Discrimination of Insulation Condition Based on Partial Discharge Waveform Using Deep Learning

The characteristics of the partial discharge (PD) waveform reflect the behavior of electron avalanches in the discharge space. When an insulator with lower surface resistivity is employed, the difference in surface resistance appears at the rising part of the waveform. In this report, we classified the insulation condition from the PD waveform using deep learning. We showed that learning by extracting only the rising part of the waveform improved the classification accuracy. Furthermore, the frequency band of the waveform was adjusted using an appropriate digital filter, and the relationship between the accuracy and waveform characteristics was shown from the point of view of the difference in the statistical distribution of waveform characteristics. Using the JS (Jensen-Shannon) divergence as an index of the similarity of the distribution, it was found that the rising part of the waveform mainly reflected the surface resistivity difference. This result agrees with the tendency read from the classification test using deep learning. In addition, it implied that it is possible to estimate the waveform characteristic strongly reflecting the insulation condition using the classification test.