Data-driven Non-linear Motion Modelling Of High-precision GNSS Reference Stations

As GNSS technology continues to develop,the accumulated observations provide a valuable data base for research in areas such as geodesy and geodynamics.The GNSS coordinate sequences are inevitably coarse and missing,and contain not only tectonic signals but also non-tectonic signals.These factors affect the accuracy and reliability of the GNSS coordinate sequences.Therefore,pre-processing the GNSS coordinate sequences,investigating the non-linear variation of the kinematic law and establishing the corresponding error correction models will help to obtain more reliable coordinate sequences,which in turn will help the separation of other signals such as period and velocity.In recent years,GNSS technology has made great progress,however,the non-linear variation of GNSS stations has yet to be further investigated due to the complexity of their motion.In this thesis,we study the nonlinear motion of GNSS coordinate sequences by modelling the trajectories based on nearly 9 years of observations from 20 GNSS reference stations in the Japanese region from January 2010 to December 2018 provided by the Nevada Laboratory as the basic data.Specifically,the main research work of this thesis is as follows.1.interpolating the missing terms present in the GNSS coordinate sequences,this thesis adopts an SVT algorithm based on the matrix filling method to implement the missing value filling for GNSS coordinate sequences.Due to the non-parametric nature of this method,human interference is greatly reduced.The method is able to interpolate the missing values of GNSS coordinate sequences of multiple stations at the same time,and can take into account the correlation between stations.The method is validated in both the original coordinate sequence and the residual sequence in terms of continuous missing ephemerides and random missing ephemerides respectively.The experimental results show that the interpolation results have RMSE values between 1-2.5mm in the horizontal direction and between 4-5mm in the U direction.2.Principal component analysis was applied to remove the common mode error existing between the stations.The overall volatility of the GNSS coordinate sequence was significantly reduced after removing the common mode error,and the reliability of the coordinate sequence was significantly improved.The noise model in the region was estimated,and it was found that the noise model was not a single noise model,and the FNWN model dominated.Analyzing the impact of the difference in noise models on the velocity of the stations and their uncertainty estimation,the experimental results showed that the velocity parameter valuation under the best noise model and the white noise model differed on average by 1.07mm/a,0.23mm/a,and0.38mm/a in the E/N/U direction respectively.0.38mm/a;and the uncertainty of the velocity under the best noise model will be about 10 times higher than that under the white noise model on average.3.Based on the principle of non-parametric function estimation of the Gaussian kernel,the GNSS residual series is modelled and the model is sparse by L1 parametrization to achieve a better description of the non-linear motion variation.The modeling of the residuals is done in terms of both model fit and model sparsity.The experimental results show that the model obtained after the convergence of the FISTA iterative algorithm is able to extract the low-frequency signals in the residual sequence with good sparsity by selecting the appropriate regularisation parameters and kernel function density parameters.Finally,the separation of the periodic and velocity signals contained in the residual sequence is effectively accomplished by using the Gaussian integration method through the best square approximation principle.