Introduction

Diagnosing faults in power transformers is critical for ensuring their reliable and safe operation. DGA serves as a widely employed technique for diagnosing faults in transformers. The presence and concentrations of certain gases in the insulating oil of a transformer can indicate different types of faults and potential issues. Transformer faults affect the transformer oil, causing certain dissolved combustible gases (Hydrogen (H2), Mathane (CH4), Ethan (C2H6), Ethelyne (C2H4), and Acetelyne (C2H2)). These gases are used to assess the transformer’s state1. Several DGA techniques have been addressed in standards1,2, for example, the Dornenburg Technique, the Rogers three and four ratios technique, and the International Electrotechnical Committee (IEC) Standard Code (IEC 60599). Previous methods of detecting transformer flaws have depended on examining the amount of DGA present in power transformer oils as a function of mechanical, thermal, and electrical stresses3,4,5,6,7. These methods establish rules and utilize gaseous ratios based on five primary combustible gases: C2H2, C2H4, C2H6, CH4, and H2. Some interpretative techniques, like the key gas method, include carbon monoxide (CO) in conjunction with the abovementioned gases. Although non-combustible gases including carbon dioxide (CO2), nitrogen (N2), and oxygen (O2) can also be created as a result of faults, their contribution to the diagnosis of power transformer oil problems is negligible8,9,10. The pentagon graph method13and the Duval Triangle11,12 are credited with several widely used graphic techniques. In the event of a malfunction, Duval Triangle and Pentagon were utilized once the oil’s condition was examined. It is crucial to use a different technique to examine the oil’s condition to ensure the defect is present. The pentagon technique is based on the five primary gases, H2, CH4, C2H6, C2H4, and C2H2, whereas the Duval triangle relied primarily and exclusively on the concentration ratio of just three gases, CH4, C2H4, and C2H2.

Techniques for applying artificial intelligence techniques in DGA have recently been merged to generate new approaches to interpreting transformer defects. These new techniques are based on artificial neural networks3,14,15, fuzzy logic16,17,18, support vector machines19,20, and other methods can be added that contribute mainly to solving many problems for not making the right decision4,5,21,22. Moreover, a MATLAB-based fuzzy neural diagnostic system employing an Adaptive Network-Based Fuzzy Inference System (ANFIS) was presented for interpreting transformer defects23. In24, a Neural Pattern Recognition (NPR) method was utilized to predict fault types and the severity of transformer issues during different seasons. The study proposed two scenarios employing distinct data transformation methods to enhance the accuracy of the NPR method in predicting transformer fault types and severity classes. The proposed model was evaluated using 446 samples from laboratory experiments and existing literature, demonstrating higher accuracy (92.8%) than other DGA methods.

In7, the diagnostic accuracy of transformer faults was improved by implementing optimization techniques. Specifically, a Particle Swarm Optimization-Fuzzy System (PSO-FS) platform was utilized to generate new gas ratio limits. These optimized limits aimed to reduce incorrect fault diagnoses for the IEC Code and Rogers’ ratio method. In25, an efficient code matrix was developed to identify the transformer faults by adjusting the rules that mapped the gas ratio limits for each fault type. The code matrix was built by using Hybrid Grey Wolf Optimization (HGWO), which can limit the uncertainties’ effect on identifying fault types. Teaching learning-based optimization (TLBO) in8was utilized to enhance diagnostic accuracy by new gas ratio limits based on the work in4.

The artificial intelligence or optimization methods to diagnose transformer faults were very complex and required more software code development experience. Therefore, a MATLAB arrangement learner tool can be used to identify and organize transformer faults. The transformer faults accuracy of diagnostic using the support vector machine (SVM)-bat (BA) algorithm in9. BA can adjust the form penalty parameter “c” and the SVM’s conditioning parameter “l” to develop a maximum diagnostic accuracy rate.

In26, the factors affecting the effect of transforming data for transformer faults on diagnostic accuracy have been studied. Six Optimized machine learning methods were used to diagnose the transformer faults, such as decision trees, KNN, SVM, Naïve Bayes, ensemble trees, and discrimination analysis. Data transformation was used to address its effect on the diagnostic accuracy of the classifiers. The ensemble method developed the accuracy required for high diagnosing power transformer faults, which was about 90.61%.

In this work, an intelligent system that enhances the high diagnostic accuracy of the power transformer faults was built on the outputs of four DGA methods. These methods are the Code Tree 202026, IEC modified7, Rogers ratio modified7, and neural pattern recognition24. The intelligent system compared the outputs of the four methods to develop the dominant faults. 386 DGA samples were applied to the four methods to obtain the results of each one. The intelligent systems achieved an accuracy that was better than that developed by each technique, where the diagnostic accuracy for the intelligent system was 89.12%. On the other hand, the highest diagnostic accuracy of the four methods was neural pattern recognition (86.01%).

