Knowledge, Innovation and Technology in Scientific Research Vol. 5

Aims: This chapter evaluates, in simulation, a gyro-free attitude-determination architecture for a guided rocket. Neural estimators provide the body-axis gravity direction and GNSS carrier-phase integer ambiguities; these estimates are combined with GNSS-derived baselines and, during terminal flight, semi-active laser (SAL) line-of-sight data. The gravity estimate is treated as one reference direction and does not by itself determine full three-axis attitude.

Study Design: A numerical study coupling a high-fidelity flight simulator with supervised learning experiments.

Place and Duration of Study: Aerospace Systems and Transport Research Group, Rey Juan Carlos University (URJC), Madrid, Spain. The satellite scenario covers one full day (24 h) above the URJC campus.

Methodology: Training data came from a MATLAB/Simulink six-degree-of-freedom (6-DoF) model of a canard-steered artillery rocket and from a GNSS signal generator. One network maps accelerometer readings to body-axis gravity (10 000 samples); a family of 35 networks recovers double-difference ambiguities for eight antennas and six satellites (576 000 epochs from 100 runs). Four experimental phases, implemented in TensorFlow/Keras, compared classification with regression, feedforward with recurrent models, and variations in window length, network size, activation, learning rate and number of epochs. The chosen estimators then fed a least-squares attitude solver, which was tested with Monte Carlo simulation and compared with a real-time kinematic (RTK) reference.

Results: Recurrent classifiers were the most accurate ambiguity estimators, reaching about 0.92 during screening and at least 99% on the internal model-development evaluation subset after final training; regression models generally produced smaller-magnitude errors when incorrect. The selected architecture used dense layers of 800, 400, 200 and 100 units plus a 100-unit tanh recurrent layer. Errors clustered both in time and across networks. Using the classification models, mean attitude RMSE values averaged across the tested command amplitudes were 0.55° in roll, 0.48° in pitch and 0.82° in yaw, numerically similar to the RTK reference values of 0.56°, 0.47° and 0.83°. Regression models gave 6.80°, 5.59° and 10.25°. The RTK comparison is descriptive and does not constitute a formal equivalence test.

Conclusion: Within the simulated conditions, learned gravity and GNSS ambiguity estimates, augmented by terminal laser-seeker vectors, produced mean attitude RMSE values below one degree when averaged across the tested command amplitudes for the classification configuration, although some high-amplitude yaw cases exceeded 1°. The findings support the architecture as a proof-of-concept for low-cost, highly dynamic GNC applications; independent holdout evaluation and hardware or flight validation remain necessary before deployment claims can be made.