GNSS NLOS Signal Classification Based on Machine Learning and Pseudorange Residual Check

Global navigation satellite system (GNSS) positioning has recently garnered attention for autonomous driving, machine control, and construction sites. With the development of low-cost multi-GNSS receivers and the advent of new types of GNSS, such as Japan's Quasi-Zenith Satellite System, the potential of GNSS positioning has increased. New types of GNSS directly increase the number of line-of-sight (LOS) signals in dense urban areas and improve positioning accuracy. However, GNSS receivers can observe both LOS and non-line-of-sight (NLOS) signals in dense urban areas, and more NLOS signals are observed under static conditions than under dynamic conditions. The classification of LOS and NLOS signals is important, and various methods have been proposed, such as C/N0, using three-dimensional maps, fish-eye view, and GNSS/inertial navigation system integration. Multipath detection based on machine learning has also been reported in recent years. In this study, we propose a method for detecting NLOS signals using a support vector machine (SVM) classifier modeled with unique features that are calculated by receiver independent exchange format-based information and GNSS pseudorange residual check. We found that using both the SVM classifier and GNSS pseudorange residual check effectively reduced the error due to NLOS signals. Several static tests were conducted near high-rise buildings that are likely to receive some NLOS signals in downtown Tokyo. For all static tests, the percentage of positioning errors within 10 m in the horizontal positioning error was improved by >80% by detecting and eliminating satellites receiving NLOS signals.