Momentum Gradient Descent With Kinematic Prediction-Based Initialization for High-Accuracy Star Sensor Attitude Estimation

The traditional gradient descent (GD) method for attitude estimation is prone to local optimum, and the real-time performance is poor in dynamic environments. To address these issues, this article proposes an improved attitude estimation algorithm based on a star sensor, which integrates momentum GD (MGD) with kinematic prediction-based initialization (KPI). First, the momentum term is introduced to accumulate the historical gradient information, so that the iterative direction could be optimized and the iterative oscillations could be effectively suppressed. Furthermore, the initial value prediction mechanism is designed based on the continuity of spacecraft dynamics. The attitude quaternion of the previous moment is fused with the angular velocity measured by the gyroscope, and the recursive estimation of the initial value is dynamically achieved through the attitude kinematic equation. Key MGD parameters are determined via dedicated sensitivity analysis to balance convergence speed and accuracy. During the attitude maneuver mode, MGD reduces the mean absolute error (MAE) by 73.28% and the root-mean-square error (RMSE) by 72.48% compared with GD. For the attitude stability control mode, MGD decreases the MAE by 72.20% and the RMSE by 72.09% relative to GD. Furthermore, in contrast to MGD with temporal recursive initialization (TRI), the proposed MGD-KPI further reduces the MAE by 79.86% and the RMSE by 79.76% while achieving the shortest execution time. The simulation results demonstrate that the proposed MGD with the KPI algorithm achieves significant improvements in accuracy and computational efficiency for attitude estimation.