Emergency Operation Scheme Generation for Urban Rail Transit Train Door Systems Using Retrieval-Augmented Large...

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Safety Science

Background:

• Urban rail transit (URT) train-door failures are safety-critical, causing service disruptions.
• Existing emergency operation schemes (EOSs) are static, difficult to adapt, and hard to verify.
• Need for dynamic, verifiable, and evidence-traceable EOS generation.

Purpose of the Study:

• Propose a retrieval-augmented large language model (LLM) framework for executable and evidence-traceable EOS generation.
• Improve adaptability and verifiability of safety-critical operation schemes.
• Address limitations of static EOSs in evolving fault patterns.

Main Methods:

• Normalize multi-source heterogeneous incident evidence into a structured representation.
• Employ a hybrid retriever (dense + BM25) with cross-encoder reranking for evidence selection.
• Fine-tune a generator with structured objectives for schema compliance, role assignment, and citation grounding.

Main Results:

• Hybrid retriever with reranking achieved high retrieval quality (Recall@5 = 0.78).
• The full LLM framework significantly improved operational usability metrics (SchemaPass = 0.88, RoleAcc = 0.91, CiteCov = 0.73, UsableAns = 0.83).
• Outperformed pure LLM baseline (UsableAns = 0.15) and RAG-only prompting (UsableAns = 0.26).

Conclusions:

• Combining high-utility retrieval with structure- and citation-aware fine-tuning enhances EOS executability and verifiability.
• The proposed framework offers a substantial improvement for safety-critical operation scheme generation in URT.
• Demonstrates the potential of LLMs for dynamic and reliable safety management in critical infrastructure.