Analysis and comparison of three unsupervised learning clustering methods for GNSS multipath signals

In urban environments, the GNSS signals could be easily blocked and reflected by buildings, which leads to low positioning accuracy and discontinuity. This paper employs k-means++, Gaussian mixed clustering (GMM) and fuzzy c-means (FCM) clustering methods to separate the line-of-sight signals (LOS), multi-path and non-line-of-sight signals (NLOS). The pseudorange residuals, signal-to-noise ratio, elevation angle, as well as pseudorange rate consistency are considered in the three methods. The performance comparisons of the three methods are carried out by static testing of GPS/BDS integrated system in known position point. The results show that FCM has the greatest silhouette coefficient and the best clustering performance. The positioning performance is therefore been greatly improve by eliminating the NLOS. The root mean squares error (RMSE) of point positioning indicates that positioning results can be improved by 50% after NLOS been eliminated by k-means++ and FCM. Compared with the supervised clustering methods, the unsupervised clustering is easy to implement without prior information, has lower computation burden and has certain advantages in improving the positioning accuracy.