Adaptive Neural Unscented Kalman Filter
Amit Levy, Itzik Klein
TL;DR
The paper tackles the sensitivity of nonlinear UKF-based INS/DVL fusion to time-varying process noise by introducing ANUKF, which uses ProcessNet to regress the process noise covariance in real time from inertial data. Two parallel networks (accelerometer and gyroscope) provide diagonal Q components that are integrated into the UKF, enabling robust navigation for autonomous underwater vehicles. Evaluation on real Snapir AUV data shows substantial improvements over non-adaptive UKF and an adaptive EKF, including during DVL outages, and demonstrates the feasibility of using lower-cost IMUs in high-precision fusion. The work advances practical, robust inertial sensor fusion by coupling neural estimation of process noise with a principled unscented Kalman framework.
Abstract
The unscented Kalman filter is an algorithm capable of handling nonlinear scenarios. Uncertainty in process noise covariance may decrease the filter estimation performance or even lead to its divergence. Therefore, it is important to adjust the process noise covariance matrix in real time. In this paper, we developed an adaptive neural unscented Kalman filter to cope with time-varying uncertainties during platform operation. To this end, we devised ProcessNet, a simple yet efficient end-to-end regression network to adaptively estimate the process noise covariance matrix. We focused on the nonlinear inertial sensor and Doppler velocity log fusion problem in the case of autonomous underwater vehicle navigation. Using a real-world recorded dataset from an autonomous underwater vehicle, we demonstrated our filter performance and showed its advantages over other adaptive and non-adaptive nonlinear filters.