A Domino Effect-Driven Knowledge Graph for Large Language Model-Based Risk Identification in Natural Gas Pipeline Operations

In light of the hallucination issue frequently encountered by large language models (LLMs) in risk identification, a domino effect–based approach is introduced for constructing knowledge graphs that represent risk events, contributing factors, and corresponding mitigation strategies. These knowledge graphs serve as external knowledge bases for LLMs, supported by carefully designed prompt words to enhance retrieval and reasoning capabilities. A System-Theoretic Process Analysis (STPA) of natural gas pipeline operations was employed as a case study to evaluate the effectiveness of this method in improving the risk identification performance of LLMs. The findings indicate that the knowledge graph–based Retrieval-Augmented Generation (RAG) approach significantly reduces the occurrence of hallucinations in LLM outputs, thereby increasing the precision of STPA. This approach presents a novel avenue for utilizing LLMs in risk identification tasks for complex industrial systems.

Keywords: Risk identification, LLMs, Domino effect, Natural gas pipelines, STPA, Knowledge graph.