数学物理学报  2018, Vol. 38 Issue (3): 588-598   PDF    
扩展功能
加入收藏夹
复制引文信息
加入引用管理器
Email Alert
RSS
本文作者相关文章
李向东
傅勤
吴健荣
基于有限差分法的二阶双曲型分布参数系统的迭代学习控制
李向东, 傅勤, 吴健荣     
苏州科技大学 数理学院 苏州 215009
摘要:研究了一类分布参数系统基于有限差分法的迭代学习控制问题,该类分布参数系统由二阶双曲型偏微分方程所构建.针对系统所满足的初、边值条件,基于有限差分法构建得到迭代学习控制律,利用迭代收敛原理,证明这种学习律能使得系统的状态跟踪误差沿迭代轴方向收敛到原点的一个小邻域内.数值仿真验证了所提算法的有效性.
关键词双曲型分布参数系统    学习律    迭代学习控制    有限差分法    
Iterative Learning Control for Second Order Hyperbolic Distributed Parameter Systems Based on Finite Differential Method
Li Xiangdong, Fu Qin, Wu Jianrong     
School of Mathematics and Physics, Suzhou University of Science and Technology, Suzhou 215009
Abstract: In this paper, the iterative learning control for distributed parameter systems based on finite differential method is considered. Here, the considered distributed parameter systems are composed of second order hyperbolic partial differential equations. According to the initialboundary conditions of the systems, iterative learning control laws are proposed based on finite differential method. Using the iterative converge principle, we show that the schemes can guarantee the state tracking errors can converge to a small region of the origin along the iteration axis. Finally, the simulation result verifies the effectiveness of the proposed algorithm.
Key words: Hyperbolic distributed parameter systems     Learning scheme     Iterative learning control     Finite differential method    
1 引言

迭代学习控制由Arimoto等人[1]首次提出完整的控制算法后, 已成为近年来控制理论研究的热点问题, 并引起人们的广泛关注[2-6].针对在有限时间区间上具有相同重复运动性质的受控系统, 迭代学习控制利用目标轨迹与系统输出信号之间的误差来构建控制输入, 通过反复迭代学习修正, 使得系统在有限时间区间上渐近跟踪给定的目标轨迹.由于迭代学习控制具有算法简单有效的特点, 被广泛应用于各大领域中, 见文献[7].如今, 迭代学习控制已经成为智能控制领域中的一个重要研究分支.

由偏微分方程, 或偏微分-积分方程, 或偏微分方程与常微分方程耦合的方程描述的控制系统称为分布参数系统.许多实际问题都可以用分布参数系统模型来刻画[8], 如弹性振动系统的控制、温度场的控制、核反应堆的控制、带柔性连杆的机器人等.近年来, 分布参数系统的迭代学习控制引起了人们的关注[8-14], 涉及到两类常见的系统:二阶抛物型(由热传导方程构建而成)系统和二阶双曲型(由波动方程构建而成)系统.文献[8-11]研究了二阶正则(系统的输入输出有直输通道)抛物型系统的迭代学习控制问题; 文献[12]则将迭代学习控制方法应用到了二阶正则双曲型系统上; 文献[13]进一步于Sobolev空间$W^{1, 2}$中研究了由抛物型方程或双曲型方程构成的正则分布参数系统的迭代学习控制问题.另一方面, 对于非正则(系统的输入输出无直输通道)分布参数系统, 由于其控制输入须通过状态方程作用于输出上, 而状态方程中又含有状态变量对空间自变量的偏导数, 由此导致其迭代学习控制设计相应于正则系统要困难的多.最近, 利用泛函分析中的弱收敛性质, 文献[14]解决了二阶非正则抛物型系统的迭代学习控制问题.值得关注的是, 对于二阶非正则双曲型系统, 在迭代学习控制设计方面, 至今尚无任何相关的研究成果.众所周知, 偏微分方程有多种数值解法, 如常用的有限差分法、有限元法等.近年来, 基于数值解法的迭代学习控制设计已引起了人们的重视[15-18].针对一类二阶非正则抛物型系统, 基于有限差分法, 文献[15-16]分别利用显格式和Crank-Nicholson格式对系统离散化, 再由离散化差分方程构建基于时空动态模型的线性离散时间系统, 由此解决了非正则离散抛物型系统的迭代学习控制问题.文献[17-18]则借助于离散形式的Gronwall不等式解决了一类正则离散抛物型系统的迭代学习控制问题.基于上述分析, 自然而然, 二阶非正则双曲型系统的迭代学习控制问题引起了我们的关注.

