Machine learning based satellite compass

Abstract

This thesis documents the development, implementation, and evaluation of an ML- based Satellite Compass System customized for marine navigation. Motivated by the necessity for a low-cost, highly accurate, and reliable navigation sensor unaffected by interferences, the project addresses limitations present in conventional navigation sensors like Magnetic compass sensors and GPS-based satellite compasses. Central to the endeavor is the concept of "heading," representing a ship's direction relative to a reference point, such as magnetic or true north. Our choice of an ML-based approach stems from its potential to handle the intricate relationships inherent in sensor data more effectively than traditional methods. Throughout the project, we meticulously selected and trained machine learning models, including LSTM, Random Forest, and XGBoost, to predict gyro readings from heading data. The integration of these models facilitated a comprehensive analysis of prediction accuracy, model robustness, and generalizability. The results of our study underscore the efficacy of the ML-based Satellite Compass System. Through rigorous testing and evaluation, the XGBoost model emerged as the standout performer, demonstrating superior prediction accuracy and robustness compared to alternative models. This achievement represents a significant advancement in marine navigation technology, offering safer, more efficient, and technologically advanced solutions for maritime operations.