Design of a novel control system using neural network for a VTOL drone

The control of dynamic systems is a highly complex discipline largely due to the non-linearity of the processes involved. Over time, numerous methodologies have been proposed to achieve efficient control, often relying on approximating models or linearizations, which, while effective, are quite distant from direct real-world control. This project aims to implement a new methodology that is increasingly gaining traction today, namely the use of neural networks for problem-solving. While dynamics that are difficult to model with standard mathematical interpretations will still be approximated in their behavior, the approach shifts towards approximators whose internal logic more closely resembles their highly non-linear nature, thus incredibly increasing the possibilities. This is due to their ability to automatically organize internal logic through learning, with inner relationships between neurons. Despite demonstrating their efficiency, these networks are still heavily studied to understand why they are organized in these ways. In particular, radial basis function networks will be used, which are slightly different from all the others and are among the best approximators of linear and non-linear functions. This new control methodology will be based on a composite approach, leveraging information on estimations about the states and trajectory errors plus the implementation of a disturbance observer to better tailor the control on the different sources of disturbance. It will be applied to a simulation model of a UAV currently in the design and development phase by Sky Eye Systems, a company specialized in the production of military drones, which already boasts certifications that are difficult to obtain in the aerospace industry.