希尔伯特变换在MATLAB中的应用

一、基本理论

  A-Hilbert变换定义

对于一个实信号x(t)x(t),其希尔伯特变换为:

x~(t)=x(t)1πtx~(t)=x(t)∗1πt

式中*表示卷积运算。

Hilbert本质上也是转向器,对应频域变换为:

1πtjsign(ω)1πt⇔j⋅sign(ω)

即余弦信号的Hilbert变换时正弦信号,又有:

1πt1πtjsign(ω)jsign(ω)=11πt∗1πt⇔j⋅sign(ω)⋅j⋅sign(ω)=−1

即信号两次Hilbert变换后是其自身相反数,因此正弦信号的Hilbert是负的余弦。

对应解析信号为:

z(t)=x(t)+jx~(t)z(t)=x(t)+jx~(t)

此操作实现了信号由双边谱到单边谱的转化。

  B-Hilbert解调原理

设有窄带信号:

x(t)=a(t)cos[2πfst+φ(t)]x(t)=a(t)cos⁡[2πfst+φ(t)]

其中fsfs是载波频率,a(t)a(t)x(t)x(t)的包络,φ(t)φ(t)x(t)x(t)的相位调制信号。由于x(t)x(t)是窄带信号,因此a(t)a(t)也是窄带信号,可设为:

a(t)=[1+m=1MXmcos(2πfmt+γm)]a(t)=[1+∑m=1MXmcos⁡(2πfmt+γm)]

式中,fmfm为调幅信号a(t)a(t)的频率分量,γmγmfmfm的各初相角。

x(t)x(t)进行Hilbert变换,并求解解析信号,得到:

z(t)=ej[2πfs+φ(t)][1+m=1MXmcos(2πfmt+γm)]z(t)=ej[2πfs+φ(t)][1+∑m=1MXmcos⁡(2πfmt+γm)]

A(t)=[1+m=1MXmcos(2πfmt+γm)]A(t)=[1+∑m=1MXmcos⁡(2πfmt+γm)]

Φ(t)=2πfst+φ(t)Φ(t)=2πfst+φ(t)

则解析信号可以重新表达为:

z(t)=A(t)ejΦ(t)z(t)=A(t)ejΦ(t)

对比x(t)x(t)表达式,容易发现

a(t)=A(t)=x2(t)+x~2(t)a(t)=A(t)=x2(t)+x~2(t)

φ(t)=Φ(t)2πfst=arctanx(t)x~(t)2πfstφ(t)=Φ(t)−2πfst=arctan⁡x(t)x~(t)−2πfst

由此可以得出:对于窄带信号x(t)x(t),利用Hilbert可以求解解析信号,从而得到信号的幅值解调a(t)a(t)和相位解调φ(t)φ(t),并可以利用相位解调求解频率解调f(t)f(t)因为:

f(t)=12πdφ(t)dt=12πdΦ(t)dtfsf(t)=12πdφ(t)dt=12πdΦ(t)dt−fs

  C-相关MATLAB指令

  • hilbert

功能:将实数信号x(n)进行Hilbert变换,并得到解析信号z(n).

调用格式:z = hilbert(x)

  • instfreq

功能:计算复信号的瞬时频率。

调用格式:[f, t] = insfreq(x,t)

示例

1
2
z = hilbert(x);
f = instfreq(z);

 

二、应用实例

 例1:给定一正弦信号,画出其Hilbert信号,直接给代码:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
clc
clear  all
close  all
ts = 0.001;
fs = 1/ts;
N = 200;
f = 50;
k = 0:N-1;
t = k*ts;
% 信号变换
% 结论:sin信号Hilbert变换后为cos信号
y =  sin (2* pi *f*t);
yh = hilbert(y);     % matlab函数得到信号是合成的复信号
yi =  imag (yh);       % 虚部为书上定义的Hilbert变换
figure
subplot (211)
plot (t, y)
title ( '原始sin信号' )
subplot (212)
plot (t, yi)
title ( 'Hilbert变换信号' )
ylim ([-1,1])

  对应效果图:

例2:已知信号x(t)=(1+0.5cos(2π5t))cos(2π50t+0.5sin(2π10t))x(t)=(1+0.5cos⁡(2π5t))cos⁡(2π50t+0.5sin⁡(2π10t)),求解该信号的包络和瞬时频率。

分析:根据解包络原理知:

