Noise Reduction of Partial Discharge in Power Cable Insulation Based on Deep Learning

With the rapid development of ultra-high voltage (UHV) transmission technology in China, insulation diagnosis for power equipment has been applied widely to improve the reliability of the giant power system. Partial discharge (PD) detection, with the advantages of high precision and convenient deployment, plays an important role in defect detection, location, and classification. However, the low signal to noise ratio (SNR) of PD signals and poor electromagnetic environment make it difficult to extract useful features from PD signals. To distinguish the PD signals from noise, a noise reduction method based on deep learning is proposed in this paper. With short-time Fourier transform (STFT) technology, time-domain PD signals with various noise were first transformed into time-frequency domain patterns. After preprocessing the raw pattern, a deep convolutional neural network (CNN) was built, trained, and tuned for noise reduction of PD signals. Finally, the noise reduction model was verified and compared with the traditional method based on wavelet analysis. The results demonstrate that noise reduction mothed based on deep learning can restore the PD signals with more than 90% noise reduced. Given the advantage that the convolution kernel could sense both the time-related information and the frequency-related information in a spectrum simultaneously, the DnCNN model performs better than wavelet analysis. The noise reduction model proposed in this paper is based on new generation artificial intelligence, providing a new approach for PD signal restoration, which promotes the reliability of the power system.