Explainable Machine Learning-Based Layered CSI Feedback in Massive MIMO System
Gang Li, Shuangfeng Han, Zhuo Chen, Yupeng Li, Zirui Wen, Qixing Wang · 2024
In massive MIMO system, it is important for the base station (BS) to obtain channel state information (CSI) for precoding to improve beamforming gain in frequency division duplexing (FDD) systems. One of the biggest challenges is that CSI feedback overhead is remarkably huge especially for high rank number and meanwhile significantly fluctuates with the varying rank number. The uplink(UL) resources allocation is hard to match accurately the need for CSI feedback. Recent studies show that machine learning (ML) based CSI feedback has demonstrated impressive performance gain in CSI feedback overhead reduction. However, the compressed CSI feedback bits by ML models are not explainable and can not represent any physical meaning, and thus the feedback bits can't be treated dif-ferently by discarding insignificant beam coefficiencies. Inspired by scalable video coding (SVC) and semantic communication, in order to generate layered CSI feedback bits, we propose a segment-level loss design and an iterated training strategy for ML based CSI feedback. Simulation results show that the different segments of the compressed CSI feedback bits demonstrate distinct importance weight for the CSI recovery accuracy, which brings great transmission efficiency improvement. Moreover, adaptive modulation and coding (AMC) enabled CSI segments is proposed to carry as more segments as possible and simulation results show that the CSI feedback accuracy can be improved greatly especially under relatively good SNR.