信号包络(1+0.5cos(2π5t))(1+0.5cos⁡(2π5t))

瞬时频率2π50t+0.5sin(2π10t)2π2π50t+0.5sin⁡(2π10t)2π

那么问题来了,实际情况是:我们只知道x(t)x(t)的结果,而不知道其具体表达形式,这个时候,上文的推导就起了作用:可以借助信号的Hilbert变换,从而求解信号的包络和瞬时频率。

对应代码:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
clear  all clc close  all ;
 
fs=400;                                  % 采样频率
N=400;                                   % 数据长度
n=0:1:N-1;
dt=1/fs;
t=n*dt;                                  % 时间序列
A=0.5;                                   % 相位调制幅值
x=(1+0.5* cos (2* pi *5*t)).* cos (2* pi *50*t+A* sin (2* pi *10*t));   % 信号序列
z=hilbert(x');                           % 希尔伯特变换
a= abs (z);                                % 包络线
fnor=instfreq(z);                        % 瞬时频率
fnor=[fnor(1); fnor; fnor( end )];         % 瞬时频率补齐
% 作图
pos =  get ( gcf , 'Position' );
set ( gcf , 'Position' ,[pos(1), pos(2)-100,pos(3),pos(4)]);
subplot  211;  plot (t,x, 'k' );  hold  on;
plot (t,a, 'r--' , 'linewidth' ,2);
title ( '包络线' );  ylabel ( '幅值' );  xlabel ([ '时间/s'  10  '(a)' ]);
ylim ([-2,2]);
subplot  212;  plot (t,fnor*fs, 'k' );  ylim ([43 57]);
title ( '瞬时频率' );  ylabel ( '频率/Hz' );   xlabel ([ '时间/s'  10  '(b)' ]);

  其中instfreq为时频工具包的代码,可能有的朋友没有该代码,这里给出其程序:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
function  [fnormhat,t]=instfreq(x,t,L, trace );
%INSTFREQ Instantaneous frequency estimation.
%   [FNORMHAT,T]=INSTFREQ(X,T,L,TRACE) computes the instantaneous
%   frequency of the analytic signal X at time instant(s) T, using the
%   trapezoidal integration rule.
%   The result FNORMHAT lies between 0.0 and 0.5.
%
%   X : Analytic signal to be analyzed.
%   T : Time instants           (default : 2:length(X)-1).
%   L : If L=1, computes the (normalized) instantaneous frequency
%       of the signal X defined as angle(X(T+1)*conj(X(T-1)) ;
%       if L>1, computes a Maximum Likelihood estimation of the
%       instantaneous frequency of the deterministic part of the signal
%       blurried in a white gaussian noise.
%       L must be an integer        (default : 1).
%   TRACE : if nonzero, the progression of the algorithm is shown
%                                   (default : 0).
%   FNORMHAT : Output (normalized) instantaneous frequency.
%   T : Time instants.
%
%   Examples :
%    x=fmsin(70,0.05,0.35,25); [instf,t]=instfreq(x); plot(t,instf)
%    N=64; SNR=10.0; L=4; t=L+1:N-L; x=fmsin(N,0.05,0.35,40);
%    sig=sigmerge(x,hilbert(randn(N,1)),SNR);
%    plotifl(t,[instfreq(sig,t,L),instfreq(x,t)]); grid;
%    title ('theoretical and estimated instantaneous frequencies');
%
%   See also  KAYTTH, SGRPDLAY.
 
%   F. Auger, March 1994, July 1995.
%   Copyright (c) 1996 by CNRS (France).
%
%   ------------------- CONFIDENTIAL PROGRAM --------------------
%   This program can not be used without the authorization of its
%   author(s). For any comment or bug report, please send e-mail to
 
if  ( nargin  == 0),
  error ( 'At least one parameter required' );
end ;
[xrow,xcol] =  size (x);
if  (xcol~=1),
  error ( 'X must have only one column' );
end
 
if  ( nargin  == 1),
  t=2:xrow-1; L=1;  trace =0.0;
elseif  ( nargin  == 2),
  L = 1;  trace =0.0;
elseif  ( nargin  == 3),
  trace =0.0;
end ;
 
if  L<1,
  error ( 'L must be >=1' );
end
[trow,tcol] =  size (t);
if  (trow~=1),
  error ( 'T must have only one row' );
end ;
 
