We propose a computational method for verifying a state-space safety constraint of a network of interconnected dynamical systems satisfying a dissipativity property. We construct an invariant set as the sublevel set of a Lyapunov function comprised of local storage functions for each subsystem. This approach requires only knowledge of a local dissipativity property for each subsystem and the static interconnection matrix for the network, and we pose the safety verification as a sum-of-squares feasibility problem. In addition to reducing the computational burden of system design, we allow the safety constraint and initial conditions to depend on an unknown equilibrium, thus offering increased flexibility over existing techniques.
Index Terms-Lyapunov methods, safety verification, sum-of-squares programming.
I. INTRODUCTIONM ANY complex engineered systems result from the interconnection of well understood subsystems, however the interconnection itself results in global behavior not readily apparent from the constituent subsystems. A standard approach to safety verification of such systems relies on set invariance [1], [2] where invariant sets are established by considering sublevel sets of Lyapunov functions [1]. However, computation of Lyapunov functions is often elusive without exploiting structural system properties, and standard Lyapunov theory requires explicit knowledge of the system equilibrium.The main contribution of this work is to propose a computational method for finding an invariant set that avoids an unsafe region of the state space by parameterizing the search for an appropriate Lyapunov function using local dissipativity storage functions and sum-of-squares (SOS) techniques [3]. We consider networked dynamical systems composed of subsystems, each of which is assumed to satisfy a dissipativity property, interconnected through a static feedback matrix. In particular, the subsystems are assumed to satisfy an equilibrium-independent dissipativity property introduced in [4]. Examples of practically important networks that are composed of passive subsystems have been exhibited in [5]-[8].This work is particularly motivated by networks which induce a unique equilibrium such as certain classes of communication networks [9] and biological networks with inhibitory feedback [10], and several cooperative control problems [11], [12]. The primary contribution of this work compared to the conference version [13] is the application of our approach to equilibrium-independent subsystems for which the steady state of the local subsystem may not be uniquely defined by the Manuscript