of the operating conditions of the power transformers by assessing paper degradation. The presented model achieved an accuracy of 95.7%.
[58]	52	2021	SVM-BA	Agilent Chemical Laboratory and Literature	It proposes a fault diagnostic method based on DGA data using a coupled system of support vector machine (SVM)-bat algorithm (BA) and Gaussian classifiers. The proposed model achieved an accuracy of 93.75%.
[59]	39	2021	Improved Gray Wolf Optimizer and PNN	Jiangxi Power Supply Company	It proposes a probabilistic neural network-based fault diagnosis model and utilizes the improved gray wolf optimizer as a smoothing factor of the pattern layer of PNN to improve the classification accuracy. The proposed model achieved an accuracy of 99.71%.
[60]	9	2023	SVM	Turkish Electricity Transmission Company	It designs a hybrid fault diagnosis system based on DGA methods and fuzzy logic and uses a mathematical approach and support vector machines as decision-making methods. The designed model achieved an accuracy of 96.23%.
[61]	6	2022	ANN	South African Independent Power Producer	It proposes a multilayer artificial neural network and gas ratio method for identifying inchoate faults by employing the oil properties of dissolved gas analysis. The proposed model achieved an accuracy of 99.98%.
[62]	7	2023	ANN	Botnet DGA Dataset	It employs an artificial neural network model to enhance the accuracy of the Rogers ratio method by optimizing the hyperparameters of the model to maximize the accuracy and minimize the architectural complexity of an ANN. The proposed model achieved an accuracy of 90.70%.
[63]	55	2021	ANN	Egyptian Electricity Transmission Company	It introduces a combination of conventional DGA techniques (Duval, Roger’s four ratios refined, and IEC refined techniques) which is constructed based on the outputs of each technique, this was further refined using the artificial neural network. The introduced model achieved an accuracy of 84.96%.
[64]	11	2021	PNN and bio-inspired optimizer	Jiangxi Province Power Supply Company	It proposes an artificial intelligence fault diagnosis method based on a probabilistic neural network (PNN) and bio-inspired optimizer, where the PNN is used as a classifier the Improved Salp Swarm Algorithm is used for optimization. The proposed model achieved an accuracy of 97.44%.
[65]	41	2015	ANN	IEC TC10	It presents a fault prognosis method combining evolutionary and connectionist mechanisms into a hybrid model. The presented model achieved an accuracy of 98%.
[66]	86	2014	Fuzzy logic	Various Industrial Transformers	It introduces a new fuzzy logic approach that aids in standardizing DGA interpretation and identifies transformer critical ranking based on DGA data. The approach relies on incorporating all traditional DGA interpretation techniques (Roger, Doerenburg, IEC, key gas, and Duval triangle methods) into one expert model. The proposed model achieved an accuracy of 88%.
[67]	16	2022	SVM	State Grid Corporation of China and IEC TC10	It applies the support vector machine algorithm in a transformer fault diagnosis model aimed at improving the accuracy of DGA interpretation methods which are known for militations. The proposed model achieved an accuracy of 97.35%.
[68]	5	2023	SVM, NBC, and BPNN	Literature	It proposes the application of the Artificial Bee Colony algorithm to select the optimum dissolved gas feature for the computation of the model. The proposed model achieved an accuracy of 90.49%.
[69]	5	2023	Intelligent Voting Ensemble Learning	State Grid Corporation of China, IEC TC10, and Literature	It proposes a feature selection and ensemble learning-based methodology for fault diagnosis by selecting efficient features from classifiers and using an Intelligent Voting Ensemble Learning technique to generate a diagnosis model. The proposed model achieved an accuracy of 92.6%.
[70]	92	2017	Fuzzy logic	Not Specified	It presents a model that detects and classifies faults based on the state of transformers based on the Mamdani model and Sugeno model in a fuzzy expert system for fault probability. A fuzzy logic model is then applied in deciding on testing. The presented model achieved an accuracy of 95%.
[71]	22	2021	MCSVM	State Grid of Zibo Power Company and IEC TC10	It develops a hybrid fault diagnosis model based on the C-set&FCM-MCSVM algorithm, the combination of C-set&FCM optimizes the model by solving issues associated with dataset imbalance and outliers and improves the performance. The developed model achieved an accuracy of 88.9%.
[72]	2	2024	MLANN	Laboratories and Literature	It proposes a hybrid fault detection method based on multilayer neural networks and Expert Knowledge to enhance the performance of the model by assessing the characteristics of DGA datasets. The proposed model achieved an accuracy of 92.8%.
[73]	12	2023	Deep Residual Network	China Power Grid Company	It proposes a fault diagnosis method based on data augmentation with the application of kernel principal component analysis for mapping original data to a high-dimensional feature space to address data imbalance. The proposed model achieved an accuracy of 88%.
[74]	1	2023	CNN	Literature	It proposes a fault diagnosis method based on a Conditional Variational Automatic Encoder composed of full connection layers for expanding datasets. Once datasets are augmented, the convolutional neural network as a classifier due to its superior feature extraction capabilities. The proposed model achieved an accuracy of 90%.
[75]	32	2023	Fuzzy C-means clustering algorithm	Not Specified	It develops a fault diagnosis and prognosis model using fuzzy logic controller (FLC) and fuzzy clustering means (FCM) and simulates it to examine the health status of the insulation. The developed model achieved an accuracy of 97.5%.
[76]	7	2023	HATT-RLSTM	Not Specified	It develops an online monitoring system that combines Hierarchical Attention Network (HATT) and Recurrent Long Short-Term Memory Network (RLSTM) for condition monitoring. The developed model achieved an accuracy of 88.15%.
[77]	9	2023	AdaBoost-WBLS	State Grid Corporation of China and Literature	It proposes a novel transformer fault diagnosis model that utilizes a Filter-Wrapper Combined Feature Selection method and an AdaBoost integrated weighted broad learning system (AdaBoost-WBLS) to select significant features and eliminate redundant ones. The proposed model achieved an accuracy of 95.05%.
[78]	13	2021	kNN, RT, RF	Not Specified	It investigates the application of random tree, random forest, and k-Nearest Neighbors algorithms for classifying cross-country faults from transformer disturbances. The investigated model achieved an accuracy of 98.68%.
[79]	4	2024	RF	IEC TC10 and Literature	It presents a multi-method technique that uses Scoring Index and random forest principles to combine conventional DGA interpretation methods into a multi-method technique. The presented model achieved an accuracy of 92.57%.
