An LSTM-assisted 23-dimensional GNSS/INS tightly integrated system for train localization

The integrated GNSS/INS navigation system has been widely used for train positioning. However, when the train travels through areas such as tunnels and dense forests where GNSS signals are easily blocked, the positioning accuracy of the tightly coupled GNSS/INS system faces a severe challenge. To address this issue, this paper proposed an LSTM-assisted 23-dimensional GNSS/INS tight integration navigation system, which increases the scale factor and bias of the IMU into the state vector compared to the traditional 17-dimensional navigation system. Additionally, when the train is in an open environment, the LSTM is in training mode, with input being the measured information, and the precise position and velocity increments output by the Kalman filter used as the target for training. When the train operates in GNSS-limited scenarios, it switches to prediction mode, predicting accurate position and velocity increments from the current observation information to correct the system's navigation information. The superiority of the proposed model was verified by the data of the Qinghai-Tibet line collected from an actual train in 2023, and the results show that the proposed LSTM-assisted 23-dimensional model improves the horizontal maximum error in position and velocity over the traditional model by 74.13% and 64.28%.