Data Augmentation Fault Diagnosis of Rolling Machinery Using Condition Denoising Diffusion Probabilistic Model and Improved CNN

Imbalanced data present a significant challenge in intelligent fault diagnosis due to limited sample availability, and various generative models have been proposed for data augmentation. However, these training processes are often complex, prone to model collapse, and fail to improve the classifier’s ability to categorize fault signals. Consequently, a bearing fault diagnosis method based on condition denoising diffusion probabilistic model (DDPM) and improved convolution neural network (CNN) is proposed to address class-imbalanced. First, the condition DDPM is introduced to generate high-quality and diverse vibration synthetic data through accurate physical simulation. Second, the synthetic data’s effectiveness is assessed using the proposed improved CNN, which offers enhanced flexibility in perception. Third, the GAN-train/GAN-test is used to provide a comprehensive evaluation method for synthetic data and indicate the diversity and similarity of synthetic sample. Finally, two machinery datasets, five classifier models and seven data augmentation methods are compared under varying imbalance ratios of 1:2, 1:4, and 1:10. These results confirm the sample generation capabilities of condition DDPM and the high classification performance of improved CNN.