An intelligent detection method for underground conveyor belt deviation based on parameter regression in intelligent manufacturing

Belt deviation is one of the most common faults in belt conveyor systems, and may lead to material spillage, equipment damage, and production interruption, thereby directly affecting industrial safety and efficiency. Existing belt deviation detection methods usually rely on auxiliary references or local area information, making it difficult to balance detection accuracy and real-time performance in industrial environments with dust, uneven lighting, or occluded edges. This paper proposes an end-to-end belt deviation detection method, ATBR-Net (Anisotropic Transformer Belt Regression Network). Specifically, a multidimensional feature enhancement module is designed to adaptively strengthen belt-edge representations, which enables the network to adaptively learn edge features through a dual-domain dynamic calibration mechanism. Second, to exploit the structural continuity of conveyor belt edges, a lightweight context aggregation strategy is introduced. By exploiting the anisotropic dependency characteristics of conveyor belt edges, the proposed strategy reduces computational complexity while improving long-range contextual modeling of continuous and slender structures. Furthermore, conveyor belt deviation detection is formulated as a constrained geometric parameter regression problem, enabling the direct prediction of belt-edge geometric parameters for efficient and accurate deviation determination. We constructed a real industrial image dataset containing 6100 images with edge annotations, providing a standardized benchmark for the development and validation of conveyor belt deviation detection algorithms. Comparative experiments with six mainstream methods on Self-Belt demonstrate that ATBR-Net achieves a favorable balance between detection accuracy and real-time performance, with an accuracy of 95.52% and an inference speed of 149 FPS on an edge device. Further experiments on the CUMT-Belt dataset demonstrate its generalization performance, providing a robust and efficient end-to-end solution for the intelligent monitoring of belt conveyor systems.