A Convolutional-Transformer Residual Network for Channel Estimation in Intelligent Reflective Surface Aided MIMO...

Area of Science:

• Wireless Communications
• Signal Processing
• Artificial Intelligence

Background:

• Intelligent Reflective Surface (IRS)-aided Multiple-Input Multiple-Output (MIMO) systems offer enhanced spectral and energy efficiency for wireless networks.
• Accurate channel estimation is crucial but challenging due to the passive nature and high dimensionality of IRS channels.

Purpose of the Study:

• To propose a lightweight hybrid framework for efficient cascaded channel estimation in IRS-aided MIMO systems.
• To enhance the accuracy and efficiency of channel estimation compared to existing methods.

Main Methods:

• A hybrid framework combining a physics-based Bilinear Alternating Least Squares (BALS) algorithm with a deep neural network (ConvTrans-ResNet).
• The ConvTrans-ResNet integrates convolutional embeddings and Transformer modules within a residual learning architecture.
• Ablation studies were performed to optimize the network's architecture for reduced complexity and parameter count.

Main Results:

• The proposed ConvTrans-ResNet method significantly outperforms state-of-the-art neural models (HA02, ReEsNet, InterpResNet) in Normalized Mean Squared Error (NMSE).
• Achieved superior estimation accuracy and efficiency across various Signal-to-Noise Ratio (SNR) levels and IRS element sizes.
• Demonstrated a compact network configuration with low computational complexity.

Conclusions:

• The hybrid BALS and ConvTrans-ResNet framework provides a practical and efficient solution for channel estimation in IRS-aided MIMO systems.
• The optimized lightweight network is suitable for real-world deployment, advancing future wireless communication technologies.