YOLOv11 Model Detects Unsafe Coal Mining Behaviors

Why It Matters

The breakthrough delivers near‑human level detection of hazardous actions, enabling continuous, automated safety monitoring in hazardous mining environments. Faster, more accurate alerts can lower injury rates and operational downtime, a critical advantage for the global coal‑mining sector.

Key Takeaways

• YOLOv11 model reaches 95.7% mean average precision on mining safety data
• SimAM attention module boosts feature saliency for complex underground scenes
• K‑means++ anchor selection improves detection accuracy and training efficiency
• Dual YOLOv11‑Pose integration enables real‑time unsafe‑behavior alerts
• Study provides a new dataset for miner behavior classification

Pulse Analysis

Underground coal mining remains one of the most hazardous occupations, with accidents often stemming from momentary lapses in worker behavior. Traditional safety oversight relies on manual video review or periodic inspections, which are labor‑intensive, subjective, and prone to blind spots in the labyrinthine tunnels. As sensor networks proliferate, the industry is turning to computer vision to provide continuous, objective monitoring. However, the low‑light, dust‑filled conditions and frequent occlusions in mines have limited the effectiveness of conventional deep‑learning models, creating a clear demand for a more robust solution.

The study published in Scientific Reports tackles these obstacles by extending the YOLOv11 detector with a parameter‑free SimAM attention module and a feature‑enhancement block that prunes noisy channels. Anchor boxes are generated via K‑means++ clustering, which diversifies initial centroids and sharpens localization in cluttered scenes. A dual‑model pipeline fuses YOLOv11’s object detection with YOLOv11‑Pose’s skeletal tracking, delivering precise identification of three behavior categories—area, action, and item types. On a custom‑built dataset of underground footage, the system recorded 95.7% mean average precision, surpassing prior CNN and SVM baselines.

The practical impact of such a system extends beyond a single mine. Real‑time alerts can trigger automated ventilation adjustments, equipment shutdowns, or dispatch of rescue crews before a hazardous act escalates into a fatal incident. For operators, the high accuracy reduces false alarms, preserving workforce trust and avoiding costly production interruptions. Moreover, the publicly released dataset and open‑source code provide a foundation for further research in other high‑risk sectors such as construction and oil‑gas. As AI‑driven safety tools mature, regulators are likely to incorporate them into compliance frameworks, reshaping industry standards.