Localization of Partial Discharge Defects in High Voltage Specialized Transformers Embedded in Graph Transformer Networks

The operating environment of high-voltage specialized transformers is complex, with weak partial discharge defect signals and severe propagation path distortions, resulting in low accuracy of traditional single positioning methods. A method for locating partial discharge defects in high-voltage dedicated transformers using an embedded graph converter network is proposed. Firstly, a collaborative acquisition framework for ultra-high frequency, high-frequency current, and ultrasonic sensors is constructed, and a cascaded preprocessing strategy of variational mode decomposition and adaptive threshold denoising is designed to effectively separate defect features from background noise. Secondly, dynamic time warping is used to achieve time reference alignment of multi-source signals, and graph attention networks are further utilized to map heterogeneous features to a unified space. Finally, design a graph transformer network embedded with layered propagation of physical information in oil solid media, which uses propagation time constraints as edge feature residual correction terms, directly regresses the three-dimensional coordinates of defects, and forms an end-to-end localization model. Experiments have shown that under normal operating conditions, the maximum deviation is only 0.28 m, and under strong electromagnetic interference conditions, the average error is 0.41 m, significantly better than traditional methods.