本文研究二阶非正则双曲型分布参数系统基于有限差分法的迭代学习控制问题.先对由空间变量和时间变量组建的矩形区域进行矩形网格剖分, 再针对迭代系统和理想轨迹分别构建相应的差分求解格式(差分方程), 然后在节点处分别求得迭代系统和理想轨迹的有限差分解.一方面, 利用差分格式的收敛性, 说明当网格剖分很细(即每个小矩形网格的直径足够小)时, 迭代系统和理想轨迹在节点处的有限差分解均能无限逼近各自系统准确解在相应节点处的值.另一方面, 针对由迭代系统所构建的差分方程, 利用迭代系统和理想轨迹的有限差分解构建迭代学习控制律, 并进一步证明, 当该迭代学习控制律作用于该差分方程, 且迭代次数趋于无穷时, 迭代系统的有限差分解能无限逼近理想轨迹的有限差分解.综上两方面可知, 在迭代学习控制律作用下, 当网格剖分很细, 且迭代次数趋于无穷时, 迭代系统准确解在节点处的值能无限逼近于理想轨迹准确解在相应节点处的值, 由此从数值分析角度解决了二阶非正则双曲型分布参数系统的迭代学习控制问题.最终, 所得结论由数值例子验证.

本文给出如下符号约定:对无穷小$\alpha $, 用$O(\alpha )$表示$\alpha $的同阶无穷小, 对$C \in \mathbb{R}^{n \times n}$, 用$\rho (C)$表示矩阵$C$的谱半径, 用$I$表示适维单位矩阵.

2 问题描述

考虑如下形式的二阶双曲型分布参数系统

$ \frac{{{\partial ^2}Q(x, t)}}{{\partial {t^2}}} = {a^2}\frac{{{\partial ^2}Q(x, t)}}{{\partial {x^2}}} + u(x, t)\begin{array}{*{20}{c}} , &{0 < x < 1, 0 < t \le T, } \end{array} $ (2.1)

这里$a$为大于$0$的常数, $Q(x, t), u(x, t)\in \mathbb{R}$分别是系统的状态和控制输入.

对系统(2.1), 给出如下假设条件:

假设2.1  对于给定的初值$Q(x, 0)$$\left. {\frac{\partial Q(x, t)}{\partial t}} \right|_{t=0} $, 边值$Q(0, t)$$Q(1, t)$, 系统(2.1)的解$Q(x, t)$$(0, 1)\times [0, T]$内存在唯一.

假设2.2  对于给定的理想轨迹$Q_r (x, t)$, 存在唯一的控制输入$u_r (x, t)$, 使得

$ \frac{{{\partial ^2}Q_r(x, t)}}{{\partial {t^2}}} = {a^2}\frac{{{\partial ^2}Q_r(x, t)}}{{\partial {x^2}}} + u_r(x, t)\begin{array}{*{20}{c}} , &{0 < x < 1, 0 < t \le T.} \end{array} $ (2.2)

设动态系统(2.1)在有限区间$t\in [0, T]$内是可重复的, 在迭代学习过程中, 重写系统(2.1)为

$ \frac{{{\partial ^2}Q_k(x, t)}}{{\partial {t^2}}} = {a^2}\frac{{{\partial ^2}Q_k(x, t)}}{{\partial {x^2}}} + u_k(x, t)\begin{array}{*{20}{c}} , &{0 < x < 1, 0 < t \le T, } \end{array} $ (2.3)

$k=0, 1, 2, \cdots $.

假设2.3  系统的初值定位条件为: $Q_k (x, 0)=Q_r (x, 0)$, $\left.{\frac{\partial Q_k (x, t)}{\partial t}} \right|_{t=0} =\left. {\frac{\partial Q_r (x, t)}{\partial t}} \right|_{t=0}$, 边值定位条件为: $Q_k (0, t)=Q_r (0, t)$, $Q_k (1, t)=Q_r (1, t)$, $k=0, 1, 2, \cdots $.

用平行直线族$x = jh(j = 0, 1, \cdots , N)$, $t = i\tau (i = 0, 1, \cdots , M)$对平面区域$[0, 1] \times [0, T]$作矩形网格剖分, 其中$\tau$$h$为正常数, 分别称为时间步长与空间步长, 且$h = \frac{1}{N}$, $\tau = \frac{T}{M}$, 而$N\mbox{、} M$为正整数.

本文考虑平面区域$[0, 1] \times [0, T]$内部(含上边界)节点$({x_j}, {t_i})$ ($j = 1, 2, \cdots , N - 1$, $i = 1, 2, \cdots , M $)处, 系统(2.3)对系统(2.2)基于迭代学习的状态跟踪问题.寻找适当的学习律, 使得迭代系统(2.3)的状态${Q_k}({x_j}, {t_i})$在该学习律作用下跟踪理想轨迹(2.2)的状态${Q_r}({x_j}, {t_i})$, 即对任意给定的正数$\varepsilon $, 当网格($\tau$$h$)足够小, 迭代次数$k$足够大时, 能使

$ \mathop {\max }\limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_r}({x_j}, {t_i}) - {Q_k}({x_j}, {t_i})} \right| < \varepsilon $

成立.

