Abstract
This study introduces a novel deep-learning-assisted Kalman filter (DLKF) for integrating global navigation satellite system (GNSS) and inertial navigation system (INS) measurements, aiming to improve navigation accuracy and robustness in GNSS-denied environments. GNSS/INS integration is critical for autonomous vehicles to maintain continuous navigation when GNSS signals are obstructed. Traditional model-based Kalman filters (MBKFs), while widely employed for GNSS/INS integration, face challenges such as system model deficiencies, incorrect noise assumptions, and complex inertial measurement unit (IMU) error characteristics, especially for micro-electromechanical system IMUs. To address these challenges, the DLKF integrates a deep neural network (DNN) with the Kalman filter, leveraging its ability to learn nonlinear system dynamics and noise statistics to provide adaptive Kalman gain and accurate IMU error estimation. Our approach further addresses practical challenges such as training with asynchronous sensor data and adapting to an error-state MBKF structure. This method also leverages DNN capabilities while retaining the interpretability of model-based filtering. In addition, we design specialized training strategies to enhance model stability and performance, particularly under GNSS outage conditions. The proposed method was evaluated on simulated three-dimensional data sets and real-world scenarios. The results demonstrate that the DLKF achieves improved navigation accuracy and reduced position errors during GNSS outages compared with the MBKF. More importantly, the proposed approach provides a consistent and interpretable framework for integrating DNNs into the well-established MBKF-based GNSS/ INS integration framework. These findings highlight the potential of combining data-driven methods with traditional model-based approaches to address the limitations of GNSS/INS integration and to provide enhanced solutions for autonomous systems in challenging environments.