if  (L==1),
  if  any (t==1)| any (t==xrow),
   error ( 'T can not be equal to 1 neither to the last element of X' );
  else
   fnormhat=0.5*( angle (-x(t+1).* conj (x(t-1)))+ pi )/(2* pi );
  end ;
else
  H=kaytth(L);
  if  any (t<=L)| any (t+L>xrow),
   error ( 'The relation L<T<=length(X)-L must be satisfied' );
  else
   for  icol=1:tcol,
    if  trace , disprog(icol,tcol,10);  end ;
    ti = t(icol); tau = 0:L;
    R = x(ti+tau).* conj (x(ti-tau));
    M4 = R(2:L+1).* conj (R(1:L));
    
    diff =2e-6;
    tetapred = H * ( unwrap ( angle (-M4))+ pi );
    while  tetapred<0.0 , tetapred=tetapred+(2* pi );  end ;
    while  tetapred>2* pi , tetapred=tetapred-(2* pi );  end ;
    iter = 1;
    while  ( diff  > 1e-6)&(iter<50),
     M4bis=M4 .*  exp (- j *2.0*tetapred);
     teta = H * ( unwrap ( angle (M4bis))+2.0*tetapred);
     while  teta<0.0 , teta=(2* pi )+teta;  end ;
     while  teta>2* pi , teta=teta-(2* pi );  end ;
     diff = abs (teta-tetapred);
     tetapred=teta; iter=iter+1;
    end ;
    fnormhat(icol,1)=teta/(2* pi );
   end ;
  end ;
end ;

  对应的结果图为:

可以看到信号的包络、瞬时频率,均已完成求解。

 例3:例2中信号包络为规则的正弦函数,此处给定任意形式的包络(以指数形式为例),并利用Hilbert求解包络以及瞬时频率,并给出对应的Hilbert谱。

程序:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
clc
clear  all
close  all
ts = 0.001;
fs = 1/ts;
N = 200;
k = 0:N-1;
t = k*ts;
% 原始信号
f1 = 10;
f2 = 70;
% a = cos(2*pi*f1*t);       % 包络1
a = 2 +  exp (0.2*f1*t);      % 包络2
% a = 1./(1+t.^2*50);       % 包络3
m =  sin (2* pi *f2*t);          % 调制信号
y = a.*m;   % 信号调制
figure
subplot (241)
plot (t, a)
title ( '包络' )
subplot (242)
plot (t, m)
title ( '调制信号' )
subplot (243)
plot (t, y)
title ( '调制结果' )
% 包络分析
% 结论:Hilbert变换可以有效提取包络、高频调制信号的频率等
yh = hilbert(y);
aabs =  abs (yh);                  % 包络的绝对值
aangle =  unwrap ( angle (yh));      % 包络的相位
af =  diff (aangle)/2/ pi ;          % 包络的瞬时频率,差分代替微分计算
% NFFT = 2^nextpow2(N);
NFFT = 2^ nextpow2 (1024*4);       % 改善栅栏效应
f = fs* linspace (0,1,NFFT);
YH =  fft (yh, NFFT)/N;            % Hilbert变换复信号的频谱
A =  fft (aabs, NFFT)/N;           % 包络的频谱
subplot (245)
plot (t, aabs, 'r' , t, a)
title ( '包络的绝对值' )
legend ( '包络分析结果' '真实包络' )
subplot (246)
plot (t, aangle)
title ( '调制信号的相位' )
subplot (247)
plot (t(1: end -1), af*fs)
title ( '调制信号的瞬时频率' )
subplot (244)
plot (f, abs (YH))
title ( '原始信号的Hilbert谱' )
xlabel ( '频率f (Hz)' )
ylabel ( '|YH(f)|' )
subplot (248)
plot (f, abs (A))
title ( '包络的频谱' )
xlabel ( '频率f (Hz)' )
ylabel ( '|A(f)|' )

  对应结果图:

从结果可以观察,出了边界误差较大,结果值符合预期。对于边界效应的分析,见扩展阅读部分。注意:此处瞬时频率求解,没有用instfreq函数,扩展阅读部分对该函数作进一步讨论

 

