MACHINE LEARNING ENHANCED GYRO CALIBRATIONS BASED ON EXTENDED KALMAN FILTER

The paper proposes a novel machine learning (ML) gyro calibration method that can achieve higher accuracy gyro calibration and attitude estimation accuracy than the standard extended Kalman filter (EKF) when high-accuracy measurements are only available in a part of the orbital period. The ML calibration method does not make assumptions on the form of the sensor error model but directly learns it from data. Using a simulated torque-free satellite motion, we demonstrate the ML calibration method is effective. The results show that the gyro calibration achieves smaller residual errors compared with using standard EKF. Meanwhile, the attitude accuracy is also improved since less noise is introduced by the gyro measurement. The impact of different simulated gyro error models and the structure of the ML models are also analyzed through a series of experiments.