Reinforcement Learning-Aided Performance-Driven Fault-Tolerant Control of Feedback Control Systems

• 成果类型: Article
• 署名作者: Hua, Changsheng; Li, Linlin; Ding, Steven X.
• 署名单位: University of Duisburg Essen; University of Science & Technology Beijing
• 刊物名称: IEEE TRANSACTIONS ON AUTOMATIC CONTROL
• ISSN/ISSBN: 0018-9286
• DOI: 10.1109/TAC.2021.3088397
• 发表日期: 2022
• 页码: 3013-3020
• 关键词: degradation; system performance; Fault tolerant systems; Fault tolerance; trajectory; Stochastic processes; estimation; data-driven; fault-tolerant control (FTC); performance degradation recovery; reinforcement learning (RL)

摘要

This article is concerned with a fault-tolerant control (FTC) scheme for feedback control systems with multiplicative faults by optimizing system performance with the aid of a reinforcement learning (RL) approach. To be specific, initially, based on the Youla-Kucera (YK) and dual YK parameterizations, a new performance-driven FTC method is proposed and its capability in dealing with multiplicative faults is proven. Then, data-driven implementation of this method using RL is elaborated. This implementation shows that RL can be applied efficiently by utilizing both plant model and data to recover the fault-induced system performance degradation. Finally, a benchmark study on an inverted pendulum system demonstrates the application of the proposed performance-driven FTC method.