Defect Recognition for Partial Discharge Patterns of Gas Insulated Switchgear and Cable Joint Based on Deep Learning Methods

An analysis aimed at improving the resilient method for defect recognition of partial discharge (PD) is presented. The PD datasets combine three different defects in gas insulated switch and three different defects in cable joints, for a total of six types of defects. Each PD data is converted into phase resolved partial discharge (PRPD) and pulse sequence analysis (PSA) patterns. Three different deep learning models, convolutional neural networks (CNN), Residual Network (ResNet), and MobileNet, are used to train the defect recognition models. The ensemble method is proposed by learning noises and voting to improve the resilience for PD data containing three types of noises that cannot be filtered using wavelet transform. The result shows that the accuracy is increased from 86.9% to 97.5% using the ensemble model.