Research On Precoding In Cell-Free Massive MIMO Systems Assisted By Intelligent Reflecting Surface

Cell-Free Massive Multiple-Input Multiple-Output(MIMO)revolutionizes the cellular network architecture and effectively addresses the spectrum efficiency and energy efficiency bottlenecks faced by mobile communications.Cell-Free Massive MIMO distributes a large number of Access Points(APs)with one or more antennas over a large area,transmits data to the Central Processing Unit(CPU)via a backhaul link,and provides uniform and reliable service to multiple users using the same timefrequency resources.Intelligent Reflective Surface(IRS)is a new revolutionary technology that can significantly improve the performance of wireless communication networks by intelligently reconfiguring the wireless propagation environment through the integration of a large number of lowcost passive reflective elements on a flat surface.Cell-Free networks are highly resistant to inter-cell interference,yet further network capacity increases require the deployment of more APs,resulting in high cost and power consumption.Inspired by the IRS technology,the key idea of introducing it into the Cell-Free Massive MIMO system is to replace some APs with low-cost and energy-efficient IRS to increase the network capacity.Since the wireless environment can be manipulated efficiently with low cost and low energy consumption,IRS can be used to increase channel capacity,reduce transmit power,enhance transmission reliability,and expand wireless coverage in Cell-Free Massive MIMO system.In this thesis,we study the precoding problem in downlink IRS-assisted Cell-Free Massive MIMO systems,and maximize the system sum rate by jointly optimizing the precoding and phase shift matrices to ultimately improve the system performance.The main research includes the following two points:First,we studied IRS-assisted Cell-Free Massive MIMO system,in which low-resolution digitalto-analog converters(DACs)are used at each AP.On the one hand,the IRS is combined with the Cell-Free Massive MIMO system to reduce the system hardware cost and power consumption by replacing some APs with low-cost and low-power IRSs.Alternatively,the utilization of lowresolution DACs in APs can decrease hardware expenses and power consumption even further.In this thesis,we adopt an additive model for quantization noise to mathematically characterize the expressions for the sum rate of downlink users,and the peak of the sum rate is achieved by alternating active and passive beam assignment optimization.Due to the nonconvexity and high complexity of the formulation,we propose an alternating optimization framework to solve this complex problem.In particular,we decouple this problem by fractional programming and use the Lagrange multiplier method and Semi-definite Programming(SDP)method to derive the expressions of the precoding and phase shift matrices.Lastly,simulation results demonstrate that the proposed approach enhances network capacity considerably in comparison to the traditional Cell-Free network.Second,since the above study requires complex mathematical derivations and formula operations,and the computational effort increases exponentially with the increase of the IRS reflection elements,so we propose to use Deep Reinforcement Learning(DRL)to solve this sum rate maximization nonconvex optimization problem,and investigates the joint design of the precoding matrix and phase shift matrix.Our first contribution is the introduction of a Deep Reinforcement Learning(DRL)algorithm that learns the optimal joint design by interacting with the environment and observing predefined rewards in a continuous state-action space.Unlike the above work,which alternately optimizes the precoding and phase shift matrices,the proposed DRL-based algorithm is less complex and can simultaneously obtain the joint design of the precoding and phase shift matrices as the output of the DRL neural network.The experimental results show that the algorithm is able to learn the environment by observing the instantaneous returns and improve its behavior gradually to obtain the optimal precoding and phase shift matrices and improve the system performance by maximizing the user sum rate.