Machine Learning Assisted Efficient Beam Training and Channel Estimation for RIS-mmWave Communications

‍ ‍In this paper， we study high-efficiency beam training and channel estimation for the reconfigurable intelligent surface-assisted millimeter-wave （RIS-mmWave） system， which is the key to achieving ultra-high wireless information transmissions in B5G and 6G wireless networks. In particular， noting the intrinsic connection between RIS-mmWave beam management and effective channel acquisition， a novel machine learning-assisted efficient beam training and channel estimation method is proposed for RIS-mmWave systems. Specifically， in the first stage， a novel semi-supervised learning model is designed to achieve fast online beam training by exploiting users' location information， which directly obtains coarse angular domain information. In the second stage， a semi-supervised learning-assisted compressed sensing is proposed to estimate the RIS-dependent cascade channel， which uses the coarse angular domain information predicted by the learning model to drive the block orthogonal matching pursuit algorithm for channel estimation. Simulation results show that the performance of the proposed beam training and channel estimation method outperforms the reference scheme in terms of system overhead and channel estimation error.