Enhancing GNSS Positioning Using Transformer-Based Correction Network

Global Navigation Satellite System (GNSS) is extensively used in high-precision positioning services, such as autonomous driving, automotive navigation, and unmanned aerial vehicles (UAVs). It has become an essential component of daily life and various production activities. Traditional GNSS-based systems often struggle in challenging environments like urban canyons due to multipath and line-of-sight obstructions. In recent years, deep learning has been applied to this field due to its ability to learn features from data. However, GNSS observational data require models to satisfy permutation invariance, and the positioning regression problem often faces challenges in convergence. To address these issues, we propose a novel data preprocessing method combined with a Transformer-based GNSS Positioning Correction Network (T-GPCN) to enhance GNSS positioning accuracy. Our method mitigates the need for permutation invariance by sorting and padding GNSS data, allowing for broader neural network model selection. Additionally, we transform the positioning regression problem into a classification task by discretizing positioning errors into intervals, combined with a hybrid loss function integrating regression and classification losses. Experiments using the real-world Android Raw GNSS Measurements Dataset demonstrate improvements in positioning accuracy over traditional Weighted Least Squares (WLS) methods. Specifically, the Mean Squared Error (MSE) for the T-GPCN model demonstrated a significant reduction. In the test dataset, the overall 3D MSE decreased by 39.3%. When considering the components separately, the MSE for horizontal positioning was reduced by 25.9%, and the MSE for vertical positioning saw a reduction of 42.3%. These results highlight the effectiveness of our approach in enhancing GNSS positioning accuracy.