Incremental Learning for LOS/NLOS Classification of Global Navigation Satellite System

Yuan Sun, Shang Li, Zhongliang Deng

Abstract: Global navigation satellite system (GNSS) is widely used to provide high-precision positioning services. However, the presence of non-line-of-sight (NLOS) signals can introduce significant errors in the position estimation, particularly in urban canyons. Therefore, it is necessary to detect NLOS signals and mitigate their impact on the positioning system. Existing machine learning-based approaches have shown great potential for the classification of line-of-sight (LOS) and NLOS signals, while the detection accuracy of these methods is highly dependent on the environmental conditions. Thus, it is necessary to train a classification model with diverse environmental samples to improve the generalization capability. In this paper, we propose an Incremental Learning (IL) method for LOS/NLOS classification, which adapts pre-trained deep learning networks with both the Kullback-Leibler divergence and the cross-entropy loss. Specifically, the deep learning networks are employed to extract signal and environmental features comprehensively, while the dual loss functions are applied to constrain the model parameter updates with mitigating the catastrophic forgetting. Experimental results on real-world datasets demonstrate that our proposed IL approach outperforms the pre-training, fine-tuning, and baseline machine learning methods in terms of the average classification accuracy on both new and old datasets.
Published in: Proceedings of the 36th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+ 2023)
September 11 - 15, 2023
Hyatt Regency Denver
Denver, Colorado
Pages: 231 - 244
Cite this article: Sun, Yuan, Li, Shang, Deng, Zhongliang, "Incremental Learning for LOS/NLOS Classification of Global Navigation Satellite System," Proceedings of the 36th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+ 2023), Denver, Colorado, September 2023, pp. 231-244. https://doi.org/10.33012/2023.19314
Full Paper: ION Members/Non-Members: 1 Download Credit
Sign In