CN103472013A - Visible-near infrared spectrum PLS-DA modeling method combining Adaboost algorithm - Google Patents

Abstract

The invention belongs to the field of visible-near infrared spectroscopic analysis technology, and more specifically relates to a visible-near infrared spectrum PLS-DA modeling method combining Adaboost algorithm. The visible-near infrared spectrum PLS-DA modeling method is characterized in that: a classifier with high performances is obtained by integration of a plurality of PLS-DA models through Adaboost algorithm. PLS-DA models are common identification models used in infrared spectrum technology. PLS-DA models are not capable of reflecting the non-linear relationship between the visible-near infrared spectrum and samples to be analyzed effectively, so that the accuracy of PLS-DA models on data with high nonlinearity is reduced. Adaboost algorithm is capable of providing a framework, and a plurality of methods can be used for construction of sub-classifiers. The high-performance classifier, which is capable of realizing accurate classification of the data with high nonlinearity, is obtained by using PLS-DA models as the sub-classifiers, and combining Adaboost algorithm, so that popularized application of PLS-DA models in identification of the data with high nonlinearity is realized.

Description

A kind of PLS-DA of the visible-near-infrared spectrum in conjunction with Adaboost algorithm modeling method

Technical field

The invention belongs to visible-near-infrared spectrum identification field, a kind of data processing method that can promote visible-near-infrared spectrum partial least squares discriminant analysis (Partial least squares dis-criminationanalysis, PLS-DA) modeling effect specifically.

Technical background

In the multivariable visible-near-infrared spectrum data of small sample, the PLS-DA model can well solve variable collinearity problem and the dimension disaster that other modeling method runs into, and therefore in infrared spectrum identification, has obtained using widely.But the PLS-DA model, as a kind of linear model, can not effectively reflect the nonlinear relationship between near infrared spectrum and analyzing samples classification, therefore to non-linear stronger data, the accuracy of PLS-DA model can descend.

Self-adaptive enhancement algorithm (Adaboost) is a kind of algorithm of applying under various classification scenes that is suitable for.The core concept of Adaboost algorithm is to train different sorter (Weak Classifier) for same training set, then these Weak Classifiers is gathered, and forms a stronger final sorter (strong classifier).Adaboost has many good qualities, and as the Adaboost algorithm provides framework, can make the structure sub-classifier that ins all sorts of ways; Building method is simple; Not there will be over-fitting etc.

Summary of the invention

The PLS-DA model is owing to can not effectively reflecting that the nonlinear relationship between visible-near-infrared spectrum data and sample class label causes the accuracy of model to be limited by the nonlinear degree of visible-near-infrared spectrum data.

The present invention is directed to this problem and proposed a kind of PLS-DA of the visible-near-infrared spectrum in conjunction with Adaboost algorithm modeling method, the PLS-DA model is generalized in the identification application of non-linear stronger visible-near-infrared spectrum data.

The present invention adopts following technical scheme to realize: a kind of PLS-DA of the visible-near-infrared spectrum in conjunction with Adaboost algorithm modeling method comprises the steps:

Step 1, given training sample, S={ (x ₁, y ₁) ..., (x _m, y _m), wherein, x _i∈ X, label y _i∈ Y={1,2,3 ..., N}, m means number of training, N means the classification number of training sample;

Step 2, the weight coefficient ω of each sample of initialization _i=1/m, i=1 ..., m;

Step 3, the t=1 that circulates each time ..., T does following steps;

Step 3.1, used partial least square method to carry out modeling to the training sample that weight distribution is arranged, and obtains a PLS-DA model h _t;

Step 3.2, calculate h _ttraining error

symbol in algorithm " [] " is defined as follows: for logical expression e, as crossed e, be true, and [e]=1, otherwise [e]=0;

Step 3.3, if ε _tset T=t-1 for>1/2 and then skip to step 4;

Step 3.4, make β _t=ε _t/ (1-ε _t), α _t=1n (1/ β _t);

Step 3.5, upgrade the sample weights coefficient

${\mathrm{\ω}}_i^{t+1}=\frac{{\mathrm{\ω}}_i^t}{Z_t}\×{\mathrm{\β}}_t^{1-[h_t\left(x_i\right)\≠y_i]}$

Wherein, Z _tfor normalization coefficient, can make

Step 4, the output strong classifier: $H\left(x\right)=\arg \underset{y\∈Y}{\max }\left(\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_t\right[h_t\left(x\right)=y\left]\right);$

Step 5, test to the classification accuracy of H (x).

The present invention uses PLS-DA as sub-classifier, in conjunction with the Adaboost algorithm, has obtained the strong classifier that also can accurately classify to non-linear stronger data, thereby the PLS-DA model can be generalized in the identification application of non-linear stronger data.

The accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention.

Embodiment

Visible-near-infrared spectrum identification example below in conjunction with a laver is further described technical scheme of the present invention.

Embodiment: a kind of PLS-DA of the visible-near-infrared spectrum in conjunction with Adaboost algorithm modeling method, the identification for laver has following steps:

Step 1, the division of sample.One has four class lavers, is respectively laver a, laver b, laver c, laver d.The row matrix that the ir data of each laver sample is 1 * 1735.Every class laver sample number is 30.Every class laver sample is random is divided into two parts: 20 foundation for model, and 10 tests for model, the matrix consisted of first's sample is training set X, the matrix consisted of the second portion sample is checking collection V.The class label value is set.Because one has 4 kinds, thus laver a, laver b, laver c, the class label of laver d is set to respectively 1,2,3,4.Form matrix Y by 4 kind label values.Be training sample set S={ (x ₁, y ₁) ..., (x ₈₀, y ₈₀), x wherein _i∈ X, y _i∈ Y={1,2,3,4}.

Step 2, the weight coefficient ω of each sample in the initialization training set ¹ _i=1/80, i=1 ..., 80.Maximum iteration time T is set, and in this example, T is set to 300.

Step 3, the t=1 that circulates each time ..., T does following steps:

Step 3.1, the weight coefficient of each training sample represents the probability that this sample is selected.This example is used roulette method to select 80 training samples, and the training sample of selecting is set up to the PLS-DA model: the visible-near-infrared spectrum of 80 training samples forms dependent variable matrix X _t, corresponding class label value forms dependent variable matrix Y _t, set up regression equation X _t* b _t=Y _t, use partial least square method to try to achieve regression coefficient

obtain a sub-classifier h _t:

round values after rounding up.

Step 3.2, calculate h _ttraining error

Step 3.3, detect ε _tvalue.If ε _tset T=t-1 for>1/2 and then skip to step 4.

Step 3.4, make β _t=ε _t/ (1-ε t), α _t=ln (1/ β _t).

Step 3.5, upgrade the sample weights coefficient

${\mathrm{\ω}}_i^{t+1}=\frac{{\mathrm{\ω}}_i^t}{Z_t}\×{\mathrm{\β}}_t^{1-[h_t\left(x_i\right)\≠y_i]}$

Wherein, Z _tfor normalization coefficient, can make

Step 4, the output strong classifier: $H\left(x\right)=\arg \underset{y\∈Y}{\max }\left(\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_t\right[h_t\left(x\right)=y\left]\right).$

Step 5, tested the classification accuracy of H (x) with the sample of test set V, if higher than the accuracy of single PLS-DA model, accepts H (x) model, otherwise do not accept H (x) model.The accuracy rate of H in this example (x) classification is 95%, and the accuracy of single PLS-DA category of model is 65%.The accuracy of H (x) model is higher than the accuracy of single PLS-DA model, accepts H (x) model.

Claims (1)

1. the PLS-DA of the visible-near-infrared spectrum in conjunction with an Adaboost algorithm modeling method, its feature comprises following steps:

Step 1, given training sample, S={ (x ₁, y ₁) ..., (x _m, y _m), wherein, x _i∈ X, label y _i∈ Y={1,2,3 ..., N}, m means number of training, N means the classification number of training sample;

Step 2, the weight coefficient ω of each sample of initialization _i=1/m, i=1 ..., m;

Step 3, the t=1 that circulates each time ..., T does following steps;

Step 3.1, used partial least square method to carry out modeling to the training sample that weight distribution is arranged, and obtains a PLS-DA model h _t;

Step 3.2, calculate h _ttraining error

symbol in algorithm " [] " is defined as follows: for logical expression e, as crossed e, be true, and [e]=1, otherwise [e]=0;

Step 3.3, if ε _tset T=t-1 for>1/2 and then skip to step 4;

Step 3.4, make β _t=ε _t/ (1-ε _t), α _t=1n (1/ β _t);

Step 3.5, upgrade the sample weights coefficient

${\mathrm{\ω}}_i^{t+1}=\frac{{\mathrm{\ω}}_i^t}{Z_t}\×{\mathrm{\β}}_t^{1-[h_t\left(x_i\right)\≠y_i]}$

Wherein, Z _tfor normalization coefficient, can make

Step 4, the output strong classifier: $H\left(x\right)=\arg \underset{y\∈Y}{\max }\left(\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_t\right[h_t\left(x\right)=y\left]\right);$

Step 5, test to the classification accuracy of H (x).

Images