[80]	43	2016	PNN	Not Specified	It proposes a fault prediction method using a probabilistic neural network algorithm which applies its non-linear processing capabilities to predict internal faults. The proposed model achieved an accuracy of 80%.
[81]	7	2023	Res-Variational-Auto-Encoder and ensemble learning	Guangzhou Substation	It proposes a hybrid fault diagnosis method comprising Res-Variational-Auto-Encoder and ensemble learning algorithms. This method extracts fault characteristics using a residual convolutional neural network. The proposed model achieved an accuracy of 98%.
[82]	27	2023	RF	Not Specified	It develops a fault identification method based on the combination of Random Forest and SMOTE algorithms to reduce the limitations and complexities associated with implementing graphical DPM. The developed model achieved an accuracy of 96.5%.
[83]	3	2022	BPMF and GRU-NN	Industrial Transfromer	It proposes a transformer oil dissolved gas concentration for data loss scenarios using Bayesian probabilistic matrix factorization (BPMF) and gated recurrent unit (GRU) neural network.
[84]	13	2022	Saerima model	Industrial Transfromer	It presents a time series forecasting model to forecast the dissolved gas concentration based on the Seasonal Autoregressive Integrated Moving Average (SARIMA).
[85]	10	2023	Deep belief network and BPNN	Electric Power Research Institute of Yunnan Power Grid Co., Ltd.	It introduces a deep belief in network-based fault diagnosis methods based on the characteristics of the precipitated gases. The introduced model achieved an accuracy of 90%.
[86]	17	2022	Deep Coupled Dense CNN	IEC TC10	It reconstructs the transformer oil characteristic gas to improve its features and compare fault types against various gases using the Adaptive Synthetic Oversampling (ADASYN) algorithm. The proposed model achieved an accuracy of 94.05%.
[87]	2	2024	Bagged Trees, kNN, and QSVM	Various South African Substations	It proposes a fault diagnosis model used to analyze oil samples that showed stray gassing characteristics by combining four different classifiers to enhance fault diagnosis accuracy. The proposed model achieved an accuracy of 87.7%.
[88]	1	2024	MLANN	Not Specified	It presents a multilayer artificial neural network model for fault detection which utilizes a meta-heuristic optimizer to improve its accuracy. The presented model achieved an accuracy of 87.8%.
[89]	862	2023	Snake Optimization–random forest	Zhejiang Power Supply Company	It presents a fault diagnosis technique based on TLR-ADASYN balanced data to address the shortcomings associated with DGA sample imbalance and fault misjudgment. The presented model achieved an accuracy of 97.06%.
[90]	0	2020	Deep belief network and ANN	Punjab State Electricity Board	It presents a transformer oil aging assessment model based on an RM-LA-based DBN + NN algorithm with the application of RM-LA-based optimization. The presented model achieved an accuracy of 97.4%.
[91]	14	2020	Genetic algorithm fuzzy Q-learning	HPSEB and Literature	It develops a generic algorithm-tuned fuzzy classifier for fault classification by segregating fault types using DGA samples. The proposed model achieved an accuracy of 97.51%.
[92]	14	2016	MLSVM	Turkish Electricity Transmission Corporation and Literature	It presents a fault diagnosis method based on DGA data using a multilayer support vector machine algorithm which is optimized by a hybrid combination of generic, Differential Evolution, and Particle Swarm Optimization. The presented model achieved an accuracy of 92.00%.
[93]	5	2018	Fuzzy c-means algorithm and WNN	Not Specified	It proposes a fuzzy clustering and Flower Pollination algorithm to address the shortcomings associated with the wavelet neural network by improving the convergence and lowering diagnosis error. The proposed model achieved an accuracy of 88.50%.
[94]	8	2019	BASA-MkRVR	Panzhihua Electric Power Bureau	It presents a hybrid fault prediction method comprising Bettle Antennae Search Algorithm–mixed kernel relevance vector regression to select optimum parameters and to test the superiority of the activation functions.
[95]	4	2017	PSO-SVM	IEC TC10 Database	It proposes a fault diagnosis model based on the three DGA ratios and applies the optimization technique of Particle Swarm Optimization to the support vector machine algorithm. The proposed model achieved an accuracy of 85.71%.
[96]	2	2023	SVM, DT, and kNN	Not Specified	It reviews and compares the support vector machine, decision tree, and k-Nearest Neighbor algorithms utilized to evaluate the performance of diagnostic fault determination by generating reliable data samples for minority and majority classes to balance the diversity and distribution of DGA datasets. The proposed model achieved an accuracy of 95.70%.
[97]	3	2016	Fuzzy logic	Not Specified	It compares the conventional method and artificial intelligence method, IEC 60599 gas ratios, and Fuzzy Evidential for fault detection using DGA. The proposed model achieved an accuracy of 90.60%.
[98]	2	2024	ANN and NB	Literature	It proposes a hybrid method consisting of the artificial neural networks and Naïve Bayes techniques used to obtain statistical parameters accuracy, recall, precision, and F-measure. The proposed model achieved an accuracy of 97.22%.
[99]	55	2014	SVM and kNN	IEC10DB	It proposes the application of the k-Nearest Neighbor imputation method to predict missing DGA dataset values and use these datasets with the predicted values to classify faults using a support vector machine.
[100]	59	2015	CANN-PSO	Electrical Utility	It proposes a hybrid incipient fault prediction method consisting of artificial neural networks and various Particle Swarm Optimization techniques. The proposed model achieved an accuracy of 98.00%.
[101]	17	2020	Self-Paced Ensemble algorithm	IEC TC10 Database, Ministry of Electricity and Energy of Egypt, North China Grid and Zhejiang Grid	It investigates and presents a fault diagnosis method based on a Self-paced Ensemble using classification hardness to improve diagnosis accuracy. The investigated model achieved an accuracy of 95.10%.
[102]	16	2018	Fuzzy c-means clustering algorithm	China Southern Power Grid and IEC TC10	It presents a fault detection model based on the fuzzy c-means clustering algorithm and the dimensionality reduction techniques Improved Principal Component Analysis. The presented model achieved an accuracy of 95.10%.
[103]	4	2022	Improved Gray Wolf Optimizer and LS-SVM	Not Specified	It proposes a fault diagnosis hybrid support vector machine model with improved gray wolf optimization optimized least squares–support vector machine to address convergence and population diversity shortcomings of traditional support vector machine algorithms. The proposed model achieved an accuracy of 95.85%.
