A novel lightweight DDPM-based data augmentation method for rotating machinery fault diagnosis with small sample

Abstract

Efficient and intelligent fault diagnosis of rotating machinery is crucial for ensuring the safety and reliability of industrial systems. In practical engineering, data collected from mechanical equipment is often scarce, making accurate fault identification under small sample conditions challenging. Denoising diffusion probability model (DDPM), as a new paradigm of generative models, is widely used for data generation. However, DDPM involves multiple reverse diffusion steps in the generation process, which requires substantial computational resources. To address this issue, a novel lightweight DDPM is proposed for generating fault signals in limited sample scenarios. This method incorporates multi-dconv head transposed attention (MHTA) into the U-Net architecture, shifting attention calculations from the pixel dimension to the channel dimension by combining point-wise and depth-wise convolution, thereby significantly reducing computational complexity. Meanwhile, this method enables the model to capture global context information while also focusing on local details, enhancing the model’s representational capability. Additionally, a composite indicator is developed to evaluate the quality of the synthesized signals across multiple feature dimensions. The effectiveness of the proposed MHTA-DDPM is validated through three cases: simulated signals, real gear and bearing fault signals. The results indicate that the proposed model can generate high-quality and diverse fault signals with good generalization capability, improving fault diagnosis accuracy even under limited sample conditions.