Optimal controller tuning for nonlinear processes (Q705441)

The authors point out that traditional tuning of nonlinear control systems was based on heuristics and experience. Nonlinearities generally made it impossible to use Laplace transform and to talk immediately about zeros and poles in the frequency domain, where outstanding results have emerged in the period 1960--2005. R. E. Kalman and others stressed the Lyapunov functional approach, and several articles and monographs, such as the one written by Matrosov (in Russian, no English translation), appeared in the period 1960--2000, discussing the Lyapunov functions which seemed to be a very important tool in studies of stability. Most engineers and probably a majority of mathematicians involved in control theory (including the reviewer) discussed the concepts of regions of stability, based on positivity and of a negative derivative with respect to time, along the controlled path of an ad hoc generated Lyapunov function that seemed to fit a particular control problem. An outstanding work by Zubov [\textit{V. I. Zubov}, Methods of A. M. Lyapunov and their applications. Noordhoff, Groningen (1964; Zbl 0115.30204); Russian original: Izdat. Leningrad. Univ., Moskva (1957; Zbl 0078.07902)] introduced the idea of generating such functions by solving a partial differential equation, now known as the Zubov equation. This exactly is the main topic of this article. The authors consider a nonlinear system following the space representation: \(\partial x/\partial t = f(x,u)\), \(x\in\mathbb{R}^n\), \(u\in\mathbb{R}^q\), with \(f(0,0) = 0\), \(y=h(x)\), with \(h(0) = 0\). A closed-loop feedback system is to be designed such that the system is locally asymptotically stable at the origin and the observed output \(y(t)\) asymptotically decays to zero as \(t\to\infty\). The performance index for the controlled system is defined by: \(J(p) =\int^\infty_0\{\| y\|^2+\rho\| u\|^2\}\,dt\), where \(\rho\) is a weight parameter, and where \(p\) is the vector of controller parameters. \(p\) obeys the relation: \(\partial x/\partial t =F(x,p)\), while \(x\) is the augmented state vector, following ideas of \textit{R. Sepulchre, M. Janković}, and \textit{P. Kokotović} [Constructive nonlinear control. Springer Verlag, New York (1997; Zbl 1067.93500)] which include enlarging the state space. We can write: \(J(p) = \int^\infty_0Q(x,p)\,dt\). Now, under some assumptions, the Zubov equation: \(\partial V(x,p)/\partial x= -Q(x,p)\) admits a unique locally analytic solution, which is the required Lyapunov function. An illustrative example describes a reactor dynamics where a van de Vusse reaction takes place in an isothermally stirred tank reactor. The Zubov equation was solved using the Maple code.