Abstract
Rolling bearing fault diagnosis under limited labeled data remains a challenging problem in industrial vibration monitoring. Traditional deep learning methods often suffer from performance degradation when training samples are insufficient or noisy. To address this issue, this paper proposes a multi-stage diffusion-based vibration signal representation method for few-shot rolling bearing fault diagnosis. The proposed method leverages a diffusion restoration process to progressively enhance fault-related components in vibration signals, enabling more discriminative feature extraction under data scarcity conditions. A multi-stage representation strategy is further designed to capture complementary information across different diffusion steps, where early stages preserve global signal structure and later stages emphasize local fault characteristics. Finally, a lightweight classifier is employed for fault identification based on the fused representations. Extensive experiments conducted on benchmark bearing datasets demonstrate that the proposed method achieves superior diagnostic performance compared with several state-of-the-art methods, especially under few-shot settings. The results verify the effectiveness of diffusion-based signal representation for robust fault diagnosis in complex industrial environments.