[104]	3	2021	Improved Firefly Algorithm and SVM	IEC TC10	It proposes a fault diagnosis technique based on Improved Firefly Algorithm–Linear Programming Boosting–Classification and Regression Tree hybrid combination. It adjusts the weights of weak classifiers builds stronger ones and then optimizes them for computation. The proposed model achieved an accuracy of 91.72%.
[105]	20	2018	Parzen Window estimation	Industrial Transformers	It proposes a model to address shortcomings of Roger’s ratios, IEC ratios, and Duval triangles methods based on the Parzen Window Estimation. This model utilizes five gases to compute fault probability. The proposed model achieved an accuracy of 95%.
[106]	56	2018	GA-SVM	IEC TC10 and Local Power Utilities	It investigates the use of a hybrid feature selection technique combined with the generic algorithm and support vector machine. This is further backed by an adaptive synthetic technique and arctangent transformation technique to improve training set properties.
[107]	34	2017	RVM–ANFIS algorithm	IEC TC 10 and Jinliang	It proposes a hybrid technique that combines the Relevant Vector Machine and the Adaptive Neural Fuzzy Inference System to address ambiguous data characteristics related to conventional DGA interpretation methods. The proposed model achieved an accuracy of 95%.
[108]	2	2022	Isomap and IChOA—LSSVM	Not Specified	It presents a fault diagnosis method for improving diagnosis accuracy based on Isomap and IChOA-LSSVM. The model starts by reducing dimensionality and then optimizes the least square support vector machine algorithm to improve the overall model accuracy. The presented model achieved an accuracy of 90.83%.
[109]	5	2022	ISSA—SVM	Shaanxi Province Transformers (China)	It proposes a fault diagnosis method based on the Improved Sparrow Search Algorithm and the optimized support vector machine. Various activation functions are used to enhance the accuracy of the model. The proposed model achieved an accuracy of 91.43%.
[110]	9	2021	Refined Deep Residual Shrinkage Network	IEC TC10, State Grid Hubei Company, and State Grid Shandong Company	It proposes a fault identification model to improve the performance of transformers by coupling a high-dimensional feature map of gas concentrations and associated variables. The proposed model achieved an accuracy of 97.68%.
[111]	49	2014	PSO-SVM	UK National Grid	It presents a feature prioritization and detection model that assesses fault types, specific gas ratios, and statistical characteristics of the DGA data. The presented model achieved an accuracy of 96.50%.
[112]	12	2021	IVSVM	China Southern Power Grid and Literature	It introduces a combined fault diagnosis model based on improved information entropy and a vague support vector machine in which it uses improved information entropy to obtain gas weights from DGA datasets and the trained data to obtain classifiers using a vague support vector machine. The introduced model achieved an accuracy of 90.60%.
[113]	9	2020	MLP	Electric Technology Research Association of Japan and IEC TC10	It constructs a multilayer perceptron model for fault detection in transformers using the Keras Library, The model is tested using two datasets and it achieved an accuracy of 96.70%.
[114]	17	2020	MVO-PNN	China Electric Power Research Institute	It presents a hybrid model for fault diagnosis based on probability neural network optimized by Multi-Verse Optimizer algorithm which has a strong global convergence from improving fault diagnosis accuracy. The presented model achieved an accuracy of 97.49%.
[115]	0	2023	Improved SVM	Not Specified	It introduces a fault location recognition model based on k-means and constructs a high-precision with fast convergence diagnosis. The proposed model achieved an accuracy of 98.40%.
[116]	0	2022	SVM	IEC TC10 Database and Literature	It proposes an improved hybrid fault diagnosis model comprising the Slime Mold Optimized support vector machine and logarithmic arctangent transform with data transformation and optimization capabilities of this model equipped to improve the diagnosis accuracy. The proposed model achieved an accuracy of 90.10%.
[117]	32	2017	Fuzzy inference system	Iran Power Grid	It studied a large dataset of more than 3000 transformers and applied a fuzzy inference system to validate the state of suspicious transformers by validating the fault detection accuracy. Also, the proposed fuzzy inference system was able to detect the decomposition of the insulation paper. The proposed model achieved an accuracy of 91.20%.
[118]	33	2018	BN, kNN, MPL, and J48 DT	Not Specified	It compares and analyzes various machine learning algorithms used for DGA fault classification, i.e., Bayes network, multilayer perceptron, k-Nearest Neighbor, and J48 decision tree. It further assesses the impact of applying the AdaBoost algorithm and bootstrap aggregation (bagging) on the performances of these algorithms. The J48 decision tree algorithm was the best-performing algorithm and achieved an accuracy of 94.50%.
[119]	16	2020	ANN-GSA and SVM-GSA	Not Specified	It proposes a hybrid feature selection technique for determining fault types from DGA data using a Gravitational Search Algorithm. The proposed model achieved an accuracy of 98%.
[120]	25	2019	Fuzzy inference system and ANN	Literature	It proposes a multiple fault detection technique to improve the conventional interpretation methods to also cater for multiple fault diagnosis in non-binary situations. The proposed model achieved an accuracy of 97%.
[121]	6	2015	Hybrid fuzzy logic	IEC TC10 and Substation Equipment Maintenance and Management Web System of Taiwan Power Company	It utilizes a hybrid fuzzy technology to integrate five conventional dissolved gas analysis interpretation approaches for transformer fault diagnosis. The model achieved an accuracy of 94.36%.
[122]	3	2021	Modified Clustering	Not Specified	It proposes an unsupervised methodology for diagnosing transformer conditions using a modified clustering technique to classify the condition of transformers using DGA datasets.
[123]	3	2022	Improved Harris Hawks Optimized Random Forest	Literature	It proposes a fault diagnosis method based on improved Harris Hawks optimized random forest using logistic chaotic mapping to adjust the key parameters, non-linear energy factor to control the algorithm to transition, and Gaussian Mutation for strengthening the searchability. The proposed model achieved an accuracy of 90.18%.
[124]	13	2022	HHO-KELM	IEC TC10	It presents a fault diagnosis based on the Harris Hawks optimization algorithm optimized kernel extreme learning machine, this optimization enhances the fault diagnostic performance of the model which achieved an accuracy of 95.70%.
