LAUSR

Convolutional neural networks (CNNs) have promoted the development of insulation defect diagnosis for gas-insulated switchgear (GIS) because of their excellent feature extraction and classification capabilities. However, CNN ignores the correlation between the local areas of the feature space, resulting in insufficient feature utilization. Moreover, deploying and applying the methods explored in massive laboratory data to complex and few-shot conditions on-site is a difficult problem currently. Therefore, this study proposes a novel domain adaptation graph convolutional network (DAGCN) for GIS insulation defect diagnosis. First, the graph signal and graph convolution network are used to take advantage of the correlation between the local areas of the feature space while employing the numerical characteristics of the insulation defect signal. Then, the learned diagnostic method is deployed to the on-site GIS insulation defect diagnosis with domain adaptive transfer learning (TL). The difference between the two tasks is reduced by minimizing the difference between the marginal and conditional distributions of the source and target domains, thus, realizing the feature migration. The experimental verification shows that DAGCN has higher diagnostic accuracy and robustness than traditional methods, especially in diagnosing few-shot on-site. It provides a reliable reference for high-precision and robust diagnosis of few-shot GIS insulation defect diagnosis.