Interpretable Long-FIR Bi-Mamba Network for Noise-Robust Bearing Fault Diagnosis

Abstract

\begin{abstract} Reliable bearing fault diagnosis from vibration measurements remains challenging when fault-induced impulses are weak and measured signals are contaminated by severe noise. Existing methods often improve diagnostic accuracy through time--frequency image transformation, synthetic sample generation, graph construction, or few-shot adaptation, but their reliability boundaries under explicitly controlled low signal-to-noise ratio (SNR) conditions are not always quantified. This paper proposes an interpretable long finite-impulse-response (FIR) hybrid Bi-Mamba network for noise-robust bearing fault diagnosis directly from raw one-dimensional vibration measurements. The model integrates a learnable long-FIR measurement front end, a dual-branch local--global sequence encoder, and cross-gated fusion. The long-FIR stem acts as an adaptive filter bank whose frequency responses can be visualized, providing an interpretable link between data-driven diagnosis and measurement signal processing. Multi-view representation construction and prototype regularization are further used to enhance feature organization and class separability. Experiments are conducted on CWRU, PU, and HUST-70Hz datasets. The proposed method achieves clean-condition accuracies of 100.00\%, 98.24\%, and 100.00\%, respectively. Under severe noise, it maintains 91.39\% accuracy at $-2$ dB on PU, 87.94\% at $-6$ dB on CWRU, and 94.44\% at $-6$ dB on HUST-70Hz. Same-protocol reproduced comparisons show that the proposed method obtains the best average low-SNR accuracy while using raw vibration signals without DFT-based multi-channel preprocessing, time--frequency image conversion, or synthetic fault generation. Additional limited-sample, non-Gaussian noise, FIR sensitivity, ablation, and complexity analyses further demonstrate its robustness, interpretability, and practical feasibility for low-SNR vibration measurement diagnosis. \end{abstract}