Deep Learning-based CSI Feedback for RIS-assisted Multi-user Systems

TL;DR

This paper proposes RIS-CoCsiNet, a novel deep learning-based framework aimed at significantly enhancing feedback efficiency in reconfigurable intelligent surface (RIS)-assisted wireless communications, and proposes magnitude-dependent phase feedback strategies that incorporate statistical or instantaneous CSI magnitude information into the phase feedback process.

Abstract

In the realm of reconfigurable intelligent surface (RIS)-assisted wireless communications, efficient channel state information (CSI) feedback is paramount. This paper introduces RIS-CoCsiNet, a novel deep learning-based framework designed to greatly enhance feedback efficiency. By leveraging the inherent correlation among proximate user equipments (UEs), our approach strategically categorizes RIS-UE CSI into shared and unique data sets. This nuanced understanding allows for significant reductions in feedback overhead, as the shared data is no longer redundantly relayed. Setting RIS-CoCsiNet apart from traditional autoencoder systems, we incorporate an additional decoder and a combination neural network at the base station. These enhancements are tasked with the precise retrieval and fusion of shared and individual data. And notably, all these innovations are achieved without modifying the UEs. For those UEs boasting multiple antennas, our design seamlessly integrates long short-term memory modules, capturing the intricate correlations between antennas. With a recognition of the non-sparse nature of the RIS-UE CSI phase, we pioneer two magnitude-dependent phase feedback strategies. These strategies adeptly weave in both statistical and real-time CSI magnitude data. The potency of RIS-CoCsiNet is further solidified through compelling simulation results drawn from two diverse channel datasets.

Figures (12)

• Figure 1: Representation of two scenarios using RIS-enhanced communication.
• Figure 2: Depiction of an RIS-enhanced multi-user setup featuring two proximate UEs.
• Figure 3: Schematic of the cooperative NN structure for two adjacent UEs, comprising two encoders and three decoders: two distinct and one shared decoder. The UEs' encoders produce the bitstream, which is then sent back through the uplink channel. At the BS, decoders utilize FC layers to extract shared RIS-UE CSI and specific information from the feedback bitstreams.
• Figure 4: Diagram of the NN architecture comprising the encoder and decoder. The encoder at the UE forms the bitstream using the FC layers and a quantization/binary layer, then relays this bitstream through the uplink channel. The BS decoder retrieves the RIS-UE CSI from the feedback bitstreams utilizing the FC layers.
• Figure 5: Schematics of MDPF-1 and MDPF-2. In MDPF-1, during training, the loss function is a weighted MSE, influenced by the RIS-UE CSI magnitude. This introduces statistical RIS-UE CSI magnitude data to the NNs. The quantization module post-encoder is omitted in this depiction. In MDPF-2, the RIS-UE CSI magnitude is fed into the encoder, and the final loss function for the RIS-UE CSI phase feedback NNs is the weighted MSE, contingent on the immediate CSI magnitude.