DiagAgent: An Agent Framework for Power Equipment Fault Diagnosis by Integrating RAG and MCP Tools

Large language models (LLMs) show promise for diagnosing power equipment faults. However, they face issues such as hallucinations, knowledge lag, and a lack of specialized analytical capabilities. This paper introduces DiagAgent, an innovative Artificial Intelligence (AI) agent framework. It synergizes Retrieval-Augmented Generation (RAG) with domainspecific small models through the Model Context Protocol (MCP). The framework first constructs a knowledge base containing equipment information and historical cases. Next, it employs a multi-level hybrid retrieval strategy to obtain relevant information as the intelligent core. The LLM then dynamically invokes specialized small-model tools encapsulated via MCP, which include analyzers for operating data,$\text{SF}_{6}$pressure, and simulations, to conduct in-depth analysis. Finally, it generates structured diagnostic reports. A power bushing diagnosis case, implemented on the Dify platform, demonstrates that DiagAgent significantly outperforms baseline methods, achieving a diagnostic accuracy of 91.3% and a report quality score of$4.4 / 5$. This work provides a practical and scalable blueprint for developing reliable, domain-specific AI assistants in the power industry and beyond.