基于卷积特征提取及深度降噪网络的大规模MIMO系统信号检测
Massive MIMO System Signal Detection Based on Convolutional Feature Extraction and Deep Denoising Network
-
摘要: 传统多输入多输出（Multiple-Input Multiple-Output， MIMO）信号检测算法受到天线数量和收发天线比例的限制，一般仅适用于少量天线、收发天线比例较低的情况。本文提出一种基于深度学习（Deep Learning，DL）的稀疏连接卷积降噪网络模型，用于大规模MIMO系统上行链路信号检测。首先，通过简化经典的检测网络（Detection Network， DetNet），改进ScNet（Sparsely Connected Neural Network）检测算法，引入卷积神经网络（Convolutional Neural Networks，CNN）对三通道输入数据提取特征以减少训练参数，提出一种SConv（Sparsely Connected Convolutional Neural Network）检测算法。与DetNet算法相比，该算法可同时降低计算复杂度和提高检测精度。在此基础上，进一步基于CNN构建信号降噪模块，并嵌入SConv网络，提出一种卷积神经降噪（Sparsely Connected Convolutional Denoising，SConv-D）网络辅助的大规模MIMO检测算法。此算法检测过程分为两级，第一级由SConv算法提供初始估计值，再将初始估计值作为降噪过程的输入，并由此构成算法第二级。实验结果表明，本文提出的SConv-D算法适用于QPSK、4QAM及16QAM等多种信号调制模式，在高阶调制模式下获得的性能增益尤为明显。此外，该算法能够适应各种比例的收发天线及数量规模的系统配置，尤其是在收发天线数量相等的情况下亦能获得更优的性能。本文算法还克服了MMNet在高阶调制情况下的性能平台效应，在16QAM调制、收发天线数量相等的情况下，SConv-D在10-2误比特率上能获得接近2 dB的性能增益。Abstract: The performance of traditional multiple-input multiple-output （MIMO） signal detection algorithms is influenced by factors such as the number of antennas and the ratio of transmitting to receiving antennas. Therefore， it applies to scenarios with a few antennas and a low ratio of antennas. This study proposes a DL-based sparsely connected convolutional denoising network model for uplink signal detection in a massive MIMO system. First， by simplifying the classical detection network （DetNet）， improving the detection algorithm sparsely connected neural network （ScNet）， and introducing convolutional neural networks （CNNs） to extract features from three-channel input data and reduce training parameters， this paper proposes a detection algorithm called sparsely connected convolutional neural network （SConv）. SConv enhances detection accuracy and reduces computational complexity compared to DetNet. Building on this， the paper proposes the SConv-D denoising network， a convolutional denoising neural network-assisted massive MIMO detection algorithm， by integrating a CNN-based signal denoising module within the SConv framework. The detection process of this algorithm is divided into two stages. The first stage is provided by the SConv algorithm for the initial solution. Subsequently， the initial solution is used as the input of the denoising process， which constitutes the second stage of the algorithm. Simulation results show that the SConv-D algorithm is suitable for QPSK， 4QAM， and 16QAM signal modulation models， and the performance gain obtained in the high-order modulation mode is particularly apparent. In addition， this algorithm can adapt to system configurations with different proportions of the receiving and transmitting antennas and different numbers of antennas. Particularly when the number of receiving and transmitting antennas is equal， SConv-D can also achieve better performance. The algorithm also overcomes the performance plateau effect of MMNet under high-order modulation scenarios. Compared with MMNet， a gain of approximately 2 dB is achieved in bit error rate when employing 16QAM modulation， while the number of receiving and transmitting antennas remain the same.
下载: