Construction and Application of Multi-Agent Large Language Model for Power Transformer Fault Diagnosis

To address the limited applicability of traditional small language models and the unsatisfactory diagnostic effects caused by the multi-dimensional heterogeneity of state parameters and data loss in power transformers,a multi-agent large language model framework for transformer fault diagnosis is proposed.The framework establishes a collaborative reasoning system with fault cases knowledge graph and three agents.The fault cases knowledge graph constraint mechanism significantly enhances the large language model's understanding of power engineering expertise while effectively mitigating machine hallucination.Three agents simulate expert diagnostic processes by decomposing complex transformer fault diagnosis tasks.Primary diagnostic agents perform preliminary fault identification through feature threshold analysis, expert diagnostic agents conduct uncertainty reasoning on typical fault patterns using probabilistic graphical models,and case analysis agents access a historical fault case database to enable knowledge retrieval and diagnostic result validation.Validation results show that the proposed model achieves excellent performance in diagnosing 100 fault cases,with an accuracy rate of 86%,representing a 33% improvement over the BERT model.The integration of the fault cases knowledge graph enhances the large language model's output in terms of completeness,semantic consistency,and professional depth,with an expert rating average performance increase of 50%. This multi-agent large language model demonstrates superior performance in reducing misjudgment rates compared to monolithic large language models,and can provide technical support for intelligent fault diagnosis and operation and maintenance of power transformers.