-
摘要: 面对中国社会快速老龄化现状和庞大的残疾人群，康复辅助机器人研究具有重要学术价值和广阔的应用前景.康复辅助机器人研究涉及神经科学、生物力学、机器人自动控制等领域知识，是机器人最具挑战性和最受关注的研究领域之一.与其他机器人不同，康复辅助机器人的作用对象是人，存在人与机器人的信息交流和能量交换，物理人机交互控制方法是其研究核心和关键技术.本文以神经康复机器人、穿戴式外骨骼、智能假肢等应用为例，介绍当前的研究现状，并重点介绍人体运动意图识别方法和交互控制方法等研究重点和难点.最后展望该领域的未来技术发展方向.Abstract: With the rapid growth of China's aging and disabled population, researches on rehabilitation and assistive robots have great value to technical study and good prospects for applications. Rehabilitation and assistive robots involve the domain knowledge of neuroscience, biomechanics, robot automatic control, etc., thus are one of the most challenging and most concerned research areas of robotics. As the application contexts of rehabilitation and assistance entail information communication and energy exchange between man and robot, the physical human-robot interaction control method is the key theme. This article briefly introduces the states of the art of neurorehabilitation robots, wearable exoskeletons, and intelligent prostheses, and focuses on the key issues and research priorities in human motion intent detection methods and interaction control methods. Finally, the future development directions are prospected.
-
工业过程运行反馈控制包括底层回路关键被控变量的反馈控制和上层运行指标的反馈控制, 也就是说, 工业过程运行反馈控制不仅包括保证过程控制系统关键被控变量的跟踪控制, 而且还要选择合适的关键被控变量设定值, 实现运行指标目标值的跟踪[1].如图 1所示, 传统的工业过程运行反馈控制过程中, 关键被控变量的设定值$ {Y^*} = y_j^*$, $j=1, 2$, $\cdots$, $n$由工序作业班的运行工程师根据运行指标目标值${R^*} = R_i^*$, $j=1, 2, \cdots, m$和多年积累的人工操作经验, 并结合各种运行工况信息人为给出.为实现工业运行过程的自动控制, 自上世纪80年代末以来, 很多学者开展了工业运行过程控制方法的研究.文献[2]基于分层递阶控制的架构和多层优化理论, 提出了反馈优化控制的思想.文献[3]通过离线选择与工业过程经济效益相关的被控变量的设定值, 提出了自优化控制的概念.文献[4]将底层回路控制与过程运行优化相结合, 提出了具有两层结构的实时优化(Real time optimization, RTO)控制方法, 上层采用静态模型优化经济性能指标, 产生底层控制回路的设定值, 通过底层控制器使被控变量跟踪设定值, 从而尽可能使过程运行在经济指标目标值附近.文献[5]将RTO与模型预测控制相结合, 提出了具有三层结构的运行反馈控制方法.此外, 还有一些基于神经网络、模糊推理、案例推理等智能技术的运行反馈控制方法, 例如文献[6]将案例推理、规则推理以及神经网络相结合, 提出了工业运行过程的混合智能控制方法; 文献[7]将神经网络与模糊推理相结合, 提出了一种设定值的混合监控方法.
上述运行反馈控制方法均假设底层过程控制可以跟踪运行控制给出的设定值, 没有考虑底层跟踪设定值的动态误差对整个运行过程优化和控制的影响.为解决这一问题, 文献[8-9]提出了使运行指标实际值与目标值偏差和控制回路输出与设定值跟踪误差的二次性能指标极小化的运行反馈控制方法.文献[10]提出了运行反馈解耦控制方法.上述方法均假设运行层的模型由底层关键被控变量与运行指标之间的静态模型精确描述.实际上, 运行指标反映产品在加工过程中的质量、效率、消耗等, 与底层控制回路的被控变量之间往往具有动态特性, 并且很难用精确的数学模型描述.
本文针对一类运行层为未知动态模型的工业运行过程, 提出一种新的多模型自适应控制方法.最早的多模型自适应控制方法通过线性模型和基于神经网络的非线性模型之间的切换不仅可以保证自适应系统有界输入和有界输出(Bound-input and bound-output, BIBO)稳定, 而且可以改善系统的跟踪性能[11], 但该方法只适用于单输入、单输出系统, 并且是在系统的未建模动态全局有界这一假设下实现的.文献[12]将上述方法推广到多变量系统, 提出了基于多模型与神经网络的多变量自适应控制方法, 放松了文献[11]对系统未建模动态全局有界的假设.文献[13-14]提出了多变量强耦合系统的多模型自适应解耦控制方法.文献[15]提出了参数跳变系统的多模型自适应控制方法.文献[16]提出了具有未知执行器非线性的多变量自适应控制方法.
上述多模型自适应控制方法都是针对底层被控对象设计的.这些方法采用带死区的投影算法对未知参数进行在线辨识.投影算法收敛速度慢, 对参数初值十分灵敏, 实际使用中只有当参数初值接近真值时, 才具有良好的收敛效果, 因此投影算法对过程的先验知识要求较高, 不适合应用于动态未知的工业运行过程.相比较, 递推最小二乘算法具有较快的收敛速度, 对参数初值不灵敏.本文提出的运行过程多模型自适应控制方法采用带死区的递推最小二乘方法对未知参数进行在线辨识.理论分析和仿真实验验证了所提方法的有效性.
1. 问题描述
工业运行过程动态模型由上层运行层的动态模型和底层被控对象的动态模型两部分组成.在本文中, 为了将问题简化, 底层被控对象由如下离散时间线性状态空间模型描述.
$ x(k+1)=\bar{A}x(k)+\bar{B}u(k) $
(1a) $ y(k)=\bar{C}x(k) $
(1b) 其中, $x\in{\bf R}^n$为被控对象状态, $u\in{\bf R}^m$为被控对象的控制输入, $y\in{\bf R}^m$是被控对象的测量输出, $\overline{A}$ $\in$ ${\bf R}^{n\times n}$, $\overline{B}\in{\bf R}^{n\times m}$, $\overline{C}\in{\bf R}^{m\times n}$为时不变矩阵.针对底层被控对象(1)设计极点配置控制器.
$ u(k)=-Kx(k)+L{{y}^{*}}(k) $
(2) 其中, $y^{*}(k)$为底层回路设定值, $K\in{\bf R}^{m\times n}$, $L$ $\in$ ${\bf R}^{m\times m}$为时不变矩阵.
为获得控制器参数矩阵$K$和$L$, 将式(2)代入式(1)得到闭环系统方程为
$ x(k+1)=(\bar{A}-\bar{B}K)x(k)+\bar{B}L{{y}^{*}}(k) $
(3a) $ y(k)=\bar{C}x(k) $
(3b) 为使闭环系统稳定, 并实现稳态跟踪, 应选择控制器参数矩阵和满足:
1) 矩阵$\overline{A}-\overline{B}K$稳定;
2) $\lim\nolimits_{z\rightarrow1}\overline{C}(zI_{n}-(\overline{A}-\overline{B}K))^{-1}\overline{B}L=I_{m}$, $L=$ $\lim\nolimits_{z\rightarrow1}(\overline{C}(zI_{n}- (\overline{A}-\overline{B}K))^{-1}\overline{B})^{-1}$.
由于上层运行层动态模型是底层关键被控变量与运行指标之间的函数, 它的输出与底层控制系统输出相关.在本文中, 考虑运行层模型为如下带有未建模动态的动态模型.
$ r(k+1)=Mr(k)+Ny(k)+\nu (k) $
(4) 其中, $r(k)$为运行过程输出, 即运行过程的工艺指标, $\nu(k)\in {\bf R}^m$为外部干扰或未建模动态, $M$, $N$ $\in$ ${\bf R}^{m\times m}$为时不变矩阵.工业过程运行控制系统涉及到底层关键被控变量的反馈控制和上层运行指标的反馈控制, 为充分考虑底层跟踪设定值的动态误差对整个运行过程控制的影响, 运行过程动态模型可看作是由底层基础反馈控制系统(3)和运行层动态模型(4)构成的广义模型.
$ x(k+1)=\widetilde{A}x(k)+\widetilde{B}{{y}^{*}}(k) $
(5a) $ r(k+1)=Mr(k)+\widetilde{C}x(k)+\nu (k) $
(5b) 其中, $\widetilde{A}=\bar{A}-\bar{B}K$, $\widetilde{B}=\bar{B}L, $ $\widetilde{C}=N\bar{C}, $满足${{\widetilde{C}}^{\text{T}}}\widetilde{C}$可逆.
假设 1. 未建模动态$\nu(k)$的差分项或变化率全局有界, 即对任意的$k > 0$, $\|\nu(k)-\nu(k-2)\|\leq\Gamma$, 其中, $\Gamma$为正常数.
