CN113162883B - 一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法 - Google Patents

一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法 Download PDF

Info

Publication number
CN113162883B
CN113162883B CN202110418198.5A CN202110418198A CN113162883B CN 113162883 B CN113162883 B CN 113162883B CN 202110418198 A CN202110418198 A CN 202110418198A CN 113162883 B CN113162883 B CN 113162883B
Authority
CN
China
Prior art keywords
densenet
array
trainable
output
following formula
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110418198.5A
Other languages
English (en)
Other versions
CN113162883A (zh
Inventor
赵春明
蔡欢
姜明
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Southeast University
Original Assignee
Southeast University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Southeast University filed Critical Southeast University
Priority to CN202110418198.5A priority Critical patent/CN113162883B/zh
Publication of CN113162883A publication Critical patent/CN113162883A/zh
Application granted granted Critical
Publication of CN113162883B publication Critical patent/CN113162883B/zh
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L27/00Modulated-carrier systems
    • H04L27/26Systems using multi-frequency codes
    • H04L27/2601Multicarrier modulation systems
    • H04L27/2647Arrangements specific to the receiver only
    • H04L27/2655Synchronisation arrangements
    • H04L27/2689Link with other circuits, i.e. special connections between synchronisation arrangements and other circuits for achieving synchronisation
    • H04L27/2691Link with other circuits, i.e. special connections between synchronisation arrangements and other circuits for achieving synchronisation involving interference determination or cancellation
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L27/00Modulated-carrier systems
    • H04L27/26Systems using multi-frequency codes
    • H04L27/2601Multicarrier modulation systems
    • H04L27/2626Arrangements specific to the transmitter only
    • H04L27/2627Modulators
    • H04L27/2628Inverse Fourier transform modulators, e.g. inverse fast Fourier transform [IFFT] or inverse discrete Fourier transform [IDFT] modulators
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L27/00Modulated-carrier systems
    • H04L27/26Systems using multi-frequency codes
    • H04L27/2601Multicarrier modulation systems
    • H04L27/2647Arrangements specific to the receiver only
    • H04L27/2649Demodulators
    • H04L27/265Fourier transform demodulators, e.g. fast Fourier transform [FFT] or discrete Fourier transform [DFT] demodulators

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • Discrete Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本发明公开了一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法,通过两级基于CNN框架的神经网络DenseNet实现,第一级神经网络DenseNet‑1对已消除符号间干扰的接收OFDM符号实现粗略检测;基于第一级神经网络的粗略检测,对已消除符号间干扰的接收OFDM符号并行地进行部分子载波间干扰消除,然后由第二级神经网络DenseNet‑2对干扰消除后的接收OFDM符号实现进一步的精确检测。本发明能够应用于高阶调制无CP OFDM系统,从存在ICI和ISI的接收OFDM符号中以较高的精确度检测出发送OFDM符号,最终实现OFDM系统频谱效率的提高。

