Exploring the Potential of the Deep-Learning-Aided Kalman Filter for GNSS/INS Integration: A Study on 2-D Simulation Datasets

In the field of positioning and navigation, global navigation satellite system (GNSS) and inertial navigation system (INS) integration is a widely used approach to provide continuous and accurate navigation solutions. The most commonly used algorithm is the error-state extended Kalman filter (ES-EKF). As a model-based algorithm, the ES-EKF estimates the accumulated error in the strapdown computation based on a priori system models and noise statistics. The performance of the conventional ES-EKF is influenced by the noise model assumptions, system model deficiencies, and a variety of errors in the inertial measurement units (IMU). To improve the accuracy of GNSS/INS integration, this article proposes a hybridization of model-based and learning-based algorithms to learn the optimal Kalman gain together with the errors in the IMU measurements. The hybridization algorithm reduces the training effort while maintaining interpretability compared to learning-based algorithms. We evaluate the proposed algorithm on the simulated 2-D dataset and demonstrate its superior performance over the ES-EKF in terms of both estimated navigation solutions and IMU errors.