Title:Training Enhancement of Deep Learning Models for Massive MIMO CSI Feedback with Small Datasets

Abstract:Accurate downlink channel state information (CSI) is vital to achieving high spectrum efficiency in massive MIMO systems. Existing works on the deep learning (DL) model for CSI feedback have shown efficient compression and recovery in frequency division duplex (FDD) systems. However, practical DL networks require sizeable wireless CSI datasets during training to achieve high model accuracy. To address this labor-intensive problem, this work develops an efficient training enhancement solution of DL-based feedback architecture based on a modest dataset by exploiting the complex CSI features, and augmenting CSI dataset based on domain knowledge. We first propose a spherical CSI feedback network, SPTM2-ISTANet+, which employs the spherical normalization framework to mitigate the effect of path loss variation. We exploit the trainable measurement matrix and residual recovery structure to improve the encoding efficiency and recovery accuracy. For limited CSI measurements, we propose a model-driven lightweight and universal augmentation strategy based on decoupling CSI magnitude and phase information, applying the circular shift in angular-delay domain, and randomizing the CSI phase to approximate phase distribution. Test results demonstrate the efficacy and efficiency of the proposed training strategy and feedback architecture for accurate CSI feedback under limited measurements.