Channel Estimation for RIS-Assisted Time-Varying MIMO System: An Attention-Based Learning Approach

The challenges of time-varying channel estimation in reconfigurable intelligent surface (RIS)-based MIMO systems are focused on in this paper. We propose a scheme based on deep learning, which includes RIS element selection model (RESM) and a full channel estimation model (FCEM). Firstly, The RESM is designed for choosing the optimal subset of all RIS elements to reduce system overhead. A convolutional network based scorer is established to evaluate relationship between optimal partial channels and full channels, and the differential Top-N operation is used to select the optimal subset of RIS elements, where perturbed maximum method is utilized to ensure end-to-end learning is feasible. Then, the FCEM is developed to realize accurate time-varying channel estimation by exploiting the strong nonlinear fitting capability of attention based deep networks. We develop network structures including improved transformers and residual blocks in the FCEM to counteract the channels' stochastic characteristic, so as to recover the full channels corresponding to the optimal subset. The numerical results demonstrate the proposed scheme outperforms the benchmark schemes under various comparison conditions and is suitable for high-speed mobile scenarios.