Assessing the probability of fire and explosion accidents from unsafe behaviors in laboratories: an innovative approach integrating object detection and behavior deduction

Abstract

Unsafe behaviors are usually recognized as a critical factor in process safety, as they often serve as the primary trigger for accidents and hazardous events. To address the risks associated with unsafe behaviors in laboratory environments, an innovative approach is proposed for identifying and assessing behaviors that could lead to fires or explosions. In a case study focusing on seven specific unsafe behaviors in a chemistry laboratory, YOLOv11, an object detection model, was trained on a realistic video dataset. The model demonstrated high precision in identifying these behaviors, with most detections achieving scores above 0.9, and even small target instances scoring above 0.8. Recall rates are also notably high, with F1 scores consistently exceeding 0.8. To further assess the risks associated with these detected unsafe behaviors, a Bayesian network model is established to conduct scenario analysis, revealing that coexisting multiple target features may lead to coupling risks, thereby significantly increasing risk levels. Overall, this integrated approach offers a robust framework for the identification, assessment, and mitigation of unsafe behaviors in laboratories, ultimately enhancing safety and mitigating potential hazards.