本文的目标是将设定值$y^{*}(k)$看作控制输入, 确定一个多模型自适应控制器, 当其应用于不确定的运行过程(5)时, 闭环运行过程的输入、输出信号有界, 即闭环系统BIBO稳定, 并且运行过程输出$r(k)$尽可能跟踪事先指定的运行指标目标值$r^{*}(k)$的变化.由于未建模动态的存在, 单独使用线性控制器即使能保证闭环运行过程BIBO稳定, 也很难满足一定的跟踪性能.本文将基于带死区的递推最小二乘算法的线性鲁棒自适应控制器和具有未建模动态补偿的非线性自适应控制器与切换机制相结合, 提出的多模型自适应控制器不仅能够保证闭环运行过程BIBO稳定, 而且可使其具有良好的跟踪性能.
2. 基于带死区的递推最小二乘算法的线性鲁棒自适应控制
2.1 一步超前控制器设计
为进行控制器设计, 首先需要将广义模型(5)转化成差分方程形式, 为此引入后移算子$z^{-1}$, 于是式(5)可以重新整理为
$ A({{z}^{-1}})r(k+2)=B{{y}^{*}}(k)+C({{z}^{-1}})\nu (k+1) $
(6) 其中,
$ A({{z}^{-1}})=\widetilde{C}[{{I}_{n}}-\widetilde{A}{{z}^{-1}}]{{({{\widetilde{C}}^{\text{T}}}\widetilde{C})}^{-1}}{{\widetilde{C}}^{\text{T}}}({{I}_{m}}-M{{z}^{-1}}) $
$ B=\widetilde{C}\widetilde{B} $
$ C({z^{ - 1}}) = \widetilde C[{I_n} - \widetilde A{z^{ - 1}}]{({\widetilde C^{\rm{T}}}\widetilde C)^{ - 1}}{\widetilde C^{\rm{T}}} $
下面针对模型(6)设计一步超前控制器.引入如下一步超前最优性能指标:
$ J(k) = {\left\| {T({z^{ - 1}})r(k + 2) - R({z^{ - 1}}){r^*}(k + 2)} \right\|^2} $
(7) 其中, $r^{*}(k)=[r^{*}_{1}(k), r^{*}_{2}(k), \cdots, r^{*}_{m}(k)]^{\rm T}\in{\bf R}^m$为已知有界的运行指标目标值, $T(z^{-1})\in{\bf R}^{m\times m}$为稳定的对角加权多项式矩阵, 满足$T(0)$非奇异; $R(z^{-1})$ $\in$ ${\bf R}^{m\times m}$为对角加权多项式矩阵.引入方程
$ T({z^{ - 1}}) = H({z^{ - 1}})A({z^{ - 1}}) + {z^{ - 2}}G({z^{ - 1}}) $
(8) 为使$H(z^{-1})$和$G(z^{-1})$为唯一解或最小阶解, 由文献[17]可知, $H(z^{-1})$和$G(z^{-1})$都为1阶多项式矩阵, $T(z^{-1})$的阶次小于或等于3.易知, $H(0)=T(0)$.将式(6)两边乘$H(z^{-1})$并利用式(8), 得
$ \begin{array}{l} T({z^{ - 1}})r(k + 2) = G({z^{ - 1}})r(k) + \\ \;\;\;\;\;\;H({z^{ - 1}})B{y^*}(k) + H({z^{ - 1}})C({z^{ - 1}})\nu (k + 1) \end{array} $
(9) 定义时滞-差分算子$\Delta=1-z^{-2}$, 则式(9)转化为
$ \begin{array}{l} T({z^{ - 1}})r(k + 2) = G({z^{ - 1}})\Delta r(k) + \\ \;\;\;\;\;H({z^{ - 1}})B\Delta {y^*}(k) + T({z^{ - 1}})r(k) + \rho (k) \end{array} $
(10) 其中, $\rho(k)=H(z^{-1})C(z^{-1})[\nu(k+1)-\nu(k-1)]$.于是, 使性能指标(7)最小的一步超前最优控制$y^{*}(k)$通过下式计算.
$ \begin{array}{l} G({z^{ - 1}})\Delta r(k) + H({z^{ - 1}})B\Delta {y^*}(k) + \rho (k) = \\ \;\;\;\;\;\;R({z^{ - 1}}){r^*}(k + 2) - T({z^{ - 1}})r(k) \end{array} $
(11) 将式(11)代入模型(6), 得到运行过程闭环方程
$ T({z^{ - 1}})r(k + 2) = R({z^{ - 1}}){r^*}(k + 2) $
(12) 由式(12)可知, 若选择$R(z^{-1})=T(z^{-1})$, 则可消除运行过程的跟踪误差.
由于外部干扰或未建模动态往往是未知的, 当不考虑它对运行过程闭环系统的影响时, 可采用下面的线性控制器方程求取控制输入$y^{*}(k)$.
$ \begin{array}{l} G({z^{ - 1}})\Delta r(k) + H({z^{ - 1}})B\Delta {y^*}(k) = \\ \;\;\;\;\;\;R({z^{ - 1}}){r^*}(k + 2) - T({z^{ - 1}})r(k) \end{array} $
(13) 2.2 线性鲁棒自适应控制
运行过程的动态模型往往是未知的, 因此需要采用自适应方法在线获得控制器参数, 当组成$A(z^{-1})$, $B$, $C(z^{-1})$的参数阵未知时, 式(10)可看作控制器参数辨识方程, 为此记$\phi(k)=T(z^{-1})r(k)$, $G(z^{-1})$ $=G_0+G_1(z^{-1})$, $Q(z^{-1})=H(z^{-1})B=$ $Q_0 +Q_1(z^{-1})$, 并定义数据向量和参数矩阵分别为$\varphi(k)$ $=[\Delta r^{\rm T}(k), \Delta r^{\rm T}(k-1), \Delta {y^{*}}^{\rm T}(k)$, $\Delta {y^{*}}^{\rm T}(k-1)]^{\rm T}$和$\theta=[G_0, G_1, Q_0, Q_1]^{\rm T}$, 则控制器参数辨识方程(10)可以写为
$ \phi (k + 2) = {\theta ^{\rm{T}}}\varphi (k) + \phi (k) + \rho (k) $
(14) 线性控制器方程(13)可重新写为
$ \theta^{\rm T}\varphi(k)=R(z^{-1})r^{*}(k+2)-T(z^{-1})r(k) $
(15) 对于未知的参数矩阵$\theta$, 采用带死区的递推最小二乘方法进行在线辨识.
$ \hat \theta (k) = {\rm{proj}}\{ {\hat \theta ^ + }(k)\} $
(16a) $ \begin{array}{l} {{\hat \theta }^ + }(k) = \hat \theta (k - 2){\mkern 1mu} + \\ \;\;\;\;\;\;\;\;\;\;\;\frac{{\lambda (k)P(k - 2)\varphi (k - 2){e^{\rm{T}}}(k)}}{{1 + {\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)}} \end{array} $
(16b) $ \begin{array}{l} P(k) = P(k - 2) - \\ \;\;\;\;\;\;\;\;\;\;\frac{{\lambda (k)P(k - 2)\varphi (k - 2){\varphi ^{\rm{T}}}(k - 2)P(k - 2)}}{{1 + {\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)}} \end{array} $
(16c) $ e(k) = \phi (k) - \hat \phi (k) $
(16d) $ \hat \phi (k) = {\hat \theta ^{\rm{T}}}(k - 2)\varphi (k - 2) + \phi (k - 2) $
(16e) $ \lambda \left( k \right) = \left\{ \begin{array}{l} \frac{1}{2}, \;\;如果\left\| {e\left( k \right)} \right\|＞2E\\ 0, \;\;否则 \end{array} \right. $
(16f) $ {\rm{proj}}\left\{ {{{\hat \theta }^ + }\left( k \right)} \right\} = \left\{ \begin{array}{l} {{\hat \theta }^ + }\left( k \right), \;\;\;\;\;\;\;\;\;\;\;\;\;\hat Q_0^ + \left( k \right)非奇异\\ {\left[ { \ldots , {Q_{\min }}, \ldots } \right]^{\rm{T}}}, \;\;\;否则 \end{array} \right. $
(16g) 其中, $[\varphi(0), \widehat{\theta}(0), P(0)]$为初始条件, $P(0)>0$为正定矩阵, $E$为$\rho(k)$的已知上界, $\widehat{\theta}(k)= [\widehat{G}_0(k)$, $\widehat{G}_1(k)$, $\widehat{Q}_0(k), \widehat{Q}_1(k)]^{\rm T}$为$k$时刻未知参数矩阵$\theta$的估计, $\widehat{\theta}^{+}(k)=[\widehat{G}_0(k), \widehat{G}_1(k), \widehat{Q}_0^{+}(k), \widehat{Q}_1(k)]^{\rm T}$, ${\rm proj}\{\cdot\}$为一投影算子, 满足式(16g).
