数学物理学报  2018, Vol. 38 Issue (3): 599-612   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
杨轩
阮小娥
王彭
时变切换信号驱动的线性连续切换系统的迭代学习控制收敛性分析
杨轩1, 阮小娥2, 王彭3     
1. 西安工程大学理学院 西安 710048;
2. 西安交通大学数学与统计学院 西安 710049;
3. 中国人民解放军63771部队 陕西渭南 714000
摘要:针对一类由任意时变切换信号驱动并在某个时间区间可重复运行的切换系统,该文研究一阶和高阶PD-型迭代学习控算法.利用卷积积分的广义Young不等式,在Lebesgue-p范数意义下分析跟踪误差性态,得出算法收敛的充分条件,并量化了状态矩阵对学习效果的影响.数值仿真验证了理论结果的可行性和有效性.
关键词迭代学习控制    切换系统    切换律    Lebesgue-p范数    广义Young不等式    收敛性    
Convergence Analysis of Iterative Learning Control for Linear Continuous-Time Switched Systems with Arbitrary Time-Driven Switching Rules
Yang Xuan1, Ruan Xiaoe2, Wang Peng3     
1. School of Science, Xi'an Polytechnic University, Xi'an 710048;
2. School of Mathematics and Statistics, Xian Jiaotong University, Xi'an 710049;
3. Troops 63771 of PLA, Shaanxi Weinan 714000
Abstract: This paper addresses the convergence performance of first-order and higher-order PD-type iterative learning control strategies for a class of linear continuous-time switched systems. The manipulated systems are elaborated by arbitrary time-driven switching signals and can repetitively operate over a finite time interval. By employing the generalized Young inequality of convolution integral theoretical analysis is launched in the sense of Lebesgue-p norm. Simultaneously, sufficient convergence conditions of the algorithms are derived and the effect of the state matrices on the learning performance is quantized. To illustrate the validity and effectiveness of the theoretical results, numerical simulations are conducted.
Key words: Iterative learning control     Switched systems     Switching rules     Lebesgue-p norm     Generalized Young inequality     Convergence    
1 引言

在工业过程中, 控制器往往根据现实条件在不同模型间进行切换, 以使整个系统运行在最佳状态[1-2], 这涉及到切换系统.切换系统是一个由系列连续或离散子系统以及协调这些子系统之间进行切换的规则组成的混合系统[3-4].由于具有广泛的应用背景, 近年来关于切换系统的研究越来越受到人们的关注.

在切换系统的应用研究中, 轨迹跟踪是一个广受关注的问题.例如, 文献[5]应用Lyapunov函数分析了一类切换系统的输出跟踪控制问题; 当系统由异步切换规则支配时, 文献[6]基于平均驻留时间分析了一类鲁棒跟踪控制问题; 系统存在时滞时, 文献[7]基于观测信号, 设计了一种切换控制律.另外, 针对离散切换系统, 文献[8]基于Lyapunov-Krasovskii函数, 研究了一种指数Lebesgue-2-Lebesgue-$\infty$输出跟踪控制问题.其它有关切换系统的跟踪控制问题参见文献[9-11].这些控制器大都与系统模型信息相关, 其控制性能对模型未知系统无法保证, 因此有必要寻求不依赖系统信息的控制策略.

迭代学习控制(ILC)是满足上述条件的控制策略之一. ILC利用之前的学习信息更新当前的控制信号, 构造下一个学习周期的控制信号, 以提升系统的跟踪性能[12-15]; 如此反复最终实现对目标轨线的精确跟踪. ILC本质上是一种数据驱动的控制策略, 其突出的优点是, 不需要知道模型的具体信息即可达到较好的跟踪效果.跟踪误差是更新控制器的关键信息, 这意味着, 对切换系统而言, 在不同的迭代过程中某个时间区间上运行的子系统应该是相同的, 这要求切换信号是时变-迭代不变函数.目前, 在这种假设条件下, 对切换系统的ILC研究已取得了一定的成果.

针对线性离散切换系统, 文献[16-19]分别提出了比例-微分型(PD -型)、比例型(P -型)型和微分型(D -型)迭代学习控制策略.针对有时滞的非线性离散系统, 文献[20-21]分别分析了D -型和PID -型ILC的学习性态.另外, 为了减弱噪声影响, 文献[22]提出了带衰减因子的ILC, 文献[23-24]提出了带状态信息的混合ILC.纵观上述文献, 其分析学习性能的理论依据主要有Shifted理论、2D理论以及$\lambda$ -范数理论等.然而, Shifted理论需要将原先时间-迭代二维动态系统转化为一维代数系统, 不适用于连续切换系统, 因为连续系统的输出响应具有卷积积分. 2D理论需要构造Lyapunov函数以度量控制算法的稳定性; 但是, 其相关参数的选取较棘手. $\lambda$ -范数理论在分析ILC学习性能时, 结果虽然比较完美, 但也存在缺陷.例如, $\lambda$往往既与被控系统信息无关也与学习增益无关, 其取值仅为得到完美的理论结果, 有时为了保证算法收敛, 要取充分大的$\lambda$.这样可能造成的结果是, 系统动态学和学习增益对学习性能的影响会被压制[14], 以致掩盖了ILC的自然属性.为了避免上述方法的不足, 有必要寻求其它的理论或技术研究ILC过程. Lebesgue-$p$范数通过某时间区间内的积分度量某个信号, 其中信号的振幅、带宽等性态得以综合体现, 具有现实的物理意义.因此, 在工程实际中, 研究ILC在Lebesgue-$p$范数意义下的性态是一个合理的选择.

基于上述分析, 本文针对一类由任意时变切换信号驱动的线性连续切换系统, 研究一阶和高阶PD -型ILC策略, 借助卷积积分的广义Young不等式在Lebesgue-$p$范数意义下分析误差性态, 并得出控制算法收敛的一个充分条件, 同时也分析了学习增益和子系统动力学对学习效果的影响.

本文的其它内容组织如下:第2部分描述基本问题及相关定义, 第3部分分析ILC的收敛性.为了验证理论结果的可行性和有效性, 第4部分进行了数值仿真.最后, 第5部分总结全文.

2 问题描述

考虑如下线性连续时不变SISO切换系统

$ \begin{equation}\label{eq:de1} \left\{ \begin{array}{ll} \dot{{\mathit{\boldsymbol{x}}}}_{k}(t)&={\mathit{\boldsymbol{A}}}_{s(t)}{\mathit{\boldsymbol{x}}}_{k}(t)+{\mathit{\boldsymbol{B}}}_{s(t)}u_{k}(t), \\ y_{k}(t)&={\mathit{\boldsymbol{C}}}_{s(t)}{\mathit{\boldsymbol{x}}}_{k}(t), t\in \Omega=[0, T], \end{array} \right. \end{equation} $ (2.1)

其中, $\Omega=[0, T]$是时间区间, $t\in\Omega$是时间变量, $k$表示迭代次数. ${\mathit{\boldsymbol{x}}}_{k}(t)\in \mathbb{R}^{n}$, $u_{k}(t)\in \mathbb{R}$, $y_{k}(t)\in \mathbb{R}$分别表示状态变量、输入和输出变量.下标$s(t)$表示切换信号, 其定义如下

$ \begin{equation} s(\cdot):\Omega\rightarrow Q=\{1, 2, \cdots, q\}(q\in Z_{+}, q<\infty). \end{equation} $ (2.2)

${\mathit{\boldsymbol{A}}}_{s(t)}$${\mathit{\boldsymbol{B}}}_{s(t)}$${\mathit{\boldsymbol{C}}}_{s(t)}$分别为具有适当维数的常矩阵.

需要说明的是, 系统${\rm (2.1)}$的动态学可能未知; 但假设其可在$\Omega$上重复运行.

给定一个目标轨线$y_{{\rm d}}(t)\in \mathbb{R}(t\in \Omega)$和任意的时变切换规则$s(t)(t\in \Omega)$, ILC的目标是在学习的过程中构造一个递归的控制信号序列$\{u_{k}(t)\}(t\in \Omega)$, 使得在其控制下, 系统的输出轨线随着迭代次数的增加渐近跟踪$y_{{\rm d}}(t)\in \mathbb{R}(t\in \Omega)$, 即

$ \begin{equation} \lim\limits_{k\rightarrow \infty}||e_{k}(\cdot)||_{p}=0~~~(p\geq 1), \end{equation} $ (2.3)

其中, $e_{k}(t)=y_{{\rm d}}(t)-y_{k}(t)(t\in \Omega)$是跟踪误差, $||\cdot||_{p}$表示Lebesgue-$p$范数, 其定义如下:

定义2.1[25]  设连续数量函数$f(t):I\subseteq \mathbb{R}^{+}\rightarrow \mathbb{R}$, 其Lebesgue-$p$范数定义为

$ \begin{equation} ||f(\cdot)||_{p}=\Big(\int_{I}|f(t)|^{p}{\rm d}t\Big)^{1/p}~~~(1\leq p\leq +\infty). \end{equation} $ (2.4)

另外, 基于定义2.1, 关于卷积积分和广义Young不等式有如下定义和引理.

