MATLAB BP神经网络两层全连接 找了好多遍不知道哪有问题 求DS解答

Question

% 批量方式训练BP网络，实现病症识别
%% 清理
clear all 
clc 
%% 读入数据
xlsfile='train_dataset_washed.xls';
[~,label] = xlsread(xlsfile,'DataSet','J1:J541');
[data,~] = xlsread(xlsfile,'DataSet','A1:I541');
l = zeros(size(label));
for i = 1:length(l)
    if label{i} == 'C'
        l(i,:) = 1;
    end
    if label{i} == 'D'
        l(i,:) = 1;
    end
end
label = l;
%% 划分数据
%正常与不正常各取81个,352个进行训练
TRAIN_NUM_N = 81;
TRAIN_NUM_U = 352;
%正常与不正常分开
n_data = data(label == 1,:);
u_data = data(label == 0,:);

NUM_N = length(n_data);%正常的个数
%正常
r = randperm(NUM_N);
traind(1:TRAIN_NUM_N,:) = n_data(r(1:TRAIN_NUM_N),:);
testd(1:NUM_N - TRAIN_NUM_N,:) = n_data(r(TRAIN_NUM_N + 1:NUM_N),:);

NUM_U = length(u_data);%不正常的个数
%不正常
r = randperm(NUM_U);
traind(TRAIN_NUM_N + 1:TRAIN_NUM_U + TRAIN_NUM_N,:) = u_data(r(1:TRAIN_NUM_U),:);
testd(NUM_N - TRAIN_NUM_N + 1:NUM_N - TRAIN_NUM_N + NUM_U - TRAIN_NUM_U,:) = u_data(r(TRAIN_NUM_U + 1:NUM_U),:);

% 赋值
trainl=zeros(1,TRAIN_NUM_N+TRAIN_NUM_U);
trainl(1:TRAIN_NUM_N)=1;

testl=zeros(1,NUM_N+NUM_U-TRAIN_NUM_N-TRAIN_NUM_U);
testl(1:NUM_N-TRAIN_NUM_N)=1;
%% 设置参数
rng('default')
rng(0)
nTrainNum = TRAIN_NUM_N + TRAIN_NUM_U; % 433个训练样本
%% 构造网络
net.nIn = 9;
net.nHidden1 = 12;    % 第一个隐藏层有?个神经元
net.nHidden2 = 4;    % 第二个隐藏层有?个神经元
net.nOut = 1;         % 1x个输出层节点

w1 = rand(net.nHidden1,net.nIn);  
b1 = rand(net.nHidden1,1); 
net.w1 = [w1,b1]; 

w2 = rand(net.nHidden2,net.nHidden1);  
b2 = rand(net.nHidden2,1); 
net.w2 = [w2,b2]; 

W = rand(net.nOut,net.nHidden2); 
B = rand(net.nOut,1); 
net.w3 = [W,B]; 

%% 训练数据归一化
[traind_s,ps] = mapminmax(traind',0,1);

%% 训练
SampInEx = [traind_s;ones(1,nTrainNum)]; 
expectedOut = trainl;

eb = 0.01;                   % 误差容限 
eta = 0.01;                   % 学习率 
mc = 0.01;                    % 动量因子 
maxiter = 1000;               % 最大迭代次数 
iteration = 0;               % 第一代

errRec = zeros(1,maxiter); 
outRec = zeros(nTrainNum, maxiter); 
NET=[]; % 记录
% 开始迭代
for i = 1 : maxiter 
    hid1_input = net.w1 * SampInEx;     % 第一层隐含层的输入
    hid1_out = logsig(hid1_input);       % 第一层隐含层的输出 
    
    hid2_input = [hid1_out;ones(1,nTrainNum)];
    hid2_input2 = net.w2 * hid2_input;     % 第二层隐含层的输入
    hid2_out = logsig(hid2_input2);       % 第二层隐含层的输出     
    
    ou_input1 = [hid2_out;ones(1,nTrainNum)];   % 输出层的输入
    ou_input2 = net.w3 * ou_input1;
    out_out = logsig(ou_input2);                  % 输出层的输出 
    
    outRec(:,i) = out_out';                       % 记录每次迭代的输出
    err = expectedOut - out_out;                  % 误差
    sse = sumsqr(err);       
    errRec(i) = sse;                              % 保存误差值  
    fprintf('第 %d 次迭代   误差:  %f\n', i, sse);
    iteration = iteration + 1;   
    % 判断是否收敛
    if sse <= eb
        break;
    end     
    % 误差反向传播
    % 第二层隐含层与输出层之间的局部梯度
    % DELTA = err;
     DELTA = err.*dlogsig(ou_input2,out_out);
    % 第一层隐含层与第二层隐含层之间的局部梯度
    delta1 = net.w3(:,1:end-1)'*DELTA.*dlogsig(hid2_input2,hid2_out);
    % 输入层与第一层隐含层之间的局部梯度
    delta2 = net.w2(:,1:end-1)'* delta1.*dlogsig(hid1_input,hid1_out);
     
    % 权值修改量
    dWEX = DELTA * ou_input1'; 
    dwex2 = delta1 * hid2_input'; 
    dwex1 = delta2 * SampInEx'; 
     
    %  修改权值，如果不是第一次修改，则使用动量因子 
    if i == 1 
        net.w3 = net.w3 + eta * dWEX; 
        net.w2 = net.w2 + eta * dwex2; 
        net.w1 = net.w1 + eta * dwex1;
    else    
        net.w3 = net.w3 + (1 - mc) * eta * dWEX + mc * dWEXOld; 
        net.w2 = net.w2 + (1 - mc) * eta * dwex2 + mc * dwex2Old; 
        net.w1 = net.w1 + (1 - mc) * eta * dwex1 + mc * dwex1Old; 
    end 
    % 记录上一次的权值修改量 
    dWEXOld = dWEX; 
    dwex1Old = dwex1; 
    dwex2Old = dwex2;
end     

Answer 1

训练得到的误差很大，改变隐藏层神经元数量，会导致把另1类完全判断成0类。数据集没有问题，已经通过matlab自带的工具箱达到90%以上的正确率。