On-Orbit Geometric Distortion Correction on Star Images through 2D Legendre Neural Network

Abstract

The star tracker is a prerequisite device to realize high-precision attitude determination for a spacecraft. However, due to the errors in optical lens machining, optical path assembly, and temperature alternation, optical instruments suffer from some amount of optical geometric distortion, resulting in declining star tracker accuracy. The on-orbit distortion correction of star images is indispensable for precise performance. In this paper, a novel single-layer 2D Legendre neural network (2DLNN) to automatically correct the geometric distortion of the star tracker is proposed. An offline training method grounded on batch star images and an online training algorithm based on sequential star images are designed, respectively. The 2DLNN realizes the ground-based and on-orbit online correction of optical geometric distortion for the star tracker. The 2DLNN features self-learning, lifelong learning, and good adaptability. The single-layer neural network is simple, quick convergence, which is suitable for on-orbit implementation. The simulations demonstrate that the average distortion error can be reduced to less than 0.04 px after ground-based training. In the earth-orientation mode of the LEO satellite, the on-orbit sequential training algorithm can converge in 2500 star images under 1 frame/s. The proposed 2DLNN can achieve high-precision correction at the sub-pixel level, effectively improving the star tracker’s attitude determination accuracy.