大时滞系统的神经网络控制方法研究
|
摘要
工业生产的快速发展使得工业系统更为复杂,大时滞、时变性、严重非线性给工业控制系统提出了更高的要求。具有时滞特性的控制系统是普遍存在的,对于大时滞系统的控制是很困难的。传统的控制方法往往建立在被控对象的数学模型的基础之上,然而在实际中,由于控制对象是电子、机械、软件及其现场环境的复合体,因此很难以建立准确的数学模型,有些甚至根本无法建立模型。由于神经网络具有自学习、自组织以及能逼近任意非线性函数等诸多优良的特点,因而应用神经网络理论与方法对大时滞工业系统的控制已成为了热门研究课题。大量的研究成果表明基于神经网络的控制系统为大时滞系统的控制提供了新的途径。
目前基于神经网络的大时滞控制系统主要分为两种结构,一种是把神经网络与常规PID控制器相结合的控制结构,另一种是无需附加PID控制器,仅用神经网络本身的控制结构。本文对这两种结构各选取一种神经网络对大时滞系统进行研究。论文中的BP神经网络PID控制系统就属于第一种结构,在这种控制结构中,首先通过神经网络自学习的功能自动调整控制参数,然后把这些参数送入PID控制器中进行整合,从而形成控制律对被控对象进行控制。本文分析了BP神经网络的结构,详细的给出了BP的前向算法、反传算法,对BP算法的易陷入局部极小值和收敛速度慢两大缺点进行了重点分析,并提出了相应的改进方法,确定了BP神经网络PID控制系统的结构及算法,最后分别用标准的BP算法与改进Vogl算法对一阶大时滞的加热炉进行了仿真,仿真结果表明改进的Vogl算法的控制品质明显好于标准的BP算法。PID神经元网络控制系统就属于第二种控制结构,在这种控制结构中,分析了具有动态处理能力的PID神经元,给出了PID神经元网络控制系统的框图以及算法,并对一个典型的二阶大时滞对象进行了仿真,从阶跃响应,抗干扰能力,鲁棒性等方面分析该算法的实用性。
文章最后还重点讨论了PID神经元网络与Smith预估器相结合的控制系统,给出了结合后的算法及相关程序并进行了仿真,仿真表明结合后的控制系统在快速性、抗干扰性方面比仅用PID神经元网络控制有很大的提高。
The rapid development of industrial production made the industrial systems more complex. The characters of large time delay, time-varing and high nonlinear present the higher requirements for the control systems of industry. The control systems with these characters widely exist. It is very difficult to control the system with large time delay. Traditional control methods generally base on the mathematical model of controlled object. In fact, because the control object is a complex of the electronic, mechanical, software and on-site environment. it is difficult to establish accurate mathematical model, it is also impossible to do so for some controlled objects. The neural networks have many good features such as the self-learning and self-organization, they can approximate any nonlinear function, so the application and theory of the neural network are used for the control of the large-time delay industrial systems. Extensive research results show that control systems based on neural network provide new ways for the control of large delay system.
Now, the control systems of large time delay based on neural network have two types of structures, one is the combination of the neural network and conventional PID controller, another has the structure which only use neural network itself. The paper selects a style of neural network from these two structures. BP neural network with PID controller is the first structure, in which the control parameters are firstly adjusted by BP neural network with self-learning function, then these parameters are put into the PID controller to form the control laws to control the object. This paper analyzed the structure of BP neural network, presented the forward and back algorithm of BP in detail, and discussed the two major shortcomings of the BP algorithm which are the problem of falling into local minima and slow convergence. It presented the improved methods, gave the structures and algorithms of the BP neural network with PID control system. The paper used the standard BP algorithm and improved Vogl algorithm to stimulate the Furnace control with the first order large time delay. The simulation results show that the control quality in Vogl algorithm is better than the standard Bp algorithm. PID neural network control system belongs to the second control structure, in this structure, the paper analyzed the PID neural with the dynamic processing power, presented the block diagram and algorithm of the PID neural control system. By making simulations for the object of the typical second-order large time-delay, the paper discussed the practicality of the algorithm from the step response, anti-jamming capability and robustness.
