Research on Hybrid Diagnosis Method of Transformer Faults Driven by Knowledge Graph

As a core component of the power system, the accuracy and timeliness of transformer fault diagnosis directly affect grid security and economic operation. Traditional diagnostic methods exhibit significant limitations: static models struggle to adapt to the differentiated operational characteristics across varying service lifespans, while regional data isolation further constrains the efficiency of knowledge reuse. Therefore, developing a dynamic diagnostic technique that incorporates lifespan-based stratification, integrates fault mechanisms with data modeling, and enables cross-regional knowledge sharing is of great significance for guiding transformer maintenance, enhancing transformer reliability, and improving grid economy. To address the aforementioned issues, this paper constructs a dynamically stratified knowledge graph and employs Graph Attention Networks (GAT) to quantify weight migration between adjacent sub-graphs, thereby overcoming the inability of traditional static models to adapt dynamically to equipment aging patterns. Subsequently, a bidirectional reasoning framework combining operational mechanisms and data-driven models is established, integrating the interpretability of mechanism-based inference with the high accuracy of data-driven approaches. Finally, a federated graph learning framework is introduced, where a global model is generated through aggregation at a central server, and only encrypted sub-graph features are shared among regions, achieving secure knowledge transfer. This work provides a novel solution for intelligent transformer operation and maintenance.