Domain Adaptation from Simulation to Reality: A GAN- and MK-MMD-Based Transfer Learning Approach for Bearing Fault Diagnosis

Rolling bearings are critical components in industrial machinery, and their failures can lead to equipment downtime or safety hazards, making accurate fault diagnosis vital. While data-driven intelligent methods perform well with sufficient labeled data, acquiring large-scale fault data in real-world scenarios remains challenging. To address this issue, this paper proposes a fault diagnosis method combining finite element simulation and deep domain adaptation transfer learning. First, a finite element model of rolling bearings under normal, outer race, inner race, and rolling element fault conditions is developed, and ANSYS/LS-DYNA simulates motion to generate labeled synthetic fault data. The model’s reliability is validated through time-domain, frequency-domain, and time-frequency analyses. A lightweight 1D convolutional neural network (1D CNN) is then designed for fault diagnosis. When trained solely on simulated data, the model achieves only 61.4% accuracy on real data due to domain discrepancies. To bridge this gap, a transfer learning approach integrating generative adversarial networks (GANs) and multi-kernel maximum mean discrepancy (MK-MMD) is proposed: GANs synthesize data resembling real distributions, while MK-MMD minimizes domain shifts between simulated and actual data. This improves the model’s accuracy to 93.8% on real fault datasets. Performance evaluation under variable working conditions and bearing types demonstrates the method’s robustness, providing a practical solution for fault diagnosis in industrial applications with limited data.