Channel Estimation For Millimeter-Wave Massive MIMO Systems Based On Deep Learnin

Millimeter-wave massive MIMO and reconfigurable intelligent surface are considered key technologies for future wireless communications as they improve the spectral efficiency and energy efficiency of the system.However,the deployment of large-scale antenna arrays and the increase of passive elements in the reflective surface will make channel estimation challenging,while the inability to obtain accurate channel state information can further decrease the overall performance of the communication system.Inspired by the powerful nonlinear fitting ability of deep learning,this paper adopts it to improve the channel estimation accuracy.The channel estimation algorithm based on residual learning and multi-path feature fusion is proposed for the problem that the conventional channel estimation algorithm faces huge pilot overhead due to the increased number of antennas in millimeter-wave massive MIMO systems.The received quantized signal at the base station is considered as a low-resolution image,and then the channel matrix is reconstructed using a deep learning-based image recovery technique.Specifically,the introduction of residual learning initially allows the estimator to focus on learning the high-frequency residual part between the quantized signal and the channel matrix to alleviate the training difficulty;secondly,dense connection is added to the residual block to maximize the information flow at each layer of the estimator;finally,the features of different scales extracted from the residual blocks are fused and preserved by multi-path feature fusion for channel prediction at the reconstructed module.Experiments show that this method can effectively save the pilot overhead and achieve good estimation performance in the low signalto-noise range.The channel estimation algorithm based on deep residual attention is proposed for reconfigurable intelligent surface-assisted wireless communication system,where the channel matrix dimension increases due to the increase of passive elements on the reflective surface.In order to improve the estimator’s ability to directly learn the non-trivial mapping of the received measurement signal at the base station to the cascaded channel,a residual attention block based on spatial and channel attention is designed to efficiently learn the characteristic distribution of the measurement signal and noise;the feature fusion is also improved by fusing the features extracted from neighboring residual attention blocks until all the features are used.The performance of the proposed method outperforms traditional algorithms and other deep learning methods.