A dynamic risk assessment method integrating deep learning with Bayesian network: A case study in machining workshop

Given the complex multi-trade operations, high-density layouts, and synergistic hazards with catastrophic potential in machining workshops, risk assessment demands urgent attention. Critical research gaps remain, particularly in dynamic risk assessment based on continuously updated monitoring data. To address existing challenges, an innovative method is proposed to identify, analyze, and evaluate three distinct types of risk by continuously updating both the probability and consequence based on visual and environmental data. In a case study conducted in a machining workshop, a deep learning-based method was employed to detect 13 types of unsafe behaviors and 8 types of unsafe conditions. The identification results were then input into the Signal Layer of a 31-node Bayesian network (BN) for risk analysis. Risk evaluation was subsequently conducted based on a risk matrix. The method was validated using nine representative scenarios, and the YOLOv11 model (selected for its superior balance between detection accuracy and inference speed) demonstrated excellent detection performance, with most precision scores exceeding 0.9. Even for difficult small targets, the precision reached 0.87 or higher. Both recall rates and F1 scores consistently exceeded 0.8. The risk analysis and evaluation results show that the co-occurrence of unsafe behaviors or hazards exerts a synergistic effect; even two concurrent hazards can at least double the baseline level of one risk factor. Overall, this integrated approach provides a strong framework for risk assessment in machining workshops, effectively enhancing operational safety and mitigating potential hazards.