Memory Shared Spatial Attention Neural Network for Reliable Beam Prediction

Download PDF
Open Webpage

Haohui Jia, Na Chen, Minoru Okada

Published: 2022, Last Modified: 29 Jan 2026GCCE 2022EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: In this paper, we propose a memory shared spatial attention neural network for beam prediction based on a millimeter wave (mmWave) massive multiple input multiple output (MIMO) system. The proposed method learns temporal features based on the autoregressive model from the frequency meanwhile extract the spatial characteristics from the antenna indices. Specifically, we apply a weight sharing long-short term memory (LSTM) to explore inherent delay features from real and imaginary of mmWave channels among antenna indices. Due to the spatial coherence of massive MIMO, we further apply the attention mechanism to extract the spatial feature from the temporal features. To verify the proposal, we compare the proposed model with conventional deep neural network (DNN), Transformer, unshared weight, and shared weight proposal underlying different signal-to-noise ratio (SNR) levels. According to the experiment results, the accuracy can be improved from 66% to 92% comparing with the others under 20 dB SNR. And the average accuracy can achieve 65% under 0 dB SNR.