Research On Neural Network Assisted Integrated Navigation During GNSS Outages

With the popularity of the concept of smart city and smart driving,as well as the rapid development of navigation technology,the requirements for navigation accuracy in various fields were getting higher and higher.Among them,the Inertial Navigation System based on MEMS was widely used in the military field,and the Global Navigation Satellite System was widely used because of its long-term high precision and low cost advantages.In addition,INS and GNSS have good complementary characteristics.On the one hand,INS autonomous navigation system can achieve high precision measurement in a short time;On the other hand,GNSS can maintain good accuracy for a long time,and the combination of the two can obtain better stability and higher navigation accuracy than a single navigation system.However,in integrated navigation systems,the accuracy of inertial components and the quality of GNSS satellite signals played a decisive role in navigation results.When satellite signals were interfered by high-rise buildings,trees,tunnels,multipath effects,etc.,GNSS receiver cannot receive correct navigation information and the system enters pure inertial navigation mode,resulting in the failure of Kalman filter to correctly complete state estimation.GNSS outage cause navigation to plummet.In order to improve the navigation accuracy when the satellite was blocked,two neural network aided integrated navigation models were proposed from different perspectives.Firstly,a Radial Basis Function static network-assisted integrated navigation model was designed for simple trajectory prediction and sufficient training samples.The accuracy of the algorithm in navigation was verified by a land vehicle navigation experiment.The experimental results showed that the average velocity error was less than 0.36m/s,and the average position error was less than 3.14 m,and the analysis showed that the velocity and position errors calculated by INS+RBFNN were much smaller than those calculated by INS.The experiment verified that the static network can also predict the trajectory well when the predicted trajectory was simple and the training samples were sufficient.Compared with the complex network structure,the static network can not only keep the training time of the model short,but also maintain a higher trajectory prediction accuracy in a short time.Secondly,a GWO-GRU +AKF model was proposed for complex predicted trajectories(two bends in different directions,etc.).In this model,AKF matching noise matrix can improve the filtering accuracy and reduce the influence of noise matrix mismatch on the prediction accuracy.GRU had a strong ability to extract time series information,and can accurately establish the nonlinear fitting relationship of input and output information.Grey Wolf optimization algorithm was used to fine-tune GRU network hyperparameters to solve the problem that manually setting hyperparameters was not optimal.INS,RNN+AKF,LSTM+AKF,GRU+AKF,GWO-RNN +AKF,GWO-LSTM +AKF were taken as comparison models,and verified by land vehicle navigation experiment.Experiment 1 was to predict the linear trajectory of 30 seconds,and the INS system had the largest error in predicting the easterly position.Compared with the best model GRU+AKF in the comparison model,the proposed model was 2.095 m lower.In terms of the maximum north position error,the proposed model was 0.151 m lower than the best model GWO-LSTM +AKF in the comparison model.In experiment 2,the curve trajectory with two different directions was predicted at 74 seconds.Compared with the best model GRU+AKF in the comparison model,the maximum east error and the maximum north position error of the proposed model were reduced by 2.648 m and 2.295 m.In experiment 3,the trajectory containing a curve in 140 seconds was predicted.Compared with the best model GRU+AKF model in the comparison model,the maximum east error and the maximum north position error of the proposed model were reduced by 4.358 m and 14.023 m respectively.In general,the GWO-GRU +AKF model can not only improve the prediction accuracy of complex trajectory,but also show good predictability and robustness when the trajectory was blocked for a long time.