Neural Network‐Based Runtime Monitoring and Control for Unknown Nonlinear Systems

Real‐time system monitoring and stabilization become especially challenging when the system dynamics are unknown. This article introduces a novel design for monitoring and stabilizing unknown nonlinear systems with measurement noise. The design utilizes a generic modelling framework by decomposing the system into a known tunable linear component and an unknown nonlinear component, which is approximated using a deep neural network (DNN) with bounded errors. A robust interval observer is designed to produce tight state bounds for the original system state, with the centre of the interval serving as an estimate of the true state. Based on this estimate, a feedback control law is designed to stabilize the unknown nonlinear system. The observer and controller gains are both solved from convex optimization problems. Extensive simulations demonstrate the effectiveness of the proposed method and its robust performance across DNNs with varying structures and approximation errors.