RSRP-Based Online Beam Synthesis Using a Model-Aided Autoencoder

This paper presents a self-supervised learning approach for online beam synthesis using the reference signal received power (RSRP) measurements in mmWave band communications. We propose a sparse structure which can effectively approximate the true antenna space covariance matrix (ASCM) of the mmWave channel with only a few effective paths. The sparse structure is then forced within the latent space of a model-aided autoencoder (MAE), whose encoding part is a neural network (NN) which estimates the parameters of the effective ASCM by the measured RSRPs, while the decoding part is a closed-form model which reconstructs the RSRPs from the estimated effective ASCM. The MAE is trained based on the similarity loss between the measured RSRPs and the reconstructed RSRPs, such that the ground truth channel state information (CSI) is not needed during the training. The predicted effective spatial covariance matrix is then used to compute the optimal communication beam. Simulations show that the proposed scheme can significantly improve the beamforming performance especially in non-line-of-slight (NLOS) channel conditions.