Sequential learning navigation method and general correction model for Autonomous Underwater Vehicle

The errors introduced in the modeling and state estimation by conventional navigation methods will affect the AUV navigation accuracy. Therefore, this paper proposes a sequential learning method for AUV navigation. Firstly, a Temporal Convolutional Network (TCN) module is used to learn the temporal relationships from raw sensor data. After the feature representation, a Self-Attention (SA) mechanism is implemented to weigh the data at different time steps. Finally, the Fully Connected (FC) layers are adopted to regress the AUV displacements. Since the sequential learning navigation method does not require state estimation computation and can calculate the AUV position by displacement accumulation, it can avoid introducing related errors. Meanwhile, this paper proposes a general correction model based on the sequential learning method to correct the errors of various state estimation techniques. The experimental results based on actual sea trial data from Sailfish 210 AUV show that the average AUV navigation accuracy of the presented sequential learning method is improved by 56.78% compared to EKF, 57.64% compared to UKF, and 47.85% compared to CKF. Meanwhile, the navigation accuracy of the proposed general correction model is improved by 62.88% compared to EKF, 49.45% compared to UKF, and 42.86% compared to CKF.