Identification des coefficients aerodynamiques et commande de vol non lineaire

Classical flight control is achieved with linear controllers that are scheduled the entire flight envelope. This approach, although quite straightforward, is tedious and may not be adequate for flying operations at a distance from the linearised region. Thus modern flight control aims at developing non linear and/or adaptive control laws, based on non linear plant models, which are valid over the entire flight envelope.; The objective of this theses is to develop such a control law, is order to command the aircraft's angle of attack alpha, sideslip angled beta and roll angle o.; In the first place, the aircraft's aerodynamic coefficients are identified with precision over the entire flight envelope. To do so, a novel architecture based on neural networks is developed. It achieves the desired performances, and is of a reasonable size with regards to the number of neural networks it contains. This architecture can be easily extended to the on line identification needed for reconfigurable flight control.; In the second place, a dynamic inversion control law, the parameters of which are issued by the neural identification module, is developed. The use of precise aerodynamic coefficient estimates enables limited model errors, and provides an improved robustness of the control law when the desired dynamics is generated by a proportionel+integrator feedback.; The research ends on an extension of this control law by a predictive command. This command, based on the feedback linearised model of the plant, anticipates changes in the reference trajactory and limits the control effort. However the robustness of this control law was not part of this research and remains to be studied.; The theoretical results are applied to the Fighting Falcon F-16 through simulation on a MatLab and Simulink platform. The neural identification module contains a hunted number of networks, and achieves the desired precision performance. The dynamic inversion command is then tested on both a perfect model and a "reel" model of the plant. In both cases the results confirm that the reference was correctly tracked. This is a positive premise for the robustness of this control law.