A Bearing Cross-Domain Fault Diagnosis Method Based on Dynamic Simulation and an Improved Domain Adversarial Network
In practical industrial scenarios, high-quality bearing fault samples are often difficult to obtain. Although simulation methods can generate a certain amount of data, obvious distribution shifts still exist between simulated and measured signals due to model simplification, parameter uncertainty, and differences in noise structures, which limit their direct application in measured scenarios. To address this issue, this article proposes a cross-domain bearing fault diagnosis method based on dynamic simulation and a multiscale beta-distribution-guided domain-adversarial neural network (MBDANN), denoted as Sim2Real-MBDANN. First, a rolling bearing dynamic model is established based on actual operating parameters, and simulation is used to generate data samples covering different operating states, thereby alleviating the shortage of real fault data and the difficulty of annotation. Second, the MBDANN model is constructed using simulated data to achieve feature transfer and knowledge alignment from the simulation domain to the measured domain. The network adopts a feature extractor that integrates multiscale convolution kernels and a temporal attention mechanism to enhance the representation capability of multilevel fault features; meanwhile, a Beta-distribution-based feature alignment mechanism is introduced to improve the stability of adversarial alignment for marginal distributions, and a dynamically threshold-driven local maximum mean discrepancy (LMMD) loss is designed to improve the reliability of class-conditional distribution alignment. Finally, experimental results on the CWRU public dataset and a self-built experimental dataset show that, using only simulated signals as the source domain, the proposed method achieves average diagnostic accuracies of 94.71 % and 87.37 %, respectively, demonstrating its good generalization performance and cross-domain diagnosis capability.