<p>Channel estimation provides the channel state information (CSI) required for coherent detection and precoding in multiple-input multiple-output (MIMO) systems. Accurate CSI is particularly critical under strong noise, limited pilot overhead and model mismatch, where conventional estimators often exhibit significant performance degradation. This work introduces a hybrid residual U-Net–bidirectional long short-term memory with attention (ResUNet–BiLSTM–Attention) channel estimator that learns a nonlinear mapping from noisy pilot observations to MIMO channel coefficients. The architecture combines a ResUNet encoder–decoder for multi-scale spatial feature extraction, a BiLSTM module for capturing structured dependencies in the unfolded feature sequence and a self-attention layer that emphasizes globally informative channel components before reconstruction. The model is trained on a synthetically generated <InlineEquation ID="IEq1"><EquationSource Format="TEX">\(4\times 4\)</EquationSource></InlineEquation> MIMO dataset with 20, 000 training and 2, 000 validation samples over a wide SNR range, using a normalized mean square error (NMSE) loss for stable convergence. Extensive simulations show that the proposed estimator consistently outperforms both conventional and learning-based baselines. The comparison includes LS, LMMSE, orthogonal matching pursuit (OMP), simultaneous OMP (SOMP), beamspace-based dynamic support detection with windowing (BSP–DSDW), CNN-CE, U-Net-CE, and lightweight attention-based CE. At 25 dB SNR, it attains an NMSE of about <InlineEquation ID="IEq2"><EquationSource Format="TEX">\(-24\)</EquationSource></InlineEquation> dB, corresponding to an NMSE gain of about 9–10 dB over BSP–DSDW and a clear improvement over the added learning-based baselines. A module-wise ablation study further verifies the individual contribution of the ResUNet, BiLSTM, and attention blocks, while the runtime evaluation is conducted under a common GPU-enabled benchmarking setup using repeated inference trials. Additional studies on training set size, different user channels, computational complexity, parameter count, FLOPs, memory requirement, inference latency, pilot length, noise factor, and spatial correlation further confirm the robustness and practical feasibility of the proposed design. With a fixed computational structure, the proposed estimator achieves an observed inference latency range of approximately 2.5–4.0 ms per channel realization under the considered compact MIMO setup.</p>

Learning-driven MIMO channel estimation using a residual U-Net–BiLSTM–attention hybrid model

期刊论文
出版模式：

开放获取
发表日期：

2026年7月18日

查看全文

Mohammad Zubair Khan,
Ibrahim Aljubayri,
Chander Prabha,
Turki Alghamdi,
Adnan Nadeem

摘要

Channel estimation provides the channel state information (CSI) required for coherent detection and precoding in multiple-input multiple-output (MIMO) systems. Accurate CSI is particularly critical under strong noise, limited pilot overhead and model mismatch, where conventional estimators often exhibit significant performance degradation. This work introduces a hybrid residual U-Net–bidirectional long short-term memory with attention (ResUNet–BiLSTM–Attention) channel estimator that learns a nonlinear mapping from noisy pilot observations to MIMO channel coefficients. The architecture combines a ResUNet encoder–decoder for multi-scale spatial feature extraction, a BiLSTM module for capturing structured dependencies in the unfolded feature sequence and a self-attention layer that emphasizes globally informative channel components before reconstruction. The model is trained on a synthetically generated \(4\times 4\) MIMO dataset with 20, 000 training and 2, 000 validation samples over a wide SNR range, using a normalized mean square error (NMSE) loss for stable convergence. Extensive simulations show that the proposed estimator consistently outperforms both conventional and learning-based baselines. The comparison includes LS, LMMSE, orthogonal matching pursuit (OMP), simultaneous OMP (SOMP), beamspace-based dynamic support detection with windowing (BSP–DSDW), CNN-CE, U-Net-CE, and lightweight attention-based CE. At 25 dB SNR, it attains an NMSE of about \(-24\) dB, corresponding to an NMSE gain of about 9–10 dB over BSP–DSDW and a clear improvement over the added learning-based baselines. A module-wise ablation study further verifies the individual contribution of the ResUNet, BiLSTM, and attention blocks, while the runtime evaluation is conducted under a common GPU-enabled benchmarking setup using repeated inference trials. Additional studies on training set size, different user channels, computational complexity, parameter count, FLOPs, memory requirement, inference latency, pilot length, noise factor, and spatial correlation further confirm the robustness and practical feasibility of the proposed design. With a fixed computational structure, the proposed estimator achieves an observed inference latency range of approximately 2.5–4.0 ms per channel realization under the considered compact MIMO setup.