zhuyuan.m

经过主元分析后的神经网络算法Matlab代码

clc;
f1=fopen('inputx.txt','r');
input=fread(f1,[8,200],'float');
fclose(f1);
f2=fopen('outputy.txt','r');
output=fread(f2,[1,200],'float');
fclose(f2);

[pn,meanp,stdp,tn,meant,stdt]=prestd(input,output);
[ptrans,transmat]=prepca(pn,0.11); 

% 根据拉依达准则，剔除异常数据  
std=blkproc(ptrans,[1 200],'std2');
a=zeros(1,200);b=a;
[M,N]=size(ptrans);
for i=1:M
    for j=1:N
        if abs(ptrans(i,j))>3*std(i,1)
            a(j)=j;
        end
    end
end
Q=0;
for i=1:N
        if a(i)==0
Q=Q+1;
  end
end
PTRANS=zeros(M,Q);TN=zeros(1,Q);j=1;
    for i=1:N
        if a(i)==0
        PTRANS(:,j)=ptrans(:,i);
        TN(1,j)=tn(1,i);j=j+1;
    end
    end

[SAMPLE2,meanp,stdp,TN2,meant,stdt]=prestd(PTRANS, TN); % 归一化
A=floor(2*Q/3); % A训练用样本行
B=Q-A; % B测试用样本行
X=SAMPLE2(:,1:A); % 训练样本
Y=TN2(1,1:A); % 输出
X_test=SAMPLE2(:,1:B); % 测试样本
Y_test=TN2(1,1:B); % 输出

net=newff(minmax(X),[18,1],{'tansig','purelin'},'traingdx');
net.trainParam.epochs=1000;%训练次数
net.trainParam.goal=0.0001;%训练目标0.0001

net=init(net);
net_point=train(net,X,Y);
out_point=sim(net_point,X_test);
er_point=abs(out_point-Y_test);

figure;
plot(1:B,out_point,'k-*',1:B,Y_test,'r-^',1:B,er_point,'b--square');
legend('二线凝固点值训练结果','原始值','误差绝对值');
Merr_point=0.0;
for i=1:B;
Merr_point= Merr_point+er_point(i);
end
err_point= Merr_point/B   %显示均值误差
%计算均方根差
msn_point=0.0;
for i=1:B;
msn_point=msn_point+(er_point(i)- err_point)*(er_point(i)- err_point);
	end
	msn_point= sqrt(msn_point/B)% 显示均方根差
pause;