Description

一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除 检测方法
技术领域
本发明涉及一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法,属于无线移动通信技术领域。
背景技术
正交频分复用(OFDM)技术频谱利用率高,实现简单,在资源管理方面具有较大的灵活性,被广泛的应用于4G LTE和5G NR。每个OFDM符号的开始均附有长于信道冲激响应长度的循环前缀(CP)用于保证抵抗前一OFDM符号的干扰,保持子载波间的正交性。然而,CP承载的是一段重复的信息,会造成频谱效率的降低。因此有必要研究一种无CP OFDM通信系统。
当CP不存在时,接收OFDM符号中存在严重的子载波间干扰(ICI)和符号间干扰(ISI),导致常用的单抽头检测技术会完全不能工作。针对ICI和ISI,有国内外学者提出了串行干扰消除技术(SIC),然而SIC技术只能处理CP短于但与信道冲激响应长度差距不大的场景,在高阶QAM调制的无CP OFDM系统中SIC技术由于存在严重的误码传播无法工作。针对无CP OFDM系统,误码率最小的检测方法为最大似然序列检测(MLSE)技术,然而MLSE技术的复杂度随着调制阶数和子载波数目的增加呈现指数性的增长,在实际的子载波数目较多的高阶调制无CP OFDM系统中难以应用。
发明内容
本发明的目的是提供一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法,能够应用于高阶调制无CP OFDM系统,从存在ICI和ISI的接收OFDM符号中以较高的精确度检测出发送OFDM符号,最终实现OFDM系统频谱效率的提高。
为实现上述目的,本发明采用的技术方案为:
一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法,其特征在于:包括以下步骤:
步骤1,对发送机的M-QAM符号进行串并变换并映射到频域OFDM符号Xk的可用子载波上,Xk经过快速反傅里叶变换后得到时域OFDM符号,将该时域OFDM符号经过并串变换、数模处理后调制到载频上,然后由发送天线送入时变多径信道;其中,Xk中包含的总子载波数为N,位于中间位置的为Nu个可用子载波,位于两侧位置的为虚拟子载波,虚拟子载波个数分别为Nv
步骤2,对接收机天线上的接收信号进行解调后得到基带信号,对该基带信号进行模数变换、同步、快速傅里叶变换后得到频域接收OFDM符号Yk,取接收OFDM符号Yk中的有用子载波上的接收信号
Figure RE-GDA0003100638070000021
用于发送OFDM符号的检测,其中
Figure RE-GDA0003100638070000022
上式中,
Figure RE-GDA0003100638070000023
表示造成有用子载波间干扰的频域信道矩阵,
Figure RE-GDA0003100638070000024
表示造成符号间干扰的频域信道矩阵,
Figure RE-GDA0003100638070000025
表示有用子载波上的频域加性高斯白噪声,
Figure RE-GDA0003100638070000026
表示第k个频域发送OFDM符号Xk在有用子载波上发送的M-QAM符号;
步骤3,对
Figure RE-GDA0003100638070000027
按下式进行符号间干扰消除
Figure RE-GDA0003100638070000028
其中,
Figure RE-GDA0003100638070000029
表示已消除符号间干扰的OFDM接收符号,
Figure RE-GDA00031006380700000210
表示对第k-1个OFDM接收符号有用子载波上发送的M-QAM符号的硬判决结果;
步骤4,根据
Figure RE-GDA00031006380700000211
Figure RE-GDA00031006380700000212
生成第一级神经网络DenseNet-1的输入数据
Figure RE-GDA00031006380700000213
Figure RE-GDA00031006380700000214
输入DenseNet-1得到实数输出
Figure RE-GDA00031006380700000215
然后将DenseNet-1的实数输出
Figure RE-GDA00031006380700000216
转化为对应的复数,得到DenseNet-1的检测输出
Figure RE-GDA00031006380700000217
Figure RE-GDA00031006380700000218
进行硬判决得到
Figure RE-GDA00031006380700000219
其中
Figure RE-GDA00031006380700000220
其中,1≤n≤Nu,[·]n,m表示矩阵的第n行第m列元素,b1表示输入DenseNet-1的信道参数宽度,
Figure RE-GDA00031006380700000221
的复数化按照下式进行
Figure RE-GDA00031006380700000222
[·]:,i表示矩阵的第i列,i=1,2,j表示虚数单位;
步骤5,基于DenseNet-1输出的判决结果
Figure RE-GDA0003100638070000031
Figure RE-GDA0003100638070000032
按下式进行部分干扰消除,
Figure RE-GDA0003100638070000033
其中,
Figure RE-GDA0003100638070000034
表示消除了部分子载波间干扰的接收OFDM符号,
Figure RE-GDA0003100638070000035
含义同
Figure RE-GDA0003100638070000036
b2与b1的取值相同或不同;
步骤6,基于
Figure RE-GDA0003100638070000037
Figure RE-GDA0003100638070000038
生成第二级神经网络DenseNet-2的输入数据
Figure RE-GDA0003100638070000039
Figure RE-GDA00031006380700000310
输入DenseNet-2得到实数输出
Figure RE-GDA00031006380700000311
然后将DenseNet-2的实数输出
Figure RE-GDA00031006380700000312
参考
Figure RE-GDA00031006380700000313
复数化方式转化为对应的复数,得到DenseNet-2的检测输出
Figure RE-GDA00031006380700000314
Figure RE-GDA00031006380700000315
进行硬判决得到最终的检测输出
Figure RE-GDA00031006380700000316
所述步骤4和步骤6中,第一级神经网络DenseNet-1和第二级神经网络DenseNet-2的输入数据通过下式获得:
Figure RE-GDA00031006380700000317
Figure RE-GDA00031006380700000318
其中,以
Figure RE-GDA00031006380700000319
统一表示
Figure RE-GDA00031006380700000320
Figure RE-GDA00031006380700000321
统一表示
Figure RE-GDA00031006380700000322
Figure RE-GDA00031006380700000323
b统一表示b1和b2,1≤n≤Nu
Figure RE-GDA00031006380700000324
表示取元素实部运算,
Figure RE-GDA00031006380700000325
表示取虚部运算。
所述步骤4和步骤6中,第一级神经网络DenseNet-1和第二级神经网络DenseNet-2的输出通过以下步骤计算得到:
因DenseNet-1和DenseNet-2架构相同,唯一区别在于输入数据的不同,下面以DenseNet统一表示DenseNet-1和DenseNet-2,
Figure RE-GDA00031006380700000326
统一表示
Figure RE-GDA00031006380700000327
Figure RE-GDA00031006380700000328
Figure RE-GDA00031006380700000329
统一表示
Figure RE-GDA0003100638070000041
Figure RE-GDA0003100638070000042
b统一表示b1和b2
Figure RE-GDA0003100638070000043
统一表示DenseNet-1和DenseNet-2中间层的输出
Figure RE-GDA0003100638070000044
Figure RE-GDA0003100638070000045
其中,
Figure RE-GDA0003100638070000046
统一表示DenseNet-1和DenseNet-2 中总的DenseNet块数
Figure RE-GDA0003100638070000047
Figure RE-GDA0003100638070000048
1≤ld≤Ld+1,Ld统一表示DenseNet-1和DenseNet-2 中第d个DenseNet块内总的微网络数
Figure RE-GDA0003100638070000049
Figure RE-GDA00031006380700000410
m1=1,2,3,对于可训练参数同样忽略 DenseNet-1的上标标识-1和DenseNet-2的上标标识-2统一表示;
计算步骤包括:
步骤I,对
