waveletnn_time_feature_morlet_4f2pog16q.m

小波神经网络的分类识别

%% this procedure is a simulation of wavelet artificial neural network (WNN) for modulated signal classification
%% the weights of hidden layer are wavelet functions and the mother wavelet is 'morlet'function.

clear all;

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%  define the necessary parameters for WNN
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
global node_in node_media node_out  alfa belta w u o layer_j shift_k;     % % node_in:represent 'I' in the paper,node_media:'J',node_out:'K',alfa:the rate of step change

% % 输入神经网络的学习参数和各层节点数
% % 输入节点数node_in,隐层节点数node_media,输出节点数node_out
alfa=input('请输入网络学习率因子 alfa=')
belta=input('请输入网络动量项因子 belta=')
layer_j=input('请输入网络多分辨级数 layer_j=')    % % 首次值可以选3
shift_k=input('请输入网络平移参数 shift_k=')      % % 首次值可以选1
node_in=5;
node_media=(2*layer_j+1)*(2*shift_k+1);
node_out=3;
% node_out=5;
n=5                                         % % 神经网络识别模式总数 

%% 初始化网络
K=10;
w=rand(node_out,node_media)/K;     % % w represents the weights between media layer and out layer and w(i,j) stands for the weight value between ith out-layer and jth media-layer:(-0.5,0.5)/K
u=rand(node_media,node_in)/K;

w1=w;
u1=u;
% load('e:\lcy\wnn-network\dwnn-recognize\morning-1-1.mat','w1')
% load('e:\lcy\wnn-network\dwnn-recognize\morning-1-1.mat','u1')
% 
% w=w1;
% u=u1;

o=rand(1,node_out)/K;                     % % o stores the values of WNN output layer

% % 设定网络训练参数
sse_draw=0;                                 % % 全局误差向量初设为零
train_max=2000;                             % % 设定最大训练次数为2500
qwsse=0.001;                                % % 期望误差为0.001,则每一个模式的误差必须小于0.01,否则全局误差会大于0.001

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% simulating data for learning
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% load('d:\lcy\AMR\design-programme-regulator\DWNN\data\learn-data-5-4f2pog16qam-20av.mat','ppp1')
load('d:\lcy\AMR\design-programme-regulator\DWNN\data\learn-data-5-4f2pog16qam-20av.mat','ppp1')
% load('d:\lcy\AMR\design-programme-regulator\DWNN\data\filter_8db_learn_4f2pog16qam_unav_1.mat','ppp1')
% load('e:\work_files\design-programme-regulator\DWNN\data\4db_learn_4f2pog16q_20av_4.mat','ppp1')
% load('d:\lcy\AMR\design-programme-regulator\DWNN\data\4db_learn_4f2pog16q_20av_4.mat','ppp1')

low=[0 0 0.1 0.05 0.6];                      % % 用线性方法归一化 SNR=8dB  
up=[35 2.5 0.45 1.2 6];

% low=[0 0.1 0.1 0.1 1.1];                      % % 用线性方法归一化 snr=5db
% up=[36 2.3 0.45 1.4 6.8];

% low=[0 0 0.1 0 0]';                      % % 用线性方法归一化  snr=15db
% up=[55 2.2 0.45 1.2 5.1]';

% low=[0 0.3 0.1 0.2 1.6];                      % % 噪声不加带限滤波时特征范围,用线性方法归一化 SNR=8dB
% up=[25 2.9 0.35 1.25 7.8];

t=up-low;
for i=1:n
    ppp1(i,:)=ppp1(i,:)./t+1-up./t;    
end
% d=ppp1;
% for i=1:n                                 % % 用除以各分量和的方法归一化
%     for j=1:node_in
%         ppp1(i,j)=ppp1(i,j)/(sum(d(i,:))); 
%     end
% end

% for i=1:n
%     ppp1(i,:)=ppp1(i,:)/max(d(i,:));
% end

qw_o=[0 0 1                                            % % 各种调制模式的期望输出值
      0 1 0
      0 1 1
      1 0 0
      1 0 1];
% qw_o=[0 0 0 0 1                                            % % 各种调制模式的期望输出值
%       0 0 0 1 0
%       0 0 1 0 0
%       0 1 0 0 0
%       1 0 0 0 0];
  
for train=1:train_max    
    sse=0;                                            % %　全局误差清零
    gw_l=zeros(node_media,node_out);                  % % gw_l,gb_l,ga_l 分别存放的是此次训练的全局误差对w,b,a的偏导数
    gu_l=zeros(node_media,node_in);
       
  for i=1:n                                           % %　每次训练交叉输入n个模式特征向量　          
        out=wnn_implementation(ppp1(i,:));            % % 计算对应的神经网络分类器实际输出  
        
