This paper proposes a novel machine learning (ML) gyro calibration method that can achieve higher-accuracy gyro calibration and attitude estimation than the classical extended Kalman filter (EKF) approach when high-accuracy measurements are unavailable. The gyro calibration process is modeled as a time-series problem based on the standard EKF output. Then, a designed ML model is trained by the collected time-series data so that it can conduct gyro calibration by generating an ML correction to the gyro measurement. The proposed ML calibration method does not make assumptions in the form of gyro measurement errors, but directly learns it from data when high-accuracy information is available. Therefore, it is possible to outperform the EKF bias calibration when there is only low-accuracy information. To validate the method, a torque-free CubeSat is simulated using sun sensors and magnetometers to generate higher- and lower-accuracy attitude measurements, respectively. The simulation results show that the ML gyro calibration achieves smaller residual errors compared with the standard EKF. Meanwhile, the EKF attitude estimation accuracy is also improved, as the attitude integration is more accurate using the ML-calibrated gyro measurement. Four ML models based on different principles are examined, including multilayer perceptron, convolutional neural network, recurrent neural network, and Gaussian processes. It is found that, usually, a more sophisticated ML model can capture more gyro error information, but all four models can achieve similar performance with well-tuned parameters.

本文提出了一种新的机器学习 (ML) 陀螺仪校准方法，当无法获得高精度测量值时，该方法可以实现比经典扩展卡尔曼滤波器 (EKF) 方法更高精度的陀螺仪校准和姿态估计。陀螺仪校准过程被建模为基于标准 EKF 输出的时间序列问题。然后，通过收集的时间序列数据训练设计的 ML 模型，以便它可以通过对陀螺仪测量生成 ML 校正来进行陀螺仪校准。所提出的 ML 校准方法不会以陀螺仪测量误差的形式进行假设，而是在有高精度信息时直接从数据中学习。因此，当只有低准确度信息时，有可能优于 EKF 偏差校准。为了验证该方法，使用太阳传感器和磁力计模拟无扭矩 CubeSat，分别生成更高和更低精度的姿态测量值。仿真结果表明，与标准 EKF 相比，ML 陀螺仪校准实现了更小的残差。同时，EKF 姿态估计精度也得到了提高，因为使用 ML 校准陀螺仪测量的姿态积分更加准确。检查了基于不同原理的四种 ML 模型，包括多层感知器、卷积神经网络、循环神经网络和高斯过程。发现，通常，更复杂的 ML 模型可以捕获更多的陀螺仪误差信息，但所有四种模型都可以通过调整良好的参数实现相似的性能。