针对现有输送带跑偏状态监测方法实用性、鲁棒性不足及数据集制作难度大的问题，提出了一种基于深度学习的输送带跑偏状态智能监测方法。将输送带边缘识别问题看作特定场景下的直线检测问题，提出以目标检测网络预测框的对角线表征输送带边缘直线的检测策略，以预测框的右上−左下对角线表征输送带左边缘，以左上−右下对角线表征输送带右边缘；通过YOLOv5模型对输送带边缘进行检测，并设计了跑偏量计算方法和跑偏状态判定规则。实验结果表明，利用目标检测网络预测框的对角线特征可稳定高效地实现输送带边缘识别和跑偏量量化，简化了图像数据处理流程和数据标注，具有较强的泛化能力和快速迁移学习能力；结合直线检测策略的YOLOv5模型对料流边界、支柱等其他直线的抗干扰能力强，在CUMT−BELT数据集上的检测精度达99%以上，检测速度最快达148帧/s，实时性好。

Existing methods for monitoring conveyor belt misalignment face challenges in terms of practicality, robustness, and the difficulty of dataset creation. This paper proposed an intelligent monitoring method for conveyor belt misalignment based on deep learning. First, the conveyor belt edge recognition problem was treated as a line detection issue in a specific scenario. A strategy was proposed to detect the straight lines of the conveyor belt edges using the diagonal features of the bounding box predicted by the object detection network. Specifically, the top-right to bottom-left diagonal of the predicted bounding box was used to represent the left edge of the conveyor belt, and the top-left to bottom-right diagonal was used to represent the right edge. The YOLOv5 model was employed to detect the conveyor belt edges, and a misalignment calculation method and misalignment state determination rules were developed. Experimental results demonstrated that the diagonal features of the predicted bounding box could stably and efficiently achieve conveyor belt edge recognition and misalignment quantification, thereby simplifying image data processing and annotation tasks. The method exhibited strong generalization ability and rapid transfer learning capability. The YOLOv5 model, combined with the line detection strategy, showed excellent anti-interference performance for detecting material flow boundaries, support pillars, and other straight lines. On the CUMT-BELT dataset, the detection accuracy exceeded 99%, with a maximum detection speed of 148 frames per second, ensuring excellent real-time performance.