Flight simulation and hardware implementation of deep model predictive control

This thesis presents the flight simulation and hardware implementation of Deep Model Predictive Control (DMPC) on an experimental setup, which consists of a quadcopter and motion capture system. DMPC aims to adapt abrupt state-dependent matched uncertainties arising due to faults, collects training data for a deep neural network (DNN) to learn slowly varying features, and ensures safety during the learning phase. Training of DNN to learn features is carried out on a parallel machine, while the actual system is controlled by a tube MPC and an adaptive mechanism with fixed features. Under certain verifiable technical conditions, DMPC ensures the asymptotic stability of closed-loop states. Through simulations presented in this thesis, it is shown that DMPC can outperform other control architectures when utilizing an affine model with additive nonlinear disturbance to the control input, and is able to guarantee stability while avoiding unwanted behavior in early learning phases while having long-term learning capabilities. This study demonstrates that DMPC is a powerful and safe control architecture for nonlinear systems.LimitedAuthor requested closed access (OA after 2yrs) in Vireo ETD syste