3 有限差分法

对定义在$[0, 1] \times [0, T]$上的二元函数$f(x, t)$, 在节点$({x_j}, {t_i})$处, 用$f(j, i)$表示$f({x_j}, {t_i})$的近似值.

对系统(2.3),采用最简显格式[19],有

$ \begin{array}[b]{rl} {Q_k}(j, i + 1) = &{s^2}{Q_k}(j - 1, i) + 2(1 - {s^2}){Q_k}(j, i) + {s^2}{Q_k}(j + 1, i)\\ &-{Q_k}(j, i - 1) + {\tau ^2}{u_k}({x_j}, {t_i}), \end{array} $ (3.1)

$j = 1, 2, \cdots , N - 1$, $i = 1, 2, \cdots , M - 1$, 其中$s = \frac{{a\tau }}{h}$, 称为网比[19].结合假设2.3可知, 系统(3.1)的边值条件为

$ {Q_k}(0, i) = {Q_k}(0, {t_i}) = {Q_r}(0, {t_i}), ~{Q_k}(N, i) = {Q_k}(1, {t_i}) = {Q_r}(1, {t_i}), ~i= 0, 1, \cdots , M. $ (3.2)

由于(3.1)式中的迭代关于时间变量是两步的, 需给出$i = 0, 1$时的迭代初值, 迭代方能进行.结合假设2.3和(2.3)式, 借助于泰勒展开公式, 给出$i = 1$时的迭代初值的估算值(见文献[19]), 由此迭代系统(3.1)的初值取为

$ {Q_k}(j, 0) = {Q_k}({x_j}, 0) = {Q_r}({x_j}, 0), $
$ \begin{array}[b]{rl} {Q_k}(j, 1) =& {Q_k}({x_j}, 0) + \tau {\left. {\frac{{\partial {Q_k}({x_j}, t)}}{{\partial t}}} \right|_{t = 0}} + \frac{{{\tau ^2}}}{2}{\left. {\frac{{{\partial ^2}{Q_k}({x_j}, t)}}{{\partial {t^2}}}} \right|_{t = 0}} \\[4mm] =& {Q_r}({x_j}, 0) + \tau {\left. {\frac{{\partial {Q_r}({x_j}, t)}}{{\partial t}}} \right|_{t = 0}} + \frac{{{\tau ^2}}}{2}\left[ {{a^2}{{\left. {\frac{{{\partial ^2}{Q_r}(x, 0)}}{{\partial {x^2}}}} \right|}_{x = {x_j}}} + {u_k}({x_j}, 0)} \right], \\ &j = 1, 2 \cdots , N - 1. \end{array} $ (3.3)

同样,对系统(2.2)采用最简显格式,有

$ \begin{array}{rl} {Q_r}(j, i + 1) = &{s^2}{Q_r}(j - 1, i) + 2(1 - {s^2}){Q_r}(j, i) + {s^2}{Q_r}(j + 1, i)\\ &- {Q_r}(j, i - 1) + {\tau ^2}{u_r}({x_j}, {t_i}), \\&j = 1, 2, \cdots , N - 1, ~i= 1, 2, \cdots , M - 1. \end{array} $

边值条件为

$ {Q_r}(0, i) = {Q_r}(0, {t_i}), {Q_r}(N, i) = {Q_r}(1, {t_i}), ~i= 0, 1, \cdots , M. $ (3.4)

初值条件为

$\begin{array}{c}&{Q_r}(j, 0) = {Q_r}({x_j}, 0), \\ &{Q_r}(j, 1) = {Q_r}({x_j}, 0) + \tau {\left. {\frac{{\partial {Q_r}({x_j}, t)}}{{\partial t}}} \right|_{t = 0}} + \frac{{{\tau ^2}}}{2}\left[ {{a^2}{{\left. {\frac{{{\partial ^2}{Q_r}(x, 0)}}{{\partial {x^2}}}} \right|}_{x = {x_j}}} + {u_r}({x_j}, 0)} \right], \\&j = 1, 2, \cdots , N - 1.\end{array} $ (3.5)

选取时间步长$\tau $与空间步长$h$, 使$s = \frac{{a\tau }}{h}$为定常数, 记${e_k}(j, i) = {Q_k}({x_j}, {t_i}) - {Q_k}(j, i)$, ${e_r}(j, i) = {Q_r}({x_j}, {t_i}) - {Q_r}(j, i)$, $i = 0, 1, \cdots , M$, $j = 0, 1, \cdots , N$.由(3.3)和(3.5)式可知${e_k}(j, 0) = 0$, ${e_r}(j, 0) = 0$, 由(3.2)和(3.4)式有${e_k}(0, i) = {e_k}(N, i) = 0$, ${e_r}(0, i) = {e_r}(N, i) = 0$, 所以当$s = \frac{{a\tau }}{h} < 1$时, 有[19-20]

