Skip to content
Open access

Interpretable Long-FIR Bi-Mamba network for noise-robust bearing fault diagnosis

Oct 2026 · Measurement science and technology · Vol 37 · 0 citations · 27 references
Physics

Abstract

Reliable bearing fault diagnosis remains difficult when defect-induced impulses are masked by measurement noise. This paper studies reliability under explicitly controlled low signal-to-noise ratio (SNR) conditions. The proposed long finite-impulse-response (FIR) hybrid bidirectional Mamba network operates directly on one-dimensional vibration segments. A learnable FIR bank is modulated by signal-quality descriptors, parallel local convolutional and bidirectional state–space paths extract impulsive and long-range evidence, and a cross gate fuses them. Time, spectral, and envelope views are combined, with prototype regularization used only during training. FIR responses reveal selected carrier bands, whereas bearing characteristic frequencies describe impact repetition or envelope modulation. The method is evaluated independently on Case Western Reserve University, Paderborn University (PU), and the 70 Hz subset of HUSTbearing. On PU, three runs give 98.04%±0.30% clean accuracy and 88.40%±1.28% at −2 dB. In same-protocol comparisons, the proposed method obtains the highest mean accuracy at 0 dB and remains competitive at −2 dB. With 30% of the PU training data, nine runs give 97.62%±0.45% clean accuracy and 87.85%±1.46% at −2 dB. Under colored, impulsive, and mixed disturbances at −2 dB, the proposed method obtains 91.10%±0.74%, 86.72%±1.01%, and 84.21%±1.15%, respectively. PU leave-one-operating-condition-out evaluation yields four-fold mean accuracies of 90.4% clean and 81.8% at −2 dB. Class-conditioned FIR analysis increases the envelope peak-to-background ratio by 4.6–4.7 dB for fault classes; occluding the three most active channels reduces class accuracy by 4.1–5.3 points, compared with 0.8 points for matched control channels. These results support low-SNR and cross-condition reliability on public laboratory datasets while not constituting direct industrial-field validation.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.