A novel dimensional variational prototypical network for industrial few-shot fault diagnosis with unseen faults

Abstract

A Dimensional Variational Prototypical Network (DVPN) is proposed to learn transferable knowledge from a largescale dataset containing sufficient samples of diverse faults, enabling few-shot diagnosis on new faults that are unseen in the dataset. The network includes a multiscale feature fusion module with shared weights to extract fault features, followed by a dimensional variational prototypical module that uses variational inference to determine metric scaling parameters. This adaptive approach accurately measures feature similarity between samples and fault prototypes. To enhance discriminability, a representation learning loss is employed, distinguishing between the least similar samples within the same class (hard positive samples) and the most similar samples across different classes (hard negative samples). The network combines representation learning and prototypical learning through the joint representation learning (JRL) module, acquiring both task-level and feature-level knowledge for a more discriminative metric space and improved classification accuracy on unseen faults. Experimental evaluations on datasets from the Tennessee Eastman process and a real-world polyester esterification process show that the proposed DVPN achieves high diagnostic performance and is comparable to state-of-the-art methods for few-shot fault diagnosis (FSFD).