A Diffusion-Based Data Augmentation Framework for Few-Shot Fault Diagnosis of Intelligent High-Speed Train Components

Few-shot fault diagnosis of intelligent high-speed train components remains challenging because fault samples are scarce and highly imbalanced. To address this issue, this paper proposes MR-DDIM, a class-conditional diffusion-based data augmentation framework for generating high-fidelity fault vibration signals from limited labeled data. A WT-UNet denoising backbone is developed by combining one-dimensional wavelet convolution with Feature-Wise Linear Modulation (FiLM) to capture multiscale time-frequency structures and enable class-controllable generation. To improve training stability and spectral fidelity, log-σ regularization and a multi-resolution STFT consistency loss are introduced into the optimization process. In addition, this paper proposed the multi-resolution spectral correlation coefficient (MR-SCC) and class-intrinsic maximum mean discrepancy (cMMD) to evaluate generation quality from spectral and distributional perspectives. Experiments on the BJTU-RAO datasets show that the proposed method can generate fault samples with high spectral consistency and reasonable intra-class diversity, thereby improving the robustness of downstream few-shot fault diagnosis. The results indicate that MR-DDIM provides an effective data augmentation solution for intelligent fault diagnosis in high-speed railway systems.