Channel Estimation for Semi-Passive Reconfigurable Intelligent Surfaces With Enhanced Deep Residual Networks

Abstract

Reconfigurable intelligent surface (RIS) is envisioned as an essential paradigm for realizing the sixth-generation networks, due to the use of low-cost reflecting elements for establishing programmable and favourable wireless environment. However, accurate channel estimation is a fundamental technical challenge for achieving large performance gains brought by RIS. To address this challenge, we first integrate a RIS with a small number of uniformly distributed active sensing devices, which are equipped with active radio frequency chains for acquiring partial channel state information (CSI). Then, by leveraging the rank-deficient structure of RIS channels, two practical residual neural networks, named single-scale enhanced deep residual (EDSR) and multi-scale enhanced deep residual (MDSR), are proposed to obtain accurate CSI, which can strike a balance between the system complexity and estimation performance. Simulation results reveal the cost-performance trade-off of the two proposed methods and unveil their superior performance compared with existing baseline schemes.