Multi-Scale Attention Based Channel Estimation for RIS-Aided Massive MIMO Systems

A multi-scale attention based channel estimation framework is proposed for reconfigurable intelligent surface (RIS) aided massive multiple-input multiple-output systems, in which hardware imperfections and time-varying characteristics of the cascaded channel are investigated. By exploiting the spatial correlations of different scales in the RIS reflection element domain, we construct a Laplacian pyramid attention network (LPAN) to realize the high-dimensional cascaded channel reconstruction with limited pilot overhead. In LPAN, we leverage the multi-scale supervision learning to progressively capture the spatial correlations of the cascaded channel, where the attention mechanism based dual-branch architecture is designed. To balance network performance and complexity of LPAN, we further propose a lightweight LPAN-L architecture. In LPAN-L, the partial standard convolutional layers are decomposed into the group convolution, dilated convolution and point-wise convolution, which forms a sparse convolutional filter set to extract the channel feature with less computation cost. Furthermore, we leverage parameter sharing and recursion strategy to reduce the space complexity. Moreover, a selective fine-tuning strategy is developed to realize the domain adaption. Simulation results show that the proposed LPAN can achieve higher estimation accuracy than the existing estimation schemes, while the LPAN-L architecture with a close performance to LPAN efficiently reduces the network complexity. The code is available at
https://github.com/Holographic-Lab/LPAN
.