A Novel Bearing Fault Data Generation Strategy Combining Physical Modeling and CycleGAN Variant for Fault Diagnosis Without Real Samples

The number and quality of fault samples significantly impact the effectiveness of mechanical component fault diagnosis. However, fault data are scarce in real scenarios due to the high cost and risk of obtaining such samples. In addition, collected data often exhibit distribution bias and are therefore unsuitable for real engineering applications. This article proposes a novel fault data generation strategy combining the physical model and the cycle generative adversarial network (CycleGAN) variant for bearing fault diagnosis. First, a bearing dynamics model is developed to obtain sufficient labeled fault simulation samples as the source domain. Healthy operational signals in real scenarios are then collected as the target domain. After that, a CycleGAN variant is designed to minimize the gap between simulation and reality. Specifically, the generator is built upon bidirectional long short-term memory (LSTM) to capture the temporal dependencies of vibration signals. Meanwhile, a multihead self-attention mechanism is introduced to enhance multiscale feature extraction and reduce computational complexity. Moreover, auxiliary classification labels are embedded to generate type-specific fault data. Finally, the proposed method achieved satisfactory performance in both nontransfer and transfer diagnosis scenarios, demonstrating its potential as a reliable solution when real fault data are unavailable.