Research On GNSS/INS Integrated Navigation Algorithm Assisted By GRU Recurrent Neural Network

GNSS/INS integrated navigation system,which is composed of GNSS and INS,can make up for each other's shortcomings.The system can provide accurate navigation results in open scenes.But in complex scenes,GNSS signal is usually blocked by trees,buildings and other objects.The observation conditions of the integrated navigation system are changed greatly.In some scenarios,system even completely loses GNSS signal,and the results of integrated navigation are completely dependent on the solution of INS.Therefore,in this paper,the method of reducing navigation error in complex scenes is studied from two aspects:adaptive integrated navigation,parameter adjustment and neural network assisted integrated navigation after GNSS signal is unlocked.The main research contents are as follows:(1)Before the integrated navigation,the Nelder-Mead algorithm is used to optimize the device parameters calculated by Allan variance.The navigation error is reduced from 0.19 m to 0.13 m.(2)In order to deal with complex scenes,fading filter based on adaptive factor and Sage Husa filter are used.The former is used for the case where INS result is not accurate and GNSS observation is accurate,while the latter is used when INS result is accuracy and GNSS observation has large errors.After adding errors to INS,the experiments are carried out by using fading filter and conventional Kalman filter respectively.The horizontal error of the conventional Kalman filter is 0.99 m,and the result of fading filter decreases to 0.37 m.When the initial covariance is not set correctly,the horizontal error of Sage-Husa filter is 0.59 m,which is lower than that of the conventional Kalman filter which is 1.03 m.(3)In order to solve the problem of the complete loss of GNSS signal in more complex scenes,the integrated navigation system is assisted by the gate recurrent unit(GRU).After the GNSS signal is unlocked,the trained GRU model is used to output the pseudo position of GNSS,and Kalman filtering can be continued.When the integrated navigation is selected as loose coupling and GNSS signal is out of lock for 200 seconds,the maximum horizontal error is 87.14 m,which is 82.3% lower than the inertial navigation result which is 492.40 m.The maximum horizontal error of integrated navigation which is assisted by Multilayer Perceptron(MLP)is 406.06 m,and the results of GRU are 78.5% lower than that of the former.(4)For the complex scenes with less than 4 GNSS satellite observations in some time periods of training set,the GRU assisted tight coupling model is used to ensure better training set data,and is compared with the GRU assisted loose coupling.When GNSS signal is unlocked for 100 seconds,the maximum horizontal error of the pure inertial navigation is460.45 m,the maximum horizontal error of GRU assisted loose coupling is 124.39 m,and the maximum horizontal error of the GRU assisted tight coupling model is 44.33 m.The error of GRU assisted integrated navigation is significantly reduced,while the error of GRU assisted tight coupling is 64.4% lower than that of GRU assisted loose coupling.It is proved that using GRU assisted tight coupling in complex scenarios is better.