Neural Networks for Estimating Attitude, Line of Sight, and GNSS Ambiguity Through Onboard Sensor Fusion

Accurate estimation of attitude, line of sight (LOS), and carrier-phase ambiguity is essential for the performance of Guidance, Navigation, and Control (GNC) systems operating under highly dynamic and uncertain conditions. Traditional sensor fusion and filtering methods, although effective, often require precise modeling and high-grade sensors to maintain robustness. This paper investigates a deep learning-based estimation framework for attitude, LOS, and GNSS ambiguity through the fusion of onboard sensors-GNSS, IMU, and semi-active laser (SAL)-and remote sensing information. Two neural network estimators are developed to address the most critical components of the navigation chain: GNSS carrier-phase ambiguity and gravity-vector reconstruction in the body frame, which are integrated into a hybrid guidance and navigation scheme for attitude and LOS determination. These learning-based estimators capture nonlinear relationships between sensor measurements and physical states, improving generalization under degraded conditions. The proposed system is validated in a six-degree-of-freedom (6-DoF) simulation environment that includes full aerodynamic modeling of artillery guided rockets. Comparative analyses demonstrate that the learning-based ambiguity and gravity estimators reduce overall latency, enhance estimation accuracy, and improve guidance precision compared to conventional networks. The results suggest that deep learning-based sensor fusion can serve as a practical foundation for next-generation low-cost GNC systems, enabling precise and reliable operation in scenarios with limited observability or sensor degradation.