针对含噪输入输出导致 Koopman 模型有偏及动力学失配的问题,本文提出 KROC 双回路鲁棒控制框架,并通过 LMI 保证稳定性与性能,在 Van Der Pol 振子上验证了性能提升。
摘要:
用于非线性系统数据驱动控制设计的 Koopman 算子方法因其能够捕捉全局动力学行为而日益兴起。然而,输入和输出的测量噪声会使 Koopman 模型辨识产生偏差,并导致模型与实际非线性动力学之间失配。当前工作评估了 Koopman 算子模型由噪声诱发的模型偏差界限,并针对该有偏模型提出了一种数据驱动的鲁棒双回路控制框架(基于 Koopman 的鲁棒控制—KROC)。首先,在径向基函数(RBF)和有界噪声条件下,证明模型失配是有界的,并评估了模型失配的界。其次,展示了 Van Der Pol 振子基于有偏 Koopman 模型的线性二次型高斯(LQG)控制所存在的缺陷。受这些缺陷启发,提出了双回路控制,它由基于标称 Koopman 模型的基于观测器的状态反馈控制,以及用于补偿模型失配的附加鲁棒回路组成。推导了一个线性矩阵不等式(LMI),该不等式能够保证有限维 Koopman 算子模型在有界噪声下的鲁棒稳定性和性能。最后,将所提出的框架应用于非线性 Van Der Pol 振子,以证明双回路鲁棒控制带来的控制性能提升。
英文摘要:
The Koopman operator approach for data-driven control design of a nonlinear system is on the rise because of its capability to capture the behaviours of global dynamics. However, the measurement noises of inputs and outputs will bias the Koopman model identification and cause model mismatch from the actual nonlinear dynamics. The current work evaluates the bounds of the noise-induced model bias of the Koopman operator model and proposes a data-driven robust dual-loop control framework (Koopman based robust control-KROC) for the biased model. First, the model mismatch is found bounded under radial basis functions (RBF) and the bounded noises, and the bound of model mismatch is assessed. Second, the pitfalls of linear quadratic Gaussian (LQG) control based on the biased Koopman model of Van Der Pol oscillator are shown. Motivated from the pitfalls, the dual-loop control is proposed, which consist of an observer-based state-feedback control based on the nominal Koopman model and an additional robust loop to compensate model mismatch. A linear matrix inequality (LMI) is derived, which can guarantee robust stability and performance under bounded noises for the finite-dimensional Koopman operator model. Finally, the proposed framework is implemented to a nonlinear Van Der Pol oscillator to demonstrate enhanced control performance by the dual-loop robust control.
A Koopman Operator-based NMPC Framework for Mobile Robot Navigation under Uncertainty
Department of Electrical and Computer Engineering, University of California, Riverside(加州大学河滨分校电气与计算机工程系);Polytechnic School of Tunisia(突尼斯多技术学院);Institute of AI Industry Research, Tsinghua University(清华大学人工智能产业研究院)