A small sample transfer diagnosis method driven by a high-precision dynamic model for rolling bearings

Although intelligent diagnosis methods based on deep learning have achieved significant theoretical advancements, the application in real-world industrial scenarios remains challenging. The scarcity of fault samples in real-world environments hinders the effective training of deep learning models, thereby compromising their diagnostic accuracy and generalization ability. To address this issue, a small sample transfer diagnosis method driven by a high-precision dynamic model for rolling bearings is proposed this paper. Firstly, a high-precision dynamic model of rolling bearing with defects is constructed to simulate vibration signals under various fault types and severities, generating a large-scale and diverse library of simulated fault samples. Then, an improved Transformer-based deep transfer learning network is developed to extract features from both simulated samples and measured small samples at local and global scales, and perform multi-layer deep domain adaptation to minimize the distribution discrepancy between the simulated and measured data, facilitating accurate fault diagnosis under small sample conditions. Finally, experimental verification on two datasets with different tasks demonstrated that the proposed method not only effectively diagnoses rolling bearing faults but also exhibits excellent generalization ability on small sample datasets.