Automated knowledge graph-based risk assessment for fall-from-height accidents in construction

Fall-from-height (FFH) accidents remain a leading cause of fatalities in the construction industry. To systematically extract and analyze risk factors from unstructured FFH accident reports, this paper employed large language models (LLMs) to enable zero-shot automated fall-from-height knowledge graph (FFHKG) construction. By clustering FFHKG entities to merge semantically similar factors, a weighted complex network is formed, enabling topological analysis for quantitative risk assessment. A case study on 1097 FFH accident reports validates the proposed framework. Results demonstrate that GPT-4o achieves high extraction accuracy, with an F1 score of 0.94 in named entity recognition and a precision of 0.90 in relationship extraction. Key risk factors, such as poor safety management, lack of training, insufficient edge protection, etc., are quantitatively identified across multiple perspectives. Observable unsafe behaviors are also detected, offering insights for behavior-based safety monitoring. The proposed framework provides a data-driven solution for more effective safety management on construction sites. • An LLM-driven approach is proposed for automatic FFHKG construction without annotated data. • Accimap model is integrated into KG ontology to capture hierarchical interactions in FFH accident. • Quantitative analysis of risk factors is conducted to provide managerial insights. • The proposed approach advanced data-driven safety management practices on construction sites.