Channel Charting-Based Channel Prediction on Real-World Distributed Massive MIMO CSI

TL;DR

It is demonstrated that channel charting can be used to predict future CSI by exploiting spatial relationships between known estimates that are embedded in the channel chart, and is validated on a real-world distributed massive MIMO dataset.

Abstract

Distributed massive MIMO is considered a key advancement for improving the performance of next-generation wireless telecommunication systems. However, its efficacy in scenarios involving user mobility is limited due to channel aging. To address this challenge, channel prediction techniques are investigated to forecast future channel state information (CSI) based on previous estimates. We propose a new channel prediction method based on channel charting, a self-supervised learning technique that reconstructs a physically meaningful latent representation of the radio environment using similarity relationships between CSI samples. The concept of inertia within a channel chart allows for predictive radio resource management tasks through the latent space. We demonstrate that channel charting can be used to predict future CSI by exploiting spatial relationships between known estimates that are embedded in the channel chart. Our method is validated on a real-world distributed massive MIMO dataset, and compared to a Wiener predictor and the outdated CSI in terms of achievable sum rate.

Figures (5)

• Figure 1: Channel charting-based channel prediction: The UE position within the channel chart (black dots) is predicted (dashed red line) from its previous trajectory (solid red line). Delaunay triangulation generates a mesh of triangles among known channel chart positions. The triangle containing the predicted position (orange) is formed by the three samples used for CSI interpolation. Outdated UL beams are visualized in blue, and predicted DL beams in violet.
• Figure 2: Information about the environment the dataset was measured in: The figure shows (a) a photograph of the environment, (b) a top view map and (c) a scatter plot of colorized "ground truth" positions of datapoints in $\mathcal{D}_\mathrm{train}$, colorized with the measured delay spread. The antenna arrays in the map are drawn to scale as black rectangles and their viewing direction is indicated by the green sectors.
• Figure 3: The three major steps of channel charting-based channel prediction: Learning the FCF (red), predicting the UE's location within the channel chart (green), and interpolating between known CSI samples (red).
• Figure 4: Visual evaluation: The figure shows (a) the channel chart positions with preserved coloring from Fig. \ref{['fig:groundtruth-map']}, (b) the absolute latent space prediction error over the channel chart positions, and (c) the sum rate achieved with the predicted CSI from CC-interp over the channel chart positions.
• Figure 5: Average sum rate achieved with the CSI obtained by different channel prediction methods depending on the prediction horizon $p$.