An optimal real-time neural-network-based controller for an on-orbit service mission is proposed. The problem can be mathematically formulated as a complex optimal control problem due to the coupling nature of the orbit, the attitude, and the manipulator of the chaser spacecraft. A hierarchical optimization procedure is developed to efficiently solve the problem by recursively applying three optimization modules. The relative motion characteristics and the decoupled wrist–arm feature are used to transform the solved optimal solutions into easier learnable samples. A series of deep neural network (DNN) controllers are designed to learn from these samples. The performances of the trained network controllers are analyzed by altering the number of learned trajectories and the structure of networks. Simulation results illustrate that the designed DNN controllers can successfully guide a chaser to approach and capture an uncontrolled target with tolerable errors.

提出了一种最优的基于实时神经网络的在轨服务任务控制器。由于轨道，姿态和追赶航天器的操纵器的耦合特性，该问题可以用数学公式化为复杂的最优控制问题。通过递归应用三个优化模块，开发了一种分层优化程序来有效地解决该问题。相对运动特性和解耦的腕臂特性用于将求解的最优解转换为更容易学习的样本。设计了一系列的深度神经网络（DNN）控制器以从这些样本中学习。通过改变学习轨迹的数量和网络结构来分析训练有素的网络控制器的性能。