Agile control of test mass based on PINN-DDPG for drag-free satellite.

Abstract

Agile control after the release of test mass is related to the success or failure of China's space gravitational wave detection program, such as TianQin and Taiji. In the release process, the test mass's motion state is complex and susceptible to collisions with the satellite cavity. In addition, the release capture control of the test mass uses electrostatic force, which is extremely small. These factors pose a significant challenge to the control system design. For this purpose, this paper proposes a real-time predictive control method for PINN-DDPG based on Physical Information Neural Network (PINN), Long Short-Term Memory (LSTM), and Deep Deterministic Policy Gradient (DDPG) to solve the problem of agile capture control under weak electrostatic force. First, a PINN-LSTM network for real-time state prediction is designed based on PINN and LSTM to solve the problems of interpretability and time-dependent state prediction. Subsequently, a DDPG controller was designed to solve the reinforcement learning control problem in continuous action space. Finally, simulation results demonstrate that, in comparison to the traditional PINN, the PINN-LSTM markedly hastens the training convergence, cutting the time by 60 %. Compared to traditional DDPG control, the PINN-DDPG diminish the stabilization time of position and velocity errors by 70 %.