LSTM-based Process Noise Covariance Prediction for AUV Navigation

The significance of process noise covariance cannot be overstated when conducting Extended Kalman Filter (EKF) for Autonomous Underwater Vehicles (AUVs). Regrettably, there currently exists no direct methods to accurately determine the values of process noise covariance in a specific AUV motion model. As a consequence, the selection of process noise covariance values is often reliant on experiential knowledge or specific experiments, leading to sub-optimal outcomes. Consequently, the implementation of EKF typically falls short of achieving optimal results. In this study, a novel approach for predicting the process noise covariance in an AUV navigation system based on a Long Short-Term Memory (LSTM) network is presented. The LSTM network takes into account the velocity measurements obtained from the Doppler Velocity Log (DVL), as well as the combined acceleration and attitude angle data obtained from the Inertial Navigation System (INS), and provides predictions for the process noise covariance. We have verified the effectiveness of the proposed algorithm through sea trials conducted in Tuandao Bay, Qingdao, China, using the Sailfish-210 AUV. The results demonstrate that the LSTM-based method outperforms the traditional EKF approach, efficaciously reducing errors in AUV position prediction and enhancing navigation accuracy.