Abstract: Aiming at the problem of sudden changes in the attitudes encountered by satellites while performing complex tasks such as discarding a payload or capturing a target, a satellite attitude control method based on the deep reinforcement learning is proposed to restore the satellite to a stable state. Concretely, the attitude dynamics environment of the vehicle is firstly established, and the output of continuous control torque is discretized. Deep Q Network algorithm is then performed to train the autonomous attitude control of the satellite for further processing, and the optimal intelligent output of discrete behavior is rewarded with the stabilization of attitude angular velocity. Finally, the validity of the mechanism is verified by the simulation test. Results analysis illustrates that the deep reinforcement learning algorithm for satellite attitude control can stabilize satellite attitude after the satellite is disturbed by sudden random disturbance, and it can effectively solve the problem of traditional PD controller depending on the mass parameters of the controlled object. The proposed method adopts selflearning to control the satellite attitude, which has strong intelligence and universal applicability, and has a strong application potential for future intelligent control of satellites performing complex space tasks.

Key words: deep reinforcement learning, satellite attitude control, dynamic environment, autonomous attitude control, mass parameters