Enhanced dual-channel feature fusion approach for rolling bearing fault diagnosis

Bearing fault diagnosis models predominantly rely on vibration signals for signal processing. The presence of noise interference and limited feature extraction capacity significantly impedes effective feature representation and compromises model generalisation, consequently leading to suboptimal diagnostic accuracy in real-world operational scenarios. To address these issues, a novel dual-channel bearing vibration signal fault diagnosis method, termed Relative Position Matrix – Swin Transformer + CNN2D combined with Global Attention Mechanism (RSTCG), is proposed in this paper. First, the Relative Position Matrix (RPM) method is used to convert the original temporal sequence into a two-dimensional feature map while preserving its time correlation. Second, A dual-channel network framework is designed to extract features from the two-dimensional feature map, consisting of a lightweight Swin Transformer and a CNN2D network enhanced with a Global Attention Mechanism (GAM). Finally, the features of different scales from the dual-channel outputs are fused to produce the recognition results. Experimental evaluations demonstrate that the proposed method achieves remarkable accuracy rates of 99.83% and 98.85% on the laboratory dataset and the University of Paderborn dataset, respectively. Compared to traditional models, the proposed method effectively extracts bearing fault features, achieving superior accuracy and generalization in noise and small-sample scenarios, demonstrating robustness and adaptability.