Few-Shot Bearing Fault Diagnosis Based on Finite Impulse Response Generative Adversarial Networks With Frequency-Domain Loss Constraints

The data augmentation quality of vibration signals has a great impact on the results of few-shot bearing fault diagnosis. However, existing vibration signal generation methods have problems such as time domain information loss, instability, and difficulty in maintaining the physical consistency of frequency components. In order to meet the challenge of generating high-quality vibration signals, this paper proposes a finite impulse response generative adversarial networks with frequency-domain loss constraints (FIR-FreqGAN), which introduces a frequency loss constraint based on physical information to improve the quality of the generated signal. At the same time, a physically meaningful finite impulse response convolution kernel is added to the discriminator for enhancement, thereby improving its ability to identify fault-sensitive features. Extensive experiments were conducted on the Case Western Reserve University (CWRU) dataset and a self-collected bearing fault dataset under various imbalance ratios, demonstrating that FIR-FreqGAN produces high-fidelity signals and significantly improves fault classification accuracy over conventional oversampling and state-of-the-art GAN-based methods. In addition, an autoencoder-based reconstruction experiment was performed on the MFPT dataset, which contains unlabeled real-world bearing signals. The results show that signals generated by FIR-FreqGAN achieve lower reconstruction loss than those generated by baseline models, further confirming its generalization ability and frequency-domain consistency under realistic conditions.