APPLICATION OF NEURAL NETWORK SYSTEM TOSPACECRAFT ATTITUDE ESTIMATION

摘要

A module structured neural network (MNN) toestimate the rotational motion of a satellite isinvestigated. This system consists of NeuralNetwork for Prediction of Time Series Data (NNT)and Neural Network for Prediction of Weightsthrough NNTs (NNW), and is expected to becapable of estimating nonlinear motion. In order totrain the MNN system, sequential sets of visualinformation associated with geometrical features ofthe satellite are supplied to the MNN as input data,and the position, velocity and angular velocity of thesatellite, estimated by an Extended Kalman Filter(EKF), are used as data to train the MNN. If visualinformation cannot be obtained, the MNN functionsas a predictor of the state values. The results ofnumerical simulations show that the MNN hasexcellent robustness to noise, short convergencetime, and accuracy for estimation, in comparison toa single module NN system.