• Proposed a knowledge extraction strategy for human-LLM collaboration. • Established a concept-fact-norm knowledge graph framework for tunnel fires. • Fused KG and LLM enabling second-level decision-making responses. • Integrated RAG with prompt, enabling over 70% reliability on professional QA. • Validated the reduced workload in a real tunnel design case of the optimized system. Tunnel fires pose severe threats due to the confined nature of tunnel environments, resulting in significant casualties, property damage, and transportation network disruptions. Effective fire prevention, emergency response, and post-disaster recovery rely heavily on comprehensive and well-structured knowledge. However, existing knowledge on tunnel fire safety is highly fragmented, dispersed across research studies, technical manuals, and regulatory documents, making it difficult to retrieve and apply in real-time scenarios. The lack of a structured and integrated knowledge system hinders efficient decision-making and emergency management. To address this issue, this paper proposes an intelligent decision-support method that integrates dynamic knowledge graphs and large language models. By leveraging ontology modeling, natural language processing, and graph database technologies, a domain-specific knowledge graph for tunnel fires is constructed, encompassing conceptual, factual, and normative knowledge. This framework facilitates the semantic organization and dynamic expansion of multi-source heterogeneous information. Building on this foundation, an intelligent system capable of natural-language question answering is developed. By applying retrieval-augmented generation and prompt engineering techniques, the system combines the precision of domain knowledge graphs with the semantic understanding of large language models, thereby significantly enhancing the reliability and practicality of question answering in professional settings. A case study involving the ventilation system design of a highway tunnel engineering project demonstrates the feasibility and practical value of the proposed system for decision-making in specific tunnel fire prevention scenarios. By integrating scattered expertise, the system replaces the traditional manual approach with a second-scale automated process that enables a compliant, scientific design for the case project while dramatically improving efficiency. This research provides systematic knowledge support and an intelligent interactive tool for tunnel fire disaster prevention design, emergency response, and safety management, thereby offering a feasible reference for enhancing tunnel fire prevention and control capabilities.
Lin et al. (Tue,) studied this question.