【OFDM通信】基于深度学习的OFDM系统信号检测附matlab代码

matlab科研助手

已于 2022-08-04 11:02:27 修改

阅读量1.2k

点赞数

CC 4.0 BY-SA版权

分类专栏：通信文章标签：深度学习 matlab 机器学习

于 2022-07-06 10:05:22 首次发布

本文链接：https://blog.youkuaiyun.com/matlab_dingdang/article/details/125633919

通信专栏收录该内容

81 篇文章

订阅专栏

1 简介

提供了一种基于深度学习的时变OFDM系统信号检测方法,生成信号检测网络模型输入数据集,构建一个信号检测网络模型,在网络训练前需预设好训练和测试参数,采用在线生成训练数据和测试数据的方式训练网络,测试数据喂入信号检测网络,信号检测网络模型根据喂入的特征向量,产生预测出的发送数据比特,与真实的发送数据比特进行对比,测试网络当前的性能.本发明于针对快速时变OFDM系统,结合深度学习方法,利用循环神经网络处理时间序列的优势,简化了接收机架构,成功实现信号的解调,改进了快速时变OFDM系统中的信号检测性能,本发明有效减小了系统实现复杂度,同时也提升了系统整体的误比特率性能.

2 部分代码

function ReceivedPacket = genTransmissionReceptionOFDM(TransmittedFrame,LengthCP,h,NoiseVar) % This function is to model the transmission and reception process in OFDM systems.

% Extract parameters [NumSym,NumSC,NumPacket] = size(TransmittedFrame);

%% Transmitter

PhaseShift = exp(-1j*rand(1,NumPacket)*2*pi);

for p = 1:NumPacket % 1. IFFT x1 = ifft(TransmittedFrame(:,:,p),NumSC,2); % 2. Inserting CP x1_CP = [x1(:,NumSC-LengthCP+1:end) x1]; % 3. Parallel to serial transformation x2 = x1_CP.'; x = x2(:); % 4. Channel filtering y_conv = conv(h*PhaseShift(p),x); y(:,p) = y_conv(1:length(x)); end %% Adding noise

SeqLength = size(y,1); % Calculate random noise in time domain NoiseF = sqrt(NoiseVar)/sqrt(2).*(randn(NumPacket,NumSC)+1j*randn(NumPacket,NumSC)); % Frequency-domain noise NoiseT = sqrt(SeqLength)*sqrt(SeqLength/NumSC)*ifft(NoiseF,SeqLength,2); % Time-domain noise

% Adding noise y = y+NoiseT.';

%% Receiver

ReceivedPacket = zeros(NumPacket,NumSym,NumSC); for p = 1:NumPacket % 1. Serial to parallem transformation y1 = reshape(y(:,p),NumSC+LengthCP,NumSym).';

% 2. Removing CP y2 = y1(:,LengthCP+1:LengthCP+NumSC);

% 3. FFT, # x NymSym x 64 ReceivedPacket(p,:,:) = fft(y2,NumSC,2); % NumSym x 64 end

ReceivedPacket = permute(ReceivedPacket,[2,3,1]);

%% Testing
%
% This script
% 1. generates testing data for each SNR point;
% 2. calculates the symbol error rate (SER) based on deep learning (DL),
% least square (LS) and minimum mean square error (MMSE).

%% Clear workspace

clear variables;
close all;

%% Load common parameters and the trained NN

load('SimParametersPilot64.mat');
load('TrainedNetPilot64.mat');

%% Other simulation parameters

NumPilot = length(FixedPilot);
PilotSpacing = NumSC/NumPilot;
NumOFDMsym = NumPilotSym+NumDataSym;
NumClass = length(Label);
NumPath = length(h);

% Load pre-calculated channel autocorrelation matrix for MMSE estimation
% This autocorrelation matrix is calculated in advance using the 3GPP
% channel model, which can be replaced accordingly.
load('RHH.mat');

%% SNR range

Es_N0_dB = 0:2:20; % Es/N0 in dB
Es_N0 = 10.^(Es_N0_dB./10); % linear Es/N0
N0 = 1./Es_N0;
NoiseVar = N0./2;

%% Testing data size

NumPacket = 10000; % Number of packets simulated per iteration

%% Simulation

% Same pilot sequences used in training and testing stages
FixedPilotAll = repmat(FixedPilot,1,1,NumPacket);

% Number of Monte-Carlo iterations
NumIter = 1;

% Initialize error rate vectors
SER_DL = zeros(length(NoiseVar),NumIter);
SER_LS = zeros(length(NoiseVar),NumIter);
SER_MMSE = zeros(length(NoiseVar),NumIter);

for i = 1:NumIter

for snr = 1:length(NoiseVar)

%% 1. Testing data generation

noiseVar = NoiseVar(snr);

% OFDM pilot symbol (can be interleaved with random data symbols)
PilotSym = 1/sqrt(2)*complex(sign(rand(NumPilotSym,NumSC,NumPacket)-0.5),sign(rand(NumPilotSym,NumSC,NumPacket)-0.5));
PilotSym(1:PilotSpacing:end) = FixedPilotAll;

% OFDM data symbol
DataSym = 1/sqrt(2)*complex(sign(rand(NumDataSym,NumSC,NumPacket)-0.5),sign(rand(NumDataSym,NumSC,NumPacket)-0.5));

