Channel Estimation for Intelligent Reflecting Surface Aided Communication via Graph Transformer

Intelligent reflecting surface (IRS) is a potential technology for enhancing communication systems' performance. Accurate cascaded channel estimation between the base station (BS), IRS, and the user is vital for optimal system performance. However, incorporating IRS increases channel estimation complexity due to additional dimensions from each element, leading to higher training overhead. To reduce training overhead, existing approaches assume the sparse cascaded channel which may not be valid in dense multipath propagation and non-line-of-sight settings. We propose a novel technique to address this issue by leveraging the spatial correlation among IRS elements' channels. By dividing the IRS surface into groups, we estimate the channel for some groups via the least square (LS) method. To estimate the channels for the remaining groups, a graph transformer-based IRS channel estimation (G-TIRC) model is proposed, which includes a graph neural network (GNN) and transformer model. The GNN finds the correlations among the different groups by embedding the channel information. Then, the attention mechanism within the transformer extracts useful correlations to accurately predict the channels for the unknown groups. The experiments demonstrate the effectiveness of the G-TIRC model in achieving accurate channel estimation with reduced pilot overhead compared to other state-of-the-art methods.