Research On Ultra-Tightly Coupled GNSS/SINS Integration System Assisted By Neural Network

Global Positioning System(GPS)and the Strapdown Inertial Sensors System(SINS)integration navigation systems could provide reliable navigation solutions by overcoming each of their shortcomings;it is widely used and studied.However,there are some problems to be solved: Kalman filter requires accurate mathematical modelling of error characteristics of GPS and INS sensors for error prediction,which are difficult to model precisely.As the loosely or tightly coupled integration has no error correction for GPS's Doppler Shift,navigation performance is bad in high dynamic and weak signal circumstances.During GPS signal outages,the error of SINS will accumulate.To solve the above problems,the main research contents are as follows:To model the errors of navigation system's state equation,this paper introduces the principle and structure of SINS,then detailed establishes its position error equation,attitude error equation,velocity error equation,the error modle of gyroscopes and accelerometers.This paper also studies the postioning principle,the capturing and tracking process of GPS.The paper introduces the basic theory and training process of Artificial Neural Network(ANN).Two major neural networks,Back Propagation Neural Network(BPNN)and Radial Basis Function Neural Network(RBFNN)is studied,the simulation on the non-linear training and prediction is conducted.The analysis shows that the RBFNN is faster in leaning and more accurate than BPNN.To improve the performance of loosely and tightly coupled integration in high dynamic and weak signal circumstances,the paper adopts ultra-tight integration architecture to assist GPS receiver in tracking process.Kalman filter is suitable for solving the linear model,but ultra-tight integration architecture is deeper integration,linearized model can't reflect characteristics of integration system.So this paper studies Unscented Kalman Filter(UKF).As UKF has filtering divergence,an adaptive UKF algorithm aiding by RBFNN is designed,the simulation results show that the accuracy of proposed RBF-UKF is better than UKF,and filter divergence of UKF is solved.To solve the decreasing precision of SINS during GPS signal outages,this paper exploits the idea of incorporating RBF neural network to ultra-tight integration architecture.The input for neural network are the accelerometer and gyroscope measurements of SINS,the output is the error correction information for SINS.When GPS satellites are available,the neural network works in training mode.In case the signal of GPS satellites is blocked,the trained RBFNN is used to estimate the error correction for SINS.Results of the mathematic simulation show that ultra-tight integration system with RBFNN's assistance has less velocity and position error than integration system without RBFNN's assistance during long time GPS outages.Also the position and velocity errors of SINS growing with time are under control.