Dual Adversarial Contrast for Train Bearing Fault Diagnosis under Extremely Low Label Rate

Fault diagnosis of train bearings is a promising approach to avoid catastrophic disasters and save train maintenance costs. Supervised fault diagnosis model is preferred because the labels of the training dataset are of great use in model training. However, the labeling of a large amount of data requires a lot of manpower and expert knowledge. The dataset in industry generally has a low label rate. Therefore, how to utilize the datasets with low label rates has become a concern. This paper proposes a new semi-supervised fault diagnosis model, which is termed dual adversarial contrast (DAC), to learn discriminative representations of train bearing health states from the dataset with extremely low label rate. Specifically, the DAC performs adversarial contrastive twice to extract features, which are more representative for bearing fault diagnosis. The proposed model is first pre-trained in an unsupervised way, and is then fine-tuned by the labeled data. A train bearing dataset with an extremely low labeling rate, i.e., 1%, is used to validate the proposed model. The results confirmed the advantages of the DAC in both diagnostic accuracy and convergence speed in comparison with the existing contrastive learning models.