Channel Charting Resources

locating user equipment within the radio environment using channel state information

What is Channel Charting?

Channel charting learns a mapping from channel state information (CSI) to a so-called channel chart in which nearby datapoints indicate nearness in real space. In other words, the learned channel chart captures the nearby spatial geometry of the transmitting user equipments (UEs), effectively encoding relative (or logical) UE locations. Channel charting is self-supervised as the mapping from CSI to the channel chart is learned only using a database of passively collected CSI information. Such a data-driven localization approach has the advantages of being scalable and avoiding reference location information, e.g., from global navigation satellite systems (GNSSs). The self-supervised nature of channel charting also avoids the need for line-of-sight (LoS) propagation conditions or (costly) measurement campaigns, while enabling the infrastructure basestations or access points to perform cognitive and predictive radio access network (RAN) tasks which are tied to UE location.

Typical Channel Charting Pipeline

channel charting pipeline

An infrastructure basestation (BS) or access point (AP) passively collects high-dimensional CSI (describing complex-valued frequency and time coefficients at possibly multiple antennas) from a large number transmitting UEs and/or UE locations. The BS or AP then extracts CSI features, which describe large-scale fading properties contained in the collected CSI. Finally, dimensionality reduction (DR)-techniques are applied to the CSI-feature database in order to learn a low-dimensional description, which is the channel chart. The channel chart has the key property that nearby points correspond to nearby locations in real space.

Typical Channel Chart from Real-World Measurements

channel charting pipeline

The above figure shows channel charting results obtained from real-world measurements (240,000 CSI samples) acquired with a 32-antenna BS operating at 2.5 GHz. The left part shows the measurement campaign consisting of a loop-shaped path acquired over 20 minutes with four quadrants colored differently. The right part shows the resulting channel chart obtained using a triplet-loss-based neural network. One can see that local geometry is very well preserved in the channel chart. Consequently, tracking user equipments in the channel chart will enable location dependent tasks in a purely self-supervised fashion. The above figure is courtesy of Ferrand, Decurninge, Ordoñez, and Guillaud, 2021.

Publications

2018

2019

2020

2021

2022

2023

2024

Tutorials

Software

CSI Datasets

Below is a compilation of links to some real-world measurement datasets that contain CSI data. These datasets have previously been used to showcase Channel Charting:

Patents

Contact

This website is maintained by the Integrated Information Processing (IIP) Group, led by Prof. Christoph Studer, in the Department of Information Technology and Electrical Engineering at ETH Zürich, Switzerland.

Suggest a Resource

In case you would like to suggest a publication, patent, dataset, or software link, then please contact Christoph Studer and provide all the necessary details required to create an item in the above lists. Please note that we prefer open-access resources and discourage the use of papers behind a paywall.

Website last updated by Florian Euchner on Apr. 18, 2024.