Receivers can usually estimate CSI with the use of known pilot signals. The BS is able to measure the down-link CSI using pilots in uplink broadcasts in the special case of time-division duplex (TDD) mode, courtesy of channel reciprocity. The BS uses UE feedback to approximate downlink CSI in frequency division duplex (FDD) mode because the channel reciprocity of the uplink channel and the downlink channel is less than 100 rather limited. One important consideration is how to reduce feedback bandwidth while keeping downlink CSI predictions correct; this is known as the CSI feedback scheme. Deep learning-based alternatives to CS approaches for compressed CSI estimation have been proposed in recent studies. In order to derive compressed versions of high-dimensional CSI, these suggested approaches usually make use of convolutional neural networks (CNNs). The CNN-based research has two broad types of deep learning architecture. Convolutional neural networks (CNNs) networks for autoencoders fall into the first group. These networks have two subnetworks: one that learns a low-dimensional representation from the original data and estimates the original data using this representation. The second kind of CNN is an unrolled optimization network, which takes its cues from the aforementioned CS techniques by constructing the CNN based on a set of small number of repeated blocks, which resembles an iteration of a particular CS algorithm. We also enhance those which came before where the homogeneous differential encoders are used by the heterogeneous differential encoder which utilizes different network architectures per time slot, giving more accurate estimations.
Pandey et al. (Wed,) studied this question.