A data augmentation-based fault diagnosis method for rotating machinery

Abstract Efficient and robust intelligent fault diagnosis for rotating machinery plays a crucial role in ensuring the reliability of industrial equipment and enhancing economic efficiency. However, in real-world operating conditions, fault samples are often scarce, making fault diagnosis tasks under small-sample conditions highly challenging. Recently, due to the significant achievements of the denoising diffusion model in other fields, it has become the most powerful generative model. This study introduces a discriminator to facilitate a competitive interaction between the two, thereby generating high-quality synthetic data. After increasing the number of fault data samples using generative methods, the symmetric hybrid fusion network is proposed. Based on the encoded two-dimensional image format, symmetric convolution kernels are used for feature extraction, followed by hierarchical fusion of the extracted features. Finally, the features extracted from multiple branches are fused at the decision layer for diagnosis. The effectiveness of the proposed method was validated on two datasets, achieving recognition accuracies of 97.6% and 96.3% on the WT-Planetary gearbox dataset and Paderborn University bearing dataset, respectively. This study successfully achieved data augmentation and high-accuracy fault diagnosis under small-sample conditions for rotating machinery.