Series-parallel approach to on-line observer based neural control of a helicopter system

Abstract

This paper is concerned with the control of an under-actuated, uncertain, delayed non-linear system through the implementation of arti cial neural networks(ANNs). The aim is the development of a series-parallel training scheme for the on-line observer to ensure faster convergence and more accurate estimations. Reinforcement learning is used to improve future performance and maintain stability while an estimated tracking error is minimised. Lyapunov stability measures are employed to guarantee the uniform ultimate boundedness of the closedloop tracking error. Real-time learning algorithms are derived for the individual components (observer, actor, critic). Final performance is tested on a mathematical helicopter model and real-world helicopter ight data