Intelligent Decision Support for Safety Risk Management of Urban Gas Pipeline Network Leakage by Integrating Large Language Models and Knowledge Graph Technologies

To address the systemic challenges in risk prevention and control of urban gas pipeline network leakage accidents (UGPNLAs), this study conducts a systematic analysis of 209 accident investigation reports. Nine categories of core entities, including disaster type, risk factor, accident location, and leakage location are identified, along with nine types of entity relationships, such as "risk factor-risk factor," "risk factor-disaster type," and "disaster type-leakage location." By integrating the STAMP with association rule mining, this study identifies the risk factors and their interaction relationships in UGPNLAs. On this basis, a multi-dimensional knowledge graph for UGPNLAs is constructed to deeply represent the accident causation chains and evolutionary patterns. Furthermore, a knowledge graph-large language model (LLM) intelligent auxiliary decision-making system based on “knowledge graph retrieval-Cypher query-evidence-enhanced generation” was developed to support precise accident knowledge retrieval and risk prevention and control decision-making. The results indicate that this study not only reveals the coupling mechanisms among risk factors, but also identifies 17 key risk factors based on complex network theory (e.g., insufficient safety supervision and guidance, workers' non-compliant operations, equipment aging). Moreover, compared with general LLMs, the knowledge-graph-enhanced approach provides more evidence-based, reliable, and interpretable support for risk identification, accident tracing, and risk prevention and control decision-making.. Overall, this study provides an innovative theoretical framework and technological paradigm for the systematic cognition and intelligent management of risks in urban gas pipeline networks, demonstrating significant engineering value for improving the reliability and safety resilience of urban lifeline infrastructures.