DGA Method Implementation for Incipient Fault Analysis using Gas Concentrations

: Power transformers are essential devices for the durable and reliable performance of an electrical system. the main objective of this study is to analyze three classical diagnosis techniques to identify incipient faults in Transformer oil using Rogers’s Ratio Method, Doernenburg Ratio Method, and ANN which is a type of artificial intelligence learning method. Implementation of the system in MATLAB software for each diagnosis method and compare their accuracy and efficiency and hence design three diagnosis methods of DGA for condition assessment of Power Transformer. And the analysis on the MATLAB software shall be carried so as to detect the best method for detection of a certain type of fault and the best suited method for overall fault analysis for a certain data sets out of the three methods. This technique utilizes the learning capacity of that artificial neural network has been shown to be more efficient in detecting different mistakes. The overall error detection accuracy of such gas neural network study was found to be 73.8 percent.


INTRODUCTION
Power transformers are essential devices for the durable and reliable performance of an electrical system. Their continuity is therefore the daily activity of the energy suppliers. Consequently, their proper maintenance based on detected incipient failures and/or degraded conditions is essential to achieve this goal. When failures occur in a transformer, planning for subsequent maintenance is of the utmost importance; otherwise, malfunctions will occur which could lead to system failure.
Even under typical operating circumstances, a lengthy transformer produces gas. Nevertheless, it is subjected to electronics, physical, biochemical, and thermally pressures on a routine basis throughout time, resulting in a significant degree of gas generation in the transformers insulating system [1]. The gas concentration, on either hand, will rise if there is any abnormality. Whenever oil decomposes, gases such as hydrogen (H 2 ), methane (CH 4 ), acetylene (C 2 H 2 ), ethylene (C 2 H 4 ), and ethane (C 2 H 6 ) are generated, while when cellulose breaks down, methane (CH4) and hydrogen (H 2 ) are formed. Carbon monoxide (CO) and carbon dioxide (CO 2 ) are two gases that may be found in the atmosphere (CO 2 ). Carbon monoxide (CO) and carbon dioxide (CO 2 ) imply problem caused by decay of the article, ethylene (C 2 H 4 ) and ethane (C 2 H 6 ) indicate an increase in fuel temperature, incomplete discharges (which supply hydrogen (H 2 ) and methane (CH 4 ) at low energy), and the arc can be confirmed by the release of acetylene (C 2 H 2 ) and hydrogen (H 2 ) [ approach is based on the concentration of dissolved gases in the insulating oil and the specific limits of dissolved gas ratios. Some gas ratio limits and rules have been developed to specify transformer failures. These gas rules and ratios are assumed to be based on five main fuel gases: hydrogen (H 2 ), methane (CH 4 ), ethane (C 2 H 6 ), ethylene (C 2 H 4 ) and acetylene (C 2 H 2 ). In some interpretation techniques, carbon monoxide (CO) has been added to these gases. Some non-flammable gases such as carbon dioxide (CO 2 ), nitrogen (N 2 ) and oxygen (O 2 ) have been developed in the event of a malfunction, but have no influence on fault diagnosis. After the extraction of the dissolved gases from the mineral oil, the gas concentration For volume Several methods have been proposed to find dissolved gases, but among them the direct injection technique and the headspace method are the most accurate methods and directly provides the dissolved gas volume by injecting mineral oil, by entering the gas concentration in ppm (parts per million) for each diagnostic method there is the type of incipient error, the next step is to compare all types of incipient errors after each method and finally the conclusion will give the final unit error ( mineral oil) according to the percentage given by each method.
The main reason for combining three methods is: "None of the methods provide all kinds of examples. Reporting processes have high thermal, partial, and arc discharges, but no combination of thermal and electrical failures.
The severity of the transformer fault has been divided into three types based on the level of danger it presents, namely low, high and medium severity of the fault. The contribution of the total dissolved burnt gas during the analysis was immense when examining the state of the transformer. These two states (1) gravity, (2) total dissolved combustible gases (TDCG) were used throughout the study. For example, if the rate of change of TDCG> 30 ppm, this is considered a high severity thermal failure (T3). To achieve maximum inclusion in the work, the framework uses 9 gas concentration ratios, namely H 2 , CH 4 , C 2 H 6 , C 2 H 2 , C 2 H 4 , CO, CO 2 , N 2 and O 2 . When parsing different methods, any entries with a missing value are represented by -1 during parsing. The program takes the inputs required for each method and ignores the unnecessary ones. The analysis uses the categorization of the results of the whole method into six main failures, listed in Table 1. Low energy discharges D2 High energy discharges T1 Low Thermal faults not exceeding 300 0 C temperature T2 Medium Thermal faults between 300 0 C to 700 0 C temperature T3 High Thermal faults above 700 0 C temperature

