Machine learning based on-orbit distortion calibration technique for large field-of-view star tracker

As on-orbit missions of satellites are increasingly complicated, the demand for satellites' attitude determination accuracy is becoming higher. Star sensor is the most accurate attitude sensor on the satellite, so calibration of the star sensor is critical to further improvements of satellites' attitude determination. However, conventional on-orbit method which take advantage of inter-star cosine angle to solve the calibration problem of large field-of-view (FOV) star sensor was not sufficient due to the complexity of lens distortions. Motivated by this observation, a novel technique for the large FOV star sensor's calibration based on machine learning theory was presented. The technique used the created features to build complex distortion models. These created features were redundant, so Principal Components Analysis (PCA) was employed to reduce the useless features. Moreover, two methods were employed to judge the efficiency and the effectiveness of the calibration algorithms. The results show that machine-leaning based on-orbit calibration method can achieve remarkable improvement when the distortion of the large FOV star sensor is relatively huge. The calibration error is less than 0.8 under the working condition in the paper. Compared with current algorithms, this algorithm can achieve a higher accuracy and is more robust.