Position Accuracy Improvement of the Inertial Navigation System using LSTM Algorithm without GNSS Signals

Currently, satellite navigation systems on cars provide a promising means of making these vehicles unmanned in the future. There is a common problem with these systems due to the unavailability of satellite signals in tunnels, forests, and noisy areas. One of the methods used to solve this problem is by using inertial navigation systems as an auxiliary system. This system works by modeling errors and correcting them when GNSS signals are absent. A number of methods are available for error modeling such as Kalman filters, neural networks, and so on, each of which has its own advantages and disadvantages. The error of inertial navigation is modeled using LSTM deep neural networks in this article. In this neural network, the relationship between current and past data is modeled as long as the GNSS satellite signal is available to improve the output position of the inertial navigation system when the GNSS satellite signal can no longer be received. The proposed method has been tested on real car driving data, and the calculated position from the inertial navigation system in four maneuvers has been compared with the Extended Kalman Filter method outputs. According to the results of the experiments, the proposed algorithm has improved the position estimation by 60% on average during 30, 60, and 120 seconds without GNSS signals, compared to the inertial navigation system based on Extended Kalman filter.