Multilevel SVM and AI based Transformer Fault Diagnosis using the DGA Data

1Research Scholar, JNTUH & Assistant Professor, Dept. of Electrical and Electronics Engineering, Matrusri Engineering College, Saidabad, Hyderabad, India. 2Deputy Executive Engineer & JNTUH Univerity, Kukatpally, Hyderabad, India. 3Additional Director, UHV Research Laboratory, CPRI, Hyderabad, India. 4Associate Professor, Dept. of Electrical and Electronics Engineering, Matrusri Engineering College, Saidabad, Hyderabad, India. 1music.sarma2016@gmail.com, 2bumanapalli_brreddy@yahoo.co.in, 3pmnirgude@cpri.in, 4pvdnaidu81@gmail.com


Introduction
In power systems, transformers play a critical role. Any malfunction in the transformer could result in an outage and, as a result, a disruption in service. It's critical to spot a transformer issue early on. The key-gas ratios C2H2/C2H4, CH4/H2, and C2H4/C2H6 have a strong relationship with transformer failures. The ability to forecast these key-gas ratios in transformer oil is critical for detecting and identifying transformer incipient breakdowns early. Because of its nonlinearity and the limited amount Open Access

Journal of Informatics Electrical and Electronics
Engineering (JIEEE) A2Z Journals of training data, forecasting key-gas ratios in transformer oil is a difficult challenge. The use of software-based diagnosis tools to monitor the condition of transformers is critical, and DGA is considered an excellent test for detecting incipient transformer defects. The examination of certain dissolved gas concentrations in a transformer's insulating oil provides information about the state of the transformer, allowing required preventive steps to be taken. Due to the variability of gas data, standard approaches fail to identify the faulty condition, making it a difficult and time-consuming task. The key gas analysis, Rogers Ratio technique, IEC gas ratio code (IEC-60599), Doernenburg Ratio method, and Duval triangle method are all traditional ways to diagnose a transformer issue utilizing the DGA method. These diagnostic approaches do not provide much information regarding incipient defects, which can lead to a state of indecision. The support vector machine with genetic algorithm (SVMG) developed by Sheng-wei Fei et al [1] is applied to fault diagnosis of a transformer, in which the genetic algorithm (GA) is used to select appropriate free parameters of SVM, and the experimental results showed that the SVMG method provided higher diagnostic accuracy. According to A. Akbari [2] et al, condition-monitoring and software-based diagnosis tools are effective maintenance management strategies for transformers. Agent-based systems have been developed for complex systems with relatively simple individual agents, and Multiagent systems have been used to overcome complexities. An intelligent fault classification approach to power transformer dissolved gas analysis (DGA) was given by A. Shintemirov et al [3] and Y.-j. Sun et al [4], which dealt with highly varied or noise-corrupted data. To increase the interpretation accuracy for DGA of power transformers, bootstrap, and genetic programming (GP) are used. For fault classification, the features retrieved by GP are fed into artificial neural networks (ANN), support vector machines (SVM), and K-nearest neighbor (KNN) classifiers. The combined classification accuracies are compared and determined to be good. Tang [5], [6] introduced a Parzen-Windows (PW)-based transformer failure diagnostic classifier that uses a probabilistic scheme to interpret transformer dissolved gas analysis (DGA). To increase fault classification accuracies, a global optimizer called particle swarm optimizer (PSO) is used to tune the parameters of PW. with genetic algorithm (SVMG) to forecast key-gas ratios in power transformer oil, and genetic algorithm (GA) is used to determine free parameters of support vector machine. The SVMG model is shown to be a proper alternative for forecasting key-gas ratios in power transformer oil. R.N. Afiqah [11,12] stated that "Detection of new faults in the transformer early is very important to prevent accidents and to reduce related material losses," and "Conventional methods used in the implementation of DGA are improved by using intelligent systems." In this study, the IEC ratio method, one of the classical methods is used, and an intelligent fuzzy-based analysis is performed using MATLAB. The remainder of this paper is structured as follows: Section 2 presents DGA fundamentals. Section 3 explains concepts of SVM and its mathematics, section 4 details transformer faults classification based on SVM, section 5 gives results and discussions and finally Section 6 describes conclusions.

