Application of Machine Learning for Partial Discharge Classification under DC Voltage

Partial discharge measurement is one of the most important diagnosis methods and well investigated under AC voltage. Furthermore, machine learning is established and has been used successfully already many years for automated recognition of PD defects. For AC voltage, there are several diagnosis methods and interpretation tools. In the field of DC voltage this is not the case, so it needs significant tools to interpret the results. In this contribution typical PD defects of HVDC GIS/GIL are investigated, but the methods can be adopted to other HV equipment as well. The machine learning techniques were realized with MATLAB and WEKA. Statistical parameters, derived from the PD pulse sequences, were used as features. A hierarchical clustering of the features was performed to analyse the separability between the PD defects. Classification was done with three popular algorithms (SVM, k-NN, ANN). The parameters of these algorithms were varied and compared to each other’s. SVM clearly outperformed the other classifiers.