定义2.2[14, 25]  给定两个${\rm Lebesgue}$可积数量函数$f(\cdot):I\rightarrow \mathbb{R}$$g(\cdot):I\rightarrow \mathbb{R}$, 其卷积积分定义为

$ \begin{equation} (f*g)=\int_{I}f(t-\tau)g(\tau){\rm d}\tau. \end{equation} $ (2.5)

引理2.1[14]  给定两个${\rm Lebesgue}$可积数量函数$f(\cdot):I\rightarrow \mathbb{R}$$g(\cdot):I\rightarrow \mathbb{R}$, 其卷积积分的广义${\rm Young}$不等式

$ \begin{equation} ||f*g||_{l}\leq ||f||_{q}||g||_{p} \end{equation} $ (2.6)

对满足$\frac{1}{p}+\frac{1}{q}=\frac{1}{l}+1$的所有$1\leq p, q, l\leq +\infty$均成立.特别地, 当$l=p$时, 不等式(2.6)变为

$ \begin{equation} ||f*g||_{p}\leq ||f||_{1}||g||_{p}. \end{equation} $ (2.7)

本文首先利用当前学习周期内的跟踪误差及其导数, 提出了一阶PD -型学习律

$ \begin{equation} u_{k+1}(t)=u_{k}(t)+\Gamma_{{\rm p}}e_{k}(t)+\Gamma_{{\rm d}}\dot{e}_{k}(t), \end{equation} $ (2.8)

其中, $\Gamma_{{\rm p}}$$\Gamma_{{\rm d}}$分别表示比例型和微分型学习增益.其次, 利用前$r(r\in Z_{+}, r<k)$个学习周期内的控制信号和跟踪误差信号, 提出了一种$r$阶PD -型学习律

$ \begin{eqnarray} u_{k+1}(t)&=& \theta_{1}[u_{k}(t)+\Gamma_{{\rm p}, 1}e_{k}(t)+\Gamma_{{\rm d}, 1 }\dot{e}_{k}(t)]\\ & &+ \theta_{2}[u_{k}(t)+\Gamma_{{\rm p}, 2}e_{k-1}(t) +\Gamma_{{\rm d}, 2}\dot{e}_{k-1}(t)]\\ &&\quad \quad \quad \quad \quad \quad \vdots\\ &&+\theta_{r}[u_{k}(t)+\Gamma_{{\rm p}, r}e_{k-r+1}(t)+\Gamma_{{\rm d}, r}\dot{e}_{k-r+1}(t)], \end{eqnarray} $ (2.9)

这里, $r$表示学习律的阶数, $\theta_{i}$ $(i=1, 2, 3, \cdots, r)$表示权重系数, 满足$0\leq \theta_{i}\leq 1$ $(i=1, 2, 3, \cdots, r), $ $\sum\limits_{i=1}^{r}\theta_{i}=1$; $\Gamma_{{\rm p}, i}$$\Gamma_{{\rm d}, i}$分别表示比例型和微分型学习增益.

系统(2.1)的每个子系统, 满足如下假设:

假设2.1  任意给定目标轨线$y_{{\rm d}}(t)(t\in \Omega)$, 存在一个目标控制信号$u_{{\rm d}}(t)(t\in \Omega)$和一个适当的目标状态信号${\mathit{\boldsymbol{x}}}_{{\rm d}}(t)(t\in \Omega)$, 使得

$ \begin{eqnarray*} \left\{ \begin{array}{ll} \dot{{\mathit{\boldsymbol{x}}}}_{{\rm d}}(t)&={\mathit{\boldsymbol{A}}}_{s(t)}{\mathit{\boldsymbol{x}}}_{{\rm d}}(t)+{\mathit{\boldsymbol{B}}}_{s(t)}u_{{\rm d}}(t), \\ y_{{\rm d}}(t)&={\mathit{\boldsymbol{C}}}_{s(t)}{\mathit{\boldsymbol{x}}}_{{\rm d}}(t), t\in \Omega=[0, T], \end{array} \right. \end{eqnarray*} $

假设2.2  每次迭代的初始状态相同, 本文假设初始状态为${\mathit{\boldsymbol{x}}}_{k}(0)={\mathit{\boldsymbol{x}}}_{{\rm d}}(0), \forall k\in N_{+}$.

假设2.3  切换信号$s(t)$为定义在$\Omega$上迭代不变的任意分段常函数.

假设2.4  乘积${\mathit{\boldsymbol{C}}}_{s(t)}{\mathit{\boldsymbol{B}}}_{s(t)}(s(t)\in Q)$非零且性质符号已知.

注2.1  假设${\rm 2.1}$表明了${\rm ILC}$策略对切换系统进行控制的可解性; 假设${\rm 2.2}$指明了初始状态选取方法; 假设${\rm 2.3}$意味着在不同学习过程的某个子区间上运行的子系统相同, 这是满足迭代学习控制条件的基本假设. ${\mathit{\boldsymbol{C}}}_{s(t)}{\mathit{\boldsymbol{B}}}_{s(t)}(s(t)\in Q)$的结果决定着控制的方向, 虽然系统${\rm (2.1)}$的动态学信息可能未知, 但假设${\rm 2.4}$表明, 控制的方向是已知的.

3 主要结果

根据(2.2)式和假设2.3, 切换信号可以记作

$ \begin{equation} s(t)=i= \left\{ \begin{array}{ll} &1, t\in \Omega_{1}=[0, t_{1}), \\ &2, t\in \Omega_{2}=[t_{1}, t_{2}), \\ &\vdots\quad \quad \quad\quad\vdots\\ &q, t\in \Omega_{q}=[t_{q-1}, t_{T}], \end{array} \right. \end{equation} $ (3.1)

在(3.1)式的支配下, 被控系统(2.1)可变为

$ \begin{equation} \left\{ \begin{array}{ll} \dot{{\mathit{\boldsymbol{x}}}}_{k}(t)&={\mathit{\boldsymbol{A}}}_{i}{\mathit{\boldsymbol{x}}}_{k}(t)+{\mathit{\boldsymbol{B}}}_{i}u_{k}(t), \\ y_{k}(t)&={\mathit{\boldsymbol{C}}}_{i}{\mathit{\boldsymbol{x}}}_{k}(t), t\in \Omega=[0, T], i\in Q. \end{array} \right. \end{equation} $ (3.2)

任意给定一个初始控制信号序列, 系统(3.2)在第$i$个子区间$\Omega_{i}$上的状态响应如下

$ \begin{equation} {\mathit{\boldsymbol{x}}}_{k}(t)={\rm exp}({\mathit{\boldsymbol{A}}}_{i}(t-t_{i-1})) {\mathit{\boldsymbol{x}}}_{k}(t_{i-1})+\int_{t_{i-1}}^{t}{\rm exp}({\mathit{\boldsymbol{A}}}_{i}(t-\tau)){\mathit{\boldsymbol{B}}}_{i}u_{k}(\tau){\rm d}\tau, i\in Q, \end{equation} $ (3.3)

其中$t_{0}=0, t_{q}=T$, ${\mathit{\boldsymbol{x}}}_{k}(0)={\mathit{\boldsymbol{x}}}_{{\rm d}}(0)$.

等式(3.3)表明, 除控制信号外, 系统的状态响应还与初始状态, 子系统的动态学信息有关.由于任意一个子区间上的终止状态是其下一个子系统的初始状态, 每次迭代开始的初始状态只在第一个子区间内设定, 后续时间子区间上的初始状态没必要重新设置.如果初始状态已提前设定, 且各个子系统的动态学信息固定不变, 切换信号对系统的动态行为起决定性作用, 它指定在子区间$\Omega_{i}$上运行的子系统.以下的分析和讨论是在切换信号任意给定的条件下展开的.

