SVM classification of contaminating particles in liquid dielectrics using higher order statistics of electrical and acoustic PD measurements

Electrical and acoustic partial discharge (PD) measurement and pattern recognition procedures are described for detecting and identifying contaminating particles in transformer mineral oils. This work introduces the use of support vector machines (SVM), a nonlinear non-parametric automatable machine learning algorithm, for the purpose of classifying the size and composition of such particles. The training and validation of acoustic and electrical PD measurement data, which are contaminated by time varying noise, are first filtered adaptively using wavelet decomposition. Statistics of a particle's impact upon collision with the walls of a tank, containing the electrode test assembly and the inter arrival time between collisions constitute the features for the SVM classifier. These statistics include higher order moments and the entropy of the estimated density function of the features. Results based on experimental training and testing data indicate that fusing of the acoustic and electric PD information at the features level provides a nearly perfect classification success rate. These observations demonstrate that, while electrical and acoustic PD data are correlated, they contain individually independent and complementary information regarding the state and condition of transformer type mineral oils.