$ \mathop {\max }\limits_{1 \le j \le N - 1} \left| {{e_k}(j, i)} \right| = O({\tau ^2} + {h^2}) , ~\mathop {\max }\limits_{1 \le j \le N - 1} \left| {{e_r}(j, i)} \right| = O({\tau ^2}+ {h^2}), ~i = 1, 2, \cdots , M. $

进一步我们有

$ \mathop {\max }\limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_k}({x_j}, {t_i}) - {Q_k}(j, i)} \right| = O({\tau ^2}+ {h^2}), $ (3.6)
$ \mathop {\max }\limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_r}({x_j}, {t_i}) - {Q_r}(j, i)} \right| = O({\tau ^2}+ {h^2}). $ (3.7)
4 迭代学习控制

$\delta {Q_k}(j, i) = {Q_{k + 1}}(j, i) - {Q_k}(j, i)$, $\delta {u_k}({x_j}, {t_i}) = {u_{k + 1}}({x_j}, {t_i}) - {u_k}({x_j}, {t_i})$, 则由(3.1)式可得

$ \begin{array}[b]{rl} \delta {Q_k}(j, i + 1) =& {s^2}\delta {Q_k}(j - 1, i) + 2(1 - {s^2})\delta {Q_k}(j, i) + {s^2}\delta {Q_k}(j + 1, i)\\ &- \delta {Q_k}(j, i - 1) + {\tau ^2}\delta {u_k}({x_j}, {t_i}), \end{array} $ (4.1)

$j = 1, 2, \cdots , N - 1$, $i = 1, 2 \cdots , M - 1$.由(3.2)和(3.3)式可知, 系统(4.1)满足如下初、边值条件

$ \delta {Q_k}(j, 0) = 0, ~\delta {Q_k}(j, 1) = \frac{{{\tau ^2}}}{2}\delta {u_k}({x_j}, 0), ~j= 1, 2, \cdots, N - 1; $ (4.2)
$ \delta {Q_k}(0, i) = 0, ~\delta {Q_k}(N, i) = 0, ~i= 0, 1, \cdots, M. $ (4.3)

$ \delta \vec Q_k^i = {\left[ {\begin{array}{*{20}{c}} {\delta {Q_k}(1, i)}, &{\delta {Q_k}(2, i)},&\cdots , &{\delta {Q_k}(N - 1, i)} \end{array}} \right]^{\rm T}}, $
$ \delta \vec u_k^i = {\left[ {\begin{array}{*{20}{c}} {\delta {u_k}({x_1}, {t_i})}, &{\delta {u_k}({x_2}, {t_i})},&\cdots , &{\delta {u_k}({x_{N - 1}}, {t_i})} \end{array}} \right]^{\rm T}}, $
$ ~i = 0, 1, \cdots , M. $

结合边值条件(4.3)式, 将系统(4.1)写成向量形式[19]

$ \delta \vec Q_k^{i + 1} = [2(1 - {s^2})I + {s^2}B]\delta \vec Q_k^i - \delta \vec Q_k^{i - 1} + {\tau ^2}\delta \vec u_k^i, $

$i = 1, \cdots , M - 1$, 其中$I$$N - 1$阶单位阵, 且

$ B = {\left[ {\begin{array}{*{20}{c}} 0&1&{~~~~~}&{~~~~~}&{~~~~~}\\ 1&0&1&{~~~~~}&{~~~~~}\\ {~~~~~}&1& \ddots&\ddots &{}\\ {~~~~~}&{~~~~~}& \ddots &0&1\\ {~~~~~}&{~~~~~}&{}&1&0 \end{array}} \right]_{(N - 1) \times (N - 1)}}. $

结合初值条件(4.2)式, 进一步有

$ \left[ {\begin{array}{*{20}{c}} {{{\delta \vec Q_k^2}}}\\ {{{\delta \vec Q_k^3}}}\\ { \cdot \cdot \cdot }\\ {{{\delta \vec Q_k^M}}} \end{array}} \right] = \left[ {\begin{array}{*{20}{c}} A&{~~~~~}&{~~~~~}&{~~~~~}\\ -I&A&{~~~~~}&{~~~~~}\\ {~~~~~}&\ddots& \ddots&{~~~~}\\ {~~~~~}&{~~~~}& -I &A \end{array}} \right]_{(M - 1) \times (M - 1)} \times \left[ {\begin{array}{*{20}{c}} {{{\delta \vec Q_k^1}}}\\ {{{\delta \vec Q_k^2}}}\\ { \cdot \cdot \cdot }\\ {{{\delta \vec Q_k^{M - 1}}}} \end{array}} \right] + {\tau ^2}\left[ {\begin{array}{*{20}{c}} {{{\delta \vec u_k^1}}}\\ {{{\delta \vec u_k^2}}}\\ { \cdot \cdot \cdot }\\ {{{\delta \vec u_k^{M - 1}}}} \end{array}} \right], $ (4.4)