        sse=sse+0.5*(qw_o(i,:)-o)*(qw_o(i,:)-o)';  % % 计算批训练的全局误差
        
        for t1=1:node_media                        % %  计算当前训练全局误差的偏导数 
            
            if mod(t1,2*shift_k+1)==0
                j1=t1/(2*shift_k+1)-1-layer_j;
                k1=shift_k;
            else 
                j1=floor(t1/(2*shift_k+1))-layer_j;
                k1=mod(t1,2*shift_k+1)-shift_k-1;
            end
            
            gw_l(t1,:)=-(qw_o(i,:)-o)*wavelet_function(ppp1(i,:)*u(t1,:)',j1,k1);       
            gu_l(t1,:)=-(sum((qw_o(i,:)-o).*w(:,t1)'))*(2^(j1/2))*wavelet_derivative(ppp1(i,:)*u(t1,:)',j1,k1)*ppp1(i,:);
       end
       if train==1                            
           sw_l=-alfa*gw_l;
           su_l=-alfa*gu_l;
       else 
              sw_l=-alfa*gw_l+belta*sw_l;
              su_l=-alfa*gu_l+belta*su_l;
       end
       
       w=w+sw_l';                                                     % % 调整网络权值
       u=u+su_l;
       
       
  end  
    
    sse_draw(train)=sse;                             % % 记录每次训练的全局误差值
    
%    if train>=2                                     % % 自适应调整学习率
%        if sse>sse_draw(train-1) 
%            alfa=alfa*0.2;
%        elseif sse<sse_draw(train-1)
%            alfa=alfa*1.15;
%        end
%    end    
    
  
   if sse<qwsse&&train<train_max                                    % % 如果达到期望误差要求,结束网络训练过程
       disp('the number of learning times train=')
       train
       sse
       break;
   end
   
end

%训练结束
x=1:1:train;                                  % % 画全局误差与训练次数的关系曲线
% plot(x,sse_draw,'r--');
semilogy(x,sse_draw,'r--');
xlabel('训练次数(*5)')
ylabel('全局误差')
title('The Maximum Error is 0.001')
hold on

% disp('输出层权值为')
% w
% u
% % 进行分类器识别性能评估：

% 设每种调制方式400个样本，其特征值向量构成的输入矩阵pppf为一个400行5列的矩阵。

% 各个pppf矩阵分别从data_2ASK_8dB.mat、data_4ASK_8dB.mat、data_2FSK_8dB.mat、data_4FSK_8dB.mat、
% data_2PSK_8dB.mat、data_4PSK_8dB.mat、data_16QAM_8dB.mat文件中加载.

% 每次选一种方式,将相应文件的特征矩阵加载到pppf矩阵:

% 说明：此程序的检测识别成功率的测试集是由400个2ASK信号的特征向量构成的400行5列的矩阵ppp

num_e=zeros(1,n);
out1=zeros(400,node_out,n);
o1=zeros(400,node_out,n);

for p=1:5
   
if p==1
%      load('d:\lcy\AMR\design-programme-regulator\DWNN\data\4db-10-4fsk-400-1.mat','ppp');
      load('d:\lcy\AMR\design-programme-regulator\DWNN\data\data-10-1-4fsk.mat','ppp');
elseif p==2
%      load('d:\work_files\design-programme-regulator\DWNN\data\4db-10-2psk-400-1.mat','ppp');
       load('d:\lcy\AMR\design-programme-regulator\DWNN\data\data-10-1-2psk.mat','ppp');
elseif p==3
%      load('d:\work_files\design-programme-regulator\DWNN\data\4db-10-oqpsk-400-1.mat','ppp');
       load('d:\lcy\AMR\design-programme-regulator\DWNN\data\data-10-1-oqpsk.mat','ppp');
elseif p==4
%      load('d:\work_files\design-programme-regulator\DWNN\data\4db-10-gmsk-400-1.mat','ppp');
     load('d:\lcy\AMR\design-programme-regulator\DWNN\data\data-10-1-gmsk.mat','ppp');
elseif p==5
%      load('d:\work_files\design-programme-regulator\DWNN\data\4db-10-16qam-400-1.mat','ppp');
     load('d:\lcy\AMR\design-programme-regulator\DWNN\data\data-10-1-16qam.mat','ppp');

end

% c=ppp;
for i=1:400
    ppp(i,:)=ppp(i,:)./t+1-up./t;    
end

% for i=1:400
%     ppp(i,:)=ppp(i,:)/max(c(i,:));    
% end

% 
% for i=1:400                                 % 用除以各分量和的方法归一化
%     for j=1:node_in
%         ppp(i,j)=ppp(i,j)/(sum(c(i,:))); 
%     end
% end

% ppp=(ppp2+ppp3)/2;

for i=1:400                                      % % 统计此小波神经网络分类器对2ASK调制制式的识别性能       
    if size(ppp)~=[400,5]
        disp('输入测试样本集有误');
        break;
    end
    out=wnn_implementation(ppp(i,:));
    out1(i,:,p)=out;
    o1(i,:,p)=o;
    if out==qw_o(p,:)
        num_e(p)=num_e(p)+1;
    end
end

disp('此小波神经网络分类器调制制式的正确识别率为:')
num_e/400
end

return;