Nonlinear control of UAVs using multi-layer perceptrons with off-line and on-line learning

This paper presents the research on the development of neural network based non-linear controllers for an airplane UAV. Multi-layer perceptrons are used for the training of networks, both off-line and on-line. The data required for off-line training is generated from a validated non-linear flight dynamics model of the Cal Poly Pomona 12´ Telemaster UAV. The off-line trained network using multi-layer perceptrons replaces the inverse transformation required for feedback linearization. On-line training is then accomplished to account for the inversion and modeling error. The controllers are tested in the software-in-the-loop simulation environment using FlightGear Flight Simulator. Simulation results compared with flight data are shown. Also shown are the results in the presence of sensor noise.