Data collection

The DGA samples were taken from the Holding Company for Electricity in the Chemical Laboratory in the Arab Republic of Egypt, recently published works, and from literature. Table 1 provides the distribution and fault types for the 386 DGA samples. Table 1 gives information about the dataset samples related to different types of transformer heat levels and discharge types. In particular, 43, 69, 115, 81, 24, and 54 are the suggested numbers of samples for partial discharge (PD), low energy discharge (D1), high energy discharge (D2), low thermal (T1), medium thermal (T2), and high thermal (T3), respectively. The distribution of training samples across various fault types (PD, D1, D2, T1, T2, T3) with their respective references was reported. The total number of samples is 386.

DGA methods

Four DGA methods were used in this work. They were the Neural Pattern Recognition (NPR)24, Modified IEC Code7, modified Rogers’ ratio method7, and code tree 202025. The 386 are applied to the four selected DGA methods. The diagnosis of each method is obtained and a comparison between them to built the proposed model.

Table 1 Training examples were distributed based on the categories of defects and their references.
Full size table

Neural pattern recognition (NPR)

In24, A neural pattern recognition (NPR) method was exploited to predict the classes and severity of transformer faults. NPR is used to admit an item using classification, feature extraction, and filtering. The first process, filtering, removes unwanted data or features. Information can easily be extracted from the remaining data to develop distinctive features. The last process is the classification that identifies the group of similar samples.

In31, how to identify directing state parameters on the fault and the Huffman tree with the principle of the neural network was used in this work. State parameters are transferred within some vector length as input to reduce dimensionality. Furthermore, the Huffman fault tree was built for the output layer to decrease the scheming task through repropagation. The intercom and cosine correlation results between state indices were obtained from the trained parameter vector method.

Modified IEC Code (MIEC)

The International Electrotechnical Commission issued IEC 60599 in 19781. Five main combustible dissolved gases were used to draw based on gas concentration through follow codes using only three ratios, as in (1):

$$\:\left[R2\:R3\:R4\right]=\left[\begin{array}{ccc}\frac{{C}_{2}{H}_{4}}{{C}_{2}{H}_{6}}&\:\frac{{C}_{2}{H}_{2}}{{C}_{2}{H}_{4}}&\:\frac{{CH}_{4}}{{H}_{2}}\end{array}\right]$$
(1)

In7, Table 2 explains the ratio limits derived from weak ratio limits and their impact on the diagnostic accuracy of transformer faults. The Particle Swarm Optimization with Fuzzy Logic System (PSO-FS) was employed to enhance diagnostic accuracy. These columns outline the current thresholds or limits defined by the conventional IEC method for each gas ratio. For example, for R2, if the ratio is less than 1.0, it falls under code 0; if it’s between 1.0 and 3.0, it falls under code 1; and if it’s greater than or equal to 3.0, it falls under code 2. The analysis of faults in transformers is determined using the IEC approach. It delineates the gas ratios under examination, the existing current limits as per conventional IEC standards, and the refined limits achieved through the PSO-FS method. By comparing the traditional limits with the enhanced ones, it becomes possible to discern potential improvements in fault diagnosis accuracy. This optimization method was used to modify the ratio limits, resulting in improved diagnostic accuracy. The updated ratio limits for Modified IEC (MIEC) are presented in Table 2based on the findings in7. For further details on the fault types of the current IEC code and the modified version, refer to Table 3, which corresponds to Table 4in7. Table 3 provides a comprehensive overview of fault types, alongside their corresponding codes derived from the conventional IEC 60599 standard and the enhanced IEC method. This comparison enables an understanding of how the refined coding system can enhance fault classification accuracy.

Table 2 Ratios, current limits and limits of enhanced PSO-FS approaches for the IEC method.
Full size table
Table 3 Current and enhanced IEC 60,599 form codes.
Full size table

Modified Rogers’ ratio method (MR4R)

Rogers’ four-ratio method is used as one of the DGA methods for the basis of four ratios that make up the five combustible gases, transformer faults are diagnosed as follows:

$$\:\left[R1\:R2\:R3\:R4\right]=\left[\begin{array}{cc}\begin{array}{ccc}\frac{{C}_{2}{H}_{6}}{{CH}_{4}}&\:\frac{{C}_{2}{H}_{4}}{{C}_{2}{H}_{6}}&\:\frac{{C}_{2}{H}_{2}}{{C}_{2}{H}_{4}}\end{array}&\:\frac{{CH}_{4}}{{H}_{2}}\end{array}\right]$$
(2)

It delivers low diagnostic accuracy of transformer faults, so PSO-FS in7 has been used to improve diagnostic accuracy. Tables 4 and 5 show the code and fault types of the current and improved PSO-FS rogers’ ratio technique. Table 4 is on the gas ratios and their associated limits, as determined by both the conventional Rogers’ method and the enhanced PSO-FS method. By contrasting the limits derived from these two approaches, one can evaluate the efficacy of the PSO-FS method in refining the diagnostic accuracy of fault analysis in transformers.