% Transmitted OFDM frame
TransmittedPacket = [PilotSym;DataSym];

% Received OFDM frame
ReceivedPacket = genTransmissionReceptionOFDM(TransmittedPacket,LengthCP,h,noiseVar);

% Collect the data labels for the selected subcarrier
DataLabel = zeros(size(DataSym(:,idxSC,:)));
for c = 1:NumClass
DataLabel(logical(DataSym(:,idxSC,:) == 1/sqrt(2)*Mod_Constellation(c))) = Label(c);
end
DataLabel = squeeze(DataLabel);

% Testing data collection
XTest = cell(NumPacket,1);
YTest = zeros(NumPacket,1);
for c = 1:NumClass
[feature,label,idx] = getFeatureAndLabel(real(ReceivedPacket),imag(ReceivedPacket),DataLabel,Label(c));
featureVec = mat2cell(feature,size(feature,1),ones(1,size(feature,2)));
XTest(idx) = featureVec;
YTest(idx) = label;
end
YTest = categorical(YTest);

%% 2. DL detection

YPred = classify(Net,XTest,'MiniBatchSize',MiniBatchSize);
SER_DL(snr,i) = 1-sum(YPred == YTest)/NumPacket;

%% 3. LS & MMSE detection

% Channel estimation
wrapper = @(x,y) performChanEstimation(x,y,RHH,noiseVar,NumPilot,NumSC,NumPath,idxSC);
ReceivedPilot = mat2cell(ReceivedPacket(1,:,:),1,NumSC,ones(1,NumPacket));
PilotSeq = mat2cell(FixedPilotAll,1,NumPilot,ones(1,NumPacket));
[EstChanLS,EstChanMMSE] = cellfun(wrapper,ReceivedPilot,PilotSeq,'UniformOutput',false);
EstChanLS = cell2mat(squeeze(EstChanLS));
EstChanMMSE = cell2mat(squeeze(EstChanMMSE));

% Symbol detection
SER_LS(snr,i) = getSymbolDetection(ReceivedPacket(2,idxSC,:),EstChanLS,Mod_Constellation,Label,DataLabel);
SER_MMSE(snr,i) = getSymbolDetection(ReceivedPacket(2,idxSC,:),EstChanMMSE,Mod_Constellation,Label,DataLabel);

end

end

SER_DL = mean(SER_DL,2).';
SER_LS = mean(SER_LS,2).';
SER_MMSE = mean(SER_MMSE,2).';

figure();
semilogy(Es_N0_dB,SER_DL,'r-o','LineWidth',2,'MarkerSize',10);hold on;
semilogy(Es_N0_dB,SER_LS,'b-o','LineWidth',2,'MarkerSize',10);hold on;
semilogy(Es_N0_dB,SER_MMSE,'k-o','LineWidth',2,'MarkerSize',10);hold off;
legend('Deep learning (DL)','Least square (LS)','Minimum mean square error (MMSE)');
xlabel('Es/N0 (dB)');
ylabel('Symbol error rate (SER)');

function [EstChanLS,EstChanMMSE] = performChanEstimation(ReceivedData,PilotSeq,RHH,NoiseVar,NumPilot,NumSC,NumPath,idxSC)
% This function is to perform LS and MMSE channel estimations using pilot
% symbols, second-order statistics of the channel and noise variance [1].

% [1] O. Edfors, M. Sandell, J. -. van de Beek, S. K. Wilson and
% P. Ola Borjesson, "OFDM channel estimation by singular value
% decomposition," VTC, Atlanta, GA, USA, 1996, pp. 923-927 vol.2.

PilotSpacing = NumSC/NumPilot;

%%%%%%%%%%%%%%% LS estimation with interpolation %%%%%%%%%%%%%%%%%%%%%%%%%

H_LS = ReceivedData(1:PilotSpacing:NumSC)./PilotSeq;
H_LS_interp = interp1(1:PilotSpacing:NumSC,H_LS,1:NumSC,'linear','extrap');
H_LS_interp = H_LS_interp.';

%%%%%%%%%%%%%%%% MMSE estimation based on LS %%%%%%%%%%%%%%%%

[U,D,V] = svd(RHH);
d = diag(D);

InvertValue = zeros(NumSC,1);
if NumPilot >= NumPath

InvertValue(1:NumPilot) = d(1:NumPilot)./(d(1:NumPilot)+NoiseVar);

else

InvertValue(1:NumPath) = d(1:NumPath)./(d(1:NumPath)+NoiseVar);

end

H_MMSE = U*diag(InvertValue)*V'*H_LS_interp;

%%%%%%%%%%%%%%% Channel coefficient on the selected subcarrier %%%%%%%%%%%

EstChanLS = H_LS_interp(idxSC);
EstChanMMSE = H_MMSE(idxSC);

end

function SER = getSymbolDetection(ReceivedData,EstChan,Mod_Constellation,Label,DataLabel)
% This function is to calculate the symbol error rate from the equalized
% symbols based on hard desicion.

EstSym = squeeze(ReceivedData)./EstChan;

% Hard decision
DecSym = sign(real(EstSym))+1j*sign(imag(EstSym));
DecLabel = zeros(size(DecSym));
for c = 1:length(Mod_Constellation)
DecLabel(logical(DecSym == Mod_Constellation(c))) = Label(c);
end

SER = 1-sum(DecLabel == DataLabel)/length(EstSym);

end