Transfer Learning-Based NLOS Identification for UWB in Dynamic Obstructed Settings

Research output: Journal Publications and ReviewsRGC 21 - Publication in refereed journalpeer-review

View graph of relations

Related Research Unit(s)

Detail(s)

Original languageEnglish
Journal / PublicationIEEE Transactions on Industrial Informatics
Online published15 Nov 2023
Publication statusOnline published - 15 Nov 2023

Abstract

Positioning with ultrawideband (UWB) is prominent among industrial localization systems, due to its high-range resolution attributes and lower cost. However, one notable challenge with UWB positioning in industrial environments is the prevalent presence of nonline-of-sight (NLOS) components or signals that degrade localization performance drastically. Coupled with this, industrial settings tend to constantly change making NLOS signals challenging to characterize each time these changes occur. Recently, promising approaches have been proposed to identify NLOS components, however their performances are limited to the specific environment where measurements were performed. Their performance cannot be extended to other unknown environments due to the distribution divergence problem, owing to differences in environments captured by the channel impulse response (CIR) waveforms. This therefore requires laborious processes of data collection and training environment-specific models for NLOS identification. In this article, we propose a robust transfer learning-based NLOS identification approach, which harnesses transition information via cross-domain mappings from both source and target domains, to construct representative homogeneous features of both domains. The representative homogeneous features capture discriminative information of both domains, while reducing the distribution divergence between the domains, making it easy to classify LOS and NLOS components from both environments together. To test the robustness of our approach, we perform extensive simulations with CIR data collected from two distinct environments—“hard NLOS” (characterized by high relative permittivity of surrounding objects, e.g., thick concrete walls, metallic objects, etc.) and “soft NLOS” (characterized by low relative permittivity of surrounding objects, e.g., plasterboard walls). Our proposed approach is not just effective in transferring knowledge between distinct environments, but significantly outperforms state-of-the-art works to NLOS identification in UWB positioning networks, while reducing the laborious process of data collection in the target domain. © 2023 IEEE.

Research Area(s)

  • Data models, Feature extraction, Informatics, Localization, Location awareness, nonline-of-sight (NLOS), Nonlinear optics, Permittivity, ranging, time-of-flight (ToF), Training, transfer learning (TL)