引理3.1[26]  设$\{a_{k}\}$$\{\varepsilon_{k}\}$是两个非负序列, 满足

$ \begin{equation} a_{k}\leq \sigma_{1}a_{k-1}+\sigma_{2}a_{k-2}+\cdots +\sigma_{n}a_{k-n}+\varepsilon_{k}, k=n+1, n+2, \cdots, \end{equation} $ (3.4)

其中, $a_{l}(l=1, 2, 3, \cdots, n)$为初始条件.如果系数满足$\sigma_{j}\geq 0, \sum\limits_{j=1}^{n}<1$, $\lim\limits_{k\rightarrow +\infty}\varepsilon_{k}\leq \varepsilon$, 则

$ \limsup\limits_{k\rightarrow +\infty}a_{k}\leq \frac{\varepsilon}{1-\sigma}. $

推论3.1  设非负序列$\{a_{k}\}$满足

$ a_{k}\leq \sigma_{1}a_{k-1}+\sigma_{2}a_{k-2}+\cdots +\sigma_{n}a_{k-n}, k=n+1, n+2, \cdots, $

其中, $a_{j}(j=1, 2, \cdots, n)$是初始值.如果$\sigma_{j}\geq 0, \sum\limits_{j=1}^{n}<1$, 则

$ \begin{eqnarray*} \lim\limits_{k\rightarrow +\infty}a_{k}=0. \end{eqnarray*} $

引理3.2  考虑序列$\{b_{j}\}, b_{j}\in [0, 1), j=1, 2, \cdots, n$.如果存在序列$\{\sigma_{j}\}$, 满足$0\leq \sigma_{j}<1, $ $\sum\limits_{j=1}^{n}\sigma_{j}=1$, 则不等式

$ \sum\limits_{j=1}^{n}\sigma_{j}b_{j}<1 $

成立.

  记$\widetilde{b}={\rm max}\{b_{1}, b_{2}, \cdots, b_{n}\}$, 不等式

$ \begin{eqnarray*} \sum\limits_{j=1}^{n}\sigma_{j}b_{j}\leq \Big (\sum\limits_{j=1}^{n}\sigma_{j}\Big)\widetilde{b}<\widetilde{b}<1 \end{eqnarray*} $

显然成立.证毕.

定理3.1  将一阶${\rm PD}$ -型${\rm ILC}$学习律$({\rm 2.8})$应用到切换系统$({\rm 3}.{\rm 2})$, 如果各子系统动态学${\mathit{\boldsymbol{A}}}_{i}, {\mathit{\boldsymbol{B}}}_{i}, {\mathit{\boldsymbol{C}}}_{i}$以及学习增益$\Gamma_{{\rm p}}$$\Gamma_{{\rm d}}$满足

$ \begin{equation} \rho_{i}=|1-{\mathit{\boldsymbol{C}}}_{i}{\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm d}}|+ ||{\mathit{\boldsymbol{C}}}_{i}{\rm exp}({\mathit{\boldsymbol{A}}}_{i} (\cdot))({\mathit{\boldsymbol{A}}}_{i}{\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm d}}+{\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm p}})||_{1}<1, \end{equation} $ (3.5)

则随着迭代次数的增加, 系统输出能够渐近跟踪到目标轨线$y_{{\rm d}}(t)(t\in \Omega)$, 即

$ \begin{eqnarray*} \lim\limits_{k\rightarrow +\infty}||e_{k+1}(\cdot)||_{p}=0. \end{eqnarray*} $

  根据跟踪误差定义和状态响应(${\rm 3.3}$), 得

$ \begin{eqnarray} e_{k+1}(t)&=&y_{{\rm d}}(t)-y_{k+1}(t)=e_{k}(t)-[y_{k+1}(t)-y_{k}(t)]\\ &=&e_{k}(t)-{\mathit{\boldsymbol{C}}}_{i}{\rm exp}({\mathit{\boldsymbol{A}}}_{i}(t-t_{i-1}))[{\mathit{\boldsymbol{x}}}_{k+1}(t_{i-1})-{\mathit{\boldsymbol{x}}}_{k}(t_{i-1})]\\ &&-{\mathit{\boldsymbol{C}}}_{i}\int_{t_{i-1}}^{t}{\rm exp}({\mathit{\boldsymbol{A}}}_{i}(t-\tau)){\mathit{\boldsymbol{B}}}_{i} [u_{k+1}(\tau)-u_{k}(\tau)]{\rm d}\tau. \end{eqnarray} $ (3.6)

将学习律(${\rm 2.8}$)代入(${\rm 3.6}$)式, 得

$ \begin{eqnarray} e_{k+1}(t)&=&e_{k}(t)-{\mathit{\boldsymbol{C}}}_{i}{\rm exp}({\mathit{\boldsymbol{A}}}_{i}(t-t_{i-1}))[{\mathit{\boldsymbol{x}}}_{k+1}(t_{i-1})-{\mathit{\boldsymbol{x}}}_{k}(t_{i-1})]\\ &&-{\mathit{\boldsymbol{C}}}_{i}\int_{t_{i-1}}^{t}{\rm exp}({\mathit{\boldsymbol{A}}}_{i}(t-\tau)){\mathit{\boldsymbol{B}}}_{i} [\Gamma_{{\rm p}}e_{k}(\tau)+\Gamma_{{\rm d}}\dot{e}_{k}(\tau)]{\rm d}\tau. \end{eqnarray} $ (3.7)

由分部积分法, 得

$ \begin{eqnarray} &&{\mathit{\boldsymbol{C}}}_{i}\int_{t_{i-1}}^{t}{\rm exp}({\mathit{\boldsymbol{A}}}_{i} (t-\tau)){\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm d}}\dot{e}_{k}(\tau){\rm d}\tau\\ &=& {\mathit{\boldsymbol{C}}}_{i}{\rm exp}({\mathit{\boldsymbol{A}}}_{i}(t-\tau){\mathit{\boldsymbol{B}}}_{i} \Gamma_{{\rm d}}e_{k}(\tau)|_{\tau=t_{i-1}}^{\tau=t} +{\mathit{\boldsymbol{C}}}_{i}\int_{t_{i-1}}^{t}{\rm exp}({\mathit{\boldsymbol{A}}}_{i}(t-\tau)){\mathit{\boldsymbol{A}}}_{i}{\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm d}}e_{k}(\tau){\rm d}\tau, \end{eqnarray} $ (3.8)

将(${\rm 3.8}$)式代入(${\rm 3.7}$)式, 得

$ \begin{eqnarray} e_{k+1}(t)&=&e_{k}(t)-{\mathit{\boldsymbol{C}}}_{i}\exp({\mathit{\boldsymbol{A}}}_{i}(t-t_{i-1}))[{\mathit{\boldsymbol{x}}}_{k+1}(t_{i-1})-{\mathit{\boldsymbol{x}}}_{k}(t_{i-1})] \\ &&- {\mathit{\boldsymbol{C}}}_{i}\exp({\mathit{\boldsymbol{A}}}_{i}(t-\tau) {\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm d}}e_{k}(\tau)|_{\tau=t_{i-1}}^{\tau=t} \\ &&-{\mathit{\boldsymbol{C}}}_{i}\int_{t_{i-1}}^{t}{\rm exp}({\mathit{\boldsymbol{A}}}_{i}(t-\tau))[{\mathit{\boldsymbol{A}}}_{i}{\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm d}}+{\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm p}}]e_{k}(\tau){\rm d}\tau. \end{eqnarray} $ (3.9)

(Ⅰ)当$t\in\Omega_{1}$时, 第一个子系统运行, 记$t_{0}=0$, 跟踪误差递归关系(${\rm 3.9}$)可变为

$ \begin{eqnarray} e_{k+1}(t)&=&e_{k}(t)-{\mathit{\boldsymbol{C}}}_{1}\exp({\mathit{\boldsymbol{A}}}_{1}t)[{\mathit{\boldsymbol{x}}}_{k+1}(0)-{\mathit{\boldsymbol{x}}}_{k}(0)] - {\mathit{\boldsymbol{C}}}_{1}\exp({\mathit{\boldsymbol{A}}}_{1}(t-\tau) {\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm d}}e_{k}(\tau)|_{\tau=0}^{\tau=t} \\ &&-{\mathit{\boldsymbol{C}}}_{1}\int_{0}^{t}\exp({\mathit{\boldsymbol{A}}}_{1} (t-\tau))[{\mathit{\boldsymbol{A}}}_{1}{\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm d}}+{\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm p}}]e_{k}(\tau){\rm d}\tau. \end{eqnarray} $ (3.10)

进而, 考虑假设2.2, 得

$ \begin{equation} e_{k+1}(t)=(1-{\mathit{\boldsymbol{C}}}_{1}{\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm d}})e_{k}(t)- \int_{0}^{t}{\mathit{\boldsymbol{C}}}_{1}\exp({\mathit{\boldsymbol{A}}}_{1}(t-\tau))[{\mathit{\boldsymbol{A}}}_{1}{\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm d}}+{\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm p}}]e_{k}(\tau){\rm d}\tau. \end{equation} $ (3.11)

对(3.11)式两边取Lebesgue-$p$范数并利用卷积积分广义Young不等式, 得

$ \begin{eqnarray} ||e_{k+1}(\cdot)||_{p}&\leq &(|1-{\mathit{\boldsymbol{C}}}_{1}{\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm d}}|+ ||{\mathit{\boldsymbol{C}}}_{1}\exp({\mathit{\boldsymbol{A}}}_{1}(\cdot))[{\mathit{\boldsymbol{A}}}_{1}{\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm d}}+{\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm p}}]||_{1})||e_{k}(\cdot)||_{p} \\ &=&\rho_{1}||e_{k}(\cdot)||_{p}\leq \rho_{1}^{2}||e_{k-1}(\cdot)||_{p}\leq \cdots \leq \rho_{1}^{k}||e_{1}(\cdot)||_{p}. \end{eqnarray} $ (3.12)

$\rho_{1}<1$知, $\lim\limits_{k\rightarrow +\infty}||e_{k+1}(\cdot)||_{p}=0$$[0, t_{1}]$成立.

