Method for stabilizing near-infrared models and determining their applicability

• US 5,668,374 A
• Filed: 05/07/1996
• Issued: 09/16/1997
• Est. Priority Date: 05/07/1996
• Status: Expired due to Fees

• Current Assignee: Core Laboratories NV
• Original Assignee: Core Laboratories NV
• Inventors: Ranc, Martha L., DiFoggio, Rocco
• Primary Examiner(s): Porta, David P.
• Assistant Examiner(s): Bruce, David Vernon
• Application Number:
  US08/646,082
• Time in Patent Office: 497 Days
• Field of Search: 250/339.07, 250/339.09, 250/339.12, 250/343, 364/498, 364/502
• US Class Current: 250/339.12
• CPC Class Codes: G01N 21/3563 for analysing solids; Preparation of samples therefor
• CPC Class Codes: G01N 21/3577 for analysing liquids, e.g. polluted water
• CPC Class Codes: G01N 21/359 using near infrared light
• CPC Class Codes: G01N 33/2829 Mixtures of fuels

Abstract

This invention provides a method for infrared and near-infrared estimations of properties of materials with improved model stability against random and pseudo-random effects and improved assessment of the appropriateness of the model for an unknown sample. Statistical tests are defined for quantifying the model'"'"'s stability against random effects (B-statistic) and the probability of substantial interpolation error (I-statistic). The method of this invention involves building many different infrared models for each property of interest, either randomly or by other strategies. The models with the lowest calibration errors for their respective calibration sets whose errors are least correlated to one another are averaged to create a composite model. Such composite models have improved stability and accuracy. The consistency with which the individual models (making up the composite model) estimate the same property value for an unknown sample is a measure of the improbability of substantial interpolation error.

Claims

1. A method for determining a property of interest of a test sample by utilizing a plurality of samples, each having a known value of the property of interest, comprising:

(a) subjecting each sample in the plurality of samples to a known light energy and collecting a corresponding spectrum for each such sample;

(b) grouping the samples in the plurality of samples into a plurality of subsets;

(c) determining a model for each such subset from the corresponding spectra of the samples in such subset;

(d) selecting a group of models from the models determined for the each of the subsets based on a predetermined criterion;

(e) retaining a subset of models from the group of models having residuals that are least correlated to one another; and

(f) determining a composite model from the retained models to determine the property of interest of the test sample.

2. The method of claim 1 further comprising determining the applicability of the composite model to the test sample by utilizing a disparity in predictions of the retained models for the test sample.

3. The method of claim 1 further comprising:

(i) determining a spectrum of the test sample; and

(ii) determining the property of interest of the test sample by utilizing the composite model.

4. The method of claim 3, wherein the property of interest is an octane number.

5. The method of claim 3, wherein the property of interest is a cetane number.

6. The method of claim 3, wherein the property of interest is a chemical composition.

7. The method of claim 3, wherein each sample in the plurality of samples is a sample of gasoline.

8. The method of claim 3, wherein each sample in the plurality of samples is a sample of diesel fuel.

9. The method of claim 3, wherein each sample in the plurality of samples is a sample of refinery component stream.

10. The method of claim 3, wherein the samples are subjected to a near-infrared energy and the spectrum is a corresponding near-infrared spectrum.

11. The method of claim 10, wherein each model is a regression model.

12. The method of claim 1, wherein the predetermined criterion is a value of the standard errors of calibration.

13. The method of claim 1, wherein the property of interest is a physical property of the test sample.

14. The method of claim 1, wherein the property of interest is a chemical property of the test sample.

15. A method for developing a model for determining a property of interest of a test sample:

(a) obtaining a plurality of samples of a material having known values of the property of interest;

(b) collecting a near-infrared spectrum of each of the samples in the plurality of samples;

(c) grouping the samples in the plurality of samples into a plurality of subsets;

(d) determining a regression model for each subset from the near-infrared spectra of the samples forming such subset;

(e) selecting regression models which have a standard error of calibration less than a predetermined value;

(f) retaining regression models from the selected models which are relatively poorly correlated to each other;

(g) averaging the retained models to determine a composite model;

(h) determining a near-infrared spectrum of the test sample; and

(i) determining the property of interest of the test sample by applying the composite model to the near-infrared spectrum of the test sample.

16. A method of determining the value of a property of interest of a test material, comprising:

(a) obtaining a plurality of samples of the material, each such sample having a known value of the property of interest;

(b) collecting a near-infrared spectrum for each sample in the plurality of samples using near-infrared radiation;

(c) determining a separate regression model for a plurality of subsets of the plurality of samples;

(d) selecting regression models having a standard error of calibration less than a predetermined value;

(e) determining a residual for each sample and each selected regression model as the difference between a value of the property of interest determined by such selected regression model and a value determined by a primary measurement technique for the parameter of interest;

(f) pairing the regression models having any common samples;

(g) determining a correlation coefficient between residuals of common samples in each of the pairs of models;

(h) selecting a subset of selected regression models, wherein the residuals of the regression models in the subset are the least correlated to one another;

(i) averaging the models in the selected subset (Step h) to obtain a composite model;

(j) estimating the degree of improvement of the composite model estimates using a B-statistic, wherein the B-statistic is defined as 1-R², where R is the average correlation coefficient between residuals of the subset of the least-correlated models;

(k) collecting a near-infrared spectrum of the test sample; and

(l) computing the value of the parameter of interest of the test sample using the composite model.

17. The method of claim 16 further comprising comparing the computed value of the parameter of interest as calculated by each model in the subset of least-correlated models to determine the likelihood of interpolation error.

18. The method of claim 16, further comprising:

(i) determining whether the property of interest of the test sample is an outlier; and

(ii) Determining an I-statistic for the computed value of the parameter of interest of the test sample using the subset of least correlated models, wherein the I-statistic is defined as a ratio of the variance of the subset models'"'"' estimates of the parameter of interest to the square of the average standard error of calibration of the subset models.

19. The method of claim 18, wherein an I-statistic greater than about 1 indicates that the computed value of the parameter of interest of the test sample is more likely an inlier than not.

20. The method of claim 19, wherein the test sample material is gasoline and the property of interest is an octane number of the gasoline.

21. A method for determining a property of interest of a test sample by utilizing a plurality of samples, each having a known value of the property, comprising:

(a) subjecting each sample in the plurality of samples to a known light energy and collecting a corresponding spectrum for each such sample;

(b) determining a composite model from the spectra of the plurality of samples;

(c) determining a spectrum of the test sample;

(d) determining the property of interest of the test sample by applying the composite model to the spectrum of the test sample;

(e) determining whether the property of interest of the test sample is an outlier; and

(f) determining an I-statistic for the parameter of interest of the test sample if the test sample is not an outlier, wherein the I-statistic is defined as a ratio of the variance of the subset models'"'"' estimates of the parameter of interest to the square of the average standard error of calibration of the subset models.

(g) predicting the test sample to be an inlier if the value of the I-statistic of the parameter of interest of the test sample is greater than about 1.

22. The method of claim 21, wherein the samples are subjected to near-infrared light energy and the spectra are near-infrared spectra.

23. The method of claim 21, wherein each model is a regression model.