[Paper Review] Deep Learning-Based Channel Estimation for Double-RIS Aided Massive MIMO System

Mengbing Liu, Xin Li+4 more|arXiv (Cornell University)|Oct 22, 2022

Advanced Wireless Communication Technologies4 citations

TL;DR

This paper proposes a skip-connection attention (SC-attention) network for deep learning-based channel estimation in double-reconfigurable intelligent surface (RIS)-aided massive MIMO systems. By integrating self-attention layers and skip-connection structures, the network enhances feature representation and mitigates vanishing gradients, achieving up to 16% lower normalized mean square error (NMSE) than baseline methods, especially at low SNR, demonstrating superior accuracy and robustness in noisy, high-dimensional cascaded channel estimation.

ABSTRACT

Reconfigurable Intelligent Surface (RIS) is considered as an energy-efficient solution for future wireless communication networks due to its fast and low-cost configuration. In this letter, we consider the estimation of cascaded channels in a double-RIS aided massive multiple-input multiple-output system, which is a critical challenge due to the large number of antennas equipped at the base station and passive RIS elements. To tackle this challenge, we propose a skip-connection attention (SC-attention) network that utilizes self-attention layers and skip-connection structure to improve the channel estimation performance from the noisy pilot-based observations. Simulation results compare the proposed SC-attention network with other benchmark methods and show that SC-attention network can effectively improve the accuracy performance on normalized mean square error (NMSE) for cascaded links in a double-RIS aided system.

Motivation & Objective

• Address the high-dimensional, noisy channel estimation challenge in double-RIS-aided massive MIMO systems with large-scale antennas and passive RIS elements.
• Overcome the limitations of conventional linear estimators (e.g., LS, LMMSE) in low SNR environments and high-dimensional cascaded channel estimation.
• Improve estimation accuracy and robustness by leveraging deep learning to model complex non-linear spatial features in noisy pilot observations.
• Integrate self-attention mechanisms and skip-connection structures to enhance receptive field size, model capacity, and gradient flow in the neural network.
• Demonstrate superior performance over existing DL-based and conventional estimators across varying SNR conditions.

Proposed method

• Model the channel estimation problem as a denoising task, where the input is a noisy observation of the cascaded channel matrix from pilot signals.
• Design a skip-connection attention (SC-attention) network that combines self-attention layers to capture long-range spatial dependencies and improve feature representation.
• Incorporate skip-connections to preserve low-level features and prevent gradient vanishing or exploding during training, enhancing training stability and performance.
• Use a residual-like architecture where the input is directly concatenated with the output of attention blocks to facilitate learning of residual mappings.
• Train the network end-to-end using a large dataset of 120,000 samples with SNR ranging from -10 dB to 15 dB to ensure generalization across diverse conditions.
• Evaluate performance using normalized mean square error (NMSE) as the primary metric for both single- and double-reflection links.

Figure 1: Illustration of the Double-RIS aided system model.

Experimental results

Research questions

• RQ1 Can a deep learning-based approach outperform conventional linear estimators (LS and LMMSE) in estimating high-dimensional cascaded channels in double-RIS-aided massive MIMO systems?
• RQ2 To what extent does integrating self-attention layers improve the estimation accuracy by enhancing the receptive field and model capacity?
• RQ3 How effective is the skip-connection structure in mitigating gradient issues and improving estimation performance, especially in low SNR regimes?
• RQ4 Does the proposed SC-attention network generalize well across different SNR levels and maintain robustness in noisy environments?
• RQ5 How does the performance of the SC-attention network compare to existing deep learning baselines like SF-CNN in terms of NMSE across various channel conditions?

Key findings

• The SC-attention network achieves up to 16% lower NMSE than the attention-only baseline at -10 dB SNR, demonstrating significant improvement in low-SNR environments.
• At high SNR (15 dB), the performance gap between the SC-attention network and conventional estimators narrows, as expected due to reduced noise, but the network still maintains a slight edge.
• Visualization of the attention blocks shows a progressive reduction in noise intensity from input to output, confirming the effectiveness of the block-by-block denoising mechanism.
• The SC-attention network outperforms both the LS and LMMSE estimators across all SNR levels, particularly in low-SNR conditions where linear methods struggle.
• The network also surpasses the SF-CNN baseline, indicating that the integration of self-attention and skip-connections leads to superior feature learning and estimation accuracy.
• The proposed method maintains robustness and generalization across the entire SNR range from -10 dB to 15 dB, validated on independent test data.

Figure 2: Overview of the self-attention layer.