Improving Smartphone Positioning by Adapting Measurement Noise Covariance using Machine Learning

This study presents a machine-learning-based method to improve GNSS Precise Point Positioning (PPP) accuracy by scaling the measurement covariance matrix, particularly for signals affected by multipath. The approach utilizes features such as carrier-to-noise density ratio (C/No), elevation angle, residuals, innovations, and pseudorange consistency to classify signals into line-of-sight (LOS) (good) and non-line-of-sight (NLOS) (bad) signals. Range errors derived from post-processed ground truth data serve as labels, effectively distinguishing good and bad measurements. A Random Forest model was trained using these features, achieving a classification accuracy of approximately 75%. The key innovation of this study is the use of the probability of a signal being LOS or NLOS to scale the measurement covariance matrix, enhancing the PPP solution's horizontal accuracy for smartphones. The machine-learning-assisted PPP sequential least squares (SLS) (PPP ML-SLS) method shows improvements of 51%, 37%, 32%, and 13% for the 50th, 68th, 95th percentile errors, and root mean square (rms) error, respectively, compared to traditional PPP SLS for processed datasets. Further evaluations with other datasets, demonstrates the method's consistent performance, with rms error improvements ranging from 11% to 69% at the 95th percentile. However, challenges remain due to the significant overlap between LOS and NLOS signals in terms of C/No and pseudorange consistency, highlighting the need for additional features to further enhance classification accuracy. Future work will focus on refining the feature selection process and labeling for smartphone GNSS receivers to improve accuracy and reliability across varying environments.