CDDPM Data Augmentation and Fault Diagnosis for Small-Sample Vibration Signals

Vibration data from industrial equipment commonly suffers from issues such as sparse samples and class imbalance, which limit the performance of fault diagnosis models based on convolutional neural networks (CNNs). To address this issue, this paper proposes a vibration signal data augmentation method based on the Conditional Denoising Diffusion Probabilistic Model (CDDPM). The method first converts one-dimensional vibration signals into frequency-domain visualizations (FDVIs), then trains a conditional diffusion model to generate high-quality fault samples, and finally uses these generated samples to augment the training set, thereby improving the diagnostic accuracy of the CNN classifier. Experimental results on a dataset of rolling bearing failure data from oil transfer pumps collected in an actual petrochemical production environment demonstrate that the proposed method significantly improves classification performance, achieving an accuracy of 85.7% and with a macro F1-score of 0.84. This outperforms traditional data augmentation methods as well as deep generative models based on Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), validating its effectiveness and practicality in fault diagnosis tasks under low-sample-size and class-imbalanced scenarios, and providing new technical support for the predictive maintenance of industrial equipment.