Deep Learning-Assisted Attitude Measurement Optimization Method in Highly Dynamic Environments

Abstract

The noncommutative errors arising from highly dynamic environments severely affect attitude errors. Conventional rotation vector algorithms struggle to adequately cope with intense motion and rapid angular variations. Thus, we propose a deep learning (DL) assisted attitude measurement optimization method (MS). The nonlinear coupling relationship between angular rates and rotation vector is deeply mined by Transformer network. Transformer uses triaxial angular rates of gyro as network input and rotation vector as network output. A state equation is formulated based on the rotation vector differential equation, while a measurement equation is devised in accordance with the Transformer prediction. Unscented Kalman filter (UKF) is adopted to realize deep fusion of Transformer and traditional rotation vector, reducing spatial noncommutative errors and enhancing overall attitude estimation performance. Simulation and experimental results demonstrate that the proposed MS reduces the root mean square error (RMSE) of pitch, roll, and yaw by at least 79.16%, 71.61%, and 85.94%, respectively.