Adaptive Attitude Maneuver Control of a Rigid–Flexible Satellite Based On Deep Reinforcement Learning
Liang Sun, Zelin Zhao, Xurui Zhao, Yu Liu
Abstract
In this paper, the attitude maneuver system is studied for a rigid-flexible satellite with a large fixed space net. The rigid-flexible satellite has significant advantages in deep space exploration, space-based power generation and space debris removal. The attitude maneuver model is formulated as a highly nonlinear and coupled system subject to parametric perturbations, external disturbances, flexible vibration and input saturation. To realize large-angle and rapid attitude maneuver control, a dual-loop adaptive attitude maneuver controller based on deep reinforcement learning (DARL) is designed, which consists of a robust observer-based compensator, a dual-loop feedback controller and an adaptive parameter regulator. The robust observer-based compensator is designed to compensate for uncertainties from multiple sources within a finite time. And the dual-loop feedback controller that accounts for input saturation is proposed for large-angle and rapid attitude maneuver. To further enhance the adaptability and robustness of the control system, an adaptive parameter regulator utilizing deep reinforcement learning is designed, which adopts a structured staged training strategy in the training process. Compared with traditional control algorithms, the simulation results demonstrate that the proposed controller not only achieves the high-precision attitude maneuver control, but also suppresses the vibration of the flexible space net effectively.