Model-Driven Lightweight Network for CSI Feedback in Time-Varying Massive MIMO Systems

Channel state information (CSI) is vital for massive multiple-input multiple-output (MIMO) systems to provide high channel capacity and energy efficiency. However, the massive antennas will lead to high feedback overhead in frequency division duplex (FDD) MIMO systems. Recent applications of recurrent neural networks have demonstrated promising results of feedback CSI compression but with tremendous complexity. The markovian-based framework reduces the overall complexity but owns the high complexity of the user equipment (UE), which is not suitable for practical deployment. In this work, we aim to reduce the complexity of UE and improve the reconstruction performance. Leveraging the channel temporal correlation, we propose two differential frameworks called DIFNet1 and DIFNet2 to improve feedback accuracy and efficiency. Moreover, We unfold the Iterative Shrinkage-Thresholding Algorithm (ISTA) to provide excellent reconstruction performance. Simulation results demonstrate that the proposed framework improves the reconstruction performance while greatly reducing the complexity of UEs.

I. Introduction

Massive MIMO system is widely recognized as a key technology in the next-generation mobile communication systems, which can improve channel capacity by exploiting spatial diversity [1] [2]. Moreover, by taking advantage of beamforming, a base station (BS) can focus signal power on a specific user to improve its signal-to-noise ratio (SNR) and reduce interference. However, the high performance of beamforming depends on that BS can obtain accurate downlink CSI. In FDD mode, which is operated in most cellular systems, the channel is weak in reciprocity, so the UE has to feed the downlink CSI back to the BS. A large number of antennas in massive MIMO systems result in a high-dimensional channel matrix, which causes a heavy burden on the feedback. Thus, the downlink CSI compression is crucial to reduce feedback overload.