Relation Awareness Network for Few-Shot Fine-Grained Fault Diagnosis

Abstract

Few-shot fine-grained fault diagnosis aims at identifying faults at a fine-grained level with limited training samples, which is challenged by subtle category differences inherent in fine-grained fault diagnosis. To address this limitation, we introduce Relation Awareness Network (RAN) to explore multi-level semantic relationships in time-frequency images. RAN integrates two novel strategies: Contextual Relation Learning (CRL) and Relation Collaboration Optimization (RCO). CRL strategy analyzes internal variations within samples as well as the correlations among different samples, effectively capturing the unique characteristics of each sample but also identifying the patterns shared across various samples. Concurrently, RCO strategy enhances the model's ability to discriminate between classes by optimizing the inter-class relationships. Experiments on two public and one lab-built bearing dataset demonstrate RAN's effectiveness. Quantitative results show that compared with existing methods, RAN achieves a diagnosis accuracy improvement of up to 2.27% and 1.01% on the PU bearing dataset in the 10-way 1-shot and 10-way 5-shot settings, respectively.