CNN based Sparse IRS Design for Channel Estimation in Assisted Uplink Communications

Intelligent reflecting surface (IRS) has become a promising technology in wireless communications benefiting from its low cost and high communication efficiency. Passive IRS encounters the challenge of channel state estimation, which can be overcome by connecting IRS elements to receive channels to construct semi-active IRS, thus enabling to estimate the user-to-IRS and IRS-to-base station (BS) channels separately. As the channel estimation is directly related to IRS configuration, we propose a convolutional neural network (CNN) based sparse IRS design to assist the uplink communications in this work. A given number of IRS elements are switched on to connect with the receive channels for enhanced user-to-IRS channel state estimation while reducing the cost simultaneously. The sparse IRS design is formulated as a classification problem and we focus on designing the CNN structure. In addition, we generate the dataset using the Cramer-Rao Bound (CRB) of direction of arrival (DOA) estimation as an indicator to train the CNN. When labeling the training data, a new labeling strategy that can greatly reduce the number of classes is proposed and thus improving the classification accuracy. Numerical results are provided to demonstrate the adaptability and effectiveness of the proposed CNN based sparse IRS design in arbitrary unkown environment.