Imbalanced fault identification via embedding-augmented Gaussian prototype network with meta-learning perspective

Abstract

In practical engineering, the number of acquired fault samples from different categories can be vastly different due to the low probability of key equipment malfunctioning. When training the imbalanced data, many methods focus on balancing the number of samples or weights between different categories, which may be time-consuming and easy to over-fit. To address this problem, we propose the embedding-augmented Gaussian prototype network (EGPN), which applies a new training mechanism from the perspective of meta-learning. We only train the categories with large samples and the remaining categories only appear in the testing process to calculate untrained prototypes. EGPN includes a feature-embedding augmented module, weighted prototype module and metric module. Firstly, ordinary convolution and dilated convolution are mixed to capture different frequency bands simultaneously, and the residual attention module is added to highlight key features and suppress unimportant features. Different prototypes are calculated by weighting to the embedding vectors through the Gaussian covariance matrix. Finally, the classification is done according to the modified distance. The experiments in the two datasets indicate that the proposed method can effectively recognize the untrained categories with only a few samples used as the prototypes, and can tackle the problem of identifying imbalanced fault data efficiently.