Figure RE-GDA00031006380700000411
按照下式进行补零操作并重新赋值给
Figure RE-GDA00031006380700000412
Figure RE-GDA00031006380700000413
步骤II,
Figure RE-GDA00031006380700000414
按照下式与卷积核进行运算得到第一个DenseNet块第一个微网络的输入数组
Figure RE-GDA00031006380700000415
Figure RE-GDA00031006380700000416
其中,
Figure RE-GDA00031006380700000417
的数组大小为
Figure RE-GDA00031006380700000418
reLU(x)为对单个数组元素运算的激活函数,当x>0时,该函数输出x,当x≤0时,该函数输出0,W1
Figure RE-GDA00031006380700000419
的可训练参数数组,b1为大小为
Figure RE-GDA00031006380700000420
可训练参数数组,
Figure RE-GDA00031006380700000421
表示索引为i1-n+1,i2,i3的数组元素,
Figure RE-GDA00031006380700000422
表示索引为i3的数组元素;
步骤III,初始化d=1,ld=1;
步骤IV,
Figure RE-GDA00031006380700000423
按照下式与卷积核进行运算得到
Figure RE-GDA00031006380700000424
Figure RE-GDA00031006380700000425
其中,
Figure RE-GDA00031006380700000426
为Nu×2R的微网络第1层卷积层的输出数组,R表示DenseNet块每个微网络最终输出数组第2维的大小,设置为大于0的整数,
Figure RE-GDA00031006380700000427
表示
Figure RE-GDA00031006380700000428
的第n行,
Figure RE-GDA00031006380700000429
Figure RE-GDA0003100638070000051
的可训练参数数组,
Figure RE-GDA0003100638070000052
Figure RE-GDA0003100638070000053
第2维的大小,
Figure RE-GDA0003100638070000054
为大小为1×2R的可训练参数数组;
步骤V,对
Figure RE-GDA0003100638070000055
按照下式进行补零操作并重新赋值给
Figure RE-GDA0003100638070000056
Figure RE-GDA0003100638070000057
步骤VI,
Figure RE-GDA0003100638070000058
按照下式与卷积核进行运算得到
Figure RE-GDA0003100638070000059
Figure RE-GDA00031006380700000510
其中,1≤i3≤2R,
Figure RE-GDA00031006380700000511
为Nu×2R的微网络第2层卷积层的输出数组,
Figure RE-GDA00031006380700000512
为3×2R×2R 的可训练参数数组,
Figure RE-GDA00031006380700000513
为大小为1×2R的可训练参数数组;
步骤VII,对
Figure RE-GDA00031006380700000514
按照下式进行补零操作并重新赋值给
Figure RE-GDA00031006380700000515
Figure RE-GDA00031006380700000516
步骤VIII,
Figure RE-GDA00031006380700000517
按照下式与卷积核进行运算得到
Figure RE-GDA00031006380700000518
Figure RE-GDA00031006380700000519
其中,1≤i3≤R,
Figure RE-GDA00031006380700000520
为Nu×R的微网络输出层输出数组,
Figure RE-GDA00031006380700000521
为3×2R×R的可训练参数数组,
Figure RE-GDA00031006380700000522
为大小为1×R的可训练参数数组;
步骤Ⅸ,将
Figure RE-GDA00031006380700000523
Figure RE-GDA00031006380700000524
在第2维进行拼接并赋值给
Figure RE-GDA00031006380700000525
Figure RE-GDA00031006380700000526
其中,赋值后的
Figure RE-GDA00031006380700000527
为大小为
Figure RE-GDA00031006380700000528
的数组,
Figure RE-GDA00031006380700000529
步骤X,ld=ld+1,如果ld>Ld则进入步骤XI,否则进入步骤IV;
步骤XI,
Figure RE-GDA00031006380700000530
按照下式与卷积核进行运算得到
Figure RE-GDA00031006380700000531
Figure RE-GDA0003100638070000061
其中,
Figure RE-GDA0003100638070000062
Figure RE-GDA0003100638070000063
的第d个DenseNet块最后过渡层的输出数组,
Figure RE-GDA0003100638070000064
设置为大于 0小于
Figure RE-GDA0003100638070000065
的整数,
Figure RE-GDA0003100638070000066
Figure RE-GDA0003100638070000067
的可训练参数数组,
Figure RE-GDA0003100638070000068
为大小为
Figure RE-GDA0003100638070000069
的可训练参数数组;
步骤XII,d=d+1,如果
Figure RE-GDA00031006380700000610
则进入步骤XIII,否则,令ld=1;
步骤XIII,对
Figure RE-GDA00031006380700000611
按照下式进行补零操作并重新赋值给
Figure RE-GDA00031006380700000612
Figure RE-GDA00031006380700000613
步骤XIV,
Figure RE-GDA00031006380700000614
按照下式与卷积核进行运算得到
Figure RE-GDA00031006380700000615
Figure RE-GDA00031006380700000616
其中,
Figure RE-GDA00031006380700000617
为Nu×Co的卷积层输出数组,1≤i3≤Co,Co设置为大于0的整数,Wo
Figure RE-GDA00031006380700000618
的可训练参数数组,bo为大小为1×Co的可训练参数数组;
步骤XV,对
Figure RE-GDA00031006380700000619
按照下式进行补零操作并重新赋值给
Figure RE-GDA00031006380700000620
Figure RE-GDA00031006380700000621
步骤XVI,
Figure RE-GDA00031006380700000622
按照下式与卷积核进行运算得到
Figure RE-GDA00031006380700000623
Figure RE-GDA00031006380700000624
其中,
Figure RE-GDA00031006380700000625
为Nu×2的DenseNet输出,1≤i3≤2,Wout为3×Co×2的可训练参数数组,bout为小为1×2的可训练参数数组,a为OFDM系统发送M-QAM调制符号的实部或者虚部的幅度最大值,M表示调制阶数,tanh(x)为激活函数,
Figure RE-GDA00031006380700000626
所述第一级神经网络DenseNet-1和第二级神经网络DenseNet-2的训练阶段包括如下步骤:
步骤一,获得训练样本
Figure RE-GDA0003100638070000071
组成训练样本集Ψ1,其中,训练样本集Ψ1的生成信噪比在工作信噪比区间SNRlow至SNRhigh内均匀分布,训练样本集Ψ1的大小为BS1, Zk表示训练标签,为Nu×2的数组,由下式生成
Figure RE-GDA0003100638070000072
步骤二,应用xavier初始化方法初始化DenseNet-1所有的可训练参数,然后根据训练数据完成对DenseNet-1的训练;
步骤三,基于已训练DenseNet-1检测输出的硬判决结果对
Figure RE-GDA0003100638070000073
做干扰消除得到
Figure RE-GDA0003100638070000074
进而得到DenseNet-2的训练样本
Figure RE-GDA0003100638070000075
组成训练样本集Ψ2,其中,训练样本集Ψ2的生成信噪比在工作信噪比区间SNRlow至SNRhigh内均匀分布,训练样本集Ψ2的大小为 BS2
步骤四,应用xavier初始化方法初始化DenseNet-2所有的可训练参数,然后根据训练数据完成对DenseNet-2的训练。
有益效果:本发明提供的一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法,与现有技术相比,具有以下优点:
