Imitation Learning of MPC for Fault-Tolerant Control of an Octorotor UAV

Artificial Neural Networks (ANNs) have been increasingly researched in aerospace applications due to promising advantages such as their ability to model nonlinear functions, recognize patterns, generalize to unseen data and fast execution. Aiming to investigate the use of ANNs for flight control problems, the contribution of this work is the development of a neural network-based fault-tolerant control scheme for a Vertical Take-Off and Landing (VTOL) aircraft via imitation learning of a linear Model Predictive Control (MPC) expert. While the MPC offers high control performance and explicitly handles constraints, it demands considerable computational power. The neural network, then, offers the potential benefit of imitating the MPC's behavior while reducing the computational cost. The simulation results demonstrate that the neural network closely replicates the expert's behavior across a good range of scenarios, achieving comparable trajectory tracking performance under nominal conditions and under single-rotor failures. It is also capable of maintaining the hover of the octorotor under two simultaneous total rotor failures. Moreover, the neural network showed good generalization to new trajectories, and implicitly learned the rotor angular speed and attitude angle constraints that were defined for the MPC expert. Finally, the neural network significantly outperformed the MPC in computational efficiency, thus showing good potential for real-time control.