A data generation method with a novel loss for fault diagnosis without target-domain data

Data-driven bearing fault diagnosis methods have achieved significant success under single working conditions, where training and testing data follow similar distributions. However, changes in bearing working conditions often cause data distribution shifts, leading to significant performance degradation. Although transfer learning methods have been widely studied to address this issue, they typically rely on abundant and high-quality target-domain data, which are difficult to obtain in practice. To overcome this limitation, this paper proposes a data generation method based on a residual autoencoder with a novel physics-consistent loss function. The proposed loss simultaneously accounts for data distribution similarity, alignment of fault characteristic frequencies, and amplitude variations, guiding the model to learn a simulation-to-reality mapping that preserves essential physical properties. Based on data from a single working condition, the proposed method can generate fault frequency spectra under unseen conditions. Experimental results on a benchmark dataset demonstrate that the generated signals effectively reproduce the physical characteristics of vibration responses under new working conditions, and using these signals for fault diagnosis achieves excellent identification performance.