【作者】 易军；

【导师】 费敏锐；

【作者基本信息】 上海大学 ， 控制理论与控制工程， 2007， 博士

【摘要】 控制回路通过通信网络闭环的控制系统称为网络控制系统。网络控制系统中的网络诱导延时一般是不可避免的，并且往往是一种不确定性的随机延时，这给控制系统的精确设计带来了很大的的困难。此外，不确定性非线性对象在工程实际中大量存在。因此，不确定非线性网络控制系统的鲁棒分析和设计十分重要。当前，不确定非线性网络控制系统的鲁棒稳定性研究是其首要解决的学术难点问题。因此，本文针对一类不确定性非线性对象，通过Lyapunov定理来估计保证系统鲁棒稳定的最大网络延时，由此提出了网络控制系统鲁棒渐近稳定的充分条件。在网络控制系统中，控制回路完全通过网络形成闭环，由于存在网络延时、丢包，甚至瘫痪的危险，控制的稳定性和安全性很难保证。另一方面，随着生产规模的扩大，被控对象的不断拓展、复杂化且存在时变性，为了进一步提高其控制性能并保持优化状态，越来越多的带有学习功能的控制器纷纷出现。并且随着控制要求的高性能化，学习算法复杂度也在不断提高，当需要耗用相当的计算资源和存储资源时，现有的现场控制设备的计算和存储资源将难以胜任实施过于复杂的学习算法。即使有这样高性能的现场智能控制仪表，其价格也是非常昂贵的。为此，本文提出一种现场的控制单元，采用简单、易于实现的方式，而复杂的辨识、学习算法则通过网络连接的远端计算机实现，以求实现一种低成本、高性能、充分利用网络资源的控制系统，并称之为网络学习控制系统。基于上述原因，如果只采用确定性的、线性的控制策略将很难获得较好的控制性能。因此，针对不确定非线性对象进行网络学习控制系统的研究具有重要的价值，且网络学习控制系统的稳定性及学习收敛性研究是首要问题。为此，本文从以下几个方面进行了深入研究：首先，针对未知数学模型的被控对象，提出了基于三次样条插值的一步预测延时补偿模型，在此基础上为了提高延时补偿精度，进一步提出了三次样条滚动优化多步预测延时补偿算法。同时提出了神经网络作为网络学习的复合控制，并对网络环境下的复杂、时变被控对象进行了控制仿真。仿真结果验证了控制策略的有效性。其次，针对不确定非线性被控对象，提出了基于三次样条滚动优化多步预测算法校正非线性模型的综合预测延时补偿算法，以及相应的网络学习算法，并对网络环境下的不确定非线性对象进行了控制仿真。仿真结果验证了算法的有效性。再次，针对一类包含未知非线性函数的不确定性对象，采用动态递归神经网络进行本地控制，采用均匀三次样条插值算法针对未知非线性函数进行远程辨识，并通过三次样条滚动优化多步预测算法对网络延时进行实时补偿。最后针对有界不确定性网络延时，通过Lyapunov稳定性定理得到了保证系统稳定的动态递归神经网络的在线最大学习率，由此提出了网络学习控制系统渐近稳定的充分条件，并用仿真实验验证了理论分析的正确性。最后，开展了网络学习控制策略针对发电厂中的循环流化床锅炉(CFBB)的燃烧系统和氢氧燃料电池测试系统研究，通过实验仿真和现场测试验证了网络学习控制策略的可实现性。更多还原

【Abstract】 Networked control system is a kind of feedback control systems, wherein the control loop is closed via the communication network, and the network-induced delay is inevitable. Network-induced delay is a class of uncertain and random delay, which results in the difficulty for the control system to be designed. In practice, moveover, a lot of uncertain and nonlinear factors come into existence. So the robust analysis and design of uncertain nonlinear networked control system need to be emphasizd. Obviously, the first issue needs to be solved is the robust stability analysis. In the thesis, a class of uncertain nonloinear networked control system is discussed and the sufficient condition for robust asymptotic stability is presented. And it is also given the maximum network-induced delay that is allowable for the stability to be remained. The Lyapunov theorem is employed to solve the problems.In networked control systems, control loop is closed via the communication network entirely. The network-induced delay as well as the possible data loss will always deteriorate the control performance. And in addition, the possibility of network paralysis makes the system security be hard to guarantee. On the other hand, for the expanding of the production scope, continuously developing complication and time varying of the controlled plant, more and more controller with some learning ability, such as the ability of online system identification, optimazation and system diagnosis, are applied to adapt itself to the time varying environment to improve the control performance and to keep its optimal operation condition. The learning algorithm complexity, however, is also continuously increased as the demand on control performance going higher and higher. Existing field control equipments’ computing and memory resource will be exhausted by so complicated learning algorithm. If the high capability field control equipments would be produced, they will be very expensive. In the thesis, we design a systerm that the local control unit is simple and ease of realization, while the complicated identification and learning algorithm are realized by the remote network unit. Named as networked learning control system, this system could be a control system with low cost, high capability and full use of network resource.As mentioned before, if only apply certain and linear control strategy, it would be very hard to improve the control system performance. Therefore for the uncertain nonlinear plant, the research on networked learning control system is valuable. And the stability and learning algorithm convergence should be main focus. In this thesis, some following aspects have been carefully researched. First, for the control object with unknown mathematical model, the network-induced delay compensation strategy, which is based on the cubic spline one-step predictive algorithm, is proposed. In order to improve the compensation precision of network-induced delay, multi-step predictive delay compensation model of cubic spline rolling optimization is also designed, and the corresponding networked learning stratery is put up. The networked learning control simulation is studied for complex, time-varying control object with unknown mathematical mode, and the simulation results prove the validity of this control strategy.Second, for the uncertain nonlinear object, the composite predictive algorithm in which the nonlinear model predictive is compensated by error prediction based on multi-step predictive algorithm using cubic spline rolling optimization is presented. And the corresponding networked learning algorithm is also presented. The networked learning control simulations are studied for uncertain nonlinear object and the simulation results prove the validity of the control strategy.Third, for a class of uncertain control object with unknown nonlinear function, dynamic recurrent neural network (DRNN) is applied in local controller, and unknown nonlinear function is remotely identified by uniformly partition cubic spline interpolating function. The network-induced delay is compensated by multi-step predictive algorithm of cubic spline rolling optimization in the networked learning unit. Finally, for uncertain bounded delay, the maximum online learning rate of DRNN that insures the system stability is obtained by Lyapunov stability theorem, and then the sufficient condition for the asymptotic stability is presented.At last, the networked learning control strategy is applied in the circulating fluidized bed boiler (CFBB) of the electricity generating burning systems and the H2-O2 fuel cell test system. The validity of the networked learning control strategy is also testified by the simulations and field testing.更多还原