DS-GCMAF: a dual-stream gated cross-modal attention fusion network for robust bearing fault type and severity classification under complex noise
TL;DR
A Dual-Stream Gated Cross-Modal Attention Fusion Network (DS-GCMAF), which simultaneously processes one-dimensional raw vibration sequences and two-dimensional time–frequency representations, and demonstrates the superior robustness and its effectiveness in joint fault-type and severity classification under highly noisy environments.
Abstract
Accurate assessment of fault severity in rolling bearings under strong noise remains a critical challenge for intelligent predictive maintenance. In this study, the diagnostic task is formulated as joint fault type and severity classification, with emphasis on severity-level discrimination. In practical industrial environments, vibration signals are frequently contaminated by complex noise, including Gaussian noise, pink noise, Laplace (impulsive) noise, and combined interference, which severely mask transient fault impulses and distort time–frequency representations. To tackle this problem, we propose a Dual-Stream Gated Cross-Modal Attention Fusion Network (DS-GCMAF). The framework simultaneously processes one-dimensional (1D) raw vibration sequences and two-dimensional (2D) time–frequency representations. A bidirectional cross-modal multi-head attention mechanism is introduced to facilitate effective information interaction and alignment across heterogeneous feature spaces. Meanwhile, a quality-aware adaptive gating strategy is employed to dynamically regulate the contribution of each modality based on its reliability. In low signal-to-noise ratio (SNR) conditions, the less reliable 2D branch is selectively attenuated, and the more robust 1D stream serves as an anchor for cross-modal calibration. Extensive experiments were carried out on the Paderborn University (PU) dataset and HUSTbearing dataset under four complex noise types. On the PU dataset across varying SNR levels (−8 dB to 8 dB), DS-GCMAF achieves 85.80% accuracy at −8 dB, surpassing the strongest baseline by 2.69 percentage points. On the HUSTbearing dataset under −10 dB, the proposed method attains 74.04% accuracy under the most challenging combined noise condition, significantly outperforming existing single-stream and conventional fusion approaches. These results demonstrate the superior robustness and its effectiveness in joint fault-type and severity classification under highly noisy environments.