Neural Network-Based 3D Velocity-Assisted GNSS/SINS Vehicle-Integrated Navigation Algorithm Research

Three-dimensional velocity-assisted GNSS/SINS positioning is a common enhancement technique for in-vehicle combined navigation and positioning,including vehicle non-integrity constraints(NHC)and odometry(ODO).Due to the inevitable special motions of vehicle body such as sideslip and bounce and the existence of mounting angle,the assumption that the lateral and vertical velocities of the traditional NHC constraint are zero no longer holds,and the complex mapping relationship between IMU output and NHC pseudo-observations can be established with machine learning to adjust the size of NHC pseudo-observations directly from the observation domain,which improves the traditional NHC constraint accuracy to some extent,but the existing NHC However,the existing NHC neural network models do not consider the motion state of the carrier,and the current three-dimensional velocity-assisted GNSS/SINS localization methods based on machine learning often separate NHC and ODO without considering the coupling relationship between them,resulting in the low accuracy of3 D velocity prediction and the inaccuracy of the stochastic model,especially in the case of complex motion such as turning and fast forward velocity of the carrier.In response to the above problems,this paper takes into account the motion state of the carrier and carries out the research of 3D velocity-assisted GNSS/SINS on-board combined navigation algorithm based on neural network,and the main research contents and contributions are as follows:(1)The effectiveness of GNSS/SINS on-board navigation enhancement algorithm and anti-difference adaptive filtering algorithm in complex environments is verified.Due to the weak GNSS satellite signals or even GNSS unavailability in urban canyons,viaducts,tunnels and other environments,the non-integrity constraint and the addition of an odometer to form a3 D velocity constraint can limit the accumulation of inertial guidance errors in the case of satellite loss of lock.The anti-difference adaptive algorithm increases or decreases the contribution of measured and predicted values in the combination by adjusting the variance expansion matrix and the adaptive factor,thus reducing the impact of inaccurate measurement model or dynamics model on the combined navigation system and improving the stability of the system.Experimental results show that the planimetric accuracy of GNSS/SINS positioning by NHC with ODO assistance is improved from 40.65 m to 3.45 m in the case of GNSS shorttime loss of lock.By adding GNSS coarse difference and inertial guidance prediction error to simulate the situation of poor GNSS observation conditions and inaccurate dynamics model,the plane accuracy of GNSS/SINS positioning with anti-difference adaptive filtering algorithm is also significantly improved compared with the traditional filtering algorithm.(2)The NHC convolutional neural network(CNN)based on motion state classification is designed to improve the accuracy of NHC prediction,especially in the case of turning,compared with the traditional method that does not consider the vehicle motion state;since the current research mainly deals with the case of missing GNSS signals,it lacks the effective verification of the combined GNSS/SINS positioning performance with full driving process NHC assistance.In this paper,we propose an end-to-end NHC-assisted full-travel GNSS/SINS combined positioning framework.For complex scenarios,mainly tunnels,high buildings,bridges,and tree blockage caused by the reduced number of GNSS satellites,serious multipath and other anomalies,the NHC adaptive adjustment of its variance domain,compared with the traditional fixed variance method,can further improve the NHC constraint effect.And the effectiveness of this paper’s method is verified by the complex urban vehicle-mounted GNSS/SINS tight combination RTK positioning.The experimental results show that the average accuracy of NHC predicted by this method is about 2 cm/s,and the IMU/NHC horizontal positioning accuracy improves by 91.42% compared with the inertial guidance projection when the simulated GNSS signal is completely missing for 60 s.During the full driving process,this method can improve the on-board GNSS/SINS positioning accuracy,especially in the heavily occluded area.It can also improve the GNSS ambiguity fixation rate and satellite availability to a certain extent due to the relatively high IMU recurrence accuracy.(3)For the existing machine learning-based NHC or ODO prediction is generally separate,without considering the coupling between the two,resulting in the prediction accuracy is not very high,the reliability of the constraint is not strong,this paper considers the influence of different forward speed,based on LSTM neural network designed to predict both NHC and ODO of 3d-nets,virtual odometer to a certain extent can replace the The virtual odometer can replace the role of odometer to a certain extent,which can reduce the cost of sensors,and realize the neural network prediction of 3D velocity-assisted GNSS/SINS on-board combined navigation.The on-board experimental results show that the average accuracy of the forward velocity predicted by this method is about 0.4 m/s,and the average accuracy of the lateral and skyward velocities is about 2 cm/s.The horizontal positioning accuracy of the 3D velocity constraint is improved by 99.40% with respect to the inertial guidance projection when the simulated GNSS signal is completely missing for 460 seconds.