Finally, the paper mainly discussed the control system of the PID neural network combined with the Smith predictor, gave the combination algorithm and associated procedures and simulation. The simulation results show that the combined control system has been improved more than the one which only uses PID neural network control system.
目前基于神经网络的大时滞控制系统主要分为两种结构,一种是把神经网络与常规PID控制器相结合的控制结构,另一种是无需附加PID控制器,仅用神经网络本身的控制结构。本文对这两种结构各选取一种神经网络对大时滞系统进行研究。论文中的BP神经网络PID控制系统就属于第一种结构,在这种控制结构中,首先通过神经网络自学习的功能自动调整控制参数,然后把这些参数送入PID控制器中进行整合,从而形成控制律对被控对象进行控制。本文分析了BP神经网络的结构,详细的给出了BP的前向算法、反传算法,对BP算法的易陷入局部极小值和收敛速度慢两大缺点进行了重点分析,并提出了相应的改进方法,确定了BP神经网络PID控制系统的结构及算法,最后分别用标准的BP算法与改进Vogl算法对一阶大时滞的加热炉进行了仿真,仿真结果表明改进的Vogl算法的控制品质明显好于标准的BP算法。PID神经元网络控制系统就属于第二种控制结构,在这种控制结构中,分析了具有动态处理能力的PID神经元,给出了PID神经元网络控制系统的框图以及算法,并对一个典型的二阶大时滞对象进行了仿真,从阶跃响应,抗干扰能力,鲁棒性等方面分析该算法的实用性。
文章最后还重点讨论了PID神经元网络与Smith预估器相结合的控制系统,给出了结合后的算法及相关程序并进行了仿真,仿真表明结合后的控制系统在快速性、抗干扰性方面比仅用PID神经元网络控制有很大的提高。
The rapid development of industrial production made the industrial systems more complex. The characters of large time delay, time-varing and high nonlinear present the higher requirements for the control systems of industry. The control systems with these characters widely exist. It is very difficult to control the system with large time delay. Traditional control methods generally base on the mathematical model of controlled object. In fact, because the control object is a complex of the electronic, mechanical, software and on-site environment. it is difficult to establish accurate mathematical model, it is also impossible to do so for some controlled objects. The neural networks have many good features such as the self-learning and self-organization, they can approximate any nonlinear function, so the application and theory of the neural network are used for the control of the large-time delay industrial systems. Extensive research results show that control systems based on neural network provide new ways for the control of large delay system.
Now, the control systems of large time delay based on neural network have two types of structures, one is the combination of the neural network and conventional PID controller, another has the structure which only use neural network itself. The paper selects a style of neural network from these two structures. BP neural network with PID controller is the first structure, in which the control parameters are firstly adjusted by BP neural network with self-learning function, then these parameters are put into the PID controller to form the control laws to control the object. This paper analyzed the structure of BP neural network, presented the forward and back algorithm of BP in detail, and discussed the two major shortcomings of the BP algorithm which are the problem of falling into local minima and slow convergence. It presented the improved methods, gave the structures and algorithms of the BP neural network with PID control system. The paper used the standard BP algorithm and improved Vogl algorithm to stimulate the Furnace control with the first order large time delay. The simulation results show that the control quality in Vogl algorithm is better than the standard Bp algorithm. PID neural network control system belongs to the second control structure, in this structure, the paper analyzed the PID neural with the dynamic processing power, presented the block diagram and algorithm of the PID neural control system. By making simulations for the object of the typical second-order large time-delay, the paper discussed the practicality of the algorithm from the step response, anti-jamming capability and robustness.
Finally, the paper mainly discussed the control system of the PID neural network combined with the Smith predictor, gave the combination algorithm and associated procedures and simulation. The simulation results show that the combined control system has been improved more than the one which only uses PID neural network control system.
