Proceedings of the Mansoura 4th Int. Engineering Conference,Egypt , 20-22 April, 2004
DESIGN OF A NEURAL NETWORK SYSTEM
REALIZING PREDETERMINED AIRCRAFT TRAJECTORIES
Attia E. M. *, Badr M.A. **, Awad T. H.*
* Faculty of Engineering –Alexandria University –Alexandria, Egypt.
** Arab Academy for science and Technology- Alexandria Egypt.
Abstract - The aim of this work is to design and test the implementation of the ANN (artificial neural network) in controlling the motion of aircraft to realize predetermined spatial maneuvers. This replaces the classic solutions of the inverse problem, which may lead to inaccurate guidance of the aircraft. An aircraft model is derived that simulates both the longitudinal and the lateral motions. For different input control surfaces the output trajectories are calculated, registered and used to train the selected neural networks in the inverse direction. Three NN designs are applied and tested. The first NN is a single feed-forward NN trained for all possible spatial aircraft trajectories. The second design is a group of parallel feed-forward neural networks. Each one are trained to obtain the control surfaces to realize special type of spatial maneuvers. The third design is a combination of two groups of parallel neural networks, one group for longitudinal motion and the other for the lateral. In both groups each network are trained for a special types of trajectories.
Different layers, neuron types and numbers are applied in each design .The accuracy and training time are compared together. The third NN design with separate longitudinal and lateral motions shows more accurate trajectories and the training time needed is less than the other two designs.