Abstract
Intelligent beamforming algorithms which can enable ultra-reliable low-latency communication, spectral efficiency, and increased network capacity in a highly dynamic channel environment are needed for Beyond-5G (B5G) wireless networks. Even though the Massive Multiple-Input Multiple-Output (MIMO) system has been found to be an important tool in achieving these goals, the existing beamforming algorithms which use iterative methods show high computation cost, poor convergence, and little adaptivity to the channel environment changes which makes them incapable of being used in real-time applications. In this regard, we propose in this paper a new architecture for a deep learning-based Beamforming called Lightweight Transformer-Based Dynamic Beamforming Network (LT-DBNet). In addition, the designed framework is able to capture the long-range spatial relationships between the antennas in a manner that reduces complexity to an acceptable level for B5G Massive MIMO systems. Besides, there exists an adaptive beam refinement module that dynamically adjusts the beamforming weight based on the changing nature of the channel state. In the evaluation of the proposed approach, a simulation was carried out using the B5G Massive MIMO scenario where the proposed model is compared to the CNNs, LSTMs, and conventional Transformer beamforming approaches. Results show that LT-DBNet is able to achieve a beam selection accuracy of 97.82%, spectral efficiency of 27.1 bps/Hz, network throughput of 13.1 Gbps, energy efficiency of 31.2 bits/J, and computational complexity of 15.4 GFLOPs.
Keywords
Beyond-5G Networks, Massive MIMO, Transformer Networks, Dynamic Beamforming, Deep Learning