A Deep Learning-Based Method for Conveyor Belt Deviation Detection with Geometric Constraints

Abstract Belt conveyors are widely utilized in material handling applications because of their high efficiency and substantial capacity. However, belt conveyor deviation typically leads to transmission system failures, which in turn impact production efficiency and may even cause serious safety incidents. Traditional sensor-based detection methods are sensitive to environmental noise, lighting changes, and material interference. Deep learning approaches face challenges such as high computational complexity, limited edge localization accuracy, and baseline drift due to camera position shifts. This paper proposes a deviation detection method that integrates deep learning with geometric constraints. The DC-YOLO-seg model, based on enhanced YOLOv11-seg, is combined with deformable convnets v3 (DCNv3) and coordinate attention (CA) mechanisms for high-precision instance segmentation of conveyor belts and rollers. Subsequently, the belt centerline was extracted using the Canny edge detection algorithm and random sample consensus (RANSAC) fitting method. The drive centerline was estimated based on the alignment of the roller centers, thereby quantifying the offset distance. This approach effectively reduces dependency on camera position and minimizes environmental interference. Experimental results on a self-constructed dataset demonstrate that DC-YOLO-seg achieves a Mask (mAP50-95) of 0.948, representing a 2.5% improvement over baseline models. The deviation detection error is generally maintained within 5 mm. This research provides a robust solution for intelligent operation and maintenance, establishing a foundation for cross-scenario generalization and real-time deployment.