Applying Deep Learning for Phase-Array Antenna Design

Abstract

Hybrid beamforming (HBF) can provide rapid data transmission rates while reducing the complexity and cost of massive multiple-input multiple-output (MIMO) systems. However, channel state information (CSI) is imperfect in realistic downlink channels, introducing challenges to hybrid beamforming (HBF) design. For HBF designs, we had a hard time finding the proper labels. If we use the optimized output based on the traditional algorithm as the label, the neural network can only be trained to approximate the traditional algorithm, but not better than the traditional algorithm. This thesis proposes a hybrid beamforming neural network based on unsupervised deep learning (USDNN) to prevent the labeling overhead of supervised learning and improve the achievable sum rate based on imperfect CSI. Compared with the traditional HBF method, the unsupervised learning-based method can avoid the labeling overhead as well as obtain better performance than the traditional algorithm. The network consists of 5 dense layers, 4 batch normalization (BN) layers and 5 activation functions. After training, the optimized beamformer can be obtained, and the optimized beamforming vector can be directly output. The simulation results show that our proposed method is 74% better than manifold optimization (MO) and 120% better than orthogonal match pursuit (OMP) systems. Furthermore, our proposed USDNN can achieve near-optimal performance.