Self-supervised Contrastive Learning Approach for Bearing Fault Diagnosis with Rare Labeled Data

Data-driven intelligent fault diagnosis methods based on Deep Learning algorithms have been widely studied in recent years. These methods can help reduce costly breakdowns. However, most of such deep-learning methods are supervised, which need numerous labeled data for model training. Learning from just a few labeled examples while making the best use of a large amount of unlabeled data is the target we aim to achieve. In this paper, we innovatively introduce the unsupervised learning method for fault diagnosis. We construct a deep learning structure using the typical contrastive learning method MoCo and a linear classifier. Our structure consists of the encoder network based on ResNet-18 and a simple linear classifier. Experiments are made on the CRWU dataset. The model based on unsupervised features can reach 75.86% accuracy using only a few labeled data, in comparison with the traditional supervised method at 75.86% on average. The use of unsupervised learning also improves the generalization ability of the model.