Active nonlinear vibration control of space membrane structure based on deep reinforcement learning

Abstract

To maintain the working performance of membrane spacecraft, active nonlinear vibration control of space membrane structure is a bottleneck problem of great value and research interest. The traditional model-based vibration control method usually requires a fine dynamic model which is very hard to establish in practice especially when it comes to large-amplitude nonlinear vibration. In this paper, a model-free active vibration control method for space membrane structure based on deep reinforcement learning (DRL) is presented. The proper orthogonal decomposition (POD) modal coordinates obtained from the nonlinear vibration of the space membrane structure caused by attitude maneuvering are selected as the observations. Two stayed cables are used as vibration control actuators, and the control action is applied by adjusting the tension forces in the cable actuators. A DRL agent is trained by using the deep deterministic policy gradient (DDPG) algorithm to suppress the nonlinear vibration of the space membrane structure. Simulation results show that the convergence rate of the training process for the DDPG agent can be improved significantly by choosing the low-order POD modal coordinates as observations, the DRL-based active controller can suppress the nonlinear vibration of the space membrane structure under attitude maneuvering effectively, and the DRL-based vibration controller can even out-perform the model-based controller for some cases.