Classical least squares multivariate spectral analysis

• US 6,415,233 B1
• Filed: 03/03/2000
• Issued: 07/02/2002
• Est. Priority Date: 03/04/1999
• Status: Expired due to Term

• Current Assignee: National Technology & Engineering Solutions Of Sandia LLC (Honeywell International Inc.)
• Original Assignee: Sandia Corporation (Martin Marietta Corporation)
• Inventors: Haaland, David M.
• Primary Examiner(s): Hoff, Marc S.
• Assistant Examiner(s): Raymond, Edward
• Application Number:
  US09/518,351
• Time in Patent Office: 851 Days
• Field of Search: 702/22, 702/23, 702/24-26, 702/30, 702/181, 702/189-190, 702/196-199, 430/12
• US Class Current: 702/22
• CPC Class Codes: G01J 3/28 Investigating the spectrum using colour filters G01J3/51

Abstract

An improved classical least squares multivariate spectral analysis method that adds spectral shapes describing non-calibrated components and system effects (other than baseline corrections) present in the analyzed mixture to the prediction phase of the method. These improvements decrease or eliminate many of the restrictions to the CLS-type methods and greatly extend their capabilities, accuracy, and precision. One new application of PACLS includes the ability to accurately predict unknown sample concentrations when new unmodeled spectral components are present in the unknown samples. Other applications of PACLS include the incorporation of spectrometer drift into the quantitative multivariate model and the maintenance of a calibration on a drifting spectrometer. Finally, the ability of PACLS to transfer a multivariate model between spectrometers is demonstrated.

Claims

1. A method for performing a classical least squares estimation of the value of at least one source of spectral variation of a sample comprising:

utilizing a previously constructed calibration data set wherein the constituents of such calibration data set include at least one of the sources of spectral variation that affect the optical response of the sample to be measured, such calibration data set yielding a matrix {circumflex over (K)} representing the combination of the vectors containing the at least one spectral shape of the measured sources of spectral variation;

measuring the optical response of a sample set that contains the at least one of the sources of spectral variation in the calibration data set and at least one additional source of spectral variation whose value was not represented in the original calibration data set, said measurement forming a prediction data set;

adding at least one vector representing a spectral shape that is representative of the at least one additional source of spectral variation in the prediction data set to the matrix {circumflex over (K)} to form an augmented matrix {tilde over ({circumflex over (K)})}; and

estimating the value of the at least one of the sources of spectral variation in the calibration data set or the at least one additional source of spectral variation in the sample by a classical least squares prediction utilizing the augmented matrix {tilde over ({circumflex over (K)})}, wherein the at least one additional source of spectral variation is a non-baseline source of spectral variation.

2. The method of claim 1 wherein the spectral shape that is representative of the at least one additional source of spectral variation is generated from spectral measurements of pure samples of chemical constituents that are the at least one additional source of spectral variation, for constituents that are not affected by interactions with the sample.

3. The method of claim 1 wherein the spectral shape that is representative of the at least one additional source of spectral variation is generated from spectral measurements of representative calibration samples that have been spiked with a chemical constituent or parameter of interest, then subtracting out the representative calibration sample spectrum to leave the spectral shape of the chemical constituent or the parameter of interest.

4. The method of claim 1 wherein the spectral shape that is representative of the at least one additional source of spectral variation is generated from taking repeated spectral measurements of a representative calibration sample and then subtracting the original spectral measurement of the representative calibration sample from the rest of the measurements.

5. The method of claim 4 wherein the representative sample is a sample near the midpoint of the concentration range for that representative sample.

6. The method of claim 5 wherein the representative sample further includes samples at the extreme ends of its concentration range.

7. The method of claim 1 wherein a particular vector selected to represent a spectral shape to add to the matrix {circumflex over (K)} corresponds to an eigenvector that represents at least one additional source of spectral variation in the prediction data set whose value was not represented in the calibration data set.

8. The method of claim 7 wherein the eigenvector is selected based on a functional form relating a prediction error from a calibration model to the eigenvector and a net-analyte signal.

9. The method of claim 1 wherein the at least one additional source of spectral variation is due to a change in spectrometers between a primary spectrometer and a secondary spectrometer and the at least one spectral shape representing this change is the difference between a spectral shape of at least one sample measured on the primary spectrometer and a spectral shape of the same at least one sample measured on the secondary spectrometer or is an eigenvector or other linear combination of spectral shapes obtained from the spectral differences.

10. The method of claim 1 wherein the at least one additional source of spectral variation is due to a plurality of individual effects from several different sources that are not analyzed in the prediction data set and the spectral shape of this plurality is generated from the several different sources of spectral variations that are varied while the sources of spectral variation in the calibration data set that will be measured in the prediction data set are held at zero value or a constant value.

