Aerodynamic Identification and Control Law Design of a Missile Using Machine Learning

Abstract

The new generation of air vehicle is confronted with a more intricate environment and challenging missions, which puts forward higher requirements for the flight control system. In this study, the aerodynamic identification and control law design based on machine learning for a missile configuration is investigated through numerical simulations. The missile pitch and elevator deflection are realized via the combination of a rigid dynamic grid method and an overlapping grid technology, while the computational fluid dynamics/rigid body dynamics (CFD/RBD) strong coupling method is implemented to simulate the unsteady flows associated with the motion of the missile. Firstly, the aerodynamic data of the missile are gathered through forced pitching motion involving elevator deflection, and an aerodynamic model is constructed using a deep neural network to identify the aerodynamic moment with only a small number of unsteady aerodynamic data. Then, the accuracy and fidelity of the model are checked with the open-loop control law. Afterward, a missile pitch control law is generated through deep reinforcement learning based on the aerodynamic model, which enables the realization of a robust and exact angle-of-attack control process. Finally, the control law is transferred to a numerical environment and numerical virtual flight based on CFD is conducted, which demonstrates that stable control can be maintained even in continuous maneuvering. This study verifies the possibility of applying a deep neural network to air vehicle aerodynamic identification and deep reinforcement learning to a complicated flight control law design with excellent generalization ability. Machine learning is expected to play a significant role in the design and research for the novel generation of air vehicle.