Abstract
Ultra-wideband (UWB) technology, with its strong penetration capability and high time resolution, has become a popular wireless positioning method widely used in fields such as smart home, industrial automation, health monitoring, and logistics tracking. However, experiments indicate that the ranging accuracy of UWB devices significantly decreases in complex scenarios. To achieve accurate positioning in complex environments, our research introduces an innovative channel state identification method as a key prerequisite for improving ranging accuracy. This paper proposes a novel Channel state identification method, CIR-SSL (Channel Impulse Response-Semi-Supervised Learning), considering the high cost of obtaining label data in practical scenarios. Firstly, the CIR of a one-dimensional time series is transformed into a two-dimensional image representation with more information by Continuous Wavelet Transform (CWT) and Gramian Angular Field (GAF). Secondly, we design a network architecture, SEWideResViT (Wide Residual Vision Transformer with Squeeze-and-Excitation), which integrates Wide residual network (WideResNet), Visual Transformer (ViT), and Squeeze-and-Excitation (SE) attention mechanism, and apply it to SoftMatch, a semi-supervised learning algorithm. The experimental results show that the classification accuracy of the CIR-SSL method based on GAF-CWT image representation and SEWideResViT network can reach 98.95% in the channel state identification task with only 60 labeled samples per class.
Funding Information
  • Fundamental Research Funds for the Central Universities Science and Technology Leading Talent Team Project (2022JBXT003)
  • Tianjin Science and Technology Plan Project (24YDLQYS00110)

This publication has 34 references indexed in Scilit: