Bayesian contrastive Learning: An augmentation-free fault diagnosis method with limited labels and uncertainty quantification

Fault diagnosis under limited labeled data remains a significant challenge for deep learning-based methods, especially when uncertainty is not estimated, compromising model trustworthiness. While contrastive learning (CL) has been used for self-supervised learning in label-scarce scenarios, most methods rely on augmentations, which may limit robustness in industrial applications. This paper proposes an augmentation-free CL framework for fault diagnosis. A Bayesian Multiscale Convolutional Neural Network (BMCNN) encoder, a linear projector, and a Bayesian classifier are jointly pre-trained using contrastive, cross-entropy losses, and Kullback-Leibler divergence, with the encoder updated via a momentum encoder. Subsequently, pseudo-labeling-based semi-supervised learning is employed for effective fine-tuning, which uses uncertainty weighting to achieve the pseudo-labeling-based classification objective. The Bayesian nature of encoder and classifier results in well-calibrated uncertainty, thereby improving performance of semi-supervised learning using pseudo-labels. Furthermore, Monte Carlo sampling is employed for final uncertainty estimation. Experimental results demonstrate that the proposed method achieves superior diagnostic accuracy under extreme label scarcity, outperforming state-of-the-art contrastive learning, Bayesian variants, and pseudo-labeling approaches. Moreover, the predicted uncertainty exhibits a strong correlation with prediction accuracy, significantly improving model reliability and decision-making confidence in intelligent fault diagnosis systems.