Neural Network-Based Formation Flying Using Aerodynamic Forces via Variable Shape Function

Abstract

Conventional formation flying Earth-orbiting satellites control their orbits to perform their missions using thrusters, but the amount of propellant loaded into a satellite is limited. Therefore, the use of aerodynamic forces for orbit control has been attracting attention, particularly in low Earth orbit. The orbit control can be achieved by appropriately changing the state of satellite, such as attitude and shape, to meet the aerodynamic requirements. In modeling the relationship between satellite state and aerodynamic forces, conventional methods ignore the shielding caused by the nonconvexity of the satellite’s appearance. Ignoring the shielding creates a gap between the modeled and the real aerodynamic forces, resulting in poor control performance. To solve this problem, we propose an aerodynamic force modeling method that incorporates a neural network to estimate the shielding. We train the neural network using data from an aerodynamics simulator. The optimal state that not only generates the required aerodynamic forces but also improves controllability under various mechanical constraints is obtained by solving an optimization problem that incorporates the proposed aerodynamic model. We conduct numerical simulations for establishing and maintaining general circular orbit formations. The results show convergence and continuous stable control of the deputy satellite to the ideal orbit.