Model Identification of Space Mechanisms by Using NARX Neural Network

Space mechanisms are usually affected by rigid-flexible coupling characteristics and special space environment when they are in orbit. Therefore, their models have very strong nonlinear characteristics and uncertainties. So, it is a key scientific problem that how to realize the efficient and accurate identification of in-orbit models so that they can survive and keep good performance in space. In this paper, based on flexible hub-beam unit structures in spacecraft, Nonlinear AutoRegressive models with eXogenous inputs (NARX) are used to implement the autonomous evolution of the model. A modified Lipschitz algorithm is utilized to determinate the model order in advance. Then Mini-batch Gradient Descent Method is combined with efficient Automatic Differential Algorithm to make the network parameters converge to the optimal value rapidly. Finally, a simplified First-Order Approximation Coupling dynamic model is built to simulate practical system. By comparing the response results of trained NARX model with those of dynamic model, it can be seen that the methods in the paper are able to realize the online models identification of space mechanisms efficiently and accurately.