Augmented classical least squares multivariate spectral analysis

• US 6,687,620 B1
• Filed: 07/31/2002
• Issued: 02/03/2004
• Est. Priority Date: 08/01/2001
• Status: Active Grant

• Current Assignee: National Technology & Engineering Solutions Of Sandia LLC (Honeywell International Inc.)
• Original Assignee: Sandia Corporation (Martin Marietta Corporation)
• Inventors: Haaland, David M., Melgaard, David K.
• Primary Examiner(s): Barlow, John
• Assistant Examiner(s): LE, JOHN H
• Application Number:
  US10/209,841
• Time in Patent Office: 552 Days
• Field of Search: 702/22-26, 702/30, 702/181, 702/189-190, 702/196-199, 700/266-269, 73/233.6, 73/233.7, 73/234.1, 73/240.2, 250/339.08, 250/339.09, 250/338.5, 250/338.1, 250/339.07, 250/341.1, 250/339.11, 250/339.12, 250/339.13, 250/339.14
• US Class Current: 702/22
• CPC Class Codes: G01N 21/274 Calibration, base line adjustment, drift correction
• CPC Class Codes: G01N 21/359 using near infrared light
• CPC Class Codes: G01N 21/64 Fluorescence; Phosphorescence
• CPC Class Codes: G01N 21/65 Raman scattering
• CPC Class Codes: G01N 2201/1293 resolving multicomponent spectra

Abstract

A method of multivariate spectral analysis, termed augmented classical least squares (ACLS), provides an improved CLS calibration model when unmodeled sources of spectral variation are contained in a calibration sample set. The ACLS methods use information derived from component or spectral residuals during the CLS calibration to provide an improved calibration-augmented CLS model. The ACLS methods are based on CLS so that they retain the qualitative benefits of CLS, yet they have the flexibility of PLS and other hybrid techniques in that they can define a prediction model even with unmodeled sources of spectral variation that are not explicitly included in the calibration model. The unmodeled sources of spectral variation may be unknown constituents, constituents with unknown concentrations, nonlinear responses, non-uniform and correlated errors, or other sources of spectral variation that are present in the calibration sample spectra. Also, since the various ACLS methods are based on CLS, they can incorporate the new prediction-augmented CLS (PACLS) method of updating the prediction model for new sources of spectral variation contained in the prediction sample set without having to return to the calibration process. The ACLS methods can also be applied to alternating least squares models. The ACLS methods can be applied to all types of multivariate data.

Claims

1. A method for analyzing multivariate spectral data, comprising the steps of:

a) creating a calibration model for a calibration set of multivariate spectral data A by;

i) obtaining a set of reference component values C representative of at least one of the spectrally active components in the calibration set of multivariate spectral data A,

ii) estimating pure-component spectra {circumflex over (K)} for the at least one of the spectrally active components according to {circumflex over (K)}=(C^TC)<sup>−

1</sup>C^TA=C⁺A,

iii) obtaining spectral residuals E_A according to E_A=A−

C{circumflex over (K)}, and

iv) augmenting the estimated pure-component spectra {circumflex over (K)} with at least one vector of the spectral residuals E_A to obtain augmented pure-component spectra {tilde over ({circumflex over (K)})}; and

b) predicting a set of component values {tilde over ({circumflex over (C)})} for a prediction set of multivariate spectral data A_P by;

i) further augmenting the augmented pure-component spectra {tilde over ({circumflex over (K)})} with at least one vector representing a spectral shape that is representative of at least one additional source of spectral variation in the prediction set, and

ii) predicting the set of component values {tilde over ({circumflex over (C)})} using the further augmented pure-component spectra {tilde over ({circumflex over (K)})} according to {tilde over ({circumflex over (C)})}=A_P{tilde over ({circumflex over (K)})}^T({tilde over ({circumflex over (K)})}{tilde over ({circumflex over (K)})}^T)<sup>−

1</sup>=A_P({tilde over ({circumflex over (K)})}^T)⁺.

2. The method of claim 1, wherein augmenting step a)iv) is repeated until all sources of spectral variation are accounted for in the calibration set.

3. The method of claim 1, wherein an F-test is used to select the number of vectors used to augment the estimated pure-component spectra {circumflex over (K)}.

4. The method of claim 1, further comprising the step of augmenting the estimated pure-component spectra {circumflex over (K)} with at least one vector representing at least one spectral shape representative of at least one additional sources of spectral variation in the calibration set prior to step a)iii).

