Separation of superimposed PRPD Patterns by Signal Clustering

Partial discharges are one of the main reasons of insulation failure in transformers. Multiple partial discharge sources in HV equipment can be active at the same time. Diagnosis is complicated by superimposed or overlapping patterns in the PRPD patterns. In order to classify the fault source correctly, the patterns must be separated first. In this paper we propose a method to separate partial discharge signals from different sources by their impulse waveform. A feature-set is constructed by selecting relevant features from a pool of 700 identified possible features. Then, different sources are separated by clustering techniques. Centroid-based clustering methods like the k-means algorithm are compared to density-based approaches like DBSCAN. After clustering multiple PRPD patterns can be calculated – one for each source. This method is developed on artificial partial discharge data, measured in the laboratory. Four different typical defect types are considered. Through multiple case studies we show, that the separation process works on different kinds of PD problems. This method works for all kinds of PD faults. It is shown, that PD sources not present in the training data used for development can also be separated reliably.