1、相比于CP充分OFDM系统,本发明设计方案可以在提高OFDM系统的频谱效率的同时达到相似的检测性能;
2、相比于传统的干扰消除算法,本发明设计方案检性能大大提高;
3、相比于最大似然检测,本发明设计方案的计算复杂度随着子载波数目呈线性增长,复杂度较低,可应用于实际系统;
4、相比于无干扰消除的基于DenseNet的检测方法,相同可训练参数量下,本发明设计方案具有更优的检测性能,尤其是在子载波数目较多的时候;
5、针对信道延迟功率谱以及OFDM系统的子载波数目,本发明设计方案具有良好的鲁棒性,不需要重复训练神经网络;
6、本发明同样适用于CP存在的OFDM系统,但其中CP长度小于信道冲激响应长度。
附图说明
图1是本发明设计DenseNet-PIC算法流程框图;
图2是本发明设计应用实施例一的BER仿真曲线图;
图3是本发明设计应用实施例二的BER仿真曲线图;
图4是本发明设计应用实施例三的BER仿真曲线图;
图5是本发明设计应用实施例四的BER仿真曲线图。
具体实施方式
下面结合附图对本发明作更进一步的说明。
本发明的一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法,用于对无CP OFDM系统接收OFDM符号的均衡检测,本发明通过两级基于CNN框架的神经网络DenseNet实现,第一级神经网络DenseNet-1对已消除符号间干扰的接收OFDM符号实现粗略检测;基于第一级神经网络的粗略检测,对已消除符号间干扰的接收OFDM 符号并行地进行部分子载波间干扰消除,然后由第二级神经网络DenseNet-2对干扰消除后的接收OFDM符号实现进一步的精确检测。具体包括以下步骤:
步骤1,对发送机的M-QAM符号进行串并变换并映射到频域OFDM符号Xk的可用子载波上,Xk经过快速反傅里叶变换后得到时域OFDM符号,将该时域OFDM符号经过并串变换、数模处理后调制到载频上,然后由发送天线送入时变多径信道;其中,Xk中包含的总子载波数为N,位于中间位置的为Nu个可用子载波,位于两侧位置的为虚拟子载波,虚拟子载波个数分别为Nv
步骤2,对接收机天线上的接收信号进行解调后得到基带信号,对该基带信号进行模数变换、同步、快速傅里叶变换后得到频域接收OFDM符号Yk,取接收OFDM符号Yk中的有用子载波上的接收信号
Figure RE-GDA0003100638070000081
用于发送OFDM符号的检测,其中
Figure RE-GDA0003100638070000082
上式中,
Figure RE-GDA0003100638070000083
表示造成有用子载波间干扰的频域信道矩阵,
Figure RE-GDA0003100638070000084
表示造成符号间干扰的频域信道矩阵,
Figure RE-GDA0003100638070000085
表示有用子载波上的频域加性高斯白噪声,
Figure RE-GDA0003100638070000086
表示第k个频域发送OFDM符号Xk在有用子载波上发送的M-QAM符号;
步骤3,对
Figure RE-GDA0003100638070000091
按下式进行符号间干扰消除
Figure RE-GDA0003100638070000092
其中,
Figure RE-GDA0003100638070000093
表示已消除符号间干扰的OFDM接收符号,
Figure RE-GDA0003100638070000094
表示对第k-1个OFDM接收符号有用子载波上发送的M-QAM符号的硬判决结果;
步骤4,根据
Figure RE-GDA0003100638070000095
Figure RE-GDA0003100638070000096
生成第一级神经网络DenseNet-1的输入数据
Figure RE-GDA0003100638070000097
Figure RE-GDA0003100638070000098
输入DenseNet-1得到实数输出
Figure RE-GDA0003100638070000099
然后将DenseNet-1的实数输出
Figure RE-GDA00031006380700000910
转化为对应的复数,得到DenseNet-1的检测输出
Figure RE-GDA00031006380700000911
Figure RE-GDA00031006380700000912
进行硬判决得到
Figure RE-GDA00031006380700000913
其中
Figure RE-GDA00031006380700000914
其中,1≤n≤Nu,[·]n,m表示矩阵的第n行第m列元素,b1表示输入DenseNet-1的信道参数宽度,
Figure RE-GDA00031006380700000915
的复数化按照下式进行
Figure RE-GDA00031006380700000916
[·]:,i表示矩阵的第i列,i=1,2,j表示虚数单位;
步骤5,基于DenseNet-1输出的判决结果
Figure RE-GDA00031006380700000917
Figure RE-GDA00031006380700000918
按下式进行部分干扰消除,
Figure RE-GDA00031006380700000919
其中,
Figure RE-GDA00031006380700000920
表示消除了部分子载波间干扰的接收OFDM符号,b2表示输入DenseNet-2的信道参数宽度,
Figure RE-GDA00031006380700000921
含义同
Figure RE-GDA00031006380700000922
唯一区别在于b2与b1的取值可以相同也可以不同;
步骤6,基于
Figure RE-GDA00031006380700000923
Figure RE-GDA00031006380700000924
生成第二级神经网络DenseNet-2的输入数据
Figure RE-GDA00031006380700000925
Figure RE-GDA00031006380700000926
输入DenseNet-2得到实数输出
Figure RE-GDA00031006380700000927
然后将DenseNet-2的实数输出
Figure RE-GDA00031006380700000928
参考
Figure RE-GDA00031006380700000929
复数化方式转化为对应的复数,得到DenseNet-2的检测输出
Figure RE-GDA00031006380700000930
Figure RE-GDA00031006380700000931
进行硬判决得到最终的检测输出
Figure RE-GDA00031006380700000932
步骤4和步骤6中,第一级神经网络DenseNet-1和第二级神经网络DenseNet-2的输入数据通过下式获得:
Figure RE-GDA0003100638070000101
Figure RE-GDA0003100638070000102
其中,以
Figure RE-GDA0003100638070000103
统一表示
Figure RE-GDA0003100638070000104
Figure RE-GDA0003100638070000105
统一表示
Figure RE-GDA0003100638070000106
Figure RE-GDA0003100638070000107
b统一表示b1和b2,1≤n≤Nu
Figure RE-GDA0003100638070000108
表示取元素实部运算,
Figure RE-GDA0003100638070000109
表示取虚部运算。
步骤4和步骤6中,第一级神经网络DenseNet-1和第二级神经网络DenseNet-2的输出通过以下步骤计算得到:
因DenseNet-1和DenseNet-2架构相同,唯一区别在于输入数据的不同,下面以DenseNet统一表示DenseNet-1和DenseNet-2,
Figure RE-GDA00031006380700001010
统一表示
Figure RE-GDA00031006380700001011
Figure RE-GDA00031006380700001012
Figure RE-GDA00031006380700001013
统一表示
Figure RE-GDA00031006380700001014
Figure RE-GDA00031006380700001015
b统一表示b1和b2
Figure RE-GDA00031006380700001016
统一表示DenseNet-1和DenseNet-2中间层的输出
Figure RE-GDA00031006380700001017
Figure RE-GDA00031006380700001018
其中,
Figure RE-GDA00031006380700001019
统一表示DenseNet-1和DenseNet-2 中总的DenseNet块数
Figure RE-GDA00031006380700001020
Figure RE-GDA00031006380700001021
1≤ld≤Ld+1,Ld统一表示DenseNet-1和DenseNet-2 中第d个DenseNet块内总的微网络数
Figure RE-GDA00031006380700001022
Figure RE-GDA00031006380700001023
m1=1,2,3,对于可训练参数同样忽略 DenseNet-1的上标标识-1和DenseNet-2的上标标识-2统一表示;
计算步骤包括:
步骤I,对
Figure RE-GDA00031006380700001024
按照下式进行补零操作并重新赋值给
Figure RE-GDA00031006380700001025
