Partial Discharge Pattern Recognition for XLPE Cables Based on Autonomous Feature Extraction of CNN

Most pattern recognition methods employed for differentiating partial discharge caused by different types of insulation defects in XLPE power cables mainly rely on the manual extraction of partial discharge features, which is easily affected by subjective uncertainty. An efficient insulation defects recognition method based on autonomous feature extraction of convolutional neural network (CNN) is proposed in this paper. Original partial discharge signals of four defects are obtained through experiments firstly, then the time-domain waveform image is taken by the skills of graying and clipping as the input of CNN for classification. The influences of different convolution layers, pooling methods, activation functions, convolution kernel sizes and input image sizes on the network performance are studied comprehensively. Experiments demonstrate that our method could achieve the overall recognition rate of 96%, which is 3.2% and 6.0% higher than that of SVM and BP neural network, respectively. Our algorithm automatically extracts the intrinsic features of image pixel data by CNN, which avoids the uncertainty of manual feature extraction, and has higher recognition rate and better robustness.