CN102072767A

CN102072767A - Wavelength similarity consensus regression-based infrared spectrum quantitative analysis method and device

Info

Publication number: CN102072767A
Application number: CN 201010601351
Authority: CN
Inventors: 郝勇
Original assignee: East China Jiaotong University
Current assignee: East China Jiaotong University
Priority date: 2010-12-23
Filing date: 2010-12-23
Publication date: 2011-05-25

Abstract

The invention relates to a wavelength similarity consensus regression-based infrared spectrum quantitative analysis method and a wavelength similarity consensus regression-based infrared spectrum quantitative analysis device. The device comprises a spectrometer, a preprocessor, a wavelength filter and a partial least-squares regression analyzer which are connected through a data signal line; and the wavelength of the near-infrared spectrum ranges from 780 to 50,000nm. By performing cluster analysis on spectra along the wavelength direction, a spectrum is divided into different information blocks, multiple models are constructed, and the difficulty for a single model method to extract information of a spectrum signal is overcome; and by singly determining factor number for different submodels, effective information is fully extracted, and the prediction accuracy and robustness of an infrared spectrum analysis model are improved.

Description

Return the IR spectrum quantitative analysis method and apparatus based on wavelength similarity common recognition

Technical field

The present invention relates to a kind of IR spectrum quantitative analysis method and apparatus, particularly return the IR spectrum quantitative analysis method and apparatus based on wavelength similarity common recognition.

Background technology

Owing to both contained useful information in the spectral signal that spectrometer obtains, other various stochastic errors (background and noise) are also superposeing simultaneously.Therefore, when utilizing partial least-square regression method to carry out quantitative test, be difficult to obtain the higher model of precision of prediction.At this problem of precision of prediction that improves the partial least squares regression model, number of research projects is carried out, mainly comprise preprocessing procedures and variable STUDY ON SCREENING at present, these methods have been successfully used to the elimination of spectral background correction, noise removing, no information variable.Yet when utilizing said method to carry out information extraction, regular meeting face useful information extract insufficient or the information extracted in comprise noisy problem.When utilizing multivariate calibration methods to carry out the quantitative test of infrared spectrum; often face the problem (as regular parameter and nuclear parameter selection etc. in selection of factor number and the least square support vector regression in the partial least squares regression) that parameter is selected; if spectral signal is adopted unified parameter, usually can cause the information extraction of Partial Variable insufficient or introduce the phenomenon of noise information.

Cluster analysis belongs to no supervised recognition method, is usually used in the division of sample ownership.If sample belongs to same class, show that sample has similarity preferably in such.When clustering method was used for the attribution analysis of spectrum similar argument, the spectral variables in the same class included similar spectral information; And the spectral variables between class has been represented different composition informations.

Based on the analysis of above-mentioned two aspects, this paper has invented a kind of IR spectrum quantitative analysis method and apparatus that returns based on similar wavelength cluster common recognition.In this invention, directly utilize the several similar spectral variables zone after the cluster analysis to set up a plurality of recurrence submodels, last antithetical phrase predicted results is carried out integrated, can select different regression parameters at different spectral information classes like this, extract the useful information under each spectral variables class more flexibly effectively.Therefore be the infrared spectrum analysis that a kind of extremely has application prospect based on similar wavelength cluster common recognition homing method.

Summary of the invention

In the application Infrared Spectrum Technology component to be measured is carried out in the express-analysis process, in order to solve the inaccurate problem of withdrawing spectral information.The invention provides a kind of IR spectrum quantitative analysis method and apparatus that returns based on wavelength similarity cluster common recognition.This method and apparatus makes full use of the characteristics of cluster analysis " things of a kind come together, people of a mind fall into the same group ", and the message block that being divided into of the similar spectral variables of infrared spectrum is different has realized making full use of and reasonable distribution of infrared spectrum information.

The device of realizing technique scheme comprises: grow up to be a useful person through spectrometer, pretreater, cluster analysis device, partial least squares regression mode set that data signal line links to each other.

Infrared and near infrared spectrum during described infrared spectrum comprises, promptly wavelength coverage is: 780 nm-50000 nm.

Described pretreater takes centralization and vector normalization that the original signal of spectrometer collection is handled.

Described cluster analysis device, the treatment step that specifically comprises is as follows: at first calculate the Euclidean distance between spectral variables, its computing formula is:

Wherein Be

Wavelength points and Distance between the wavelength points,

With

Be respectively

Wavelength points and

Wavelength points is at spectrum

The absorbance at place.The distance of calculating between newly-built class and other wavelength points or the class adopts the mean distance method.Secondly, carry out cluster analysis and divide different spectral variables message block (block) into according to the distance between the spectral variables of aforementioned calculation.

