Fault diagnosis for driving motor with insufficient fault data: a data transfer generation method

Data-driven diagnostic methods are beneficial for fault diagnosis in driving motors. However, insufficient monitoring data during actual diagnosis limit their application. Although various methods have been proposed, they have limitations. This paper proposes a data transfer generation method based on the distribution difference metric and residual cycle-consistent generative adversarial network (Res-CycleGAN) to overcome these limitations. First, a data layer is designed to select the most similar data from other similar datasets. Then, a Res-CycleGAN-based model layer that can generate high-quality target fault data by utilizing fault features from the selected most similar data is constructed. Finally, a strategy layer for the diagnostic model is proposed to make rational use of different fault data. The proposed method is verified using a motor dataset from a mechanical fault simulation (MFS) platform. A set of comprehensive verification experiments is designed for each layer. The results demonstrate that the proposed method is highly effective.