SNAP-CSI: Personalized Neural Compression for Enhanced CSI Compression in Wireless Networks

作     者：Askar, Nurassyl Rini, Stefano 

作者机构：Natl Yang Ming Chiao Tung Univ NYCU Dept Elect & Comp Engn Hsinchu 300 Taiwan 

出 版 物：《IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS》 (IEEE Trans. Wireless Commun.)

年 卷 期：2025年第24卷第3期

页      面：2083-2093页

核心收录：

学科分类：0810[工学-信息与通信工程] 0808[工学-电气工程] 08[工学] 

基　　金：National Science and Technology Council, Taiwan [NSTC 111-2221-E-A49-068-MY3] Media Tek Advanced Research Center (MARC) 

主　　题：Clustering algorithms Wireless communication Quantization (signal) Wireless networks OFDM Computer architecture Accuracy Training Estimation Decoding Channel state information CSI compression wireless networks deep neural networks personalized models heterogeneous environments 

摘      要：This paper introduces the Selection Network Assisted Personalized CSI Compression (SNAP-CSI) algorithm, a novel approach for efficient Channel State Information (CSI) compression in wireless networks. Focusing on scenarios where CSI from User Equipment (UE) is transmitted to a Base Station (BS) via a rate-limited channel, SNAP-CSI employs Deep Neural Networks (DNNs) trained on historical CSI data for enhanced compression. Central to SNAP-CSI is the exploitation of CSI heterogeneity to cluster users, enabling the training of tailored personalized models. These models comprise an encoder at the UE and a decoder at the BS, optimized for efficient compression with minimal parameters, specific to each user cluster. A key innovation in SNAP-CSI is the development of a Selection Network (SN). This network predicts cluster membership from compressed CSI data, allowing UEs to select the most fitting personalized model for the lowest data distortion. Concurrently, the BS utilizes the SN for accurate reconstruction of compressed CSI, thus negating the need for further synchronization. The effectiveness of SNAP-CSI is validated through simulations with the ultra-dense indoor maMIMO dataset, evaluating performance across diverse heterogeneity conditions and UE-to-BS channel rates.