Using Graph Attention Network for Causal Reasoning of Construction Safety Accidents

Historical construction accident reports can provide key references for accident cause analysis. Accidents with similar accident processes and case backgrounds are often attributed to similar causal factors. This study proposes a method based on graph attention network to enhance cause reasoning. First, the graph attention network is used to aggregate the accident features of multiple similar accidents. After feature aggregation, an accident similarity network is formed to achieve feature optimization and redistribution, which helps to improve the quality of feature representation and case retrieval. Then, the relationship between accidents and causes is established to form a two-layer reasoning network, and the key causes of accidents are inferred based on network association analysis. Compared with traditional case retrieval and reasoning, this method offers flexibility by simultaneously considering features from multiple cases to comprehensively infer accident causes. The method’s feasibility is validated through two actual cases. The results showed that the accuracy of the reasoning factors for accident causes can reach 83.33%. It also provides new insights for quantitatively exploring accident process to improve the production environment.