Table 4 PSO-FS Rogers’ four-ratio method for ratio limits.
Full size table

Code Tree (CTM)25

The code tree (CTM) method aims to improve the diagnostic accuracy of power transformer faults25. The new codes of the CTM method for transformer fault detections were generated based on the membership limits identified using several optimization processes. A total of 10 membership limits and fault types of 243 fuzzy rules constitute the variables of the system inputs based on five gas ratios. The1,2,3 codes were selected for each gas built on the gas limits from zero to 100%. To achieve this, the CTM method utilizes a set of ten membership limits and fault types, which are incorporated as variables for the system inputs. These variables are derived from five gas ratios, which are typically used in DGA to assess the condition of power transformer oils. In particular, the CTM method generates codes for each gas based on the zero to 100% range. It means that the CTM method explores different possible combinations of gas limits to create codes that can better identify and diagnose transformer faults accurately. By using the CTM approach and optimizing the codes based on the membership limits, it is possible to improve the accuracy of detecting different types of faults in power transformers. It can lead to more reliable and efficient fault diagnosis, enhancing maintenance and overall reliability of power transformer systems. The results of each method individual can be explained in Table 6 as follows;

The results of each DGA method are illustrated in Table 6. The diagnostic accuracies of the four methods are 81.87, 85.23, 84.97, and 86.01% for CTM, MIEC, MR4R, and NPR methods, respectively. The NPR has the highest ability to detect PD, D1, T1, and T3, but it is distinct in detecting T1. The MIEC Code develops the highest accuracy for PD, D2, T2, and T3 but is distinct from the others for D2. The MR4R has the highest accuracy for identifying T2. Therefore, we can merge these four methods into an intelligent system from the enhancement solver for the accuracy of diagnostics for the most used transformer malfunctions.

Transformer Fault diagnosis Intelligent System (TFDIS)

The intelligent diagnosis system is based on the decision-making approach constructed using the output of the four methods. Every decision-making procedure generates a final selection that may or may not encourage a decision. Decision-making is built to identify and select alternatives depending on the values and behavior of the decision-maker system. Decision-making can be judged as one of the main actions of an organization and is a significant part of any application process. The proposed transformer fault diagnosis intelligence system (TFDIS) to determine the fault appearing in the transformer is created using the decision-making principle as in Fig. 1. The TFDIS is built based on the output of each of the mentioned four methods that are based on MIEC, MR4R, CTM, and NPR methods. Each of the four methods has an output of one of the six transformer fault types of PD, T1, T2, T3, D1, or D2.

Table 5 The new codes and fault types of the improved Rogers’ four-ratio technique.
Full size table
Table 6 The results of each DGA method.
Full size table

Figure 1 shows the flowchart for the suggested TFDIS. Five dissolved gases in ppm of the transformer oil are used as inputs for the four diagnosis methods. The five dissolved gases (H2, C2H6, C2H2, CH4, and C2H4). The different ratios for each diagnosis method are estimated and considered inputs to each. The output from each method is obtained and collected as inputs to the TFDIS, as depicted in Fig. 1. The decision process follows to combine the decision from the outputs of the four methods. The TFDIS model will operate as follows; Regarding the outputs of the four methods, if more than three outputs have the same fault type diagnosis, the TFDIS decision will be one of the four or three methods. Otherwise, if two outputs have the same diagnosis of the other two outputs, i.e. two are D1 and two are T1, the TFDIS decision will be like that of the MR4RM (It refers to the dangerous fault, i.e. if two outputs are D1 and the other two are T1, then the diagnose is D1). If only two outputs are similar, then the fault diagnosis is the same for the two outputs, i.e., if two outputs are D1 and the other two outputs are different (one PD and one T1), so the diagnosis is D1). Finally, the TFDIS decision will be like the NPR method’s diagnosis.

