A deep learning-enhanced observation-domain multipath mitigation study of BDS-3

Abstract

Multipath is the main error source limiting high-precision Global Navigation Satellite System (GNSS) positioning. Multipath repeat periods of the three kinds of orbital satellites are inconsistent based on the study of the BeiDou navigation satellite system orbital repeat period. The use of advanced sidereal filtering (ASF) requires accurate calculation of the repeat periods for each satellite, which increases the complexity of the ASF model and is insensitive to satellite orbital maneuvers. Therefore, we propose a deep learning-enhanced (DL) observation-domain multipath mitigation method, in which the single difference residuals are trained to obtain multipath models that tend to be optimal using a convolutional neural network and long short-term memory. We obtained the predicted multipath by inputting previous single difference residuals into the multipath model in the process of real-time multipath mitigation. Experiments show that the proposed method can extract multipath information for more range frequencies (0.006–0.040 Hz) than the ASF and multipath hemispherical map (MHM) methods, and avoids the process of calculating repeat periods. In the single difference residuals, the DL method improved by 8.11 % and 9.27 % over the ASF and MHM methods; in the coordinates, the accuracy improvement of the DL method increased by 8–12 % in the E, N, and U directions compared with the ASF and MHM methods. The positioning accuracy and robustness of the DL method were better than those of the ASF and MHM methods. The proposed method provides technological support for real-time, high-precision deformation monitoring and seismic research.