ENFES: ENsemble FEw-Shot Learning For Intelligent Fault Diagnosis with Limited Data

Abstract

Fault diagnosis is a key component of predictive system maintenance. Big data collected from sensors helps create data-driven fault diagnosis methods. However, it may be extremely costly to label specific fault types in a collected dataset. Hence, prediction algorithms should perform well under limited supervision. Few-shot learning (FSL) can provide a great prediction performance using very limited labeled data by discovering similarity among input pairs. But selection of a single FSL method may be arduous due to changing working conditions. Ensemble FSL solves this problem by combining a variety of FSL methods systematically. We propose an ensemble FSL framework, ENFES, where we combine 5 different Siamese neural network architectures using an iterative majority voting classifier. Our transfer learning-oriented experiments show that ENFES can improve the best algorithm significantly while using very limited labeled data. We obtain up to 16.4% improvement over the best algorithm by only using 0.3% of the training data.