[125]	30	2015	FCM-GA-BPNN	Provincial Power Company of State Grid Corporation	It integrates a combination of the conventional IEC and Doernenburg interpretation techniques with the genetic algorithm and fuzzy c-means clustering algorithm (FCM) optimized BP for improved diagnosis accuracy and generalization ability while overcoming issues associated with data noises, over-fitting, and convergence. The proposed model achieved an accuracy of 89.71%.
[126]	17	2019	SVM and kNN	KEPCO	It presents a classification system of transformer faults and defects based on transformer parts and structures using the gases obtained from the DGA datasets and applied to support vector machine and k-Nearest Neighbors algorithms. The diagnostic performance of this proposed model achieved an accuracy of 88%.
[127]	3	2022	Boosted Trees, RUS Boosted Trees, and Subspace kNN	Local Power Utility	It proposes a fault detection model that is capable of handling imbalanced DGA data while improving the accuracy of predictions using machine learning algorithms Boosted Trees, RUS Boosted Trees, and Subspace k-Nearest Neighbors. The Boosted Trees algorithm was the best-performing algorithm and achieved an accuracy of 82.60%.
[128]	12	2019	ANFIS	Iran-Transfo Company Chemical Laboratory and Damavand Power Substations	It presents a hybrid incipient faults detection method using the DGA information and Adaptive Neuro-Fuzzy Inference System algorithm. It further uses the Association Rules Mining Technique to select features from the DGA dataset. The proposed model achieved an accuracy of 99.64%.
[129]	2	2023	Fuzzy logic and deep learning	Electrical Utility	It proposes a fuzzy logic DGA interpretation technique that combines the use of gas level, gas rate, and the Duval pentagon method with the aim of increasing the reliability of fault detection and performance accuracy. The proposed model achieved an accuracy of 98.80%.
[130]	8	2017	PNN	Literature	It proposes a transformer fault diagnosis technique employing Multi-Input and Multi-Output Polynomial Neural Network. This technique is divided into two parts, the first part single output classification model is trained and the transformer state is divided into binary states normal and abnormal, the results are then used to build a fault diagnosis model.
[131]	5	2023	RF, Tree Ensemble, Gradient Boosted Tree, and Extreme Gradient Tree Algorithms	Literature	It discusses the impact that the feature input vector of traditional diagnostic methods has on the performance of DGA-based artificial intelligent fault diagnosis methods. The algorithms used to assess the performance were the random forest, Tree Ensemble, Gradient Boosted Tree, and Extreme Gradient Tree. The random forest algorithm was the best-performing algorithm and achieved an accuracy of 98.37%.
[132]	16	2022	Decision tree, support vector machine, random forest, neural network, Naïve Bayes, and AdaBoost	Not Specified	It examines machine learning algorithms commonly used for transformer fault identification using DGA data namely decision trees, support vector machine, random forest, neural networks, Naïve Bayes, and AdaBoost as well the traditional DGA fault identification methods DTM1, DTM4, and DTM5. Of all these algorithms and methods, the random forest algorithm was the best-performing method with a performance accuracy of 99.80%.
[133]	0		CVA-SVM	Not Specified	It proposes a combined fault diagnosis method based on Canonical Variate Analysis and support vector machine with the Canonical Variate Analysis adopted due to its multivariate state space model identification ability using experimental data. The proposed model achieved an accuracy of 88.74%.
[134]	5	2021	Fuzzy logic	IEC TC10	It proposes a fault diagnosis method for multiple transformer incipient faults using fuzzy logic based on the concentration of gases obtained from the transformer oil. The proposed model achieved an accuracy of 98.30%.
[135]	1	2020	Deep believe network	Power Grid Corporation and IEC TC10	It introduces a fault diagnosis and classification technique based on the combination of the Restricted Boltzmann Machine and deep belief network algorithms. The proposed model achieved an accuracy of 93.20%.
[136]	0	2022	SSA-SVM	Not Specified	It proposes a hybrid fault diagnosis method based on the kernel principal component analysis and Sparrow Search Algorithm-support vector machine to integrate the ratio of gas concentration and machine learning capabilities to improve the rate of fault diagnosis accuracy. The proposed model achieved an accuracy of 88.89%.
[137]	0	2023	PSO-SVM	Not Specified	It proposes a fault diagnosis model comprising the support vector machine algorithm and optimization technique Particle Swarm Optimization for enhancing the performance and accuracy of the model by optimizing the support vector machine parameters. The proposed model achieved an accuracy of 82%.
[138]	0	2023	GA-FCM	Literature	It proposes a hybrid fault diagnosis model with improved accuracy using the Genetic Algorithm and fuzzy c-means clustering algorithm algorithms. This model showed the capabilities of distinguishing between different fault types. The proposed model achieved an accuracy of 80%.
[139]	1	2022	AdaBoost–DT	Not Specified	It proposes a fault classification technique based on the continuous iteration of the AdaBoost–decision tree algorithm. This model further addresses inadequate accuracy levels from previous publications using a single learner. The proposed model achieved an accuracy of 97.18%.
[140]	0	2015	Coupled Hidden Markov model	State Grid Corporation of China and Literature	It proposes a coupled hidden Markov model based on dissolved gas analysis for incipient fault detection using the theory of scalar quantization to preprocess the gas concentrations. The proposed model achieved an accuracy of 92%.
[141]	0	2020	BPNN–Improved Apriori Algorithm	Not Specified	It presents improved association rules and a BP neural network model for incipient fault detection and prediction. The model combines the rough algorithm with the association rules where the rough set of data is reduced and the BP neural network is used to train and predict the operation state of the transformer. The proposed model achieved an accuracy of 86.30%.
[142]	0	2018	Discrete Hopfield Neural Network	Not Specified	It presents a discrete Hopfield neural network for identifying the condition of transformers by diagnosing it for the five possible states observable using gas concentrations. The improved Rogers three-ratio technique is then used to obtain the input vectors required for model establishment.
[143]	2	2016	PSO-RBFNN	Not Specified	It proposes a fault diagnosis model based on the RBF network algorithm optimized using Immune Particle Swarm Optimization for improving the diagnosis accuracy of incipient internal insulation faults. The proposed model achieved an accuracy of 90%.

