A dual-channel attention and feature selection-based method for rolling bearing fault diagnosis

Abstract

Rolling bearings are critical to the safe and reliable operation of rotating machinery and to predictive maintenance. However, existing fault diagnosis methods still face challenges in insufficient feature representation, inadequate heterogeneous information fusion, and limited diagnostic accuracy and stability. To address these issues, a dual-channel attention and feature selection-based fault diagnosis method for rolling bearings is proposed, and a corresponding digital twin-based interaction system is further developed. First, a hierarchical bearing digital twin architecture composed of the physical layer, data layer, model layer, function layer, and application layer is established. At the model layer, a signal channel based on multi-scale sample entropy and bidirectional long short-term memory (BILSTM) is designed to extract discriminative features from raw vibration signals, while an image channel based on Gramian angular difference field, convolutional neural network, and BILSTM is constructed to learn complementary spatial representations from two-dimensional signal images. Subsequently, multi-head self-attention (MHSA) is introduced to adaptively weight and fuse heterogeneous dual-channel features, and a Pearson-correlation-based threshold filtering strategy is employed to remove highly redundant dimensions before fault prediction. Based on the Case Western Reserve University bearing fault dataset, the performance of the proposed model was comprehensively evaluated through model training, t-SNE visualization analysis, threshold parameter sensitivity analysis, feature selection analysis under different operating conditions, comparison with conventional models, and ablation study analysis. Experimental results demonstrate that the proposed method achieves an average diagnostic accuracy of 99.05% across four working conditions. Compared with the best-performing baseline method, the proposed model improves the average diagnostic accuracy by 0.8%. In addition, ablation studies show that removing the MHSA module, the Pearson-correlation-based screening module, and both modules together reduces the average accuracy to 98.61%, 97.99%, and 97.76%, respectively, further confirming the effectiveness of the proposed components.