Exploiting Temporal Priors for Efficient Real-time Compression and Feedback of Wireless Channels

Machine learning based compression frameworks are rapidly gaining popularity as the demand for efficient storage, processing, and transmission of large-scale data continues to grow across diverse applications such as video streaming and IoT. Recently, such frameworks have also sparked significant interest in wireless communications and the task of ML based wireless channel compression is currently one of the use cases being explored by the international wireless standards body, 3GPP, for standardization. In wireless communication systems, each user device or user equipment (UE) typically estimates the wireless channel between the transmitting base station (BS) and itself and feeds back information related to the estimated channel state information (CSI) to the serving BS, which may then be utilized for downstream processing. While the current 5G communication stack employs a combination of matrix factorization and quantization approaches to compress the CSI, autoencoders (AE) have emerged as a viable option to compress the estimated spatial-frequency (SF) channel sample and send it back to the base station for reconstruction. Although the AE-based approaches have shown acceptable CSI reconstruction performance, there is still a large room for further improvement, both from an overhead reduction as well as reconstruction performance perspectives. This paper proposes a new AE framework that leverages the temporal correlation properties of the channel to enhance the compression process. In particular, we propose an AE framework that performs temporal-spatial-frequency (TSF) compression by utilizing priors based on historical CSI samples to efficiently compress the current estimated CSI sample. End-to-end simulation results on a realistic test bench demonstrate the superiority of the proposed TSF compression approach relative to the state-of-the-art methods.