A lightweight multi-scale hybrid routing-based dual-attention network for rolling bearing fault diagnosis under strong noise

Abstract Under strong noise, fault features of rolling bearings are easily submerged. This makes reliable diagnosis difficult, while balancing noise robustness and model compactness remains challenging. To address these issues, this paper proposes a lightweight multi-scale hybrid routing-based dual-attention network, termed MS-HRDNet. A lightweight multi-scale input module with branch-wise scale calibration is first designed to stabilize shallow feature extraction in noisy environments. Then, a routing-based multi-group depthwise convolution module is introduced. It combines sample-adaptive sparse routing with routing-weighted neighboring-group response aggregation to enhance information interaction while preserving lightweight characteristics. In addition, an adaptive multi-feature output fusion module dynamically integrates shallow and deep features by modeling their discrepancy and complementarity. Experiments on the public CWRU dataset and a self-collected WHUT dataset verify the effectiveness of the proposed method. At the strongest noise level of -8 dB, MS-HRDNet achieves a diagnostic accuracy of 90.20 ± 1.04% on the CWRU dataset and 89.35 ± 1.93% on the WHUT dataset, demonstrating strong robustness under both white-noise and colored-noise interference.