DPFM‐Mamba: Weak Fault‐Related Response Enhancement and Preservation for Limited‐Sample Rolling Bearing Fault Diagnosis

ABSTRACT Limited‐sample rolling bearing fault diagnosis is challenging because only a small number of labeled fault samples are available for model training. Local transient impulses and subtle fault‐induced variations can be obscured by background fluctuations and redundant components, while these responses may be further attenuated during feature propagation. To improve the utilization of such fault‐related information, this paper proposes DPFM‐Mamba. Before temporal modeling, the Dual‐Path Feature Modulation module (DPFM) constructs a local trend reference and its corresponding variation component to jointly generate adaptive modulation weights, thereby strengthening insufficiently salient fault‐related responses. The modulated features are then processed by Mamba, whose input‐dependent selective state space mechanism captures temporal dependencies across different feature patches with linear complexity. During temporal downsampling and channel mapping, the Residual‐Preserved Transition Module (RPTM) combines parallel transition and residual paths with channel‐wise adaptive fusion to reduce response attenuation, after which one‐dimensional deformable convolution captures local abnormal responses with temporal shifts. Experiments on the Jiangnan University bearing dataset show that DPFM‐Mamba achieves an average diagnostic accuracy of 93.47% at a 10% training‐sample ratio, outperforming all compared methods.