Deep-auto-encoder neural-networks based attitude control allocation for over-actuated spacecraft

Modern spacecraft attitude control system mostly adopt over-actuated configuration to improve overall performance. It is necessary to consider reducing the energy consumption of the over-actuated system due to the limited on-board power supply. This paper proposes a deep-auto-encoder (DAE) neural-network-based control allocation method for spacecraft attitude control. It can achieve optimal energy consumption with high control allocation accuracy. The DAE network is trained with data generated by the dynamics of actuators. The decoder-part network is a fitting of actuators kinetics, and the encoder-part conducts control allocation. The optimization function of the network is the weighted sum of energy loss and control allocation error. Numerical examples show that the proposed DAE based control allocation method possesses good performance in torque distribution with optimal energy distribution.