AI 기술을 접목한 사고조사 데이터 분석 및 활용

Objectives: This study proposed and evaluated an artificial intelligence (AI) framework based on GraphRAG that integrates large language models with an industrial safety knowledge graph. This framework is designed to automate the analysis of accident investigation reports and support risk assessment and the planning of preventive measures. Methods: Accident-investigation texts, images and related legislation from the Korea Occupational Safety and Health Agency were processed and standardized. Key entities, relations, and visual risk cues were extracted using named-entity recognition, dependency parsing, and vision-language models, and were then incorporated into an industry-wide safety knowledge graph. A retrieval-augmented generation pipeline, combining vector search with graph traversal, provided a Korean large language model with multi-hop context, enabling explainable question-answering. Pilot tests compared AI outputs with expert judgements on historical cases. Results : The system suggested frequent risk factors for specific industries and processes in a contextual manner, automatically detected potential legal violations and shortcomings in proposed countermeasures within new reports, produced quantitative risk ratings by analyzing historical incident patterns and visualized causal networks to facilitate expert validation and iterative knowledge enrichment. Expert reviewers confirmed that the AI markedly reduced manual screening time while maintaining explanatory detail. Conclusions: The GraphRAG approach effectively combines unstructured accident narratives with structured safety knowledge to deliver transparent insights for proactive risk management. The continuous ingestion of new reports and phased integration with the KRAS platform could help to make data-driven safety governance the norm and speed up the transition to smart occupational risk control.