Deep Learning-Based Variable Scaling Beam Training for Massive MIMO mmWave Systems

Improving the accuracy of beam training while reducing the training overhead and the influence of noise has become an important issue for massive multiple-input multiple-output (MIMO) millimeter-wave (mmWave) communication systems. In this paper, we propose a deep learning-based multi-scale neural network for beam training in massive MIMO mmWave communication system. The model predicts the orientation of narrow beams by learning the characteristics of wide beams, achieving high accuracy and low training overhead. Specifically, our model consists of three modules. In the first module, we deploy a convolutional neural network (CNN) to extract features of the instantaneous received signal of a wide beam. In the second module, we develop multi-scale convolution to extract wide beam features of the different time combination. In the third module, we conduct a long-term short-term memory (LSTM) network to calibrate the orientation of narrow beams based on previous predictions, thereby enhance the model's robustness to noise. Finally, according to the experimental results, our model improves beam training accuracy with low training overhead while reducing the influence of noise.