Rigid-flexible coupling identification and attitude control based on deep neural networks

Rigid-flexible coupling significantly affects the attitude control accuracy during attitude maneuvering of spacecraft with large flexible appendages. On-orbit identification of flexible parameters such as modal frequency is very important for attitude control. However, large errors could exist in current identification methods, especially for extreme scenarios such as large deformation or configuration transformation of spacecraft. Unlike previous studies that identified natural frequencies or modal shapes, this paper proposes a method of directly estimating the rigid-flexible coupling term based on deep neural networks (DNNs). By using the neural network (NN), the proposed identification method is independent of the specific dynamic model and has better adaptability. We establish a rigid-flexible dynamic model considering uncertainties and analyze the possibility of identifying the rigid-flexible coupling term in the attitude dynamic equation through selected system states. To verify the feasibility of training the neural networks mapping from system states to the coupling term, we design a numerical experiment. Five attitude maneuvers with different target attitude angles are conducted with the data collected for training. The experimental results show that all estimation errors of neural networks are relatively small. Based on this, we further propose an improved PD controller. We assume that the DNNs can be well trained offline, followed by online prediction of the rigid-flexible coupling term to improve attitude control accuracy. Five random attitude maneuvers are performed for offline training, and a large angle attitude maneuver under the proposed improved PD controller is simulated. The experiment proves that the improved PD controller is effective and can improve system performance. The Elman neural network appears the best choice for rigid-flexible estimation and the improved PD controller. (C) 2021 COSPAR. Published by Elsevier B.V. All rights reserved.