11. The method of claim 1 wherein there is a change in a constituent sample in the calibration data set from a measurement taken at a first time and a measurement taken at a second time due to a change in the sample itself, wherein the spectral shape that is representative of the at least one additional source of spectral variation that is due to this change is generated by taking the difference between the spectral measurements of the sample at the first time and the second time.

12. The method of claim 1 wherein the at least one additional source of spectral variation is due to a non-linear effect in a spectrographic system and the spectral shape is generated artificially by adding nonlinear effects to the calibration data set and then selecting at least one vector representing the non-linearity in the calibration data set matrix as the spectral shape.

13. The method of claim 12 wherein the non-linear effect is an effect caused by stray light.

14. The method of claim 12 wherein the non-linear effect is an effect cause by detector nonlinearities.

15. The method of claim 1 wherein the at least one additional source of spectral variation is that of a non-linear detector response and the spectral shape is generated by first subtracting a constant amount of intensity from all single-beam sample spectral intensities in the calibration data set matrix to simulate a first non-linear detector response and then selecting a vector representing error in the calibration data set matrix as the spectral shape.

16. A method for estimating a measure of at least one component within a sample including a plurality of components, comprising the steps of:

a) obtaining a calibration data set A₁ composed of first attributes K₁ of the plurality of components within a first sample set S₁ composed of at least two different samples;

b) obtaining a second data set A₂ composed of second attributes K₂ of the plurality of components within a second sample set S₂, wherein the second sample set S₂ includes at least one component not in the first sample set S₁;

c) estimating a third attribute K₃ of the at least one other component not in the first set S₁ of samples and combining it with calibration model from A₁ to obtain an augmented set of attributes {tilde over (K)}; and

d) predicting a measure of at least one component in the second sample set S₂ according to {tilde over (C)}=({tilde over (K)}^T{tilde over (K)})<sup>−

1</sup>{tilde over (K)}^TA≈

{tilde over (K)}A.

17. The method of claim 16, wherein the first attributes K₁ comprise a plurality of responses of the individual components within the first set S₁ of samples to an applied energy source selected from the group consisting of:

electromagnetic, electric, magnetic, acoustic, thermal, and mechanical.

18. The method of claim 17, wherein the first attributes K₁ can be represented as a vector or matrix K_ji, where i is one element of the response for the j^th component within the first set S₁ of samples.

19. The method of claim 17, wherein the first set of attributes K₁ comprise a spectral response.

20. The method of claim 19, wherein the estimated attribute K₃ comprises an estimated spectral shape of the spectral response of at least one component within the sample set S₂.

21. The method of claim 20, where the spectral shape is generated by first subtracting a constant amount of intensity from all single-beam sample spectral intensities in the calibration data set to simulate a first non-linear detector response.

22. The method of claim 16, wherein the estimated attribute K₃ can be represented as a vector or matrix.

23. The method of claim 22, further including selecting a vector representing error in the calibration data set A₁ as the spectral shape.

24. The method of claim 16, wherein the at least one attribute K₃ not in the first set of attributes K₁ is for at least one component not in the first set S₁ of samples.

25. The method of claim 16, wherein the at least one attribute K₃ not in the first set of attributes K₁ is for at least one source of variation not in the first calibration data set A₁.

26. The method of claim 16, wherein the calibration data set A₁ represents vectors expressing the spectral responses and concentrations of know components within the first sample set S₁.

27. The method of claim 24, wherein the at least one source of variation not in the first calibration data set is selected from the group consisting of:

spectrometer drift, changes in spectrometers, temperature variations, optical variations including sample insertion effects, chromatic variations, diffraction effects, and differences resulting from related chemical analogues in the sample.

28. The method of claim 16, wherein the estimated attribute K₂ comprises at least one vector expressing the spectral response of at least one additional component not in the prediction data set S₂.

30. The method of claim 16 wherein the functional form comprises the inverse of the ratio of the squared change in the net-analyte signal and the squared prediction error.

29. A method for estimating a measure of at least one component within a sample including a plurality of components, comprising the steps of:

a) obtaining a calibration data set A₁ composed of first attributes K₁ of the plurality of identified components within a first sample set S₁ composed of at least two different samples;

b) obtaining a measure E_A of unknown variables in the calibration data set A₁ and separating E_A into coherent P and incoherent E components;

c) augmenting the first set of attributes K₁ with at least one coherent P and at least one incoherent E components to obtain an augmented first set of attributes {tilde over (K)}₁;

d) obtaining a second data set A₂ composed of second attributes K₂ of the plurality of components within a second sample set S₂, wherein the second sample set S₂ includes at least one component not in the first sample set S₁;

e) estimating a third attribute K₃ of the at least one other component not in the first set S₁ of samples and combining it with calibration model from A₁ to obtain a second augmented set of attributes {tilde over (K)}₂; and

f) predicting a measure of at least one component in the second sample S₂ according to {tilde over (C)}=({tilde over (K)}^T{tilde over (K)})<sup>−

1</sup>{tilde over (K)}^TA≈

{tilde over (K)}A.