Few-shot Partial Discharge Pattern Recognition for Power Cables

Abstract

The data set of phase-resolved partial discharge (PRPD) grayscale image of power cables is usually small, and it is difficult to train complex deep residual networks for high-efficiency defect recognition. In this paper, a novel method combining data augmentation with network optimization is proposed to tackle this problem. Deep convolutional generative adversarial network (DCGAN) is employed firstly to augment training data set, in particular, an evaluation indicator based on the box dimension is proposed to verify the effectiveness of the generated samples. In addition, a deep residual network is optimized from the perspectives of network depth, convolution kernels and shortcut type to adapt to the small-scale PRPD grayscale image data set. Compared with the original deep residual network trained with the small sample, the proposed method achieves faster training speed and higher recognition accuracy. The recognition accuracy could reach 98.5% in the experiment, which effectively improves the effect of partial discharge pattern recognition for power cables under small sample condition.