Partial discharge pattern recognition in GIS based on a novel one-dimensional convolutional neural network

Abstract

In recent years, massive data-driven deep learning methods, especially convolutional neural networks (CNN), have made rapid progress in partial discharge (PD) diagnosis of gas-insulated switchgear (GIS). However, the two-dimensional CNN (2DCNN) converts PD signals into the form of picture, resulting in the temporal fine-grained information of PD signals being ignored. In addition, 2DCNN relies on massive data and perform poorly on unbalanced samples. To solve these problems, a novel one-dimensional convolutional neural network (1DCNN) is constructed to recognize GIS PD pattern. Firstly, a fault diagnosis method based on the combination of spatial pyramid pooling and 1DCNN is proposed. This method not only enables the network to handle data of different sizes, but also reduces the complexity of network structure and the amount of computation required. Then, the focus loss function is introduced to guide the model training, which can discover difficult-to-classify samples, alleviate the problem of sample imbalance, and further improve the model identification performance. Experimental results show that the proposed 1DCNN can achieve high accuracy and robust GIS PD diagnosis under small samples, and has good fault tolerance for unbalanced samples, which provides a reliable reference for GIS PD diagnosis.