Publication

Abstract : Massive multiple-input and multiple-output (M-MIMO) is considered a vital technology for enhancement of energy efficiency and link capacity in fifth-generation (5G) communication systems. The accessibility of the downlink channel state information (CSI) at the base station (BS) is necessary to access the potential advantages of millimeter wave (mmWave) and M-MIMO systems. The CSI matrix is usually large due to the massive scale antennas present at the BS. To carry out the downlink precoding computations, the downlink channel responses need to be estimated and fed back to the base station. However, in absence of channel reciprocity in frequency division duplexing, acquiring accurate CSI is a challenge. This paper proposes a novel network based on deep learning, VAECNN-Net. The network is the combination of convolutional neural network and variational autoencoder, hence, termed VAECNN-Net. The network compresses the CSI at the user equipment side and is reliably retrieved at the base station. The performance analysis is carried out on a synthetic data set for a single-user mmWave M-MIMO system and is also validated for the COST 2100 channel model. It is observed that the proposed network has lower computational complexity than existing models such as CsiNet, TransNet, and CsiNet+DNN. The network also provides superior cosine similarity and normalized mean square error (NMSE) performance compared to existing techniques. The model is compared with the baseline network CsiNet and it also shows superior performance in terms of NMSE for the generated data set.