Table 7 compares the proposed TFDIS model, the four mentioned methods, and the original IEC and Rogers’ four ratios (Rog4) methods based on the 386 dataset samples. Compared to other methods, a good diagnosis of the proposed PD, D2, T2 and T3 model and overall accuracy. The overall accuracy is observed at 44.82%, 50.26%, 81.87%, 84.97%, 85.23%, 86.01% and 89.12% for Rog4, IEC, CTM, MR4R, MIEC, NPR, and the proposed TFDIS, respectively; therefore, a significant enhancement will be obtained when the SSFTFDE is applied to detect the transformer fault types based on DGA approach. The results in Table 7 show that the proposed TFDIS model outperformed all the other methods regarding overall accuracy. The TFDIS model achieved an accuracy of 89.12%, significantly higher than the conventional IEC and Rog4 methods and the four individual mentioned methods (PD, D2, T2, and T3). The significant enhancement in accuracy when applying the TFDIS model to detect transformer fault types based on the DGA approach indicates that the TFDIS model has the potential to be a highly effective and reliable tool for transformer fault diagnosis. It shows promise in improving the accuracy and efficiency of transformer fault detection, leading to better maintenance and enhanced overall reliability of power transformers. These results highlight the potential of the TFDIS model as an effective and reliable tool for transformer fault diagnosis based on DGA approach. The significant enhancement in accuracy suggests that the TFDIS model could contribute to improving the overall reliability and efficiency of transformer fault detection, ultimately leading to better maintenance practices and enhanced reliability of power transformers.

Figure 2 compares the proposed TFDIS model, the four mentioned methods, and the original IEC and Rogers’ 4 ratio (Rog4) methods based on the 386 dataset samples. Compared to other methods, a good diagnosis of the proposed PD, D2, T2, and T3 model and overall accuracy. A good diagnosis performance of the proposed TFDIS model compared to the individual four methods and the conventional methods for detecting the transformer fault types based on the DGA approach. The evaluation was based on a dataset of 386 samples related to transformer faults and their respective DGA (DGA) results. The main objective was to assess the accuracy of fault diagnosis provided by each method. According to the results presented in Fig. 2, the proposed TFDIS model exhibited a good diagnostic performance compared to the individual four methods (PD, D2, T2, and T3) and the conventional IEC and Rogers’ 4 ratio methods. The TFDIS model outperformed these methods in accurately identifying and diagnosing various transformer faults based on the DGA approach. Overall, the TFDIS model demonstrated better accuracy and effectiveness in detecting transformer fault types when compared to the other methods, which indicates its potential as an advanced and reliable approach for fault diagnosis in power transformers using DGA data.

Table 8 presents 9 practical samples to discuss the proposed TFDIS model diagnosis. For example, in sample No. 1, the diagnosis of all four methods is PD so the diagnosis of the proposed TFDIS model is like that which is PD. In sample No. 2, the diagnosis of the three methods is PD so the diagnosis of the proposed TFDIS model is like to them, which is PD. In sample 3, only two methods give similar diagnoses (D1) while the other two methods are different, so, according to the flowchart of the proposed TFDIS model, the diagnosis is similar to the two methods (D1). Sample 4, each two methods gives identical diagnoses. Then, the proposed TFDIS model gives a diagnosis like the MIEC model (D2).

The authors add two new artificial techniques based on DGA data: the support vector machine (SVM) and the K-neighbors method (KNN) to investigate its performance and effect on the accuracy of TFDIS. After applying the SVM and KNN DGA models with the four methods selected in the suggested TFDIS Model, the TFDIS’s accuracies were addressed. As in Tables 7 and 9, the overall TFDIS’s accuracy was reduced from 89.12 for the four selected methods to 88.6 after adding the SVM and KNN DGA methods. So that we only apply the suggested model with the four methods.

Fig. 1
Fig. 1
Full size image

The proposed intelligent system for transformer fault diagnosis enhancement (TFDIS).

Table 7 TFDIS accuracy with the selected four methods.
Full size table
Table 8 Practical case studies.
Full size table
Fig. 2
Fig. 2
Full size image

Comparison among the fault diagnoses based on TFDIS and other methods as in Table 7.

Table 9 TFDIS accuracy after adding SVM and KNN.
Full size table

Conclusions

This work focused on enhancing the transformer fault diagnosis based on DGA methods. The traditional DGA methods have poor diagnostic accuracies, and the new recent DGA techniques did not develop satisfactory accuracies. Therefore, this work proposed an intelligent system to slightly improve the diagnostic accuracy of DGA methods using a decision-making system (TFDIS). The system was built by combining four DGA methods and diagnosing the transformer fault based on the dominant fault from the outputs of the four DGA methods. The TFDIS developed the highest accuracy (89.12%), which is higher than the other DGA method individually, where the highest accuracy can be produced by individual NPR (86.01%). Also, the study emphasizes how intelligent systems might solve persistent problems in transformer malfunction identification in a revolutionary way. The effectiveness of TFDIS as a useful instrument for improving diagnostic accuracy is demonstrated by its success, which also contributes to better maintenance procedures and increased power distribution system reliability.