Stability of the integral control of stable nonlinear systems

PI controllers are the most widespread type of controllers and there is an intuitive understanding that if their gains are sufficiently small and of the correct sign, then they 'always' work. In this paper we try to give some rigorous backing to this claim, under specific assumptions. Let P be a nonlinear system described by = f (x, u), y = g(x), where the state trajectory x takes values in Rⁿ, u and y are scalar and f, g are of class C¹. We assume that there is a Lipschitz function Ξ : [u_min, u_max] → Rⁿ such that for every constant input ^u 0 ^{[u min, u} max], Ξ(u₀) is an exponentially stable equilibrium point of P. We also assume that G(u) = gΞ(u)), which is the steady state input-output map of P, is strictly increasing. Denoting y_min = G(u_min) and y_max = G(u_max), we assume that the reference value r is in (y_min, y_max). Our aim is that y should track r, i.e., y → r as t →∞ while the input of P is only allowed to be in [u_min, u_max]. For this, we introduce a variation of the integrator, called the saturating integrator, and connect it in feedback with P in the standard way, with gain k > 0. We show that for any small enough k, the closed-loop system is (locally) exponentially stable around an equilibrium point (Ξ(u_r ), u_r), with a 'large' region of attraction X_T ⊂ Rⁿ × [u_min, u_max]. When the state (x(t),u(t)) of the closed-loop system converges to (Ξ(u_r),u_r), then the tracking error r - y tends to zero. The compact set X_t can be made larger by choosing a larger parameter T > 0, resulting in smaller k.