引文
[1]白瑞林,李军.大纯滞后系统的Smith-NN预估控制.电子测量与仪器学报,2000,14(4)：40-44
[2]汤伟.大时滞过程自整定控制及其在制浆造纸工业中的应用：[上海交通大学博士论文].上海：上海交通大学,2002,1-10
[3]黄忠霖.控制系统MATLAB计算及仿真.第一版.北京：国防工业出版社,2001,340-350
[4]喻德宏.大时滞系统智能控制研究：[昆明理工大学硕士论文].昆明：昆明理工大学,2005,1-10
[5]诸丽丽.大时滞系统参数自整定控制的研究：[西南交通大学硕士论文].成都：西南交通大学,2008,2-10
[6]高国燊,余文烋.自动控制原理.第1版.广州：华南理工大学出版社2001,1-10
[7]李剑,谷俊杰.PID参数整定方法进展.电力情报,2001,4(3)：11-13
[8]邵惠鹤.工业过程高级控制.上海：上海交通大学出版社,1997,100-110
[9]ASTROM KJ, HANG CC, LIM BC. A new Smith predictor for controlling a process with an integrator and long dead-time. IEEE Trans on Automtic Control,1996,41(8):1199-1203
[10]庞国仲,孙丽华,刘军,薛福珍.多变量时滞系统鲁棒稳定性.自动化学报,1997,23(1)：99-102
[11]TAN Yonghong, Achiel R. Van Cauwenberghe. Neural Network Based on Nonlinear Smith Predictor.控制理论与应用,2000,17(3)：410-414
[12]鲁照权,韩江洪.一种新型增益自适应Smith预估器.仪器仪表学报,2002,23(2)：195-199.
[13]HANG C C. A performance study of control system with dead time. IEEE Trans IECI,1980,27(1):234-241
[14]WATANABE K A. Process-model control for linear system with delay. IEEE Trans-AC,1981,26(6):261-269
[15]刘欣,尹绍清,肖顺达.模糊自适应Smith预估控制及应用.控制理论与应用,1993,10(4)：261-269
[16]李钟慎.加入滞后时间削弱器的大滞后系统神经网络PID控制.电子测量与仪器学报,2008,22(3)：51-54
[17]王堃,王广军.基于SMITH预估的神经网络再热汽温控制.计算机仿真,2008,25(1)：256-258
[18]刘开培,吕鹏刚,黄天成.时滞系统几种控制算法的相互关系及其近似实现.武汉大学学报,2002,35(2)：83-87
[19]张卫东,许晓鸣,孙优贤.单变量Dahlin控制器设计的新方法.自动化学报,1999,25(3)：360-364
[20]Dahlin E B. Desiging and tuning digital controllers. Instr.&. sys.1968, 41(7):77-83
[21]ZafAnou E. M Morran. Digital controllers for SISO systems:a review and a new algorithm. Int. J. Control.1985,42(4):855-877
[22]赵明旺,王杰,江卫华.现代控制理论.武汉：华中科技大学出版社,2007,3-20
[23]钱积新,赵均,徐祖华.预测控制.北京：化学工业出版社,2007,5-20
[24]郭瑞青,程启明,杜许峰,郑勇.大时滞过程的控制方法.上海电力学院学报,2008,24(3)：248-254
[25]孙秀桂,陆绮荣,赵政春.大时滞过程的控制系统的控制策略分析.安阳工学院学报,2006,21(3)：4-7
[26]易继锴,侯媛彬.智能控制技术.第1版.北京：北京工业大学出版社1999,1-10
[27]张弘.大滞后系统控制中专家一模糊PID方法的应用.计算机工程与应用,2009,45(28)：244-248
[28]吴敏,唐朝晖.锌湿法冶炼电解过程的神经网络专家控制.自动化学报2001,27(6)：867-869
[29]文定都.一类纯滞后系统Fuzzy-Dahlin控制策略的研究.工业仪表与自动化装置,2007,(4)：3-5
[30]罗安.大时滞系统模糊控制.工业加热,1996,(2)：14-18
[31]舒怀林.PID神经元网络及其控制系统.北京：国防工业出版社,2006,1-10
[32]葛哲学,孙志强.神经网络理论与MATLABR2007实现.第1版.北京：电子工业出版社,2007,1-10
