Low-complexity Algorithms And Implementations For Massive MIMO Detection

As wireless communication technology is developing dramatically in recent years,the 5-th generation(5G)mobile communication technique has been a hot topic for researchers.In this thesis,we focus mainly on the two key techniques of 5G communication systems,namely massive MIMO and polar code.For massive multiple-input multiple-output(MIMO)systems,linear minimum mean square error(MMSE)detection is near-optimal but suffers from prohibitive high complexity due to the large-scale matrix inversion.To this end,Neumann series(NS)expansion approximation,which avoids the direct computation of the matrix inversion,is recently investigated due to its low implementation complexity.Unfortunately,the complexity reduction can only be achieved well when the required number of the NS terms is small.To solve this problem,we proposed an soft-output iterative NS(SINS)algorithm for MMSE detection at a manageable complexity even for large NS terms.An approximation method based on NS is proposed to compute the log-likelihood ratios(LLRs)for channel decoders.Both analytic and numerical results have demonstrated that,the overall complexity of the proposed MMSE-SINSE algorithm is significantly reduced compared with the conventional NS method and the Cholesky decomposition method,while keeping similar detection performance.Being matrix inversion free,detection algorithms based on Gauss-Seidel(GS)method have been proved more efficient than conventional NS based ones.Therefore,we propose an improved GS-based soft-output detection(IGS)and its VLSI architecture for massive MIMO.To accelerate the convergence,a new initial solution is proposed for GS method.Several optimizations on the VLSI architecture have been conducted to further reduce the latency and area.The reference implementation results on a Xilinx Virtex-7 XC7VX690 T FPGA for a 128 × 8MIMO system show that our GS-based detector can achieve a throughput of 732 Mb/s and a close-to-MMSE performance.Numerical and implementation results have shown that our design has advantages in terms of complexity and efficiency,especially in poor propagation environments.Furthermore,based on the GS detection algorithm,we propose the parallel Gauss-Seidel(ParGS)iterative method,and achieves comparable detection performance as MMSE detection.This algorithm successfully avoids explicit matrix inverse in uplink,and therefore effectively reduces the computational complexity of the massive MIMO detection problem.The proposed efficient architecture is able to reduce the processing latency per iteration.Furthermore,this architecture is scalable and can be easily reconfigured as the number of antennas increases.Reference implementation results on a Xilinx Virtex-7 FPGA for a 128×8 massive MIMO system demonstrate the advantages of the proposed PGS detector in terms of hardware efficiency.We also propose an improved and low-complexity signal detection approach for massive MIMO systems,namely MMSE-PreGS.This approach utilizes the preconditioning technique to accelerate the conventional detection algorithm based on GS iterative method,and achieves a detection performance close to MMSE detection algorithm with relatively small iteration counts.It also outperforms the counterparts based on NS and the conjugate gradient(CG)method in poor propagation environments,such as MIMO systems with large loading or correlated factors.The corresponding architecture is also proposed with both novelty and scalability.It takes advantage of the cyclic-shift property of the GS method,and therefore facilitates the hardware implementation.Bot numerical results and complexity analysis demonstrate that the proposed detector is efficient and suitable for massive MIMO systems.Since the conventional detection approaches for massive MIMO failed to exploit neither the special structure of channel matrices nor the critical issues in the hardware implementation,which results in poorer throughput performance and longer processing delay,we propose a modified NS based on tridiagonal matrix inversion approximation(TMA)to accommodate the complexity as well as the performance issue in the conventional hardware implementation,and analyze the corresponding approximation errors.Meanwhile,we investigate the VLSI implementation for the TMA algorithm based on a Xilinx Virtex-7 XC7VX690 T FPGA platform.It is shown that for correlated massive MIMO systems,the TMA detector can achieve nearMMSE performance and 630 Mb/s throughput.Compared with other benchmark systems,the proposed pipelined TMA detector can get high throughput-to-hardware ratio.In this thesis,we also propose a family of adjustable detection approaches and their VLSI architectures for massive MIM systems,which are flexibly-configurable and low-complexity for implementation.We extend the conventional NS-based and GS/SOR-based detection approaches for massive MIMO systems,and propose the more general Adjustable Linear Detection algorithm(ALD),which enables flexible convergence rates of these iterative methods.We further improve the performance of ALD,and introduce several techniques,which are also adjustable,to achieve better detection performance than traditional methods in poor propagation environments.We generalize these adjustable detection approaches as the Fully Adjustable Linear Detection algorithm(FALD)in order to meet various requirements.In addition,we propose a low-complexity algorithm adapted to each user equipment(UE)equipped with multiple antennas in massive MIMO uplink linear detection.The NS based algorithm skillfully utilizes the serious correlation between antennas from the same UE to achieve rapid convergence and innovatively operates multiple levels iterations to obtain preferable biterror-ratio(BER)performance with relatively small iteration counts.Simulation results demonstrate that the proposed algorithm can reduce computational complexity by 44% compared with the block matrix(BM)algorithm and achieve near MMSE detection performance in spite of poor propagation environments.Also,an efficient VLSI architecture for proposed algorithm is provided.Deep learning recently shows strong competitiveness to improve polar code decoding.However,suffering from prohibitive training and computation complexity,the conventional deep neural network is only possible for very short code length.In this thesis,the main problems of deep learning in decoding are well solved.We first present the multiple scaled belief propagation(BP)algorithm,aiming at obtaining faster convergence and better performance.Based on this,deep neural network decoder(NND)with low complexity and latency,is proposed for any code length.The training only requires a small set of zero codewords.Besides,its computation complexity is close to the original BP.Experiment results show that the proposed NND with constant 5 iterations achieves even lower BER than the 50-iteration conventional BP.The hardware architecture of basic computation block is given and folding technique is also considered,saving about 50 % hardware cost.