Dynamics-assisted unsupervised domain adaptation method for rolling bearing fault diagnosis under noisy conditions

Real-time and accurate fault diagnosis of rolling bearings is crucial for the safe operation of rotating machinery. Deep learning technology has gained widespread applications in fault diagnosis due to its capability in vast data analysis and complex nonlinear modeling. However, such data-driven approaches are highly dependent on both data quantity and quality. In practical applications, they suffer from significant performance degradation due to scarce labeled data and noisy measurements. To address these issues, this paper proposes a dynamics-assisted unsupervised domain adaptation method for rolling bearing fault diagnosis. First, a dynamic model of rolling bearings is established and a parameter identification method is developed to determine its critical physical parameters, which enable the generation of high-fidelity labeled fault simulation data. Then, an adversarial unsupervised domain adaptation framework is constructed to mitigate the distribution discrepancy between simulated and measured data. Meanwhile, a deep learning model incorporating a multi-scale mode denoising network and an inverse-embedding cosine similarity attention mechanism is proposed to extract domain-invariant fault features by capturing multi-scale modal characteristics, enhancing fault-related features, and suppressing noise. The effectiveness of the proposed method is validated on two public datasets under various target-domain data availability conditions, achieving average accuracies of 81.09% and 85.00%, respectively, and outperforming the best-performing method among other advanced methods by 2.27% and 3.83%. Under −5 dB noise, the improvements further increase to 13.12% and 11.65%, respectively.