Data-Driven Differentiated Analysis of Construction Accident Causes Using Hybrid Knowledge Graphs and Bayesian Networks

Abstract Construction accidents are characterized by a high level of risk and are typically caused by the combined effects of multiple interacting factors. However, systematic investigation into the causes of construction accidents is lacking in current research, and differentiated analysis of accident types has not been provided, which renders the formulation of targeted prevention strategies challenging. To address these limitations, this study proposes a multilevel analytical framework to explore the underlying causes of construction accidents. First, accident reports of various types were collected and analyzed to extract causal factors, and a construction site accident knowledge graph (CSAKG) was established to visualize their interrelationships. Second, the decision-making trial and evaluation laboratory method was applied to simplify the complex interrelationships among causal factors within the CSAKG and to identify the key influencing elements. These elements were subsequently incorporated into a four-layer hierarchical Bayesian network (BN) model, which was developed based on the human factors analysis and classification system framework. Finally, differentiated causal pathway analyses were performed using the BN model to identify the critical causal factors and clarify their interaction mechanisms. The results demonstrate significant variations in key causal factors and their interaction pathways across different accident types, each exhibiting distinct causative logics and evolutionary characteristics. The identified pathways were further validated through tolerance design, thereby confirming the robustness and reliability of the proposed model. Overall, the proposed analytical framework effectively addresses the shortcomings of traditional approaches, offering precise, multilevel, and practical decision-support for construction safety risk management.