The square-root unscented Kalman filter （SR- UKF） for state estimation probably encounters the problem that Cholesky factor update of the covariance matrices can＇t be implemented when the zero＇th weight of sigm...The square-root unscented Kalman filter （SR- UKF） for state estimation probably encounters the problem that Cholesky factor update of the covariance matrices can＇t be implemented when the zero＇th weight of sigma points is negative or the mnnerical computation error becomes large during the faltering procedure. Consequently the filter becomes invalid. An improved SR-UKF algorithm （ISR- UKF） is presented for state estimation of arbitrary nonlinear systems with linear measurements. It adopts a modified form of predicted covariance matrices, and modifies the Cholesky factor calculation of the updated covariance matrix originating from the square-root covariance filtering method. Discussions have been given on how to avoid the filter invalidation and further error accumulation. The comparison between the ISR-UKF and the SR-UKF by simulation also shows both have the same accuracy for state estimation. Finally the performance of the improved filter is evaluated under the impact of model mismatch. The error behavior shows that the ISR-UKF can overcome the impact of model mismatch to a certain extent and has excellent trace capability.展开更多
基金Shanghai Commission of Science and Technology,China(No.08JC1408200)Shanghai Leading Academic Discipline Project,China(No.B504)
文摘The square-root unscented Kalman filter （SR- UKF） for state estimation probably encounters the problem that Cholesky factor update of the covariance matrices can＇t be implemented when the zero＇th weight of sigma points is negative or the mnnerical computation error becomes large during the faltering procedure. Consequently the filter becomes invalid. An improved SR-UKF algorithm （ISR- UKF） is presented for state estimation of arbitrary nonlinear systems with linear measurements. It adopts a modified form of predicted covariance matrices, and modifies the Cholesky factor calculation of the updated covariance matrix originating from the square-root covariance filtering method. Discussions have been given on how to avoid the filter invalidation and further error accumulation. The comparison between the ISR-UKF and the SR-UKF by simulation also shows both have the same accuracy for state estimation. Finally the performance of the improved filter is evaluated under the impact of model mismatch. The error behavior shows that the ISR-UKF can overcome the impact of model mismatch to a certain extent and has excellent trace capability.
