| Title |
Automated Safety Countermeasure Generation Using LLM-based Accident Case Analysis |
| Authors |
성지연(Sung, Ji Yeon);나종호(Na, Jong Ho);공준호(Gong, Jun Ho);정유석(Jung, Yoo Seok);신휴성(Shin, Hyu Soung);오윤석(Oh, Yoon Seuk) |
| DOI |
https://doi.org/10.12652/Ksce.2026.46.1.0111 |
| Keywords |
건설안전; 사고사례; 의미검색; LLM; 재발방지대책 Construction safety; Accident case analysis; RAG; LLM; Countermeasure generation |
| Abstract |
Construction sites in Korea continue to exhibit a high incidence rate of industrial accidents, while many accident countermeasure documents have been written in a largely formalistic manner, resulting in limited practical effectiveness. In this study, an automated system for generating recurrence prevention measures was proposed using an LLM-RAG framework based on accident reports from the Construction Safety and Health Information System (CSI) and the KOSHA construction safety guidelines. The proposed system was designed to promptly derive accident-type-specific recurrence prevention measures and provide concrete actions that can be applied to real construction sites. A Korean-specialized embedding model was employed to semantically vectorize both accident cases and guideline documents, and a retrievable knowledge base was constructed accordingly. Upon user query input, semantically similar accident cases and guideline documents were retrieved and provided to the large language model as reference materials, enabling evidence-grounded response generation. In the experimental evaluation, the suitability of the generated outputs was compared and reviewed across multiple accident-type queries, and the generated measures were assessed in terms of alignment with the recommendations specified in the guidelines. Based on an LLM-as-a-judge evaluation, an accuracy of approximately 93 % was achieved, suggesting that the generated recurrence prevention measures were appropriate for the given accident types and highly applicable in practice. In particular, the document-grounded outputs tended to demonstrate improved specificity and consistency compared to conventional countermeasure documents. The proposed approach is expected to contribute to the development of practical response strategies within increasingly digitalized construction safety management systems. |