【作者】 杜大军；

【导师】 费敏锐； 李慷； George W. Irwin；

【作者基本信息】 上海大学 ， 控制理论与控制工程， 2010， 博士

【摘要】 网络控制系统具有布线少、易于安装维护和灵活性强等诸多优点,已经在磁悬浮球系统、双轴压力位置系统、机器人等领域出现一系列应用。然而实际工业过程往往为非线性多变量系统,参数往往时变,这类复杂对象本身在空间上分布广泛,常采用多个本地控制器的分布式控制模式,但由于本地控制器自身计算资源有限,无力运行复杂的控制算法以满足高性能的控制要求,因此控制需求对单层网络控制系统提出了严峻的挑战。此外,在复杂的控制算法中,学习算法由于具有可发现潜在规律以及适应环境等特点,能有效提高系统控制性能而得到广泛应用。于是,考虑学习算法趋于复杂,而采用本地设备受成本和可靠性要求的影响其计算资源和算法复杂度有限等现实,提出了两层网络学习控制系统。本文主要研究两层网络学习控制系统的学习和控制策略设计,主要工作概括如下:首先,提出了两层网络学习控制系统(networked learning control system, NLCS)架构,分析了其与传统网络控制系统相比所具有的不同特点。针对学习算法的需求,提出了一种新的基于快速回归算法的快速RBF神经网络构造方法,不但可以选择隐节点中心而且可以计算输出层权重,并分析了计算复杂度,与常见的基于正交最小二乘算法构造RBF神经网络相比速度更快。在此基础上,提出了基于RBF神经网络的自学习模糊控制策略,利用RBF神经网络实时调整本地模糊控制器的参数,并将其应用到两层网络学习控制系统中,得到基于快速RBF神经网络的两层NLCS自学习模糊控制策略方案。其次,提出了一种新的基于带有调整因子前向回归算法的RBF神经网络构造方法,通过给每一个候选项绑定一个调整因子,采用贝叶斯证据框架推理来优化调整因子,随着调整因子优化的进行,一些在网络中起非常微弱作用的隐节点将从RBF网络中移出,从而精简了RBF神经网络的结构,解决了RBF神经网络对含有噪声的测量数据出现过度拟合的问题。接着,定义一系列中间变量并采用递归方法简化了算法的计算负担,给出了算法的计算复杂度。在此基础上,提出了改进的基于RBF神经网络的自学习模糊控制策略,得到了基于带有调整因子RBF神经网络的两层NLCS自学习模糊控制策略方案。第三,提出了一种新的基于带有调整因子前向回归算法的RBF神经网络自动构造方法,采用能够测量模型泛化能力的LOO交叉验证准则来自动终止RBF神经网络的中心选择过程,进一步运用调整因子精简径向基神经网络的结构,并通过定义一系列中间变量以及采用递归方法简化了算法的计算复杂度。在此基础上,提出了进一步改进的基于RBF神经网络的自学习模糊控制策略,得到了基于带有调整因子自构造RBF神经网络的两层NLCS自学习模糊控制策略方案。第四,在不改变本地控制器的前提下,为了提高系统的控制性能,研究了基于强化学习控制策略的两层网络学习控制方案,学习单元采用强化学习策略,本地控制器采用PI控制策略,学习单元的输出动态调整PI控制器输出,组成本地控制器和学习单元相结合的控制策略。进一步探索了适合该策略的强化学习方法,对基于Q学习算法和Actor-Critic神经网络学习算法的两层网络学习控制方案分别进行仿真研究,仿真结果验证了所提方案的有效性。第五,分析了多输入多输出网络控制系统具有的多通道网络特性,采用对角矩阵描述广义传感器和广义执行器的状态。然后分别采用丢失数据置零策略和零阶保持器策略补偿数据丢包,建立了带有任意切换序列的切换系统模型,给出了闭环系统渐近稳定的充分条件,并将其拓展到参数不确定线性系统对象,得到了具有参数不确定性的闭环系统渐近稳定的充分条件。最后,在上述理论研究并在仿真结果验证了所提方案的有效性基础上,结合上海自动化仪表股份有限公司生产的具有两层网络架构的SUPMAX分散控制系统,分析了其网络架构和网络不确定性因素处理策略,通过扩展一个学习单元构建了两层网络学习控制系统实验平台。进一步以SUPMAX系统中广泛应用的PID控制回路为例进行两层网络学习控制实验,即在SUPMAX实现基本PID控制功能的基础上,利用SUPMAX的SMCP协议和名字服务器的地址解析功能,通过Ethernet网络从分散处理单元(DPU)获得实时数据,在学习单元采用面积法和继电整定法,对SUPMAX系统DPU中的PID控制器进行参数整定,开发了SUPMAX的先进PID自整定功能库。仿真实验表明该先进控制软件能够有效地进行PID参数整定,具有良好的应用前景。更多还原

