A fine-tuning prototypical network for few-shot cross-domain fault diagnosis

Abstract

With the continuous development of computer technology, deep learning has been widely used in fault diagnosis and achieved remarkable results. However, in actual production, the problem of insufficient fault samples and the difference in data domains caused by different working conditions seriously limit the improvement of model diagnosis ability. In recent years, meta-learning has attracted widespread attention from scholars as one of the main methods of few-shot learning. It can quickly adapt to new tasks by training on a small number of samples. A fine-tuning prototypical network (FPN) is proposed on meta-learning methods to address the challenges of fault diagnosis under few-shot and cross-domain. Firstly, the shuffle attention (SA) is used to enhance the feature extraction ability of the network and suppress irrelevant features. Then, the support set of the target domain is split into two parts: pseudo support set and pseudo query set, which are used to fine-tune the prototypical network and improve the model generalization. Finally, experiments are conducted on three rotating equipment datasets to verify the method's effectiveness.