Advances in machine learning have widened the range of its applications in many fields. In particular, deep learning has attracted much interest for its ability to provide solutions where the derivation of a rigorous mathematical model of the problem is troublesome. Our interest was drawn to the application of deep learning for channel state information feedback reporting, a crucial problem in frequency division duplexing (FDD) 5G networks, where knowledge of the channel characteristics is fundamental to exploiting the full potential of multiple-input multiple-output (MIMO) systems. We designed a framework adopting a 5G New Radio convolutional neural network, called NR-CsiNet, with the aim of compressing the channel matrix experienced by the user at the receiver side and then reconstructing it at the transmitter side. In contrast to similar solutions, our framework is based on a 5G New Radio fully compliant simulator, thus implementing a channel generator based on the latest 3GPP 3-D channel model. Moreover, realistic 5G scenarios are considered by including multi-receiving antenna schemes and noisy downlink channel estimation. Simulations were carried out to analyze and compare the performance with current feedback reporting schemes, showing promising results for this approach from the point of view of the block error rate and throughput of the 5G data channel.
Keywords: 5G; CSI reporting; New Radio; convolutional neural network; deep learning.