Design and Application of an Intelligent Fault Diagnosis System for Thermal Power Units Based on Fault Knowledge Base and RAG Architecture

To address the issues of low accuracy and slow response in traditional fault diagnosis for thermal power units, as well as low knowledge base utilization and poor interpretability of existing intelligent systems, a diagnostic system integrating a fault knowledge base and the retrieval-augmented generation (RAG) model is proposed. The system adopts a four-layer architecture of "Data-Processing-Service-Application" and integrates four core modules. A knowledge base containing over 3000 fault cases and more than 800 entity relationships is constructed. Combined with a domestically developed large model enhanced by RAG, the system forms a "retrieval-reasoning-verification" closed loop. Additionally, the signed directed graph (SDG) mechanism model and long short-term memory (LSTM) data-driven model are introduced to improve diagnostic reliability. Verification on a 600 MW thermal power unit shows that the system achieves a diagnostic response time of less than 1 second, a fault localization accuracy of over 90%, and a 65% reduction in false alarm rate compared with traditional systems. It supports concurrent monitoring of more than 200 units and maintains an availability of 99.99%, realizing full-process intellectualization of fault diagnosis for thermal power units.