How Machine learning can help model segments of a GNSS System

Last Updated: 31/05/2023 07:40     Created at: 31/05/2023 07:39

Final Presentation of NAVISP Project EL1 035 ter now available:

On Friday, May 20th, 2023,  GMV NSL Ltd presented the results of the NAVISP EL1 035 ter project "Machine Learning Techniques to model GNSS (HARMONY)". Over 50 people from industry and research institutes followed the interesting presentation and the subsequent interactive Q&A session.

GNSS simulations are important tools to support system engineering trade-offs and decisions. They help analyse system performance and behaviour using synthetic scenarios or real data. Currently, different tools are used by industry and ESA for simulating GNSS systems, but they have limitations. One option is to use accurate but computationally intensive simulation chains that represent each part of the system separately. This approach is cumbersome and time-consuming. The other option is to use a faster macro-model-based tool, but it requires manual tuning and calibration, limiting its usability to specific scenarios. Machine learning (ML) algorithms can address macro model-based tool limitations by predicting the behaviour of GNSS system cores. Trained on real data from years of operation, these ML tools serve as intermediaries between existing simulation approaches. They combine user-friendly operation and low computational cost, similar to macro models, with improved accuracy across a wider range of scenarios compared to traditionally tuned macro models.

The primary goals of this activity were therefore to analyse GNSS applications that could benefit from machine learning and identify ML models, techniques, and processes suitable for these applications. The main focus was on designing, implementing, and validating an ML prototype. Furthermore, a comparison was conducted between ML models and classical models. 

In this context, the HARMONY Demonstrator was created with the purpose of collecting and storing GNSS data from various public sources to be utilized for ML models. The demonstrator has the capability to train ML models using the gathered data and host the developed ML models. It also features a user-friendly interface for interacting with the ML models. Additionally, the demonstrator can apply ML models to perform inference and store the predicted data. Overall, the demonstrator consists of three distinct layers: the UI layer, responsible for the graphical user interface and user interaction, the API layer, responsible for facilitating communication between the UI and the backend, and the backend layer, responsible for ML model predictions and data storage. Following, two specific use cases were identified to highlight the ML demonstrator performances: the International GNSS Service (IGS) Corrections forecast and Ionospheric STEC (Slant Total Electron Content) Modelling.

The IGS ML model in use case 1 was evaluated using a transformer ML model to improve satellite orbit and clock corrections for accurate position estimation. The demonstrator underwent training and validation using publicly available data. The developed model exhibited remarkable accuracy, achieving clock bias correction predictions <2 nanoseconds and orbit corrections within 2 meters. Notably, it outperformed other models by attaining up to 50% better accuracy in predicting clock bias for a two-hour future time horizon, along with 40-50% improved accuracies for orbit corrections. To evaluate the Iono ML model (use case 2) its performance was compared against the Klobuchar and Nequick models. The Machine Learning Demonstrator emerged as the top choice for STEC/VTEC forecast at high and mid-latitudes, with varying results for low latitudes. Key findings include the Demonstrator outperforming the NeQuick and Klobuchar models at mid-latitudes, achieving lower RMSE values and more consistent RMSE's std. At low latitudes, the NeQuick model performed closely, but the Demonstrator displayed better overall metrics such as RMSE and R2. During high ionospheric activity, the Demonstrator demonstrated superiority at high- and mid-latitudes, while performing similarly to NeQuick at low latitudes.

Based on these findings, there are promising paths for advancing ML models in the GNSS domain to improve accuracy, reliability, and real-time capabilities. These include expanding training data, incorporating more constellations and stations, considering additional parameters, and fine-tuning model parameters. Areas of potential application include scintillation forecast, interference characterization, NLOS detection and correction. Furthermore, real-time integration and evaluation of the impact of ML predictions on the PNT stage offer exciting opportunities for further development and practical implementation.

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.