A Deep Learning Framework for Millimeter Wave Antenna Beam Selection and Tracking

Deep learning models are being extensively explored for millimeter wave (mmWave) beamforming and tracking. As compared to traditional beam tracking methods, deep learning based schemes have the potential to offer low complexity and latency. Conventional beamforming techniques rely on beam switching protocols in order to identify the optimal beams at the transmitter and receiver which maximize the received signal-to-noise ratio. Such techniques incur large training overhead which can be prohibitive due to the need for frequent beam tracking in order to maintain a reliable link under outdated channel conditions. In this paper, a novel technique for antenna beam selection and tracking based on Recurrent Neural Network (RNN) is proposed. The change in user node location relative to its previous location is predicted based on the received signal strength difference between successive user node positions and current antenna beam indices. The optimum beam pair for the new location is thus determined based on the RNN prediction. The proposed scheme accounts for blockage of transmission link as well as channel aging thereby augmenting the beam selection method with a beam tracking mechanism. Numerical results indicate the feasibility of the proposed scheme for deployment in indoor and outdoor fixed wireless access mmWave systems.