RISNet: A DL-Based High Accuracy CSI Feedback Approach for RIS-Aided FDD Systems

Reconfigurable Intelligent Surfaces (RIS) have become a prominent research topic in recent years. However, downlink channel state information (CSI) feedback in RIS-aided frequency division duplex (FDD) system is a considerable challenge due to the expansion of the cascaded channel dimensions. In this letter, a novel deep learning-based CSI feedback architecture named RISNet is proposed to address this issue. RISNet is designed to operate in situations where the direct link between the base station (BS) and the user equipment (UE) is obstructed, and RIS is employed to direct signals towards the UE. To achieve this, an encoder is placed at the UE end to compress the BS-RIS-UE cascaded channel to a codeword and feed it back to the BS. Before this, the variation in the hybrid domain and principal component mark are designed to reduce the complexity of encoder. At the BS end, a designed decoder is used to recover rich features with various receptive fields from the feedback codeword to reconstruct high-precision CSI. Simulations show that the proposed RISNet exhibits superior accuracy compared to existing work with 3.73 dB lower mean squared error and 0.34 higher cosine similarity, while also displaying less computational complexity on the UE side.