Machine learning based GNSS signal classification and weighting scheme design in the built environment: a comparative experiment

Abstract None-Line-of-Sight (NLOS) signals denote Global Navigation Satellite System (GNSS) signals received indirectly from satellites and could result in unacceptable positioning errors. To meet the high mission-critical transportation and logistics demand, NLOS signals received in the built environment should be detected, corrected, and excluded. This paper proposes a cost-effective NLOS impact mitigation approach using only GNSS receivers. By exploiting more signal Quality Indicators (QIs), such as the standard deviation of pseudorange, Carrier-to-Noise Ratio (C/N 0 ), elevation and azimuth angle, this paper compares machine-learning-based classification algorithms to detect and exclude NLOS signals in the pre-processing step. The probability of the presence of NLOS is predicted using regression algorithms. With a pre-defined threshold, the signals can be classified as Line-of-Sight (LOS) or NLOS. The probability of the occurrence of NLOS is also used for signal subset selection and specification of a novel weighting scheme. The novel weighting scheme consists of both C/N 0 and elevation angle and NLOS probability. Experimental results show that the best LOS/NLOS classification algorithm is the random forest. The best QI set for NLOS classification is the first three QIs mentioned above and the difference of azimuth angle. The classification accuracy obtained from this proposed algorithm can reach 93.430%, with 2.810% false positives. The proposed signal classifier and weighting scheme improved the positioning accuracy by 69.000% and 40.700% in the horizontal direction, 79.361% and 75.322% in the vertical direction, and 75.963% and 67.824% in the 3D direction.