HASwinNet: A Swin Transformer-Based Denoising Framework with Hybrid Attention for mmWave MIMO Systems.

Area of Science:

• Wireless communication systems
• Signal processing
• Machine learning for communications

Background:

• Millimeter-wave (mmWave) massive MIMO is crucial for 6G integrated sensing and communication (ISAC), enabling high-capacity backhaul and environmental sensing.
• Accurate channel estimation in mmWave systems is challenging due to noise sensitivity, sparse multipath, and limited pilot resources, especially at low SNRs.

Purpose of the Study:

• To propose HASwinNet, a deep learning framework for denoising millimeter-wave channels.
• To address the challenges of accurate channel estimation under limited pilot resources and low SNR conditions in mmWave ISAC systems.

Main Methods:

• Developed HASwinNet, a deep learning framework utilizing a Swin Transformer encoder for representation learning.
• Incorporated two branches for sparse token extraction and angular-domain refinement using DFT, SE, and IDFT operations.
• Introduced an angular-domain perceptual loss to enforce spectral consistency and preserve multipath structures.

Main Results:

• HASwinNet significantly improves normalized mean squared error (NMSE) and bit error rate (BER) compared to CNN, LSTM, and U-Net baselines.
• The framework effectively exploits angular sparsity, maintaining performance advantages under pilot-limited conditions.
• Demonstrated superior performance using the Saleh-Valenzuela (S-V) channel model.

Conclusions:

• HASwinNet offers a scalable solution for practical 6G mmWave backhaul applications.
• The proposed framework shows significant potential for accurate channel recovery in ISAC scenarios, benefiting both communication and sensing functionalities.