Research On GNSS Observation-domain Multipath Reduction Method Based On Deep Learning

In recent years,with the realization of the global network of Beidou-3,the Beidou satellite navigation and positioning system provide real-time high-precision,all-weather positioning,navigation,and timing services for global users.Through the continuous improvement and in-depth research of the Global Navigation Satellite System(GNSS),GNSS has been profoundly applied to the high-precision deformation monitoring of ground surfaces and structures.Obtain accurate GNSS positioning results by using short baseline relative positioning in complex observation environments such as bridges,tunnels,and large buildings.Based on the double difference processing technique can weaken most of the errors in the observations,including receiver clock difference,satellite clock difference,tropospheric delay,ionospheric delay,and other station geometric distance-related errors.However,multipath errors are related to the station environment and cannot be weakened by double difference,so multipath errors become one of the main error sources in the short baseline relative positioning process.At present,more studies focus on multipath error modeling of single systems,and there is limited research on multi-path error weakening for tightly combined positioning of multi-GNSS systems.Meanwhile,traditional multipath error modeling studies suffer from problems such as time constraints of observing satellites,and equipment requirements.In this regard,the paper proposes a sidereal filtering(SF)method based on convolutional neural networks(CNN)and long and short-term memory networks(LSTM)and a study of GNSS multi-path error weakening in real time based on a convolutional neural network-gated recurrent network-enhanced multipath hemispherical map(MHM)with a single difference observation value domain,respectively.The main work and results of this paper are as follows:1)First,we extract the multipath errors of GPS,Galileo,and BDS-3,establish the MHM model of the multi-GNSS system and improve the short baseline positioning accuracy of GPS/Galileo/BDS-3 with realtime multipath error weakening of relative positioning under the tight combination of GPS/Galileo/BDS-3after fully considering several combined positioning modes.For Galileo and BDS-3 MEO satellites with repetition periods of 10 days and 7 days,there is a problem of large and complex workload and limited long-term weakening effect by using advanced sidereal filtering(ASF)to build the model,and MHM is better than ASF for the long-term weakening of multipath errors.After MHM treatment of 15 days of GPS/Galileo/BDS-3 data the model still has a significant effect on multipath error attenuation within 15 days.Moreover,the MHM model in this paper is built on a conventional receiver,which is a universal multipath error handling method.2)A combined algorithm of CNN and LSTM is proposed to train the multipath error data obtained from continuous observation,build a multipath error real-time prediction model,and then correct it by the SF method.First,the single-difference residual results of the B1 C and B2 a signals of BDS-3 were calculated,and the B1 C was obtained to be more effective than B2 a in resisting multipath errors.The RMS values of the original single-difference residuals and the single-difference residuals corrected by CNNLSTM-SF for all observable satellites of BDS-3 in short baseline relative positioning are analyzed,and the degree of improvement is demonstrated that this method can effectively correct the multipath errors in the single-difference residuals.The experimental results using the ASF model and MHM model are compared with the CNN-LSTM-SF method,and by solving the observed data for 15 consecutive days,it is obvious that the correction of multipath errors is better than ASF and MHM,and the degree of improvement remains stable over time.In addition,the CNN-LSTM-SF method is used to avoid calculating complex satellite repetition periods,and the method in this paper has obvious advantages for stellar day filtering of multi-GNSS systems.3)A convolutional neural network-gated recurrent network-enhanced MHM model based on a singledifference observation domain is proposed for multipath error weakening.The method extracts the correlation between multipath error and corresponding sky coordinates through the convolutional layer of CNN,reduces the data dimensionality by pooling layer,obtains the temporal characteristics of multi-path error sequence using GRU,and effectively predicts,the multi-path error,predicts the corresponding multipath error after obtaining the altitude and azimuth angles of satellites and corrects the observations in realtime.The experimental results show that Conv GRU-MHM is better than SF and MHM methods in correcting multipath errors in the long term for single-difference residuals of satellites;in the baseline coordinates,Conv GRU-MHM is basically more effective than SF and MHM in improving the results from306-310 days;after the time span of more than 30 days,the improvement effect of Conv GRU-MHM is more obvious.From the above results,we get that the method proposed in this paper can not only correct the multipath error effectively in real time,but also Conv GRU-MHM is more stable and reliable than SF and MHM for long-term multipath error weakening,and this method can be applied to long-term multipath error weakening.Figure 【54】 Table 【8】 Reference 【102】...