Deep Reinforcement Learning with Robust Spatial-Temporal Representation for Improving GNSS Positioning Correction

In complex urban environments, where GNSS positioning is severely degraded by multipath interference and non-line-of-sight reception, data-driven methods offer a promising solution by effectively modeling complex non-Gaussian errors from sufficient data for positioning correction. The inherent spatial geometric relationships among different constellations in single-epoch GNSS observations, and the temporal dependencies exhibited in sequential multi-epoch observations, contain rich spatial-temporal information that facilitates the modeling of complex stochastic noise in GNSS measurements. However, the effective extraction and correlation of these multidimensional features from GNSS observation data have not yet been sufficiently explored in existing studies. Moreover, dynamic changes in real-world environments induce data distribution shift between training and testing, requiring generalization capability for the data-driven model in unseen scenarios. In this paper, we propose a novel deep reinforcement learning model with robust spatial-temporal representation (DRL-RSTR) for GNSS positioning correction. The spatial geometric relationships among different constellations is modeled by a graph convolutional network (GCN), and the temporal dependencies of sequential observations are captured by transformer. Then, the spatial-temporal features are fused through summation, and a cross-attention network is employed to model the interactions among multi-observations to obtain a comprehensive environmental representation. Finally, we construct a multi-observation GCN-transformer (MOGT) to encode spatial-temporal representation. Additionally, a self-supervised pretext task (SST) is introduced to improve the robustness of spatial-temporal representation against data distribution shift through consistency regularization across non-augmented and augmented observations. We conduct extensive experiments on the public GSDC and built GZGNSS datasets, results show that DRL-RSTR achieves superior positioning accuracy and generalization compared to the model-based and learning-based state-of-the-art methods, with improvements of 51.2% and 41.4% on the GZGNSS dataset and 6.5% compared with kalman filters on the GSDC dataset in terms of positioning accuracy.