基于神经网络的多模型自适应控制方法研究

发布时间：2018-01-07 16:35

本文关键词：基于神经网络的多模型自适应控制方法研究　出处：《北京科技大学》2017年博士论文　论文类型：学位论文

【摘要】：随着经济的快速发展科技的进步,生产制造过程逐渐变得复杂多变,生产工况环境也越来越多样化,这就对控制品质提出了新的严格要求。在某些实际控制过程中,一些偶然突发情况(如某一个零件磨损或突然脱落),都将使被控对象的模型瞬间发生剧烈变化。传统自适应控制方法针对的被控对象通常是基于一个参数不变或缓慢变化的模型,操作环境是时不变或慢时变的。然而在诸如零部件失灵,以及一些意想不到的故障发生时,系统动力学模型将发生突变。对于系统的突发性变化,瞬态误差往往很大,传统的自适应控制算法收敛速度很低,控制效果往往表现不佳。多模型自适应控制被视为是解决上述问题最行之有效的方法,该方法的核心要点是:根据被控对象有可能存在的不同工作点,构建含有多个模型的模型集合覆盖被控对象的不确定性;根据模型集合内每一个模型设计相对应的控制器,进而形成控制器集;依据每个模型与被控对象之间的辨识误差,设计基于此误差的切换准则。一旦被控系统参数发生变化,根据切换准则,系统会从模型集中选择最接近当前被控系统的模型,并将控制器切换到该模型的控制器上。基于此思想,本文重点针对实际生产过程中的大量非线性系统,建立基于神经网络的多模型自适应控制器,在被控对象不同的工作点处,建立多个模型,将被控对象参数的不确定性转化为神经网络模型权值的不同。主要研究成果如下:1.基于动态神经网络(Dynamic Neural Networks),分别从动态神经网络的两种典型结构:并行结构和串并结构两种形式出发,考虑系统的有无未建模动态情况,建立了多个动态神经网络辨识模型(自适应模型、固定模型、重新赋初值自适应模型),并对多种动态神经网络辨识模型进行组合,构建了多种动态神经网络组合下的模型集、控制器集以及相应的切换准则,给出了多种组合下控制效果的对比。同时给出了系统的稳定性和切换稳定性的证明。2.基于静态神经网络,提出了基于OS-ELM(On-line Sequence Extreme Learning Machine)神经网络的多模型自适应控制。给出了基于OS-ELM神经网络的模型集、控制器集及切换准则,同时给出了系统的稳定性和切换稳定性证明。3.基于 OS-ELM 和 EM-ELM(Error Minimum Extreme Learning Machine)神经网络,提出了一种自组织神经网络结构即OEM-ELM(On-line Error Minimum Extreme Learning Machine)神经网络。OEM-ELM 算法的核心思想是:在线学习、网络性能评价、动态增加隐层节点数。它将OS-ELM和EM-ELM的优点相结合,既提高了网络的辨识能力又避免了网络节点的冗余。给出了基于OEM-ELM神经网络自适应控制的应用,同时分析了节点变化对系统的影响。4.在钢铁废渣循环利用矿渣微粉生产系统中,将本文的研究成果进行应用。分析了矿渣微粉的生产工艺流程,梳理了影响矿渣微粉比表面积、磨内压差的关键因素。由于实际验证的局限性,本文基于现场采集的大量实际生产数据建立组涵盖多个工况的动态神经网络模型作为生产运行环境。基于此生产运行环境,构建了一种基于OEM-ELM神经网络自适应控制器,在测试验证环境中进一步验证本文提出算法的有效性。
[Abstract]:With the rapid development of economy and the progress of technology, the manufacturing process becomes complicated, production environment is becoming more and more diversified, the quality control strictly required new. In some actual control process, some accidental contingencies (such as some parts wear or suddenly fall off), will make the object the model of controlled moment change dramatically. The traditional adaptive control method for the controlled object is usually based on a constant or slowly changing parameters of the model, the operating environment is time invariant or slowly time-varying. However, such as zero component failures, and some unexpected failures, the system dynamics model for the burst will change. The change of the system, the transient error is often large, the convergence speed of the traditional adaptive control algorithm is very low, the control effect is often poor performance. Multi model adaptive control As is the most effective way to solve the above problems, the core point of this method is: according to the object may have different working points, constructing a set covering the uncertainty of the controlled object with multiple model controller; according to the model set in each model is designed correspondingly, and the formation of controller basis; the identification error between each model and object design, the error based on the switching rule. Once the parameters of the controlled system change, according to a switching rule, the system will choose the most close to the current control system model from the model, and the controller is switched to the controller model. Based on this idea, this paper focuses on the nonlinear the system in actual production process, establish a multiple model adaptive controller based on neural network, the object in different work point, build multiple models, The object parameter uncertainty into the neural network model of different weights. The main research results are as follows: 1. based on dynamic neural network (Dynamic Neural Networks), respectively from the two kinds of typical structure of dynamic neural network: parallel structure and string and the structure of two forms of the system without considering unmodeled dynamics. To establish a number of dynamic neural network model (adaptive model, fixed model, rein itialized model), and the combination of a variety of dynamic neural network model, constructed a combination of dynamic neural network model set, the controller set and the corresponding switching rule, contrast gives the effect of a variety of combinations under control. At the same time proved the stability and stability of the system of the.2. static switch based on neural network is proposed based on OS-ELM (On-line Sequence Extreme Learning Machine) Multi model adaptive neural network control is presented. The OS-ELM neural network model based on set, controller and switching criterion, and the stability and stability of switching system prove that the OS-ELM and.3. based on EM-ELM (Error Minimum Extreme Learning Machine) neural network, a self-organizing neural network structure which is OEM-ELM (On-line Error Minimum Extreme Learning Machine) the core idea of.OEM-ELM algorithm of neural network is: online learning, network performance evaluation, the dynamic increase of the number of hidden nodes. The advantages of OS-ELM and EM-ELM combination, the identification ability of the network is improve and avoid redundant network nodes. The application of OEM-ELM neural network based on adaptive control is given. At the same time, analysis of the node changes the influence on the system of.4. in iron and steel slag recycling slag powder production system, this paper will research into The fruit of the application. Analyzed the production process of slag powder, analyzes the impact of slag powder specific surface area, the key factors in the grinding pressure. Due to the limitation of the actual verification, the large number of actual production data collected based on the dynamic conditions of God Group covers a number of the network as a production environment. This based on the production environment, to construct a neural network OEM-ELM controller based on adaptive, further verify the effectiveness of the proposed algorithm in the test environment.

【学位授予单位】：北京科技大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP273.2;TP183

【参考文献】