Numerical simulation-aided deep convolutional transfer learning networks to reliable detect bearing faults under varying working conditions

Data-driven intelligent bearing fault detection methods are only suitable for close-set detection. For open-set detection, it requires complete labelled training fault samples to preserve the reliability and accuracy of fault detection. Therefore, the improved version of finite element method (FEM) simulation-aided deep convolutional transfer learning network (SA-DCTLN) is developed to achieve agreeable fault classification precision of bearings under varying working conditions. First, FEM is employed to simulate missing fault samples to solve the problem of lacking complete labelled training fault samples. Second, random noises with certain distribution are adopted as inputs of GAN to generate a group of synthetic simulation fault samples with the same fault label. Third, synthetic simulation fault samples are combined to obtain the complete fault samples in the source domain. Finally, SA-DCTLN is applied to reduce the distribution discrepancy between samples under varying working conditions and further identify unknown samples in the target domain to obtain preliminary classification results. Furthermore, the decision-making information is fused using Dempster-Shafer (DS) evidence theory to further improve classification precision. Bearings dataset of open-source repository is utilized to verify the proposed method, and the experimental results demonstrate the reliability and generalizability of the constructed method.