由式(15)及确定性等价原则可知, 线性鲁棒自适应控制器为
$ {\hat \theta ^{\rm{T}}}(k)\varphi (k) = R({z^{ - 1}}){r^*}(k + 2) - T({z^{ - 1}})r(k) $
(17) 2.3 线性鲁棒自适应控制系统稳定性和性能
引理 1. 定义函数
$ V(k) = {\rm{tr}}\left[ {{{\widetilde \theta }^{\rm{T}}}(k){P^{ - 1}}(K)\widetilde \theta (k)} \right] $
则带死区的递推最小二乘辨识算法(16)具有如下性质:
1)
$ \begin{array}{l} V(k) - V(k - 2) \le \\ \;\;\;\; - \frac{{3\lambda (k){{\left\| {e(k)} \right\|}^2}}}{{8[1 + {\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)]}} - \\ \;\;\;\;\frac{{\lambda (k)[{{\left\| {e(k)} \right\|}^2} - 4{E^2}]}}{{4\{ 1 + [1 - \lambda (k)]{\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)\} }} \end{array} $
2)
$ \mathop {\lim }\limits_{k \to \infty } \left\| {\hat \theta (k) - \hat \theta (k - 2)} \right\| = 0 $
证明. 见附录A.
定理 1. 运行过程动态模型(5)或(6)满足假设1, 则当线性鲁棒自适应控制算法(16)应用于式(6)时, 闭环运行过程全局李雅普诺夫稳定, 并且广义跟踪误差满足${\lim _{k \to \infty }}\lambda (k)[{\left\| {\bar e(k)} \right\|^2} - 4{E^2}] = 0$, 其中, $\bar e(k): = T({z^{ - 1}})r(k) - R({z^{ - 1}}){r^*}(k)$.
证明. 由引理1的1)可知,
$ \mathop {\lim }\limits_{k \to \infty } \frac{{\lambda (k)[{{\left\| {e(k)} \right\|}^2} - 4{E^2}]}}{{4\{ 1 + [1 - \lambda (k)]{\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)\} }} = 0 $
(18) 由于$\overline{e}(k):=T(z^{-1})r(k)-R(z^{-1})r^*{(k)}$及$T(z^{-1})$的稳定性, 存在正常数$c_1$, $c_2$, $c_3$, $c_4$满足
$ \begin{array}{l} |{r_i}(k)| \le {c_1} + {c_2}\mathop {\max }\limits_{_{\scriptstyle0 \le \tau \le t\atop \scriptstyle1 \le i \le m}} |{{\bar e}_i}(\tau )|, \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;i = 1, 2, \cdots , m \end{array} $
(19) $ \begin{array}{l} |y_i^*(k - 2)| \le {c_3} + {c_4}\mathop {\max }\limits_{_{\scriptstyle0 \le \tau \le t\atop \scriptstyle1 \le i \le m}} |{r_i}(\tau )|, \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;i = 1, 2, \cdots , m \end{array} $
(20) 令
$ \begin{array}{l} X(k - 2) = \\ \;\;\;{[{r^{\rm{T}}}(k - 2), {r^{\rm{T}}}(k - 3), {y^*}^{\rm{T}}(k - 2), {y^*}^{\rm{T}}(k - 3)]^{\rm{T}}} \end{array} $
则存在正常数$c_5$, $c_6$满足
$ \left\| {X(k - 2)} \right\| \le {c_5} + {c_6}\mathop {\max }\limits_{0 \le \tau \le t} \left\| {\bar e(\tau )} \right\| $
(21) 由式(16d)可知,
$ e(k)=T(z^{-1})r(k)-R(z^{-1})r^{*}(k)=\overline{e}(k) $
(22) 因此, 由式(21)和式(22)可知
$ \left\| {X(k - 2)} \right\| \le {c_5} + {c_6}\mathop {\max }\limits_{0 \le \tau \le t} \left\| {e(\tau )} \right\| $
(23) 由式(23)可知, 单独采用线性鲁棒自适应控制算法时, 系统输入和输出信号的有界性由$e(k)$的有界性决定.下面假设$e(k)$无界.由式(16f)可知, 存在时刻$K_0>0$, 当$k>K_0$时, $\left\| {\mathit{e}(\mathit{k})} \right\| > {\rm{2}}E$并且$\lambda(k)$ $=$ $1/2$, 即式(18)的分子是一个正实序列.于是存在一单调递增序列$\left\| {\mathit{e}({\mathit{k}_n})} \right\|$, 使得
$ {\lim _{k \to \infty }}\left\| {e({k_n})} \right\| = \infty $
由式(23)可知
$ \begin{array}{l} \frac{{\lambda ({k_n})[{{\left\| {e({k_n})} \right\|}^2} - 4{E^2}]}}{{41 + [1 - \lambda ({k_n})]{\varphi ^{\rm{T}}}({k_n} - 2)P({k_n} - 2)\varphi ({k_n} - 2)}} \ge \\ \frac{{\lambda ({k_n})[{{\left\| {e({k_n})} \right\|}^2} - 4{E^2}]}}{{81 + [1 - \lambda ({k_n})][{{({c_5} + {c_6}\left\| {e({k_n})} \right\|)}^2}]\left\| {P({k_n} - 2)} \right\|}} \end{array} $
由于$\left\| {\mathit{P}({\mathit{k}_n} - {\rm{2}})} \right\|$为递减序列, 因此,
$ \left\| {P({k_n} - 2)} \right\| \le \left\| {P(0)} \right\| $
因此
$ \begin{array}{l} \mathop {\lim }\limits_{k \to \infty } \frac{{\lambda ({k_n})[{{\left\| {e({k_n})} \right\|}^2} - 4{E^2}]}}{{41 + [1 - \lambda ({k_n})]{\varphi ^{\rm{T}}}({k_n} - 2)P({k_n} - 2)\varphi ({k_n} - 2)}} \ge \\ \;\;\;\;\;\frac{1}{{8{c_6}\left\| {P(0)} \right\|}} > 0 \end{array} $
这与式(18)矛盾.故假设不成立, $e(k)$有界, 从而采用线性鲁棒自适应控制算法时, 闭环系统BIBO稳定.
注释 1. 单独使用线性鲁棒自适应控制器能够保证闭环运行过程全局李亚普洛夫稳定, 但是无法使运行过程具有良好的跟踪性能.为了改善运行过程的跟踪性能, 同时不影响其稳定性, 我们将线性鲁棒自适应控制器、基于神经网络的非线性控制器以及切换机制相结合, 提出一种新的多模型自适应控制方法.
3. 多模型自适应控制
3.1 多模型自适应控制算法
下面考虑多模型自适应控制问题, 为将问题简化, 采用两个模型, 多模型切换系统如图 2所示.
图 2中, 线性估计模型定义为
$ \widehat{\phi}_1(k)=\widehat{\theta}_1^{\rm T}(k-2)\varphi(k-2)+\phi(k-2) $
(24) 其中, $\widehat{\theta}_1(k)=[\widehat{G}_{1, 0}(k), \widehat{G}_{1, 1}(k), \widehat{Q}_{1, 0}(k), \widehat{Q}_{1, 1}(k)]^{\rm T}$为$k$时刻的基于线性模型(24)的估计, 采用式(16)在线辨识, 线性辨识误差为
$ {e_1}(k) = \phi (k) - {\hat \phi _1}(k) $
(25) 通过下式计算控制输入$y^*(k)$, 作为基于线性模型的控制器$y^*_1(k)$.
$ \hat \theta _1^{\rm{T}}(k)\varphi (k) = R({z^{ - 1}}){r^*}(k + 2) - T({z^{ - 1}})r(k) $
(26) 非线性估计模型定义为
$ {\hat \phi _2}(k) = \hat \theta _2^{\rm{T}}(k - 2)\varphi (k - 2) + \phi (k - 2) + \hat \rho (k - 2) $
(27) 其中,$\widehat{\theta}_2(k)=[\widehat{G}_{2, 0}(k), \widehat{G}_{2, 1}(k), \widehat{Q}_{2, 0}(k), \widehat{Q}_{2, 1}(k)]^{\rm T}$为$k$时刻$\theta$的基于非线性模型(27)的估计; $\widehat{\rho}(k)$为$\rho^*(k)$的神经网络估计, 其中, $\rho^*(k):=\Delta\phi(k+2)-\widehat{\theta}_2(k)^{\rm T}\varphi(k)$, 即
$ \widehat{\rho}(k)=NN[\widehat{W}(k), \varphi(k)] $
(28) 其中, $NN[\cdot]$表示神经网络结构; $\varphi(k)$为神经网络的输入向量; $\widehat{W}(k)$为$k$时刻理想权阵$W^*$的估计.与文献[12]类似, 除了要求参数阵的估计$\widehat{\theta}_2(k)$和权阵的估计$\widehat{W}(k)$有界, $\widehat{Q}_{2, 0}(k)$非奇异, 并未对$\widehat{\theta}_2(k)$的辨识算法和神经网络的结构以及权阵校正算法进行任何限制, 即
$ \widehat{\theta}_2(k), \widehat{W}(k);~~\widehat{Q}_{2, 0}(k)~\text{非奇异}, ~ \forall k $
(29) 非线性辨识误差为
$ e_2(k)=\phi(k)-\widehat{\phi}_2(k) $
(30) 因此, 根据式(10)和确定性等价原则, 通过下式计算控制输入$y^*(k)$, 作为基于非线性模型的控制器$y^*_2(k)$.
