Transformer-based conditional generative transfer learning network for cross domain fault diagnosis under limited data

Most current generative adversarial network (GAN) cannot simultaneously consider the quality and diversity of generated samples due to limited data and variable working condition. To solve the problem, a Transformer-based conditional GAN transfer learning network is proposed. Firstly, a transformer-based conditional GAN (TCGAN) generative network is constructed with sample label information, enhancing the quality of generated data while retaining the diversity of generated signals. Secondly, a transfer learning network based on TCGAN is established, and a "generation-transfer" collaborative training strategy based on the expectation maximization is introduced to realize parallel updating of the parameters of the generative network and the transfer network. Finally, the effectiveness of the proposed method is verified using bearing datasets from CWRU and the self-made KUST-SY. The results show that the proposed method can generate higher quality data than comparative methods such as TTS-GAN and CorGAN, which provides a new solution for improving the cross-domain fault diagnosis performance.