【摘要】 由于井下带式输送机输送带纵向撕裂故障检测往往无法兼顾检测速度与检测精度，当满足较高的检测精度时，可能导致撕裂未被及时发现而卷入滚筒或托辊中，引发次生安全事故；当满足较快的检测速度时，容易误报触发不必要的停机和未能识别真实撕裂，导致故障潜伏发展，最终引发灾难性破坏，所以文章提出了一种基于改进YOLOv11的AFA-YOLO深度学习算法。通过引入下采样模块(ADown)、共享卷积特征金字塔(FPSC)和辅助检测头(Aux)。在模拟矿井下的输送带纵撕数据集上进行实验，结果表明，AFA-YOLO模型的mAP@.5值达到96.30%,相比YOLOv11提升了3.55%,计算复杂度(GFLOPs)降低了17.19%。AFA-YOLO模型在精度和速度上能够实现最优平衡，满足井下复杂环境的检测需求。更多还原

【Abstract】 Longitudinal tear detection in underground belt conveyor belts often faces a trade-off between detection speed and accuracy. Achieving high detection accuracy may result in tears not being identified in time, leading to entanglement with pulleys or idlers and causing secondary safety incidents. Conversely, prioritizing fast detection can easily trigger false alarms, causing unnecessary shutdowns or failing to identify actual tears, allowing faults to develop latently and ultimately lead to catastrophic failure. Therefore, an AFA-YOLO deep learning algorithm based on an improved YOLOv11 is proposed. This method introduces a downsampling module(ADown), a shared convolutional feature pyramid(FPSC), and an auxiliary detection head(Aux). Experiments conducted on a simulated underground conveyor belt longitudinal tear dataset show that the AFA-YOLO model achieves an mAP@0.5 of 96.30%, a 3.55% improvement over YOLOv11, while reducing computational complexity(GFLOPs) by 17.19%. The AFA-YOLO model achieves an optimal balance between accuracy and speed, meeting the detection requirements of complex underground environments.更多还原