(Ⅱ)当$t\in\Omega_{2}$, 第二个子系统运行, 跟踪误差(3.9)变为

$ \begin{eqnarray} e_{k+1}(t)&=&e_{k}(t)-{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{i}(t-t_{1}))[{\mathit{\boldsymbol{x}}}_{k+1}(t_{1})-{\mathit{\boldsymbol{x}}}_{k}(t_{1})]- {\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(t-\tau)){\mathit{\boldsymbol{B}}}_{2} \Gamma_{{\rm d}}e_{k}(\tau)|_{\tau=t_{1}}^{\tau=t} \\ &&-{\mathit{\boldsymbol{C}}}_{2}\int_{t_{1}}^{t}{\rm exp}({\mathit{\boldsymbol{A}}}_{2} (t-\tau))[{\mathit{\boldsymbol{A}}}_{2}{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}}+{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm p}}]e_{k}(\tau){\rm d}\tau \\ & =&(1-{\mathit{\boldsymbol{C}}}_{2}{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}})e_{k}(t)-\int_{t_{1}}^{t}{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2} (t-\tau))[{\mathit{\boldsymbol{A}}}_{2}{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}}+{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm p}}]e_{k}(\tau){\rm d}\tau \\ &&- {\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(t-t_{1}))\triangle {\mathit{\boldsymbol{x}}}_{k}(t_{1})+{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2} (t-t_{1})){\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}}e_{k}(t_{1}), \end{eqnarray} $ (3.13)

其中, $\triangle{\mathit{\boldsymbol{x}}}_{k}(t_{1})={\mathit{\boldsymbol{x}}}_{k+1}(t_{1})-{\mathit{\boldsymbol{x}}}_{k}(t_{1})$.

对(3.13)式两边求Lebesgue-$p$范数, 并应用广义Young不等式, 得

$ \begin{eqnarray} ||e_{k+1}(\cdot)||_{p}&\leq&(|1-{\mathit{\boldsymbol{C}}}_{2}{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}}|+ ||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2} (\cdot))[{\mathit{\boldsymbol{A}}}_{2}{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}}+{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm p}}]||_{1})||e_{k}(\cdot)||_{p} \\ &&+ ||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(\cdot))||_{p}||\triangle{\mathit{\boldsymbol{x}}}_{k}(t_{1})||_{p}+||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(\cdot)) {\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}}||_{p}||e_{k}(t_{1})||_{p} \\ & =&\rho_{2}||e_{k}(\cdot)||_{p}+||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2} (\cdot))||_{p}||\triangle{\mathit{\boldsymbol{x}}}_{k}(t_{1})||_{p} \\ &&+||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(\cdot)) {\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}}||_{p}||e_{k}(t_{1})||_{p}. \end{eqnarray} $ (3.14)

由证明过程(Ⅰ)知, $\lim\limits_{k\rightarrow +\infty}||e_{k+1}(\cdot)||_{p}=0$$\Omega_{1}$上成立.这说明$\lim\limits_{k\rightarrow +\infty}||e_{k}(t_{1})||_{p}=0$$\lim\limits_{k\rightarrow +\infty}||{\mathit{\boldsymbol{x}}}_{{\rm d}}(t_{1})-{\mathit{\boldsymbol{x}}}_{k+1}(t_{1})||_{p}=0$$t_{1}$时刻都成立.因此

$ \begin{eqnarray*} \lim\limits_{k\rightarrow +\infty}||\triangle {\mathit{\boldsymbol{x}}}_{k}(t_{1})\|_p&=&\lim\limits_{k\rightarrow +\infty}||{\mathit{\boldsymbol{x}}}_{k+1}(t_{1})-{\mathit{\boldsymbol{x}}}_{k}(t_{1})||_{p}\\ &=&\lim\limits_{k\rightarrow +\infty}||{\mathit{\boldsymbol{x}}}_{{\rm d}}(t_{1}) -{\mathit{\boldsymbol{x}}}_{k}(t_{1})-({\mathit{\boldsymbol{x}}}_{{\rm d}}(t_{1}) -{\mathit{\boldsymbol{x}}}_{k+1}(t_{1}))||_{p}\\ &\leq& \lim\limits_{k\rightarrow +\infty}||{\mathit{\boldsymbol{x}}}_{{\rm d}}(t_{1})-{\mathit{\boldsymbol{x}}}_{k}(t_{1})||_{p}+ \lim\limits_{k\rightarrow +\infty} ||{\mathit{\boldsymbol{x}}}_{{\rm d}}(t_{1})-{\mathit{\boldsymbol{x}}}_{k+1}(t_{1}))||_{p}\\ &=&0. \end{eqnarray*} $

$||{\mathit{\boldsymbol{C}}}_{2}\exp ({\mathit{\boldsymbol{A}}}_{2}(\cdot))||_{p}$有界, 故$\lim\limits_{k\rightarrow +\infty}||{\mathit{\boldsymbol{C}}}_{2}\exp ({\mathit{\boldsymbol{A}}}_{2}(\cdot))||_{p}||\triangle {\mathit{\boldsymbol{x}}}_{k}(t_{1})||_{p}=0$.由引理3.1得: $\lim\limits_{k\rightarrow +\infty}||e_{k+1}(\cdot)||_{p}=0$在区间$\Omega_{2}$上成立.

重复上述过程, 易证$\lim\limits_{k\rightarrow +\infty}||e_{k+1}(\cdot)||_{p}=0$在区间$\Omega_{l}(l=3, 4, \cdots, q)$上成立, 从而结论在整个区间$\Omega=[0, T]$上均成立.证毕.

注3.1  定理3.1中$0<\rho_{i}<1$是保证子系统的输出能够在子区间$\Omega_{i}$上渐近跟踪目标轨线的充分条件.若切换信号任意给定, 子系统动态学和学习增益直接影响控制算法的收敛性态.与文献[17, 20]中$\lambda$ -范数意义下收敛条件比较, ${\rm (3.5)}$式量化了状态矩阵对学习性态的影响, 它揭示了学习系统的自然属性, 便于我们在工程实际中设计有效的控制器.

注3.2从定理${\rm 3.1}$的证明过程可以看出, 控制算法${\rm (2.8)}$除了在$\Omega_{1}$内都非单调收敛的, 其原因主要是系统在$t_{1}, t_{2}, \cdots, t_{q}$时刻的切换增加了震荡的风险.特别地, 如果系统不切换, $({\rm 3.5})$式就是单调收敛的充分条件, 这与文献[14]中定理${\rm Ⅲ.1}$的结论一致.

注3.3  在工程实际中, ${\mathit{\boldsymbol{A}}}_{i}$的特征值实部往往小于零, 这可保证$||{\mathit{\boldsymbol{C}}}_{i}{\rm exp}({\mathit{\boldsymbol{A}}}_{i} (\cdot))({\mathit{\boldsymbol{A}}}_{i}{\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm d}}+{\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm p}})||_{1}$是充分小的正数, 从而通过选取适当的学习增益使${\rm (3.5)}$式成立.因此, 本文的算法具有实用价值.

