Diffusion model and vision transformer for intelligent fault diagnosis under small samples

Abstract The existing deep learning models can achieve a high level of fault diagnosis accuracy in the case of a large number of samples. However, in actual production, data is often limited due to the difficulty of data collection and labeling. For small sample fault diagnosis, a fault diagnosis method called diffusion model-overlapping-patch vision transformer (DM-OVT) is proposed in this paper. The method adds coordinate attention to the DM, so that it can consider both channel information and spatial information. In the patch embedding part of Vision Transformer, features are first extracted using convolutional layers, and then overlapping patch divisions are used to improve the correlation between each patch. To be specific, DM-OVT first uses short-time Fourier transform to convert the one-dimensional signals into the time–frequency maps. And then inputs them into the DM to generate different classes of fault data according to labels. Finally, OVT is used to classify the expanded data. The effectiveness of the proposed method was tested on data sets from laboratory multistage centrifugal fans and Case Western Reserve University, and the highest accuracy was achieved in the comparison experiments.