5. The method of claim 1, further comprising the step of augmenting {circumflex over (K)} to account for baseline variations in the calibration set prior to step a) iii).

6. The method of claim 1, further comprising the step of augmenting {tilde over ({circumflex over (K)})} to account for baseline variations in the calibration set prior to step b).

7. The method of claim 1, wherein the reference component values C comprise concentrations of the spectrally active components.

8. The method of claim 1, wherein the spectrally active components are selected from the group consisting of chemical specie, chemical interactions among species, physical variations, temperature variations, humidity variations, spectrometer drift, changes in spectrometers, and insertion effects.

9. The method of claim 1, wherein the spectral residuals E_A are obtained from a full CLS model or a cross-validated CLS model.

10. The method of claim 1, wherein the spectral residuals E_A are decomposed according to E_A=TP+E, where T is a set of n×

r scores and P is a set of r×

p loading vectors obtained from factor analysis of the spectral residuals E_A, and E is a set of n×

p random errors and spectral variations not useful for prediction.

11. The method of claim 10, wherein the P loading vectors comprise the at least one vector of the spectral residuals E_A used to augment the estimated pure-component spectra {circumflex over (K)} in step a)iv).

12. The method of claim 10, wherein the factor analysis method comprises principal components analysis.

13. A method for analyzing multivariate spectral data, comprising the steps of:

a) creating a calibration model for a calibration set of multivariate spectral data A by;

i) obtaining a set of reference component values C representative of at least one of the spectrally active components in the calibration set of multivariate spectral data A,

ii) estimating pure-component spectra {circumflex over (K)} for the at least one of the spectrally active components according to {circumflex over (K)}=(C^TC)<sup>−

1</sup>C^TA=C⁺A,

iii) obtaining spectral residuals E_A according to E_A=A−

C{circumflex over (K)},

iv) decomposing the spectral residuals E_A according to E_A=TP+E where T is a set of n×

r scores and P is a set of r×

p loading vectors obtained from factor analysis of the spectral residuals E_A, and E is a set of n×

p random errors and spectral variations not useful for prediction,

v) augmenting the set of reference component values C with at least one vector of the T scores to obtain a set of augmented component values {tilde over (C)}, and

vi) estimating augmented pure-component spectra {tilde over ({circumflex over (K)})} according to {tilde over ({circumflex over (K)})}=({tilde over (C)}^T{tilde over (C)})<sup>−

1</sup>{tilde over (C)}^TA≈

{tilde over (C)}⁺A; and

b) predicting a set of component values {tilde over ({circumflex over (C)})} for a prediction set of multivariate spectral data A_P according to {tilde over ({circumflex over (C)})}=A_P{tilde over ({circumflex over (K)})}^T({tilde over ({circumflex over (K)})}{tilde over ({circumflex over (K)})}^T)<sup>−

1</sup>=A_P({tilde over ({circumflex over (K)})}^T)⁺.

14. The method of claim 13, wherein the augmenting step a)iv) is repeated until all sources of spectral variation are accounted for in the calibration set.

15. The method of claim 13, wherein an F-test is used to select the number of vectors used to augment the estimated pure-component spectra {circumflex over (K)}.

16. The method of claim 13, further comprising the step of augmenting the estimated pure-component spectra {circumflex over (K)} with at least one vector representing a spectral shape representative of at least one additional source of spectral variation in the calibration set prior to step a)iii).

17. The method of claim 13, further comprising the step of augmenting {circumflex over (K)} to account for baseline variations in the calibration set prior to step a)iii).

18. The method of claim 13, further comprising the step of augmenting the augmented pure-component spectra {tilde over ({circumflex over (K)})} with at least one vector representing a spectral shape representative of at least one additional source of spectral variation in the prediction set prior to step b).

19. The method of claim 13, further comprising the step of augmenting {tilde over ({circumflex over (K)})} to account for baseline variations in the calibration set prior to step b).

20. The method of claim 13, wherein the reference component values C comprise concentrations of the spectrally active components.

21. The method of claim 13, wherein the spectrally active components are selected from the group consisting of chemical specie, chemical interactions among species, physical variations, temperature variations, humidity variations, spectrometer drift, changes in spectrometers, and insertion effects.

22. The method of claim 13, wherein the step a)iv) comprises augmenting the set of reference component values C with a set of random numbers.

23. The method of claim 13, wherein the spectral residuals E_A are obtained from a full CLS model or a cross-validated CLS model.

24. The method of claim 13, wherein the factor analysis method comprises principal components analysis.