Attention-GANs: An Advanced GNSS Data Augmentation Method for Improved NLOS/LOS Classification

Abstract. Global Navigation Satellite System (GNSS) positioning in urban environments remains challenging due to signal obstructions and reflections caused by tall buildings, trees, and overpasses. Non-Line-of-Sight (NLOS) propagation leads to significant positioning errors, making accurate classification of Line-of-Sight (LOS) and NLOS signals essential for robust GNSS performance. Machine learning (ML) techniques have been widely explored for NLOS/LOS classification, yet their effectiveness is constrained by data imbalance, as acquiring labeled NLOS data is more challenging than LOS data. This imbalance reduces model generalization, leading to biased predictions. To address this challenge, we propose an Attention-GAN framework for synthetic GNSS data generation, coupled with a transformer-based encoder to enhance feature extraction. The proposed Attention-GAN incorporates Multi-Head Self-Attention (MHA) in both its generator and discriminator to improve the quality of generated data. Using the UrbanNav dataset, we validate our approach by training various ML classifiers on augmented data and comparing their performance against conventional methods. Experimental results demonstrate that our approach effectively mitigates data imbalance, improves classification accuracy, and enhances GNSS positioning robustness in complex urban environments.