Figure RE-GDA0003100638070000111
步骤II,
Figure RE-GDA0003100638070000112
按照下式与卷积核进行运算得到第一个DenseNet块第一个微网络的输入数组
Figure RE-GDA0003100638070000113
Figure RE-GDA0003100638070000114
其中,
Figure RE-GDA0003100638070000115
的数组大小为
Figure RE-GDA0003100638070000116
reLU(x)为对单个数组元素运算的激活函数,当x>0时,该函数输出x,当x≤0时,该函数输出0,W1
Figure RE-GDA0003100638070000117
的可训练参数数组,b1为大小为
Figure RE-GDA0003100638070000118
可训练参数数组,
Figure RE-GDA0003100638070000119
表示索引为i1-n+1,i2,i3的数组元素,
Figure RE-GDA00031006380700001110
表示索引为i3的数组元素;
步骤III,初始化d=1,ld=1;
步骤IV,
Figure RE-GDA00031006380700001111
按照下式与卷积核进行运算得到
Figure RE-GDA00031006380700001112
Figure RE-GDA00031006380700001113
其中,
Figure RE-GDA00031006380700001114
为Nu×2R的微网络第1层卷积层的输出数组,R表示DenseNet块每个微网络最终输出数组第2维的大小,设置为大于0的整数,
Figure RE-GDA00031006380700001115
表示
Figure RE-GDA00031006380700001116
的第n行,
Figure RE-GDA00031006380700001117
Figure RE-GDA00031006380700001118
的可训练参数数组,
Figure RE-GDA00031006380700001119
Figure RE-GDA00031006380700001120
第2维的大小,
Figure RE-GDA00031006380700001121
为大小为1×2R的可训练参数数组;
步骤V,对
Figure RE-GDA00031006380700001122
按照下式进行补零操作并重新赋值给
Figure RE-GDA00031006380700001123
Figure RE-GDA00031006380700001124
步骤VI,
Figure RE-GDA00031006380700001125
按照下式与卷积核进行运算得到
Figure RE-GDA00031006380700001126
Figure RE-GDA0003100638070000121
其中,1≤i3≤2R,
Figure RE-GDA0003100638070000122
为Nu×2R的微网络第2层卷积层的输出数组,
Figure RE-GDA0003100638070000123
为3×2R×2R 的可训练参数数组,
Figure RE-GDA0003100638070000124
为大小为1×2R的可训练参数数组;
步骤VII,对
Figure RE-GDA0003100638070000125
按照下式进行补零操作并重新赋值给
Figure RE-GDA0003100638070000126
Figure RE-GDA0003100638070000127
步骤VIII,
Figure RE-GDA0003100638070000128
按照下式与卷积核进行运算得到
Figure RE-GDA0003100638070000129
Figure RE-GDA00031006380700001210
其中,1≤i3≤R,
Figure RE-GDA00031006380700001211
为Nu×R的微网络输出层输出数组,
Figure RE-GDA00031006380700001212
为3×2R×R的可训练参数数组,
Figure RE-GDA00031006380700001213
为大小为1×R的可训练参数数组;
步骤Ⅸ,将
Figure RE-GDA00031006380700001214
Figure RE-GDA00031006380700001215
在第2维进行拼接并赋值给
Figure RE-GDA00031006380700001216
Figure RE-GDA00031006380700001217
其中,赋值后的
Figure RE-GDA00031006380700001218
为大小为
Figure RE-GDA00031006380700001219
的数组,
Figure RE-GDA00031006380700001220
步骤X,ld=ld+1,如果ld>Ld则进入步骤XI,否则进入步骤IV;
步骤XI,
Figure RE-GDA00031006380700001221
按照下式与卷积核进行运算得到
Figure RE-GDA00031006380700001222
Figure RE-GDA00031006380700001223
其中,
Figure RE-GDA00031006380700001224
Figure RE-GDA00031006380700001225
的第d个DenseNet块最后过渡层的输出数组,
Figure RE-GDA00031006380700001226
设置为大于 0小于
Figure RE-GDA00031006380700001227
的整数,
Figure RE-GDA00031006380700001228
Figure RE-GDA00031006380700001229
的可训练参数数组,
Figure RE-GDA00031006380700001230
为大小为
Figure RE-GDA00031006380700001231
的可训练参数数组;
步骤XII,d=d+1,如果
Figure RE-GDA00031006380700001232
则进入步骤XIII,否则,令ld=1;
步骤XIII,对
Figure RE-GDA00031006380700001233
按照下式进行补零操作并重新赋值给
Figure RE-GDA00031006380700001234
Figure RE-GDA0003100638070000131
步骤XIV,
Figure RE-GDA0003100638070000132
按照下式与卷积核进行运算得到
Figure RE-GDA0003100638070000133
Figure RE-GDA0003100638070000134
其中,
Figure RE-GDA0003100638070000135
为Nu×Co的卷积层输出数组,1≤i3≤Co,Co设置为大于0的整数,Wo
Figure RE-GDA0003100638070000136
的可训练参数数组,bo为大小为1×Co的可训练参数数组;
步骤XV,对
Figure RE-GDA0003100638070000137
按照下式进行补零操作并重新赋值给
Figure RE-GDA0003100638070000138
Figure RE-GDA0003100638070000139
步骤XVI,
Figure RE-GDA00031006380700001310
按照下式与卷积核进行运算得到
Figure RE-GDA00031006380700001311
Figure RE-GDA00031006380700001312
其中,
Figure RE-GDA00031006380700001313
为Nu×2的DenseNet输出,1≤i3≤2,Wout为3×Co×2的可训练参数数组,bout为小为1×2的可训练参数数组,a为OFDM系统发送M-QAM调制符号的实部或者虚部的幅度最大值,M表示调制阶数,tanh(x)为激活函数,
Figure RE-GDA00031006380700001314
第一级神经网络DenseNet-1和第二级神经网络DenseNet-2的训练阶段包括如下步骤:
步骤一,获得训练样本
Figure RE-GDA00031006380700001315
组成训练样本集Ψ1,其中,训练样本集Ψ1的生成信噪比在工作信噪比区间SNRlow至SNRhigh内均匀分布,训练样本集Ψ1的大小为BS1, Zk表示训练标签,为Nu×2的数组,由下式生成
Figure RE-GDA00031006380700001316
步骤二,应用xavier初始化方法初始化DenseNet-1所有的可训练参数,然后根据训练数据完成对DenseNet-1的训练;
步骤三,基于已训练DenseNet-1检测输出的硬判决结果对
Figure RE-GDA0003100638070000141
做干扰消除得到
Figure RE-GDA0003100638070000142
进而得到DenseNet-2的训练样本
Figure RE-GDA0003100638070000143
组成训练样本集Ψ2,其中,训练样本集Ψ2的生成信噪比在工作信噪比区间SNRlow至SNRhigh内均匀分布,训练样本集Ψ2的大小为 BS2
步骤四,应用xavier初始化方法初始化DenseNet-2所有的可训练参数,然后根据训练数据完成对DenseNet-2的训练。
下面结合实施例对本发明做进一步说明。
实施例1
实际应用当中,神经网络中采用均方误差形式的代价函数。
表1
Figure RE-GDA0003100638070000144
