Low-Complexity Deep Learning Framework for CSI Feedback in Massive MIMO System

Channel state information (CSI) feedback is critical for achieving maximum multiplexing gain and antenna diversity in massive MIMO (M-MIMO) systems. However, challenges arise due to the lack of channel reciprocity in frequency division duplexing (FDD) and the large number of antennas. In order to effectively analyze the feedback of CSI, we developed a deep learning (DL) based network DeConvD-CRNet suitable for the 5G new radio clustered delay line (nrCDL) channel model that follows 3GPP Release 18 specifications. The network utilizes convolution factorization and a joint combination of dilated channel reconstruction network, residual network, and inception module, hence, termed DeConvD-CRNet. The proposed network is compared with two baseline networks, CsiNet and CRNet with set of synthetically generated data. Simulation results exhibit superior reconstruction performance in terms of normalized mean square error (NMSE) and cosine similarity ( $\rho $ ) for different compression ratios ( $\eta $ ) at reduced computational complexity. The analysis is extended to the COST 2100 channel and a mixed dataset demonstrating the scalability, feasibility, and generalization of the network in practical scenarios.