Research And Implementation Of Integrated Navigation Algorithm Based On Deep Learning Assisting Measurement Estimation

Currently,most vehicles are all equipped with the GNSS(Global Navigation Satellite System)to give users for precise velocity and position.However,in complex urban environments,such as urban canyons,viaducts,tunnels,avenues,GNSS signal blocking will lead to positioning accuracy decline,or even loss.Therefore,GNSS is usually integrated with INS(Inertial Navigation System).GNSS/INS integrated navigation system has the advantages of low cost,low space usage and light weight.Nevertheless,due to the possible inherent errors of sensors,the accuracy of INS tends to decline over time during GNSS outage.How to improve the continuity,reliability,and precision of vehicle positioning,shorten the unnecessary time spent in the process of driving,and provide more accurate position services for vehicles performing special emergency tasks to ensure the efficient and smooth task,has become a problem that must be faced and solved.We carry out the technical research on improving the positioning performance of assisted GNSS/INS integrated navigation,and auxiliary measurement estimation predicted by deep learning is used to reduce the errors of integrated navigation.The main work completed in this paper is as follows:(1)The residual attention network is used to dynamically estimate the noise covariance of the non-holonomic constraint pseudo-observations to achieve the optimal Kalman filtering(KF)fusion.In order to emphasize more representative features with greater weights to estimate measurement noise more accurately,this paper introduces an attention mechanism which automatically allocates different weights according to the learned feature contributions.We assess our proposed model on practical road datasets and contrast with other seven techniques including the traditional KF,Pure INS,KF with three deep learning networks,K-means,and the Input-Delayed Neural Networks based method.The proposed algorithm achieves 39.39%,30.65%,56.57%positioning accuracy improvement over traditional kalman filter method when the GNSS outage durations are 10s,30s and 60s,respectively.Several experimental results show that the proposed algorithm effectively limits the speed errors of carriers and reasonably improves the accuracy of position and velocity estimation.(2)The measurement noise covariances of the pseudo-observation in zero velocity update are updated by dynamic estimation of long short-term memory network to achieve the optimal kalman filtering fusion.We evaluate the proposed algorithm on real road datasets and compare it with traditional navigation methods and other adaptive filters.The proposed algorithm achieves 68.29%positioning accuracy improvement and 87%velocity accuracy improvement respectively over traditional kalman filter during GNSS durations in stationary.Several experiments present that the proposed algorithm limits the increase of position and speed errors when the vehicle is stationary efficiently.In view of the problem that measurement estimation depends on specific road conditions and vehicles,self-adaptive predictive auxiliary measurement method by deep learning is adopted to effectively improve the accuracy and stability of positioning accuracy of integrated navigation in complex urban environment,which has far-reaching practical application value.Meanwhile,the construction and testing of the integrated navigation simulation system based on deep learning assisted measurement estimation has been completed in this paper to verify the effectiveness of those algorithms.