【Abstract】 Networked control system (NCS) has wide industrial applications, such as in ball maglev system, dual-axis hydraulic positioning system, robot and large-scale transportation vehicles, due to various advantages including low cost of installation, ease of maintenance, easy installation, and flexibility. However, many industrial systems are characterized by time-varying, nonlinear and multivariable coupling. The plants are distributed at different locations, and distributed control strategies are most common where many local controllers are implemented. While the local controllers are mostly embedded real-time controller or program logic controllers which are highly reliable, but they can not fulfill the needs for high quality control due to very low computational capability for implementation of complex control strategies. This imposes a severe challenge on the widely used single-layer NCS. Learning control strategies are capable of effectively improving the control performance by on-line mining of valuable knowledge and hidden relations, accumulating experience and adapting to environment, but it is difficult to design and implement via the local controllers. Therefore, a two-layer networked learning control system (NLCS) is proposed. The focus of this paper is on the design of control strategies of two-layer NLCS. The main work is summarized as follows:Firstly, a two-layer NLCS architecture is presented, and.the differences between two-layer NLCS and NCS are analyzed. Then, a novel fast radial basis function (RBF) neural network based on fast recursive algorithm (FRA) is proposed. The method can not only select RBF neural network centers, but also can estimate the network weights simultaneously using a back substitution approach. Unlike popular orthogonal least squares (OLS) algorithm, FRA requires less computational effort by the computational complexity analysis. Furthermore, a self-learning fuzzy control strategy based on fast RBF neural network for two-layer NLCS is presented. Under this strategy, RBFNN is employed in a learning agent, while fuzzy control strategy is adopted in the local controller. The RBF neural network is used to on-line tune the parameters of fuzzy controller. Simulation results confirm its effectiveness.Secondly, a new locally regularized recursive method for the center selection of RBF neural network is proposed. By associating each candidate center to an individually regularized parameter which is optimized within the Bayesian evidence framework, the associated weights of those nonsignificant candidate centers are effectively forced to zero as their corresponding regularization parameters become sufficiently large. Therefore, the proposed method can easily choose significant centers and hence improve the sparsity. The computational complexity analysis shows that the computational cost is significantly reduced by using a proper regression context and the recursive formulas. Then, a self-learning fuzzy control strategy based on RBF neural networks with regularized parameters for two-layer NLCS is developed, and simulation results demonstrate its effectiveness.Thirdly, a novel locally regularized automatic construction method for RBF neural models is proposed. This is achieved by combing the proposed locally regularized recursive method with the leave-one-out (LOO) cross-validation criterion. It can automatically determine the network size by iteratively minimizing a LOO mean square error (MSE) without the need to specify any additional termination criterion. By defining a proper regression context and the recursive formulas, the whole network construction process can be concisely formulated and easily implemented at significantly reduced computational expense which is not achievable using any existing approaches. Then, self-learning fuzzy control strategy based on automatic constructive RBF neural network with regularized parameters for two-layer NLCS is investigated, and its effectiveness is then confirmed by simulation results.Fourthly, to improve the control performance, a two-layer NLCS scheme without changing the local controller is studied by using reinforcement learning methods. Under this scheme, reinforcement learning methods are employed in a learning agent, while proportion integration (PI) control strategy is adopted in the local controller. Total control signal consists of control signal of learning agent plus control signal of PI controller, where control signal of learning agent dynamically tunes total control signal. Then, two-layer NLCS using Q-learning and Actor-Critic neural network are investigated respectively, and simulation results verify its effectiveness.Fifthly, the characteristics of multi-channels networks for multi-input multi-output (MIMO) NCS are analyzed. The generalized sensor and generalized actuator are defined, and the sate of generalized sensor and generalized actuator can be described using diagonal matrix. The methods of zero-set and zero order holding (ZOH) are used to compensate data loss, MIMO NCS models are then modeled as a switched system with unknown switched sequence and sufficient conditions for asymptotically stable are given. This can be further extended to uncertain MIMO system, and sufficient conditions for asymptotically stable are derived. Numerical example confirms its effectiveness.Finally, SUPMAX distributed control system (DCS) is introduced, and the network architecture and the method for tracking networked nondeterministics are presented. A two-layer NLCS experiment platform is then established by addding a learning agent to SUPMAX DCS. Using the SMCP protocol and address resolution function, and real-time data can be obtained from distributed processing unit (DPU) via Ethernet network. An advanced PID self-tuning software package is then developed using the characteristic area method and a relay tuning algorithm, which can tune the PID parameters in DPU. Simulation experiments demonstrate that the software can effectively tune the PID parameters, and shows the potentials of the software for real-life applications.更多还原