$ \begin{array}{l} {{\hat \theta }_2}(k)\varphi (k) + \hat \rho (k) = R({z^{ - 1}}){r^*}(k + 2){\mkern 1mu} - \\ \;\;\;\;\;\;\;\;T({z^{ - 1}})r(k) \end{array} $
(31) 切换准则为
$ \begin{array}{l} {J_j}\left( k \right) = \sum\limits_{i = 2}^k {\frac{{{\lambda _j}(k)[{{\left\| {{e_j}(k)} \right\|}^2} - 4{E^2}]}}{{4\{ 1 + [1 - {\lambda _j}(k)]{\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)\} }} + } \\ {c_0}\sum\limits_{l = k - N - 1}^k {\left( {\frac{1}{2} - {\lambda _j}(l){{\left\| {{e_j}(l)} \right\|}^2}} \right)} \end{array} $
(32) $ {\lambda _j}\left( k \right) = \left\{ \begin{array}{l} \frac{1}{2}, \;\;若\left\| {{e_j}(k)} \right\|＞2E\\ 0, \;\;否则 \end{array} \right. $
(33) 其中, $N$是一个正整数, $c_0$是一个大于等于0的预先确定的常数.
每一时刻$k$, 比较$J_1(k)$和$J_2(k)$, 求出最小的$J^*(k)$, 选择与$J^*(k)$对应的自适应控制器$y_i^*(k)$, 并将其应用于运行过程.
3.2 多模型自适应控制系统稳定性和性能
定理 2. 运行过程动态模型(6)满足假设1, 则当基于多模型自适应控制算法(24)~ (33)用于运行过程(6)时, 闭环切换系统BIBO稳定.此外, 对于任意给定的正数$\varepsilon$, 存在时刻$K$, 当$k>K$时, 系统的广义跟踪误差满足$\left\| {\bar e(k)} \right\| \le 2E + \varepsilon $.
证明. 由引理1可知,
$ \mathop {\lim }\limits_{k \to \infty } \frac{{{\lambda _1}(k)[{{\left\| {{e_1}(k)} \right\|}^2} - 4{E^2}]}}{{4\{ 1 + [1 - {\lambda _1}(k)]{\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)\} }} = 0 $
(34) 由式(24)和式(25)可知,
$ \begin{array}{l} {e_1}(k) = \phi (k) - {{\hat \phi }_1}(k) = \\ \;\;\;\;\;\;\;\;\;\;\Delta \phi (k) - \hat \theta _1^{\rm{T}}(k - 2)\varphi (k - 2) = \\ \;\;\;\;\;\;\;\;\;T({z^{ - 1}})r(k) - T({z^{ - 1}})r(k - 2) - \\ \;\;\;\;\;\;\;\;\hat \theta _1^{\rm{T}}(k - 2)\varphi (k - 2) \end{array} $
(35) 由式(27)和式(30)可知
$ \begin{array}{l} {e_2}(k) = \phi (k) - {{\hat \phi }_2}(k) = \\ \;\;\;\;\;\;\Delta \phi (k) - \hat \theta _2^{\rm{T}}(k - 2)\varphi (k - 2) - \hat \rho (k - 2) = \\ \;\;\;\;\;\;T({z^{ - 1}})r(k) - T({z^{ - 1}})r(k - 2) - \\ \;\;\;\;\;\;\hat \theta _2^{\rm{T}}(k - 2)\varphi (k - 2) - \hat \rho (k - 2) \end{array} $
(36) 于是, 根据确定性等价原则, 每一时刻
$ \bar{e}(k)=e_1(k)~ \mbox{或}~e_2(k) $
(37) 由于每一时刻, 系统辨识误差$e(k)=e_1(k)$或$e_2(k)$, 故由式(21)可知, 存在正常数$c_7$, $c_8$满足
$ \left\| {X(k - 2)} \right\| \le {c_7} + {c_8}\mathop {\max }\limits_{0 \le \tau \le k} \left\| {e(\tau )} \right\| $
(38) 由式(33)可知, 切换函数$J_j(k)$ $(j=1, 2)$的第2项是有界的.因此由引理1可知, $J_1(k)$有界.对于$J_2(k)$有两种情况.
1) $J_2(k)$无界.由于$J_1(k)$有界, 因此存在时刻$K_1$使得当$k\geq K_1$时有$J_1(k)\leq J_2(k)$.故根据切换机制, 当$k\geq K_1+1$时, 系统辨识误差$e(k)=e_1(k)$满足
$ \mathop {\lim }\limits_{k \to \infty } \frac{{{\lambda _1}(k)[{{\left\| {e(k)} \right\|}^2} - 4{E^2}]}}{{4\{ 1 + [1 - {\lambda _1}(k)]{\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)\} }} = 0 $
(39) 其中,
$ \lambda \left( k \right) = \left\{ \begin{array}{l} \frac{1}{2}, \;\;若\left\| {e(k)} \right\|＞2E\\ 0, \;\;否则 \end{array} \right. $
2) $J_2(k)$有界.由切换准则式(32)可知, $e_2(k)$满足
$ \mathop {\lim }\limits_{k \to \infty } \frac{{{\lambda _2}(k)[{{\left\| {{e_2}(k)} \right\|}^2} - 4{E^2}]}}{{4\{ 1 + [1 - {\lambda _2}(k)]{\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)\} }} \to 0 $
故系统辨识误差$e(k)=e_1(k)$或$e_2(k)$满足式(39).
其余部分的证明类似于定理1.
由式(39)和$X(k-2)$的有界性可知,
$ \mathop {\lim }\limits_{k \to \infty } \lambda (k)\left[ {{{\left\| {e(k)} \right\|}^2} - 4{E^2}} \right] $
即, 对任意小的正数$\varepsilon$, 存在时刻$K$, 当$k>K$时,
$ \left\| {e(k)} \right\| \le 2E + \varepsilon $
(40) 注释2. 由式(39)可知, 非线性辨识误差
$ {e_2}(k) = {\rho ^*}(k - 2) - \hat \rho (k - 2) $
(41) 适当选择神经网络结构和参数, 可以保证$\left\| {{\rho ^*}(k} \right.$ $-$ $\left. {2) - \hat \rho (k - 2)} \right\|＜\varepsilon $.因此若运行过程选择非线性自适应控制器$y_2^*(k)$作为输入信号, 则由式(35)和式(36)可知, 广义跟踪误差$\left\| {\bar e(k)} \right\|＜\varepsilon $满足.
4. 仿真实验
为验证本文所提方法的有效性, 首先考虑如下底层被控对象模型
$ \begin{array}{l} x(k + 1) = \left( {\begin{array}{*{20}{c}} {1.5}&6\\ 6&4 \end{array}} \right)x(k) + \\ \;\;\;\;\;\;\;\;\;\left( {\begin{array}{*{20}{c}} { - 4.2623}&{ - 3.8254}\\ {8.3534}&{6.1711} \end{array}} \right)u(k)\\ y(k) = \left( {\begin{array}{*{20}{c}} {0.1546}&{ - 0.012}\\ { - 0.0099}&{0.2281} \end{array}} \right)x(k) \end{array} $
(42) 其中, $x=[x_1, x_2]^{\rm T}$, $y=[y_1, y_2]^{\rm T}$, $u=[u_1, u_2]^{\rm T}$.为使底层闭环系统稳定, 并实现稳态跟踪, 选择如下极点配置控制器
$ u(t)=-Kx(t)+Ly^*(t) $
(43) 其中, $y^*(t)=[y_1^*, y_2^*]^{\rm T}$为底层回路设定值, 由后面的运行控制给出.
