Research On Hybrid Precoding Based On Neural Network In Massive MIMO Systems

In recent years,the demand for mobile data traffic has shown explosive growth.To meet the requirement of high-speed data transmission,massive MIMO and millimeter wave(mm Wave)communications have received extensive attention.Mm Wave communications need to take advantage of large antenna arrays,made possible by the decrease in wavelength,to combat high pass loss with beamforming gain obtained through precoding.However,dozens or hundreds of antennas are ususlly equipped in massive MIMO systems,which makes the power consumption,cost and implementation complexity of full digital precoding too high,while analog precoding suffers from severe spectral efficiency loss.Therefore,hybrid precoding is proposed to balance the hardware complexity and the performance.This thesis focuses on the energy-efficient hybrid precoding scheme,and designs low-complexity hybrid precoding algorithms with the help of excellent learning ability of neural networks.The main contents are as follows:Firstly,to reduce power consumption in the analog domain,instead of phase shifters,phase modulation array(PMA)is employed to implement analog precoding,on the basis of which,an energy-efficient hybrid precoding structure based on adaptive switch module and phase modulation array(ASM-PMA)is proposed,in which ASM is used to flexibly adjust the connection state between the antennas and the RF links to maximize the energy efficiency.The ASM-PMA hybrid precoder design problem related to three matrix variables is formulated.Owing to the tough task of joint optimization of matrix variables,we can divide it into two sub-problems.Correspondingly,an alternating optimization algorithm is suggested to solve PMA-based hybrid precoder.Then,matching theory and convolutional neural network(CNN)based algorithms are respectively proposed to seek the optimal ASM.For the CNN based scheme,relaxed adaptive switching CNN(RAS-CNN)is designed to solve the binary integer problem in the offline train stage,while in the online deployment stage,ideal adaptive switching CNN(IAS-CNN)is adopted.Furthermore,CNN accepts estimated channel as input to yield robust precoding performance.Simulation results verify that,the proposed hybrid precoding based on ASM-PMA can outperform other schemes in energy efficiency with satisfactory spectral efficiency.Secondly,in order to solve the problem that mm Wave signals are easily blocked by obstacles,an intelligent reflecting surface(IRS)assisted MIMO wireless communication system with low energy consumption is studied.Above all,the structure and characteristics of IRS are introduced,and the power gain brought by IRS is analyzed.Then,the IRS-assisted singleuser and multi-user hybrid precoding system architectures are expounded,and the problem models are formulated.Accordingly,maximizing array gain(MAG)based algorithm and maximizing equivalent channel gain(MECG)based algorithm are proposed to solve the reflection matrix of IRS,respectively.Aiming at reducing the computational complexity,deep neural network(DNN)based scheme is also proposed to design IRS.Furthermore,it is demonstrated that the achievable spectral efficiency of the PMA-based single-user fullyconnected structure is the same as that of digital precoder.For the multi-user scenario,analog precoder can be derived by maximizing beamforming gain,and zero-forcing(ZF)algorithm is employed to eliminate inter-user interference at baseband.Simulation results show that the IRS-assisted communication system can significantly improve the spectral efficiency at a low power consumption cost,and the enviable performance of the proposed algorithms is verified.