Phase-resolved partial discharge analysis of different types of electrode systems using machine learning classification

Partial discharge (PD) measurement is a diagnostic technique used for electrical insulating systems. The establishment of a highly accurate PD diagnostic technique has become necessary in recent years. Therefore, in this study, we analyzed a phase-resolved PD signal using machine learning classification for different types of experimental electrode systems. A polyethylene sheet was used as the sample, and PD was generated by applying a high AC voltage to four different types of electrodes. The number of PD pulses was counted from the raw data as preprocessing to calculate the feature value for the machine learning. The PD generation rate was defined for each phase angle section. Four types of machine learning algorithms were adopted for the classification of the electrode system: k-NN (k-nearest neighbor algorithm), logistic regression, decision tree, and random forest. The best accuracy was obtained by using the random forest algorithm (0.97), and it was found that k-NN also demonstrated good performance. The parameter dependencies were also evaluated for each algorithm. Based on the results generated by the random forest, it became clear that there were phase angle sections that were of high importance. The reason for the result was discussed from the perspective of (i) the difference due to the electrical circuit parameters (modified abc-model) and (ii) the stochastic fluctuation of PD signal.