Numerical virtual flight investigation for longitudinal maneuver of a generic fighter based on machine learning

Coupled with computational fluid dynamics (CFD), rigid body dynamics (RBD), and flight control system, the numerical virtual flight (NVF) technology can simulate the maneuvering flight process of an air vehicle under control. In this paper, the NVF investigation of longitudinal maneuvers with elevator and thrust vector control is performed for a generic fighter configuration. The rigid dynamic hybrid grid method is taken to realize the motion of the fighter, and the overlapping moving grid technology meets the deflection of the elevator. The Reynolds-averaged Navier–Stokes equations in arbitrary Lagrangian–Eulerian form are coupled with the RBD equations to solve aerodynamics and kinematics problems, while flight control is achieved through an advanced machine learning method. First, the fighter is forced to pitch with the periodic deflection of the elevator, and the unsteady computation is carried out to obtain aerodynamic data. Then, an artificial neural network (ANN) is adopted for aerodynamic identification and modeling, which involves establishing a model between the aerodynamic coefficient and pitching motion parameters. Afterward, the ANN-based NVF is implemented on the basis of the established model and deep reinforcement learning (DRL) is used to design the pitching control law of the fighter. The NVF results based on ANN show that the fighter has a good control effect under the action of the elevator, elevator with open-loop thrust vector, and elevator with closed-loop thrust vector, respectively, as well as the results from the CFD-based NVF system. Finally, the three-degree-of-freedom NVF based on CFD also indicates that the control law designed through DRL has good generalization characteristics. This study demonstrates the potential prospects of machine learning methods in the design and research for a novel generation of air vehicles.