Adaptive Attention-Driven Few-Shot Learning for Robust Fault Diagnosis
Zhe Wang, Yi Ding, Te Han, Qiang Xu, Hong Yan, Min Xie
Abstract
The deep learning techniques have propelled significant advancements in intelligent fault diagnosis. However, the limited labeled data due to resource-intensive labeling processes pose the challenges for actual applications. This study proposes an attention-centric model for few-shot fault diagnosis in rotating machinery. The model is informed by few-shot learning (FSL) and integrates internal and external attention (EA) mechanisms, which are leveraged to enhance the feature extraction capability. Performance evaluations under the five-way one-shot setting achieve remarkable results. The accuracy reaches 97.147% for the scenario from artificial damage to real damage, and 95.613% for the scenario of different operational conditions. The critical role of the integrated attention modules is further validated through the ablation study. Comparative analysis with state-of-the-art techniques demonstrates the superior performance of the proposed model. In short, this work provides an alternative method for fault diagnosis under the few-shot limitation.