Session B1b: Precise GNSS Positioning and Applications

Improving NLOS/LOS Classification Accuracy in Urban Canyon Based on Channel-Independent Patch Transformer with Temporal Information
Jiajun Chen, Jingran Wang, Shaolong Zheng, Yujie Liu, School of Automation, Guangdong University of Technology; Guangdong-HongKong-Macao, Joint Laboratory for Smart Discrete Manufacturing; Zhenni Li, (School of Automation, Guangdong University of Technology & Center for Intelligent Batch Manufacturing Based on IoT Technology; Shengli Xie, Guangdong Key Laboratory of IoT Information Technology; Qianming Wang, School of Automation, Guangdong University of Technology, & Techtotop Microelectronics Technology Co. Ltd.
Location: Beacon A

Peer Reviewed

The global navigation satellite systems (GNSS) positioning performance is significantly degraded due to the blocking of direct signals and errors caused by reflected signals in urban canyons. Recent studies have used deep learning or machine learning methods to distinguish the line-of-sight(LOS) and non-line-of-sight(NLOS) signals to solve multi-path problems. However, these approaches still face challenges. The visibility of satellites is relatively stable in a short period, but the features of their signals change in varying degrees. Existing methods focus on the visibility identification of satellites at a single epoch, which fails to capture the effect of the temporal features of satellite signals on visibility and ignores the complex associations between satellites at multiple continual epochs. To address the above challenge, this paper develops a novel channel-independent patch transformer neural network with temporal information, for improving the prediction of GNSS satellite visibility. Firstly, to capture the influence of individual satellite features on the classifications of NLOS signals, we adopt the concept of independent channels to disentangle the various satellite features. In this way, we construct temporal variations for each feature and then independently assess the effect of these feature variations on satellite visibility. Secondly, to account for the association of multiple continuous epochs satellites, we partition the constructed temporal window feature sequences into a collection of subsequences level patches. This patch-level structural design preserves the semantic associations of multiple epochs satellites while also maintaining the ability to focus across a sufficient number of epochs. Finally, based on the idea of channel independence and patch, we develop a novel channel-independent patch transformer (CIPT) neural network with temporal information for predicting satellite visibility, which can not only learn the effect of individual features on satellite visibility but also focus on the association of multiple epochs satellites. Experimental results on real-world urban canyon datasets demonstrate that our method can achieve more than 90% satellite visibility prediction accuracy, which is about 2.5%-15% higher than existing methods.