Development of model predictive control schemes for unmanned aerial vehicles using neural networks

Abstract

This thesis is focused on developing a data-driven nonlinear Model Predictive Control (MPC) scheme for controlling a quadcopter’s attitude dynamics. Quadcopter vehicles constitute a specific type of Unmanned Aerial Vehicles (UAVs) with high flexibility and rapid deployment that have gained significant traction in the industry over the past decade. The novel MPC controller, developed in this thesis, leverages Radial Basis Function (RBF) neural networks to model the orientation dynamics, which exhibit nonlinear behavior that traditional linear methodologies ignore. The RBF network is utilized for system identification, in order to create more precise predictive models that accurately capture the system’s orientation behavior. In addition, to train the neural networks, the Fuzzy Means (FM) learning algorithm was employed to determine the RBF centers. These models were then integrated into the MPC framework to improve control performance, robustness and adaptability. Extensive simulations were conducted to evaluate the ability of the nonlinear MPC which was based on purely data-driven models, to efficiently control a highly nonlinear system. The results from an angle-regulating scenario demonstrated the improved performance of the proposed scheme for controlling the system’s behavior compared to linear MPC schemes, underscoring the potential of RBF neural networks in advancing UAV autonomous tasks and control. Furthermore, this thesis provides a comprehensive comparison between the MPC control scheme and the traditional Proportional-Integral-Derivative (PID) control scheme for a demanding trajectory-tracking task. Based on simulation results, the optimal control MPC scheme outperformed the classic PID control architecture, highlighting its advantages in driving complex multivariate nonlinear systems along a reference trajectory.