Comparative performance analysis of machine learning algorithms in partial discharge source classification

Insulation failure is one of the most critical faults in electric switchgear. One of the foremost factors resulting in insulation degradation is partial discharge (PD). There can be multiple sources of PD in the switchgear and based on the type of defect, the characteristic patterns of the resultant PD may vary. Machine learning-based techniques are well suited for the analysis and classification of these complex patterns. In the present work, five different types of discharges such as corona discharge, floating electrode discharge, free-moving particle discharge, internal discharge, and surface discharge are considered. With the advent of data acquisition systems, it is possible to acquire multiple features in a PD signal. The classification of PD defects helps in time-bound defect-specific maintenance of switchgear. The performance of four supervised machine learning algorithms, namely, K-Nearest Neighbor (KNN), Random Forest Classifier (RFC), Linear Discriminant Analysis (LDA), and Gaussian Naive Bayes (GNB), is analyzed for the classification of PD defects. The optimal parameters for maximum classification performance are obtained for the considered algorithms. Matthews correlation, Weighted average F1 score, the area under the receiver operating characteristic (ROC) curve, and Cohen’s Kappa Score are reported for the considered algorithms and it is observed that the RFC algorithm shows the highest accuracy levels compared to the other algorithms.