Abstract
Traditional geometry-based positioning methods suffer from low accuracy in non-line-of-sight environments, whereas fingerprint-based positioning methods incur high maintenance costs due to the need for continuous database updates. Channel charting-based positioning methods adopt unsupervised learning to infer user positions from channel state information, but often suffer from low accuracy under complex propagation conditions. Channel maps enable the prediction of real-world positions by providing a priori channel information associated with the propagation environment. Building on this concept, channel charting-based positioning methods leverage this prior knowledge to enhance positioning accuracy and reduce reliance on large labeled datasets. In this paper, we propose a novel channel charting-based semi-supervised positioning algorithm with co-training, which utilizes both labeled data from the channel map and unlabeled data from practical communication systems to predict real-world geographical positions. This algorithm leverages a covariance-based channel feature and a corresponding dissimilarity metric to enhance robustness against noise and timing advance interference. Through ray tracing simulations calibrated by real-world measurements, the proposed algorithm is compared with state-of-the-art positioning methods under different noise conditions, demonstrating its effectiveness and superiority.
Funding Information
  • National Natural Science Foundation of China (62394290)
  • National Natural Science Foundation of China (62394291)
  • National Natural Science Foundation of China (62394294)
  • Major Science and Technology Project of Jiangsu Province (BG2025039)
  • National Mobile Communications Research Laboratory, Southeast University (2026A05)