The Support Vector Machine (SVM) has been adopted here to identify between four different Partial Discharge (PD) sources that can affect the insulation system of ac rotating machines. Some Roebel bars were prepared to obtain bar-to-finger, corona and slot PD in addition to distributed micro-voids. PD measurements were performed using different set-up conditions, defect location and voltage levels. The SVM was trained to differentiate between the inherent features (global and derived parameters) of each discharge source. In order to achieve the best accuracy, SVM characteristic parameters were optimized by removing features that can be affected by the measurement conditions. A cross validation has been used to obtain the highest testing accuracy. Moreover, results obtained using raw data and scaled parameters, are also compared to obtain the best performances in identification of the given defect typologies.
