An Intelligent GNSS Positioning Correction Method Based on Frequency-Domain State-Sequence Prediction

Abstract High-precision global navigation satellite system (GNSS) positioning is crucial for a wide range of location-based services and systems. In complex urban environments, GNSS signal obstruction and multipath interference significantly degrade accuracy. Recently, deep reinforcement learning (DRL) has shown potential in learning dynamic measurement noise patterns to correct positioning errors and thus improve accuracy. However, existing methods struggle to fully exploit the inherent structural information in observation data, primarily focusing on local temporal features, which makes it difficult to explore the advantages of long-term patterns. Frequency-domain characteristics, which reveal the underlying structural information and global periodicity of observation data, can potentially improve long-term GNSS positioning stability. In this paper, we propose a self-supervised Fourier prediction network, a frequency-augmented reinforcement learning framework that integrates frequency-domain state-sequence prediction. By transforming state sequences via a discrete-time Fourier transform, our method extracts underlying structural features to enrich the input representation. An auxiliary self-supervised network predicts these frequency-domain features, guiding the DRL agent to capture long-term trends. Experiments on the public Google Smartphone Decimeter Challenge data set and a Guangzhou GNSS data set demonstrate that our algorithm achieves a statistically significant improvement in positioning accuracy of approximately 16.1% compared with a model-based Kalman filter method and 10.9% compared with a real-time kinematic method.