Abstract
<title>Abstract</title> <p>Robust bearing fault diagnosis under noise interference remains challenging for lightweight industrial deployment. Existing deep models trained on clean vibration data often suffer accuracy degradation when fault-related spectral energy in the time–frequency representation is buried by broadband noise. In this paper, a lightweight Frequency-Sensitive Multi-Scale Network (FSMSN) is proposed to address this issue by exploiting the spectral distribution of bearing fault characteristic frequencies. The proposed pipeline first converts raw vibration segments into short-time Fourier transform (STFT) spectrograms with sufficient frequency resolution to distinguish the characteristic frequencies, after which a Frequency-Sensitive Multi-Scale Depthwise Convolution (FSMSDC) module captures transient energy patterns of varying widths, and a Prior-guided Frequency-Aware Channel Attention (PFACA) module reweights feature channels along the frequency axis by injecting the theoretical ball-pass frequency inner/outer races, ball spin frequency and fundamental train frequency as soft physical priors. The prior is class-agnostic and serves as a frequency-band emphasis rather than a class-specific discriminator. The model contains only 0.38 million parameters and can be trained on a single consumer-grade GPU. The method is evaluated on the CWRU and Paderborn bearing datasets against ten baselines including 1D CNN, WDCNN, ResNet18, DRSN, MobileNetV3, TCN and ViT under clean, additive-Gaussian-noise (AWGN), impulsive-noise and harmonicinterference conditions at signal-to-noise ratios from 4 dB to −4 dB. On the CWRU mixed-load protocol, FSMSN achieves 97.18% accuracy, outperforming DRSN by 1.24 percentage points while using 2× fewer parameters. At −4 dB SNR, the margin over DRSN widens to 4.69 pp. On the Paderborn bearingdisjoint protocol with unseen bearing identities, FSMSN achieves 85.27%. A prior sensitivity experiment confirms that correct physical alignment outperforms shifted, random and flat priors, ruling out the hypothesis that the gain is due to generic low-frequency amplification. Attention type comparison, noisespecific ablation, structured-noise evaluation, cross-condition generalization and complexity analyses further verify the contribution of each component. The proposed design delivers improved noise robustness while maintaining a substantially smaller footprint than ResNet- and Transformer-based baselines, providing a practical candidate for edge-deployed industrial vibration monitoring.</p>