Machine learning-based global positioning performance evaluation method of satellite navigation system
-
摘要: 针对卫星导航系统伪距相位、广播星历等观测数据，本文采用特征提取和模型回归等技术手段，从数据类型和观测时间两个维度寻找数据内在特征，挖掘出海量测站数据之间的特征关联，并采用机器学习方法评估卫星导航系统全球定位性能. 本文所提出的评估方法在实际测站数据上进行了验证，中国及周边区域12个测站模型定位精度1−平均绝对百分比误差(mean absolute percentage error, MAPE)的均值为92.36%，最差为PTGG站，1−MAPE为89.26%；全球范围120个测站模型定位精度1−MAPE的均值为86.59%，最差为SCOR站，1−MAPE为81.46%，与传统数理统计框架下得到的实测值较为吻合. 实验结果表明：基于机器学习模型评估卫星导航定位性能的方法可行有效，机器学习模型在大数据统计分析中具有强评估能力和高泛化性，突破了现仅用传统数理统计的全球定位性能评估思路.Abstract: For observation data such as pseudorange phase and broadcast ephemeris of satellite navigation systems, this paper adopts technical means such as feature extraction and model regression to find the intrinsic characteristics of the data from two dimensions of data type and observation time, excavate the feature associations between massive station data, and use machine learning methods to evaluate the global positioning performance of satellite navigation systems. The evaluation method proposed in this article has been validated on actual station data. The average positioning accuracy of 12 station models in China and surrounding areas, 1−MAPE, is 92.36%, with the worst being PTGG stations and 1−MAPE being 89.26%. The average positioning accuracy of 120 station models worldwide, 1−MAPE, is 86.59%, the worst being SCOR stations and 1−MAPE being 81.46%, which is in good agreement with the measured values obtained under the traditional mathematical statistical framework, It is shown that the method for evaluating satellite navigation and positioning performance based on machine learning models is feasible and effective. Machine learning models have strong evaluation capabilities and high generalization in big data statistical analysis, breaking through the current global positioning performance evaluation approach that only uses traditional mathematical statistics.
-
表 1 演示样本的输入特征
样本编号 Ap SN F10.7 DOY HOD TEC PDOP 1 9 16 73.3 0.87 1.07 10.01 2.60 2 4 0 74.0 −0.97 0.54 23.14 4.06 3 12 6 81.2 −0.44 1.58 35.66 2.30 
下载: 导出CSV
表 2 ENAO的训练集分布
训练集分布 Ap SN F10.7 DOY HOD TEC PDOP 均值 7.12 29.32 80.99 0.025 0.96 17.74 3.11 标准差 9.12 27.63 13.06 0.700 0.72 7.24 1.90
下载: 导出CSV
表 3 处理后的输入特征
样本编号 Ap SN F10.7 DOY HOD TEC PDOP 1 0.21 −0.48 −0.59 1.20 0.15 −1.07 −0.22 2 −0.34 −1.06 −0.53 −1.41 −0.58 0.75 0.55 3 0.53 −0.84 0.02 −0.66 0.86 2.48 −0.38
下载: 导出CSV
-
[1] 宁津生, 姚宜斌, 张小红. 全球导航卫星系统发展综述[J]. 导航定位学报, 2013, 1(1): 3-8. DOI: 10.16547/j.cnki.10-1096.2013.01.005 [2] 中国卫星导航系统管理办公室(CSNO). 北斗卫星导航系统公开服务性能规范(3.0)[S]. 2021. [3] HSU LI-TA. GNSS multipath detection using a machine learning approach[C]//IEEE 20th International Conference on Intelligent Transportation Systems(ITSC), 2017. DOI: 10.1109/ITSC.2017.8317700 [4] LINTY N, FARASIN A, FAVENZA A, et al. Detection of GNSS ionospheric scintillations based on machine learning decision tree[J]. IEEE transations on aerospace and electronic systems, 2018, 55(1): 303-317. DOI: 10.1109/TAES.2018.2850385 [5] 周相兵. GNSS数据的智能聚类学习算法研究[J]. 测绘学报, 2019, 48(8): 1072. [6] XIA Y, PAN S G, MENG X L, et al. Anomaly decetion for urban vehicle GNSS observation with a hybrid machine learning system[J]. Remote sensing, 2020, 12(6): 971. DOI: 10.3390/rs12060971 [7] KIANI S M. A precise machine learning aided algorithm for land subsidence or upheave prediction from GNSS time series[J]. arXiv: signal processing, 2020: 1-10. DOI: 10.48550/arxiv.2006.03772 [8] 骆黎明. 基于机器学习树模型的GNSS—R海面风场反演研究[D]. 北京: 中国科学院大学,2019. [9] GBDT算法原理以及实例理解(含Python代码简单实现版)[EB/OL]. (2022-08-24)[2023-04-10]. https://blog.csdn.net/wzk4869/article/details/126471404 [10] 支持向量回归(SVR)的详细介绍以及推导算法[EB/OL]. (2022-09-06)[2023-04-10]. https://cloud.tencent.com/developer/article/2095877 [11] 多层感知机(MLP)简介[EB/OL]. (2022-04-01)[2023-04-10]. https://wenku.baidu.com/view/ae6352af6b0203d8ce2f0066f5335a8102d266fa.html?_wkts_ -
点击查看大图
图(11) / 表(3)
计量
- 文章访问数: 328
- HTML全文浏览量: 153
- PDF下载量: 37
- 被引次数: 0
