Spatial Attention and Quantization based Contrastive Learning Framework for mmWave Massive MIMO Beam Training

Abstract Identifying the optimal beam is a major challenge in millimeter wave (mmWave) wireless communication systems, considering the impacts of noise and training overhead for processing massive multiple-input multiple-output (MIMO) channel streams. To address these challenges, the conventional schemes are to extract the feature of received signals and directly predict the optimal beam based on deep learning (DL) during the beam training. However, this process highly relies on the quality and prior knowledge of massive MIMO channels. In this paper, we propose a novel spatial attention and quantization based beam contrastive learning framework to draw the inherent transmitting angular relation between the beam signature and sub array channels from the contrastive environmental prediction. Specifically, we propose two DL-associated beam training solutions to improve transmission reliability and reduce the training overhead. The first solution, named quantified phase-based transformer architecture (QPTNet), applies a high spatial resolution codebook to quantify the received array signals along the subcarriers into finite categorical beamformer codeword and predict the optimal beam based on the attention mechanism. The second solution, named self enhanced QPTNet (SE-QPTNet), enhances the capability of identifying the beam signature by competitively predicting the partial channel arrays with the limited training data. Simulation results show that our proposed schemes obtain significantly higher beamforming gain with lower prior channel requirements compared to the existing DL-based schemes, and achieve highly reliable performance for mmWave massive MIMO systems.