I. Introduction
AUV plays a vital role in marine resource surveys, security and defense. Accurate navigation and positioning technology is the key to accomplishing the above tasks [1]. Extend Kalman filter (EKF) is one of the most commonly used data fusion algorithms in the field of navigation [2]. On the basis of EKF, improved kalman filters such as Unscented Kalman filter (UKF) and Cubature Kalman filter (CKF) have been developed, which have been proved to be applicable to AUV navigation to a certain extend [3] [4]. Traditional filtering algorithms have certain drawbacks in practical applications. For example, the above algorithms are based on the assumption that the measurement noise is Gaussian distribution, and accurate estimates are obtained when the measurement noise actually satisfies the Gaussian distribution, but this assumption is usually unrealistic [5]. In the AUV navigation system, the velocity of the Doppler Velocity Log (DVL) is obtained by the principle of the Doppler effect. When the AUV motion state has a sharp change, especially the attitude changes rapidly, it will affect the measurement accuracy of the DVL, and the obtained velocity information contains a lot of abnormal noise, the assumption of the Gaussian measurement noise is far from adequate for the high precision required in AUV navigation. Therefore, in this paper, we re-model the noise as Inverse-Gamma distribution and Student-t distribution, and estimate the joint posterior distribution of AUV state and improved measurement noise parameters using Variational Bayesian learning for the purpose of noise characteristic adaption. The two most commonly used nonlinear Kalman filters are improved to verify the effectiveness of the proposed algorithms in AUV navigation. It is shown experimentally that noise estimation by Variational Bayesian learning significantly improves AUV navigation performance.