A globally-locally consistent dynamically weighted joint optimization network for few-shot bearing fault diagnosis

Abstract In recent years, intelligent fault diagnosis methods have achieved significant success in the condition monitoring of mechanical systems. However, existing deep learning approaches typically rely on large amounts of labeled data for training, which is often expensive and time-consuming to acquire in practical engineering applications. As a result, few-shot learning has garnered increasing attention in fault diagnosis. This paper proposes a globally-locally consistent dynamically weighted joint optimization network (GLD-JON) for few-shot fault diagnosis. The method first applies a wavelet transform to convert the raw vibration signal into a time–frequency representation. Next, a dual-branch network structure is designed, where the global branch is used to predict the class label of the sample, and the local branch extracts local features by randomly cropping the query sample, followed by classification of the local features. Based on these two branches, a global–local consistency constraint is further designed to enhance the model’s generalization ability in few-shot cross-domain tasks. Multiple experiments were conducted on three bearing datasets. The results show that the proposed GLD-JON method outperforms various benchmark models in few-shot cross-domain learning tasks, demonstrating superior generalization ability.