Sparse Channel Estimation for IRS-Assisted Communication System Based on Denoising Autoencoder

Abstract

Intelligent reflective surface (IRS) technology, as a promising technology, can enhance the communication quality of traditional communication systems. It is very necessary to design an effective channel estimation scheme, which is the basis for IRS to adjust the amplitude and phase of the incident signal. In this paper, the denoising autoencoder (DAE) is used to solve the channel estimation problem of IRS-assisted uplink millimeter wave channel from the user to the base station (BS). The two-stage channel model of user-IRS-BS can not be directly processed by deep learning networks. Therefore, in the channel modeling part, the original channel is unified into a cascaded channel, and the problem of channel estimation is transformed into the problem of restoring undersampled signals. In the part of network model training, an end-to-end channel estimation scheme based on stacked DAE is proposed. A group of orthogonal pilot signals received by BS is the input, and the channel state information (CSI) is the output. When the normalized mean square error of the output CSI and the original CSI is minimum,the training ends and the weight matrix model including the data hiding relationship is trained. In this paper, the training and verification of the channel estimation scheme is completed based on the self-simulated channel model data set. Simulation and performance evaluation show that our scheme is superior to several traditional schemes and a classical scheme based on deep learning, which confirms its superiority.