将上述所设计适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法,具体应用于实际,当中,诸如实施例1,实际应用中CP不充分OFDM系统参数以及DenseNet-1和DenseNet-2训练参数如表1所示。
对发送机的64阶QAM符号进行串并变换并映射到频域OFDM符号Xk的有用子载波上,Xk经过快速反傅里叶变换后得到时域OFDM符号,将该时域OFDM符号经过并串变换、数模处理后调制到载频上,然后由发送天线送入时变多径信道。其中,Xk中包含的总子载波数为N,位于中间位置的为Nu个可用子载波,位于两侧位置的为虚拟子载波,虚拟子载波个数分别为Nv。其中时变多径信道的功率延时谱如表2所示。
表2
延时 0 1 2 3 9 14 17
功率(dB) 0 0.7904 3.5312 3.1232 0.4559 3.6998 0.4744
接着对接收机天线上的接收信号进行解调后得到基带信号,对该基带信号进行模数变换、同步、快速傅里叶变换后得到频域接收OFDM符号Yk
Yk=HkXkkXk-1k
其中,
Figure RE-GDA0003100638070000151
Figure RE-GDA0003100638070000152
Figure RE-GDA0003100638070000153
Figure RE-GDA0003100638070000154
G=diag{FV}
Figure RE-GDA0003100638070000155
Figure RE-GDA0003100638070000156
Figure RE-GDA0003100638070000161
其中,V=[1,0,0,...,0]T,V的大小为N×1,N表示OFDM系统子载波数,[·]T表示转置运算;diag{}表示生成对角方阵,其对角线上的值为括号内矢量的值,
Figure RE-GDA0003100638070000162
FH表示F的共轭转置;
Figure RE-GDA0003100638070000163
l=0,1,2,...,L-1,
Figure RE-GDA0003100638070000164
表示对应于第k个 OFDM符号、可分辨径l的时域信道参数,在时域多径信道参数后补零得到N×1维矢量hk, L表示多径数目,时域多径信道利用Jakes模型进行仿真。
在检测中取Yk、Xk、Hk、Φk和Θk在Nu个有用子载波上的部分,分别记为
Figure RE-GDA0003100638070000165
Figure RE-GDA0003100638070000166
Figure RE-GDA0003100638070000167
Figure RE-GDA0003100638070000168
Figure RE-GDA0003100638070000169
按下式进行符号间干扰消除
Figure RE-GDA00031006380700001610
其中,
Figure RE-GDA00031006380700001611
表示已消除符号间干扰的OFDM接收符号,
Figure RE-GDA00031006380700001612
表示对第k-1个OFDM接收符号有用子载波上发送的M-QAM符号的硬判决结果。
根据
Figure RE-GDA00031006380700001613
Figure RE-GDA00031006380700001614
生成DenseNet-1的输入数据
Figure RE-GDA00031006380700001615
Figure RE-GDA00031006380700001616
输入DenseNet-1得到实数输出
Figure RE-GDA00031006380700001617
然后将DenseNet-1的实数输出
Figure RE-GDA00031006380700001618
转化为对应的复数,得到DenseNet-1的检测输出
Figure RE-GDA00031006380700001619
Figure RE-GDA00031006380700001620
进行硬判决得到
Figure RE-GDA00031006380700001621
其中
Figure RE-GDA00031006380700001622
1≤n≤Nu,[·]n,m表示矩阵的第n行第m列元素,b1表示输入DenseNet-1的信道参数宽度,
Figure RE-GDA00031006380700001623
的复数化按照下式进行
Figure RE-GDA0003100638070000171
[·]:,i表示矩阵的第i列,i=1,2,j表示虚数单位,
基于DenseNet-1输出的判决结果
Figure RE-GDA0003100638070000172
Figure RE-GDA0003100638070000173
按下式进行部分干扰消除,
Figure RE-GDA0003100638070000174
其中,
Figure RE-GDA0003100638070000175
表示消除了部分子载波间干扰的接收OFDM符号,
Figure RE-GDA0003100638070000176
含义同
Figure RE-GDA0003100638070000177
唯一区别在于b2的取值可以不同于b1
基于
Figure RE-GDA0003100638070000178
Figure RE-GDA0003100638070000179
生成DenseNet-2的输入数据
Figure RE-GDA00031006380700001710
Figure RE-GDA00031006380700001711
输入DenseNet-2得到实数输出
Figure RE-GDA00031006380700001712
然后将DenseNet-2的实数输出
Figure RE-GDA00031006380700001713
参考
Figure RE-GDA00031006380700001714
复数化方式转化为对应的复数,得到DenseNet-2的检测输出
Figure RE-GDA00031006380700001715
Figure RE-GDA00031006380700001716
进行硬判决得到最终的检测输出
Figure RE-GDA00031006380700001717
图2为实施例1中本发明设计方案DenseNet-PIC与并行干扰消除算法PIC、MMSE 以及CNN-PIC的BER对比,其中,CNN-PIC由将DenseNet-PIC中的DenseNet-1和 DenseNet-2替换为等可训练参数量的CNN-1和CNN-2得到。由图2可知,本发明设计方案DenseNet-PIC具有最优的检测性能,并且明显优于复杂度相同的CNN-PIC。
实施例2
实施例2由将实施例1的OFDM系统参数换为表3,信道功率延迟分布换位表4得到
表3
Figure RE-GDA00031006380700001718
表4
延时 0 4 16 21 74 116 140
功率(dB) 0 0.7904 3.5312 3.1232 0.4559 3.6998 0.4744
图3为实施例2中本发明设计方案DenseNet-PIC与无干扰消除的DenseNet大网络的BER对比,其中DenseNet大网络的可训练参数量与DetseNet-1和DenseNet-2总的参数量相同,DenseNet-PIC与DenseNet大网络中神经网络均在实施例1的参数设置下训练,在实施例2的参数设置下测试。由图3可知,DenseNet-PIC的检测性能明显优于 DenseNet大网络,验证了干扰消除设计方案的有效性。此外根据图3的仿真性能可知, DenseNet-PIC在128子载波数目的OFDM系统中训练,在1024子载波数目的OFDM系统中也能取得良好检测效果,证明了DenseNet-PIC对于子载波数目具有鲁棒性,针对不同子载波数目的OFDM系统,网络DenseNet-1和DenseNet-2只需要在某个固定的子载波数目下训练一次。
实施例3
实施例3为将实施例1中的信道延迟功率分布分别换为表5和表6。
表5
延时 0 1 2 4
功率(dB) 3.4234 6.3423 0.4559 5.3901
表6
延时 0 1 2 5 8 9
功率(dB) 0 5.3460 3.1232 0.4559 3.6998 0.4744
图4为实施例3的仿真结果,其中MMSE(L=5)、DenseNet-PIC(L=5)为信道功率延迟分布采用表5,其余测试参数同实施例1的仿真结果;MMSE(L=10)、DenseNet-PIC(L=10)为信道功率延迟分布采用表6,其余测试参数同实施例1的仿真结果;MMSE(L=18)、DenseNet-PIC(L=18)为实施例1的仿真结果;其中DenseNet-PIC中DenseNet-1和DenseNet-2均根据实施例1参数训练。由图4可知,在表2的信道延迟功率分布下训练得到的DenseNet-PIC在其他的不同于表2的延迟功率分布下也能表现良好,没有出现剧烈的性能恶化,始终优于MMSE,证明了DenseNet-PIC针对不同的信道延迟功率分布具有良好的鲁棒性,可以在实际系统中使用。
实施例4
实施例4为DenseNet-PIC算法在CP存在但长度小于信道冲激响应长度的OFDM系统中的测试。在实施实施例1中的发送机中增加附加CP的模块,接收机中增加减去CP 的模块。
图5为实施例4的仿真结果。DenseNet-PIC算法中DenseNet的训练测试CP长度为9,其余训练测试参数参考实施例1。由图可知,DenseNet-PIC算法的检测性能超过了 PIC和MMSE,逼近CP充分OFDM系统的BER性能。图5的仿真结果证明了本发明提出的 DenseNet-PIC算法同样适用于CP存在但小于信道冲激相应的OFDM系统。
以上所述仅是本发明的优选实施方式,应当指出:对于本技术领域的普通技术人员来说,在不脱离本发明原理的前提下,还可以做出若干改进和润饰,这些改进和润饰也应视为本发明的保护范围。

