A survey on few-shot learning for remaining useful life prediction

Abstract

The prediction performance of most data-driven remaining useful life (RUL) prediction methods relies on sufficient training samples, which is challenging in few-shot scenarios such as time-consuming or expensive monitoring, and the lack of historical data for newer equipment. Therefore, utilizing few-shot learning (FSL) methods to accurately obtain mapping functions of equipment RUL from limited data has attracted the attention of many researchers. Despite this, a systematic review of this class of prediction methods is still lacking. To fill this gap, this review comprehensively examines numerous research findings on RUL prediction in few-shot scenarios, groups the existing FSL for RUL prediction methods into three categories based on different sources of prior knowledge, introduces the principles, and recent advances of each category in detail, and, in particular, highlights the impact and constraints of RUL prediction task characteristics on various FSL methods. Additionally, this review discusses the challenges faced by FSL-RUL prediction during development and application, and explores potential future opportunities from an informative and knowledge perspective.