Abstract: Ensuring the safe and stable operation of conveyor belts is a critical requirement in modern coal mining, as these systems serve as the core of material transportation. However, the underground mining environment is characterized by low and uneven illumination, severe dust interference, and highly variable visual backgrounds, which pose major challenges for vision−based monitoring. Traditional monitoring methods, relying on manual inspection or sensor−based measurements, suffer from high computational loads, limited real−time capability, and poor adaptability to small foreign objects. To overcome these limitations, this study proposes an intelligent recognition method for conveyor belt operational states based on an improved YOLOv11 algorithm. The improved model adopts the lightweight YOLOv11−n network as its backbone and introduces a Shape−IoU loss function to enhance bounding−box regression accuracy by improving geometric alignment between predicted and ground−truth targets. In addition, standard convolutions in the detection head are replaced with depthwise separable convolutions, thereby reducing the model’s parameter count and computational complexity while maintaining strong feature representation capability. A comprehensive dataset was constructed using images collected from real underground coal mine conveyor systems. The dataset covered typical abnormal operating states, including large gangue, rock bolts, metallic foreign objects, no−load operation, and belt tearing. The model was trained and validated on the enhanced dataset, and the experimental results show that the proposed method achieves a mean Average Precision (mAP) of 0.985 and an average Recall of 0.994. Ablation experiments further demonstrate that the optimization strategies improve the baseline precision from 0.913 to 0.983, confirming the contribution of the Shape−IoU loss and depthwise separable convolution modules. Compared with mainstream lightweight YOLO models such as YOLOv8−n and YOLOv10−n, the proposed method achieves higher Precision and Recall. The method provides a lightweight and efficient solution for real−time monitoring of conveyor belt operational states, offering reliable technical support for intelligent mine safety management and contributing to the development of smart, safe, and automated coal mining systems.