TA的每日心情 | 开心 2019-11-20 15:05 |
|---|
签到天数: 2 天 [LV.1]初来乍到
|
EDA365欢迎您登录!
您需要 登录 才可以下载或查看,没有帐号?注册
x
本帖最后由 Colbie 于 2020-3-18 09:58 编辑
; O% e4 v; r' z7 |+ N3 P+ ~3 T# z/ h7 p
核主元分析(Kernel principal component analysis ,KPCA)在降维、特征提取以及故障检测中的应用。主要功能有:(1)训练数据和测试数据的非线性主元提取(降维、特征提取)
5 p+ O7 x' X8 X Q" d7 H( m8 ~) v(2)SPE和T2统计量及其控制限的计算
) X$ L X7 g$ k: C(3)故障检测
' M' a5 k0 p! S: W/ C1 d! s: ]4 i+ g+ @% o' V
参考文献:
2 \0 q; W* |, D9 h4 MLee J M, Yoo C K, Choi S W, et al. Nonlinear process monitoring using kernel principal component analysis[J]. Chemical engineering science, 2004, 59(1) : 223-234.
9 z' y5 V" P4 \ H7 w' p* }1 R8 P. V1 b' b- G
1. KPCA的建模过程(故障检测):
2 g! ~4 h/ G2 T: r4 }1 C% R3 [(1)获取训练数据(工业过程数据需要进行标准化处理)' q( Z$ J# Y- L7 e
(2)计算核矩阵
' g R. t' h4 J7 k, p6 Z4 ?/ c(3)核矩阵中心化5 ^- ] o4 ~) W
(4)特征值分解
4 s4 m6 P2 g! ]9 Z) k0 z(5)特征向量的标准化处理
% Q) g9 D8 |2 K* r) E" o(6)主元个数的选取
6 X2 H; \+ o1 K t# ^ d8 m+ F |(7)计算非线性主成分(即降维结果或者特征提取结果)
1 l2 d4 Y& o5 D(8)SPE和T2统计量的控制限计算& Y5 W, ]4 U2 m+ _) r
- function model = kpca_train(X,options)
- % DESCRIPTION
- % Kernel principal component analysis (KPCA)
- %
- % mappedX = kpca_train(X,options)
- %
- % INPUT
- % X Training samples (N*d)
- % N: number of samples
- % d: number of features
- % options Parameters setting
- %
- % OUTPUT
- % model KPCA model
- %
- %
- % Created on 9th November, 2018, by Kepeng Qiu.
- % number of training samples
- L = size(X,1);
- % Compute the kernel matrix
- K = computeKM(X,X,options.sigma);
- % Centralize the kernel matrix
- unit = ones(L,L)/L;
- K_c = K-unit*K-K*unit+unit*K*unit;
- % Solve the eigenvalue problem
- [V,D] = eigs(K_c/L);
- lambda = diag(D);
- % Normalize the eigenvalue
- V_s = V ./ sqrt(L*lambda)';
- % Compute the numbers of principal component
- % Extract the nonlinear component
- if options.type == 1 % fault detection
- dims = find(cumsum(lambda/sum(lambda)) >= 0.85,1, 'first');
- else
- dims = options.dims;
- end
- mappedX = K_c* V_s(:,1:dims) ;
- % Store the results
- model.mappedX = mappedX ;
- model.V_s = V_s;
- model.lambda = lambda;
- model.K_c = K_c;
- model.L = L;
- model.dims = dims;
- model.X = X;
- model.K = K;
- model.unit = unit;
- model.sigma = options.sigma;
- % Compute the threshold
- model.beta = options.beta;% corresponding probabilities
- [SPE_limit,T2_limit] = comtupeLimit(model);
- model.SPE_limit = SPE_limit;
- model.T2_limit = T2_limit;
- end5 o* I6 Q6 J( o
! ]/ H: P; R, m% `6 o! B6 h. i" z: @! b! F3 \0 T G
2. KPCA的测试过程:
1 Q- b8 ~0 \6 B, {7 _9 }(1)获取测试数据(工业过程数据需要利用训练数据的均值和标准差进行标准化处理)
$ i* ?3 k" C# M2 A! ?# u(2)计算核矩阵
0 n6 H$ g# u0 Q; G) ~(3)核矩阵中心化, ~/ |" ]% R8 x7 B6 m( H
(4)计算非线性主成分(即降维结果或者特征提取结果)& B) V" u C( K2 J2 Z! M \; h9 `1 M
(5)SPE和T2统计量的计算
1 U% o. T1 o4 b! E1 l$ }8 z2 q. n: D- function [SPE,T2,mappedY] = kpca_test(model,Y)
- % DESCRIPTION
- % Compute the T2 statistic, SPE statistic,and the nonlinear component of Y
- %
- % [SPE,T2,mappedY] = kpca_test(model,Y)
- %
- % INPUT
- % model KPCA model
- % Y test data
- %
- % OUTPUT
- % SPE the SPE statistic
- % T2 the T2 statistic
- % mappedY the nonlinear component of Y
- %
- % Created on 9th November, 2018, by Kepeng Qiu.
