Research on Neural Networks Information Fusion for Multi-sensor Measurement of Satellite Attitude Determination Based on UKF
Haiyin Zhou · Jisuanji fangzhen · 2010
The fusion weight of traditional Federal Kalman Filter is difficult to be determined because of the fusion system modeling error,the inaccuracy of noise statistic characteristics as well as the dynamic variability in the fusion filtering process.In order to solve this problem,a self-adaptive fusion estimation algorithm for multi-information measurement based on neural networks was presented,which used the self-adaptive ability of neural networks to make real-time compensation and amendment for the state fusion estimation results.Combining a nonlinear optimal estimation with neural network,an online adaptive training algorithm for the weights of neuron based on Unscented Kalman filter (UKF) was researched,which could still realize the optimal fusion for the global estimation even if the accurate covariance information of each local sub-filter were absent.The performances of UKF training algorithm and the traditional EKF algorithm were analyzed and compared,and moreover taking the multi-information fusion system for satellite attitude determination as the experimental example,the simulation calculation and analysis were advanced,which show that the presented models and algorithms are effective in the actual application.