Neural Network-Based GNSS Code Measurement De-Weighting for Multipath Mitigation

This paper proposes a multi-layer perceptron (MLP)-based neural network (NN) to estimate the variance of code measurements over time for multipath de-weighting purposes. Particularly, we aim to estimate the variance of a short window of code measurements using an MLP and use the output to de-weight measurements affected by multipath. The proposed solution can be widely used in various positioning techniques. The proposed method is trained and tested using a suburban land vehicle data set that includes segments of urban canyon environments. Positions are estimated using the standard Single Point Positioning (SPP) least-squares solution with a conventional residual test approach. Furthermore, the proposed method is compared with an expert model for variance estimation for multipath de-weighting and the standard elevation-dependent variance estimation model. We demonstrate that the MLP-based model can use NN to substitute human experience in the empirical variance model and provide a better positioning performance in both suburban and light-urban environments compared with the conventional code variance model and the empirical, or expert, variance model for multipath de-weighting.