- % Compute Hotelling's T2 statistic
- % T2 = diag(model.mappedX/diag(model.lambda(1:model.dims))*model.mappedX');
- % the number of test samples
- L = size(Y,1);
- % Compute the kernel matrix
- Kt = computeKM(Y,model.X,model.sigma );
- % Centralize the kernel matrix
- unit = ones(L,model.L)/model.L;
- Kt_c = Kt-unit*model.K-Kt*model.unit+unit*model.K*model.unit;
- % Extract the nonlinear component
- mappedY = Kt_c*model.V_s(:,1:model.dims);
- % Compute Hotelling's T2 statistic
- T2 = diag(mappedY/diag(model.lambda(1:model.dims))*mappedY');
- % Compute the squared prediction error (SPE)
- SPE = sum((Kt_c*model.V_s).^2,2)-sum(mappedY.^2 ,2);
- end
6 _' P9 ~! {; x) K
. N8 Q9 B& z* l: J4 ^6 y4 J6 r0 m! p2 O) p: Q5 I" Y
) X3 R$ G8 Y( C5 s$ \' ^
3. demo1: 降维、特征提取" m& ?; _2 B7 u# F. }
(1) 源代码
2 C7 T8 ~; v: T- % Demo1: dimensionality reduction or feature extraction
- % ---------------------------------------------------------------------%
- clc
- clear all
- close all
- addpath(genpath(pwd))
- % 4 circles
- load circledata
- %
- X = circledata;
- for i = 1:4
- scatter(X(1+250*(i-1):250*i,1),X(1+250*(i-1):250*i,2))
- hold on
- end
- % Parameters setting
- options.sigma = 5; % kernel width
- options.dims = 2; % output dimension
- options.type = 0; % 0:dimensionality reduction or feature extraction
- % 1:fault detection
- options.beta = 0.9; % corresponding probabilities (for ault detection)
- options.cpc = 0.85; % Principal contribution rate (for ault detection)
- % Train KPCA model
- model = kpca_train(X,options);
- figure
- for i = 1:4
- scatter(model.mappedX(1+250*(i-1):250*i,1), ...
- model.mappedX(1+250*(i-1):250*i,2))
- hold on
- end0 O- K8 d0 S6 [% ?$ K6 K) }
- P* R" y, V7 N; ^; \2 o
/ F1 h" e3 j" x5 A- [; @0 C2 j(2)结果 (分别为原图和特征提取后的图)
. h0 L/ q0 s5 f0 O) c8 a. j* l& l$ M
" W# B! z% i- k- Q# Y4 b9 x6 Y
4. demo2: 故障检测(需要调节核宽度、主元贡献率和置信度等参数来提高故障检测效果)$ a% X, Y4 m+ h: N/ Q0 p$ S2 B9 d
(1)源代码
- |7 B2 H( k# \5 a2 h" g- % Demo2: Fault detection
- % X: training samples
- % Y: test samples
- % Improve the peRFormance of fault detection by adjusting parameters
- % 1. options.sigma = 16; % kernel width
- % 2. options.beta % corresponding probabilities
- % 3. options.cpc ; % principal contribution rate
- % ---------------------------------------------------------------------%
- clc
- clear all
- close all
- addpath(genpath(pwd))
- %
- X = rand(200,10);
- Y = rand(100,10);
- Y(20:40,: ) = rand(21,10)+3;
- Y(60:80,: ) = rand(21,10)*3;
- % Normalization (if necessary)
- % mu = mean(X);
- % st = std(X);
- % X = zscore(X);
- % Y = bsxfun(@rdivide,bsxfun(@minus,Y,mu),st);
- % Parameters setting
- options.sigma = 16; % kernel width
- options.dims = 2; % output dimension
- options.type = 1; % 0:dimensionality reduction or feature extraction
- % 1:fault detection
- options.beta = 0.9; % corresponding probabilities (for ault detection)
- options.cpc = 0.85; % principal contribution rate (for ault detection)
- % Train KPCA model
- model = kpca_train(X,options);
- % Test a new sample Y (vector of matrix)
- [SPE,T2,mappedY] = kpca_test(model,Y);
- % Plot the result
- plotResult(model.SPE_limit,SPE);
- plotResult(model.T2_limit,T2);; |! W2 H( ^) j+ F
& n+ t+ s0 s U* u
/ w* X* o' ^, t- T
(2)结果(分别是SPE统计量和T2统计量的结果图)
( [5 h+ w7 P2 l' S" c$ x ?) y6 j, ?9 I
" q1 @9 g, J1 P# S) C附件是基于KPCA的降维、特征提取和故障检测程序源代码。如有错误的地方请指出,谢谢。+ d$ }6 Q s7 j. K; n
2 J3 a+ Q: @+ ?. y8 r2 l
|
|