Q-Learning-Based Noise Covariance Matrices Adaptation in Kalman Filter for Inertial Navigation

Abstract

Velocity estimation of a rigid body using measurements from low-cost inertial and magnetic sensors plays an important role in various applications. The extended Kalman filter (EKF) is widely used for this purpose. However, EKF's estimation performance relies on the knowledge of process and measurement noise covariance matrices, and this information is generally unavailable. In this work, we introduce a solution that combines two techniques: the generation of velocity pseudo-measurements using a Bidirectional Long Short-Term Memory (BiLSTM) network, and the Q-learning method for online adaptation of noise covariance matrices. The performance of the proposed solution is validated using real experimental datasets, demonstrating that Q-learning can select appropriate noise covariance matrices to enhance velocity estimation.