Objective To improve the intelligent O&M (operation and maintenance) level of urban rail transit signal systems and address prominent issues in traditional O&M modes such as data silos, inefficient retrieval, and excessive reliance on human expertise in fault analysis, it is necessary to conduct research on intelligent O&M and fault analysis methods for signal systems based on large models.

Method Taking the intelligent O&M of metro signal systems as the research object, an intelligent analysis system consisting of data layer, service layer, application layer, and interaction layer is constructed. Four core technologies, i.e. NLP (natural language processing), Text-to-SQL (structured query language), RAG (retrieval-augmented generation), and OCR (optical character recognition) are adopted. A full-process processing mechanism is established based on semantic alignment, dynamic SQL generation and correction, hybrid retrieval, and re-ranking. Experiments are conducted to compare different retrieval methods and their performances. Meanwhile, fault scenario application tests are conducted to systematically assess the performance of each retrieval method in tasks such as fault Q&A, table parsing, and status discrimination.

Result & Conclusion  Experimental results show that the Hybrid+reranker (hybrid+BGE-reranker-v2-m3 re-ranking model) method performs excellently in six typical Q&A tasks, its overall performance is also significantly improved compared with traditional keyword-based retrieval and rule-based matching retrieval. The constructed intelligent system can achieve automatic fault analysis, precise retrieval of O&M knowledge, and rapid generation of decision suggestions, providing technical solutions and practical support for the digital and intelligent O&M of urban rail transit signal systems.