LS-DDPM: A Vibration Signal Generation Method for Small-Sample Bearing Fault Diagnosis

• 1、School of Information and Communication Engineering, Beijing University of Post and Technology, Beijing 100876
• 2、School of Mechanical and Electrical Engineering,Beijing Information Science and Technology University, Beijing 102206

Abstract：Industrial settings commonly face challenges such as scarcity of bearing fault samples, complex distribution of working conditions, and significant cross-scale signal characteristics, making it difficult for data-driven fault diagnosis models to achieve stable and reliable performance. To overcome the shortcomings of existing generative models in modeling long-range harmonic structures, capturing transient impact details, and maintaining training stability, a vibration signal generation method named LS-DDPM is proposed for small-sample scenarios. Based on the stepwise denoising generation mechanism of diffusion models, this method introduces the LSBlock, which incorporates a Large-Kernel Perception (LKP) module and a Small-Kernel Aggregation (SKA) module. An LS-UNet is constructed as the denoising network, enabling the model to simultaneously reconstruct long-range periodic modulation structures and transient impact details in vibration signals. Furthermore, by integrating conditional injection and classifier-free guidance strategies, the method achieves category control during the generation process, thereby producing high-fidelity samples for scarce fault modes in a targeted manner. Experimental results on public bearing datasets PU and CWRU demonstrate that the signals generated by LS-DDPM exhibit high consistency with real samples in terms of time-domain textures, frequency-domain harmonic components, and energy distribution. Relevant statistical metrics significantly outperform mainstream generative models such as DDPM, DCGAN, PDA-WGANGP, and CVAE-GAN. In small-sample fault diagnosis tasks, data augmented with LS-DDPM effectively enhances the classification performance of models like AA-CNN, ISONet, and ResNet. The study indicates that LS-DDPM can effectively recover key engineering characteristics of vibration signals under small-sample conditions, providing a high-fidelity, controllable, and practically applicable data generation solution for rolling bearing fault diagnosis.

Keywords： rolling bearing diffusion model few-shot learning vibration signal generation fault diagnosis