Machine Learning Based mmWave Channel Tracking in Vehicular Scenario

Abstract

Millimeter wave (mmWave) communication has become a key enabling technology for 5G and beyond networks because of its large bandwidth and high transmission rate. In a vehicular mmWave system, beam tracking is a challenging task due to the user's fast mobility and narrow beam of mmWave transmission. In this paper, we study the intelligent beam tracking scheme with low training overhead for mmWave vehicular transmission. Specifically, we utilize the past channel state information (CSI) to efficiently predict the future channel by designing a machine learning prediction model. Using such predicted CSI, the base stations (BSs) reduce the number of channel estimations and save the overhead of pilots. We build the prediction model based on a long short term memory (LSTM) structure whose dataset is composed of the channel vectors of each coherence time duration. The experiments show that the proposed LSTM can accurately predict the channel of the vehicular user and achieve satisfactory transmission rate with less pilot overhead than that of traditional beam training scheme.