Partial Discharge Classification from Highly Noise-contaminated Data Using Cascaded Two Neural Networks

Partial discharge (PD) classification is a crucial tool for assessing the reliability of insulation systems in high-voltage electric power equipment, and many denoising techniques such as the wavelet shrinkage method have been developed for dealing with external noise contamination in the preprocessing step of PD classification. However, conventional denoising methods require burdensome manual parameter tuning because the PD characteristics depend on the measurement conditions. In addition, previous studies have focused on signals with relatively high signal-to-noise ratios (SNR) obtained by intrusive sensors such as ultrahigh-frequency sensors, so those methods cannot be naively applied to low SNR signals obtained by nonintrusive sensors such as transient earth voltage sensors. Here we propose novel cascaded two neural networks for estimating PD type via reconstruction of noise reduced PD feature from highly noise-contaminated signals without manual parameter tuning. Experimental results show that the proposed method achieved higher classification accuracy than conventional methods, including neural network-based method, even for highly noise-contaminated signals.