GNSS localization by machine learning techniques

This thesis addresses the problem of accurately determining the location of a receiver using a machine learning-based neural network classifier exploiting ranging signals from a Global Navigation Satellite System (GNSS), such as GPS or Galileo. Traditional GNSS methods, such as Least Squares Positioning (LSP), often focus on minimizing the difference between observed and predicted satellite signals to estimate locations. While these methods work well under ideal conditions, they frequently encounter challenges in environments with significant signal interference, obstructions, or multipath effects, which can degrade their performance. The proposed solution involves developing a neural network that takes as input the positions of satellites and ranging measurements (pseudoranges) to predict the region, or tile, in which a receiver is located. The concept of ‘tiles’ refers to dividing the geographical area into smaller, distinct regions, which simplifies the classification task and allows the model to focus on predicting these predefined zones. Key contributions of this work include the development of a multi-input neural network architecture, tailored preprocessing strategies to ensure data consistency, and integration techniques to handle GNSS data variability effectively. Performance analysis focuses on evaluating the accuracy, robustness, and computational efficiency of the model across various scenarios. While the proposed approach shows promise as a scalable alternative to conventional GNSS localization methods, future research is aimed at refining the model architecture and expanding the dataset to improve real-world applicability. This thesis presents a step toward a more adaptable and precise GNSS localization framework, bridging the gap between traditional techniques and modern machine learning advancements.