RACNN: Residual Attention Convolutional Neural Network for Near-Field Channel Estimation in 6G Wireless Communications

Vu Tung Lam, Do Hai Son, Tran Thi Thuy Quynh ✉, Thanh Trung Le · 2025

Read RACNN: Residual Attention Convolutional Neural Network for Near-Field Channel Estimation in 6G Wireless Communications by Vu Tung Lam, Do Hai Son, Tran Thi Thuy Quynh ✉, Thanh Trung Le (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-02299.txt.

Introduction

Vu Tung Lam and co-authors present a 2025 network for estimating wireless channels in the near field, a task that grows hard in sixth generation systems. Huge antenna arrays make near field links possible, but they also make the signal math much heavier. Older model based methods cost too much and scale poorly, while learned methods can falter when channel conditions shift.

Their network joins convolutional layers, which read local patterns, with an attention step that focuses on key regions of those patterns. Leftover shortcut links help the network train in depth. In tests, it beat both classic and learned rivals across many settings. The gain was clearest in mixed far and near field scenes, the hardest case. The authors offer it as a robust step for future wireless systems.

Catalog record

CREATORVu Tung Lam, Do Hai Son, Tran Thi Thuy Quynh ✉, Thanh Trung Le
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-02299

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

RACNN: Residual Attention Convolutional Neural Network for Near-Field Channel Estimation in 6G Wireless Communications Vu Tung Lam, Do Hai Son, Tran Thi Thuy Quynh ✉, Thanh Trung Le Abstract Near-field channel estimation is a fundamental challenge in the sixth-generation (6G) wireless communication, where extremely large antenna arrays (ELAA) enable near-field communication (NFC) but introduce significant signal processing complexity. Traditional model-based methods suffer from high computational costs and limited scalability in large-scale ELAA systems, while existing learning-based approaches often lack robustness across diverse channel conditions. To overcome these limitations, we propose the Residual Attention Convolutional Neural Network (RACNN), which integrates convolutional layers with self-attention mechanisms to enhance feature extraction by focusing on key regions within the CNN feature maps. Experimental results show that RACNN outperforms both traditional and learning-based methods, including XLCNet, across various scenarios, particularly in mixed far-field and near-field conditions. Notably, in these challenging settings, RACNN achieves a normalized mean square error (NMSE) of 4.8 ∗ 10 − 3 4.8*10^{-3} at an SNR of 20dB, making it a promising solution for near-field channel estimation in 6G. 1 Introduction Sixth-generation technology (6G) is expected to be the next major advancement in telecommunications. The key performance indicators of 6G should be much superior to those of 5G in terms of peak data rate, spectral efficiency, access density, coverage, and latency [ 7 ] . To achieve these indicators, 6G requires a significantly increased number of antennas, i.e., thousands of elements, compared to hundreds of elements in 5G. Antenna arrays that meet this requirement are referred to as extremely large antenna arrays (ELAA), distinguishing them from massive multiple-input multiple-output (MIMO) in 5G. ELAA is not only more complex in design and deployment but also introduces new challenges in signal processing [ 13 ] , particularly the fundamental transition in electromagnetic propagation modeling from planar-wave (i.e., far-field communication in 5G) to spherical-wave (i.e., near-field communication in 6G). In detail, the Rayleigh distance is used to determine the boundary where propagation is divided as near-field or far-field. For instance, [ 1 , 13 ] indicate that wireless communication systems operating with ELAA and within the Terahertz frequency range (typically around 300GHz) can extend the Rayleigh distance beyond 200 meters, often covering a significant portion of user equipment (UE) devices. This implies that most UEs will operate within the near-field propagation region. Hence, the near-field communication in 6G (NFC-6G) introduces specifically two major challenges: (i) the substantial computational complexity as the number of ELAA elements reaches thousands; (ii) all existing signal processing techniques used in 5G and earlier generations must be adjusted to work with the near-field channel model. In this study, we focus on the problem of channel estimation. Channel estimation is an essential problem in wireless communication systems and many studies have been proposed to address this problem in far-field scenarios [ 3 ] . These studies can be divided into two groups: (i) traditional model-based methods, such as least squares (LS) and minimum mean square error (MMSE) [ 8 ] , and (ii) learning-based methods, which have recently