定理3.2  假设$r$${\rm PD}$ -型${\rm ILC}$学习律(2.9)应用到由切换信号$({\rm 3.1})$支配下的系统(3.2).如果各子系统动态学${\mathit{\boldsymbol{A}}}_{i}, {\mathit{\boldsymbol{B}}}_{i}, {\mathit{\boldsymbol{C}}}_{i}$以及学习增益$\Gamma_{{\rm p}, j}$$\Gamma_{{\rm d}, j}$满足

$ \begin{equation} \rho_{i, j}=|1-{\mathit{\boldsymbol{C}}}_{i}{\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm d}, j}|+ ||{\mathit{\boldsymbol{C}}}_{i}{\rm exp}({\mathit{\boldsymbol{A}}}_{i}(\cdot))({\mathit{\boldsymbol{A}}}_{i}{\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm d}, j}+{\mathit{\boldsymbol{B}}}_{i}\Gamma_{{\rm p}, j})||_{1}<1, \end{equation} $ (3.15)

其中, $j=1, 2\cdots, r$, 则随着迭代次数的增加, 系统输出能渐近跟踪到$y_{{\rm d}}(t)(t\in \Omega)$, 即

$ \begin{eqnarray*} \lim\limits_{k\rightarrow +\infty}||e_{k+1}(\cdot)||_{p}=0. \end{eqnarray*} $

  (Ⅰ) 当$t\in\Omega_{1}$, 第一个子系统运行, 将学习律$(2.9)$应用到系统(3.2), 跟踪误差为

$ \begin{eqnarray} e_{k+1}(t)&=&y_{{\rm d}}(t)-y_{k+1}(t) \\ & =&\theta_{1}[y_{{\rm d}}(t)-y_{k}(t)]+\theta_{2}[y_{{\rm d}}(t)-y_{k-1}(t)]+\cdots+\theta_{r}[y_{{\rm d}}(t)-y_{k+1-r}(t)] \\ &&- [y_{k+1}(t)-\theta_{1}y_{k}(t)-\theta_{2}y_{k-1}(t)-\cdots-\theta_{r}y_{k+1-r}(t)] \\ & =&\sum\limits_{j=1}^{r}\theta_{j}e_{k+1-j}(t)-{\mathit{\boldsymbol{C}}}_{1}[{\mathit{\boldsymbol{x}}}_{k+1}(t)-\theta_{1}{\mathit{\boldsymbol{x}}}_{k}(t)-\theta_{2}{\mathit{\boldsymbol{x}}}_{k-1}(t)-\cdots-\theta_{r}{\mathit{\boldsymbol{x}}}_{k+1-r}(t)] \\ & =&\sum\limits_{j=1}^{r}\theta_{j}e_{k+1-j}(t)-{\mathit{\boldsymbol{C}}}_{1}\int_{0}^{t}\exp({\mathit{\boldsymbol{A}}}_{1}(t-\tau)){\mathit{\boldsymbol{B}}}_{1}[u_{k+1}(\tau)-\theta_{1}u_{k}(\tau) \\ &&- \theta_{2}u_{k-1}(\tau)-\cdots-\theta_{r}u_{k+1-r}(\tau)] \\ & =&\sum\limits_{j=1}^{r}\theta_{j}e_{k+1-j}(t)-{\mathit{\boldsymbol{C}}}_{1}\int_{0}^{t}\exp({\mathit{\boldsymbol{A}}}_{1}(t-\tau)){\mathit{\boldsymbol{B}}}_{1}\{\theta_{1}[\Gamma_{{\rm p}, 1}e_{k}(\tau)+\Gamma_{{\rm d}, 1}\dot{e}_{k}(\tau)] \\ &&+ \cdots+\theta_{r}[\Gamma_{{\rm p}, r}e_{k+1-r}(\tau)+\Gamma_{{\rm d}, r}\dot{e}_{k+1-r}(\tau)]\}{\rm d}\tau \\ & =&\sum\limits_{j=1}^{r}\theta_{j}e_{k+1-j}(t)-{\mathit{\boldsymbol{C}}}_{1}\int_{0}^{t}\exp({\mathit{\boldsymbol{A}}}_{1}(t-\tau)){\mathit{\boldsymbol{B}}}_{1}[\theta_{1}\Gamma_{{\rm p}, 1}e_{k}(t)+\theta_{2}\Gamma_{{\rm p}, 2}e_{k-1}(t) \\ & &+\cdots+\theta_{r}\Gamma_{{\rm p}, r}e_{k+1-r}(t)]{\rm d}\tau-{\mathit{\boldsymbol{C}}}_{1}\int_{0}^{t}\exp({\mathit{\boldsymbol{A}}}_{1}(t-\tau)){\mathit{\boldsymbol{B}}}_{1}[\theta_{1}\Gamma_{{\rm d}, 1}\dot{e}_{k}(\tau) \\ &&+ \theta_{2}\Gamma_{{\rm d}, 2}\dot{e}_{k-1}(\tau)+\cdots+\theta_{r}\Gamma_{{\rm d}, r}\dot{e}_{k+1-r}(\tau)]{\rm d}\tau, \end{eqnarray} $ (3.16)

利用分部积分法, (3.16)式最后一部分可变为

$ \begin{eqnarray} &&{\mathit{\boldsymbol{C}}}_{1}\int_{0}^{t}\exp({\mathit{\boldsymbol{A}}}_{1}(t-\tau)){\mathit{\boldsymbol{B}}}_{1}[\theta_{1}\Gamma_{{\rm d}, 1}\dot{e}_{k}(\tau)+\cdots+\theta_{r}\Gamma_{{\rm d}, r}\dot{e}_{k+1-r}(\tau)]{\rm d}\tau \\ & =&{\mathit{\boldsymbol{C}}}_{1}\exp({\mathit{\boldsymbol{A}}}_{1}(t-\tau)){\mathit{\boldsymbol{B}}}_{1}[\theta_{1}\Gamma_{{\rm d}, 1}e_{k}(\tau)+\cdots+\theta_{r}\Gamma_{{\rm d}, r}e_{k+1-r}(\tau)]|_{\tau=0}^{\tau=t} \\ &&+ {\mathit{\boldsymbol{C}}}_{1}\int_{0}^{t}\exp({\mathit{\boldsymbol{A}}}_{1}(t-\tau)){\mathit{\boldsymbol{A}}}_{1}{\mathit{\boldsymbol{B}}}_{1}[\theta_{1}\Gamma_{{\rm d}, 1}e_{k}(\tau)+\cdots+\theta_{r}\Gamma_{{\rm d}, r}e_{k+1-r}(\tau)]{\rm d}\tau \\ & =&{\mathit{\boldsymbol{C}}}_{1}{\mathit{\boldsymbol{B}}}_{1}[\theta_{1}\Gamma_{{\rm d}, 1}e_{k}(t)+\cdots+\theta_{r}\Gamma_{{\rm d}, r}e_{k+1-r}(t)] \\ &&+ {\mathit{\boldsymbol{C}}}_{1}\int_{0}^{t}\exp({\mathit{\boldsymbol{A}}}_{1}(t-\tau)){\mathit{\boldsymbol{A}}}_{1}{\mathit{\boldsymbol{B}}}_{1}[\theta_{1}\Gamma_{{\rm d}, 1}e_{k}(\tau)+\cdots+\theta_{r}\Gamma_{{\rm d}, r}e_{k+1-r}(\tau)]{\rm d}\tau. \end{eqnarray} $ (3.17)

将(3.17)式代入(3.16)式, 并结合假设3.2, 得

$ \begin{eqnarray} e_{k+1}(t)&=&\sum\limits_{j=1}^{r}\theta_{j}[(1-{\mathit{\boldsymbol{C}}}_{1} {\mathit{\boldsymbol{C}}}_{1}\Gamma_{{\rm d}, j})e_{k+1-j}(t)\\ &&-\int_{0}^{t}{\mathit{\boldsymbol{C}}}_{1}\exp({\mathit{\boldsymbol{A}}}_{1}(t-\tau))({\mathit{\boldsymbol{A}}}_{1}{\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm d}, j}+ {\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm p}, j})e_{k+1-j}(\tau){\rm d}\tau]. \end{eqnarray} $ (3.18)

对(3.18)式两边取Lebesgue-$p$范数并利用广义Young不等式, 得

$ \begin{eqnarray} ||e_{k+1}(\cdot)||_{p}&\leq&\sum\limits_{j=1}^{r}\theta_{j}[|1-{\mathit{\boldsymbol{C}}}_{1} {\mathit{\boldsymbol{C}}}_{1}\Gamma_{{\rm d}, j}|+||{\mathit{\boldsymbol{C}}}_{1} \exp({\mathit{\boldsymbol{A}}}_{1}(\cdot))({\mathit{\boldsymbol{A}}}_{1}{\mathit{\boldsymbol{B}}}_{1} \Gamma_{{\rm d}, j}+{\mathit{\boldsymbol{B}}}_{1}\Gamma_{{\rm p}, j})||_{1}] ||e_{k+1-j}(\cdot)||_{p}\\ &=&\sum\limits_{j=1}^{r}\theta_{j}\rho_{1, j}||e_{k+1-j}(\cdot)||_{p}. \end{eqnarray} $ (3.19)

由于权重系数$\theta_{l}(l=1, 2, 3, \cdots, r)$满足$0\leq \theta_{l}\leq 1(l=1, 2, 3, \cdots, r), \sum\limits_{l=1}^{r}\theta_{l}=1$, 收敛条件(3.15)当$i=1$时成立, 故$\sum\limits_{j=1}^{r}\theta_{j}\rho_{1, j}<1$.由推论3.1知, $\lim\limits_{k\rightarrow +\infty}||e_{k+1}(\cdot)||=0$在区间$\Omega_{1}$成立.

