Abstract:

In classical filter algorithms, the predictive capabilities are limited by the uncertainty of system model and noise covariance. To solve this problem, Gaussian process regression and the square root unscented Kalman filter (UKF) are used in conjunction to derive a new filter algorithm. This new algorithm includes two parts. First, Gaussian process regression is used to learn training data, so the regression models and noise covariance of the dynamic system are gotten. Then the standard square root UKF filter is adopted, the state equation and observation equation are replaced by their regression models respectively, relevant noise covariance is adjusted by Gaussian kernel function adaptively and realtimely. Applied in SINS/GPS integrated navigation of an aerial vehicle, this new algorithm shows its strong adaptability to system noise, good convergence rate and excellent navigation accuracy.