DeepGINS: A Deep-Learning-Based Approach to Robust Virtual INS/GNSS Positioning in Non-Gaussian Noise

To improve the positioning accuracy and robustness of autonomous vehicles under the influence of non-Gaussian noise at a lower cost, this article proposes DeepGINS, a bidirectional long–short-term memory (Bi-LSTM)-based inertial navigation system (INS)/global navigation satellite system (GNSS) fusion methodology. This is the first work to combine Bi-LSTM-predicted IMU data with real-time IMU via virtual IMU (VIMU). First, a deep IMU network is trained to predict fine inertial measurements from the raw gyroscope and accelerometer outputs. These predictions are then fused with real-time IMU data through a VIMU architecture to minimize the limitations of deep learning predictions and the inability to efficiently obtain low-frequency signals. Second, an adaptive weighted average filter with dynamic parameter tuning is developed to minimize noise contamination during the VIMU data fusion process. Finally, an adaptive maximum entropy criterion iterative error state Kalman filter (AMC-IESKF) containing a Gaussian kernel function is designed to deal with nonlinear non-Gaussian noise in heterogeneous sensor fusion. The experimental results show that the proposed method improves the latitude estimation accuracy by 70% and the longitude estimation accuracy by 88% compared with the PreVINS method under non-Gaussian noise interference. In addition, the lateral position estimation accuracy of the proposed method is improved by about 58% compared to the Bi-LSTM method under GNSS rejection environment, which is more suitable for low-cost vehicle application environments.