FaultSeer: An Agentic Retrieval-Augmented Generation Framework for Defect Analysis and Diagnosis in Power Systems

Fault detection and diagnosis are critical for maintaining the safe and stable operation of power systems. Traditional methods often depend on manual rules or shallow matching, which are insufficient to address diverse defect characteristics and unstructured queries. To overcome these limitations, this paper proposes a defect analysis workflow based on the DeepSeek V3 large language model and a local knowledge base, adopting a retrieval-augmented generation (RAG) strategy. The system leverages embeddingbased retrieval to organize historical defect information and employs re-ranking models to enhance semantic relevance. With this workflow, the model can infer potential causes of newly observed defects based on their characteristics (e.g., representation, manufacturer, commissioning date) and support multidimensional queries and statistics by manufacturer, time, or region. Additionally, semantic normalization and fuzzy matching strategies are integrated to mitigate the impact of inconsistent data reporting and vague user queries, thereby enhancing retrieval accuracy. Experimental results show that the proposed system achieves significant improvements over traditional tools, particularly enhancing semantic accuracy.