【作者】 韩冰；

【导师】 韩敏；

【作者基本信息】 大连理工大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 本文针对工业过程中常见的时滞系统建模与控制方法进行研究。其目的是在动态神经网络理论的基础上,构建对未知时滞系统的参数辨识与建模方法,进而提出时滞系统的有效控制策略.本文提出两种包含可变延迟时间参数的动态神经网络算法分别用于实现对未知时滞系统的离线和在线辨识。在此基础上分别提出时滞系统控制器的设计方案,并进行相关的理论分析。主要研究内容和研究成果包括:(1)基于通用学习网络自适应算法的非线性时滞系统辨识研究。本文根据通用学习网络在对非线性时滞系统建模过程中表现出的特性,结合网络中延迟参数可以任意设定的特点,提出一种自适应选择延迟时间参数的通用学习网络学习算法。该算法利用通用学习网络的收敛速度对网络中与输出节点相连的分支上的延迟时间参数较为敏感的特性,通过误差评价函数对网络的延迟时间参数进行修正,在保证误差精度的同时加快网络的收敛速度。与此同时,通过该算法优化得到的网络延迟时间参数可以用于时滞系统延迟时间的辨识。此外,本文针对神经网络学习过程中延迟参数变化引起的网络状态扰动问题进行分析,给出网络在状态扰动情况下保持稳定的必要条件。仿真结果证明本文所提出的算法能够有效地实现对包含时滞环节的黑箱非线性系统的建模,并能对系统所包含的时滞环节进行辨识。(2)基于通用学习网络的时滞系统控制研究。针对模型未知的时滞系统控制问题,本文提出一种基于通用学习网络的模型预测控制方法。该方法利用本文提出的通用学习网络自适应算法对模型未知的时滞系统进行离线建模,得到该过程的输入输出模型以及系统的滞后时间,进而将辨识得到的神经网络模型作为模型预测器对时滞系统进行预报。在控制结构上,该方法将内模控制结构与神经网络Smith预估控制方法相结合,在神经网络控制器的作用下实现对模型未知时滞系统的控制。以工业生产过程中常见的pH中和过程为例,本文在对pH中和过程内在机理分析的基础上,对其进行系统建模与控制仿真实验,仿真结果表明本文所提出的控制方法对pH中和过程有较好的控制能力,并且控制系统具有良好的鲁棒性。(3)基于一种新型动态前向神经网络的时滞系统辨识与控制研究。针对动态递归神经网络不适用于在线建模的局限,本文提出一种动态前向神经网络用于时滞系统的在线辨识。根据动态前向神经网络的状态方程,本文给出该神经网络的稳定条件。该网络能够根据误差梯度对网络中的连接权值和延迟时间参数进行在线修正,从而能够在充分逼近时滞系统的同时,对时滞系统的延迟时间进行估计。此外,为改善该动态前向神经网络对模型未知时滞系统建模的泛化能力,本文提出一种改进的微粒群算法用于神经网络的在线训练。该算法通过引入白噪声和Logistic映射解决了一般微粒群算法提前收敛的问题,在改善网络泛化能力的同时能够提高网络的学习效率。在动态前向神经网络对时滞系统辨识的基础上,结合鲁棒容错控制结构,本文提出一种基于动态前向神经网络的非线性预测控制系统。仿真实例证明了本文所提出的控制系统具有较强的鲁棒性,能够对模型未知的非线性时滞系统进行有效的辨识和控制。更多还原

【Abstract】 In this article,the modeling and control methods are discussed for time delay systems, which is familiar in industrial process.On the base of dynamic neural network theory,the purpose of this paper is constructing a methodology to identify system parameters and model for unknown time delay systems.Furthermore,an effective control method is proposed for dynamic systems with time delay.Dynamic neural network with adaptive time delay parameters is adopted both online and offline for unknown time delay system identification and controller design in this article.Corresponding theory is also analyzed.The main research contents and research conclusions are listed as follows.(1) Research on nonlinear time delay system identification based on Universal Learning Network(ULN).Due to multiple branches and the arbitrary time delay of ULN,an adaptive algorithm is designed to choose the time delay parameters of ULN and adopt this algorithm to identify the nonlinear process with time delay.The performance of ULN has close relationship with the time delay on the branches which connect with the output node, especially.Based on this the large time delay initial value is adopted for increaseing the converge speed.At the same time,an evaluation function is designed for adjusting time delay parameters of ULN during the training the process.The optimized parameters can be used for the time delay identification of object process.Moreover,the stability analysis of ULN with state disturbance is presented,in order to confirm the stability condition of the adaptive algorithm.The simulation results show that the adaptive algorithm of ULN can not only embody the characters of a blackbox nonlinear system but also can identify the pure time delay of the object system well.(2) Research on nonlinear time delay control based on Universal Learning Network.For the time delay system control,a ULN based predictive control method is mentioned in this article.According to the method,a input/output off-line model and time delay parameter of a unknown time delay process is identified by ULN.The ULN that can be a neural network predictor is used for a prediction model of the time delay systems.For the control structure,a neural network based Smith predictive control with internal model control structure is presented for unknown time delay system control.In modern industry processes,due to nonlinear character and long time delay part,pH neutralization process is highly difficult to control such process.Based on the internal mechanism analysis of pH neutralization,a simulation of the proposed method is given.The simulation results show that the proposed controllers can stabilize the pH neutralization process effectively,and have a good robustness.(3) Research on time delay system identification and control based on a novel dynamic neural network.To overcome the limitations of Dynamic recurrent neural network that is a slow convergence method and not suitable for on-line identification,a feedforward neural network with dynamic neuron is presented in this article.The stability of the network is given by the state equation of the network.The gradient descent method is adopted to optimize both weights and time delays parameters.The proposed algorithm can not only embody the dynamic behavior of a poorly known time delay system,but also identify the pure time delay very well.In order to improve the generalization ability of neural networks for poorly known nonlinear dynamic system with long time delay,a modified Particle Swarm Optimization algorithm is proposed for neural network training.Otherwise,to overcome the particles’ premature convergence,the white noise and Logistic mapping are used to enhance the particles’ search performance and learning efficiency.Based on the dynamic nueral network and robust fault-tolerant control structure,a predictive method is proposed for unknown nonlinear systems with time delay.Finally,simulation results show that the proposed controllers can stabilize the unknown nonlinear time delay systems effectively with robustness.更多还原