Abstract: To achieve the precise regulation of air flow in mine ventilation systems and the abnormal identification of air window opening, as well as to improve the intelligent level of mines, it is necessary to conduct real-time and accurate detection of underground air window opening. Aiming at the problems of reduced object detection accuracy caused by low illumination and shadow interference in underground mines, as well as the difficulty in deployment and real-time operation of existing detection models due to large parameters and computational complexity, this paper proposes a real-time and accurate detection model CFW-YOLO (C2f_Faster and WIoUv3 Enhanced YOLO) for underground air window opening based on the improved YOLOv8n. With high precision and lightweight as objectives, the CFW-YOLO model replaces the original C2f structure in the backbone network with the C2f_Faster module, which ensures that the network can effectively extract features from input images during high-speed operation, improves detection accuracy while reducing the number of parameters and computational complexity. In the training phase, the WIoUv3 (Wise Intersection over Union version 3) loss function is adopted to replace the CIoU (Complete Intersection over Union) loss function. The learning weights of samples with different quality are balanced through a dynamic non-monotonic mechanism and gradient gain allocation strategy, reducing the influence of harmful gradients and improving the detection accuracy. Images of air windows with different openings were collected and labeled on the experimental platform, and the dataset covers 41 types of air window openings ranging from 0 to 40 cm. Data augmentation was realized by adjusting illumination and adding random shadows, which enhances the generalization ability of the model under complex lighting conditions. The air window opening detection model was trained, and the independent contributions of the two improvements were verified through ablation experiments and comparative tests. A hardware and software system architecture for adjusting air doors integrated with CFW-YOLO model was designed, and the system composition and workflow were analyzed from four aspects: perception, early warning, control, and human-computer interaction. The results show that compared with the original YOLOv8n model, the precision and mean average precision of CFW-YOLO are increased by 3.4% and 0.7%, respectively, and the number of parameters, floating-point operations, and model size are reduced by 0.47 M, 1.2 G, and 0.9 MB, respectively, verifying the accuracy and real-time performance of the improved model.