信号处理——Hilbert端点效应浅析

作者:桂。

时间:2017-03-05  19:29:12

链接:http://www.cnblogs.com/xingshansi/p/6506405.html 

声明:转载请注明出处,谢谢。


前言

本文为Hilbert变换分析的补充,主要介绍Hilbert变换中的端点效应,内容拟分两部分展开:

  1)Gibbs现象介绍;

  2)端点效应分析;

内容为自己一家之言,其中不合理的地方,希望各位告诉我,反正改不改是我的事(●'◡'●)

一、Gibbs现象

 关于Gibbs的理论推导,可以参考郑君里的《信号与系统》上册(第二版,P97~101),里边有详细的理论分析。也可参考:维基百科,里边有一张图很形象:

给出对应的代码(用该程序,借助MATLAB便可以生成.gif图):

%% 代码介绍========================================================================
%% Title: 
%  1) Fourier series approximation of square wave.
%  2) Demonstration of Gibbs phenomenon (verification of Fig. 3.9 of [1])
%% Author: 
%  Ankit A. Bhurane (ankit.bhurane@gmail.com)
%% Expression:
% The Fourier series approximation of a square wave signal existing between
% -Tau/2 to Tau/2 and period of T0 will have the form:
%
% Original signal to be approximated: 
%                         :
%               __________:__________  A
%              |          :          |
%              |          :          |
%              |          :          |
%    __________|          :          |__________
% -T0/2     -Tau/2        0        Tau/2       T0/2
%                         :
%                         :
%
% Its Fourier series approximation:
%
%                  Inf   
%                  ___ 
%         A*Tau        / sin(pi*n*Tau/T0)                    
% r(t) = -------   |   | ------------------ exp^(j*n*2*pi/T0*t) |
%          T0     /___    (pi*n*Tau/T0)                      /
%               n = -Inf 
%
% The left term inside summation are the Fourier series coeffs (Cn). The
% right term is the Fourier series kernel. 
% Tau: range of square wave, T: period of the square wave, 
% t: time variable, n: number of retained coefficients.
% 
%% Observations:
% 1) As number of retained coefficients tends to infinity, the approximated
% signal value at the discontinuity converge to half the sum of values on
% either side. 
% 2) Ripples does not decrease with increasing coefficients with
% approximately 9% overshoot.
% 3) Energy in the error between original and approximated signal, reduces
% as the number of retained coefficients are increased.
% 
%% References: 
% [1] Oppenheim, Willsky, Nawab, "Signals and Systems", PHI, Second edition
% [2] Dean K. Frederick and A. Bruce Carlson, "Fourier series" section in 
%     Linear systems in communication and control 
%% Last Modified: Sept 24, 2013.
%% Copyright (c) 2013-2014 | Ankit A. Bhurane
%% 

  

clc; clear all; close all;

% Specification
A = 1; % Peak-to-peak amplitude of square wave
Tau = 10; % Total range in which the square wave is defined (here -5 to 5)
T0 = 20; % Period (time of repeatation of square wave), here 10 
C = 200; % Coefficients (sinusoids) to retain 
N = 1001; % Number of points to consider
t = linspace(-(T0-Tau),(T0-Tau),N); % Time axis
X = zeros(1,N); X(t>=-Tau/2 & t<=Tau/2) = A; % Original signal
R = 0; % Initialize the approximated signal
k = -C:C; % Fourier coefficient number axis
f = zeros(1,2*C+1); % Fourier coefficient values

% Loop for plotting approximated signals for different retained coeffs.
for c = 0:C % Number of retained coefficients
    for n = -c:c % Summation range (See equation above in comments)
        
        % Sinc part of the Fourier coefficients calculated separately 
        if n~=0
            Sinc = (sin(pi*n*Tau/T0)/((pi*n*Tau/T0))); % At n NOTEQUAL to 0
        else
            Sinc = 1; % At n EQUAL to 0
        end
        Cn = (A*Tau/T0)*Sinc; % Actual Fourier series coefficients
        f(k==n) = Cn; % Put the Fourier coefficients at respective places
        R = R + Cn*exp(1j*n*2*pi/T0.*t); % Sum all the coefficients
    end
    
    R = real(R); % So as to get rid of 0.000000000i (imaginary) factor
    Max = max(R); Min = min(R); M = max(abs(Max),abs(Min)); % Maximum error
    Overshoot = ((M-A)/A)*100; % Overshoot calculation
    E = sum((X-R).^2); % Error energy calculation
    
    % Plots:
    % Plot the Fourier coefficients
    subplot 211; stem(k,f,'m','LineWidth',1); axis tight; grid on;
    xlabel('Fourier coefficient index');ylabel('Magnitude');
    title('Fourier coefficients');
    
    % Plot the approximated signal
    subplot 212; plot(t,X,t,R,'m','LineWidth',1); axis tight; grid on; 
    xlabel('Time (t)'); ylabel('Amplitude');
    title(['Approximation for N = ', num2str(c),...
    '. Overshoot = ',num2str(Overshoot),'%','. Error energy: ',num2str(E)])

    pause(0.1); % Pause for a while
    R = 0; % Reset the approximation to calculate new one
end

  简而言之:

