MFSFormer: A Novel End-to-End Rotating Machinery Fault Diagnosis Framework for Noise and Small Samples

Convolutional neural networks and transformers have both achieved remarkable success in the field of fault diagnosis owing to their proficiency in extracting local and global features. However, in real industrial production, the diagnostic performance of many methods is often limited by factors such as environmental noise and sample quantity. To address these challenges, this article presents a novel fault diagnosis framework called MFSFormer. First, it employs embedded convolutional layers with kernels of various sizes to extract multiscale receptive field features from vibration signals. Second, a Fuse-Shuffle attention block is employed to capture dependencies between feature channels and windows, facilitating the implementation of subfeature flow. The case studies conducted on two public datasets demonstrate that the proposed method achieves an average diagnostic accuracy that surpasses the suboptimal method by 3.91% and 7.00% in noisy environments and small-sample conditions, respectively. The results indicate that the proposed method not only enhances the robustness of fault diagnosis in noisy environments, but also improves diagnostic accuracy under limited sample conditions, and the method has practical value.