(Ⅱ)当$t\in\Omega_{2}$, 跟踪误差(3.16)变为

$ \begin{eqnarray} e_{k+1}(t)&=&\sum\limits_{j=1}^{r}\theta_{j}e_{k+1-j}(t)-{\mathit{\boldsymbol{C}}}_{2} \bigg[\sum\limits_{j=1}^{r}\theta_{j}({\mathit{\boldsymbol{x}}}_{k+1}(t)-{\mathit{\boldsymbol{x}}}_{k+1-j}(t))\bigg] \\ & =&\sum\limits_{j=1}^{r}\theta_{j}e_{k+1-j}(t)-{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(t-t_{1})) \bigg[\sum\limits_{j=1}^{r}\theta_{j}({\mathit{\boldsymbol{x}}}_{k+1}(t_{1})-{\mathit{\boldsymbol{x}}}_{k+1-j}(t_{1}))\bigg] \\ &&- {\mathit{\boldsymbol{C}}}_{2}\int_{t_{1}}^{t}\exp({\mathit{\boldsymbol{A}}}_{2}(t-\tau)){\mathit{\boldsymbol{B}}}_{2} \bigg[\sum\limits_{j=1}^{r}\theta_{j}(u_{k+1}(\tau)-u_{k+1-j}(\tau))\bigg]{\rm d}\tau \\ & =&\sum\limits_{j=1}^{r}\theta_{j}e_{k+1-j}(t)-{\mathit{\boldsymbol{C}}}_{2} \bigg[\sum\limits_{j=1}^{r}\theta_{j}\exp({\mathit{\boldsymbol{A}}}_{2}(t-t_{1}))({\mathit{\boldsymbol{x}}}_{k+1}(t_{1})-{\mathit{\boldsymbol{x}}}_{k+1-j}(t_{1}))\bigg] \\ &&- {\mathit{\boldsymbol{C}}}_{2}\int_{t_{1}}^{t}\exp({\mathit{\boldsymbol{A}}}_{2}(t-\tau)) {\mathit{\boldsymbol{B}}}_{2}\bigg[u_{k+1}(\tau)-\sum\limits_{j=1}^{r}\theta_{j}u_{k+1-j} (\tau)\bigg]{\rm d}\tau. \end{eqnarray} $ (3.20)

将学习律(2.9)代入(3.20)式并重组, 得

$ \begin{eqnarray} e_{k+1}(t)&=&\sum\limits_{j=1}^{r}\theta_{j}e_{k+1-j}(t)-{\mathit{\boldsymbol{C}}}_{2} \bigg[\sum\limits_{j=1}^{r}\theta_{j}\exp({\mathit{\boldsymbol{A}}}_{2}(t-t_{1}))({\mathit{\boldsymbol{x}}}_{k+1}(t_{1})-{\mathit{\boldsymbol{x}}}_{k+1-j}(t_{1}))\bigg] \\ &&- {\mathit{\boldsymbol{C}}}_{2}\bigg[\sum\limits_{j=1}^{r}\theta_{j}\int_{t_{1}}^{t}\exp({\mathit{\boldsymbol{A}}}_{2}(t-\tau)){\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm p}, j}e_{k+1-j}(\tau){\rm d}\tau\bigg] \\ && -{\mathit{\boldsymbol{C}}}_{2}\bigg[\sum\limits_{j=1}^{r}\theta_{j}\int_{t_{1}}^{t}\exp({\mathit{\boldsymbol{A}}}_{2} (t-\tau)){\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}, j}\dot{e}_{k+1-j}(\tau){\rm d}\tau\bigg]. \end{eqnarray} $ (3.21)

利用分部积分法, 并记$\triangle {\mathit{\boldsymbol{x}}}_{k+1-j}(t_{1}) ={\mathit{\boldsymbol{x}}}_{k+1}(t_{1})-{\mathit{\boldsymbol{x}}}_{k+1-j}(t_{1})$, (3.21)式可变为

$ \begin{eqnarray} e_{k+1}(t)&=&\sum\limits_{j=1}^{r}\theta_{j}e_{k+1-j}(t)-{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(t-t_{1})) \bigg[\sum\limits_{j=1}^{r}\theta_{j}\triangle {\mathit{\boldsymbol{x}}}_{k+1-j}(t_{1})\bigg] \\ &&-{\mathit{\boldsymbol{C}}}_{2}\bigg[\sum\limits_{j=1}^{r}\theta_{j} \int_{t_{1}}^{t}\exp({\mathit{\boldsymbol{A}}}_{2}(t-\tau)){\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm p}, j}e_{k+1-j}(\tau){\rm d}\tau\bigg] \\ &&-{\mathit{\boldsymbol{C}}}_{2}\bigg\{\sum\limits_{j=1}^{r}\theta_{j} \bigg[\exp({\mathit{\boldsymbol{A}}}_{2}(t-\tau)){\mathit{\boldsymbol{B}}}_{2} \Gamma_{{\rm d}, j}e_{k+1-j}(t)|_{\tau=t_{1}}^{\tau=t}\\ &&+\int_{t_{1}}^{t}\exp({\mathit{\boldsymbol{A}}}_{2}(t-\tau)){\mathit{\boldsymbol{A}}}_{2}{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}, j}e_{k+1-j}(\tau){\rm d}\tau \bigg]\bigg\} \\ & =&\sum\limits_{j=1}^{r}\theta_{j}(1-{\mathit{\boldsymbol{C}}}_{2}{\mathit{\boldsymbol{B}}}_{2} \Gamma_{{\rm d}, j})e_{k+1-j}(t) \\ &&-\sum\limits_{j=1}^{r}\theta_{j}\int_{t_{1}}^{t} {\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(t-\tau)) ({\mathit{\boldsymbol{A}}}_{2}{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}, j} +{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm p}, j})e_{k+1-j}(\tau){\rm d}\tau \\ &&+\sum\limits_{j=1}^{r}\theta_{j}{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2} (t-t_{1})) {\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}, j} e_{k+1-j}(t_{1}) \\ &&-{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(t-t_{1})) \bigg[\sum\limits_{j=1}^{r}\theta_{j}\triangle {\mathit{\boldsymbol{x}}}_{k+1-j}(t_{1})\bigg], \end{eqnarray} $ (3.22)

对(3.22)式两边取Lebesgue-$p$范数, 并利用广义Young不等式, 得

$ \begin{eqnarray*} ||e_{k+1}(\cdot)||_{p}&\leq&\sum\limits_{j=1}^{r}\theta_{j}[|1-{\mathit{\boldsymbol{C}}}_{2}{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}, j}|+||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(\cdot))({\mathit{\boldsymbol{A}}}_{2}{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}, j}+{\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm p}, j})||_{1}] ||e_{k+1-j}(\cdot)||_{p}\\ &&+\sum\limits_{j=1}^{r}\theta_{j}||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(\cdot)){\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}, j}||_{p}||e_{k+1-j}(t_{1})||_{p}\\ &&+\sum\limits_{j=1}^{r}\theta_{j}||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(\cdot))||_{p}||\triangle {\mathit{\boldsymbol{x}}}_{k+1-j}(t_{1})||_{p}, \end{eqnarray*} $

$ \begin{eqnarray} ||e_{k+1}(\cdot)||_{p}&\leq&\sum\limits_{j=1}^{r}\theta_{j}\rho_{2, j}||e_{k+1-j}(\cdot)||_{p}+\sum\limits_{j=1}^{r}\theta_{j}||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(\cdot)){\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}, j}||_{p}||e_{k+1-j}(t_{1})||_{p} \\ && +\sum\limits_{j=1}^{r}\theta_{j}||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(\cdot))||_{p}||\triangle {\mathit{\boldsymbol{x}}}_{k+1-j}(t_{1})||_{p}. \end{eqnarray} $ (3.23)

