Enhancing Power Transformer Fault Diagnosis Through Dynamic Knowledge Graph Reasoning

Power transformers (PTs) are pivotal in electrical power systems as they facilitate transmission, distribution, and power supply by altering the ac voltage levels. Transformer failures can lead to power system interruptions and equipment damage, impacting the reliability and stability of power supply. Conventional fault diagnosis (FD) methods often struggle to effectively use fragmented knowledge and expert experience, resulting in reduced fault localization accuracy. To address these challenges, this article proposes a fault localization model for PTs using knowledge graphs (KGs). We present a novel knowledge extraction model to extract fact triples, facilitating the automatic construction of a PT fault KG (PTF-KG). In addition, existing KG-based FD methods often perform poorly in handling complex relationships and multihop reasoning. To overcome these limitations, we propose the knowledge graph reasoning model based on the dynamic interaction of dual agents (KGR-DIDA) model. The KGR-DIDA model comprises global and local agents, aiming to explore inference paths from source entities to target entities. The global agent uses a global policy to select the most relevant relationship, while the local agent, based on the chosen relationship, applies a local policy to navigate to the next entity, bringing us closer to the target entity. The action trajectory of the dual agent forms a reasoning chain that explains the fault query results. On the PTF-KG dataset, the KGR-DIDA model achieves an average FD accuracy of 97.5%. The experimental results validate that the KGR-DIDA model effectively addresses the challenges of multihop reasoning under complex relationships and significantly improves FD accuracy.