Congealed Deep Neural Network-based System Identification for Morphing Aircraft

This paper proposes a new deep neural network (DNN) architecture called the congealed DNN for system identification of morphing aircraft. The developed DNN consists of two parts: the invariant features of the inner layers and the time-varying weights of the output layer. For the inner invariant features, a novel meta-learning with adversarial optimization framework is developed to derive a common representation function shared by different deformation conditions. For the time-varying weights, we consider them to be composed of congealed weights and time-varying perturbations. The congealed weights are estimated using standard adaptive techniques, while a sliding mode-like function is employed to attenuate time-varying disturbance terms. The experimental results indicate that the proposed method demonstrates more precise and faster adaptation capabilities to the morphing aircraft system compared to other methods.