$ \begin{array}{l} K = \left( {\begin{array}{*{20}{c}} {7.1487}&{15.3085}\\ { - 8.8017}&{ - 19.0044} \end{array}} \right)\\ \;\;L = \left( {\begin{array}{*{20}{c}} { - 14.5}&{30.25}\\ { - 20.6}&{ - 35.6} \end{array}} \right) \end{array} $
(44) 假设运行层动态模型为
$ \begin{array}{l} r(k + 1) = \left( {\begin{array}{*{20}{c}} 1&0\\ 0&1 \end{array}} \right)r(k) + \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\left( {\begin{array}{*{20}{c}} { - 1.2893}&{ - 0.0678}\\ {0.2798}&{ - 4.3693} \end{array}} \right)y(k) + \nu (k) \end{array} $
(45) 其中,
$ r(k) = {[{r_1}, {r_2}]^{\rm{T}}} $
$ \begin{array}{l} \nu (k) = \left( {\begin{array}{*{20}{c}} {{\nu _1}}\\ {{\nu _2}} \end{array}} \right) = 0.01 \times \\ \;\;\;\;\;\;\;\;\;\;\left( {\begin{array}{*{20}{c}} {{\rm{sin}}({\rm{1 + }}y_{\rm{1}}^{{\rm{*2}}}(k{\rm{ - 1}}){\rm{ + }}r_{\rm{1}}^{\rm{2}}(k{\rm{ - 1}}){\rm{ + }}}\\ {r_2^2(k) - \frac{{{r_1}(k - 1) + {r_2}(k)}}{{1 + y_1^{*2}(k - 1) + r_1^2(k - 1) + r_2^2(k)}})}\\ {{\rm{sin}}({\rm{1 + }}y_{\rm{2}}^{{\rm{*2}}}(k{\rm{ - 1}}){\rm{ + }}r_{\rm{1}}^{\rm{2}}(k{\rm{ - 1}}){\mkern 1mu} {\rm{ + }}}\\ {r_2^2(k) - \frac{{{r_1}(k) + {r_2}(k - 1)}}{{1 + y_2^{*2}(k - 1) + r_1^2(k) + r_2^2(k - 1)}})} \end{array}} \right) \end{array} $
则由式(5)可知, 运行过程广义对象模型为
$ \begin{array}{l} x(k + 1) = \left( {\begin{array}{*{20}{c}} { - 1.7}&{ - 1.45}\\ {0.6}&{ - 6.6} \end{array}} \right)x(k) + \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\left( {\begin{array}{*{20}{c}} {17}&{7.25}\\ { - 6}&{33} \end{array}} \right){y^*}(k)\\ r(k + 1) = \left( {\begin{array}{*{20}{c}} 1&0\\ 0&1 \end{array}} \right)r(k) + \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\left( {\begin{array}{*{20}{c}} { - 0.2}&0\\ 0&{ - 1} \end{array}} \right)x(k) + \nu (k) \end{array} $
(46) 选择加权多项式矩阵
$ \begin{array}{l} T({z^{ - 1}}) = R({z^{ - 1}}) = \\ \;\;\;\;\;\;\;\;\;\;\;\;\left( \begin{array}{l} - 5 - 0.1{z^{ - 1}}\;\;\;\;\;\;\;\;\;0\\ \;\;\;\;\;\;0\;\;\;\;\;\;\;\;\; - 1 - 0.1{z^{ - 1}} \end{array} \right) \end{array} $
运行指标目标值为
$ {r^*}(k) = \left( \begin{array}{l} \;\;\;\;\;\;\;\;\;\;\;\;\;0.5\\ 0.5{\rm{sign}}\left( {\cos \left( {k \times \frac{\pi }{{50}}} \right)} \right) \end{array} \right) $
易知, 控制器真实参数阵为
$ \theta = \left( {{\theta _0}\;\;\;\;\;{\theta _1}} \right) $
其中,
$ \begin{array}{l} {\theta _0} = ( - 6.53, - 10.556, 1.43, 10.556, 17, 7.25, \\ \;\;\;\;\;\;\;\; - 2.86, - 52.78{)^{\rm{T}}} \end{array} $
$ \begin{array}{l} {\theta _1} = (21.6, - 36.53, - 21.6, 35.43, - 6, 33, \\ \;\;\;\;\;\;\;43.2, - 177.15{)^{\rm{T}}} \end{array} $
本仿真实验中, 我们假设它是未知的, 并根据先验知识选择待辨识控制器初始参数阵为
$ \begin{array}{l} \\ \begin{array}{*{20}{l}} {\hat \theta (0) = {{\left( {\begin{array}{*{20}{c}} { - 4}&{ - 7}&0&4&{19}&3&{ - 1}&{ - 35}\\ {16}&{ - 31}&{ - 15}&{30}&{ - 2}&{29}&{35}&{ - 160} \end{array}} \right)}^{\rm{T}}}} \end{array} \end{array} $
选择单隐层线性输出的静态BP网对$\rho^*(k)$进行估计, 其隐元数为20, 学习率为0.1;选择$c=1$, $N$ $=$ $2$.
图 3为单独采用线性鲁棒自适应控制方法时运行过程的运行指标目标值$r^*(k)$和运行过程输出$r(k)$.从图 3可以看出, 虽然该控制器可以使运行过程稳定, 但跟踪效果很差. 图 4为采用本文所提的多模型自适应控制方法时运行过程的运行指标目标值$r^*(k)$、运行过程输出$r(k)$和运行过程控制输入, 即底层设定值$y^*(k)$.与图 3相比, 图 4中的跟踪效果明显改善. 图 5为$\widehat{\theta}_1(k)$中16个参数的在线变化曲线. 图 6为底层极点配置控制系统的跟踪曲线.为进行比较, 仍以上述矩阵为控制器初始参数阵, 采用文献[12]提出的基于投影算法的多模型自适应控制方法对运行过程进行仿真, 运行过程跟踪结果如图 7所示.相应的$\widehat{\theta}_1(k)$中各参数的在线变化曲线如图 8所示.由图 4和图 7可知, 采用本文提出的基于最小二乘算法的多模型自适应控制方法时, 即使控制器初始参数阵离控制器真实参数阵较远, 仍具有有良好的跟踪效果.两相比较, 基于投影算法的多模型自适应控制方法对初始参数阵非常灵敏, 当初始参数阵远离控制器真实参数阵时, 控制效果较差.比较图 5和图 8可以看出, 最小二乘算法与投影算法相比具有更快的收敛速度.
图 3 采用基于递推最小二乘算法的线性鲁棒自适应控制方法时, 运行过程的输出及运行指标目标值Fig. 3 Outputs of the operation process and theirs operation targets when the linear robust adaptive control method based on recursive least square algorithm is used
图 4 采用基于递推最小二乘算法的多模型自适应控制方法时, 运行过程的输出、运行指标目标值及控制输入Fig. 4 Outputs of the operation process, theirs operation targets and control inputs when the proposed multi-model adaptive control method based on recursive least square algorithm is used
图 5 采用基于递推最小二乘算法的多模型自适应控制方法时, $\widehat{\theta}_1(k)$中16个参数的在线变化曲线Fig. 5 Online curves of 16 parameters in $\widehat{\theta}_1(k)$ when the proposed multi-model adaptive control method based on recursive least square algorithm is used
图 6 底层极点配置控制系统的跟踪曲线Fig. 6 Tracking curves of the underlying pole assignment control system
图 7 采用基于投影算法的多模型自适应控制方法时, 运行过程的输出和运行指标目标值Fig. 7 Outputs of the operation process and theirs operation targets when the multi-model adaptive control method based on projection algorithm is used
图 8 采用基于投影算法的多模型自适应控制方法时, $\widehat{\theta}_1(k)$中16个参数的在线变化曲线Fig. 8 Online curves of 16 parameters in $\widehat{\theta}_1(k)$ when the multi-model adaptive control method based on projection algorithm is used5. 结论
工业运行过程应考虑底层跟踪设定值的动态误差对整个运行过程优化和控制的影响.现有的工业运行控制方法假设运行层为已知的线性静态模型.本文针对一类运行层为未知线性动态模型的工业运行过程, 提出了一种基于递推最小二乘算法的多模型自适应控制方法.通过理论分析和仿真比较, 验证了与现有的多模型自适应控制方法相比, 本文提出方法可以应用于工业运行过程, 并具有良好的跟踪效果.
工业过程运行控制是近年来控制领域比较热门的研究方向之一, 现有的方法针对的被控对象主要是线性的, 并且主要集中在理论研究上.在实际的工业过程控制中, 非线性动态无可避免, 当两层结构中的被控对象都是非线性时, 如何设计控制器, 如何将理论的研究成果进行实际应用具有一定的挑战.
