Channel Estimation For RIS-aided Mm Wave Massive MIMO Systems Based On Deep Learning

Reconfigurable intelligent surface(RIS)can meet the needs of ultra-high mobile broadband,ultra-low communication latency,and ultra-reliable communication with its unique advantages,which have become a key candidate technology for the sixth generation(6G)wireless communication.However,due to the huge number of RIS units,the difficulty of uplink and downlink channel estimation in a RIS-aided millimeter-wave(mm Wave)massive multiple-input multiple-output(MIMO)system increases sharply,which becomes the key factor to restrict the system performance.Aiming at the abovementioned issues,this paper studies the channel estimation method of the RIS system from the perspective of RIS architecture and mobility.First,a semi-passive RIS with signal sensing and processing in part of units is analyzed.Considering the low sparsity of mm Wave channels,a semi-passive RIS-aided massive MIMO channel estimation method is proposed,including the enhanced single-scale and multi-scale super-resolution neural network.It is found that when the signal-to-noise ratio(SNR)is in the range of-10 to 40 d B,the proposed method has an improvement of5.24 to 28.17 d B compared with the orthogonal matching pursuit and alternating multiplier method,and the algorithm complexity is reduced by one order of magnitude.Secondly,we analyze a fully passive RIS in which all units can only reflect but have not to sense.Taking advantage of the sparsity of cascaded channels,a multiple residual dense network(MRDN)is proposed,and compared with the convolution blind denoising channel estimation,parallel factor analysis,etc.It is found that when the SNR is in the range of 10 to 30 d B,the MRDN has an improvement of 5.61 to 10.30 d B compared with the traditional deep learning method;compared with the alternating multiplier method,it has an improvement of 21.03 to 32.55 d B,and the complexity is reduced by one order of magnitude.Then,we analyze a semi-passive RIS system based on high-speed scenarios,considering the characteristics of Doppler frequency shift,fast time-varying channel,and short coherence time,we propose a deep deterministic policy gradient(DDPG)method for channel estimation and channel tracking at multiple coherence times.Using the sparse characteristic of the mm Wave channel under Doppler frequency shift,we can estimate the mm Wave channel through the discrete angle and path gain.It is found that when the SNR is in the range of-10 to 30 d B,the DDPG method has an improvement of 8.35 to 30.69 d B compared with sparse Bayesian learning,and is better than MRDN at high SNR.Finally,based on a fully passive RIS system in high-speed scenarios,a cascaded channel based is considered.We can estimate the mm Wave channel by estimating the sparse discrete angles and joint path gains from users and base stations to the RI.And we propose a channel estimation framework based on the maximum entropy model.The study found that when the SNR is in the range of-10 to 40 d B,the proposed algorithm has an improvement of more than 10 d B compared with the iterative reweighting-based super-resolution method,and has an improvement of about 4 d B compared with DDPG.In summary,this paper studies deep learning-based channel estimation methods in low-speed and high-speed scenarios for fully passive and semi-passive RIS-aided mm Wave massive MIMO.The research results provide a basis for the design of RIS-aided mm Wave massive MIMO systems,which will help the development of future 6G mobile communication systems.