基于改进无迹Kalman滤波的小波网络算法及其应用

WNN Algorithm Based on Improved Unscented Kalman Filter and Its Applications

无迹Kalman滤波(UKF)是无迹变换(UT)和标准Kalman滤波的结合,对非线性系统具有出色的估计性能,使用UKF估计小波网络参数,速度快,精度高,无需求导计算Jacobian矩阵,但其计算量偏大.基于此,本文考虑引入一种改进的UKF来估计小波网络的参数,以提高训练效率.该改进UKF在Kalman滤波体系内应用了一种基于最小偏度单形Sigma点采样策略的UT,它继承了UKF的优点,并显著提升了计算效率.仿真结果表明,相对于EKF,采用改进UKF算法训练小波网络,速度更快,精度更高;计算精度与UKF相当,但计算效率较之更高.

Unscented Kalman Filter（UKF）,which is a combination of Unscented Transform（UT） and standard Kalman filter,has a good estimation performance to the nonlinear system.The parameters of Wavelet Neural Network（WNN） by UKF do not need to calculate the derivative of Jacobian matrix with fast speed and high accuracy.But UKF is computationally expensive.Based on this,an improved UKF is introduced into the parameters estimation of WNN to raise the training efficiency.The improved UKF adopts an UT based on minimal skew simplex Sigma point sampling strategy in the system of Kalman filter which has the merits of UKF,and improves the computational efficiency greatly.Simulation results show that WNN based on the improved UKF has faster training speed and higher accuracy than that of EKF,and has an approximately close accuracy to that of UKF but with high computation efficiency.