Abstract Rolling bearings are critical supporting components in rotating machinery, whose health conditions directly affect system reliability and operational safety. However, under practical industrial conditions, bearing vibration signals are often characterized by strong non-stationarity, low signal-to-noise ratios, and complex time–frequency coupling, posing significant challenges to robust feature extraction and long-range dependency modeling. To address these issues, this paper proposes a learnable rational-dilation wavelet enhanced Mamba network (RDEMamba), which achieves hierarchical feature extraction and fusion by stacking multiple dual-stream modeling (DSM) modules. Each DSM module consists of a multi-RDWT Mamba (MRM) block and a large-kernel depthwise separable (LDS) block. Specifically, the MRM block integrates a learnable rational-dilation wavelet transform with heterogeneous wavelet bases to capture fault-induced transient impulses from multiple time–frequency perspectives, while the Mamba operator enables efficient local-nonlocal sequence modeling. In parallel, the LDS block employs LDS convolutions to expand the temporal receptive field with low computational cost, complementing the MRM branch in modeling long-range dependencies. Furthermore, a gated multi-head self-attention block is introduced at the network tail to adaptively recalibrate the fused deep features, suppressing redundant responses and enhancing fault-relevant patterns. Extensive experiments on multiple public bearing datasets demonstrate that RDEMamba achieves superior fault diagnosis performance under complex operating conditions, with strong robustness against signal disturbances.