A Scalable Deep Learning Framework for Dynamic CSI Feedback With Variable Antenna Port Numbers
Yu-Chien Lin, Ta-Sung Lee, Zhi Ding
Abstract
Transmitter-side channel state information (CSI) is vital for large MIMO downlink systems to achieve high spectrum and energy efficiency. Existing deep learning architectures for downlink CSI feedback and recovery show promising improvement of UE feedback efficiency and eNB/gNB CSI recovery accuracy. One notable weakness of current deep learning architectures lies in their rigidity when customized and trained according to a preset number of antenna ports for a given compression ratio. To develop flexible learning models for different antenna port numbers and compression levels, this work proposes a novel scalable deep learning framework that accommodates different numbers of antenna ports and achieves dynamic feedback compression. It further reduces computation and memory complexity by allowing UEs to feedback segmented DL CSI. We showcase a multi-rate successive convolution encoder with under 500 parameters. Furthermore, based on the multi-rate architecture, we propose to optimize feedback efficiency by selecting segment-dependent compression levels. Test results demonstrate superior performance, good scalability, and high efficiency for both indoor and outdoor channels.