3.3. Results of Individual Studies

The study shows that various machine learning algorithms were used for transformer monitoring using DGA data; the support vector machine (SVM), artificial neural networks (ANN), and k-Nearest Neighbor (kNN) stand out as the commonly used and adapted algorithms with percentages of 32%, 17%, and 12%, respectively, (excluding the hybrid algorithms of the aforementioned). Some of the studied algorithms have hybrid algorithms adapted with an attempt to improve their performance such as those of ANN found in this study which includes CNN, GRNN, improved BP neural network, deep believe network, and Adaptive Neuro-Fuzzy Inference System, all of which contribute an additional combined 10%, which bring the overall total of neural networks contribution to 27%. Based on the evaluated studies, the noted contributing factors behind the dominant application of the abovementioned algorithms include their accuracy, simplicity, and computational speed when compared to other algorithms.

Other machine learning algorithms that had fewer contributions and are being explored in the studied literature include the k-Means Clustering algorithm, improved fuzzy c-means clustering, Bayesian Networks, hybrid gray wolf optimization algorithm, Gaussian Process Multi-classification, Lasso regression, Hidden Markov model, and Logistic Regression, among others. The combined contribution of these algorithms is 26%. Figure 9 shows the distribution of machine learning algorithms used for transformer monitoring using DGA data.

Figure 9. Machine learning algorithms used for transformer monitoring.

In terms of the software adopted in the literature, MATLAB 2019b is the most commonly used software for modeling and computing machine learning algorithms for transformer DGA applications with a percentage of 31%, surprisingly, Python (v.3.8) is used by only 2% of the papers studied, and LabView (v.2024) just 1%. The rest of the documents which contributed 66% did not specify the software or engineering tool used for computation. Figure 10 shows the distribution of the software used.

