机器学习提升大规模MIMO传输链路性能的方法

论文总字数：34587字

摘 要

在5G移动通信系统中，大规模MIMO传输技术是提升系统链路性能的重要手段之一，但是由于天线数量的增加导致计算复杂度成指数级增长，给系统的实现带来了困难。机器学习是通过观察系统大量输入输出，从而可以用通用计算架构模拟系统输入输出行为的方法，可以有效降低大规模MIMO传输链路的计算复杂度，因而得到了广泛研究。本文围绕用机器学习方法提升大规模MIMO传输链路的性能展开了研究，主要工作如下：

一、对机器学习方法及其在移动通信系统中的应用进行了综述。介绍了机器学习的概念、三要素以及分类，对常用的机器学习架构——神经网络架构进行了详细介绍，还对机器学习在移动通信系统中的应用进行了综述，分别从信道估计和预测、业务类型和数量预测、物理层以及无线资源分配方面进行了介绍。

二、基于卷积神经网络的大规模MIMO信道估计方法。针对大规模MIMO有导频信号的情况，提出了基于压缩图像恢复领域的DAMP算法的神经网络机器学习架构（L-DAMP）。该架构使用基于卷积神经网络的DnCNN降噪器替换DAMP算法中的降噪器。仿真结果表明，使用DnCNN降噪器的L-DAMP机器学习神经网络架构收敛速度很快，而且能够达到比最先进的使用BM3D降噪器的DAMP算法更好的性能。

三、基于神经网络的Alamouti编码信号检测方法。针对的MIMO系统，提出了一种用于未知CSI的Alamouti检测的ML-NN检测结构，该检测结构采用了四层神经网络。借助TensorFlow和MATLAB，在准静态瑞利信道条件和QPSK调制条件下，仿真了这种ML-NN结构和已知CSI的最大似然检测。仿真结果表明，对于Alamouti检测，相比于传统的未知CSI的最大似然检测，ML-NN结构检测的BER性能更好，但是是以增加训练时间和训练符号开销为代价的，除此之外，ML-NN结构检测的BER性能随着训练符号长度的增加而提高，ML-NN（32）结构检测的BER性能已经非常接近已知CSI的最大似然检测。

关键词：大规模MIMO，机器学习，神经网络，信道估计，信号检测

Abstract

In the 5G mobile communication system, massive MIMO transmission technology is one of the important means to improve the performance of the system link. However, the computational complexity increases exponentially due to the increase in the number of antennas, which brings difficulties to the implementation of the system. Machine learning is a method of observing a large number of input and output of the system, so that the general computing architecture can be used to simulate the input and output behavior of the system, which can effectively reduce the computational complexity of large-scale MIMO transmission links, and thus has been extensively studied. This paper focuses on the use of machine learning methods to improve the performance of large-scale MIMO transmission links. The main work is as follows:

First, the machine learning method and its application in mobile communication systems are reviewed. This paper introduces the concept, three elements and classification of machine learning, introduces the common machine learning architecture—the neural network architecture, and also summarizes the application of machine learning in mobile communication systems, from channel estimation and prediction, respectively. Business type and quantity prediction, physical layer, and wireless resource allocation are introduced.

Second, a massive MIMO channel estimation method based on convolutional neural networks. Aiming at the situation that the massive MIMO has pilot signals, a neural network machine learning architecture (L-DAMP) based on the DAMP algorithm in the field of compressed image restoration is proposed. The architecture replaces the denoise in the DAMP algorithm with a DnCNN denoise based on a convolutional neural network. The simulation results show that the L-DAMP machine learning neural network architecture using DnCNN denoise has a fast convergence speed and can achieve better performance than the most advanced DAMP algorithm using BM3D denoise.

Third, the neural network based Alamouti coded signal detection method. For the MIMO system, an ML-NN detection structure for Alamouti detection of unknown CSI is proposed. The detection structure uses a four-layer neural network. With TensorFlow and MATLAB, the maximum likelihood detection of this ML-NN structure and known CSI is simulated under quasi-static Rayleigh channel conditions and QPSK modulation conditions. The simulation results show that for Alamouti detection, the BER performance of ML-NN structure detection is better than that of the traditional unknown CSI maximum likelihood detection, but at the cost of increasing training time and training symbol overhead. The BER performance of the ML-NN structure detection increases with the increase of the training symbol length, and the BER performance of the ML-NN (32) structure detection is very close to the maximum likelihood detection of the known CSI.

KEY WORDS: massive MIMO, machine learning, neural network, channel estimation, signal detection

目 录

摘要 I

Abstract III

目 录 V

图目录 IX

表格目录 XI

缩略语目录 XIII

第一章 绪论 1

1.1 机器学习的概念 1

1.1.1 人工智能的概念 1

1.1.2 机器学习的概念 2

1.1.3 与人工智能的关系 3

1.2 机器学习的三要素 4

1.2.1 模型 4

1.2.2 策略 5

1.2.3 算法 5

1.3 机器学习的分类 6

1.3.1 确定算法参数的机器学习 6

1.3.2 优化算法参数的机器学习 7

1.4 神经网络 7

1.4.1 模型 7

1.4.2 策略 10

1.5 机器学习在移动通信系统中的应用 11

1.5.1 在信道估计和预测方面的应用 11

1.5.2 在业务预测方面的应用 12

1.5.3 在物理层的应用 13

1.5.4 在无线资源分配方面的应用 14

1.6 本文结构 14

第二章 基于卷积神经网络的信道估计方法 17

2.1 引言 17

2.2 系统模型 17

2.3 信道估计的卷积神经网络机器学习方法 19

2.3.1 DAMP算法 19

2.3.2 L-DAMP卷积神经网络 21

2.3.3 DnCNN降噪器结构 22

2.3.4 DnCNN的训练 23

2.4 仿真结果 23

2.4.1 参数设置 23

2.4.2 性能分析 24

2.4.3 本章小结 26

第三章 基于神经网络的Alamouti编码信号检测 27

3.1 引言 27

3.2 Alamouti编码和检测 27

3.2.1 Alamouti空时编码 27

3.2.2 信道模型 28

3.2.3 已知CSI的ML检测 29

3.3 最大似然神经网络检测方法 29

3.3.1 神经网络 30

3.3.2 ML-NN的结构 30

3.3.3 ML-NN的训练 31

3.4 仿真结果 31

3.4.1 参数设置 31

3.4.2 性能分析 32

3.5 本章小结 34

第四章 总结 35

参考文献 37

致 谢 41

剩余内容已隐藏，请支付后下载全文，论文总字数：34587字