DGA Fundamentals
Because transformer equipment is so expensive, it must be properly monitored while in use. DGA is first used to monitor the status of a specific transformer and has since received widespread approval among professionals. DGA can be used to determine the most likely situation inside transformers, provide early warnings and diagnoses, and boost the chance of acting appropriately, similar to how doctors check a human body with a stethoscope. DGA is a sensitive and reliable technology for locating problems in power transformers [20][21][22][23][24][25]. It is possible to distinguish fault in a wide range of oil-filled equipment using  The ppm concentration values range in the transformers according to IEC 60599 are given in Table.2.  The approach of graphical depiction utilizing Duval's triangle is discussed as below.  [26].
Journal of Informatics Electrical and Electronics Engineering (JIEEE) A2Z Journals

Concepts of SVM and its Mathematics
The Support Vector Machine (SVM) is a supervised machine learning model that learns how to divide various groups by establishing decision boundaries and is used for classifications. It is used to provide an ideal separating hyper-plane solution for linearly separable and non-linearly separable datasets by maximizing the margin between the separating data. Hyperplanes are decision boundaries that aid in data classification. Different classifications can be given to data points on either side of the hyperplane as shown in the figure 2. The hyperplane's dimension is also determined by the number of features. If there are only two input characteristics, the hyperplane is merely a line. The hyperplane becomes a two-dimensional plane when the number of input features reaches three.
where k X signifies the input vector, k Y ranges -1 to 1 denotes the associated output values, and m denotes the total number of data patterns, and the regression approximation estimates a function based on a given data set. It starts with data separated by a hyperplane and then uses the kernel method to extend to non-linear decision boundaries.  .

Type of Fault Zone Percentage of gases
Where w denotes the weight vector and b denotes the bias term. w and b are used to define the position of the separating hyperplane by which it should satisfy the constraints.
( ) using the Lagrangian principle, the above equation is transformed to where αk are the Lagrange coefficients ( Then the condition of optimality is applied.
We have the following equation along with the constraint: Hence the dual problem is expressed as the following equations 1 max . .
. for all k and 0 2 The ranking function Where C is the margin parameter.
A kernel is a method of placing a two-dimensional plane into a higher dimensional space, so that it is curved in the higher dimensional space. In a simple way a kernel is a function from the low dimensional space into a higher dimensional space. A Kernel Trick is a simple method where a non-Linear data is projected onto a higher dimension space so as to make it easier to

Transformer Faults Classification based on SVM
DGA training and testing data: It is based on the transformer oil's dissolved gas content data. The training data set and the testing data set are separated from the rest of the data. The training data set is assessed using several DGA methods, with judgments assigned to seven classes i.e normal, partial discharge, low energy discharge, high energy discharge, low temperature overheating, middle temperature overheating, and high temperature overheating. The transformer ratings and characteristics used for the taking of mineral oil samples are given in the with each layer's SVM being a binary classifier. The multilayer SVM is a "one-against-all" type of model. If the binary SVM in the preceding layer erroneously identifies the condition, the other SVMs will categorize the condition incorrectly, and the final classification result will be incorrect. It's difficult to detect less severe issues inside transformers with limited data and known conditions, thus it's always necessary to thoroughly inspect and evaluate conditions inside transformers for alarming or failureprone conditions.  . When SVM1 receives a sample representing the normal condition as input, output is set to +1; otherwise, output is set to 1. The data is then passed to SVM2, which has been trained to distinguish between the discharge and overheating faults.

The Functional Concept of Multi-Level SVM Classifier:
The output of SVM2 is set to +1 when the input of SVM2 is a discharge fault; otherwise, it is set to 1. This information is passed to SVM3, which has been trained to distinguish between high-energy discharge (D2), partial discharge (PD), and low-energy discharge (D1) faults. The output of SVM3 is set to +1 when the input of SVM3 is a D2 fault; otherwise, it is set to 1. SVM4 is trained to distinguish between high temperature overheating (T3) and low and moderate temperature overheating (T1 and T2). When the input of SVM4 is a sample representing the T3 fault, the output of SVM5 is set to +1; otherwise −1. Then the it is fed to SVM5 is trained to separate the middle temperature overheating (T2) fault the low temperature overheating (T1) fault and when the input of SVM5 is a sample representing the T2 fault, the output of SVM5 is set to +1; otherwise −1. Then it is fed to SVM6 which is trained to separate the partial discharge (PD) fault from the low energy discharge (D1) fault. The output of SVM6 is set to +1 when the input of SVM6 is a sample indicating the D1 fault; otherwise, it is set to 1. The parameters of the The Algorithm for the multi-layer SVM algorithm: • Initialize output SVM   Data generation: The data is generated by considering both a normal and a failed state mode.
• Load the 192-sample training data as training data.
• Polynomial and Gaussian kernel functions, with Kernel=1 and C=1 • The SVM training data set with the SVM class function.
• Display of the results as well as the most important support vectors.
• Determine the value of the hyper-parameter • Using a logarithmic scale spanning from 1 to 1000, determine the C parameter.
• The cross-validation approach is used to select the C and parameters.
• Using the suitable values for these parameters, the SVMs are trained and tested.