Figure 10. Software used for DGA analysis.

On the industrial standards utilized, many of the papers studied do not specify the globally adopted standards used in the analysis. This causes concern as having an accurate depiction of standards would aid in understanding which ones are the most effective and accurate. From the collected studies, 36% of the papers used IEC 60599, while 9% specified C57 IEEE 104 standards, and the remaining 55% of papers did not specify the standard used. Figure 11 shows the distribution of standards used for DGA interpretation.

Figure 11. Standards used for DGA interpretation.

This study shows that a majority of the DGA datasets used to conduct the simulations are obtained from the literature (34%) and the IEC TC10 Database (32%), and a combined percentage of 25% is obtained from transformers used in 16 different industrial power utilities, manufacturers, substations, and private companies. The rest of the 9% of sources of the datasets are not specified in papers. Figure 12 shows the distribution of the sources.

Figure 12. DGA datasets sources.

3.4. Results of Syntheses

This section discusses the results of the syntheses conducted to evaluate the characteristics and risk of bias among the literature studied following the PRIMSA eligibility criterion and the detailed research question defined in Section 1.4.

3.4.1. Characteristics and Risk of Bias of Contributing Studies and Their Statistical Results

To gauge the depth of machine learning applications in transformer fault detection and interpretation using dissolved gas analysis, a synthesis of the results obtained from the PRISMA process was conducted to highlight key findings, methodologies utilized, and the biases that exist in the systematic review study. This study was conducted using literature from reputable repositories in various countries with varying expertise in transformer health management. The study shows that significant research based on transformer dissolved gas analysis revolves around transformer fault diagnosis with an emphasis on improving the accuracy of known and widely used techniques such as SVM and ANN, while limited effort is put into other key algorithm metrics such as precision, Mean Squared Error, and R-Squared. The results also reveal that China is at the forefront of this research area with a dominant 53% of the total publications on transformer fault diagnosis and classification based on DGA using machine learning algorithms which slightly introduces a bias factor in the results obtained as a majority of the studies and observations are based on the Chinese context, as depicted in Figure 7. Additionally, the availability of transformer DGA datasets from manufacturers and DGA-based chromatography labs to researchers and the public introduces a risk of bias as the study shows that 57% of the datasets used in research are from the IEC TC10 database (25%) and the published literature (32%), while 11% is not disclosed in the published studies as shown in Figure 12.

3.4.2. Causes of Heterogeneity

With the dominance of China in the application of machine learning in transformer fault detection and interpretation based on DGA data due to their fast-growing economy, technology interventions, and swift adaptability to artificial intelligence applications. This brings about heterogeneity with the rest of the world which is playing catch-up to the advancements made by China. Also, the utilization of different DGA datasets from different countries, which exposes the transformers in-service to unique environmental factors that influence the accumulation of dissolved gases in the transformer oil, plays a key role in the heterogeneity of the studied results.

3.4.3. Sensitivity Analysis

To assess the robustness of the synthesized results concerning the relationship between machine learning applications in transformer fault detection and classification based on DGA, a sensitivity analysis was conducted, specifically on the type of algorithms used. This analysis was crucial in determining the impact of each algorithm on the insights required to assess the application of machine learning in transformers. Figure 9 in Section 3.3 highlights the distribution of various algorithms used, with SVM, ANN, and kNN being the commonly used methods at 32%, 17%, and 12%, respectively. The other 39% is distributed between 23 other algorithms. The common application of SVM, ANN, and kNN in transformers may be linked to their high performance in solving binary and multi-classification problems.

3.5. Reporting Biases

To provide an accurate interpretation of the systematic review, eligibility criteria with 542 biases were introduced to monitor the literature included in the study. In Table 7, the identified biases are listed with the challenges they present; these include the research domain investigated, the publication period, the methodology used, the publication type, and the communication language.

Table 7. Identified biases from studied literature.

Bias Type	Challenge
Research domain investigated	Only research related to both machine learning and the classification of faults based on transformer oil DGA may be included.
Publication period	Some studies falling outside the defined timeframe may have better results than the ones within the defined timeframe.
Methodology	Only studies with high accuracies may be published leaving the opportunity to compare the performance of another study.
Publication type	Publications from the identified reputable sources may not be the best-performing ones.
Language used	Good and eligible papers published in other native languages may be unfairly omitted from this study.

3.6. Certainty of Evidence

To assess the certainty of the studied literature, a scaling matrix was used to evaluate its eligibility in conjunction with the PRISMA flow diagram in Figure 4. A set of questions that validates the eligibility of the studied literature was thoroughly applied using a scaling matrix consisting of three criteria to rank the paper’s performance, with “yes” as a score of “1”, “partially” as a score of “0.5”, and “no” as a score of “0”. A summary of the results of the studied papers is tabulated in Table 4 under Section 2.7.

4. Discussion

To obtain insights into the research questions defined in Section 1.4, the 124 resulting papers from the screening conducted using the PRISMA flow diagram were examined considering transformer fault interpretation methods based on DGA data.

Q1. 

What are the primary objectives and research aims that current research addresses on DGA-based transformer fault classification?

The systematic review shows that the existing literature seeks to improve on the existing traditional fault interpretation methods which include SVM, ANN, kNN, etc. The systematic literature review further shows that even though research has been conducted in this research area, there are still shortcomings and limitations.

