Approximate Optimal Control for Morphing Aircraft via Attention Meta-Learning and Continual Learning

This study presents an innovative approximate optimal controller for variable-span morphing aircraft (MA), where the mapping relationship between aerodynamic parameters and deformation parameters is unknown. The foundation of our approach relies on two core observations: unknown dynamic models for MA have a common representation across different morphing scenarios, while the specific morphing condition lies in a collection of adaptive linear parameters. Guided by this dual-layer decomposition, a novel meta-learning (MetaL) training structure with an adversarial optimization framework is leveraged to extract common invariant features from offline flight data. At the same time, we incorporate the squeeze-and-excitation (SE) network, a novel architectural unit, into the MetaL training process, aiming to enhance the network's representational capability through dynamic recalibration of channel features. Regarding the new deformation situations encountered during the online process, online data are collected to update the common features via a continual learning method, and a concurrent learning (ConcL)-based methodology is introduced to update the linear coefficients. Utilizing the identified dynamical model, a model-based reinforcement learning (RL) framework is formulated to address optimal control problems of MA. By employing nonsmooth Lyapunov stability analysis, it is demonstrated that continuously updating the weights allows the implemented control strategies to converge to a neighborhood of the optimal strategies. Furthermore, numerical simulations are performed to validate the effectiveness of the proposed methodology.