Non-holonomic constraint (NHC)-assisted GNSS/SINS positioning using a vehicle motion state classification (VMSC)-based convolution neural network
Non-holonomic constraint (NHC)-assisted GNSS/SINS positioning using a vehicle motion state classification (VMSC)-based convolution neural network
复制标题
DOI：
10.1007/s10291-023-01483-9
复制
发表时间：
2023-06
期刊：
影响因子：
4.9
通讯作者：
Xin Li;Hanxu Li;Guanwen Huang;Qin Zhang;Shuolin Meng
中科院分区：
文献类型：
--
作者：
Xin Li;Hanxu Li;Guanwen Huang;Qin Zhang;Shuolin Meng
非完整约束（NHC）是一种常用的GNSS/SINS车辆定位增强技术。由于车辆在行驶过程中不可避免的侧滑、弹跳等特殊运动，以及IMU安装角度不可忽略，传统的NHC已经不能满足车辆横向和垂直速度为零的假设。最新研究表明，机器学习（ML）可以在IMU输出和NHC伪测量值之间建立映射关系，并直接预测观测域中的NHC值，从而提高NHC性能。然而，现有的基于ml的NHC预测和相应的随机模型的准确性和鲁棒性都不够，特别是对于转弯等复杂的车辆运动。此外，目前的研究主要集中在IMU-only dead estimation （DR）定位上，缺乏对nhc辅助GNSS/SINS全行程综合定位性能的可行验证。因此，我们研究了基于车辆运动状态分类（VMSC）的卷积神经网络用于NHC预测和基于NHC创新残差的方差域自适应随机模型。此外，提出了自适应nhc辅助GNSS/SINS车辆全程定位的“端到端”框架，并在城市环境下进行了车辆GNSS/SINS紧密耦合实时运动定位的验证实验。数值结果表明，基于vmsc的NHC- nets的平均预测精度可达到约2 cm/s，在GNSS中断60 s时，与仅使用imu的DR相比，使用新的NHC的定位误差降低了约91.4%；在整个行程中，GNSS/SINS的定位精度得到了显著提高，特别是在GNSS观测环境严重受限的情况下。由于DR精度较高，GNSS的歧义固定率和卫星可用性也可以在一定程度上得到提高。
Non-holonomic constraint (NHC) is a commonly used enhancement technology for GNSS/SINS vehicle positioning. Due to the vehicle's inevitable side slip, bounce, and other special movements during driving and the non-ignorable IMU installation angle, the traditional NHC can no longer satisfy the zero-value assumption for vehicle lateral and vertical velocity. The latest research demonstrates that machine learning (ML) can establish a mapping relationship between IMU output and NHC pseudo-measurements and directly predict the NHC value in the observation domain, which can improve the NHC performance. However, the existing ML-based NHC prediction and corresponding stochastic model are not sufficiently accurate and robust, especially for complex vehicle movements, such as turning. Moreover, the current studies mainly focus on IMU-only dead reckon (DR) positioning and lack feasible verification of NHC-assisted GNSS/SINS integrated positioning performance during the complete travel. Thus, we study a convolutional neural network based on a vehicle motion state classification (VMSC) for NHC prediction and an adaptive stochastic model in the variance domain using the NHC innovation residuals. Additionally, we propose an “end-to-end” framework for adaptive NHC-assisted GNSS/SINS vehicle positioning during the whole traveling, and then, the validation experiment is conducted with the vehicle GNSS/SINS tightly coupled real-time kinematic positioning in the urban environment. Numerical results show that the average predicted accuracy of VMSC-based NHC-Nets can reach approximately 2 cm/s, and by using the new NHC, the positioning error is reduced by around 91.4% compared with the IMU-only DR during a 60-s GNSS outage; for the whole traveling, the GNSS/SINS positioning accuracy achieved a significant improvement, especially in the heavily limited GNSS observation environment. Due to a relatively high DR accuracy, the GNSS ambiguity fixed rate and satellite availability can also be promoted to a certain extent.