附录A. 引理1的证明
证明. 当$\widehat{Q}_0^+(k)$非奇异时, $\widehat{\theta}(k)$与$\widehat{\theta}^+(k)$相等.由式(16d)和式(16e)可知,
$ e(k) = [{\theta ^{\rm{T}}} - {\hat \theta ^{\rm{T}}}(k - 1)]\varphi (k - 2) + \rho (k - 2) $
(A1) 令
$ L(k)=\frac{P(k-2)\varphi(k-2)}{1+\varphi^{\rm T}(k-2)P(k-2)\varphi(k-2)} $
(A2) 由式(16b)、式(16c)以及式(A1)和式(A2)可知,
$ \begin{array}{l} P(k) = P(k - 2) - \lambda (k)L(k){\varphi ^{\rm{T}}}(k - 2)P(k - 2) \Rightarrow \\ P(k) = [I - \lambda (k)L(k){\varphi ^{\rm{T}}}(k - 2)]P(k - 2) \Rightarrow \\ P(k){P^{ - 1}}(k - 2) = I - \lambda (k)L(k){\varphi ^{\rm{T}}}(k - 2) \end{array} $
(A3) 和
$ \begin{array}{l} \widetilde \theta (k) = \widetilde \theta (k - 2) + \lambda (k)L(k){e^{\rm{T}}}(k) = \\ \;\;\;\;\;\;\;\;\;\;[I - \lambda (k)L(k){\varphi ^{\rm{T}}}(k - 2)]\widetilde \theta (k - 2) + \\ \;\;\;\;\;\;\;\;\;\;\lambda (k)L(k){\rho ^{\rm{T}}}(k - 2) = \\ \;\;\;\;\;\;\;\;\;\;P(k){P^{ - 1}}(k - 2)\widetilde \theta (k - 2) + \\ \;\;\;\;\;\;\;\;\;\lambda (k)L(k){\rho ^{\rm{T}}}(k - 2) \end{array} $
(A4) 由式(A4)可知,
$ \begin{array}{l} {P^{ - 1}}(k)\widetilde \theta (k) - {P^{ - 1}}(k - 2)\widetilde \theta (k - 2) = \\ \;\;\;\;\;\;\lambda (k){P^{ - 1}}(k)L(k){\rho ^{\rm{T}}}(k - 2) \end{array} $
(A5) 由于$\varphi^{\rm T}(k-2)P(k-2)\varphi(k-2)\times I=\varphi(k-2)\varphi^{\rm T}(k-2)P(k-2)$, 因此由式(16c)可知
$ \begin{array}{l} \frac{{{P^{ - 1}}(k)P(k - 2)\varphi (k - 2)}}{{1 + {\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)}} = \\ \;\;\;\;\frac{{\varphi (k - 2)}}{{1 + [1 - \lambda (k)]{\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)}} \end{array} $
(A6) 令$Q(k):=\varphi^{\rm T}(k-2)P(k-2)\varphi(k-2)$, 则
$ \begin{array}{l} V(k) - V(k - 2) = \\ \;\;\;\;\;\;\;{\rm{tr}}\{ - \frac{{\lambda (k)[e(k){e^{\rm{T}}}(k) - 4\rho (k - 2){\rho ^{\rm{T}}}(k - 2)]}}{{4[1 + [1 - \lambda (k)]Q(k)]}} - \\ \;\;\;\;\;\;\;\left. {\frac{{3\lambda (k)e(k){e^{\rm{T}}}(k)[1 + Q(k)[1 - \frac{{4\lambda (k)}}{3}]]}}{{4[1 + [1 - \lambda (k)]Q(k)][1 + Q(k)]}}} \right\} \end{array} $
(A7) 由于
$ \frac{1+Q(k)\left[1-\frac{4\lambda(k)}{3}\right]}{1+[1-\lambda(k)]Q(k)}\geq \frac{1}{2} $
(A8) 因此
$ \begin{array}{l} V(k) - V(k - 2) \le \\ \;\;\;\;\;{\rm{tr}}\{ - \frac{{\lambda (k)[e(k){e^{\rm{T}}}(k){\rm{ - 4}}\rho (k{\rm{ - 2}}){\rho ^{\rm{T}}}(k{\rm{ - 2}})]}}{{{\rm{4}}[{\rm{1 + }}[{\rm{1 - }}\lambda (k)]Q(k)]}} - \\ \;\;\;\;\;\frac{{3\lambda (k)e(k){e^{\rm{T}}}(k)}}{{8[1 + Q(k)]}}\} \le \\ \;\;\;\;\; - \frac{{\lambda (k)[{{\left\| {e(k)} \right\|}^2} - 4{E^2}]}}{{4[1 + [1 - \lambda (k)]Q(k)]}} - \frac{{3\lambda (k){{\left\| {e(k)} \right\|}^2}}}{{8[1 + Q(k)]}} \end{array} $
(A9) 因此, 引理1中1)得证.由式(16b)可知,
$ \begin{array}{l} {\varphi ^{\rm{T}}}(k - 2)[\hat \theta (k) - \hat \theta (k - 2)] = \\ \;\;\;\;\;\;\frac{{\lambda (k){\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2){e^{\rm{T}}}(k)}}{{1 + {\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)}} \end{array} $
(A10) 因此
$ \begin{array}{l} {\varphi ^{\rm{T}}}(k - 2)[\hat \theta (k) - \hat \theta (k - 2)][\hat \theta (k) - \\ \;\;\;\;\;\;\;\hat \theta (k - 2){]^{\rm{T}}}\varphi (k - 2) \le \\ \;\;\;\;\;\;\;\frac{{\lambda (k){\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2){e^{\rm{T}}}(k)e(k)}}{{1 + {\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)}} \end{array} $
(A11) 故
$ {\left\| {\hat \theta (k) - \hat \theta (k - 2)} \right\|^2} \le \frac{{\lambda (k)\left\| {P(k - 2)} \right\|{e^{\rm{T}}}(k)e(k)}}{{1 + {\varphi ^{\rm{T}}}(k - 2)P(k - 2)\varphi (k - 2)}} $
(A12) 由引理1中的1)及$\| P(k-2)\|$的有界性, 可知2)成立.当$\widehat{Q}_0^+(k)$非奇异时, 记.由式(16g)可知${\left\| {{\theta ^ + }(k) - \theta } \right\|^2}$, 因此$V^+(k)\leq V(k)$, 故引理1\linebreak仍旧成立.
-
图 1 MIT-Manus与Armeo Power上肢康复机器人
Fig. 1 MIT-Manus and Armeo Power upper limb rehabilitation robots
图 5 基于TMR手术和表面肌电信号控制的动力型假肢
Fig. 5 Power prothesis controlled by sEMG signal via targeted muscle reinnervation
-
[1] 侯增广, 赵新刚, 程龙, 王启宁, 王卫群.康复机器人与智能辅助系统的研究进展.自动化学报, 2016, 42(12):1765-1779 http://www.aas.net.cn/CN/abstract/abstract18966.shtmlHou Zeng-Guang, Zhao Xin-Gang, Cheng Long, Wang Qi-Ning, Wang Wei-Qun. Recent advances in rehabilitation robots and intelligent assistance systems. Acta Automatica Sinica, 2016, 42(12):1765-1779 http://www.aas.net.cn/CN/abstract/abstract18966.shtml [2] Losey D P, McDonald C G, Battaglia E, O'Malley M K. A review of intent detection, arbitration, and communication aspects of shared control for physical human-robot interaction. Applied Mechanics Reviews, 2018, 70(1):Article No. 010804 http://adsabs.harvard.edu/abs/2018ApMRv..70a0804. [3] Marchal-Crespo L, Reinkensmeyer D J. Review of control strategies for robotic movement training after neurologic injury. Journal of Neuroengineering and Rehabilitation, 2009, 6:Article No. 20 doi: 10.1186/1743-0003-6-20 [4] De Santis A, Siciliano B, De Luca A, Bicchi A. An atlas of physical human-robot interaction. Mechanism and Machine Theory, 2008, 43(3):253-270 doi: 10.1016/j.mechmachtheory.2007.03.003 [5] Maciejasz P, Eschweiler J, Gerlach-Hahn K, Jansen-Troy A, Leonhardt S. A survey on robotic devices for upper limb rehabilitation. Journal of Neuroengineering and Rehabilitation, 2014, 11:Article No. 3 doi: 10.1186/1743-0003-11-3 [6] Hogan N, Krebs H I, Charnnarong J, Srikrishna P, Sharon A. Mit-manus: a workstation for manual therapy and training. Ⅰ. In: Proceedings of the 1992 IEEE International Workshop on Robot and Human Communication. Tokyo, Japan: IEEE, 1992. 161-165 http://www.mendeley.com/catalog/mitmanus-workstation-manual-therapy-training-i/ [7] Kwakkel G, Kollen B J, Krebs H I. Effects of robot-assisted therapy on upper limb recovery after stroke:a systematic review. Neurorehabilitation and Neural Repair, 2008, 22(2):111-121 doi: 10.1177/1545968307305457 [8] Nef T, Mihelj M, Colombo G, Riener R. Armin-robot for rehabilitation of the upper extremities. In: Proceedings of the 2006 IEEE International Conference on Robotics and Automation. Orlando, FL, USA: IEEE, 2006. 3152-3157 http://www.mendeley.com/research/armin-robot-rehabilitation-upper-extremities/ [9] Klamroth-Marganska V, Blanco J, Campen K, Curt A, Dietz V, Ettlin T, et al. Three-dimensional, task-specific robot therapy of the arm after stroke:a multicentre, parallel-group randomised trial. The Lancet Neurology, 2014, 13(2):159-166 http://www.sciencedirect.com/science/article/pii/S1474442213703053 [10] Jezernik S, Colombo G, Keller T, Frueh H, Morari M. Robotic orthosis lokomat:a rehabilitation and research tool. Neuromodulation:Technology at the Neural Interface, 2003, 6(2):108-115 doi: 10.1046/j.1525-1403.2003.03017.x [11] Duschau-Wicke A, von Zitzewitz J, Caprez A, Lunenburger L, Riener R. Path control:a method for patient-cooperative robot-aided gait rehabilitation. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2010, 18(1):38-48 doi: 10.1109/TNSRE.2009.2033061 [12] 明东, 蒋晟龙, 王忠鹏, 綦宏志, 万柏坤.基于人机信息交互的助行外骨骼机器人技术进展.