AutoW: Self-Supervision Learning for Weighting Estimation in GNSS Positioning

Global Navigation Satellite System (GNSS) provides essential location-based services worldwide. Position estimation is achieved using estimation approaches such as weighted least squares and Kalman filter. For these estimation methods, the weightings of measurements contribute a lot to the robustness and accuracy of the final result. Currently, the prevalent weighting schemes are based on various statistical models, such as the C/N0-based Sigma-? model. However, the performance of these variance models could be suboptimal in the presence of unknown errors and noises. Besides the statistical models, there are differentiable estimators that could learn the covariance or weighting directly from the labeled data. However, these approaches require high-quality labeled data to perform optimally. In this paper, we proposed a self-supervised framework for weighting determination, which is called AutoW. The AutoW can learn to generate the proper weightings for the pseudorange without the need for manually labeled data. Two priors of the static experiment are used to construct the AutoW based on the differentiable factor graph optimization (DFGO): clustering and zero-velocity. The performance of AutoW is evaluated using data collected with OPPO Find X6 (Mediatek 9200). For OPPO smartphone data, AutoW demonstrates 19% and 25.1% enhancements compared to Sigma-? and equal weighting models. The results show the feasibility of the AutoW and demonstrate its robustness and effectiveness even without using the labeled data.