Personalized safety training for construction workers: A large language model-driven multi-agent framework integrated with knowledge graph reasoning

Abstract

Construction sites are inherently high-risk environments, making safety training for workers crucial to enhancing operational skills, reinforcing safety awareness, and reducing accident risks. Traditional centralized training often fails to engage workers due to monotonous nature and lack of relevance, leading to low efficiency. Moreover, critical resources such as operating instructions, safety guidelines, and accident reports are frequently mismanaged or underutilized. Therefore, this study proposes ConSTRAG, an innovative personalized construction safety training framework. By integrating large language model-empowered agents with knowledge graph reasoning, ConSTRAG generates tailored training materials, significantly improving the relevance and effectiveness of safety training. Validation tests conducted on a dataset of 11,020 questions achieved an average score of 81.25, exceeding the benchmark by 6.94. The generated personalized training materials were evaluated through an expert questionnaire survey, with an average score of 4.16 out of 5 across five dimensions. This research contributes to overcoming individual heterogeneity in construction safety training, enhances training efficiency and effectiveness, and holds potential for extension to other personnel training industries.