Described partial least squares regression mode set is grown up to be a useful person, and its concrete operations are as follows: analyze each the spectral variables piece after handling through the cluster analysis device is set up corresponding partial least squares regression submodel.In each submodel, number of principal components can be selected different numerical value.Utilize the submodel of having set up respectively test set sample index to be predicted, predicted the outcome accordingly respectively.Then predicting the outcome of each submodel is weighted, obtains the predicted value of final sample index.By comparing under the different decomposition yardstick, the predicted root mean square error of model (RMSEP) value is determined class number and the model parameter and the preservation of suitable cluster analysis,, be used for the forecast analysis of follow-up fresh sample infrared spectrum.

Because the present invention adopts above technical scheme, obtains following effect:

By spectrum is carried out cluster analysis along wavelength direction, make a spectrum be divided into different message block, realized the structure of multi-model, overcome the difficulty of single model method to the spectral signal information extraction; By different submodels are determined factor number separately, realized the abundant extraction of effective information, improved the precision of prediction and the robustness of Infrared spectroscopy model.

Description of drawings

The IR spectrum quantitative analysis method and apparatus synoptic diagram that Fig. 1 wavelength similarity cluster common recognition returns;

Fig. 2 near infrared light spectrogram;

Fig. 3 wavelength similarity cluster common recognition homing method synoptic diagram;

Fig. 4 cluster analysis divided block number is to the influence of wavelength similarity cluster common recognition homing method precision of prediction

Fig. 5 wavelength similarity cluster synoptic diagram;

The IR spectrum quantitative analysis model predication value that Fig. 6 wavelength similarity cluster common recognition returns and the correlogram of reference value.

Embodiment

Embodiment describes in conjunction with following embodiment.Near infrared spectrum with pears is an example, the pol index of pears inside is carried out the structure of wavelength similarity cluster common recognition regression model.

Fig. 1 is the IR spectrum quantitative analysis method and apparatus synoptic diagram that wavelength similarity cluster common recognition returns, and Fig. 2 is the near infrared light spectrogram, and spectral range is 750 ~ 1800 nm, and every spectrum comprises 1051 data points.All samples is divided into calibration set and test set according to the ratio of 2:1.

All samples spectrum is carried out cluster analysis, in the compute classes and the distance between class adopt Euclidean distance method and average Furthest Neighbor respectively.Figure 3 shows that wavelength similarity cluster common recognition homing method synoptic diagram, as we know from the figure, the principle of this method is to select wavelength points at random, and the wavelength points that comprises in the same spectral information piece after the cluster is not necessarily adjacent.

Figure 4 shows that of the influence (wherein spectral information piece get 1 o'clock, corresponding be the predicted root mean square error of overall optical spectrum model) of cluster analysis divided block number to wavelength similarity cluster common recognition homing method precision of prediction.As we know from the figure, along with the increase of clustering information piece, predicted root mean square error reduces earlier afterwards to increase, and different clustering information pieces are counted predicting the outcome of drag and had notable difference.

Figure 5 shows that wavelength similarity cluster synoptic diagram.From figure can, spectrum is 3 message block along the wavelength direction cluster analysis.Respectively 3 message block are set up the partial least squares regression model, corresponding number of principal components is respectively 3,7,3, and the number of principal components of full spectrum is 4, as seen adopts wavelength similarity cluster common recognition homing method to be more conducive to extract at comprising useful information in the spectrum.

Fig. 6 is the correlogram of infrared spectrum wavelength similarity cluster common recognition model predication value and reference value.As we know from the figure, obtained quantitative result preferably.

Effect of the present invention is: more help to extract the useful information of the lower infrared spectrum in different wave length point zone by adopting wavelength similarity cluster common recognition homing method, obtain more accurate, stable predicting the outcome. Therefore, wavelength similarity cluster common recognition homing method is expected to become the spectroscopic analysis methods that a kind of extremely has application prospect.

Claims

1. return the IR spectrum quantitative analysis method and apparatus based on wavelength similarity common recognition, it is characterized in that: comprise through data signal line continuous spectrometer, pretreater, cluster analysis device, partial least squares regression mode set and growing up to be a useful person;

Infrared and near infrared spectrum during described infrared spectrum comprises, promptly wavelength coverage is: 780 nm-50000 nm;