Deep Learning Assisted Kalman Filter for GNSS/MEMS IMU Integration in GNSS Denied Environments
Shuo Li, Bosch, Universität der Bundeswehr München; Thomas Pany, Universität der Bundeswehr München; Maxim Mikhaylov, ETH Zurich; Nikoaly Mikhaylov, Fugro Innovation & Technology
Date/Time: Thursday, Sep. 19, 1:50 p.m.
Integrating global navigation satellite system (GNSS) and inertial navigation system (INS) can enhance the overall navigation system performance, providing continues, accurate and reliable position, velocity, and attitude navigation solutions (Groves, 2013). The micro-electromechanical system (MEMS) Inertial Measurement Units (IMU) becomes attractive due to the low- cost, small size and low power consumption but tends to suffer from drift errors over time (Godha, 2006). In the GNSS/INS integration, the model-based Kalman filter (MBKF) can be used to correct the accumulated error from the strapdown computation and estimate the IMU error. The MBKF can not always properly estimate the IMU error as a result of the non-linearity of the navigation system, noise assumptions and model deficiencies (Sorenson, 1985; Khazraj et al., 2016). The GNSS outage is a challenging scenario. In order to overcome the drawbacks of the model-based (MB) algorithms and improve the performance of the integrated system, leveraging the strong learning ability of deep neural network (DNN) to the GNSS/INS field is considered (Fang et al., 2020; Tang et al., 2022; Taghizadeh and Safabakhsh, 2023). Several studies have concentrated on the hybridization of DNN and MB algorithms to utilize the respective advantages of both methodologies, such as KalmanNet (Revach et al., 2022), ProNet (Or and Klein, 2023), A-KIT (Cohen and Klein, 2024). In our previous study we applied this approach to the application of GNSS/INS integration, we evaluated the KalmanNet alongside the MBKF to compare the Kalman gain (KG) obtained with both methods and understand it deeply (Li et al., 2023c). Following that, we customized the KalmanNet structure to adapt to GNSS/INS requirements and conducted testing on a 2D simulated dataset, as well as an initial assessment on a small real dataset (Li et al., 2023b,a). In this paper we further evaluate this algorithm on simulated 3D data and large real dataset, especially in the GNSS denied environments.
For Attendees Call for Abstracts Registration Hotel Travel and Visas Exhibits Submit Kepler Nomination For Authors and Chairs Abstract Management Student Paper Awards Editorial Review Policies Publication Ethics Policies For Exhibitors Exhibitor Resource Center Marketing Resources Other Years Future Meetings Past Meetings