自动化学报, 2017, 43(7):1089-1100 http://www.aas.net.cn/CN/abstract/abstract19085.shtmlMing Dong, Jiang Sheng-Long, Wang Zhong-Peng, Qi Hong-Zhi, Wan Bai-Kun. Review of walk assistant exoskeleton technology:human-machine interaction. Acta Automatica Sinica, 2017, 43(7):1089-1100 http://www.aas.net.cn/CN/abstract/abstract19085.shtml [13] Zoss A B, Kazerooni H, Chu A. Biomechanical design of the Berkeley lower extremity exoskeleton (BLEEX). IEEE/ASME Transactions on Mechatronics, 2006, 11(2):128-138 doi: 10.1109/TMECH.2006.871087 [14] Mertz L. The next generation of exoskeletons:lighter, cheaper devices are in the works. IEEE Pulse, 2012, 3(4):56-61 doi: 10.1109/MPUL.2012.2196836 [15] Lee S, Sankai Y. Power assist control for leg with HAL-3 based on virtual torque and impedance adjustment. In: Proceedings of the 2002 IEEE International Conference on Systems, Man, and Cybernetics. Yasmine Hammamet, Tunisia: IEEE, 2002. 6 http://ieeexplore.ieee.org/xpls/icp.jsp?arnumber=1173329 [16] Esquenazi A, Talaty M, Packel A, Saulino M. The rewalk powered exoskeleton to restore ambulatory function to individuals with thoracic-level motor-complete spinal cord injury. American Journal of Physical Medicine and Rehabilitation, 2012, 91(11):911-921 doi: 10.1097/PHM.0b013e318269d9a3 [17] Wehner M, Quinlivan B, Aubin P M, Martinez-Villalpando E, Baumann M, Stirling L, et al. A lightweight soft exosuit for gait assistance. In: Proceedings of the 2013 IEEE International Conference on Robotics and Automation (ICRA). Karlsruhe, Germany: IEEE, 2013. 3362-3369 http://www.mendeley.com/catalog/lightweight-soft-exosuit-gait-assistance/ [18] Asbeck A T, De Rossi S M M, Galiana I, Ding Y, Walsh C J. Stronger, smarter, softer:next-generation wearable robots. IEEE Robotics and Automation Magazine, 2014, 21(4):22-33 doi: 10.1109/MRA.2014.2360283 [19] Araromi O A, Walsh C J, Wood R J. Hybrid carbon fiber-textile compliant force sensors for high-load sensing in soft exosuits. In: Proceedings of the 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Vancouver, BC, Canada: IEEE, 2017. 1798-1803 http://ieeexplore.ieee.org/document/8205994/ [20] Reiter R. Eine neue elektrokunsthand. Grenzgebiete der Medizin, 1948, 1(4):133-135 [21] Kuiken T A, Li G L, Lock B A, Lipschutz R D, Miller L A, Stubblefield K A, et al. Targeted muscle reinnervation for real-time myoelectric control of multifunction artificial arms. JAMA, 2009, 301(6):619-628 doi: 10.1001/jama.2009.116 [22] 黄剑平, 李文庆, 杨琳, 朱明星, 朱小弟, 李楚炎等.上肢经肱骨截肢神经功能重建研究.集成技术, 2016, 5(5):30-37 http://d.old.wanfangdata.com.cn/Periodical/jcjx201605003Huang Jian-Ping, Li Wen-Qing, Yang Lin, Zhu Ming-Xing, Zhu Xiao-Di, Li Chu-Yan, et al. A pilot study of nerve function reinnervation on a transhumeral amputee. Journal of Integration Technology, 2016, 5(5):30-37 http://d.old.wanfangdata.com.cn/Periodical/jcjx201605003 [23] Orendurff M S, Segal A D, Klute G K, McDowell M L, Pecoraro J A, Czerniecki J M. Gait efficiency using the c-leg. Journal of Rehabilitation Research and Development, 2006, 43(2):239-246 doi: 10.1682/JRRD.2005.06.0095 [24] Lawson B E, Mitchell J, Truex D, Shultz A, Ledoux E, Goldfarb M. A robotic leg prosthesis:design, control, and implementation. IEEE Robotics and Automation Magazine, 2014, 21(4):70-81 doi: 10.1109/MRA.2014.2360303 [25] Schultz A E, Kuiken T A. Neural interfaces for control of upper limb prostheses:the state of the art and future possibilities. PM & R, 2011, 3(1):55-67 http://www.sciencedirect.com/science/article/pii/S1934148210005174 [26] Rebsamen B, Guan C T, Zhang H H, Wang C C, Teo C, Ang M H, et al. A brain controlled wheelchair to navigate in familiar environments. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2010, 18(6):590-598 doi: 10.1109/TNSRE.2010.2049862 [27] Jiang N, Englehart K B, Parker P A. Extracting simultaneous and proportional neural control information for multiple-DOF prostheses from the surface electromyographic signal. IEEE Transactions on Biomedical Engineering, 2009, 56(4):1070-1080 doi: 10.1109/TBME.2008.2007967 [28] Shen B Q, Li J F, Bai F J, Chew C M. Motion intent recognition for control of a lower extremity assistive device. In: Proceedings of the 2013 IEEE International Conference on Mechatronics and Automation (ICMA). Takamatsu, Japan: IEEE, 2013. 926-931 http://www.mendeley.com/catalog/motion-intent-recognition-control-lower-extremity-assistive-device-lead/ [29] Burdet E, Milner T E. Quantization of human motions and learning of accurate movements. Biological Cybernetics, 1998, 78(4):307-318 doi: 10.1007/s004220050435 [30] Negro F, Muceli S, Castronovo A M, Holobar A, Farina D. Multi-channel intramuscular and surface EMG decomposition by convolutive blind source separation. Journal of Neural Engineering, 2016, 13(2):Article No. 026027 http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=c15f5fbe316d25cde6279ab9d4cedca3 [31] Hargrove L J, Englehart K, Hudgins B. A comparison of surface and intramuscular myoelectric signal classification. IEEE Transactions on Biomedical Engineering, 2007, 54(5):847-853 doi: 10.1109/TBME.2006.889192 [32] Vallery H, Veneman J, van Asseldonk E, Ekkelenkamp R, Buss M, van der Kooij H. Compliant actuation of rehabilitation robots. IEEE Robotics and Automation Magazine, 2008, 15(3):60-69 doi: 10.1109/MRA.2008.927689 [33] Wolf S, Grioli G, Eiberger O, Friedl W, Grebenstein M, Höppner H, et al. Variable stiffness actuators:review on design and components. IEEE/ASME Transactions on Mechatronics, 2016, 21(5):2418-2430 doi: 10.1109/TMECH.2015.2501019 [34] Engemann D A, Gramfort A. Automated model selection in covariance estimation and spatial whitening of MEG and EEG signals. NeuroImage, 2015, 108:328-342 doi: 10.1016/j.neuroimage.2014.12.040 [35] Fougner A, Scheme E, Chan A D C, Englehart K, Stavdahl Ø. Resolving the limb position effect in myoelectric pattern recognition. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2011, 19(6):644-651 doi: 10.1109/TNSRE.2011.2163529 [36] Peng L, Hou Z G, Chen Y X, Wang W Q, Tong L N, Li P F. Combined use of sEMG and accelerometer in hand motion classification considering forearm rotation. In: Proceedings of the 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). Osaka, Japan: IEEE, 2013. 4227-4230 http://www.mendeley.com/research/combined-semg-accelerometer-hand-motion-classification-considering-forearm-rotation/ [37] Cé E, Rampichini S, Monti E, Venturelli M, Limonta E, Esposito F. Changes in the electromechanical delay components during a fatiguing stimulation in human skeletal muscle:an EMG, MMG and force combined approach. European Journal of Applied Physiology, 2017, 117(1):95-107 doi: 10.1007/s00421-016-3502-z [38] Zardoshti-Kermani M, Wheeler B C, Badie K, Hashemi R M. EMG feature evaluation for movement control of upper extremity prostheses. IEEE Transactions on Rehabilitation Engineering, 1995, 3(4):324-333 doi: 10.1109/86.481972 [39] Boostani R, Moradi M H. Evaluation of the forearm EMG signal features for the control of a prosthetic hand. Physiological Measurement, 2003, 24(2):309-319 doi: 10.1088/0967-3334/24/2/307 [40] Nazmi N, Rahman M A A, Yamamoto S I, Ahmad S A, Zamzuri H, Mazlan S A. A review of classification techniques of EMG signals during isotonic and isometric contractions. Sensors, 2016, 16(8):Article No. 1304 doi: 10.3390/s16081304 [41] Shao Q, Bassett D N, Manal K, Buchanan T S. An EMG-driven model to estimate muscle forces and joint moments in stroke patients. Computers in Biology and Medicine, 2009, 39(12):1083-1088 doi: 10.1016/j.compbiomed.2009.09.002 [42] Buchanan T S, Lloyd D G, Manal K, Besier T F. Neuromusculoskeletal modeling:estimation of muscle forces and joint moments and movements from measurements of neural command. Journal of Applied Biomechanics, 2004, 20(4):367-395 doi: 10.1123/jab.20.4.367 [43] Lloyd D G, Besier T F. An EMG-driven musculoskeletal model to estimate muscle forces and knee joint moments in vivo. Journal of Biomechanics, 2003, 36(6):765-776 doi: 10.1016/S0021-9290(03)00010-1 [44] Kiguchi K, Hayashi Y. An EMG-based control for an upper-limb power-assist exoskeleton robot. