Semi-Supervised Fault Diagnosis Method for Hydraulic Pumps Based on Data Augmentation Consistency Regularization

Due to the scarcity of labeled samples, the practical engineering application of deep learning-based hydraulic pump fault diagnosis methods is extremely challenging. This study proposes a semi-supervised learning method based on data augmented consistency regularization (DACR) to address the issue of lack of labeled data in diagnostic models. It utilizes augmented data obtained from the improved symplectic geometry modal decomposition method as additional perturbations, expanding the feature space of limited labeled samples under different operating conditions of the pump. A high-confidence label prediction process is formulated through a threshold determination strategy to estimate the potential label distribution of unlabeled samples. Consistent regularization loss is introduced in labeled and unlabeled data, respectively, to regularize model training, reducing the sensitivity of the classifier to additional perturbations. The supervised loss term ensures that the predictions of the augmented labeled samples are consistent with the true labels. Meanwhile, the unsupervised loss term can be used to minimize the difference between the distributions of unlabeled samples for different augmented versions. Finally, the proposed method is combined with Kolmogorov–Arnold Network (KAN). Comparative experiments based on data from two models of hydraulic pumps verify the superior recognition performance of this method under low label rate.