Reinforcement Learning-Based Satellite Attitude Stabilization Method for Non-Cooperative Target Capturing.

Area of Science:

• Aerospace Engineering
• Artificial Intelligence
• Control Systems

Background:

• Satellites performing complex maneuvers like payload deployment or target capture experience sudden, unpredictable changes in attitude and mass properties.
• These unpredictable changes lead to satellite instability and rolling, rendering traditional attitude control methods ineffective due to their reliance on known mass parameters.

Purpose of the Study:

• To develop and evaluate a novel reinforcement learning (RL) method for re-stabilizing satellite attitude following unknown disturbances.
• To demonstrate the efficacy of RL in autonomously controlling satellite attitude in dynamic and unpredictable space environments.

Main Methods:

• A reinforcement learning approach was employed, specifically utilizing the deep Q Network (DQN) algorithm.
• A neural network model was developed to output discretized control torques for satellite attitude control.
• A satellite dynamics simulation environment was created to train the neural network, with successful stabilization serving as the training reward.

Main Results:

• The trained neural network model demonstrated the ability to re-stabilize satellite attitude after unknown disturbances through iterative training.
• In contrast, traditional Proportion Differential (PD) controllers failed to re-stabilize the satellite due to their dependence on pre-defined mass parameters.
• The RL method exhibited self-learning capabilities, intelligence, and universality in controlling satellite attitudes.

Conclusions:

• The proposed reinforcement learning method offers a robust and intelligent solution for satellite attitude control in complex, dynamic scenarios.
• This approach shows significant potential for future applications in autonomous satellite operations, particularly for missions involving unpredictable events.
• The self-learning nature of RL overcomes the limitations of traditional controllers in handling unknown parameter variations.