Deep Learning-based mmWave MIMO Channel Estimation using sub-6 GHz Channel Information: CNN and UNet Approaches

TL;DR

This work addresses the challenge of mmWave MIMO channel estimation under low pre-beamforming SNR by leveraging out-of-band information from sub-6 GHz. It proposes two deep-learning approaches, a CNN and a UNet, to fuse in-band mmWave data with sub-6 GHz information for improved channel estimation and spectral efficiency. Through simulations, the authors demonstrate that out-of-band aided DL methods outperform in-band only approaches and non-ML baselines, achieving median SE gains of about 3–4% over MRC, with UNet providing a modest advantage due to its ability to capture fine-grained details. The methods are designed for offline training and online deployment, with robustness to varying K-factors and SNRs, albeit with higher complexity for the UNet.

Abstract

Future wireless multiple-input multiple-output (MIMO) systems will integrate both sub-6 GHz and millimeter wave (mmWave) frequency bands to meet the growing demands for high data rates. MIMO link establishment typically requires accurate channel estimation, which is particularly challenging at mmWave frequencies due to the low signal-to-noise ratio (SNR). In this paper, we propose two novel deep learning-based methods for estimating mmWave MIMO channels by leveraging out-of-band information from the sub-6 GHz band. The first method employs a convolutional neural network (CNN), while the second method utilizes a UNet architecture. We compare these proposed methods against deep-learning methods that rely solely on in-band information and with other state-of-the-art out-of-band aided methods. Simulation results show that our proposed out-of-band aided deep-learning methods outperform existing alternatives in terms of achievable spectral efficiency.

Figures (5)

• Figure 1: A point-to-point MIMO system with co-located sub-6 GHz and MMW antenna arrays. The system geometry is characterized by the number of transmit antennas $M_{\rm Tx}^{\left( \rm b \right)}$, receive antennas $M_{\rm Rx}^{\left( \rm b \right)}$ and their mutual separation $\Delta d^{\left( \rm b \right)}$, along with the corresponding AoD $\vartheta$ and AoA $\varphi$.
• Figure 2: The CNN architecture comprises nine convolutional layers, followed by an output layer that generates the estimated channel matrix $\overline{\mathbf{H}}^{\left( {\rm m} \right)} [n]$.
• Figure 3: The UNet architecture consists of two encoders in the contraction path and two decoders in the symmetric expansion path, followed by an output layer that produces the estimated channel matrix $\overline{\mathbf{H}}^{\left( {\rm m} \right)} [n]$.
• Figure 4: The proposed out-of-band aided CNN and UNet channel estimation methods achieve the lowest channel MSE across all $K$-factors. The small vertical bars within the circular markers indicate the 95% confidence intervals.
• Figure 5: The proposed out-of-band aided CNN and UNet channel estimation methods achieve a 3.35% and 4.03% higher median SE, respectively, compared to the MRC method.