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 2012, 42(4):1064-1071 doi: 10.1109/TSMCB.2012.2185843 [45] Kawamoto H, Sankai Y. Comfortable power assist control method for walking aid by HAL-3. In: Proceedings of the 2002 IEEE International Conference on Systems, Man, and Cybernetics. Yasmine Hammamet, Tunisia: IEEE, 2002. http://ieeexplore.ieee.org/xpls/icp.jsp?arnumber=1173328 [46] 彭亮, 侯增广, 王卫群.康复机器人的同步主动交互控制与实现.自动化学报, 2015, 41(11):1837-1846 http://www.aas.net.cn/CN/abstract/abstract18759.shtmlPeng Liang, Hou Zeng-Guang, Wang Wei-Qun. Synchronous active interaction control and its implementation for a rehabilitation robot. Acta Automatica Sinica, 2015, 41(11):1837-1846 http://www.aas.net.cn/CN/abstract/abstract18759.shtml [47] Yatsenko D, McDonnall D, Guillory K S. Simultaneous, proportional, multi-axis prosthesis control using multichannel surface EMG. In: Proceedings of the 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society. Lyon, France: IEEE, 2007. 6133-6136 http://www.ncbi.nlm.nih.gov/pubmed/18003415/ [48] Lenzi T, De Rossi S M M, Vitiello N, Carrozza M C. Intention-based EMG control for powered exoskeletons. IEEE Transactions on Biomedical Engineering, 2012, 59(8):2180-2190 doi: 10.1109/TBME.2012.2198821 [49] Ge S S, Li Y A, He H S. Neural-network-based human intention estimation for physical human-robot interaction. In: Proceedings of the 8th International Conference on Ubiquitous Robots and Ambient Intelligence (URAI). Incheon, South Korea: IEEE, 2011. 390-395 http://www.mendeley.com/catalog/neuralnetworkbased-human-intention-estimation-physical-humanrobot-interaction/ [50] Hogan N. Impedance control:an approach to manipulation:Part I:theory, Part Ⅱ:implementation, Part Ⅲ:applications. Journal of Dynamic Systems, Measurement, and Control, 1985, 107:1-24 doi: 10.1115/1.3140702 [51] Zlatnik D, Steiner B, Schweitzer G. Finite-state control of a trans-femoral (TF) prosthesis. IEEE Transactions on Control Systems Technology, 2002, 10(3):408-420 doi: 10.1109/87.998030 [52] Modares H, Ranatunga I, Lewis F L, Popa D O. Optimized assistive human-robot interaction using reinforcement learning. IEEE Transactions on Cybernetics, 2016, 46(3):655-667 doi: 10.1109/TCYB.2015.2412554 [53] Wen Y, Si J, Gao X, Huang S, Huang H H. A new powered lower limb prosthesis control framework based on adaptive dynamic programming. IEEE Transactions on Neural Networks and Learning Systems, 2017, 28(9):2215-2220 http://europepmc.org/abstract/MED/27416607 [54] Schaechter J D. Motor rehabilitation and brain plasticity after hemiparetic stroke. Progress in Neurobiology, 2004, 73(1):61-72 doi: 10.1016/j.pneurobio.2004.04.001 [55] Krakauer J W. Motor learning:its relevance to stroke recovery and neurorehabilitation. Current Opinion in Neurology, 2006, 19(1):84-90 doi: 10.1097/01.wco.0000200544.29915.cc [56] Carr J H, Shepherd R B. Stroke Rehabilitation: Guidelines for Exercise and Training to Optimize Motor Skill. Butterworth-Heinemann Medical, Oxford, UK, 2003. Stroke Rehabilitation: Guidelines for Exercise and Training to Optimize Motor Skill [57] Levin M F, Weiss P L, Keshner E A. Emergence of virtual reality as a tool for upper limb rehabilitation:incorporation of motor control and motor learning principles. Physical Therapy, 2015, 95(3):415-425 doi: 10.2522/ptj.20130579 [58] 胡进, 侯增广, 陈翼雄, 张峰, 王卫群.下肢康复机器人及其交互控制方法.自动化学报, 2014, 40(11):2377-2390 http://www.aas.net.cn/CN/abstract/abstract18514.shtmlHu Jin, Hou Zeng-Guang, Chen Yi-Xiong, Zhang Feng, Wang Wei-Qun. Lower limb rehabilitation robots and interactive control methods. Acta Automatica Sinica, 2014, 40(11):2377-2390 http://www.aas.net.cn/CN/abstract/abstract18514.shtml [59] Basteris A, Nijenhuis S M, Stienen A H A, Buurke J H, Prange G B, Amirabdollahian F. Training modalities in robot-mediated upper limb rehabilitation in stroke:a framework for classification based on a systematic review. Journal of Neuroengineering and Rehabilitation, 2014, 11:Article No. 111 doi: 10.1186/1743-0003-11-111 [60] Lo A C, Guarino P D, Richards L G, Haselkorn J K, Wittenberg G F, Federman D G, et al. Robot-assisted therapy for long-term upper-limb impairment after stroke. New England Journal of Medicine, 2010, 362(19):1772-1783 doi: 10.1056/NEJMoa0911341 [61] Reinkensmeyer D J, Wolbrecht E, Bobrow J. A computational model of human-robot load sharing during robot-assisted arm movement training after stroke. In: Proceedings of the 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society. Lyon, France: IEEE, 2007. 4019-4023 http://www.ncbi.nlm.nih.gov/pubmed/18002881 [62] Cai L L, Fong A J, Liang Y Q, Burdick J, Edgerton V R. Assist-as-needed training paradigms for robotic rehabilitation of spinal cord injuries. In: Proceedings of the 2006 IEEE International Conference on Robotics and Automation. Orlando, FL, USA: IEEE, 2006. 3504-3511 ssist-as-needed training paradigms for robotic rehabilitation of spinal cord injuries. [63] Krebs H I, Palazzolo J J, Dipietro L, Ferraro M, Krol J, Rannekleiv K, et al. Rehabilitation robotics:performance-based progressive robot-assisted therapy. Autonomous Robots, 2003, 15(1):7-20 doi: 10.1023/A:1024494031121 [64] Wolbrecht E T, Chan V, Reinkensmeyer D J, Bobrow J E. Optimizing compliant, model-based robotic assistance to promote neurorehabilitation. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2008, 16(3):286-297 doi: 10.1109/TNSRE.2008.918389 [65] Pehlivan A U, Losey D P, O'Malley M K. Minimal assist-as-needed controller for upper limb robotic rehabilitation. IEEE Transactions on Robotics, 2016, 32(1):113-124 http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=7dba74b5b9256a157f209f2c28b5611a 期刊类型引用(10)
1. 宋红超，王昕，王振雷. 基于切换的非线性多模型二阶段广义预测控制. 控制理论与应用. 2024(11): 2147-2156 . 
百度学术2. 赵建国，杨春雨. 复杂工业过程非串级双速率组合分散运行优化控制. 自动化学报. 2023(01): 172-184 . 本站查看3. 王建琴. 基于协同过滤算法的医院财务工作轮岗流程智能控制模型. 微型电脑应用. 2022(06): 54-56 . 百度学术4. 陈尉. INFIT系统在火电机组辅助调频控制中的应用. 工业控制计算机. 2021(01): 26-27+30 . 百度学术5. 陈铁锋，杜文. 基于距离度量的火电机组协调系统软切换预测控制. 热能动力工程. 2021(03): 87-92+121 . 百度学术6. 王素珍，刘建锋. 改进型BP神经网络的非线性多模型自适应控制. 电光与控制. 2021(08): 1-5 . 百度学术7. 李军，黄卫剑，万文军，刘哲. 一种新型反馈控制器的研究与应用. 控制理论与应用. 2020(02): 411-422 . 百度学术8. 陈钊，乔侨. 分层结构多模型预测控制在630MW火电机组协调系统中的应用. 工业控制计算机. 2020(10): 4-6+10 . 百度学术9. 王岩，王昕，王振雷. 一类时变系统的多模型切换动态调节控制算法. 控制理论与应用. 2020(12): 2501-2510 . 百度学术10. 唐伟强，龙文堃，孙丽娟，黄小丽. 基于聚类方法和神经网络的非线性系统多模型自适应控制. 系统工程与电子技术. 2019(09): 2100-2106 . 百度学术其他类型引用(7)
-
下载:
下载:
计量
- 文章访问数: 3550
- HTML全文浏览量: 1280
- PDF下载量: 1536
- 被引次数: 17
