202 / 2023-09-13 13:31:01

DeepO-SF: A deep learning-enhanced observation-domain sidereal filtering for BDS-3 multipath mitigation

GNSS position, BDS-3, multipath mitigation, deep learning, sidereal filtering

Multipath error is the main error source limiting high-precision GNSS deformation monitoring. Based on the study of BeiDou navigation satellite system (BDS) orbit repeat periods, the multipath repeat periods of the three orbital satellites are inconsistent, and the use of advanced sidereal filtering (ASF) requires the accuracy calculation of the repeat periods, which increases the complexity of the calculation and is not sensitive to satellite orbit maneuvers. Therefore, we propose a deep learning-enhanced observation-domain sidereal filtering (DeepO-SF), in which the single difference residuals are sequentially trained by two convolutional neural networks and a long short-term memory network, and a convergent optimal multipath model is obtained after multiple parameter adjustments. In the process of real-time multipath mitigation, we can predict the current multipath and mitigate it only by several past single difference residuals. Experiments show that the proposed method can avoid the effects of calculation repetition period and satellite maneuvers and can extract more multipath frequency information (0.006-0.04Hz) than ASF and multipath hemispherical map (MHM). In the satellite single difference residuals, DeepO-SF averagely improved by 8.11% and 9.27% over ASF and MHM; in terms of positioning accuracy, the mean improvements of DeepO-SF are 11.11%, 10.91% and 8.46% higher than SF, and 9.97%, 10.42% and 7.08% higher than MHM in E\N\U directions. With the DeepO-SF method, the positioning accuracy is obviously more accurate and robust than the original, SF and MHM, and the method provides essential technical support for real-time high-precision deformation monitoring and seismic research.