Publication

Abstract : The promising performance gains of massive multiple-input and multiple-output (M-MIMO) rely on the accurate downlink channel state information (CSI) at the base station (BS). In the case of frequency division duplex (FDD) systems, the user equipment (UE) has to feed the estimated downlink CSI matrix to the BS precisely due to the absence of the principle of reciprocity. However, M-MIMO systems have a large number of antennas which leads to a significant amount of CSI data. Sending all this data back to the BS creates a bottleneck, consuming a large portion of the limited bandwidth resources available. In this paper, CBAM-VAE, a novel deep learning (DL) framework that complies with the 3GPP specifications is proposed to effectively analyze the objective of CSI feedback. The model is designed to incorporate the key features of the convolutional block attention module (CBAM) integrated with the variational autoencoder (VAE) hence, termed CBAM-VAE. The experimental outcomes show the superior performance of the designed architecture in comparison to the baseline networks using cosine similarity () and normalized mean square error (NMSE) as the key performance indicators for four distinct lengths of codeword . In addition, CBAM-VAE also has less computational overhead making it acceptable for real-time scenarios.