[Paper Review] Deep Learning-based Channel Estimation for Beamspace mmWave Massive MIMO Systems

Hengtao He, Chao-Kai Wen+2 more|arXiv (Cornell University)|Feb 5, 2018

Millimeter-Wave Propagation and Modeling10 references22 citations

TL;DR

This paper proposes a deep learning-based channel estimation method using a Learned Denoising-based Approximate Message Passing (LDAMP) neural network for beamspace mmWave massive MIMO systems. By treating the channel matrix as a 2D image and leveraging large-scale training data, LDAMP outperforms state-of-the-art compressed sensing algorithms, especially with few RF chains, and achieves high accuracy even at low measurement ratios, validated by both simulations and an analytical performance framework.

ABSTRACT

Channel estimation is very challenging when the receiver is equipped with a limited number of radio-frequency (RF) chains in beamspace millimeter-wave (mmWave) massive multiple-input and multiple-output systems. To solve this problem, we exploit a learned denoising-based approximate message passing (LDAMP) network. This neural network can learn channel structure and estimate channel from a large number of training data. Furthermore, we provide an analytical framework on the asymptotic performance of the channel estimator. Based on our analysis and simulation results, the LDAMP neural network significantly outperforms state-of-the-art compressed sensingbased algorithms even when the receiver is equipped with a small number of RF chains. Therefore, deep learning is a powerful tool for channel estimation in mmWave communications.

Motivation & Objective

• To address the challenge of accurate channel estimation in beamspace mmWave massive MIMO systems with limited RF chains.
• To develop a deep learning-based approach that learns channel structure from large training datasets to improve estimation accuracy.
• To provide an analytical framework for predicting LDAMP network performance without extensive Monte Carlo simulations.
• To demonstrate the practicality and robustness of the LDAMP method under low measurement ratios (few RF chains).

Proposed method

• The channel matrix is modeled as a 2D natural image and processed through the LDAMP neural network, which integrates a denoising convolutional neural network (DnCNN) into an iterative approximate message passing (AMP) framework.
• The LDAMP network uses a learned denoiser to iteratively refine channel estimates, with each layer updating the estimate based on the residual and noise variance.
• The method is trained using stochastic gradient descent with Adam optimizer on a large dataset of channel matrices generated from the Saleh-Valenzuela model.
• An analytical framework based on approximate message passing (AMP) state evolution (SE) equations is derived to predict the network’s mean squared error (MSE) performance asymptotically.
• The training data are scaled to [0,1] and generated from a realistic mmWave channel model with four propagation paths and a 64×64 lens antenna array.
• The performance is evaluated using normalized mean squared error (NMSE), and the analytical SE equations are validated against simulation results.

Experimental results

Research questions

• RQ1 Can a deep learning-based method outperform traditional compressed sensing algorithms in beamspace mmWave channel estimation with limited RF chains?
• RQ2 How does the LDAMP network’s performance scale with varying numbers of RF chains (i.e., measurement ratios)?
• RQ3 Can an analytical state evolution (SE) framework accurately predict the performance of the LDAMP network without relying on time-consuming simulations?
• RQ4 What is the impact of training data size and network architecture on the estimation accuracy of the LDAMP model?

Key findings

• The LDAMP network significantly outperforms state-of-the-art compressed sensing algorithms such as SD, SCAMPI, and D-AMP, including BM3D-AMP, across all measurement ratios.
• Even at a very low measurement ratio of δ = 0.05 (5% of RF chains), LDAMP achieves excellent NMSE performance, demonstrating its robustness with minimal hardware overhead.
• The analytical state evolution (SE) equations accurately predict the NMSE performance of the LDAMP network, enabling fast and reliable performance evaluation without Monte Carlo simulations.
• The LDAMP network converges within five layers, indicating high efficiency and practicality for real-time implementation.
• The use of large-scale training data enables the network to learn complex channel sparsity and concentration patterns, leading to superior denoising and estimation accuracy.
• The LDAMP framework is generalizable and can be applied to various selection networks and system configurations due to its end-to-end learning approach.