Nonlinear Adaptive Control of Unmanned Aerial Vehicle Using Echo State Neural Networks

Abstract

This paper presents the use of Echo State neural Networks (ESN) for nonlinear adaptive control of a fixed-wing unmanned aerial vehicle (UAV). Both offline and online trained ESNs are investigated. Flight test data and simulated data are used to train the ESNs offline. The first offline trained ESN is used to estimate the inverse transformation function required for feedback linearization, specifically for the lateral-directional case. The second ESN is trained to estimate error in the output of the first ESN. The estimated error from the second ESN is used to make real time controller corrections as well as update the weighting parameters of the output phase of the first ESN, and thus allowing the controller to adapt in flight. Simulated results show that including online adaptive control improves performance in the presence of noise and disturbances.