The underground working environment is complex and full of safety hazards, while traditional supervision methods are limited and incapable of providing comprehensive real-time monitoring. Therefore, research on recognizing unsafe behaviors of underground personnel using machine learning has become increasingly urgent. This study proposes a recognition method and model for underground personnel's unsafe behaviors based on YOLOv11. Unsafe behaviors are categorized into 3 types: protective equipment-related, hazardous area-related, and dangerous behavior-related. A dataset was constructed according to these categories. The Albumentations was used for data augmentation, simulating underground environmental variations by adjusting lighting conditions, adding noise, and more, to enhance model generalization and robustness. Finally, a semi-automatic labeling method was adopted for data annotation, and the model was trained. The results show that the model, after data augmentation with Albumentations and using semi-automatic labeling, achieved good performance without requiring extensive manual annotation.