三、扩展阅读

  A-瞬时频率求解方法对比

对于离散数据,通常都是用差分代替微分,因此瞬时频率也可根据概念直接求解。此处对比分析两种求解瞬时频率的方法,给出代码:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
clc
clear  all
close  all
ts = 0.001;
fs = 1/ts;
N = 200;
k = 0:N-1;
t = k*ts;
% 原始信号
f1 = 10;
f2 = 70;
% a = cos(2*pi*f1*t);       % 包络1
a = 2 +  exp (0.2*f1*t);      % 包络2
% a = 1./(1+t.^2*50);       % 包络3
m =  sin (2* pi *f2*t);          % 调制信号
y = a.*m;   % 信号调制
figure
yh = hilbert(y);
aangle =  unwrap ( angle (yh));      % 包络的相位
af1 =  diff (aangle)/2/ pi ;          % 包络的瞬时频率,差分代替微分计算
af1 = [af1(1),af1];
subplot  211
plot (t, af1*fs); hold  on;
plot (t,70* ones (1, length (t)), 'r--' , 'linewidth' ,2);
title ( '直接求解调制信号的瞬时频率' );
legend ( '频率估值' , '真实值' , 'location' , 'best' );
subplot  212
af2 = instfreq(yh. ').' ;
af2 = [af2(1),af2,af2( end )];
plot (t, af2*fs); hold  on;
plot (t,70* ones (1, length (t)), 'r--' , 'linewidth' ,2);
title ( 'instfreq求解调制信号的瞬时频率' );
legend ( '频率估值' , '真实值' , 'location' , 'best' );

  结果图:

可以看出,两种方式结果近似,但instfreq的结果更为平滑一些。

  B-端点效应分析

对于任意包络,求解信号的包络以及瞬时频率,容易出现端点误差较大的情况,该现象主要基于信号中的Gibbs现象,限于篇幅,拟为此单独写一篇文章,具体请参考:Hilbert端点效应分析

  C-VMD、EMD

 Hilbert经典应用总绕不开HHT(Hilbert Huang),HHT基于EMD,近年来又出现了VMD分解,拟为此同样写一篇文章,略说一二心得,具体参考:EMD、VMD的一点小思考

   D-解包络方法

需要认识到,Hilbert不是解包络的唯一途径,低通滤波(LPF)等方式一样可以达到该效果,只不过截止频率需要调参。

给出一个Hilbert、低通滤波解包络的代码:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
function  y=envelope(signal,Fs)
 
%Example:
%   load('s4.mat');
%   signal=s4;
%   Fs=12000;
%   envelope(signal,Fs);
clc ;
close  all ;
 
%Normal FFT
y=signal;
figure ();
N=2*2048;T=N/Fs;
sig_f= abs ( fft (y(1:N)',N));
sig_n=sig_f/( norm (sig_f));
freq_s=(0:N-1)/T;
subplot  311
plot (freq_s(2:250),sig_n(2:250)); title ( 'FFT of Original Signal' );
 
 
%Envelope Detection based on Low pass filter and then FFT
[a,b]=butter(2,0.1); %butterworth Filter of 2 poles and Wn=0.1
%sig_abs=abs(signal); % Can be used instead of squaring, then filtering and
%then taking square root
sig_sq=2*signal.*signal; % squaring for rectifing
%gain of 2 for maintianing the same energy in the output
y_sq =  filter (a,b,sig_sq);  %applying LPF
y= sqrt (y_sq); %taking Square root
%advantages of taking square and then Square root rather than abs, brings
%out some hidden information more efficiently
N=2*2048;T=N/Fs;
sig_f= abs ( fft (y(1:N)',N));
sig_n=sig_f/( norm (sig_f));
freq_s=(0:N-1)/T;
subplot  312
plot (freq_s(2:250),sig_n(2:250)); title ( 'Envelope Detection: LPF Method' );
 
 
 
%Envelope Detection based on Hilbert Transform and then FFT
analy=hilbert(signal);
y= abs (analy);
N=2*2048;T=N/Fs;
sig_f= abs ( fft (y(1:N)',N));
sig_n=sig_f/( norm (sig_f));
freq_s=(0:N-1)/T;
subplot  313
plot (freq_s(2:250),sig_n(2:250)); title ( 'Envelope Detection : Hilbert Transform' )

  结果图:

效果是不是也不错?

Hilbert硬件实现思路

思路1(时域处理):借助MATLAB fdatool实现,Hilbert transform,导出滤波器系数

思路2(频域处理)

猜你喜欢

转载自blog.csdn.net/qq_24163555/article/details/80455882