Claims (4)

1.一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法,其特征在于:包括以下步骤:
步骤1,对发送机的M-QAM符号进行串并变换并映射到频域OFDM符号Xk的可用子载波上,Xk经过快速反傅里叶变换后得到时域OFDM符号,将该时域OFDM符号经过并串变换、数模处理后调制到载频上,然后由发送天线送入时变多径信道;其中,Xk中包含的总子载波数为N,位于中间位置的为Nu个可用子载波,位于两侧位置的为虚拟子载波,虚拟子载波个数分别为Nv
步骤2,对接收机天线上的接收信号进行解调后得到基带信号,对该基带信号进行模数变换、同步、快速傅里叶变换后得到频域接收OFDM符号Yk,取接收OFDM符号Yk中的有用子载波上的接收信号
Figure FDA0003026819370000011
用于发送OFDM符号的检测,其中
Figure FDA0003026819370000012
上式中,
Figure FDA0003026819370000013
表示造成有用子载波间干扰的频域信道矩阵,
Figure FDA0003026819370000014
表示造成符号间干扰的频域信道矩阵,
Figure FDA0003026819370000015
表示有用子载波上的频域加性高斯白噪声,
Figure FDA0003026819370000016
表示第k个频域发送OFDM符号Xk在有用子载波上发送的M-QAM符号;
步骤3,对
Figure FDA0003026819370000017
按下式进行符号间干扰消除
Figure FDA0003026819370000018
其中,
Figure FDA0003026819370000019
表示已消除符号间干扰的OFDM接收符号,
Figure FDA00030268193700000110
表示对第k-1个OFDM接收符号有用子载波上发送的M-QAM符号的硬判决结果;
步骤4,根据
Figure FDA00030268193700000111
Figure FDA00030268193700000112
生成第一级神经网络DenseNet-1的输入数据
Figure FDA00030268193700000113
Figure FDA00030268193700000114
输入DenseNet-1得到实数输出
Figure FDA00030268193700000115
然后将DenseNet-1的实数输出
Figure FDA00030268193700000116
转化为对应的复数,得到DenseNet-1的检测输出
Figure FDA00030268193700000117
Figure FDA00030268193700000118
进行硬判决得到
Figure FDA00030268193700000119
其中
Figure FDA00030268193700000120
其中,1≤n≤Nu,[*]n,m表示矩阵的第n行第m列元素,b1表示输入DenseNet-1的信道参数宽度,
Figure FDA0003026819370000021
的复数化按照下式进行
Figure FDA0003026819370000022
[·]:,i表示矩阵的第i列,i=1,2,j表示虚数单位;
步骤5,基于DenseNet-1输出的判决结果
Figure FDA0003026819370000023
Figure FDA0003026819370000024
按下式进行部分干扰消除,
Figure FDA0003026819370000025
其中,
Figure FDA0003026819370000026
表示消除了部分子载波间干扰的接收OFDM符号,
Figure FDA0003026819370000027
含义同
Figure FDA0003026819370000028
b2与b1的取值相同或不同;
步骤6,基于
Figure FDA0003026819370000029
Figure FDA00030268193700000210
生成第二级神经网络DenseNet-2的输入数据
Figure FDA00030268193700000211
Figure FDA00030268193700000212
输入DenseNet-2得到实数输出
Figure FDA00030268193700000213
然后将DenseNet-2的实数输出
Figure FDA00030268193700000214
参考
Figure FDA00030268193700000215
复数化方式转化为对应的复数,得到DenseNet-2的检测输出
Figure FDA00030268193700000216
Figure FDA00030268193700000217
进行硬判决得到最终的检测输出
Figure FDA00030268193700000218
2.根据权利要求1所述的适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法,其特征在于:所述步骤4和步骤6中,第一级神经网络DenseNet-1和第二级神经网络DenseNet-2的输入数据通过下式获得:
Figure FDA00030268193700000219
Figure FDA00030268193700000220
其中,以
Figure FDA00030268193700000221
统一表示
Figure FDA00030268193700000222
Figure FDA00030268193700000223
Figure FDA00030268193700000224
统一表示
Figure FDA00030268193700000225
Figure FDA00030268193700000226
b统一表示b1和b2,1≤n≤Nu
Figure FDA0003026819370000031
表示取元素实部运算,
Figure FDA0003026819370000032
表示取虚部运算。
3.根据权利要求1所述的适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法,其特征在于:所述步骤4和步骤6中,第一级神经网络DenseNet-1和第二级神经网络DenseNet-2的输出通过以下步骤计算得到:
因DenseNet-1和DenseNet-2架构相同,唯一区别在于输入数据的不同,下面以DenseNet统一表示DenseNet-1和DenseNet-2,
Figure FDA0003026819370000033
统一表示
Figure FDA0003026819370000034
Figure FDA0003026819370000035
Figure FDA0003026819370000036
统一表示
Figure FDA0003026819370000037
Figure FDA0003026819370000038
b统一表示b1和b2
Figure FDA0003026819370000039
统一表示DenseNet-1和DenseNet-2中间层的输出
Figure FDA00030268193700000310
Figure FDA00030268193700000311
其中,
Figure FDA00030268193700000312
Figure FDA00030268193700000313
统一表示DenseNet-1和DenseNet-2中总的DenseNet块数
Figure FDA00030268193700000314
Figure FDA00030268193700000315
1≤ld≤Ld+1,Ld统一表示DenseNet-1和DenseNet-2中第d个DenseNet块内总的微网络数
Figure FDA00030268193700000316
Figure FDA00030268193700000317
m1=1,2,3,对于可训练参数同样忽略DenseNet-1的上标标识-1和DenseNet-2的上标标识-2统一表示;
计算步骤包括:
步骤I,对
Figure FDA00030268193700000318
按照下式进行补零操作并重新赋值给
Figure FDA00030268193700000319
Figure FDA00030268193700000320
步骤II,
Figure FDA00030268193700000321
按照下式与卷积核进行运算得到第一个DenseNet块第一个微网络的输入数组
Figure FDA00030268193700000322
Figure FDA00030268193700000323
其中,
Figure FDA00030268193700000324
Figure FDA00030268193700000325
的数组大小为
Figure FDA00030268193700000326
reLU(x)为对单个数组元素运算的激活函数,当x>0时,该函数输出x,当x≤0时,该函数输出0,W1
Figure FDA00030268193700000327
的可训练参数数组,b1为大小为
Figure FDA00030268193700000328
可训练参数数组,
Figure FDA00030268193700000329
表示索引为i1-n+1,i2,i3的数组元素,
Figure FDA00030268193700000330
表示索引为i3的数组元素;
步骤III,初始化d=1,ld=1;
步骤IV,
Figure FDA0003026819370000041
按照下式与卷积核进行运算得到
Figure FDA0003026819370000042
Figure FDA0003026819370000043
其中,
Figure FDA0003026819370000044
为Nu×2R的微网络第1层卷积层的输出数组,R表示DenseNet块每个微网络最终输出数组第2维的大小,设置为大于0的整数,
Figure FDA0003026819370000045
表示
Figure FDA0003026819370000046
的第n行,
Figure FDA0003026819370000047
Figure FDA0003026819370000048
的可训练参数数组,
Figure FDA0003026819370000049
Figure FDA00030268193700000410
第2维的大小,
Figure FDA00030268193700000411
为大小为1×2R的可训练参数数组;
步骤V,对
Figure FDA00030268193700000412
按照下式进行补零操作并重新赋值给
Figure FDA00030268193700000413
Figure FDA00030268193700000414
步骤VI,
Figure FDA00030268193700000415
按照下式与卷积核进行运算得到
Figure FDA00030268193700000416
Figure FDA00030268193700000417
其中,1≤i3≤2R,
Figure FDA00030268193700000418
为Nu×2R的微网络第2层卷积层的输出数组,
Figure FDA00030268193700000419
为3×2R×2R的可训练参数数组,
Figure FDA00030268193700000420
为大小为1×2R的可训练参数数组;
步骤VII,对
Figure FDA00030268193700000421
按照下式进行补零操作并重新赋值给
Figure FDA00030268193700000422
Figure FDA00030268193700000423
步骤VIII,
Figure FDA00030268193700000424
按照下式与卷积核进行运算得到
Figure FDA00030268193700000425
Figure FDA00030268193700000426
其中,1≤i3≤R,
Figure FDA00030268193700000427
为Nu×R的微网络输出层输出数组,
Figure FDA00030268193700000428
为3×2R×R的可训练参数数组,
Figure FDA00030268193700000429
为大小为1×R的可训练参数数组;
步骤Ⅸ,将
Figure FDA0003026819370000051
Figure FDA0003026819370000052
在第2维进行拼接并赋值给
Figure FDA0003026819370000053
Figure FDA0003026819370000054
其中,赋值后的
Figure FDA0003026819370000055
为大小为
Figure FDA0003026819370000056
的数组,
Figure FDA0003026819370000057
步骤X,ld=ld+1,如果ld>Ld则进入步骤XI,否则进入步骤IV;
步骤XI,
Figure FDA0003026819370000058
按照下式与卷积核进行运算得到
Figure FDA0003026819370000059
Figure FDA00030268193700000510
其中,
Figure FDA00030268193700000511
Figure FDA00030268193700000512
的第d个DenseNet块最后过渡层的输出数组,
Figure FDA00030268193700000513
设置为大于0小于
Figure FDA00030268193700000514
的整数,
Figure FDA00030268193700000515
Figure FDA00030268193700000516
的可训练参数数组,
Figure FDA00030268193700000517
为大小为
Figure FDA00030268193700000518
的可训练参数数组;
步骤XII,d=d+1,如果
Figure FDA00030268193700000519
则进入步骤XIII,否则,令ld=1;
步骤XIII,对
Figure FDA00030268193700000520
按照下式进行补零操作并重新赋值给
Figure FDA00030268193700000521
Figure FDA00030268193700000522
步骤XIV,
Figure FDA00030268193700000523
按照下式与卷积核进行运算得到
Figure FDA00030268193700000524
Figure FDA00030268193700000525
其中,
Figure FDA00030268193700000526
为Nu×Co的卷积层输出数组,1≤i3≤Co,Co设置为大于0的整数,Wo
Figure FDA00030268193700000527
的可训练参数数组,bo为大小为1×Co的可训练参数数组;
步骤XV,对
Figure FDA00030268193700000528
按照下式进行补零操作并重新赋值给
Figure FDA00030268193700000529
Figure FDA00030268193700000530
步骤XVI,
Figure FDA00030268193700000531
按照下式与卷积核进行运算得到
Figure FDA00030268193700000532
Figure FDA0003026819370000061
其中,
Figure FDA0003026819370000062
为Nu×2的DenseNet输出,1≤i3≤2,Wout为3×Co×2的可训练参数数组,bout为小为1×2的可训练参数数组,a为OFDM系统发送M-QAM调制符号的实部或者虚部的幅度最大值,M表示调制阶数,tanh(x)为激活函数,
Figure FDA0003026819370000063
4.根据权利要求1所述的适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法,其特征在于:所述第一级神经网络DenseNet-1和第二级神经网络DenseNet-2的训练阶段包括如下步骤:
步骤一,获得训练样本
Figure FDA0003026819370000064
组成训练样本集Ψ1,其中,训练样本集Ψ1的生成信噪比在工作信噪比区间SNRlow至SNRhigh内均匀分布,训练样本集Ψ1的大小为BS1,Zk表示训练标签,为Nu×2的数组,由下式生成
Figure FDA0003026819370000065
步骤二,应用xavier初始化方法初始化DenseNet-1所有的可训练参数,然后根据训练数据完成对DenseNet-1的训练;
步骤三,基于已训练DenseNet-1检测输出的硬判决结果对
Figure FDA0003026819370000066
做干扰消除得到
Figure FDA0003026819370000067
进而得到DenseNet-2的训练样本
Figure FDA0003026819370000068
组成训练样本集Ψ2,其中,训练样本集Ψ2的生成信噪比在工作信噪比区间SNRlow至SNRhigh内均匀分布,训练样本集Ψ2的大小为BS2
步骤四,应用xavier初始化方法初始化DenseNet-2所有的可训练参数,然后根据训练数据完成对DenseNet-2的训练。
CN202110418198.5A 2021-04-19 2021-04-19 一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法 Active CN113162883B (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110418198.5A CN113162883B (zh) 2021-04-19 2021-04-19 一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110418198.5A CN113162883B (zh) 2021-04-19 2021-04-19 一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法