这里$A = 2(1 - {s^2})I + {s^2}B$.由(4.2)式有

$ \delta \vec Q_k^0 = \vec 0, ~\delta \vec Q_k^1 = \frac{{{\tau ^2}}}{2}\delta \vec u_k^0, $ (4.5)

再记

$ \delta {\vec Q_k} = {\left[ {\begin{array}{*{20}{c}} {{{(\delta \vec Q_k^1)}^{\rm T}}}, &{{{(\delta \vec Q_k^2)}^{\rm T}}},&\cdots , &{{{(\delta \vec Q_k^M)}^{\rm T}}} \end{array}} \right]^{\rm T}}, $ (4.6)
$ \delta {\vec u_k} = {\left[ {\begin{array}{*{20}{c}} {{{(\delta \vec u_k^0)}^{\rm T}}}, &{{{(\delta \vec u_k^1)}^{\rm T}}},&\cdots , &{{{(\delta \vec u_k^{M - 1})}^{\rm T}}} \end{array}} \right]^{\rm T}}, $ (4.7)

(4.4)式结合(4.5)-(4.7)式, 有

$ \delta {\vec Q_k} = E\delta {\vec Q_k} + {\tau ^2} G\delta {\vec u_k}, $ (4.8)

这里

$ E = {\left[ {\begin{array}{*{20}{c}} 0&{~~~~~}&{~~~~~}&{~~~~~}&{~~~~~}\\ A&0&{~~~~~}&{~~~~~}&{~~~~~}\\ {-I}&A& \ddots&{~~~~~}&{~~~~~}\\ {~~~~~}& \ddots&\ddots &0&{~~~~~}\\ {~~~~~}&{~~~~~}&{-I}&A&0 \end{array}} \right]_{[M \times (N - 1)] \times [M \times (N - 1)]}}, \\G ={\left[ {\begin{array}{*{20}{c}} { \frac{{1}}{2}I}&{~~~~~}&{~~~~~}&{~~~~~}&{~~~~~}\\ {~~~~~}&{I}&{~~~~~}&{~~~~~}&{~~~~~}\\ {~~~~~}&{~~~~~}& \ddots& {~~~~~}&{~~~~~}\\ {~~~~~}& {~~~~~}&{~~~~~} &{I}&{~~~~~}\\ {~~~~~}&{~~~~~}&{~~~~~}&{~~~~~}&{I} \end{array}} \right]_{[M \times (N - 1)] \times [M \times (N - 1)]}}. $

对系统(2.3), 在节点处构建如下学习律

$ \left\{ {\begin{array}{*{20}{l}} {{u_{k + 1}}({x_j}, {t_0}) = {u_k}({x_j}, {t_0}) + 2q({Q_r}(j, 1) - {Q_k}(j, 1))}, \\ {{u_{k + 1}}({x_j}, {t_i}) = {u_k}({x_j}, {t_i}) + q({Q_r}(j, i + 1) - {Q_k}(j, i + 1))}, \end{array}} \right. $ (4.9)

$j = 1, 2 \cdots , N - 1$, $i = 1, 2 \cdots , M - 1$, 其中$q$为学习增益.

注4.1  相应于准确解,有限差分解${Q_r}(j, i)$${Q_k}(j, i)$是易求得的, 由此可知, 借助于有限差分解构建而成的学习律(4.9)具有较强的实用性.

$ \Delta {Q_k}(j, i) = {Q_r}(j, i) - {Q_k}(j, i), $
$ \Delta \vec Q_k^i = {\left[ {\begin{array}{*{20}{c}} {\Delta {Q_k}(1, i)}, &{\Delta {Q_k}(2, i)},&\cdots , &{\Delta {Q_k}(N - 1, i)} \end{array}} \right]^{\rm T}}, $
$ \Delta {\vec Q_k} = {\left[ {\begin{array}{*{20}{c}} {{{(\Delta \vec Q_k^1)}^{\rm T}}}, &{{{(\Delta \vec Q_k^2)}^{\rm T}}},&\cdots , &{{{(\Delta \vec Q_k^M)}^{\rm T}}} \end{array}} \right]^{\rm T}}, $