gained attention for leveraging machine learning techniques in channel estimation. Methods in the former group are simple to implement in practice. However, their estimation accuracy is affected by noise and interference. Moreover, when the number of ELAA elements grows larger, the matrix inversion required in these methods is computationally challenging. The latter group is further divided into two approaches, i.e., model-driven methods and data-driven methods. The model-driven approach focuses on optimizing the model based on mathematical solutions. For example, in [ 14 ] , the authors presented learning ISTA (LISTA) and sparsifying dictionary learning-LISTA (SDL-LISTA) networks, which are neural networks based on the iterative shrinkage and thresholding algorithm (ISTA). In [ 9 ] , the authors proposed AMP-NET based on the principle of approximate message passing (AMP) for joint activity detection and channel estimation. These methods achieve higher estimation accuracy compared to the original methods with a small number of learning parameters. However, a neural network with a small number of learning parameters also has its own drawbacks. They perform well only under specific channel conditions that match their training data, such as the number of far-field paths, near-field paths, and noise distribution. The data-driven approach can mitigate this limitation by training neural networks with a large number of learning parameters. This enables the networks to learn from a wide range of channel conditions. To illustrate, in [ 2 ] , the authors proposed an extremely large-scale MIMO channel network (XLCNet) for channel estimation in NFC-6G systems. Specifically, XLCNet transforms noisy channel vectors into 2-channel images and then uses a convolutional neural network (CNN) to denoise them. Notably, this network supports not only far-field and near-field scenarios but also mixed far-field/near-field scenarios. Moreover, CNN is a widely studied architecture that has been utilized for many years. With recent advancements in learning-based techniques for image processing and natural language processing, the underlying concept of XLCNet has the potential for further enhancement. Inspired by [ 2 ] and a recent advancement in deep learning (DL) known as the attention mechanism [ 11 ] , we propose a novel network named the RACNN for NFC-6G wireless communication. 1 1 1 Simulation codes: https://github.com/DoHaiSon/RACNN . RACNN inherits the idea of estimating noise through a neural network and subtracting the estimated noise from the input at the final stage. RACNN follows the convolutional block structure of XLCNet and incorporates self-attention blocks to enhance the overall performance of noise estimation. Our numerical results show that RACNN outperforms traditional model-based methods (LS and MMSE) and XLCNet in various scenarios, e.g., (i) matched training and testing configurations in terms of far-/near-field paths; (ii) testing with only far-field or only near-field paths; and (iii) mismatched mixtures of far-field and near-field components. 2 System Model 2.1 Rayleigh Distance The Rayleigh distance of an NFC-6G system can be represented as follows D_{\text{Ray}}=\frac{2(M\frac{\lambda}{2})^{2}}{\lambda}=\frac{1}{2}M^{2}\lambda, (1) where M is the array dimension and \lambda is the signal wavelength [ 10 ] . In NFC-6G systems, the antenna array can reach thousands of elements, the transmit wavelength is shorter, and the aperture is larger. As a result, the Rayleigh distance D_{\text{Ray}} increases, extending to typical operating distances in cellular networks. This shift places UE services within the near-field range [ 10 ] . Unlike far-field channels, where signals exhibit planar wavefronts, signals in near-field channels propagate with spherical wavefronts and spread-out patterns, resulting in non-uniform distribution characteristics at the receiver. To understand far-field and near-field physical properties, we analyze them over mathematical channel functions presented in the next subsection. 2.2 Channel Models We consider an NFC-6G wireless channel with an ELAA consisting of M elements, serving K UEs. Following the article in [ 2 ] , the downlink far-field channel is modeled as \mathbf{Y}=\mathbf{H}\mathbf{P}^{\frac{1}{2}}\mathbf{S}+\mathbf{Z}, (2) where \mathbf{H}\in\mathbb{C}^{M\times K} represents the channel matrix, \mathbf{P}=\operatorname{diag}\left(\mathit{P}_{1},\mathit{P}_{2},\dots,\mathit{P}_{K}\right) is the transmit power matrix, \mathbf{S}\in\mathbb{C}^{K\times K} is the transmit pilot matrix, and \mathbf{Z}~\in~\mathbb{C}^{M\times K} is an additive white Gaussian noise matrix. We also assume that signal transmission is lossless and transmit powers across all antennas are identical, i.e., P_{k}=P~\forall k . In NFC-6G systems, the channel matrix \mathbf{H} can be expressed in a unified form for both far-field and near-field cases [ 13 ] as

Read the complete text