Semi-supervised feature contrast incremental learning framework for bearing fault diagnosis with limited labeled samples

Abstract

In real-world scenarios, rotating machinery consistently introduces new fault classes, but intelligent fault diagnosis methods mostly rely on the closed-world assumption, expecting only known fault classes during testing. Moreover, obtaining a sufficient number of labeled samples is often challenging. These challenges constrain the application and reliability of intelligent diagnosis models in real-world scenarios. Semi-supervised incremental learning enables continuous learning of new fault classes in an open environment, relying on a small number of labeled samples and a certain number of unlabeled samples. To address the semi-supervised incremental learning problem of fault classes, semi-supervised feature contrast (SSFC) is proposed, a new approach for bearing fault diagnosis with limited labeled samples. Specifically, a feature contrastive loss incorporating enhancement strategies is designed, independent of labeled sample information. This approach enables the model to retain knowledge of old classes while learning about new ones. A label reconstruction mechanism based on class centroids is utilized, effectively leveraging the structural information inherent in the samples to support supervised training. A dynamic class prototype cosine classifier initialized by class centroids is devised to mitigate interference between knowledge of fault classes. Finally, two incremental fault diagnosis case studies are designed to evaluate the effectiveness of the proposed method. The fault diagnosis results indicate that SSFC can continuously learn knowledge of new fault classes with limited labeled samples and effectively alleviate catastrophic forgetting.