[33]吴简彤,王建华.神经网络技术及其应用.第1版.北京：哈尔滨船舶工程学院出版社,1998,5-20
[34]易继锴,侯媛彬.智能控制技术.第1版.北京：北京工业大学出版社,1999,95-130
[35]阎平凡,张永水.人工神经网络与模拟进化算法.北京：清华大学出版社,2000，1-10
[36]焦李成.神经网络计算.第1版.西安：西安电子科技大学出版社,1993,13-32
[37]虞和济,陈长征,张省.基于神经网络的智能故障诊断.第1版.北京：冶金工业出版社,2002,15-40
[38]王耀南,孙炜.智能控制理论及应用.北京：机械工业出版社,2008,1-10
[39]周祖德,史玉升,陈幼平等.基于神经网络的智能故障诊断系统的开发研究.机械与电子,2001,1(4)：33-40
[40]刘金锟.先进PID控制及其MATLAB仿真.北京：电子工业出版社,2003,1-10
[41]冯少辉,周平,钱锋.一种确定神经网络初始权值的新方法.工业仪表与自动化装置,2006,22(1)：65-69
[42]乔双,董智红.BP网络初始权值的选取方法.东北师大学报自然科学版,2004,36(3)：25-31
[43]Dautenhahn K. Book review:Swarm intelligence[J]. Genetic Programming and Evolvable Machines,2002,3(1):93-97
[44]Clerc M,Kennedy J. The particle swarm-explosion,stability,and convergence in a multidimensional complex space[J]. IEEETrans. Evolutionarycomputation, 2002,6(1):58-73
[45]王晓萍,黄海,蒋化冰.BP神经网络Vogl快速算法的改进.浙江大学学报,2000,34(2)：143-146
[46]吕俊,张兴华.几种快速BP算法的比较研究.现代电子技术,2005,167(24)：96-100
[47]V Puiz de Angulo, Carem Torras. On-time learning with minimal degradation in feedforward networks. Neuro COLT Technical Report Series,1994, (16):1-30
[48]陈夕松,汪木兰.过程控制系统.第1版.北京：科学出版社,2002,129-133
[49]AstromK J Hang C C. Toards Intelligen PID Control. Automation,1992,28(1): 1-9
[50]薛定宇.反馈控制系统设计与分析-MATLAB语言应用.北京：清华大学出版社,2000,250-260
[51]舒怀林.PID神经元网络及其控制系统.北京：国防工业出版社,2006,25-50
[2]汤伟.大时滞过程自整定控制及其在制浆造纸工业中的应用：[上海交通大学博士论文].上海：上海交通大学,2002,1-10
[3]黄忠霖.控制系统MATLAB计算及仿真.第一版.北京：国防工业出版社,2001,340-350
[4]喻德宏.大时滞系统智能控制研究：[昆明理工大学硕士论文].昆明：昆明理工大学,2005,1-10
[5]诸丽丽.大时滞系统参数自整定控制的研究：[西南交通大学硕士论文].成都：西南交通大学,2008,2-10
[6]高国燊,余文烋.自动控制原理.第1版.广州：华南理工大学出版社2001,1-10
[7]李剑,谷俊杰.PID参数整定方法进展.电力情报,2001,4(3)：11-13
[8]邵惠鹤.工业过程高级控制.上海：上海交通大学出版社,1997,100-110
[9]ASTROM KJ, HANG CC, LIM BC. A new Smith predictor for controlling a process with an integrator and long dead-time. IEEE Trans on Automtic Control,1996,41(8):1199-1203
[10]庞国仲,孙丽华,刘军,薛福珍.多变量时滞系统鲁棒稳定性.自动化学报,1997,23(1)：99-102
[11]TAN Yonghong, Achiel R. Van Cauwenberghe. Neural Network Based on Nonlinear Smith Predictor.控制理论与应用,2000,17(3)：410-414
[12]鲁照权,韩江洪.一种新型增益自适应Smith预估器.仪器仪表学报,2002,23(2)：195-199.
