Unsupervised Neural Beamforming for Uplink MU-SIMO in 3GPP-Compliant Wireless Channels

Area of Science:

• Wireless Communication Systems
• Signal Processing
• Machine Learning

Background:

• Traditional linear beamforming (ZFBF, MMSE) struggles with non-ideal conditions like imperfect CSI and high dimensionality.
• Deep learning offers adaptive, data-driven beamforming solutions by leveraging neural network capabilities.
• Existing deep learning approaches lack systematic comparison and analysis in realistic wireless environments.

Purpose of the Study:

• To systematically compare two unsupervised deep learning beamforming architectures for uplink receive beamforming.
• To evaluate their performance against traditional methods (ZFBF, MMSE) in multi-user systems.
• To analyze the computational complexity and scalability of these deep learning models.

Main Methods:

• Developed and compared a simple Neural Network Beamforming (NNBF) model (CNNs, fully connected layers) and a transformer-based NNBF model.
• Evaluated models in a multi-user single-input multiple-output (MU-SIMO) system using 3GPP TDL-A and UMa channel models.
• Performed FLOPs-based complexity analysis to characterize inference-time scaling behavior.

Main Results:

• Transformer-based NNBF achieves superior performance in realistic conditions (imperfect CSI, mobility) despite higher computational cost.
• Simple NNBF shows comparable or better performance than ZFBF/MMSE with significantly lower complexity under simplified assumptions (perfect CSI, stationary UEs).
• Both NNBF architectures demonstrate potential for adaptive and scalable beamforming solutions.

Conclusions:

• Deep learning beamforming, particularly transformer-based models, provides a powerful and adaptive alternative to traditional methods for wireless communication.
• The choice between simple and transformer-based NNBF depends on the trade-off between performance requirements and computational constraints.
• This study provides valuable insights into the practical application and performance scaling of deep learning in beamforming.