Abstract: This paper presents an intelligent hidden hazard recognition model based on the Swin Transformer architecture, and an analysis of its test applications to underground engineering construction, aimed at the industry pain points at construction sites, such as manual hazard inspections, low efficiency, high rates of missed detections, and limited hazard coverage. We construct a two-stage recognition framework comprising self-supervised pre-training - supervised fine-tuning, and adopt a video frame extraction technology based on the cross-cosine similarity threshold control method to build a comprehensive hazard dataset. This model completes pre-training through semantic-constrained random mask reconstruction tasks, and uses a transfer learning strategy in its accurate recognition of 55 types of construction hazards. The test set results show that it achieves an accuracy of 89.8%, precision of 90.5%, and an F1 of 89.4%, and maintains recognition accuracies of 83.6% and 85.4% under the simulated conditions of low light and dust interference respectively. The system based on the model can be seamlessly integrated with the existing safety management platforms, forming a closed-loop governance process of AI recognition - construction rectification - online review by supervisors - random inspection by construction units. Through the test operation of 14 lines and 65 sections of the Beijing's rail transit system, we have applied the system to nearly 29,000 hazard identification cases in total, and achieved stable control effect on key risks such as high-altitude falls and object strikes. This study demonstrates key construction technologies and application paths and their importance for intelligent supervision, helping refined safety governance in underground engineering construction.

Key words: underground engineering, construction safety, hazard recognition, Swin Transformer