Deep Learning-Based Fast Beam Tracking for mmWave V2I System

This paper investigates the sensing-assisted fast beam tracking in a time-varying millimeter wave (mmWave) massive MIMO vehicular-to-infrastructure (V2I) system. Traditional Kalman filter-based beam tracking schemes rely on channel models and have huge beam training overheads. To avoid these issues, a sensing-assisted fusion neural network (S-FusNN) framework is constructed for fast beam tracking. This framework fuses the historical vehicle position and the historical optimal beam angle pairs to predict the current optimal beam angle pairs. The simulation results show superior beam tracking performance and low beam training overhead of the proposed S-FusNN over the developing baselines.