nf_train1.m

bp网络训练源代码,可用于参数识别,模式分类,如损伤识别,故障诊断等

%***Novelty Detection
%***Training nf 
%***Iput File: 1.prepared data file
%           or 2.Abaqus Data file '*.dat'
%***Then, Save trained NN and traing phase data on disk for testing use
%***for data from Abaqus data file,currently,only use the first 'inout_node' natural freq.
%clear;clc;
%=====below can be changed by user====
  alpha=3.0;
  inout_node=12; %input layer and output layer nodes number of NN
  hide_node1=6;  %nodes number of the first hiden layer of NN
  hide_node2=6;  %nodes number of the second hiden layer of NN
  structure='12-6-6-12'; %only for recording nn structure in .m file
  train_epochs =25;
  train_goal = 0.00001;
  noise=0/100;
  data_from=1;              
  data_file='DF_MS500.txt';  %%when data_from=1:get data from prepared data file
  %data_file='ksm1MS.dat'; %%when data_from=2:get data from Abaqus Data file '*.dat'
  nf_name='nftest_ksm1MS';  %mat filename of nf trained and data related for saveing on disk
  sample_train=500;
  sample_test=500;
%=====above can be changed by user====
lamda_train=zeros(1,sample_train);
lamda_test=zeros(1,sample_test);
if data_from==1 %1.get data from prepared data file
   fid=fopen(data_file,'r'); 
   f=fscanf(fid,'%g',[inout_node,sample_train]);
   %read in order of column by column,
   %element number in a column is "inout_node"
   %total number of columns is "sample_train" 
   fclose(fid);
end  %1.get data from prepared data file
if data_from==2
   %***read freq from abaqus.dat file and then add noise to form freq set
   fin=fopen(data_file,'r');
   number=-1;
   while number<=0
     line=fgetl(fin);
     matches=findstr(line,'CYCLES');
     number=length(matches);
   end
   line=fgetl(fin);
   line;
   for i=1:1:inout_node
     mode=fscanf(fin,'%d',1);
     x=fscanf(fin,'%g', 2);
     freq(i)=fscanf(fin,'%g', 1);
     z=fscanf(fin,'%g', 2);
   end 
   fclose(fin);
   freq=freq'; %change row to column
   %add noise to freq to form data set
   for i=1:1:inout_node
     ranf=sprandn([1:sample_train]); maxf=max(abs(ranf));
     ranf=ranf/maxf;
     for j=1:1:sample_train
        f(i,j)=freq(i)+noise*ranf(j)*freq(i);
     end
   end
   %******f is inout_node rows and samlp_train columns***********************************   
end %data_from=2:setup train data f from Abaqus data file

%=====training New NN=============
for i=1:1:inout_node
    m=mean(f(i,:));
    y(i,:)=(f(i,:)-m)*alpha+m;
end
for i=1:1:inout_node
    f_range(i,1)=min(f(i,:));
    f_range(i,2)=max(f(i,:));
end
novelty_filter = newff(f_range,[inout_node hide_node1 hide_node2 inout_node],{'purelin','purelin','tansig' 'purelin'},'trainlm');
novelty_filter.trainParam.epochs = train_epochs;
novelty_filter.trainParam.goal = train_goal;
novelty_filter = train(novelty_filter,f,y);
%=================================
%=====feed f into trained NN again 
%     to generate lamda in training phase=============
y1 = sim(novelty_filter,f);
for j=1:1:sample_train
   sum=0.0;
   for i=1:1:inout_node
      sum=sum+(y1(i,j)-y(i,j))^2.0;
   end
   lamda_train(j)=sqrt(sum);
end
threshold=mean(lamda_train)+4*std(lamda_train);
%*********testing, here use f too means undamage case*******
ft=f;
for i=1:1:inout_node
   mt=mean(ft(i,:));
   yt(i,:)=(ft(i,:)-mt)*alpha+mt;
end
yt1=sim(novelty_filter,ft);

for j=1:1:sample_test
   sum=0.0;
   for i=1:1:inout_node
      sum=sum+(yt1(i,j)-yt(i,j))^2.0;
   end
   lamda_test(j)=sqrt(sum);
end
%*****************************************************
lamda=[lamda_train,lamda_test];

figure(2);
plot(lamda);
hold on;
plot([1,sample_train+sample_test],[threshold threshold],'-.k');
xlabel('Training / testing data');
ylabel('Novelty index');
hold off;

%***save trained network and all data for later use
save(nf_name,'novelty_filter','structure','data_file','f','alpha','noise','train_epochs','train_goal','sample_train','inout_node');
%this information can be shown after loading .mat file saved
%some data saved here are used by nf_test