BP/RBF神经网络仿真框图

function [sys,x0,str,ts]=nnrbf_pid(t,x,u,flag,T,nn,K_pid,... eta_pid,xite,alfa,beta0,w0) switch flag, case 0, [sys,x0,str,ts] = mdlInitializeSizes(T,nn); case 2, sys = mdlUpdates(u); case 3, sys = mdlOutputs(t,x,u,T,nn,K_pid,eta_pid,... xite,alfa,beta0,w0); case {1, 4, 9}, sys = []; otherwise, error(['Unhandled flag = ',num2str(flag)]); end % 初始化函数 function [sys,x0,str,ts] = mdlInitializeSizes(T,nn) sizes = simsizes; % 读入模板，得出默认的控制量 sizes.NumContStates = 0; sizes.NumDiscStates = 3; sizes.NumOutputs = 4+5*nn; sizes.NumInputs = 9+15*nn; sizes.DirFeedthrough = 1; sizes.NumSampleTimes = 1; sys=simsizes(sizes); x0=zeros(3,1); str=[]; ts=[T 0];%初始化x为[0;0;0] % 离散状态变量更新函数 function sys = mdlUpdates(u) sys=[u(1)-u(2); u(1); u(1)+u(3)-2*u(2)];%u(1)、u(2)、u(3)分别为e(k)、e(k-1)、e(k-2) % 输出量计算函数 function sys = mdlOutputs(t,x,u,T,nn,K_pid,eta_pid,... xite,alfa,beta0,w0) ci_3=reshape(u(7: 6+3*nn),3,nn); ci_2=reshape(u(7+5*nn: 6+8*nn),3,nn);%RBF神经网络的输入为du,y(k),y(k-1);隐含层为nn，辨识网络输出为ym(k). ci_1=reshape(u(7+10*nn: 6+13*nn),3,nn); bi_3=u(7+3*nn: 6+4*nn); bi_2=u(7+8*nn: 6+9*nn); bi_1=u(7+13*nn: 6+14*nn); w_3= u(7+4*nn: 6+5*nn); w_2= u(7+9*nn: 6+10*nn); w_1= u(7+14*nn: 6+15*nn); xx=u([6;4;5]);%u(6)、u(4)、u(5)分别为du,y(k),y(k-1) if t==0 ci_1=w0(1)*ones(3,nn); bi_1=w0(2)*ones(nn,1); w_1=w0(3)*ones(nn,1); K_pid0=K_pid; else, K_pid0=u(end-2:end); end for j=1: nn h(j,1)=exp(-norm(xx-ci_1(:,j))^2/(2*bi_1(j)*bi_1(j))); end dym=u(4)-w_1'*h; w=w_1+xite*dym*h+alfa*(w_1-w_2)+beta0*(w_2-w_3); for j=1:nn d_bi(j,1)=xite*dym*w_1(j)*h(j)*(bi_1(j)^(-3))*norm(xx-ci_1(:,j))^2; d_ci(:,j)=xite*dym*w_1(j)*h(j)*(xx-ci_1(:,j))*(bi_1(j)^(-2)); end bi=bi_1+d_bi+alfa*(bi_1-bi_2)+beta0*(bi_2-bi_3); ci=ci_1+d_ci+alfa*(ci_1-ci_2)+beta0*(ci_2-ci_3); dJac=sum(w.*h.*(-xx(1)+ci(1,:)')./bi.^2); % Jacobian KK=K_pid0+u(1)*dJac*eta_pid.*x; sys=[u(6)+KK'*x; KK; ci(:); bi(:); w(:)];%KK为更新的PID参数，x为[e(k)-e(k-1); e(k); e(k)+e(k-3)-2*e(k-2)]