Q2 

. Which research repositories and publication types are commonly used to study DGA-based transformer fault classification?

The study found that among the 10 repositories used to extract papers, MDPI has the highest outputs at 28 publications, closely followed by Science Direct, Springer Link, and IEEE Xplore at 22, 20, and 19, respectively. The rest of the papers were published by the other six repositories. Figure 5 shows the distribution of publishers. Furthermore, from these repositories, it was investigated which type of research output is reported and it was found that journal articles are the most published papers at 105 outputs compared to the 18 outputs of conference proceedings shown in Figure 6.

Q3. 

Which global countries and regions are at the forefront of DGA-based transformer fault classification research?

From the study, there is compelling evidence that China is at the forefront of this research with a dominant 53% of the total publications on electrical transformers monitoring based on DGA using machine learning algorithms. China’s interest and work conducted on machine learning algorithms and transformer fault monitoring using DGA reflects their involvement in the adaptation of artificial intelligence in electrical power systems. From China, the other countries following from a distance are India at 5%, Saudi Arabia at 4%, and the UK, Algeria, Australia, and Indonesia all at 3%. The rest of the publications spread over 20 other countries including South Africa at 2%. Figure 7 shows the distribution of countries publishing works on transformer DGA and machine learning.

Q4. 

How has the research in the area of ML application in interpreting transformer faults using DGA evolved over the period 2014–2024?

With the growing development of machine learning integration into electrical power systems, there has been a substantial growth in research publications of machine learning algorithms application in electrical transformers monitoring techniques to interpret faults using DGA data. Figure 8 shows the research activities on machine learning and transformers with the presented data ranging from January 2014 to April 2024. From the figure, it is noticeable that there has been a steady incline in the publications with years 2021 and 2023 topping the charts in research outputs, while 2024, at the time of generating this graph (April 2024), only had seven reported outputs. The trends exhibited by this graph suggest that the application of artificial intelligence in electrical transformers has been attracting researchers to explore the potential it possesses in transformer monitoring. With the growing interest in artificial intelligence, there will likely be linear or exponential growth in the research outputs in this study area.

Q5. 

Which machine learning techniques are employed to investigate DGA-based transformer faults?

The study shows that various machine learning algorithms were used for transformer monitoring using DGA data, the support vector machine (SVM), artificial neural networks (ANN), and k-Nearest Neighbor (kNN) stand out as the most commonly used and adapted algorithms with percentages of 32%, 17%, and 12%, respectively (excluding hybrid algorithms of the aforementioned). Some of the studied algorithms have hybrid algorithms adapted with an attempt to improve their performance such as those of ANN found in this study which includes CNN, GRNN, improved BP neural network, deep believe network, and Adaptive Neuro-Fuzzy Inference System, all of which contributes an additional combined 10%, which bring the overall total of neural networks contribution to 27%.

Other machine learning algorithms that had fewer contributions and are being explored in the studied literature include the k-Means Clustering algorithm, improved fuzzy c-means clustering, Bayesian Networks, hybrid gray wolf optimization algorithm, Gaussian Process Multi-classification, Lasso regression, Hidden Markov model, Logistic Regression, among others. The combined contribution of these algorithms is 26%. Figure 9 shows the distribution of machine learning algorithms used for transformer monitoring using DGA data.

Q6. 

Which software and engineering tools are frequently used to compute and analyze DGA data?

From the papers studied, MATLAB is the most commonly used software for modeling and computing machine learning algorithms for transformer DGA applications with a percentage of 31%, surprisingly, Python is used by only 2% of the papers studied, and LabView just 1%. The rest of the documents, which contributed 66%, did not specify the software or engineering tool used for computation. Figure 10 shows the distribution of the software used.

Q7. 

Which methods and industrial specifications are utilized for DGA data collection, analysis, and interpretation of results?

A large number of the papers studied do not specify the globally adopted industrial standards used in the analysis. This causes concern as having an accurate depiction of standards would aid in understanding which ones are the most effective and accurate. From the collected studies, 36% of the papers used IEC 60599, while 9% specified C57 IEEE 104 standards, and the remaining 55% of papers did not specify the standard used. Figure 11 shows the distribution of standards used for DGA interpretation.

Q8. 

Which DGA gas components are analyzed?

The studied papers apply a holistic approach when describing the gases studies, with almost all of them mentioning all gases, i.e., hydrogen (H2), methane (CH4), ethylene (C2H4), ethane (C2H6), acetylene (C2H2), carbon monoxide (CO), and carbon dioxide (CO2). This could be due to the methods used in these papers being adapted from the traditional DGA interpretation techniques, namely, the Dornenburg ratio method, the Rogers ratio method, the Nomograph Method, the IEC Ration Method, the Duval triangle method, and the CIGRE method.

Q9. 

What are the sources of DGA datasets used in the literature?

The study shows that a majority of the DGA datasets used to conduct simulations are obtained from the literature (34%) and the IEC TC10 Database (32%), and a combined percentage of 25% is obtained from transformers used in 16 different industrial power utilities, manufacturers, substations, and private companies. The rest of the 9% of sources of the datasets are not specified in the papers. Figure 12 shows the distribution of the sources.

