Unsupervised Deep Learning for Detecting Number of Partial Discharge Sources in Stator Bars

Regular maintenance and testing of high-voltage assets are important for avoiding outages. Measurement of partial discharges (PDs) is commonly employed to monitor the health of the insulation materials of such assets, and the detection of the PD sources would help justify remedial actions such as inspection and repair. Before classifying the PD sources, the identification of the number of PD sources is necessary (clustering of the PD pulses). Different clustering techniques have been used for this purpose such as time–frequency maps,
${K}$
-means clustering, and principal component analysis. However, there are limitations in these techniques in regard to the additional information that these algorithms need. For example, the number of clusters and the minimum distance between data points must be assumed in some algorithms. In order to overcome the mentioned limitations, a robust unsupervised deep learning model is proposed for unsupervised clustering, based on a convolutional AE and an adaptive clustering technique. In a laboratory setup, defects are introduced to a generator stator bar, simulating common PD sources, and PD-induced pulses in the ground connection of the bar are measured. The inputs to the deep learning model are the unlabelled, time-series PD pulses, and the output is the predicted number of PD sources. The proposed system showed better prediction of the number of PD sources compared to the existing techniques, which require human judgment. In addition, the proposed system demonstrated immunity to noise, where additive white Gaussian noise (AWGN) is added artificially to the measured PD waveforms.