A. Roger's Ratio Method
The Rogers Ratio method diagnoses failures by taking the ranges of the gas ratios in mineral oil; Four gas ratios are required: C 2 H 2 / C 2 H 4 , C 2 H 6 / CH 4 , C 2 H 4 /C 2 H 6 and CH 4 /H 2 . Using these four gas ratios, diagnose faults such as normal aging, winding circuit currents, electrical faults, low energy discharge faults, and thermal faults in various areas (1500c-7000c) [9]. The main advantage of gas ratio analysis is "it is independent of the amount of oil involved and depends only on the proportions of the gases involved" In this technique, four relationships are identified to diagnose the types of transformer failures. These relationships are structured as follows according to five main gases:    9) and (see 10).
The error types and error codes of the Rogers Four Ratio method are also shown in Table 1. The states (see 4), (see 5), and (see 6) have been combined to refer to a low thermal error (T1). , the states (p.7) and (p.8) have been merged to refer to the mean thermal error (T2) and the state (p.11) refers to the high energy discharge error (D2), state (s 3) refers to the PD error, the states (s.10), s.12) and (s.13) refer to the low energy discharge D1 and the state (s.9) refers to the high thermal error T3. In the Duval triangle method, the electrical-thermal error (DT) is divided between the errors D1 and D2 so that a comparison can be made with the other methods.

B. Doernenburg Ratio Method
This technique uses four different gas ratios such as C2H6 / C2H2, CH4 / H2, C2H2 / CH4 and C2H2 / C2H4. Based on these gas velocity ranges, it diagnoses various fault conditions such as partial discharges, arcing and thermal failures with varying degrees of severity. This method is one of the oldest methods for detecting initial failures in transformers. To use this procedure, the first condition to be met is that at least one of the key gas concentrations (H 2 ; CH 4 ; C 2 H 6 ; C 2 H 4 ; C 2 H 2 ) exceeds twice the concentration limits (L1) indicated in Table 4.3 and that one of the other two gases exceed the limit value L1. Once this condition is met, the four typical gas ratios (CH 4

C. ANN
ANNs are information processing systems that simulate human behavior. The ANNs obtain information about the characteristics considered and learn from the input data, even if our model contains noise. The structure of the ANN consists of essential information processing units, which are neurons.
They are divided into several layers and linked together by defining weights. The synaptic weights show the interaction between each pair of neurons. These structures distribute information across neurons. The mapping of the input and the estimated responses of the output are calculated by combining several transfer functions. We can use the self-adaptive information shape recognition method to analyze the learning algorithms of artificial neural networks. The most commonly used calculation algorithm is the backward error propagation algorithm.
Neural networks can be divided into single-layer perception networks and multilayer perception networks (MLPs). The multilayer perceptual network comprises several layers of simple two-state sigmoid transfer functions with processing neurons interacting by applying weighted connections. A typical multilayer neural network consists of the input layer, the output layer, and the hidden layer. Multilayer perception (MLP) with the back-propagation learning algorithm is used in this study because many previous researchers have used this type of ANN and it is also a general function approximator. The study of the connection among both input factors with output variables formed the foundation for the creation of ANN models. As illustrated in Fig. 2, the neural architecture consisting of 3 or even more layers, namely the H. Input layer with output layer, and then hidden layer. The following is a description of that network's function: f(.) here is transfer function, and Yj means output of that node j. Xij here the input signal through node i at the lower layer to this node j, While wij is just the weight of that link between node j with node i at the lower layer.
In DGA-based intelligent mistake detection, several researchers utilized neural network ideas. Prediction, segmentation, regression, and time series prediction these are all common applications to multiple layer of perceptron neural networks. In these neural networks, the training algorithm is much essential. The training algorithm changes the distortions as well as weight so this input-output mapping takes as little time as possible. That back propagation algorithm (BPN) was an earliest learning algorithm, also it is used to under supervision learning of a multiple layered neural-network with prediction. This same back-propagation of output-level unit error determines hidden layers unit errors, which is also the fundamental concept underlying the word back-propagation. 2 Back-propagation is used in a variety of learning methods. A steepest descent algorithm, often known as gradient descent algorithm, is perhaps the most widely used. The goal is to make that variation in weight proportional to a negative of derivatives of the observed error for every weight of the present model: The gradient descent technique, also the gradient descent method using adaptive learning method, as well as the Levenberg-Marquardt method are all topology of that back-propagation learning algorithm described in this paper for the detection of VOL.7, ISSUE 10, OCTOBER 2021 www.ijoscience.com 5 transformer faults. Every arrangement takes as input concentration ratio (ppm) of following gases (Acetylene, Ethylene, Ethane, Hydrogen and Methane, i.e. R1 (C2H2-Acetylene/ C2H4-Ethylene, R2 (CH4-Methane / H2-Hydrogen) and R3 (C2H4-Ethylene / C2H6-Ethane), and incorporates all kinds of mistakes within system learning variables. This chapter discusses the interpretation of three techniques for two designed data set values. The dataset is first selected for the preventive implementation of the techniques and its performance is examined with limited values of the available gas concentrations. Subsequently, the study was extended to a greater amount of data, then the effectiveness of the two methods was examined and compared with that of the learning method, among all the classic diagnostic techniques, the Rogers ratio method, the Doernenburg ratio. It also describes the representation of the zone for all types of errors and their software implementation in MATLAB.

