Double-branch cooperative diagnosis method for roller fault of belt conveyor in industrial environment

Abstract The roller of belt conveyor is susceptible to rust, wear, and other faults in long-term operation. The early fault acoustic signal is weak and easily submerged by strong noise in the industrial environment. The traditional diagnosis method has the problems of insufficient feature extraction and poor robustness. An early fault diagnosis method of belt conveyor based on parallel architecture and multi-scale attention fusion (CST, CNN-Swin Transformer) is proposed. A double-branch parallel network is constructed. The CNN branch combines lightweight VGG and channel attention Squeeze-and-Excitation (SE) to extract local features, and the Swin Transformer branch incorporates multi-scale dilated convolution attention module (MDAM) to extract global features. The content guided attention fusion module (CGA) performs double-branch feature fusion. The experiment is based on the acoustic data set of belt conveyor operation collected in a real industrial environment. The results show that the proposed method achieves an average diagnostic accuracy of 98.11% under noise interference, outperforming a single model by 7.88%. The ablation experiment verified that the three-level attention coordination strategy (SE-MDAM-CGA) improved the model accuracy by 3.08%. The proposed Performance-Efficiency Score evaluation index shows that the CST model achieves a balance between computational efficiency and diagnostic accuracy. The noise robustness test demonstrates that the method maintains an accuracy rate above 84% even with additional noise (SNR ⩾ 0 dB), on top of the inherent industrial noise in the signal. Cross-dataset generalization experiments further demonstrate the method’s generalization performance in multi-type fault diagnosis tasks. This method provides a high-precision and high-robustness solution for early weak fault diagnosis of industrial equipment.