A Semi-supervised Fault Diagnosis Method for Rolling Bearing Using Time-Frequency Transform Enhanced Contrastive Learning

This paper proposes a novel semi-supervised framework, time-frequency Contrastive Learning (CL), to address the challenge of accurate rolling bearing fault diagnosis under industrial small-sample conditions. Raw vibration signals are transformed into discriminative time-frequency images using short-time Fourier transform (STFT). A CL network with a ResNet18 model is pre-trained on a lot of unlabeled samples to learn generalized feature, and the ResNet18 model is fine-tuned using small labeled samples for fault classification. Experimental validation on bearing fault datasets demonstrates that the proposed STFT-CL method achieves above 99% diagnosis accuracy with only 50 labeled samples per fault type, outperforming conventional semi-supervised methods by 6-12%. The proposed method provides a potential solution to the "small sample dilemma" in industrial applications through the synergistic effect of physically driven signal processing and self supervised representation learning.