面向工业监控典型监督任务的深度迁移学习方法:现状、挑战与展望

Deep transfer learning methods for typical supervised tasks in industrial monitoring:state-of-the-art,challenges,and perspectives

基于深度迁移学习的工业监控方法在近年来获得了大量研究关注,特别是在以故障诊断、软测量等为代表的工业监控典型监督任务中.通过挖掘与迁移相似源域的知识来完成对目标域的建模,这类方法为实际工业场景中变工况等原因导致的跨域监控问题提供了新的思路.本文系统梳理了面向工业监控典型监督任务的深度迁移学习方法,并将其分为基于模型迁移、基于样例迁移与基于特征迁移的工业监控方法.在此基础上,对不同类方法的基本研究思想在故障诊断与软测量任务中的研究进展进行了详细阐述.随后,从实际工业场景的复杂欠数据问题、可迁移性的量化与负迁移问题、工业过程的动态特性问题等角度,指出了当前基于深度迁移学习的工业监控研究中存在的挑战,并对该领域的未来研究方向做出进一步展望.

Deep transfer learning-based industrial monitoring methods have received considerable research attention in recent years,especially in typical industrial monitoring tasks,including fault diagnosis and soft sensor developments.Such methods mine and transfer knowledge from similar source domains to model the data in the target domain.They provide a new perspective for cross-domain industrial monitoring problems caused by varying conditions in actual scenarios.This survey systematically sorts the deep transfer learning methods for typical supervised tasks in industrial monitoring and classifies them into model-based,instance-based,and feature-based approaches.Subsequently,it introduces the basic ideas and state-of-the-art approaches in fault diagnosis and soft sensor development of different categories.Finally,from the perspectives of complexly limited data,evaluation of transferability and negative transfer problems,and the dynamic characteristics of industrial processes,the survey highlights the current challenges in cross-domain industrial monitoring and points to future research areas in this field.