Related Experiment Videos
Learning-driven MIMO channel estimation using a residual U-Net-BiLSTM-attention hybrid model
Mohammad Zubair Khan1, Ibrahim Aljubayri2, Chander Prabha3
1Faculty of Computer Science and Information Systems, Islamic University of Madinah, Madinah, 42351, Saudi Arabia.
Scientific Reports
|July 18, 2026
Summary
A novel hybrid deep learning channel estimator (ResUNet-BiLSTM-Attention) significantly improves accuracy in multiple-input multiple-output (MIMO) systems. This advanced method outperforms existing techniques, offering robust performance even in noisy conditions.
Area of Science:
- Wireless Communications
- Signal Processing
- Machine Learning
Background:
- Accurate channel state information (CSI) is crucial for multiple-input multiple-output (MIMO) systems, but conventional estimators struggle with noise, limited pilots, and model mismatch.
- Existing methods often show performance degradation in challenging wireless environments, necessitating more robust estimation techniques.
Purpose of the Study:
- To introduce a hybrid deep learning model, the Residual U-Net-Bidirectional Long Short-Term Memory with Attention (ResUNet-BiLSTM-Attention), for enhanced MIMO channel estimation.
- To evaluate the proposed estimator's performance against conventional and learning-based baselines across various signal-to-noise ratio (SNR) conditions.
Main Methods:
- Developed a hybrid ResUNet-BiLSTM-Attention architecture combining U-Net for spatial features, BiLSTM for temporal dependencies, and self-attention for global context.
- Trained the model on a synthetic MIMO dataset (20,000 samples) using normalized mean square error (NMSE) loss.
- Conducted extensive simulations comparing the proposed estimator with Least Squares (LS), LMMSE, Orthogonal Matching Pursuit (OMP), and other deep learning approaches.
Main Results:
- The ResUNet-BiLSTM-Attention estimator achieved superior performance, outperforming all conventional and learning-based baselines.
- At 25 dB SNR, the proposed method attained an NMSE of approximately [Formula: see text] dB, showing a 9-10 dB gain over the BSP-DSDW baseline.
- Ablation studies confirmed the contribution of each module (ResUNet, BiLSTM, Attention), and runtime evaluations demonstrated practical feasibility with inference latency around 2.5-4.0 ms.
Conclusions:
- The hybrid ResUNet-BiLSTM-Attention model offers a robust and efficient solution for accurate MIMO channel estimation.
- The proposed deep learning approach significantly enhances CSI acquisition, particularly under adverse channel conditions.
- The model's design is validated for practical deployment due to its strong performance, low computational complexity, and fast inference times.
