Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GAN-based Massive MIMO Channel Model Trained on Measured Data

Florian Euchner, Janina Sanzi, Marcus Henninger, Stephan ten Brink

TL;DR

A GAN architecture for a massive MIMO channel model is proposed, and the GAN’s generator may be viewed as a channel model that can replace conventional stochastic or raytracer-based models.

Abstract

Wireless channel models are a commonly used tool for the development of wireless telecommunication systems and standards. The currently prevailing geometry-based stochastic channel models (GSCMs) were manually specified for certain environments in a manual process requiring extensive domain knowledge, on the basis of channel measurement campaigns. By taking into account the stochastic distribution of certain channel properties like Rician k-factor, path loss or delay spread, they model the distribution of channel realizations. Instead of this manual process, a generative machine learning model like a generative adversarial network (GAN) may be used to automatically learn the distribution of channel statistics. Subsequently, the GAN's generator may be viewed as a channel model that can replace conventional stochastic or raytracer-based models. We propose a GAN architecture for a massive MIMO channel model, and train it on measurement data produced by a distributed massive MIMO channel sounder.

GAN-based Massive MIMO Channel Model Trained on Measured Data

TL;DR

A GAN architecture for a massive MIMO channel model is proposed, and the GAN’s generator may be viewed as a channel model that can replace conventional stochastic or raytracer-based models.

Abstract

Wireless channel models are a commonly used tool for the development of wireless telecommunication systems and standards. The currently prevailing geometry-based stochastic channel models (GSCMs) were manually specified for certain environments in a manual process requiring extensive domain knowledge, on the basis of channel measurement campaigns. By taking into account the stochastic distribution of certain channel properties like Rician k-factor, path loss or delay spread, they model the distribution of channel realizations. Instead of this manual process, a generative machine learning model like a generative adversarial network (GAN) may be used to automatically learn the distribution of channel statistics. Subsequently, the GAN's generator may be viewed as a channel model that can replace conventional stochastic or raytracer-based models. We propose a GAN architecture for a massive MIMO channel model, and train it on measurement data produced by a distributed massive MIMO channel sounder.
Paper Structure (22 sections, 9 equations, 8 figures, 1 table)

This paper contains 22 sections, 9 equations, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work and Contributions
  3. Purpose of a GAN-based channel model
  4. System Model and Measurement Dataset
  5. GAN-Based Channel Model
  6. Linear Interpolation Baseline
  7. Evaluation
  8. Evaluated Channel Parameters and Statistics
  9. Received Power
  10. Delay Spread
  11. Angle of Arrival
  12. Jensen-Shannon Distance
  13. GAN Evaluation Procedure
  14. Linear Interpolator Evaluation Procedure
  15. Spatial Distribution Analysis
  16. ...and 7 more sections

Figures (8)

  • Figure 1: Structural overview of the Wasserstein GAN for channel modeling. The generator $G$ learns the CSI distribution of the channel.
  • Figure 2: Different types of GAN-based channel models.
  • Figure 3: 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 with datapoint positions $\mathbf x^{(l)}$ in $\mathcal{S}_\mathrm{meas}$, colorized with the total received power $\lVert \mathbf H^{(l)} \rVert_\mathrm{F}^2$. The antenna arrays in the map are drawn to scale as black rectangles and their directivity is indicated by the green sectors. The received power is normalized such that the maximum received power is $0\,\mathrm{dB}$.
  • Figure 4: Structure of neural networks. Reshaping layers and the conversion from complex representation to separate real / imaginary parts are omitted.
  • Figure 5: Spatial distribution of received powers $\lVert \mathbf H^{(l)}_{b} \rVert_\mathrm{F}^2$, for measured dataset $\mathcal{S}_\mathrm{meas}$ and generated dataset $\mathcal{S}_\mathrm{GAN, fixed}$. The power is normalized such that the maximum received power (over all values $\lVert \mathbf H^{(l)}_{b} \rVert_\mathrm{F}^2$, measured and generated) is $0\,\mathrm{dB}$.
  • ...and 3 more figures