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

Area of Science:

• Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Few-shot fault diagnosis for intelligent high-speed trains is hindered by scarce and imbalanced fault data.
• Existing methods struggle with generating high-fidelity vibration signals for limited fault samples.

Purpose of the Study:

• To propose MR-DDIM, a class-conditional diffusion-based data augmentation framework.
• To generate high-fidelity fault vibration signals from limited labeled data for improved fault diagnosis.

Main Methods:

• Developed a WT-UNet denoising backbone integrating 1D wavelet convolution and Feature-Wise Linear Modulation (FiLM).
• Incorporated log-σ regularization and multi-resolution STFT consistency loss for training stability and spectral fidelity.
• Introduced multi-resolution spectral correlation coefficient (MR-SCC) and class-intrinsic maximum mean discrepancy (cMMD) for quality evaluation.

Main Results:

• MR-DDIM successfully generated fault samples with high spectral consistency and intra-class diversity.
• The generated data significantly improved the robustness of downstream few-shot fault diagnosis models.
• Experimental validation on BJTU-RAO datasets confirmed the method's effectiveness.

Conclusions:

• MR-DDIM offers an effective data augmentation solution for intelligent fault diagnosis in high-speed railway systems.
• The framework addresses the challenge of limited and imbalanced fault data in critical infrastructure monitoring.