[13]HANG C C. A performance study of control system with dead time. IEEE Trans IECI,1980,27(1):234-241
[14]WATANABE K A. Process-model control for linear system with delay. IEEE Trans-AC,1981,26(6):261-269
[15]刘欣,尹绍清,肖顺达.模糊自适应Smith预估控制及应用.控制理论与应用,1993,10(4)：261-269
[16]李钟慎.加入滞后时间削弱器的大滞后系统神经网络PID控制.电子测量与仪器学报,2008,22(3)：51-54
[17]王堃,王广军.基于SMITH预估的神经网络再热汽温控制.计算机仿真,2008,25(1)：256-258
[18]刘开培,吕鹏刚,黄天成.时滞系统几种控制算法的相互关系及其近似实现.武汉大学学报,2002,35(2)：83-87
[19]张卫东,许晓鸣,孙优贤.单变量Dahlin控制器设计的新方法.自动化学报,1999,25(3)：360-364
[20]Dahlin E B. Desiging and tuning digital controllers. Instr.&. sys.1968, 41(7):77-83
[21]ZafAnou E. M Morran. Digital controllers for SISO systems:a review and a new algorithm. Int. J. Control.1985,42(4):855-877
[22]赵明旺,王杰,江卫华.现代控制理论.武汉：华中科技大学出版社,2007,3-20
[23]钱积新,赵均,徐祖华.预测控制.北京：化学工业出版社,2007,5-20
[24]郭瑞青,程启明,杜许峰,郑勇.大时滞过程的控制方法.上海电力学院学报,2008,24(3)：248-254
[25]孙秀桂,陆绮荣,赵政春.大时滞过程的控制系统的控制策略分析.安阳工学院学报,2006,21(3)：4-7
[26]易继锴,侯媛彬.智能控制技术.第1版.北京：北京工业大学出版社1999,1-10
[27]张弘.大滞后系统控制中专家一模糊PID方法的应用.计算机工程与应用,2009,45(28)：244-248
[28]吴敏,唐朝晖.锌湿法冶炼电解过程的神经网络专家控制.自动化学报2001,27(6)：867-869
[29]文定都.一类纯滞后系统Fuzzy-Dahlin控制策略的研究.工业仪表与自动化装置,2007,(4)：3-5
[30]罗安.大时滞系统模糊控制.工业加热,1996,(2)：14-18
[31]舒怀林.PID神经元网络及其控制系统.北京：国防工业出版社,2006,1-10
[32]葛哲学,孙志强.神经网络理论与MATLABR2007实现.第1版.北京：电子工业出版社,2007,1-10
[33]吴简彤,王建华.神经网络技术及其应用.第1版.北京：哈尔滨船舶工程学院出版社,1998,5-20
[34]易继锴,侯媛彬.智能控制技术.第1版.北京：北京工业大学出版社,1999,95-130
[35]阎平凡,张永水.人工神经网络与模拟进化算法.北京：清华大学出版社,2000，1-10
[36]焦李成.神经网络计算.第1版.西安：西安电子科技大学出版社,1993,13-32
[37]虞和济,陈长征,张省.基于神经网络的智能故障诊断.第1版.北京：冶金工业出版社,2002,15-40
[38]王耀南,孙炜.智能控制理论及应用.北京：机械工业出版社,2008,1-10
[39]周祖德,史玉升,陈幼平等.基于神经网络的智能故障诊断系统的开发研究.机械与电子,2001,1(4)：33-40
[40]刘金锟.先进PID控制及其MATLAB仿真.北京：电子工业出版社,2003,1-10
[41]冯少辉,周平,钱锋.一种确定神经网络初始权值的新方法.工业仪表与自动化装置,2006,22(1)：65-69
[42]乔双,董智红.BP网络初始权值的选取方法.东北师大学报自然科学版,2004,36(3)：25-31
[43]Dautenhahn K. Book review:Swarm intelligence[J]. Genetic Programming and Evolvable Machines,2002,3(1):93-97
[44]Clerc M,Kennedy J. The particle swarm-explosion,stability,and convergence in a multidimensional complex space[J]. IEEETrans. Evolutionarycomputation, 2002,6(1):58-73
[45]王晓萍,黄海,蒋化冰.BP神经网络Vogl快速算法的改进.浙江大学学报,2000,34(2)：143-146
[46]吕俊,张兴华.几种快速BP算法的比较研究.现代电子技术,2005,167(24)：96-100
[47]V Puiz de Angulo, Carem Torras. On-time learning with minimal degradation in feedforward networks. Neuro COLT Technical Report Series,1994, (16):1-30
[48]陈夕松,汪木兰.过程控制系统.第1版.北京：科学出版社,2002,129-133
[49]AstromK J Hang C C. Toards Intelligen PID Control. Automation,1992,28(1): 1-9
[50]薛定宇.反馈控制系统设计与分析-MATLAB语言应用.北京：清华大学出版社,2000,250-260
[51]舒怀林.PID神经元网络及其控制系统.北京：国防工业出版社,2006,25-50