则(4.9)式可写为

$ \delta {\vec u_k} =q \times{\left[ {\begin{array}{*{20}{l}} {2I}&{~~~~~}&{~~~~~}&{~~~~~}&{~~~~~}\\ {~~~~~}&I&{~~~~~}&{~~~~~}&{~~~~~}\\ {~~~~~}&{~~~~~}& \ddots &{~~~~~}&{~~~~~}\\ {~~~~~}&{~~~~~}&{~~~~~}&I&{}\\ {~~~~~}&{~~~~~}&{~~~~~}&{~~~~~}&I \end{array}} \right]_{[M \times (N - 1)] \times [M \times (N - 1)]}}\times\Delta {\vec Q_k}. $

将上式代入(4.8)式, 可得

$ \delta {\vec Q_k} = E\delta {\vec Q_k} + {\tau ^2}q\Delta {\vec Q_k}. $

进一步有

$ \delta {\vec Q_k} = {\tau ^2}qF\Delta {\vec Q_k}, $ (4.10)

其中

$ F = {({I_{{{_{[M \times (N - 1)] \times[M \times (N - 1)] }}}}} - E)^{ - 1}}\\= \left[ {\begin{array}{*{20}{c}} I&{~~~~~}&{~~~~~}&{~~~~~}&{~~~~~}\\ { - A}&I&{~~~~~}&{~~~~~}&{~~~~~}\\ I&{ - A}& \ddots &{~~~~~}&{~~~~~}\\ {~~~~~}& \ddots&\ddots &I&{~~~~~}\\ {~~~~~}&{~~~~~}&I&{ - A}&I \end{array}} \right]_{{{_{[M \times (N - 1)] \times[M \times (N - 1)] }}}}^{ - 1}\\= {\left[ {\begin{array}{*{20}{c}} I&{~~~~~}&{~~~~~}&{~~~~~}&{~~~~~}\\ A&I&{~~~~~}&{~~~~~}&{~~~~~}\\ {{A^2} - I}&A& I &{~~~~~}&{~~~~~}\\ \vdots&\ddots&\ddots &\ddots&{~~~~~}\\ \ast& \cdots &{{A^2} - I}&A&I \end{array}} \right]_{{{_{[M \times (N - 1)] \times[M \times (N - 1)] }}}}} . $ (4.11)

注意到

$ \Delta {\vec Q_{k + 1}} = \Delta {\vec Q_k} - \delta {\vec Q_k}. $

将(4.10)式代入上式, 有

$ \Delta {\vec Q_{k + 1}} = ({I_{{{_{[M \times (N - 1)] \times [M \times (N - 1)]}}}}} - {\tau ^2}qF)\Delta {\vec Q_k}. $

结合(4.11)式, 可知

$ \rho ({I_{{{_{[M \times (N - 1)] \times [M \times (N - 1)]}}}}} - {\tau ^2}qF) = \left| {1 - {\tau ^2}q} \right|. $

由迭代收敛原理(迭代矩阵的谱半径小于1, 则迭代收敛[19]), 当

$ \left| {1 - {\tau ^2}q} \right| < 1 $ (4.12)

时, 有

$ \mathop {\lim }\limits_{k \to \infty } \Delta {\vec Q_k} = \vec 0, $

$ \mathop {\lim }\limits_{k \to \infty } \mathop {\max }\limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_r}(j, i) - {Q_k}(j, i)} \right| = 0. $ (4.13)

由(3.6), (3.7), (4.12)和(4.13)式可得本文的主要定理:

定理4.1  假设2.1-2.3成立, 对任意给定的正数$\varepsilon $, 存在正数${\tau _0}, {h_0}$, 正整数$K$, 取$0< \tau < {\tau _0}$, $0< h < {h_0}$, 且$h, \tau $满足$\frac{{a\tau }}{h} < 1$, 则当$k > K$, $0 < q < \frac{2}{{\tau _0^2}}$时, 系统(2.3)在学习律(4.9)的作用下, 有

$ \mathop {\max }\limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_r}({x_j}, {t_i}) - {Q_k}({x_j}, {t_i})} \right| < \varepsilon. $

  对任意给定的正数$\varepsilon $, 由(3.6)和(3.7)式可知, 存在正数${\tau _0}, {h_0}$, 当$0< \tau < {\tau _0}$, $0< h < {h_0}$时, 有

$ \mathop {\max }\limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_k}({x_j}, {t_i}) - {Q_k}(j, i)} \right| < \frac{\varepsilon }{3}, $ (4.14)
$ \mathop {\max }\limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_r}({x_j}, {t_i}) - {Q_r}(j, i)} \right| < \frac{\varepsilon }{3} $ (4.15)

成立.由$0 < q < \frac{2}{{\tau _0^2}}$可知, (4.12)式成立, 由此(4.13)式成立, 即存在正整数$K$, 当$k > K$时, 能使

