| 利用LSTM网络预测月球物理天平动参数 |
| |
| 引用本文： | 魏二虎,任晓斌,刘经南,李连艳,武曙光,聂桂根.利用LSTM网络预测月球物理天平动参数[J].武汉大学学报(信息科学版),2022,47(11):1815-1822. |
| |
| 作者姓名： | 魏二虎 任晓斌 刘经南 李连艳 武曙光 聂桂根 |
| |
| 作者单位： | 1.武汉大学测绘学院, 湖北 武汉, 430079 |
| |
| 基金项目： | 国家重点研发计划2018YFC1503600国家自然科学基金41874036 |
| |
| 摘 要： | 利用中国探月甚长基线干涉测量（very long baseline interferometry, VLBI）观测数据改进月球物理天平动参数的预测精度，对于着陆器和巡视器的精密定位具有重要意义。利用VLBI单点定位模型解算得到“嫦娥三号”（Chang’E-3，CE-3）着陆器的坐标和物理天平动，分别采用循环神经网络（recursive neural network, RNN）和长短期记忆（long-short term memory, LSTM）网络进行物理天平动的预测。选取月球着陆器的坐标和VLBI观测量作为输入量，将3个欧拉角Ω, i, μ作为输出量，将11 323个样本用于训练，2 315个样本用于测试，2 315个样本用于验证，1 000个样本用作与预测结果进行对比。结果显示, 验证集的数据经过1 000次训练和9次迭代训练后的梯度约为6.2×10-5(″)/s，证明了LSTM网络与RNN的可靠性。LSTM网络和RNN的3个欧拉角的预测精度分别达到了97.8%、99.7%、97.2%和95.2%、98.5%、95.8%，LSTM网络的预测精度更高。与DE421星历对欧拉角的预测结果进行比较，结果证明了LSTM网络预测精度更高。
|
| 关 键 词： | 长短期记忆网络 循环神经网络 月球物理天平动 甚长基线干涉测量 |
| 收稿时间： | 2021-07-09 |
Prediction of Lunar Libration Parameters Using LSTM |
| |
| Affiliation: | 1.School of Geodesy and Geomatics, Wuhan University, Wuhan 430079, China2.GNSS Research Center, Wuhan University, Wuhan 430079, China3.School of Civil Engineering, Wuhan University, Wuhan 430072, China |
| |
| Abstract: | Objectives The prediction of lunar physical liberation is very important for precise positioning of lander and rover. The coordinates and physical libration of Chang'E-3 lander are calculated based on the very long baseline interferometry(VLBI) single point positioning model. Meanwhile, the parameters of the lunar libration are obtained in the 7-day arc. Methods Based on the recursive neural network(RNN), a long-short term memory(LSTM) neural network is used to establish a prediction model of the lunar libration parameters. Using the coordinates of the lunar lander and the VLBI observation as input, selecting three Euler angles as output, 2 424 samples are used for training, 519 samples are used for testing, and 519 samples are used for verification. Results The results show that the data of the verification set has a gradient of about 1 000 training and 9 iterations, and the maximum absolute error after training is only 6.2×10-5(″)/s. The experimental results show that the three Euler angles Ω、i、μ of LSTM are accurate to 97.8%, 99.7%, 97.2%, and the three Euler angles of RNN networks are accurate to 95.2%, 98.5%, 95.8%. Conclusions It is proved that the LSTM network has higher prediction accuracy than the RNN network. Compared with the predicted results of JPL DE421 ephemeris, it is proved that the LSTM network has the higher prediction accuracy. |
| |
| Keywords: | |
|
| 点击此处可从《武汉大学学报(信息科学版)》浏览原始摘要信息 |
|
点击此处可从《武汉大学学报(信息科学版)》下载全文 |
|
