Optimal performance and robust stabilization (Q1075984): Difference between revisions

This paper deals with the design of feedback controllers which minimize the \(H_{\infty}\)-norm of the sensitivity matrix under the constraint that the closed-loop system remains stable. Let P and C be the plant and the controller respectively, where P is a \(p\times q\) transfer matrix whose elements are strictly proper rational functions and C is a \(q\times p\) transfer matrix whose elements are proper rational functions. In the following, the closed-loop system \[ y=(I+PC)^{-1}PC\cdot r+(I+PC)^{- 1}P\cdot d_ 1 \] is considered, where y, r and \(d_ 1\) are a plant output, a reference input and a disturbance input respectively. Assume that P is described via a nominal plant \(P_ 0\) together with m which is the given function, representing the radius of uncertainty. Associated with \(P_ 0\), \(S_ 0:=(I+PC)^{-1}\) is called the nominal sensitivity matrix. Define \(K(P_ 0,m)\) to be the class of all plants which satisfy \(P=(I+M)P_ 0\) for some rational matrix M such that \(\sigma [M(j\omega)]<m(j\omega)\), \(\omega\in {\mathbb{R}}\), where \(\sigma\) [M] denotes the largest singular value of M. Let the operating band of the feedback system be \([0,\omega_ 1]\). The problem is to find the infimum \[ \nu:=\inf \{\| S_ 0\|_{[0,\omega_ 1]}:\quad C\quad stabilizes\quad all\quad P\in K(P_ 0,m)\}. \] For mathematical convenience, all functions of s are transformed into functions of z via \(s=(1-z)/(1+z)\). Thus, \([-j\omega_ 1,j\omega_ 1]\) is mapped onto an arc \(\{e^{j\theta}:\) \(| \theta | \leq \theta_ 1\}.\) The polynomial matrix coprime factorization of \(P_ 0\) is as follows: \(P_ 0=\tilde B^{-1}\tilde A=AB^{-1}\), \(\tilde A\tilde X+\tilde B\tilde Y=I\), \(XA+YB=I\). Then the family of all stabilizing controllers is parametrized by \(C=(\tilde X+BZ)\cdot (\tilde Y-AZ)^{-1}\), \(Z\in RH_{\infty}^{q\times p}\) where \(RH_{\infty}\) is a real rational subspace of \(H_{\infty}\). Next, factor A and B with inner and outer factors: \(A=A_ iA_ e\), \(B=B_ iB_ e\). Define \(H:=A_ i^{- 1}\tilde Y\tilde B_ e\), \(J:=A_ i^{-1}\tilde B_ i^{-1}\), \(Q:=A_ eZ\tilde B_ e\). For the case \(| \theta | \leq \theta_ 1\) dedine \(v_{\epsilon}(e^{j\theta}):=\epsilon^{-1}\), \(E(e^{j\theta}):=H(e^{j\theta})\). For the case \(| \theta | >\theta_ 1\) define \(v_{\epsilon}(e^{j\theta}):=m(e^{j\theta})\), \(E(e^{j\theta}):=(H-J)(e^{j\theta})\). The function \(v_{\epsilon}\) is an outer \(H_{\infty}\) function and the matrix E belongs to \(L_{\infty}^{p\times p}\). Let \(\Gamma_{\epsilon}\) be the Hankel operator with symbol \(v_{\epsilon}E\). Define \(\epsilon^ 0:=\inf \{\| \Gamma_{\epsilon}\| <1\}.\) The main results are given by the following theorems. Theorem 1. \(\nu >\epsilon^ 0.\) Theorem 2. Assume that \((1+\epsilon^ 0)\| m\|_{[-\theta_ 1,\theta_ 1]}<1\). For every \(\theta_ 2\) such that \(\theta_ 1>\theta_ 2>0\) and for every \(\delta >0\), there exists a controller which stabilizes every \(P\in K(P_ 0,m)\) and for which the corresponding sensitivity matrix satisfies \[ \| S_ 0\|_{[-\theta_ 2,\theta_ 2]}<\epsilon^ 0+\delta. \] A computational approach and a scalar example are presented to illustrate the trade-offs between achievable performance and plant uncertainty.