Scenario-based multimodal deep learning framework for simultaneous detection of construction accident causal factors and risk evaluation

Construction industry's high hazard rates require intelligent safety management systems. Previous studies relied on manual inspection and single unified detection models, which showed limited accuracy in complex construction environments due to diverse hazard types and environmental variations. This paper developed a parallel-scenario-based architecture to simultaneously detect 24 accident causal factors. The factors were strategically categorized into four specialized scenarios based on work environments, with optimized Faster R-CNN parallel models built for each scenario. The parallel approach significantly outperformed conventional methods: F1-score improved by 66.7% compared to single unified models, achieving individual scores of 0.80–0.93 and mean average precision of 0.67–0.86. The system detected up to 12 accident causal factors simultaneously from single images. The multimodal framework integrated deep-learning visual detection with construction accident statistics to calculate real-time occurrence probability. This paper contributes a paradigm shift from qualitative risk identification to quantitative risk assessment, enabling data-driven decision-making in construction safety management.