Low-Overhead Beam Selection for mmWave Massive MIMO Systems by Deep Learning

Abstract

In 5G communication systems, massive multiple-input multiple-output (MIMO) technology could compensate for the severe path-loss of millimeter-wave by high-gain beams. However, the excessive overhead of beam measurement induces a significant challenge for practical systems. To reduce the measurement overhead and improve the accuracy of beam selection, deep learning (DL) is employed to predict the signal quality of all beam pairs by measuring a few selected beam pairs. The proposed beam prediction network is mainly composed of input module, residual module and multi-layer perceptron (MLP) module, therefore denoted as ReMBP net. After the preprocessing for the input signal quality of measured beam pairs, the designed residual module is adopted to extract the high-dimensional feature vector, then the signal quality of all beam pairs could be predicted by the designed MLP module. Moreover, to improve the generalization ability of the proposed ReMBP net for different communication environments, the input module with a flexible input size of signal quality data is designed. Experiment results indicate that a high-accuracy beam selection could be achieved with low overhead compared with other traditional artificial intelligence (AI) models.