Research on Partial Discharge Pattern Recognition of Small Samples Based on STFT and CNN-SVM

In actual power production, it is difficult to collect large samples of partial discharge signals from power equipment. Therefore, in order to achieve partial discharge pattern recognition in small sample situations, this paper first applies wavelet denoising to the collected signals of three types of discharges: surface discharge, air-gap discharge, and corona discharge. Then, the time-frequency maps of the three defect models are obtained through short-time Fourier transform (STFT). Finally, four network recognition models, CNN-SVM, CNN, CNN-BIGRU, and CNN-LSTM, were trained separately. The results show that the STFT time-frequency map after wavelet denoising contains rich features of partial discharge signals, which can effectively support subsequent recognition models for pattern recognition; Compared to other recognition models, CNN-SVM can maintain high recognition accuracy and correctness even in small sample situations, achieving good pattern recognition performance.