由证明过程(Ⅰ)知, $\lim\limits_{k\rightarrow +\infty}y_{k}(t_{1})=y_{{\rm d}}(t_{1})$$\lim\limits_{k\rightarrow +\infty}{\mathit{\boldsymbol{x}}}_{k}(t_{1})={\mathit{\boldsymbol{x}}}_{{\rm d}}(t_{1})$$t_{1}$时刻成立.这意味着$\lim\limits_{k\rightarrow +\infty}||e_{k+1-j}(t_{1})||_{p}=0$$\lim\limits_{k\rightarrow +\infty}||\triangle{\mathit{\boldsymbol{x}}}_{k+1-j}(t_{1})||_{p}=0$对任意$j=1, 2, \cdots, r$都成立.另外, $r$为有限正整数, 因此

$ \begin{equation} \begin{array}{l} \lim\limits_{k\rightarrow +\infty} \bigg(\sum\limits_{j=1}^{r} \theta_{j}||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(\cdot)){\mathit{\boldsymbol{B}}}_{2}\Gamma_{{\rm d}, j}||_{p}||e_{k+1-j} (t_{1})||_{p}\bigg)=0, \\[4mm] \lim\limits_{k\rightarrow +\infty} \bigg(\sum\limits_{j=1}^{r}\theta_{j}||{\mathit{\boldsymbol{C}}}_{2}\exp({\mathit{\boldsymbol{A}}}_{2}(\cdot))||_{p}||\triangle {\mathit{\boldsymbol{x}}}_{k+1-j}(t_{1})||_{p}\bigg)=0. \end{array} \end{equation} $ (3.24)

又因为权重系数$\theta_{i}(j=1, 2, 3, \cdots, r)$满足$0\leq \theta_{j}\leq 1, \sum\limits_{j=1}^{r}\theta_{j}=1$, 结合条件(3.15)得, $\sum\limits_{j=1}^{r}\theta_{j}\rho_{2, j}<1$.根据(3.23)-(3.24)式、引理3.1和推论3.1知, $\lim\limits_{k\rightarrow +\infty}||e_{k+1}(\cdot)||=0$在区间$\Omega_{2}$成立.

重复上述过程, 可以类似地证明$\lim\limits_{k\rightarrow +\infty}||e_{k+1}(\cdot)||=0$在区间$\Omega_{j}(j=3, 4, \cdots, q)$上成立, 因此结论在整个学习区间$\Omega$上成立.证毕.

注3.4  定理3.2表明, 高阶${\rm PD}$ -型迭代学习控制算法对切换系统${\rm (2.1)}$是有效的.与一阶算法比较, 高阶算法以加权均值的方式充分利用了之前的学习信息, 这便于我们灵活地选择学习增益和权重参数, 以最大限度地减弱系统不确定性对学习效果的影响.至于如何选择阶数$r$, 从理论上讲, 没有统一的方法; 但是在实际应用中, 可以根据实际经验机动选取.一般而言, 结合实际情况, 阶数的选取不超过3[21].

注3.5  $({\rm 3.5})$$({\rm 3.15})$式可知, 虽然Lebesgue-$p$范数意义下的收敛条件比$\lambda$ -范数意义下保守; 但是它利用学习区间上的积分度量跟踪误差的动态行为, 具有实际的物理意义, 且揭示了学习过程的自然属性以及学习增益和系统动态学对学习性能的影响.这便于我们在实际应用中设计更加有效的控制器.

4 数值仿真

例1  考虑具有两个二阶子系统的切换系统, 其动态学如下

$ {\mathit{\boldsymbol{A}}}_{1}= \left[\begin{array}{cc} -0.5&0\\ 0.3&-0.4\\ \end{array}\right], ~~~ {\mathit{\boldsymbol{B}}}_{1}= \left[\begin{array}{cc} 1\\ 0.3\\ \end{array}\right], ~~~ {\mathit{\boldsymbol{C}}}_{1}= \left[\begin{array}{cc} 1&0.5 \end{array}\right], $
$ {\mathit{\boldsymbol{A}}}_{2}= \left[\begin{array}{cc} -0.6&0.3\\ 0.2&0\\ \end{array}\right], ~~~ {\mathit{\boldsymbol{B}}}_{2}= \left[\begin{array}{cc} 1\\ 0.2\\ \end{array}\right], ~~~ {\mathit{\boldsymbol{C}}}_{2}= \left[\begin{array}{cc} 0.5&1 \end{array}\right], $

其中, 切换信号如图 4.1所示.设时间区间$\Omega=[0, 1]$, 采样步长$\triangle t=0.02$秒, 初始状态${\mathit{\boldsymbol{x}}}_{k}(0)={\mathit{\boldsymbol{x}}}_{{\rm d}}(0)=\bf{0}\in \mathbb{R}^{2}$, 初始控制信号$u_{0}(t)=0$, 目标轨线为$y_{{\rm d}}(t)=-t^{2}+t (t\in \Omega)$.

图 4.1 例1的切换律$s(t)$

对于一阶算法(2.8), 学习增益分别选为$\Gamma_{{\rm d}}=0.6$, $\Gamma_{{\rm p}}=0.2$, 在此条件下, 收敛条件(3.5)为$\rho_{1}=0.4102$, $\rho_{2}=0.6494$.

系统的跟踪效果如图 4.2所示, 其中点线表示目标轨线, 点划线表示系统第5次迭代的输出轨线, 实线是第10次迭代的输出轨线.显然, 随着迭代次数的增加系统的输出轨线可以渐近跟踪到目标轨线. Lebesgure-2范数意义下跟踪误差曲线如图 4.3所示.显然, 当$k=15$次时, 跟踪误差已接近于0.这也表明系统可渐近跟踪目标轨线.

图 4.2 一阶算法控制下系统跟踪性能

图 4.3 $l_{2}$范数意义下的误差曲线

对于(2.9)式, 选$r=2$和两组不同的学习增益进行仿真.第一组为$\Gamma_{{\rm d}, 1}=0.8$, $\Gamma_{{\rm p}, 1}=0.3$, 分别得收敛条件$\rho_{1, 1}=0.2734$$\rho_{2, 1}=0.5925$; 另一组为$\Gamma_{{\rm d}, 2}=1.2$, $\Gamma_{{\rm p}, 2}=0.4$, 分别得收敛条件$\rho_{1, 2}=0.5805$$\rho_{2, 2}=0.2988$.权重系数取三组值, 分别为$\theta_{1}=0.3$, $\theta_{2}=0.7$, $\theta_{1}=\theta_{2}=0.5$$\theta_{1}=0.7$, $\theta_{2}=0.3$.

$\theta_{1}=\theta_{2}=0.5$时, 系统的跟踪情况如图 4.4所示, 误差曲线如图 4.5所示.结果表明, 在二阶算法控制下, 系统可以渐近跟踪目标轨线.

图 4.4 权重系数相同时二阶算法控制下系统跟踪性能

图 4.5 二阶算法生成的误差曲线

例2[27]  考虑具有两个具有三阶子系统的切换系统, 其动态学如下

$ {\mathit{\boldsymbol{A}}}_{1}= \left[\begin{array}{ccc} 0.5108&-0.9147&-0.2000\\ -0.6563&0.1798&0.1130\\ 0.8810&-0.7841&0.1000 \end{array}\right], \\ {\mathit{\boldsymbol{B}}}_{1}= \left[\begin{array}{c} 0.1056\\ 0.1284\\ 0.1000 \end{array}\right], {\mathit{\boldsymbol{C}}}_{1}= \left[\begin{array}{ccc} 0.0100&0.0600&0.0300 \end{array}\right], $
$ {\mathit{\boldsymbol{A}}}_{2}= \left[\begin{array}{ccc} -0.1250&-0.9833&-0.3400\\ -0.5305&0.3848&0.5800\\ 1.0306&0.6521&0.1000 \end{array}\right], \\ {\mathit{\boldsymbol{B}}}_{2}= \left[\begin{array}{c} 0.7425\\ 0.1436\\ 0.1000 \end{array}\right], {\mathit{\boldsymbol{C}}}_{2}= \left[\begin{array}{ccc} 0.0100&0.0200 &0.0500 \end{array}\right], $

其中, 切换信号如图 4.6所示, 设时间区间$\Omega=[0, 1]$, 采样步长$\triangle t=0.01$秒, 初始状态${\mathit{\boldsymbol{x}}}_{k}(0)={\mathit{\boldsymbol{x}}}_{{\rm d}}(0)=\bf{0}\in \mathbb{R}^{3}$, 初始控制信号$u_{0}(t)=0$, 目标轨线为$y_{{\rm d}}(t)=\sin(10t) (t\in \Omega)$.

