MLP neural network-based decision for power transformers fault diagnosis using an improved combination of Rogers and Doernenburg ratios DGA

Abstract

Dissolved gas analysis (DGA) is a widely-used method to detect the power transformer faults, because of its high sensitivity to small amount of electrical faults. The DGA is exploited for fault classification tools implementation using the artificial intelligence techniques. In this study, we use the Rogers ratios, the Doernenburg ratios methods and our proposed combination of Rogers and Doernenburg ratios DGA methods as gas signature. The multi-layer perceptron neural network (MLPNN) is applied for decision making. The paper presents a comparative study on one hand for the choice the most appropriate DGA method and to resolve the problem of conflict between the Rogers and Doernenburg ratios methods. On the other hand, it compares the various MLP architectures by comparing two output data types and three hidden layer types with the aim to establish the most appropriate MLP model. Before testing, the proposed structures are trained and tested by the experimental data from Tunisian Company of Electricity and Gas (STEG). The test results suggest that MLPNN ratios combination can generalize better than other MLPNN models. The approach has the advantages of high accuracy. The other advantage is that the model is practically applicable and may be utilized for an automated power transformer diagnosis. The classification accuracies of the MLPNN classifier are compared with fuzzy logic (FL), radial basis function (RBF), K-nearest neighbor (KNN) and probabilistic neural network (PNN) classifiers. The test results indicate that the developed preprocessing approach can significantly improve the diagnosis accuracies for power transformer fault classification.