Accurate power transformer PD pattern recognition via its model

In this study, a transformer model is proposed to simulate the behaviour of a real transformer, under presence of different types of defects which contribute to partial discharge (PD) generation, as closely as possible. Five different types of defects (scratch on winding insulation, bubble in oil, moisture in insulation paper, very small free metal particle in transformer tank and fixed sharp metal point on transformer tank) are implemented artificially into these transformer models to investigate the resultant PD current signal magnitude and characteristics. Time-domain PD current waveforms are recorded on those transformer models which have one type of those defects. The resultant statistical PD current wave shapes and texture features are extracted from these captured PD current signals. The principal component analysis (PCA) is used to reduce the dimension of feature spaces which are required to develop the inputs for the classifier. The principal components obtained through PCA are applied to the support vector machine classifier, as an input. The classification results indicate that the extracted texture features (using grey-level covariance matrix) preserve the best characteristics for separation of the related patterns of those five defect models, accurately.