Deep FBSDE Controller for Attitude Control of Hypersonic Aircraft

Abstract

Attitude control of Hypersonic Aircraft is a very challenging subject due to the uncertainties and various noises of the system. In this paper, we propose a new methodology to solve this problem. Firstly, the attitude control of Hypersonic Aircraft is reformulated as a system of Forward-Backward Stochastic Differential Equations. Deep Neural Networks (DNNs) are used to get optimal solution of the equations. We have studied several deep neural networks, including FC-based architecture and LSTM-based architecture and proposed a new FC-based architecture that shares the weights between different time steps, which performed satisfactorily in this problem. The performance and universality of the algorithm are tested in both unconstrained and control-constrained cases. Simulation and experimental results verify the superiority of the algorithm.