Shortcomings and Limitations Identified in the Literature Relating to DGA-Based Transformer Fault Classification

• Unspecified transformer type and capacity MVA sizes: A great majority of the papers do not specify the transformer type and their capacity; this could have provided indicative insights into which transformer types and sizes are most prone to which transformer fault based on the DGA interpreted results.
• Unspecified oil type of DGA samples: Many of the papers do not provide the type of insulation oil that was used from which the DGA results were extracted. This could have provided valuable insights into which insulation oil between mineral oil, vegetable oil, esters oils, etc., is the most reliable and which ones exhibit high concentrations of gases over a period of usage.
• Transformer age: Very few of the studied papers specified the age of the transformers from which the oil samples were taken; this could have helped provide insight into the relationship between transformer age and its gas concentration performance based on DGA sample results.
• Lack of data preprocessing and feature extraction techniques: Some of the papers do not specify the techniques used for data preprocessing and feature techniques; this could have helped to provide the readers with insights as to which data preprocessing and feature extraction techniques are the most effective in relation with the machine learning algorithms employed. To optimize the performance of machine learning models, before processing raw datasets for training, it is a pre-requisite for them to be thoroughly cleaned by removing as much noise and redundant information as possible so that only useful dataset attributes are evaluated [144,145]. This process is essential to eliminate the inconsistent and unwanted noise often hidden in the input raw data [146,147]. In addition to data preprocessing, another critical step required for data preparation is feature extraction. Feature extraction is a dimension-reduction technique used to scale down datasets by eliminating non-essential features and properties in datasets used to train machine learning models and include only key, important features from the raw datasets to help narrow the focus of the algorithms towards selected features and obtain the final class of the features after eliminating unwanted and non-essential features [144,146,148]. In general, feature extraction identifies the characteristic features that better represent the dataset by minimizing the within-class pattern variability while improving the between-class variability and preventing dimensionality [146,147,148,149]. However, studies have indicated the issues associated with raw data preprocessing and feature extraction inclusion in machine learning models. These issues include navigating around large datasets with many categorized features and expanding feature sets which all serve as stumbling blocks for researchers when designing models as these issues lead to long computational and processing times of models [150]. As a result, researchers refrain from disclosing to what extent data pre-processing and feature extraction were conducted for their trained models to overcome the risk of dimensionality and overfitting.
• Lack of specifications used: A large number of the papers studied do not specify the globally adopted industrial standards used in the analysis. This causes concern as having an accurate depiction of standards would aid in understanding which ones are the most effective and accurate.
• Lack of distinction in the DGA gases studies: The studied papers apply a holistic approach when describing the gases studies, with almost all of them mentioning all gases, i.e., hydrogen (H₂), methane (CH₄), ethylene (C₂H₄), ethane (C₂H₆), acetylene (C₂H₂), carbon monoxide (CO), and carbon dioxide (CO₂).

From the shortcomings and limitations identified in the literature and listed above, the instances in which the authors did not report critical elements of their study, which may have been deemed as not essential by them, compromised the analysis and understanding of readers and presented a potential gap in reporting and documentation essential information.

5. Conclusions

In this systematic literature review (SLR), a comprehensive survey of existing literature was conducted to gauge the current state of the art of oil-immersed transformers’ DGA-based fault classification using machine learning algorithms. In the assessment of published studies, a total of 124 screened papers published between 2014 and 2024 were surveyed using the developed PRISMA framework using a statistics-driven quantitative approach, the defined inclusion and exclusion criteria, and the set of key search words identified.

In addressing the research questions developed to guide this study, the SLR results show that the majority of the research conducted for transformer fault classification using DGA employs the support vector machine (32%), artificial neural network (17%), and k-Nearest Neighbor (12%) algorithms. However, the introduction of hybrid optimization algorithms by some studies suggests further developments into hybrid algorithms which could be used for binary and multi-classification projects.

The SLR also revealed that China is at the forefront of this research area with a dominant 53% of the total publications on transformer fault diagnosis and classification based on DGA using machine learning algorithms. Furthermore, it reveals that the majority of the research conducted revolves around fault diagnosis with an emphasis on improving the accuracy of known and widely used techniques such as SVM and ANN. At the same time, limited effort is put into other key algorithm metrics such as precision, Mean Squared Error, and R-Squared. Also, the current works surveyed do not explore regularization techniques for preventing overfitting and underfitting of the proposed models. In addition to these findings, a vast majority of the studies omit the description of the type of data pre-procession method, feature selection, and feature extraction methods used; this presents a gap in determining the most efficient and adaptable method that other researchers can use for their respective projects.

Additionally, based on the results obtained from this systematic literature review, there is a need for further developments in hybrid machine learning algorithms that could be developed in conjunction with the conventional fault classification methods to find the most optimum approach to classify faults at a high accuracy level.

6. Registration and Protocol

6.1. Registration Information

This systematic review, titled “Dissolved Gas Analysis-Based Fault Classification in Power Transformers: A Systematic Literature Review”, is registered with the Open Science Framework (OSF). The registration was completed on 12 November 2024, under the https://osf.io/7hbc3/ (accessed on 12 November 2024) [151].

6.2. Protocol Accessibility

The protocol for this review is publicly accessible through the Open Science Framework (OSF) platform. Interested parties can access the full protocol and registration details via the associated project link: OSF Project.

6.3. Protocol Amendments

No significant amendments were made to the original registered protocol. Any minor changes or clarifications that occurred during the review process have been documented within the manuscript in the relevant sections to maintain transparency and ensure alignment with the registered methodology.