Multi-Branch Unsupervised Learning-Based Beamforming in mm-Wave Massive MIMO Systems With Inaccurate Information

Abstract

In mobile millimeter wave (mm-Wave) systems, most deep learning-based beamforming models only input channel state information (CSI). However, as user speed increases, CSI inaccuracy increases, leading to severe performance degradation. Their single model structures cause a low generalization in large-scale networks. In this paper, a multi-branch unsupervised learning model, named MB-IncepNet, is established for mobile user beamforming, where inaccurate user location information (ULI) is extra considered to improve the beamforming robustness, and an Inception-Shortcut block is rationally constructed to improve the generalization of MB-IncepNet. Specifically, MB-IncepNet has two sub-networks for ULI and CSI inputs, which are processed first by the Inception-Shortcut processing and then fused to correct beamforming results by full-connection processing. Furthermore, the Inception-Shortcut block has multiple parallel convolution branches with convolution kernels of different sizes and a shortcut, which indicates MB-IncepNet can adapt to networks of different scales. Besides, the base station power constraint is incorporated into the model as a power layer, and the inverse of the sum-rate is chosen as the loss function for unsupervised training. The simulation results show that, under inaccurate ULI and CSI, MB-IncepNet can still achieve more than 90% effective sum-rate compared with the ideal iterative algorithm.