基于LSTM的空间机器人系统惯性张量在轨辨识

在空间机器人抓捕目标的过程中，整个系统的惯性张量会随时间变化且在目标被捕获瞬间发生突变，这会严重影响整体姿态控制的精度。针对以上问题，提出了一种基于长短期记忆（LSTM）的系统惯性张量在轨实时辨识方法。首先，对于目标捕获前后的2个阶段，利用拉格朗日方程建立了空间机器人的动力学模型；然后，基于所建空间机器人模型采用域随机化方法生成足量训练数据，并用其对由LSTM网络与多层全连接网络构建的参数辨识网络进行训练；最后，使用训练好的参数辨识网络对系统惯性张量进行辨识。数值仿真结果表明：所提方法能够精确辨识空间机器人抓捕过程中的系统惯性张量，所研究系统的主惯量平均相对辨识误差小于0.001，惯性积的平均相对辨识误差小于0.01。

LSTM-based on-orbit identification of inertia tensor for space robot system

The system inertia tensor of the space robot is time-varying in the process of an out-of-control target capture and even undergoes abrupt changes at the moment of capture, seriously affecting the accuracy of its overall attitude control. To address the above problem, we propose an on-orbit real-time identification method for the system inertia tensor based on Long-Short Term Memory (LSTM). According to the two stages of pre-capture and post-capture, the dynamic model of the space robot is firstly developed using the Lagrangian equation. Based on the proposed model, the domain randomization method is then adopted to generate sufficient training data to train the parameter identification network constructed by an LSTM network and a multilayer fully connected network. Finally, the trained parameter identification network is used to identify the system inertia tensor. The test results demonstrate that the proposed method can accurately identify the system inertia tensor during the capture process of the space robot. The average relative identification error of the main moment of inertia is less than 0.001, and that of the product of inertia less than 0.01.