Comparison of clustering techniques of multiple partial discharge sources in high voltage transformer windings

Modern high voltage (HV) insulation systems consist of multiple dielectric media - multi-source partial discharge (PD) data discrimination is required. The ability to accurately distinguish between the PD signals generated from different sources is seen as a critical function of future diagnostic systems. Two model PD sources were utilized in this investigation to replicate void and surface discharges. The proposed processing technique relies on the assumption that the PD pulses generated from different sources exhibit unique waveform characteristics. Several clustering techniques have been employed to identify and separate multiple PD sources recently. However, for further analysis, the techniques used must produce a significant separation between the clustered data as the phase resolved patterns produced by multiple PD sources overlap - inhibiting automated classification. An experiment has been designed to activate a pair of PD sources and inject the signals simultaneously into an HV transformer winding. After the PD pulses were extracted from the measurement data recorded using two wideband radio frequency current transformers (RFCTs) positioned at the neutral to earth point and the bushing tappoint to earth. Principal Component Analysis (PCA) and t- Distributed stochastic neighbor embedding (t-SNE) were applied to the Wavelet Energy (WE) data and their performance at discriminating between the two PD sources assessed.