Research On Tropospheric Delay Based On Deep Learning Algorithms

The troposphere is the lowest portion of the atmosphere,which is closely linked with human's existence.The influence of troposphere on satellite signals is the important sources of error in measurement.But the tropospheric delay is independent of the frequency of signals,so it is also difficult to eliminate this effects in dual-frequency combinations.Only through model correction method,parameter estimation method,extermal correction method can correct it.At present,the rapid development of artificial intelligence has greatly promoted the study of the artificial neural network(ANN),which is a nonlinear dynamical system.It began in the 1940s,and developed from the shallow layer of network structure to today's deep learning algorithm,playing a great role in image recognition,speech recognition,machine translation,and medical research and other aspects and have gradually penetrated into all aspects of people's life.It also provides a new method for tropospheric modeling.By considering the shallow layer of network structure with the deep learning algorithm,this paper mainly introduces the atmospheric structure and the common tropospheric estimation model in satellite navigation,proposing two models against regional precision troposphere and empirical troposphere model:regional precision troposphere model based on BP(Back Propagation)neural network and empirical regional troposphere model based on deep learning algorithm.Two aspects of primary researches are as follows:(1)Build a regional precision tropospheric delay model based on the improved BP neural network.1)To solve the problem of BP neural network applied in regional tropospheric modeling,some improvement measures are proposed:the optimal number of hidden layer nodes is obtained by numerical simulation,the model parameters are reduced,and the calculation speed is improved.2)The improved model has small parameters and is suitable for signal dissemination,which can estimate the precise troposphere in real time.3)The effect of the number of stations on the modeling accuracy is discussed.With the increase in the number of stations,the modeling accuracy increases gradually.When the number of stations reaches a certain number,the prediction accuracy of the improved BP model is not obviously improved.4)The improved BP neural network model has a large lifting precision in fitting and forecasting.The RMSE(Root Mean Square Error)is 7.83 mm and 8.52 mm respectively,while the four-parameter model fitting and the predicted RMSE is 18.03mm and 16.60 mm respectively.(2)Build a regional empirical tropospheric delay model based on deep learning.1)It is worth trying to introduce depth learning algorithm into the research work of satellite signal tropospheric delay.For the deep learning method has a strong nonlinear modeling ability,yet the tropospheric delay is nonlinear,it is very suitable to explore the distribution characteristics of tropospheric information based on abundant historical data of tropospheric delay.2)The paper using regional tropospheric data from China and Japan,modeling an empirical troposphere model based on deep learning algorithm,explores how the estimated latency compares with the real product of two different models of a single station in China and statistic analysis of Bias and RMSE results.For the small regional data but a large area in China,the RMSE was 2.26 mm larger than the GPT2w(Global Pressure and Temperature 2 Wet).3)The paper contrasted the estimated value of the single station with the real value in Japan,concluding that the precision of GPT2w model is approximately the same as that of FCN model.And the statistical analysis of Bias and RMSE is carried out on several stations,showing that the Bias of the new model is 9.15mm smaller than that of the GPT2w model,which greatly eliminates the systematic error.The RMSE of the new model is larger 0.07mm than that of the GPT2w model,whose influence on precision positioning accuracy can be neglected.4)In addition,the new model has the benefit of the low hardware precision,simple modeling process and few model parameters,which can be applied to mobile devices conveniently.