FITAI aims to improve the accuracy of a GNSS navigation engine by using deep learning techniques. By utilizing GNSS receiver measurements and information such as SNR, post-fit residuals and other information external to the receiver (such as e.g. 3D maps, land cover, …), a deep neural network is trained to correct the positioning engine's outputs, leading to a more precise solution.
A general overview of the technique proposed in FITAI is outlined in the diagram below: a neural network accepts as inputs the post-fit residuals (and other parameters) and outputs the corrections to be applied to the position in the East, North and Up components.
The Driving Force Behind FITAI: Tackling Unmodelled GNSS Errors:
In navigation Positioning Engines (PE), post-fit residuals are the errors that remain after a solution has been estimated by the filter: they represent all the unmodelled errors that the estimated parameters could not absorb. A positioning engine typically uses measurements from satellites or other sources to determine a user's position, velocity, and time (PVT). After the solution is computed, the residuals represent the difference between the actual measurements and the predicted measurements based on the obtained solution. These residuals can indicate errors in the solution and provide valuable information for further correction. Among other sources of error that end up with a significant contribution in the post-fit residuals, one of the most common unmodelled errors in GNSS in urban environments is multi-path, which has a random nature that makes it difficult to use a model for its correction. FITAI project aims at exploring the use of deep learning techniques to use the information in the GNSS post-fit residuals to generate position corrections that can palliate the multi-path generated errors.
Project Timeline and Partnerships: The FITAI Journey:
Starting in October 2022 and scheduled for completion in March 2024, this project is part of the ESA OSIP programme. Serving as the primary gateway for pioneering concepts to join ESA, the initiative welcomes submissions in response to specific challenges or through open solicitations for innovative ideas. In the FITAI case, the purpose was to deep into a patent issued by ESA and apply it taking advantage of the vast amount of GNSS data collected and processed by Jason, Rokubun’s cloud GNSS PPK service.
Demonstration Setup: Integrating FITAI into JASON:
In the long run, the expected outcome of this activity is to use the models trained within FITAI in our service JASON, in order to improve the position obtained without the need to use CORS stations or precise products, hence the engine will not need external information to increase the accuracy of the estimated position.
Preliminary Outcomes: The Impact of Deep Learning on GNSSs:
Initial results showcase that the SPP solution can be greatly improved by using Deep Learning techniques but requires a diverse training dataset to generalize to not seen data. It is challenging to not overfit to the training data and make the Neural Network work under different conditions than the one seen during the training stage.
Future Directions: Expanding FITAI's Deep Learning Capabilities:
Different approaches are being considered to improve the current performance of FITAI, one of the most interesting being to increase the training dataset size by using Data Augmentation techniques. A key issue in every machine learning project is ensuring an adequate supply of high-quality data for training models. In the context of GNSS, a significant challenge involves replicating urban multipath effects, where GNSS signals bounce off in buildings and trees generating random patterns in the received signals.
Looking ahead, the team recognizes the importance of addressing challenges related to data availability and diversity to avoid overfitting.
Conclusion - FITAI: A Step Forward in Advanced GNSS Navigation.:
In conclusion, the FITAI project, undertaken by Rokubun as part of the ESA OSIP programme, represents a promising leap forward in enhancing the accuracy of GNSS navigation engines through the integration of deep learning techniques. By leveraging GNSS information, including SNR and post-fit residuals, this project has successfully trained a deep neural network to correct positioning engine outputs, resulting in a more precise solution. Looking ahead, the team recognizes the importance of addressing challenges related to the training datasets, the overfitting of the neural network and ensuring adaptability to diverse conditions.
