Abstract: To tackle the challenge of rolling bearing fault diagnosis under conditions of weakly labeled data samples,a novel approach utilizing self-supervised learning within the time-frequency was developed to unearth latent fault features in fault-free labeled samples.Initially,feature representations were extracted from both the time and frequency domains by constructing respective time and frequency encoders. A model for the time-frequency self-supervised learning was designed to augment the mutual predictive capabilities between features across these domains. Furthermore,to refine the model learning process,a novel cross-correlation matrix loss function was devised,significantly enhancing the model′s proficiency in identifying complex failure modes. The efficacy of this method was confirmed by using both the Case Western Reserve University bearing fault open dataset and the Paderborn University bearing open dataset,where the experimental outcomes indicated superior diagnostic results with minimal fault-labeled data.

Key words: rolling bearing, fault diagnosis, time-frequency domain characteristics, self-supervised learning, weakly labeled samples