A Novel Domain Adversarial Graph Convolutional Network for Insulation Defect Diagnosis in Gas-Insulated Substations

The data-driven models have achieved remarkable advancements in diagnosing gas-insulated substation (GIS) insulation defects on specific massive datasets. However, restricted by field operating conditions and sample scarcity of GIS, the above methods are difficult to achieve high-precision and robust diagnosis of GIS insulation defects on-site. To settle these problems, we propose an innovative domain adversarial graph convolutional network (DAGCN) for few-shot GIS insulation defects on-site. To improve the accuracy of GIS insulation defects diagnosis, a graph convolutional network (GCN) with multiple attention mechanisms is constructed to take advantage of node features and topology to adaptively mine defect features. Then, a novel domain adversarial (DA) tactic is introduced for GIS insulation defect diagnosis in few-shot conditions through the transfer of diagnostic knowledge. In the proposed DA, the source and target domains classifier loss are jointly optimized to realize feature migration and domain matching. Experiment results demonstrate that the DAGCN achieves reliable diagnosis of GIS insulation defect, which has obvious advantages over existing methods. Simultaneously, the DAGCN proposed can achieve satisfactory diagnosis including unlabeled field samples. Moreover, it has development potential in diagnosing unlabeled samples, thus providing a feasible solution for GIS insulation defect diagnosis and GIS operation and maintenance.