A. Rogers ratio method for minimum data (Data Set-1):
The Rogers ratio method compares the quantities of different key gases by dividing them against each other. This gives a ratio of the amount of one key gas to another. Report methods are only valid when there is a significant amount of gas used in the report.  The results of the Rogers ratio method were examined and then compared to the actual error, after which the accuracy of the Rogers ratio method for each error determination is examined and plotted on a graph. The analysis showed that the Rogers ratio method was best suited for detecting fault type D2, which is a high thermal fault in the transformer, with a single input phrase for analysis. Finally, the efficiency of the method for determining all types of errors for the dataset was found to be 16.66%. Using gas velocity ranges, it diagnoses various fault conditions such as partial discharges, arcing and thermal failures of varying degrees of severity. The method is first used to process less gas data than input into the Matlab script to examine its performance under this condition.   The analysis of the results shows all relevant gas key number ranges and their code and mineral oil status. Table 9 Types of fault prediction using roger's ratio method in matlab script for Data Set-2 AC T P D The results of the Rogers ratio method when used in MATLAB demonstrated its percentage accuracy in determining various types of errors. The dataset with 42 samples with the Roger ratio method of the DGA analysis was used for dissolved gas analysis. The Rogers ratio results were reviewed and then compared to the actual error, after which the accuracy of the Rogers ratio technique for each error determination is examined and graphed. Analysis showed that Roger's ratio method efficiency in detecting T1 and T3 types was maximal and was 62.5%. The efficiency of this method for determining the type of T2 error was also 33.3%. Finally, the effectiveness and efficiency of the Rogers ratio method in determining all types of errors for the entire data set was found to be 52.38%.

D. Doernenburg Ratio implementation with data set-2
Using gas velocity ranges, it diagnoses various fault conditions such as partial discharges, arcing and thermal failures of varying degrees of severity. Therefore, by implementing this method in MATLAB, incipient errors can be diagnosed in a window. The method is considered for implementation in the analysis of the selected data sets, so its accuracy is calculated in the determination of various errors. The comparative study was carried out in the book. The analysis was also more effective in formulating various types of error hypotheses.  The results of the Doernenburg ratio method were examined and then compared to the actual error, after which the accuracy of the Doernenburg ratio method for each error determination is examined and graphed. The analysis showed that the efficiency of the Doernenburg ratio method in recognizing T1 and T3 types was maximal and was 87.5%. The efficiency of this method for determining the type of D1 error was also the lowest at 33.33%. Finally, the effectiveness and efficiency of the Doernenburg ratio method in determining all types of errors for the entire data set was found to be 69.04%.   The results of the ANN method were examined and then compared to the actual error, after which the accuracy of the ANN method technique is examined and graphed for each error determination. The analysis showed that the efficiency of the ANN method in detecting type D2 was maximal and was 91.6%. The efficiency of this method for determining the type of T2 error was also lower with 33.33%. Finally, the effectiveness and efficiency of the Doernenburg ratio method in determining all types of errors for the entire data set was found to be 73.80%.  This technique utilizes the learning capacity of that artificial neural network has been shown to be more efficient in detecting different mistakes. The overall error detection accuracy of such gas neural network study was found to be 73.8 percent.

Fault detection Methods
Overall efficiency %