  • 对于跳变的点,傅里叶变换的分量只能是能量收敛,而不是一致(幅度)收敛;
  • 对于跳变的点,如门函数/方波,信号由低频到高频众多分量组合而成。

故对于连续时域信号,Fourier级数只能无限项逼近,不能完全一致。 

二、端点效应

  A-理论解释

借用之前博文的一张图:

连续信号到离散信号需要进行采样,由于有无穷多项,采样率对于高频的部分(因为无穷多项)难以满足Nyquist采样定理,因此会出现失真,这也就是端点效应的理论解释。失真的根本原因在时域采样,频域采样步骤只不过影响频域分辨率,即所谓的栅栏效应,但不是造成失真的根本原因。

  B-现象分析

为了验证理论,此处采样一个正弦信号进行分析,按连续信号来看其Hilbert变换应该是余弦信号。

给出代码:

clc;clear all;close all;
fs = 2000;
f0 = 80;
t = 0:1/fs:.1;
sig = sin(2*pi*f0*t);
sig_ref = -cos(2*pi*f0*t);
sig_hilbert = hilbert(sig);
figure;
subplot 311
plot(t,sig,'k','linewidth',2);hold on;
plot(t,sig_ref,'r.-','linewidth',2);hold on;
plot(t,imag(sig_hilbert),'b','linewidth',2);hold on;
legend('原信号','余弦信号','Hilbert变换信号','localization','best');
ylim([-2,2]);
grid on;
%频谱
f = linspace(0,fs,length(t));
subplot 312
plot(f,abs(fft(sig)),'k');hold on;
plot(f,abs(fft(sig_hilbert)),'r');hold on;
legend('原信号','Hilbert变换信号','localization','best');
grid on;

  对应的结果图:

时域

可以看到明显的端点效应(说是端点效应,如果间断点在中间位置,一样会有该效应,毕竟都是Nyquist定理下的误差嘛)。

频域

采样率取得够大了,理论上正弦信号Hilbert之后应该是一个单边的冲击。但频谱显示,峰值旁边的一个点,幅度达到160.6,这也印证了能量的泄露。

发现端点效应,与前文理论对应:Hilbert就是一个频域的变相器,本质也是fourier变换,是不是也因为跳变点(即信号中的Gibbs问题)?观察频谱,峰值旁边的点,幅值有160.5大小,我们计算原信号

2*sig(1)-sig(2)

  得到结果是:-0.4820,而对于没有的点,应该是默认是0的,因此端点算是一个跳变点了

注意:端点效应是相对于连续信号来讲,存在失真,从而出现误差。对于数字信号,Hilbert变换结果就是该正弦数字信号的精确表达,即逆Hilbert可以完全恢复出正弦信号给一张Hilbert变换后,进行逆Hilbert变换的code及示意图:

clc;clear all;close all;
fs = 2000;
f0 = 40;
t = 0:1/fs:.1;
sig = sin(2*pi*f0*t);
sig_hilbert = -imag(hilbert(imag(hilbert(sig))));
figure;
plot(t,sig,'k','linewidth',2);hold on;
plot(t,sig_hilbert,'r--','linewidth',2);
ylim([-2,2]);
grid on;

  从图中,可以看到从Hilbert结果恢复的信号与原信号完全一致!这也印证了上文的观点:失真在时域采样,而不在频域采样。


下面再做一组实验,验证推断
思路——只要保证2sig(1)-sig(2)逼近0,理论上是不是就没有边界效应了?
此时:2*sig(1)-sig(2)=1.5873e-05,再来看看结果图:

 再来看看频谱,没错,峰值旁边的点压的很小(6.183e-13),符合预期。

此处代码:

clc;clear all;close all;
fs = 20000;
f0 = 80;
t = 0.01255:1/fs:.1;%t = 0.0252:1/fs:.1则没有跳变。
sig = sin(2*pi*f0*t);
2*sig(1)-sig(2)
sig_ref = -cos(2*pi*f0*t);
sig_hilbert = hilbert(sig);
figure;
subplot 211
plot(t,sig,'k','linewidth',2);hold on;
plot(t,sig_ref,'r.-','linewidth',2);hold on;
plot(t,imag(sig_hilbert),'b','linewidth',2);hold on;
legend('原信号','余弦信号','Hilbert变换信号','localization','best');
ylim([-2,2]);
grid on;
%频谱
f = linspace(0,fs,length(t));
subplot 212
plot(f,abs(fft(sig)),'k');hold on;
plot(f,abs(fft(sig_hilbert)),'r');hold on;
legend('原信号','Hilbert变换信号','localization','best');
ylim([0 6000]);
grid on;

基于这个基本要点,很多方法对信号:预测、延拓、镜像等操作,以消除或者降低端点效应。为什么用2*sig(1)-sig(2)呢,这是因为实验的函数为sin,在0附近sinx~x ,严格意义上应该是2*f(1)-f(2)≈ 0 。f是函数表达式,而且实际中信号不是单一频点,仅仅通过两个点判断也是不够的。
只是消除端点方法不同,但端点现象的本质一致。
多说一句:Hilbert将双边谱压成单边,这让频域滤波等操作方便了不少。

参考:

Gibbs phenomen:https://en.wikipedia.org/wiki/Gibbs_phenomenon

原文地址:https://www.cnblogs.com/xingshansi/p/6506405.html