Event Triggered Neuro Controller for a Class of Nonlinear Systems

Abstract

In this thesis, identification and control strategies for nonlinear systems are tailored for those newlinescenarios where only limited computation and/or communication resources are available. The newlinecontrol strategies under consideration can be employed in both networked control systems newline(NCSs) and event-triggered control systems (ETCSs). newline newlineA novel approach is proposed for system identification in event-triggered systems to effec- newlinetively minimize computational costs. Unlike conventional neural network training methods newline newlinethat utilize all available training data, specifically the input-output pairs, the proposed ap- newlineproach trains the network only when the identification error for a particular input-output newline newlinepair surpasses a predefined threshold. It is demonstrated that the training algorithm with newlineevent-triggering is statistically comparable to the same algorithm without event-triggering. newlineMoreover, the online sequential learning algorithm (OSLA) outperforms backpropagation newlinelearning in terms of convergence time. The reduced complexity of OSLA-1, which requires newlinefewer samples, makes it an ideal choice in scenarios where computational and communication newlinecosts are limiting factors. newline newlineEvent-triggered identification is combined with event-triggered control for nonlinear sys- newlinetems in this thesis. This novel combination results in a smaller computational and commu- newlinenication footprint. This is achieved with minimal degradation in the tracking performance. newline newlineIt is observed that OSLA-1 is best suited to this scenario. The observations are true for newlineaperiodic and periodic event-triggering. newline