Leveraging Knowledge Graph and Large Language Model Synergies for Intelligent Fault Analysis in Urban Rail Transit Signaling Systems

To address the challenges of insufficient text data utilization and low manual diagnosis efficiency in urban rail transit signaling system fault analysis, particularly the difficulty in handling novel faults under unmanned train operations, this study proposes a dual-engine framework integrating Knowledge Graph (KG) and Large Language Model (LLM). Based on preprocessed fault diagnosis logs from Nanjing Metro's signaling system, a multi-dimensionally annotated dataset was constructed. The CasRel joint extraction model was employed to overcome the limitations of traditional pipeline methods in recognizing nested terms and mitigating error propagation, building a high-density domain-specific KG. Subsequently, a three-tiered question-answering paradigm (fact query, enumeration verification, causal verification) and multi-granularity intent classification system were innovatively integrated to enable semantic retrieval and dynamic reasoning over the KG. To tackle data dependence and logical hallucinations in vertical domain LLMs, a domain-adapted LLM was developed using QLoRA parameter-efficient fine-tuning technology, complemented by a zero-shot autonomous reasoning mechanism based on prompt engineering. By deconstructing signal transmission logic, cross-system association rules, and device temporal relationships, causal chain prompts, analogical inference templates, and dynamic temporal context cues were generated to guide the LLM in performing root cause deduction along domain knowledge-constrained pathways without historical data. Concurrently, the reasoning logic underwent real-time verification leveraging the KG's security protocol rule library to filter domain-violating hypotheses. Graph Retrieval-Augmented Generation (GraphRAG) technology further embedded KG retrieval results and dynamic prompts into the LLM's input, forming a closed-loop “knowledge anchoring, prompt guidance, verification and generation” decision process. Experimental results demonstrate that the framework significantly enhances novel fault diagnosis accuracy and reliability, achieving security-compliant root cause localization in zero-shot scenarios while outperforming traditional methods. It autonomously transforms new fault resolutions into KG nodes, triggering LLM incremental learning to enable continuous evolution of the knowledge system. This research provides an interpretable and scalable diagnostic decision-support paradigm for intelligent rail transit operations, offering technical substantiation for deploying industrial LLMs in safety-critical systems.