Consistency-regularized-label-aware contrastive learning with uncertainty-aware periodic pseudo-labeling for machinery fault diagnosis under limited labeled data

Deep learning-based fault diagnosis methods face challenges due to limited labeled samples in industrial settings. Traditional Contrastive Learning (CL) methods, based on self-supervised pre-training and fine-tuning, have shown promising results. However, they fail to effectively utilize labeled and unlabeled data during the pre-training and fine-tuning. Additionally, conventional 1D-convolutional neural network (1D-CNN) encoders capture only either fine- or coarse-grained features, limiting their effectiveness. To address these issues, this study proposes Consistency-Regularized-Label-aware Contrastive Learning (CR-LaCL) and Uncertainty-aware Periodic-Pseudo-Labeling (UaPPL) for a novel diagnostic framework. CR-LaCL integrates the contrastive loss for unlabeled samples with classification loss and consistency regularization for labeled samples, training the Multi-scale CNN-based encoder and a classifier for improved feature representation. The classifier is then fine-tuned only on the labeled samples. UaPPL uses entropy-based uncertainty weighting for pseudo-labels and employs a periodic pseudo-labeling strategy to mitigate the noisy labels and stabilize training. The proposed method is evaluated on nuclear circulation water pump gearbox experiments, as well as public gearbox and bearing datasets. Comparative analyses revealed that the proposed method shows higher diagnosis accuracy across all label ratios (0.05 to 0.20) as compared to supervised baselines, CL, and semi-supervised fault diagnosis methods.