基于子空间特征相似性的数据驱动主动学习控制

It has become one of the most important scientific and technological problems in the era of Big Data and Artificial to contribute the control theory with high-grade “learning intelligence” similar to the human. Iterative learning control applies “learning” to the automatic control of dynamic systems, however, it cannot utilize the control knowledge learned from one control scenario to help the control tasks in another similar scenario, which is the most difference from human learning. .Therefore, this project proposes a new data-driven active learning control based on subspace feature similarity. According to the similarity between the query data and the subspace database, an optimal control input can be found actively. In the sequence, it is used as a feedforward term of a control system to fast the convergence and to enhance the control performance. .First, this project considers repetitive dynamic systems operating over a finite time interval. The aim is to propose a data-driven active learning control method to learn from the similar but arbitrarily iteration-varying features. Apparently, the proposed method breaks through the fixed learning pattern of ILC that requires strictly repetitive scenarios..Second, the research results will be extended to the non-repetitive systems continuously running along an infinite time interval. A novel data-driven active learning method will be proposed to learning from the similar time instants instead of the similar iterations. In this section, the control system does not operate repetitively over a finite time interval any longer and thus the repetition limitation of ILC is completely overcome which broaden its practical applications. .Third, the self-tuning of the controller parameters is also discussed in this project based on subspace feature similarity and the robust theoretical results will also be explored by considering the data noises and data loss. Both of them will greatly facilitate the engineering applications in practice. .Finally, the urban traffic system is considered in this project from the viewpoint of practical applications. An experimental plat of traffic big data analysis and intelligent control will be built consequently to verify the proposed methods.

发展具备人类高级“学习智能”的控制理论已成为人工智能和大数据时代的重要科学问题。迭代学习控制(ILC)将“学习”应用到动态系统的自动控制中，但还不能跟人类一样，将从一个环境中学到的控制知识用来帮助新的相似环境中的控制任务。为此，项目提出基于子空间特征相似性的数据驱动主动学习控制方法，在每个操作点根据查询数据与经验数据库的相似性分析结果，主动寻找最优控制输入作为前馈，加快收敛速度、提高控制性能。首先，考虑有限区间上重复运行的动态系统，从迭代变化但相似的特征中主动学习，突破ILC严格重复的固定学习模式。然后，将结果推广到无穷区间上连续运行的非重复系统，提出从相似时段中主动学习的控制策略，克服ILC要求系统有限区间运行的局限。进而，还将讨论基于子空间特征相似性的控制器的参数自整定问题以及数据噪声和不完备情形下的理论结果，方便实际工程应用。最后，建立交通大数据分析与智能控制平台，验证方法的有效性。

发展具备人类高级“学习智能”的控制理论已成为人工智能和大数据时代的重要科学问题。迭代学习控制(ILC)将“学习”应用到动态系统的自动控制中，但还不能跟人类一样，将从一个环境中学到的控制知识用来帮助新的相似环境中的控制任务。本项目提出了基于子空间特征相似性的数据驱动主动学习控制方法，本项目按照计划已全面完成，取得了一些进展，在IEEE汇刊、IEEE TAC等期刊上发表论文46篇，发表中文核心7篇、会议论文11篇,获中国自动化学会自然科学奖(协会级)二等奖1项、山东省高等学校科学技术奖(市厅级)三等奖1项。授权国家发明专利4项，已申请国家发明专利8项。培养博士生2名，硕士生9名。具体包括：（1）提出了基于重复数据相似性动态的非线性系统数据驱动控制设计，建立了重复数据相似性动态的线性映射关系，提出了迭代动态线性化模型，为学习控制系统的设计提供了基础工具。（2）提出了非重复不确定系统的数据驱动学习控制设计，解决了数据丢包、数据量化、变长度操作等对数据不确定性的影响，提出了鲁棒增强的设计和分析方法。（3）提出了基于多智能体相似性动态数据关联关系的协同控制方法，分析了单智能体I/O动态时间演化关系，单智能体I/O动态在重复操作间的相似性关联关系，以及多智能体网络中各个智能体之间的动态相似性关联关系，以此提出了相应的协同控制策略，提高了控制性能。所取得成果得到了国内外学者的认可，所提出的方法可以避免复杂的系统建模过程，直接应用于相应的复杂实际过程中，并能有效克服网络环境下数据受扰对控制性能的影响，最终为现代工业自动化提供了理论支持和方法保障，具有重要的理论意义和实际应用价值。