DCSaNet: Dilated Convolution and Self-Attention-Based Neural Network for Channel Estimation in IRS-Aided Multi-User Communication System

Abstract

Channel estimation (CE) is a critical part for intelligent reflecting surface (IRS) aided multi-user communication (MUC) systems. However, the cascaded channel of the IRS-aided MUC (IRS-MUC) is a complex multi-dimensional channel, which is difficult to estimate the precise channel matrix in practice. In this letter, we propose a dilated convolution and self-attention based neural network (DCSaNet) to handle the CE in the IRS-MUC system. Specifically, the dilated convolution block is used to improve the feature extraction for CE during the network training. Further, the weighed features are obtained by the self-attention block. Last, a lightweight model, DCSaNet-l is applied to reduce the network parameters for the practical IRS deployment. Experimental results show that the proposed DCSaNet can significantly lower the normalized mean square error (NMSE), accelerate the training speed under different SNR cases and different channel dimensions. The results also verify that the lightweight DCSaNet-l can reach a near optimal performance of the proposed DCSaNet, but further significantly reduce the parameter amount by more than 83.7%.