Signal Detection Algorithms And Implementations For Massive MIMO Systems

In recent years,global information and communication industry has become increasingly intelligent.As an extension version of the 4th generation mobile communication system(4G),the 5th generation mobile communication system(5G)has become a hot topic in the field of communication at home and abroad.Among many key technologies,massive MIMO is the most concerned physical layer technology below 6 GHz in the next generation wireless communication system,whose core idea is to employ a small number of antennas at the base station side to serve multiple independent user terminals at the same time.Nowadays,large-scale MIMO system has attracted the attention of industry and academia due to its remarkable advantages in spectrum efficiency,energy efficiency and link reliability,and signal detection at the receiver side is an important link of large-scale MIMO system.Taking massive MIMO system uplink as an example,the received signal at the base station is not only from a single user terminal,but also the superposition of multiple users'signals.Therefore,the computational complexity of signal detection in large-scale MIMO system is much greater than that of traditional small-scale MIMO system.In this thesis,massive MIMO detection is deeply studied in terms of message passing algorithm and linear detection methods.To balance the performance and complexity,several effective optimization schemes is firstly proposed.With the development of information technology and the interdisciplinary integration,academia and industry have gradually realized that many problems in the field of information engineering are essentially probabilistic inference problems.Information propagation algorithm is an important approach to solve the probabilistic inference problem,which is represented by belief propagation(BP)algorithm proposed by Pearl in 1988.With the growing of the number of antennas in large-scale MIMO systems,the information trans-mission network of BP algorithm becomes more and more complex,then more probability information needs to be processed.In order to further reduce the computational complexity of conventional BP detection al-gorithm,a low-complexity BP detection based on max-sum(MS)algorithm is proposed for massive MIMO systems in this paper.In addition,by introducing the normalization factor and the offset factor,two effective strategies are employed to improve the approximate accuracy of MS algorithm,which leads to normalized max-sum(NMS)algorithm and offset max-sum(OMS)algorithm.Additionally,an efficient hardware archi-tecture is designed for the proposed NMS/OMS detector.Meanwhile,FPGA implementation results show that the proposed NMS/OMS detector has higher hardware efficiency and greater data throughput than the traditional BP detector.When the number of antennas is relatively large,large-scale MIMO system exhibits channel hardening characteristics,which means the eigenvalues of the channel matrix tend to be more stable as the number of antennas increases.In other words,the channel matrix is increasingly approaching a diagonal matrix if MIMO system gets more“massive”.In this thesis,a new system model is established based on the channel hardening characteristics of massive MIMO.Based on the new system model,an improved low complexity BP-CH detection algorithm is firstly proposed.For different antenna configurations and modulation modes,we compare the BER performance of the traditional BP algorithm and proposed BP-CH algorithm.The simulation results show that,under the condition of a large number of antennas,the BP-CH detection algorithm can obtain almost the same error rate as the traditional BP algorithm while keeping lower computational complexity.In addition,based on the reduced factor graph of BP-CH algorithm,an efficient folded hardware architecture of BP-CH detector is proposed.The FPGA implementation results show that,compared with the traditional BP detector,the BP-CH detector and the folded BP-CH detector have obvious advantages in hardware efficiency,especially the folded BP-CH detector.The expectation propagation(EP)algorithm is a Bayesian network-based inference algorithm proposed by Tomas P Minka in his doctoral thesis in 2001.Since the large-scale MIMO signal detection problem is essentially a Bayesian inference problem,the EP algorithm can be applied to solve MIMO signal detection problem.However,since the received signal is a multi-dimensional vector for large-scale MIMO system,EP algorithm always involves the intractable inversion operation of the high-dimensional matrix when calculating the characteristic parameters of the posterior probability distribution.In order to solve the high-dimensional matrix inversion problem,an improved EP detection is first proposed in this paper,which takes into account both high performance and low complexity for massive MIMO,namely EP-NSA detection algorithm.The proposed EP-NSA detection algorithm innovatively combines the traditional polynomial series expansion with the EP algorithm to achieve an approximate inversion process for the covariance matrix,thereby greatly reducing the computational complexity of the traditional EP detection algorithm.In this thesis,two impor-tant factors affecting the convergence rate of the EP-NSA algorithm are analyzed.We also give suggestions that how many Neumann-series expansion terms to be retained according to different scenario requirements.Therefore,EP-NSA algorithm with different Neumann terms can flexibly adapt to different system configu-rations and propagation environments.In reality,only two Neumann-series expansion terms are needed,EP-NSA detection have lower com-putational complexity than traditional exact EP detection.For EP-NSA detection with three Neumann-series expansion terms,its complexity is on the same order of magnitude as that of the exact matrix inversion.In this thesis,EP-NSA detection is further improved by iterative Neumann series expansion,namely EP-INSA algorithm.In the EP-INSA detection,we no longer view the high-dimensional matrix inversion as a single problem,but jointly consider the matrix inversion and the characteristics statistics updating.Specifically,we replaced one-time inversion with iterative inversion,and the result of the previous iteration can be used to calculate the result of the next iteration.Regardless of the number of Neumann expansions terms to be retained,the computational complexity of EP-INSA detection is controlled at the O(M~2)level.Owing to that the update formula of mean and variance in EP-INSA algorithm has iterative characteristics,an efficient IIR-like filter hardware architecture is designed for EP-INSA detector,and the timing schedule of EP-INSA detector is analyzed.As an important branch of the MIMO detection family,linear detection algorithms(ZF,MMSE,etc.)have high application value in small-scale MIMO systems because they can achieve better detection per-formance with lower complexity.However,due to the inevitable inversion operation of high-dimensional matrices in linear detection methods,excessive computational complexity has become a primary factor lim-iting their widespread application in large-scale MIMO systems,especially when there are many users on the uplink.In this paper,we discussed MMSE detection based on iterative numerical method in large-scale MIMO systems,and proposed an improved Jacobi iterative algorithm.Firstly,we synthetically discuss the characteristics and convergence of various MMSE detection algorithms based on iterative values,and com-pare their complexity and BER performance under different correlated channels.Then an improved damped Jacobi iteration algorithm is proposed,which takes account of both the fast convergence of GS iteration algo-rithm and the easy parallelization of DJ iteration.The core idea of the improved DJ is to utilize GS iteration to calculate the fast convergence initial solution.Finally,a general linear detector architecture based on iterative numerical algorithm is designed,which includes preprocessing module,main computing module and iteration update module.The hardware resource consumption and scheduling are analyzed in detail.