A control architecture for fixed-wing aircraft based on the convolutional neural networks

Abstract

This paper develops a nonlinear architecture to control different fixed-wing aircraft. This architecture has inner and outer loops. The inner loops, designed based on the convolutional neural networks, control the internal dynamics of the aircraft, and the outer loops, which use linear controllers, are designed to control the kinematic states, which are the same for all aircraft. So, the inner loops are designed to have adaptation mechanisms based on the convolutional neural networks to control the internal dynamics of different aircraft. The networks are trained offline based on a generated database to avoid time-consuming online procedures. The database is created by simulating simple linear training models. Then, the input-output data of these training models are preprocessed and mapped to an appropriate form to be fed to convolutional neural networks. After that, an appropriate network structure is selected, and the networks are trained based on the mapped database. These trained networks, along with linear controllers in a cascade form, are applied to control nonlinear simulations of 15 different fixed-wing aircraft. Then, the performance of the represented controller is compared to two adaptive controllers. The promising results show that the controller can sufficiently be utilized in controlling fixed-wing aircraft, even in slow or sudden changes in aircraft dynamics.