Neural Filtering for Spacecraft Attitude Determinaiton and Sensor Calibration

A neural network approach to spacecraft attitude estimation and sensor calibration and alignment is proposed. This general and systematic approach will synthesize simulated attitude and sensor data into a neural filter, which is expected to be optimal for the given neural network architecture and to approach the optimal filter in performance as the given neural network size increases. The massively parallel processing capability of the neural filter makes it ideally suitable for real-time in-flight application. In the proposed work, such a neural filter will be developed for a commonly used spacecraft attitude measuring unit, which is composed of 3 single-axis gyros and 2 fixed head star trackers. The developed neural filter will be able to perform attitude estimation and sensor calibration and alignment in the presence of the gyro misalignments and drifts and the star tracker misalignments and observational errors. The performance of the neural filter will be compared with that of the extended Kalman filter in a Monte Carlo study. In view of its parallel processing capability, a demonstration of the neural filter's better performance will firmly establish the feasibility of the proposed approach.