【概率最小均方（PLMS）自适应滤波器】PLMS对高斯和非高斯噪声具有较强的鲁棒性（Matlab代码实现）

最新推荐文章于 2024-11-11 10:06:31 发布

Ps.729

最新推荐文章于 2024-11-11 10:06:31 发布

阅读量749

点赞数 16

文章标签： matlab 开发语言

本文链接：https://blog.csdn.net/weixin_67304359/article/details/135982553

版权

💥💥💞💞欢迎来到本博客❤️❤️💥💥

🏆博主优势：🌞🌞🌞博客内容尽量做到思维缜密，逻辑清晰，为了方便读者。

⛳️座右铭：行百里者，半于九十。

📋📋📋本文目录如下：🎁🎁🎁

目录

💥1 概述

📚2 运行结果

🎉3 参考文献

🌈4 Matlab代码实现

💥1 概述

在最大后验估计框架下，本研究提出了概率最小均方（PLMS）自适应滤波器，用于从噪声数据中估计未知参数向量。PLMS将参数空间和信号空间相结合，将过程的概率分布的先验知识与信号中存在的证据相结合。利用核密度估计来估计先验分布，PLMS对高斯和非高斯噪声具有较强的鲁棒性。为了实现这一点，一些中间估计被缓冲，然后用于估计先验分布。尽管存在偏差补偿算法，但无需估计输入噪声方差。对PLMS进行了理论分析。系统识别和预测的模拟结果显示，PLMS在噪声稳态和非稳态环境中的性能可接受，而且优于偏差补偿和归一化的LMS算法。

📚2 运行结果

部分代码：

%% Initialization
N = 1;
M = 4; % Number of unknown parameters or Optimal Weight Vector Length
observations = 1000; % Number of time observation
LMSiterations = observations; % Number of LMS Iteration
experiments = 100; % Number of experiments for averaging over on.
Wo = ones(M,1)*sqrt(M); % Optimal Weight Vector
stationary = 1; % Choose 1 for stationary model, 0 for non-stationary model
N0 = .5; % Input noise variance
i = 1;
leg=[];
Xbar = [];
ix = 1:observations;
zinp = [];
for N1 = [.12 .38 1.3 ] %output noise variance
display(strcat('N1= ', num2str(N1),'...'))
for experiment = 1 : experiments
%% Generate Input Signal and Desired Output
inputNoise = 'none';
outputNoise = 'gaussian';
N0 = N1;
[X,d,w,zinp,zout,snr_inp] = createData(Wo,N,M,observations,N0,N1,stationary,inputNoise,outputNoise);
snr_out = 10*log10(var(d-zout)/var(zout));
%% PLMS Adaption
muN =0.1; % Step size
bufferL = 30; % Buffer length
[MSD(experiment,:),W,er(experiment,:)] = PLMS(X,d,LMSiterations,N,M,w,muN,bufferL);
%% Performance Analysis of PLMS
for j = 1 : observations
nrmW(j) = norm(W(:,j)-Wo);
end
p = 1;
for j = observations:-1:bufferL+1
b(p) = sum(nrmW(j-bufferL:j-1))/nrmW(j);
b(p) = b(p);
p = p +1;
end
beta = mean(b);
R = corr(reshape(X(1,:,:),M,observations)');
I = eye(M);
I2 = kron(I,I);
f = I2-kron(I,muN*(I+R))-kron(muN*(I+R),I);
q1 = muN*beta/bufferL;
MSDT(experiment) = muN^2*N1^2*R(:)'*inv(I2-f-q1*I2)*I(:);%MSD Theory
end
%% Expectation over different experiments
MeanMSDMS = mean(MSD);
MeanMSDTheory = 10*log10(mean(MSDT)*ones(1,observations));

%% Plot Results
hold on
plot(ix,10*log10(MeanMSDMS),clr1{i});
leg{end+1} = strcat('PLMS, SNR_v=',num2str(floor(snr_out)));
plot(ix,MeanMSDTheory,clr2{i})
leg{end+1} = strcat('PLMS, SNR_v=',num2str(floor(snr_out)),'- Theory');
i=i+1;
end
legend(leg,'Orientation','vertical','FontSize',7);
axis square
grid on
box on
xlabel('Iteration','fontsize',10)
ylabel('MSD(dB)','fontsize',10);

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Estimation of Weights, Using PLMS Algorithm %
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [MSDSingle,W,er] = PLMS(X,d,LMSiterations,N,M,Wo,muN,bufferL)
W = zeros(M,1);
bufferW = zeros(M,bufferL);
h = 1;% .5;
%% Non-cooperative LMS
for n = 1 : LMSiterations
for m = 1 : N
ebar = [d(m,n)-X(m,:,n)*W(:,n)]*X(m,:,n)';
WbarNodesingle(m,n) = norm(Wo(:,n)-W(:,n));
% WbarNodesingle(m,n) = norm(Wo-W(:,n));
er2(m,n) = (d(m,n)-X(m,:,n)*W(:,n))^2;
WBni = zeros(M,1); %similarity of wn and buffered wi
sumB = 0; %normalizing factor
for i = 1 : bufferL
simWni = exp(-(bufferW(:,i)-W(:,n))'*(bufferW(:,i)-W(:,n))/h);
weightedWi(:,i) = simWni*bufferW(:,i);
sumB = sumB + simWni;
end
sumWi = sum(weightedWi./sumB,2);
W(:,n+1) = W(:,n)+ muN*(ebar + sumWi - W(:,n));%gradiant ascend
bufferW = add2buffer(bufferW,W(:,n+1));
end
%% Error analysis of Network
MSDSingle(n)=norm(WbarNodesingle(:,n))^2; % Scalar of error
end
er = mean(er2,1);