Research On Application Of Neural Network In GNSS Navigation Location

In GNSS navigation calculation,the model which linearizes the observation equation is usually used.In complex environment,the precision of navigation solution will be affected.Kalman filter is an efficient recursive filter.In navigation and positioning,Kalman filter can make full use of the dynamic model information and observation information of the moving carrier to estimate the state of the carrier.The neural network can approximate the nonlinear mapping with any precision,and has good ability of denoising,learning and adaptive.Therefore,this paper studies the application of neural network in navigation and positioning.(1)A fixed linearize model is needed to solve the general positioning problem,which can not approach the real scene.A navigation and localization algorithm based on neural network is studied.The algorithm does not need a fixed mathematical model,and it approximates the observation equation by training a large number of data.The algorithm not only has a good ability to suppress the observation noise,but also has a good learning ability to the regular errors in the positioning equation.(2)Considering that the standard Kalman filter can not deal with the nonlinear system well,the problem of dynamic model mismatch exists in practical application,a dynamic model compensation algorithm based on neural network is proposed.Through training,the complex mapping relationship between Kalman filter gain and innovation product and dynamic error is studied,and the dynamic model of Kalman filter is compensated in the prediction stage.The simulation results show that the horizontal and vertical accuracy of the algorithm is improved by about 10% compared with the adaptive robust filter,and the network has good adaptability to various motion states.(3)Aiming at the non-linearity of observation error under the realistic condition,an observational error compensation algorithm based on neural network is proposed.By training and learning the complex mapping relationship between Kalman filter gain and innovation product and observation error,the observational error is compensated in the prediction stage.Simulation results show that compared with adaptive robust filtering,the horizontal precision and vertical precision of the algorithm are improved by 20% and 30% respectively.The measured data show that both the algorithm and the dynamic model compensation algorithm can basically eliminate the positioning deviation in all directions.Compared with the adaptive robust filtering,the standard deviation of horizontal and vertical errors is reduced by about 20% and 10% respectively.(4)Aiming at the problem of positioning error disturbance caused by abnormal observation error,the Doppler frequency shift observation is used to assist pseudorange positioning,and the DOP value is considered as the input of neural network in the dynamic model compensation algorithm based on neural network.The abnormal error is effectively suppressed and the positioning node is located.The fruit is more accurate.