Semi-supervised prototype optimization network for rolling bearing fault diagnosis with limited labeled samples

Abstract In practical industrial scenarios, rolling bearings serve as critical rotating components, whose operational status is directly linked to the stability and safety of equipment. However, due to the high cost and low frequency of fault data acquisition, existing deep learning methods struggle to balance diagnostic accuracy and generalization capability under sample-scarce conditions. To address this challenge, this paper proposes a semi-supervised prototypical optimization network for rolling bearing fault diagnosis with limited labeled samples. Specifically, a multi-scale residual attention feature network (MRAFN) is designed, which incorporates an improved squeeze-and-excitation module and a multi-path feature enhancement mechanism to fuse shallow local perception with deep semantic information, thereby improving the selectivity and fusion expression of inter-channel features for effectively mining latent time-frequency diagnostic characteristics in bearing vibration signals. Meanwhile, a prototype optimization strategy based on optimal transport theory is constructed to dynamically refine the initial class prototypes by leveraging the underlying structural relationship between labeled and unlabeled samples, thus improving the model’s discrimination and robustness in fault classification. Finally, extensive validation is conducted on two bearing fault datasets, and the experimental results demonstrate the superior diagnostic performance of the proposed method under limited labeled sample conditions.