$ \mathop {\max }\limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_r}(j, i) - {Q_k}(j, i)} \right| < \frac{\varepsilon }{3} $ (4.16)

成立.由(4.14)-(4.16)式, 可推得

$ \begin{eqnarray*} &&\mathop {\max }\limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_r}({x_j}, {t_i}) - {Q_k}({x_j}, {t_i})} \right| \\ & \le& \mathop {\max } \limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_r}({x_j}, {t_i}) - {Q_r}(j, i)} \right| + \mathop {\max }\limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_r}(j, i) - {Q_k}(j, i)} \right| \\ &&+ \mathop {\max } \limits_{1 \le j \le N - 1, 1 \le i \le M} \left| {{Q_k}({x_j}, {t_i}) - {Q_k}(j, i)} \right|\\ & <& \varepsilon. \end{eqnarray*} $

证毕.

5 仿真算例

构建如下形式的二阶双曲型分布参数迭代系统

$ \frac{{{\partial ^2}{Q_k}(x, t)}}{{\partial {t^2}}} = \frac{{{\partial ^2}{Q_k}(x, t)}}{{\partial {x^2}}} + {u_k}(x, t)\begin{array}{*{20}{c}} , &0 < x < 1, 0 < t \le 1, \end{array} $ (5.1)

取初值为${Q_k}(x, 0) = {Q_r}(x, 0) = 0$, ${\left. {\frac{{\partial {Q_k}(x, t)}}{{\partial t}}} \right|_{t = 0}} = {\left. {\frac{{\partial {Q_r}(x, t)}}{{\partial t}}} \right|_{t = 0}} = x$, 边值为${Q_k}(0, t) = {Q_r}(0, t) = 0$, ${Q_k}(1, t) = {Q_r}(1, t) = \sin t$.给定理想轨迹${Q_r}(x, t) = \sin xt$, 则存在控制输入${u_r}(x, t) = \left( {{t^2} - {x^2}} \right)\sin xt$, 使得

$ \frac{{{\partial ^2}{Q_r}(x, t)}}{{\partial {t^2}}} = \frac{{{\partial ^2}{Q_r}(x, t)}}{{\partial {x^2}}} + {u_r}(x, t)\begin{array}{*{20}{c}} , &{0 < x < 1, 0 < t \le 1, } \end{array} $ (5.2)

相应初值为${Q_r}(x, 0) = 0$, ${\left. {\frac{{\partial {Q_r}(x, t)}}{{\partial t}}} \right|_{t = 0}} = x$, 边值为${Q_r}(0, t) = 0$, ${Q_r}(1, t) = \sin t$.选取初始输入控制${u_0}(x, t) = - x\sin t$, 并取$h = 0.025$, $\tau = 0.000625$, 则有$s = 0.025 < 1$.分别对系统(5.1)和(5.2)采用最简显格式, 取$q = 512/3$, 则有$1 - {\tau ^2}q = 0.999933 < 1$.于是在节点处, 系统(2.3)在学习律(4.9)的作用下, 定理4.1成立.仿真结果见图 1-4.

图 1 $k = 0$时, 跟踪误差:${Q_r}(x_j, t_i) - {Q_k}(x_j, t_i)$

图 2 $k = 3$时, 跟踪误差:${Q_r}(x_j, t_i) - {Q_k}(x_j, t_i)$

图 3 $k= 13$时, 跟踪误差:${Q_r}(x_j, t_i) - {Q_k}(x_j, t_i)$

图 4 最大误差: $\max\limits_{1 \le j \le N - 1, 1 \le i \le M} \left| Q_r(x_j, t_i) - Q_k(x_j, t_i) \right|$
6 结论

本文研究了一类分布参数系统基于有限差分法的迭代学习控制问题, 该系统由二阶双曲型偏微分方程所构建, 并具有适定的初、边值定解条件.本文借助于差分方程的解, 构建起迭代系统的状态与理想轨迹状态之间的联系, 由此说明在网格节点处, 系统跟踪误差的最大值可以小于任意给定的正数, 得到了很好的结论, 仿真结果说明如此.

