Evaluation of a Gaussian Mixture Model-based Channel Estimator using Measurement Data

Nurettin Turan, Benedikt Fesl, Moritz Grundei, Michael Koller, Wolfgang Utschick

Abstract

In this work, we use real-world data in order to evaluate and validate a machine learning (ML)-based algorithm for physical layer functionalities. Specifically, we apply a recently introduced Gaussian mixture model (GMM)-based algorithm in order to estimate uplink channels stemming from a measurement campaign. For this estimator, there is an initial (offline) training phase, where a GMM is fitted onto given channel (training) data. Thereafter, the fitted GMM is used for (online) channel estimation. Our experiments suggest that the GMM estimator learns the intrinsic characteristics of a given base station's whole radio propagation environment. Essentially, this ambient information is captured due to universal approximation properties of the initially fitted GMM. For a large enough number of GMM components, the GMM estimator was shown to approximate the (unknown) mean squared error (MSE)-optimal channel estimator arbitrarily well. In our experiments, the GMM estimator shows significant performance gains compared to approaches that are not able to capture the ambient information. To validate the claim that ambient information is learnt, we generate synthetic channel data using a state-of-the-art channel simulator and train the GMM estimator once on these and once on the real data, and we apply the estimator once to the synthetic and once to the real data. We then observe how providing suitable ambient information in the training phase beneficially impacts the later channel estimation performance.

Introduction

In massive MIMO systems, accurately estimating the "channel"—the way a signal changes as it travels from a transmitter to a receiver—is crucial for performance. Mathematically, the best possible estimate depends on knowing the Probability Density Function (PDF) of the channel. However, in the real world, this PDF is unknown and analytically unreachable. Traditional methods often assume simple Gaussian distributions, ignoring "ambient information": the unique characteristics of a specific base station’s environment, such as the arrangement of buildings and streets. This study, conducted by researchers at the Technical University of Munich (TUM), evaluates whether a Gaussian Mixture Model (GMM) can implicitly learn this complex PDF from real-world data.

Methods

The authors employ a GMM to approximate the unknown channel PDF through a two-phase process:

  1. Offline Training: The Expectation-Maximization (EM) algorithm is used to fit a GMM with $K$ components to a large set of training channel samples.
  2. Online Estimation: The trained model operates as a weighted sum of Linear Minimum Mean Square Error (LMMSE) estimators. It determines which mixture component an observation likely belongs to and applies the corresponding filter.

To ensure practicality for large antenna arrays, structural constraints like Block-Toeplitz or Block-circulant structures are enforced on the GMM covariance matrices. This reduces computational complexity to $O(KN \log N)$ with minimal accuracy loss.

Experimental Setup

The study utilizes two distinct datasets to validate the model's learning capabilities:

  • Real-World Measurement Data: Collected at the Nokia campus in Stuttgart, Germany, using a 64-element antenna array (4x16 URA) at 2.18 GHz. Data was gathered from a mobile terminal moving at 25 km/h.
  • Synthetic Data: Generated using the QuaDRiGa channel simulator, configured to mimic the Nokia campus's Urban Microcell (UMi) scenario.
Measurement campaign at the Nokia campus showing LOS and NLOS paths.

Results and Discussion

The core finding is the "Ambient Information" hypothesis: a GMM acts as a universal approximator that learns the specific geometry and reflections of its environment.

Performance Comparison

The GMM estimator (with 64 components) achieved the lowest Normalized Mean Squared Error (NMSE) across the entire SNR range (-15 dB to 20 dB). It outperformed standard techniques like Least Squares (LS), Compressive Sensing (OMP), and specialized Convolutional Neural Networks (CNN).

NMSE performance on real-world measurement data comparing GMM to other estimators.

The Mismatch Proof

A critical experiment involved cross-testing: training on synthetic data and testing on real data (labeled "GMM (QuaDRiGa)" in Figure 2), and vice versa. Performance dropped significantly in both cases, proving that the GMM captures site-specific details that stochastic simulators like QuaDRiGa fail to replicate perfectly.

Spectral and Data Efficiency

The GMM approach yielded the highest achievable spectral efficiency, particularly in low-to-mid SNR regimes. While performance improves with more components ($K$), even 16 or 32 components perform remarkably well given sufficient training data (typically 100,000+ samples).

Conclusion and Future Work

The GMM-based estimator effectively learns the statistical "fingerprint" of a real-world radio environment. While highly accurate, the model is sensitive to environmental changes and requires substantial ground-truth data for training. Future work may leverage "Digital Twins"—high-fidelity 3D virtual city replicas—to generate training data that bridges the gap between synthetic models and real-world deployment.

Summary updated

Abstract

In this work, we use real-world data in order to evaluate and validate a machine learning (ML)-based algorithm for physical layer functionalities. Specifically, we apply a recently introduced Gaussian mixture model (GMM)-based algorithm in order to estimate uplink channels stemming from a measurement campaign. For this estimator, there is an initial (offline) training phase, where a GMM is fitted onto given channel (training) data. Thereafter, the fitted GMM is used for (online) channel estimation. Our experiments suggest that the GMM estimator learns the intrinsic characteristics of a given base station's whole radio propagation environment. Essentially, this ambient information is captured due to universal approximation properties of the initially fitted GMM. For a large enough number of GMM components, the GMM estimator was shown to approximate the (unknown) mean squared error (MSE)-optimal channel estimator arbitrarily well. In our experiments, the GMM estimator shows significant performance gains compared to approaches that are not able to capture the ambient information. To validate the claim that ambient information is learnt, we generate synthetic channel data using a state-of-the-art channel simulator and train the GMM estimator once on these and once on the real data, and we apply the estimator once to the synthetic and once to the real data. We then observe how providing suitable ambient information in the training phase beneficially impacts the later channel estimation performance.
BibTeX
@inproceedings{turan2022evaluation,
  author = {Nurettin Turan and Benedikt Fesl and Moritz Grundei and Michael Koller and Wolfgang Utschick},
  title = {{Evaluation of a Gaussian Mixture Model-based Channel Estimator using Measurement Data}},
  booktitle = {ISWCS},
  year = {2022},
  doi = {10.1109/iswcs56560.2022.9940363},
}
Figure