Lightweight physical information diffusion model for fault diagnosis of rolling bearings with imbalanced sample

Abstract Intelligent fault diagnosis of rolling bearings under data imbalance remains a critical challenge in industrial environments. A lightweight physical information diffusion model (LPIDM) is proposed to address the scarcity and imbalanced distribution of fault samples. Firstly, a region-adaptive noise schedule is introduced to replace the conventional linear schedule, enabling targeted augmentation of fault-relevant regions. Secondly, a depthwise separable residual (DSR) structure is incorporated into the U-Net architecture to reduce model complexity and number of parameters. Finally, a multi-objective, collaborative optimization loss function is designed to improve time–frequency fidelity of generated signals. The performance of the proposed method is evaluated through experiments on public and laboratory bearing datasets. The results demonstrate that LPIDM can generate high-quality fault samples, improve diagnostic accuracy and effectiveness under imbalanced conditions, and offer a practical solution for intelligent fault diagnosis in actual industrial scenarios.