参考文献
[1] Arimoto S, Kawamura S, Miyazaki F. Bettering operation of robots by learning. Journal of Robotic Systems, 1984, 1: 123–140. DOI:10.1002/(ISSN)1097-4563
[2] Sun M, Wang D. Iterative learning control with initial rectifying action. Automatica, 2002, 38: 1177–1182. DOI:10.1016/S0005-1098(02)00003-1
[3] Freeman C T, Cai Z L, Rogers E, Lewin P L. Iterative learning control for multiple point-to-point tracking application. IEEE Transaction on Control Systems Technology, 2011, 19: 590–600. DOI:10.1109/TCST.2010.2051670
[4] Ruan X E, Bien Z, Wang Q. Convergence characteristics of proportional-type iterative learning control in the sense of Lebesgue-p norm. IET Control Theory and Applications, 2012, 6: 707–714. DOI:10.1049/iet-cta.2010.0388
[5] Xu J X, Tan Y. Robust optimal design and convergence properties analysis of iterative learning control approaches. Automatica, 2002, 38: 1867–1880. DOI:10.1016/S0005-1098(02)00143-7
[6] Xu J X, Tan Y. On the P-type and Newton-type ILC schemes for dynamic systems with non-affine-in-input factors. Automatica, 2002, 38: 1237–1242. DOI:10.1016/S0005-1098(02)00021-3
[7] Bristow D A, Tharayil M, Alleyne A G. A survey of iterative learning control:A learning-method for high-performance tracking control. IEEE Control Syst Mag, 2006, 26: 96–114. DOI:10.1109/MCS.2006.1636313
[8] 戴喜生, 李政, 田森平. 基于向量图分析的分布参数系统迭代学习控制. 控制理论与应用, 2009, 26: 619–623.
Dai X S, Li Z, Tian S P. Iterative learning-control for distributed parameter systems based on vector-plot analysis. Control Theory & Applications, 2009, 26: 619–623.
[9] 谢胜利, 谢振东, 韦岗. 非线性分布参数系统跟踪控制的学习算法. 自动化学报, 1999, 25: 627–632.
Xie S L, Xie Z D, Wei G. Learning algorithm for tracking control of nonlinear distributed parameter systems. Acta Automatica Sinica, 1999, 25: 627–632.
[10] 田森平, 吴忻生. 基于几何分析的分布参数系统的迭代学习控制. 控制理论与应用, 2012, 29: 1082–1085.
Tian S P, Wu X S. Iterative learning control of distributed parameter systems based on geometric analysis. Control Theory & Applications, 2012, 29: 1082–1085.
[11] 谢振东, 刘永清. 分布参数系统目标跟踪的二阶P-型学习算法. 暨南大学学报(自然科学与医学版), 1998, 19: 60–64.
Xie Z D, Liu Y Q. Two order P-type learning control algorithm of distributed parameter systems. Journal of Jinan University (Natural Science & Medicine Edition), 1998, 19: 60–64.
[12] 傅勤. 二阶非线性双曲型分布参数系统的迭代学习控制. 系统科学与数学, 2014, 34: 284–293.
Fu Q. Iterative learning control for second order nonlinear hyperbolic distributed parameter systems. Journal of Systems Science and Mathematical Sciences, 2014, 34: 284–293.
[13] 傅勤. 分布参数系统于W 1, 2空间中的迭代学习控制. 数学物理学报, 2016, 36: 267–286.
Fu Q. Iterative learning control for distributed parameter systems in space W 1, 2. Acta Mathematica Scientia, 2016, 36: 267–286. DOI:10.3969/j.issn.1003-3998.2016.02.006
[14] 傅勤. 非正则抛物型分布参数系统的迭代学习控制. 控制与决策, 2016, 31: 114–122.
Fu Q. Iterative learning control for irregular distributed parameter systems. Control and Decision, 2016, 31: 114–122.
[15] Cichy B, Galkowski K, Rogers E, Kummert A. An approach to iterative learning control for spatio-temporal dynamics using nD discrete linear systems models. Multidimensional Systems and Signal Processing, 2011, 22: 83–96. DOI:10.1007/s11045-010-0108-1
[16] Cichy B, Galkowski K, Rogers E. Iterative learning control for spatio-temporal dynamics using CrankNicholson discretization. Multidimensional Systems and Signal Processing, 2012, 23: 185–208. DOI:10.1007/s11045-010-0132-1
[17] Dai X S, Tian S P, Guo Y J. Iterative learning lontrol for discrete parabolic distributed parameter systems. International Journal of Automation and Computing, 2015, 12: 316–322. DOI:10.1007/s11633-015-0892-z
[18] 戴喜生, 郭亚君, 田森平, 李克讷. 非线性抛物型偏差分系统迭代学习控制. 控制理论与应用, 2015, 32: 1607–1612.
Dai X S, Guo Y J, Tian S P, Li K N. Iterative learning control of nonlinear parabolic partial difference systems. Control Theory & Applications, 2015, 32: 1607–1612.
[19] 孙志忠. 偏微分方程数值解法. 北京: 科学出版社, 2005.
Sun Z Z. Numerical Methods for Partial Differential Equations. Beijing: Science Press, 2005.
[20] 陈小燕. 波动方程的高阶差分格式及其外推算法[D]. 南京: 东南大学, 2011
Chen X Y. High order difference schemes and extrapolation algorithms for wave equations[D]. Nanjing: Southeast University, 2011 http://www.wanfangdata.com.cn/details/detail.do?_type=degree&id=Y1976815