Dual space data fusion framework for bearing fault diagnosis under imbalanced data

Abstract In practical rotating machinery measurement systems, fault diagnosis models are often limited by imbalanced data and discrepancies between simulated and real sensor signals. Existing dynamic models or generative networks alleviate data scarcity but usually overlook measurement properties, such as preserving fault-related physical features and aligning spectral distributions between simulated and measured signals. To address this issue, this paper proposes a dual-space fusion framework for bearing fault diagnosis under imbalanced data. A bearing dynamic model first generates simulated fault signals in the virtual space, while a simulation-to-real distributional discrepancy extractor extracts background components and frequency distributions from measured signals in the physical space. The simulated signals and extracted background components form the data basis for dual-space fusion, while the frequency distributions provide fault-related prior knowledge to guide the fusion process. Specifically, a frequency prior Chebyshev mask-UNet generates multi-scale masks using frequency priors and Chebyshev polynomial-induced nonlinear responses to highlight critical regions. In addition, a collaborative two-stage structure-frequency constraint further improves training by preserving structural features while promoting information independence. Validation on three bearings demonstrates the framework’s ability to generate low-noise, realistic synthetic fault samples with distributions closely aligned to real data. Experiments show that the framework reduces maximum mean discrepancy by approximately 10 times and improves learned perceptual image patch similarity by up to 0.14. Under noisy conditions, synthetic samples achieve over 24 dB SNR gain. Under a highly imbalanced condition, the proposed method still achieves an accuracy exceeding 95%, underscoring its effectiveness and reliability.