SDMT-Net: a hybrid Spiral Dual Mamba and Transformer model for noise-robust rolling bearing fault diagnosis

Abstract Rolling bearing fault diagnosis, a core technology for the intelligent maintenance of mechanical equipment, faces severe challenges in accurately extracting fault features and processing non-stationary signals under strong background noise. Existing deep learning methods often struggle to balance high-resolution detail capture with long-range global dependency modeling, thereby limiting their generalization ability under complex operating conditions. To surmount these bottlenecks, this paper innovatively proposes SDMT-Net, a novel architecture that deeply integrates the content-aware characteristics of Mamba with the global attention mechanism of Transformers. The core Spiral Dual Mamba module employs a spiral bidirectional scanning strategy to repair spatial topological discontinuities and construct a non-causal analytical perspective, facilitating the precise capture of both fault onset transients and decay textures. Concurrently, a Spatially Enhanced Mamba submodule incorporates physical priors from both time and frequency domains for dynamic gating, significantly bolstering feature robustness under low signal-to-noise ratio conditions. Furthermore, the Transformer component introduces a spatial interaction block to further optimize spatiotemporal feature interactions. Experimental validation demonstrates that SDMT-Net achieves an average fault classification accuracy of 98.8% ± 0.3% across datasets from Jiangnan University, the University of Ottawa, and a self-built laboratory dataset. In environments simulating real-world industrial scenarios characterized by mixed Gaussian and impulsive noise, the proposed model outperforms state-of-the-art benchmark models, including Swin-FFRN, BDMamba, Diagnosisformer, BMTM-Net, and RSMamba, yielding a performance improvement of 0.8% to 4.5%. Comprehensive ablation studies and complexity analyses further verify the theoretical contributions of each module and the optimal trade-off between computational efficiency and diagnostic performance.