这学期的课程选择神经网络。最后的作业处理ECG信号,并利用神经网络识别。
1 ECG引进和阅读ECG信号
1)ECG介绍
详细ECG背景应用就不介绍了,大家能够參考百度 谷歌。仅仅是简单说下ECG的结构:
一个完整周期的ECG信号有 QRS P T 波组成,不同的人相应不用的波形,同一个人在不同的阶段波形也不同。我们须要依据各个波形的特点,提取出相应的特征,对不同的人进行身份识别。
2)ECG信号读取
首先须要到MIT-BIH数据库中下载ECG信号,具体的下载地址与程序读取内容介绍能够參考一下地址(讲述的非常具体):http://blog.csdn.net/chenyusiyuan/article/details/2027887。
读代替码(基于MATLAB)例如以下:
clc; clear all;
%------ SPECIFY DATA ------------------------------------------------------
%%选择文件名称
stringname='111';
%选择你要处理的信号点数
points=10000;
PATH= 'F:\ECG\MIT-BIH database directory'; % path, where data are saved
HEADERFILE= strcat(stringname,'.hea'); % header-file in text format
ATRFILE= strcat(stringname,'.atr'); % attributes-file in binary format
DATAFILE=strcat(stringname,'.dat'); % data-file
SAMPLES2READ=points; % number of samples to be read
% in case of more than one signal:
% 2*SAMPLES2READ samples are read
%------ LOAD HEADER DATA --------------------------------------------------
fprintf(1,'\\n$> WORKING ON %s ...\n', HEADERFILE);
signalh= fullfile(PATH, HEADERFILE);
fid1=fopen(signalh,'r');
z= fgetl(fid1);
A= sscanf(z, '%*s %d %d %d',[1,3]);
nosig= A(1); % number of signals
sfreq=A(2); % sample rate of data
clear A;
for k=1:nosig
z= fgetl(fid1);
A= sscanf(z, '%*s %d %d %d %d %d',[1,5]);
dformat(k)= A(1); % format; here only 212 is allowed
gain(k)= A(2); % number of integers per mV
bitres(k)= A(3); % bitresolution
zerovalue(k)= A(4); % integer value of ECG zero point
firstvalue(k)= A(5); % first integer value of signal (to test for errors)
end;
fclose(fid1);
clear A;
%------ LOAD BINARY DATA --------------------------------------------------
if dformat~= [212,212], error('this script does not apply binary formats different to 212.'); end;
signald= fullfile(PATH, DATAFILE); % data in format 212
fid2=fopen(signald,'r');
A= fread(fid2, [3, SAMPLES2READ], 'uint8')'; % matrix with 3 rows, each 8 bits long, = 2*12bit
fclose(fid2);
M2H= bitshift(A(:,2), -4);
M1H= bitand(A(:,2), 15);
PRL=bitshift(bitand(A(:,2),8),9); % sign-bit
PRR=bitshift(bitand(A(:,2),128),5); % sign-bit
M( : , 1)= bitshift(M1H,8)+ A(:,1)-PRL;
M( : , 2)= bitshift(M2H,8)+ A(:,3)-PRR;
if M(1,:) ~= firstvalue, error('inconsistency in the first bit values'); end;
switch nosig
case 2
M( : , 1)= (M( : , 1)- zerovalue(1))/gain(1);
M( : , 2)= (M( : , 2)- zerovalue(2))/gain(2);
TIME=(0:(SAMPLES2READ-1))/sfreq;
case 1
M( : , 1)= (M( : , 1)- zerovalue(1));
M( : , 2)= (M( : , 2)- zerovalue(1));
M=M';
M(1)=[];
sM=size(M);
sM=sM(2)+1;
M(sM)=0;
M=M';
M=M/gain(1);
TIME=(0:2*(SAMPLES2READ)-1)/sfreq;
otherwise % this case did not appear up to now!
% here M has to be sorted!!!
disp('Sorting algorithm for more than 2 signals not programmed yet!');
end;
clear A M1H M2H PRR PRL;
fprintf(1,'\\n$> LOADING DATA FINISHED \n');
%------ LOAD ATTRIBUTES DATA ----------------------------------------------
atrd= fullfile(PATH, ATRFILE); % attribute file with annotation data
fid3=fopen(atrd,'r');
A= fread(fid3, [2, inf], 'uint8')';
fclose(fid3);
ATRTIME=[];
ANNOT=[];
sa=size(A);
saa=sa(1);
i=1;
while i<=saa
annoth=bitshift(A(i,2),-2);
if annoth==59
ANNOT=[ANNOT;bitshift(A(i+3,2),-2)];
ATRTIME=[ATRTIME;A(i+2,1)+bitshift(A(i+2,2),8)+...
bitshift(A(i+1,1),16)+bitshift(A(i+1,2),24)];
i=i+3;
elseif annoth==60
% nothing to do!
elseif annoth==61
% nothing to do!
elseif annoth==62
% nothing to do!
elseif annoth==63
hilfe=bitshift(bitand(A(i,2),3),8)+A(i,1);
hilfe=hilfe+mod(hilfe,2);
i=i+hilfe/2;
else
ATRTIME=[ATRTIME;bitshift(bitand(A(i,2),3),8)+A(i,1)];
ANNOT=[ANNOT;bitshift(A(i,2),-2)];
end;
i=i+1;
end;
ANNOT(length(ANNOT))=[]; % last line = EOF (=0)
ATRTIME(length(ATRTIME))=[]; % last line = EOF
clear A;
ATRTIME= (cumsum(ATRTIME))/sfreq;
ind= find(ATRTIME <= TIME(end));
ATRTIMED= ATRTIME(ind);
ANNOT=round(ANNOT);
ANNOTD= ANNOT(ind);
%------ DISPLAY DATA ------------------------------------------------------
figure(1); clf, box on, hold on ;grid on ;
plot(TIME, M(:,1),'r');
if nosig==2
plot(TIME, M(:,2),'b');
end;
for k=1:length(ATRTIMED)
text(ATRTIMED(k),0,num2str(ANNOTD(k)));
end;
xlim([TIME(1), TIME(end)]);
xlabel('Time / s'); ylabel('Voltage / mV');
string=['ECG signal ',DATAFILE];
title(string);
fprintf(1,'\\n$> DISPLAYING DATA FINISHED \n');
% -------------------------------------------------------------------------
fprintf(1,'\\n$> ALL FINISHED \n');
以MIT-BIH数据库中111.dat 为例。
2 去除高频噪声与基线漂移
ECG读取完后,原始ECG信号含有高频噪声和基线漂移,利用小波方法能够去除对应噪声。
详细原理例如以下:将一维的ECG信号进行8层的小波分解后(MATLAB下wavedec函数,小波类型是bior2.6)得到对应的细节系数与近似系数。依据小波原理我们能够知道。1,2层的细节系数包括了大部分高频噪声,8层的近似系数包括了基线漂移。
基于此。我们将1,2层的细节系数(即高频系数置0),8成的近似系数(低频系数)置0。在对应进行小波重构,重构后我们能够明显得到去噪信号。信号无基线漂移。
以下通过图片与代码进一步解说:</