Machine-Learning to model GNSS systems – EGNOS Performance Prediction System

Last Updated: 30/08/2024 07:41     Created at: 30/08/2024 07:41

Final Presentation of NAVISP Project EL1-035 now available:

On Wednesday, August 28th, 2024, Integricom together with GeoX, and Iguassu presented the results of the NAVISP EL1-035 project "Machine Learning to model GNSS - EGNOS Performance Prediction (EPP)”.

The European Geostationary Navigation Overlay Service (EGNOS) is a satellite-based augmentation system (SBAS) that enhances GPS performance across Europe by providing range corrections and protection levels as upper bounds of position errors. This system plays a crucial role in ensuring safety for aviation, maritime, and land-based navigation services throughout the continent. To further improve EGNOS usability, an EGNOS performance prediction (EPP) system has been developed and evaluated in the past.

The EPP project originally employed two baseline macro models: one for User Differential Range Error Index (UDREI) and another for Grid Ionospheric Vertical Error Indicator (GIVEI). These models are based on the number of Ranging and Integrity Monitoring Stations (RIMS) and Ionospheric Piercing Points, respectively. The NAVISP activity updated the EPP system with Machine learning algorithms to model the EGNOS central processing facility output, predicting ionospheric and orbit/clock residual-errors. These predictions were then translated into protection levels, enabling comparative assessment of EGNOS performance for various operational scenarios for the macro models and for the machine learning EPP.

The project adopted an agile CRISP-DM (Cross-Industry Standard Process for Data Mining) methodology, progressively increasing model complexity and feature inclusion. Key features for UDREI prediction included RIMS locations, satellite positions, geometry metrics, weather indicators, and space weather data. For GIVEI prediction, features encompassed IGP locations, IPP locations, satellite information, and both global and local space weather indicators.

Multiple machine learning algorithms were evaluated, including decision trees, gradient boosted decision trees, XGBoost, LightGBM, CatBoost, and deep learning models such as CNN and CNN+MLP. The algorithms were trained on over five years of historical data, with LightGBM ultimately selected for both UDREI and GIVEI prediction due to its superior performance and acceptable training time.

The study demonstrated that machine learning models significantly enhanced EGNOS performance prediction compared to existing macro models. Substantial improvements were observed in both the UDRE/GIVEI domains and in predicting protection levels and availability. However, the study also indicated potential for further refinement of the models and highlighted opportunities for improving computational efficiency in future iterations of the system.

These findings underscore the potential of machine learning techniques in enhancing satellite-based augmentation systems performance assessment and predictions, paving the way for more accurate and reliable navigation services in various sectors.

The project was carried out in the scope of NAVISP Element 1, which is dedicated to technology innovation of the European industry in the wide PNT sector.

More detailed information can be found in the slides of the Final Presentation.