基于机器学习的卫星导航系统全球定位性能评估方法研究

Machine learning-based global positioning performance evaluation method of satellite navigation system

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.

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.