Due to the longitudinal tear fault detection of the conveyor belt of the underground belt conveyor, it is often impossible to take into account the detection speed and detection accuracy. When the detection accuracy is high, it may lead to tears that are not detected in time and are rolled into the roller or roller, causing secondary safety accidents. When fast detection speeds are met, false alarms are prone to trigger unnecessary downtime and failure to identify true tears, leading to the latent development of faults and ultimately catastrophic damage. Therefore, this paper proposes an AFA-YOLO deep learning algorithm based on the improved YOLOv11. By introducing the Downsampling Module (ADown), the Shared Convolutional Feature Pyramid (FPSC) and the Auxiliary Detection Head (Aux). Experiments on the longitudinal tearing dataset of the conveyor belt simulated underground mine show that the mAP@0.5 value of the AFA-YOLO model reaches 96.30%, which is 3.55% higher than that of YOLOv11, and the computational complexity (GFLOPs) is reduced by 17.19%. The AFA-YOLO model can achieve the optimal balance of accuracy and speed to meet the detection needs of complex downhole environments.