Data scarcity remains a major obstacle to reliable fault diagnosis in industrial applications, where vibration signals are often affected by background noise. Although generative data augmentation provides a practical approach to addressing the scarcity of fault samples, existing methods still exhibit notable limitations. Adversarial models often suffer from unstable training and mode collapse, while diffusion-based approaches typically rely on costly iterative sampling. The performance of existing generative methods may further deteriorate in the presence of industrial noise. To overcome these limitations, a Laplace-Guided Dual-Branch Flow Matching framework (Laplace-FM) is proposed for few-shot fault data generation. The framework employs conditional Flow Matching to learn a continuous transport process from Gaussian noise to the real fault data distribution, thereby providing stable training and efficient sample generation. To improve feature extraction from industrial vibration signals, a dual-branch structure is incorporated into a one-dimensional U-Net. One branch directly processes the raw signal to retain complete waveform information, while the other applies a learnable Laplace kernel with damped-oscillatory characteristics to emphasize fault-related impulsive and resonant responses while reducing the influence of unstructured noise components. Experiments conducted on real-world industrial data demonstrate that Laplace-FM can generate realistic fault samples and provide effective data augmentation for downstream diagnosis. The results further show that the proposed framework achieves favorable performance in both generation efficiency and fault information preservation.

Flow Matching,Limited data,Data generation,GAN,DDPM,Fault diagnosis