Artificial neural networks based incipient fault diagnosis for power transformers

Abstract

Dissolved gas analysis is a most prevailing and effective condition monitoring tool now a days, used for the detection of incipient faults in the oil-immersed transformer in service, as it is highly sensitive to small amount of thermal and electrical faults. In this work, a system based on condition diagnosis is developed to combine two classical techniques for Dissolved gas analysis (DGA) assessment viz. Rogers ratios and Doernenburg ratios to form a third method i.e., combination of Doernenburg and Rogers ratios in order to have most appropriate method of DGA and to solve the conflict problems between the Doernenburg and Rogers ratios methods i.e., these two methods give two different output for the same input. In this study, Multi Layer Perceptron neural network (MLPNN) is used for faults classification using MATLAB platform. To test the proposed structure, training and then test the data which is taken from IEEE Electrical Insulation Magazine, IEC publication 60599. The other advantage of this approach is that it may be utilized for an automated diagnosis and can be practically applicable for real time application of power transformer. The diagnosis accuracies of MLP neural network Classifier are compared with various other soft computing methods like Fuzzy logic (FL), K-nearest neighbor (KNN), probabilistic neural network (PNN) and Radial basic function (RBF). The results obtained from tests gives indication that the pre-processing approach developed using MLPNN can significantly improve the diagnosis accuracies for incipient faults classification of power transformer.