Embedding to Metric Model for Few-Shot Cross-Domain Fault Diagnosis

Abstract

Few-shot fault diagnosis aims to address the issue of data scarcity in fault diagnosis. Pioneering studies typically employ meta-learning frameworks due to their elegant formalization and effective properties. However, this structure will also increase the complexity of model training and limit the design of the model. Surprisingly, a new perspective is that models in which meta-train algorithms and meta-test algorithms that are completely uncorrelated can outperform all meta-learning methods. Building on this line of inquiry, we propose a novel model termed Embedding to Metric (E2M), with a new framework for cross-domain few-shot fault diagnosis. In this new framework, the meta-train algorithm focuses on obtaining good embeddings to leverage the metric in the meta-test. Finally, we evaluate the proposed model on a public dataset, demonstrating that our framework outperforms state-of-the-art algorithms with complex structures. This result confirms the viability of the new framework and may contribute to a better understanding of the relationship between few-shot fault diagnosis and other fields, such as fault feature learning and transfer learning.