Non-line-of-sight GNSS signal classification for urban navigation based on machine learning: Comparison and validation

Non-Line-Of-Sight (NLOS) observations are a challenge for satellite navigation and positioning in urban environments. Machine learning (ML)-based NLOS detection does not require 3D maps or additional hardware, and offers substantial practical advantages. In this study, the performance of different types of ML algorithms for NLOS detection, such as supervised and unsupervised algorithms, are investigated. The elevation angle, C/N0 value, pseudorange residuals as well as the difference between C/N0 observations and its nominal values are taken as the features for training and validation of different ML-based NLOS detection methods for the geodetic receiver Trimble Alloy, the low-cost receiver u-blox F9P, and the Huawei P40 mobile phone. The results show that the XGBoost has the highest successful detection rate of 98.6 % among the selected supervised learning algorithms. For the unsupervised classification algorithms, the K-means algorithm achieves the highest accuracy rate of 87.5 % and demonstrated higher computational efficiency. In addition, the models trained on low-cost devices were more universal. The static and kinematic positioning based on XGBoost and K-means for NLOS identification demonstrates that 10 % to 50 % improvement in positioning performance can be obtained. Particularly, the higher computational efficiency and no need for labelled data collection make the unsupervised algorithms more suitable for NLOS detection.