Meta-Learning Guided Few-Shot Learning Method for Gearbox Fault Diagnosis Under Limited Data Conditions

Abstract

Recently, intelligent fault diagnosis technology based on deep learning has been extensively researched and applied in large industrial equipment system for ensuring safe and stable production. However, these deep models only effective when enough data for each observed failure category are available in the training durations. Otherwise, the performance of these models will notably decrease. As the critical component in large machinery, the gearbox often changes the speed and load along with the production demand in the practical application, which caused few data samples to be collected at certain conditions. This phenomenon introduces the few-shot fault diagnosis, and its goal is to identify the fault types with extremely limited data samples. To address this problem, a Meta-learning guided Few-shot Fault Diagnosis method, named MFFD, is proposed for gearbox fault diagnosis under limited data conditions. The results verify the effectiveness of our MFFD method at one-shot and five-shot fault diagnosis tasks under different speed and load conditions.