Simulation-Driven Subdomain Adaptation Network for bearing fault diagnosis with missing samples

Deep learning has been employed widely in bearing fault diagnosis. However, in industry, sufficient and complete fault samples are hard to acquire and the working conditions vary frequently, which result in significant diagnosis performance deterioration. To solve these problems, a fully dynamic model and subdomain adaptation based intelligent fault diagnosis framework is proposed. Firstly, the fully dynamic bearing models are used to fill in some missing fault samples. Then, a simple but effective subdomain adaptation method is developed to align class-level subdomain distributions of source and target domains with consideration of samples transferability. Contrast experiments on two bearing test rigs are carried out to verify the proposed method, which contain different working conditions and missing fault types. Experimental results indicate that the proposed method is superior to non-adapted and four widely used transfer learning models, which is promising for real industrial applications.