SwinLSTM Autoencoder for Temporal-Spatial-Frequency Domain CSI Compression in Massive MIMO Systems

This study presents a parameter-light, low-complexity artificial intelligence/machine learning (AI/ML) model that enhances channel state information (CSI) feedback in wireless systems by jointly exploiting temporal, spatial, and frequency (TSF) domain correlations. While traditional frameworks use autoencoders for CSI compression at the user equipment (UE) and reconstruction at the network (NW) side in spatial-frequency (SF), massive multiple-input multiple-output (mMIMO) systems in low mobility scenarios exhibit strong temporal correlation alongside frequency and spatial correlations. An autoencoder architecture alone is insufficient to exploit the TSF domain correlation in CSI; a recurrent element is also required. To address the vanishing gradients problem, researchers in recent works have proposed state-of-the-art TSF domain CSI compression architectures that combine recurrent networks for temporal correlation exploitation with deep pre-trained autoencoder that handle SF domain CSI compression. However, this approach increases the number of parameters and computational complexity. To jointly utilize correlations across the TSF domain, we propose a novel, parameter-light, low-complexity AI/ML-based recurrent autoencoder architecture to compress CSI at the UE side and reconstruct it on the NW side while minimizing CSI feedback overhead.