Reconstruction Of Gappy GNSS Coordinate Time Series Using Bidirectional LSTM Neural Network

The coordinate time series of the GNSS base station can reflect the change of the station position with time.It is widely used in plate tectonic movement,regional deformation and geological disaster monitoring,dynamic earth reference frame establishment and maintenance,and ice back.Research on bombs,climate change,etc.In the long-term observation of the GNSS observatory,due to factors such as receiver failure,satellite anomaly and subsequent gross error elimination,the inevitable loss and discontinuity in the observation data are caused.Due to the lack of GNSS coordinate time series,it will seriously affect the correlation of data and further principal component analysis,spectrum analysis,etc.Therefore,it is very important to interpolate and delete the missing segments of GNSS coordinate time series.This paper summarizes the existing research on traditional interpolation methods,empirical orthogonal function method,singular spectrum analysis method and other time series interpolation methods.Based on the bidirectional cyclic neural network model,it is used for the GNSS of the reference station in the "China Mainland Tectonic Environment Monitoring Network".Coordinate time series interpolation completion study.In the experiment,in view of the serious data loss in each observation site,it is impossible to directly select the complete sample data of consecutive years.Therefore,the main work of this paper is as follows:Firstly,the complete coordinate time series of a single year of an observation station is taken as the experimental sample,and three sets of control experiments are set according to the missing sequence located at three different positions on the right,left and middle of the original sequence.In each of the control experiments,interpolation was performed for 3,6,9,12,and 15 days of serial deletion.Comparing the experimental results of this model with the results of traditional polynomial interpolation method,the maximum effective interpolation of the model is about 9 days,the root mean square error of the experimental results is lower,and the missing fragments can be well fitted.The local fluctuation trend,thus preliminarily verifying the accuracy and validity of the experimental model for sequence interpolation completion.Finally,based on the valid experimental model of sample data verification in a single year,multiple deletions in the multi-year observation data of the same observatory are completed,and a complete GNSS coordinate time series sample of successive years is constructed.As a result of the foregoing experiments,the size of the data sample is the main factor that restricts the accuracy of the interpolation of the experimental model,and the positional correlation with the missing fragment is small.Therefore,in the large-scale samples of consecutive years,only the missing fragments are located.The situation on the left side of the original sequence is analyzed.The experimental results show that when the GNSS coordinate time series is continuously missing for 30 days,the root mean square error of the experimental model is 2.640 mm.Comparing the experimental results of related literatures,the root mean square error of the 31-day sequence using the recurrent neural network interpolation complementing the continuous deletion is 3.429 mm.Obviously,the error level is lower,and the prediction curve can still fit the original in this experiment.Local fluctuation trend characteristics of the curve.Based on the above research and analysis,the bidirectional recurrent neural network model can be applied to the interpolation complement of GNSS coordinate time series missing segments,and has certain accuracy and effectiveness.