Observation à échantillonnage événementiel d'un système cyber-physique

Abstract

This thesis focuses on event-triggered state estimation problems within the context of cyber-physical systems (CPSs), aiming to develop new event-triggered estimators for nonlinear and Gaussian/nonGaussian systems. Event-triggered state estimation has been a prominent area of systems research for several decades, with successful applications in diverse elds such as signal processing, target tracking, and navigation systems. This approach o ers a promising solution to data tra c congestion by facilitating aperiodic, event-triggered information exchange between sensors and estimators. The motivation for this research stems from the resource limitations inherent in CPS applications, such as wireless sensor networks, and the increased computational burden associated with calculating optimal state estimates under event-triggering conditions. In this work, we address several practical challenges encountered in the eld and endeavor to advance the state of the art in event-triggered state estimation. In the rst part, we provide a brief introduction to the problem of systems under event-triggering conditions and outline the main theory using probabilistic inference, where the problem is addressed with Bayesian state estimation. In the second part, we present the necessary theory for event-triggered state estimation, where the Gaussian assumption are discussed to approximate the posterior Probability Density Function (pdf). Here, the problem is reduced to one of the approximated nonlinear type of Kalman lters. In the next part of this research, we assume that the posterior pdf is no longer to be Gaussian. Therefore, we develop an event-triggered particle lter to approximate the non-Gaussian posterior. The pdfs are approximated based on Monte Carlo simulations using a set of particles and weights. However, computing the particle weights based on the event trigger condition can lead to a computational burden. To address this issue, a Bayesian constraint is developed. Finally, we study the e ect of packet dropouts on the performance of state estimators, speci cally focusing on particle lters. Packet dropouts, caused by imperfect communication channels, are unavoidable when information is transmitted through a communication network. We rst develop a nonlinear particle lter to reduce the estimation error. Using a special form of the sequential Monte Carlo algorithm, the posterior distribution is approximated, and the corresponding minimum meansquared error is derived. By contrasting the error covariance matrix with the posterior Cramér-Rao lower bound, the estimator's performance is assessed.