Publications (2)

Publication Number Publication Date
CN113162883A CN113162883A (zh) 2021-07-23
CN113162883B true CN113162883B (zh) 2022-05-06

Family

ID=76868570

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110418198.5A Active CN113162883B (zh) 2021-04-19 2021-04-19 一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法

Country Status (1)

Country Link
CN (1) CN113162883B (zh)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109617847A (zh) * 2018-11-26 2019-04-12 东南大学 一种基于模型驱动深度学习的无循环前缀ofdm接收方法
CN111431837A (zh) * 2020-03-31 2020-07-17 东南大学 一种应对子载波和符号间干扰的ofdm信号迭代检测方法

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109617847A (zh) * 2018-11-26 2019-04-12 东南大学 一种基于模型驱动深度学习的无循环前缀ofdm接收方法
CN111431837A (zh) * 2020-03-31 2020-07-17 东南大学 一种应对子载波和符号间干扰的ofdm信号迭代检测方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
"DNN Based Iterative Detection for High Order QAM OFDM Systems With Insufficient Cyclic Prefix";Huan Cai等;《2020 IEEE/CIC International Conference on Communications in China (ICCC)》;20201109;全文 *
"基于复杂度聚类的自适应遥感场景分类";梁文韬 等;《计算机工程》;20201231;全文 *

Also Published As

Publication number Publication date
CN113162883A (zh) 2021-07-23

Similar Documents

Publication Publication Date Title
CN109922020B (zh) 一种计算复杂度低的正交时频空调制的均衡方法
Muquet et al. Subspace-based blind and semi-blind channel estimation for OFDM systems
Thompson Deep learning for signal detection in non-orthogonal multiple access wireless systems
CN111865863B (zh) 一种基于rnn神经网络的ofdm信号检测方法
JP4147193B2 (ja) マルチキャリヤ拡散スペクトル信号の受信
CN108881080B (zh) 一种基于滑动窗与深度学习的ofdm抗ici检测方法
CN113852575A (zh) 一种基于时域信道均衡辅助的迭代otfs符号检测方法
CN115296970B (zh) 基于逐元素外部信息的迭代正交时频空波形检测方法
CN100477651C (zh) 一种基于组合导频的高性能ofdm信道估计方法
CN112564830B (zh) 一种基于深度学习的双模正交频分复用索引调制检测方法及装置
CN112636855A (zh) 一种ofdm信号检测方法
CN113285902A (zh) 一种ofdm系统检测器设计方法
CN113162883B (zh) 一种适用于无CP OFDM系统的基于DenseNet的并行干扰消除检测方法
CN111641576A (zh) 一种基于索引调制的降低ofdm信号峰均比值的方法
CN107231323B (zh) 可见光通信系统中基于可靠判决反馈的信道估计方法
CN116388800A (zh) 基于快速贝叶斯匹配追踪的脉冲噪声抑制方法
CN111431837B (zh) 一种应对子载波和符号间干扰的ofdm信号迭代检测方法
CN113556305B (zh) 适用于高频率选择性衰落的fbmc迭代信道均衡方法及系统
CN110958205B (zh) 一种基于共享cp的多符号联合均衡混合载波传输方法
CN107171989A (zh) 可见光通信系统中基于dft的信道估计方法
CN109412659B (zh) 多天线ofdm指数调制方法
Soman et al. Improved DFT-based channel estimation for spatial modulated orthogonal frequency division multiplexing systems
Abd El-Hamid et al. FFT/DWT/DCT OFDM channel estimation using EM algorithm in the presence of chaotic interleaving
CN114143145B (zh) 一种基于深度学习的信道估计方法
Meng et al. MMSE channel shortening equalization for OFDM systems with unequal subcarrier powers

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant