DCsiNet: Effective Handling of Noisy CSI in FDD Massive MIMO System

To enhance the performance of frequency division duplex (FDD) massive multiple-input multiple-output (MIMO) systems, it is crucial to relay downlink channel state information (CSI) from user equipment (UE) to the base station (BS). However, minimizing feedback overhead poses a significant challenge. In recent years, various deep learning (DL) models have been introduced to efficiently compress CSI into codewords at the UE and subsequently reconstruct the information at the BS, aiming to reduce feedback overhead. Notably, previous studies overlooked real-world scenarios, such as codeword corruption by noise. In this study, we propose a novel deep learning model, denoising CSI network (DCsiNet), designed to address noise-corrupted codewords, wherein the decoder serves the dual purpose of decompressing and denoising the CSI. Our proposed network incorporates a multi-scale feature extractor and an attention-based refiner, contributing to improved denoising and reconstruction outcomes. Numerical results are presented to showcase the superior performance of our proposed network compared to other deep learning-based models, particularly in reconstructing CSI from noisy codewords.