Abstract: The predicted residual vectors should be zero-mean Gaussian white noise, which is the precondition for multiple fading factors adaptive filtering algorithm based on statistical information in GPS/INS integration system. However the abnormalities in observations will affect the distribution of the residual vectors. In this paper, a neural network aided adaptive unscented Kalman filter (UKF) algorithm with multiple fading factors based on singular value decomposition(SVD) is proposed. The algorithm uses the neural network algorithm to weaken the influence of the observed abnormalities on the residual vectors. Singular value decomposition instead of unscented transformation is adopted to suppress negative definite variation in priori covariance matrix of UKF. Since single fading factor in poor tracking of multiple variables has the limitation, multiple fading factors to adjust the predicted-state covariance matrix are constructed with better robustness so that each filter channel has different adjustability. Finally, vehicle measurement data are collected to validate the proposed algorithm. It shows that the neural network algorithm can prevent the observed abnormalities from affecting the distribution of the residual vectors, expanding the applied range of the adaptive algorithm. The neural network algorithm aided SVD-UKF algorithm with multiple fading factors is able to remove influences of state anomalies on condition of the observed abnormalities. The accuracy and reliability of the navigation solution can be improved by this algorithm.

Key words: GPS/INS integrated navigation, unscented Kalman filter, radial basis function neural network, multiple fading factors