Results and Discussions
The classification of SVM errors is done using DGA algorithms as a gas signature. Statistical diagnosis results of the ratio methods is shown given table 7.

SVM key gas
The defects are classified using the gas as input data. The words false alarm and non-detection rate are used to examine it.
Polynomial and Gaussian kernels are put to the test, shown in table.8.

TABLE 8. The matrix condition output of SVMS
The Gaussian kernel function is more efficient for system problem diagnostics, however it does not produce outstanding results. As a result, another method, SVM ratios, is proposed. The flaws in this method are dependent on the ratios that serve as input data. The test data is given back into the SVM to see if it correctly classifies and return results. For the two kernel functions polynomial and Gaussian kernels, the table.9 shows the results of the same false alarm rate and non-detection rate.  The Gaussian kernel function is found to be a more efficient approach for defect diagnostics.

Graphical representation of SVM
The defects are categorized using the same Polynomial and Gaussian kernel functions as input and the same graphical representation. The classification performance of the SVM graphical representation method is shown in the table.10.

Analysis on SVM Classification
The Gaussian kernel appears to be producing good results based on the four inputs of the data classified by the SVM. For the gas analysis approach, the false alarm rate and non-detection rate of four input data types are used. We analyse the false alarm rate and non-detection rate of four input data types to determine the most significant gas analysis method. The real result shows that the classification accuracies obtained by combining ratios and graphical representation methods are higher than those produced by SVM for gas signature classification. The performance of the multilayer SVM is compared to AI methods such as fuzzy logic and MLP. Though the MLP has the advantage of a quick learning process and no iteration for updating weights, it requires a significant quantity of training data and requires adjusting the hidden activation function's parameters.
The Fuzzy logic method requires linguistic variables, membership functions for each gas signature with "low," "medium," and "high" descriptions, and an inference rule basis. Finally, the SVM technique is used to classify errors using a combination of ratios and graphical representation. Gaussian, trapezoidal, and triangle functions are the most utilized membership functions It can be observed that the IEC ratio method gives correct diagnosis and finds No Fault (NF) cases but T1 fault cases of cannot be detected. The other two methods Doernenburg ratio and the Roger ratio methods are able to detect the T1 cases. Out of the Doernenburg ratio and the Roger ratio methods, the Roger ratio method finds good number of discharge faults, and the Doernenburg ratio method is the best to detect the thermal faults.

Conclusion
The MLSVM technique is used and compared with basic SVM and conventional IEC methods in this paper for fault classification in transformers utilizing dissolved gas measurements. The key gas, ratios, graphical representation, and combination ratios and graphical representation approach are the DGA methodologies investigated. Polynomial and Gaussian functions are used to examine the effectiveness of SVM diagnosis. The real data sets are utilized to test the DGA algorithms' capability in forecasting transformer oil. According to test results, the combination, ratios and graphical representation approach is more suitable as a gas signature and the MLSVM with the Gaussian function performs better in diagnosis accuracy than the other kernel functions.
Due to their extensive study capabilities, the accuracy of multi-layer SVM for fault identification is comparable to that of conventional approaches. It can be observed that conventional ratio methods are not able to detect the internal faults and hence given Not Detected (ND) and some cases which cannot be diagnosed are mentioned as No Fault (NF) cases. It is observed that the ratios methods are inferior to the MLSVM method. The MLSVM outperforms previous AI approaches when it comes to fault diagnostics. The MLSVM method can be used to diagnose incipient defects in transformers in real time. The results show that the MLSVM method has the ability to anticipate the DGA method in transformer oil. The SVM method has a greater precision but a poorer recall over positive classes, whereas the MLP approach has a better classification performance, correctly predicting two of the four failures that happened within the chosen period. Future work will focus on improving the overall performance metric by incorporating a more robust data set as well as a different set of features to improve and lessen the potential for bias in the results. It will be highly promising in the future to develop new intelligent comprehensive fault diagnostic systems by adding new ML theories and frameworks, as well as a new ML based on multi-layer ANN to transformer fault diagnosis based on DGA. In some circumstances, such systems can detect and discard bad data automatically, and they have better real-time capabilities and self-adaptation. Doernenburg  Rogers  IEC  ANN  SVM  MLSVM   1  T1  T1&T2  ND  ND  NF  T2  T1   2  T1  T1&T2  T1  T1  T1  T1  T1   3  T1  T1&T2  T1  ND  T1  NF  T1   4  T2  T1&T2  T2  T2  T2  T2  T2   5  T2  T1&T2  ND  ND  D2  T2