图 4.6 例2的切换信号

对于一阶算法(2.8), 学习增益分别选为$\Gamma_{{\rm d}}=1.6$, $\Gamma_{{\rm p}}=-0.3$, 在此条件下, 收敛条件(3.5)为$\rho_{1}=0.9563$, $\rho_{2}=0.9837$.仿真结果如图 4.7-4.8所示. 图 4.7是迭代50次、150次和500次系统的跟踪效果, 图 4.8是学习800次的跟踪误差曲线.结果表明, 在一阶算法控制下, 系统可以在学习区间$\Omega$内渐近跟踪目标轨线.

图 4.7 一阶算法控制下系统的跟踪性能

图 4.8 一阶算法产生的误差曲线

对于二阶算法(2.9), 选两组不同的学习增益.第一组为$\Gamma_{{\rm d}, 1}=2.5$, $\Gamma_{{\rm p}, 1}=-0.2$, 分别得收敛条件$\rho_{1, 1}=0.9419$$\rho_{2, 1}=0.9945$; 另一组为$\Gamma_{{\rm d}, 2}=1.5$, $\Gamma_{{\rm p}, 2}=-0.4$, 分别得收敛条件$\rho_{1, 2}=0.9545$$\rho_{2, 2}=0.9758$.权重系数分别取$\theta_{1}=0.2$, $\theta_{2}=0.8$, $\theta_{1}=\theta_{2}=0.5$$\theta_{1}=0.8$, $\theta_{2}=0.2$.

$\theta_{1}=\theta_{2}=0.5$时, 系统的跟踪效果如图 4.9所示, 跟踪误差曲线如图 4.10所示.结果表明, 随着学习次数的增加, 系统输出能够渐近跟踪到目标轨线.

图 4.9 二阶算法控制下系统的跟踪性能

图 4.10 二阶算法产生的误差曲线
5 结论

针对一类线性连续切换系统, 分析了一阶和二阶PD -型迭代学习控制算法的收敛性, 在Lebesgue-$p$范数意义下讨论了跟踪误差的动态性质, 并利用广义Young不等式推导了算法收敛的充分条件, 该收敛性条件揭示了系统动态学和学习增益对学习收敛性的影响.数值仿真表明, 算法对由任意时变切换信号支配的切换系统是可行且有效的.

参考文献
[1] Chai T Y, Ding J L, Wang H, Su C Y. Hybrid intelligent optimal control method for operation of complex industrial processes. Acta Automatical Sinica, 2008, 34(5): 505–515.
[2] Zhang X, Xue D F, Wang J Y, Feng X Q. Improved growth technology of large MgO single crystals. Journal of Crystal Growth, 2006, 292(2): 505–509. DOI:10.1016/j.jcrysgro.2006.04.058
[3] Lin H, Antsaklis P J. Stability and stabilizability of switched linear systems:a survey of recent results. IEEE Transactions on Automatic Control, 2009, 54(2): 308–322. DOI:10.1109/TAC.2008.2012009
[4] Zhao X, Yin S, Li H, et al. Switching stabilization for a class of slowly switched systems. IEEE Transactions on Automatic Control, 2015, 60(1): 221–226. DOI:10.1109/TAC.2014.2322961
[5] Niu B, Zhao J. Barrier Lyapunov functions for the output tracking control of constrained nonlinear switched systems. Systems and Control Letters, 2013, 62(10): 963–971. DOI:10.1016/j.sysconle.2013.07.003
[6] Lian J, Ge Y. Robust output tracking control for switched systems under asynchronous switching. Nonlinear Analysis:Hybrid Systems, 2013, 8: 57–68. DOI:10.1016/j.nahs.2012.10.003
[7] Li Q K, Zhao J, Liu X J, et al. Observer-based tracking control for switched linear systems with timevarying delay. International Journal of Robust and Nonlinear Control, 2011, 21(3): 309–327. DOI:10.1002/rnc.v21.3
[8] Hou L, Zong G, et al. Exponential l1-l output tracking control for discrete-time switched system with time-varying delay. International Journal of Robust and Nonlinear Control, 2012, 22(11): 1175–1194. DOI:10.1002/rnc.v22.11
[9] Tong S, Zhang L, Li Y. Observed-based adaptive fuzzy decentralized tracking control for switched uncertain nonlinear large-scale systems with dead zones. IEEE Transactions on Systems, Man, and Cybernetics:Systems, 2016, 46(1): 37–47. DOI:10.1109/TSMC.2015.2426131
[10] Niu B, Zhao J. Tracking control for output-constrained nonlinear switched systems with a barrier Lyapunov function. International Journal of Systems Science, 2013, 44(5): 978–985. DOI:10.1080/00207721.2011.652222
[11] Zhao X, Shi P, Zheng X, et al. Adaptive tracking control for switched stochastic nonlinear systems with unknown actuator dead-zone. Automatica, 2015, 60: 193–200. DOI:10.1016/j.automatica.2015.07.022
[12] Tharayil D A B, Alleyne A G. A survey of iterative learning control:a learning-based method for high performance tracking control. IEEE Control Systems Magazine, 2006, 26(3): 96–114. DOI:10.1109/MCS.2006.1636313
[13] Ahn H S, Chen Y Q, Koore K L. Iterative learning control:brief survey and categorization. IEEE Transactions on Systems Man and Cybernetics Part C Applications and Reviews, 2007, 37(6): 1099–1121. DOI:10.1109/TSMCC.2007.905759
[14] Ruan X, Bien Z Z, Wang Q. Convergence properties of iterative learning control processes in the sense of the Lebesgue-p norm. Asian Journal of Control, 2012, 14(4): 1095–1107. DOI:10.1002/asjc.425
[15] Bu X H, Hou Z S, Jin S T. An iterative learning control design approach for networked control systems with data dropouts. International Journal of Robust and Nonlinear Control, 2016, 26(1): 91–109. DOI:10.1002/rnc.3300
[16] Yang X. A PD-Type iterative learning control for a class of switched discrete-time systems with model uncertainties and external noises. Discrete Dynamics in Nature and Society, 2015, Article ID: 410292 http://www.researchgate.net/publication/282209779_A_PD-Type_Iterative_Learning_Control_for_a_Class_of_Switched_Discrete-Time_Systems_with_Model_Uncertainties_and_External_Noises
[17] Bu X H, Yu F S, Hou Z S, et al. Iterative learning control for a class of linear discrete-time switched systems. Acta Automatica Sinica, 2013, 39(9): 1564–1569. DOI:10.1016/S1874-1029(13)60073-7
[18] Yang X, Ruan X. Analysis of iterative learning control for a class of linear discrete-time switched systems. Abstract and Applied Analysis, 2015, Article ID: 464175 http://www.researchgate.net/publication/271730482_Analysis_of_Iterative_Learning_Control_for_a_Class_of_Linear_Discrete-Time_Switched_Systems
[19] Bu X, Hou Z, Yu F, Fu Z. Iterative learning control for a class of non-linear switched systems. IET Control Theory and Applications, 2013, 7(3): 470–481. DOI:10.1049/iet-cta.2012.0714
[20] Bu X, Yu F, Hou Z. Iterative learning control for linear switched systems with arbitrary switched rules. Intelligent Control and Automation (WCICA), 2012 World Congress on, 2012, DOI: 10.1109/WCICA.2012.6358060 http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6358060
[21] Bu X, Yu F, Fu Z, et al. Stability analysis of high-order iterative learning control for a class of nonlinear switched systems. Abstract and Applied Analysis, 2013, Article ID: 684642 http://www.ams.org/mathscinet-getitem?mr=3081572
[22] Cao W, Sun M. Iterative learning control for discrete-time switched systems with attenuation factor. Journal of Vibration and Control, 2016, 22(12): 2898–2906. DOI:10.1177/1077546314555424
[23] Shao Z, Xiang Z. Design of an iterative learning control law for a class of switched repetitive systems. Circuits, Systems, and Signal Processing, 2017, 36(2): 845–866. DOI:10.1007/s00034-016-0331-6
[24] Wang J, Luo J, Hu Y. Monotonically convergent hybrid ILC for uncertain discrete-time switched systems with state delay. Transactions of the Institute of Measurement and Control, 2017, 39(7): 1047–1058. DOI:10.1177/0142331215627002
[25] Pinsky M A. Introduction to Fourier Analysis and Wavelets. Providence, RI: American Mathematical Soc, 2009
[26] Ruan X, Bien Z Z. Pulse compensation for PD-type iterative learning control against initial state shift. International Journal of Systems Science, 2012, 43(11): 2172–2184. DOI:10.1080/00207721.2011.569586
[27] Yuan C, Wu C F, Wu F. Hybrid control for switched linear systems with average dwell time. IEEE Transactions on Automatic Control, 2015, 60(1): 240–245. DOI:10.1109/TAC.2014.2322941