Conveyor Belt Deviation Detection for Mineral Mining Applications Based on Attention Mechanism and Boundary Constraints

To address the issue of material spillage and equipment wear caused by conveyor belt deviation in complex industrial scenarios, this study proposes a detection method based on an improved U-Net. The approach adopts U-Net as the backbone network, with a ResNet34 encoder to enhance feature extraction capability. At the skip connections, a Multi-scale Adaptive Guidance Attention (MASAG) module is embedded to strengthen the fusion of semantic and detailed features. In the loss function design, a boundary loss is incorporated to improve edge segmentation accuracy. Furthermore, the segmentation results are refined via edge detection and RANSAC regression, and a reference line is constructed based on the physical stability of rollers in the image to enable quantitative measurement of deviation. Experiments on a self-constructed dataset demonstrate that the proposed method achieves higher accuracy (99.77%) compared with the baseline U-Net (99.65%) and also surpasses other categories of approaches, including detection-based (YOLOv5s), anchor-point-based (UFLD), and segmentation-based approaches represented by SEU-Net and DeepLabV3+, thereby exhibiting strong robustness and real-time performance across diverse complex operating conditions. The results validate the effectiveness of this method in practical applications and provide a reliable technical pathway for the development of intelligent monitoring systems for mining conveyor belts.