AI-Enhanced Multi-Constellation Satellite Fusion for Centimeter-Level Positioning in Signal-Degraded Environments

Satellite navigation has evolved from standalone GPS into multi-constellation GNSS Satellits, enhancing coverage, availability, and precision for modern applications operating within complex environments. However, urban canyons, signal blockage, multipath effects, and limited infrastructure still degrade accuracy and reliability, constraining many safety critical systems. This study addresses this challenge by developing an AI enhanced multi-constellation fusion framework for centimeter level positioning under signal degraded conditions. The study seeks to deliver continuous, robust positioning with minimal reliance on dense ground networks. The framework integrates MEO GNSS, LEO signals, IMU inputs, and contextual features through graph neural networks and attention based fusion, supported by adaptive loss optimization. Dual frequency data from Toronto, Tokyo, Berlin, and Singapore were collected across 120 hours, comprising 45,000 labeled epochs with controlled blockage levels reaching 80 percent. Results indicate a median 3D error of 6.2 cm across all sites, rising modestly to 8.9 cm under 70 percent blockage, outperforming EKF and deep learning baselines by 42 to 57 percent. The method achieved 95 percent positioning availability within 14.8 cm and real time inference below 15 ms. These findings demonstrate that AI driven multi-constellation fusion substantially enhances urban positioning resilience. The study concludes that combining adaptive learning, constellation diversity, and edge deployment enables reliable navigation and recommends extension with IMU, 5G, and benchmarks.