Study on CSI Compression and Restoration with Deep Learning in RIS-Assisted MIMO Systems

Intelligent reflective surfaces (IRS) have been widely studied due to their advantages such as low cost, low power consumption, and ability to improve communication quality. In this paper, an IRS-assisted multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing (OFDM) communication system is considered. In order to improve the performance gain of the system, the user (UE) needs to send its channel state information (CSI) of several channels to the base station (BS) via feedback link. Therefore, the data volume and feedback overhead of CSI in this system will undoubtedly be much huger, as compared to the conventional MIMO systems. To address this problem, we propose an attention-based deep residual network named IARNet (In-ception-Attention-Residual-Net) to compress and reconstruct the CSI with large data volume. The IARNet combines several sub-modules based on the traditional Inception network, such as the multi-convolutional feature fusion module, the hybrid attention module, and the residual module, etc. This hybrid structure can effectively compress and reconstruct the CSI of large data volumes. Simulation results show that with the warm-up training scheme, IARNet can significantly improve the reconstruction quality of CSI of large data volumes, as compared to two existing deep learning networks.