Augmented classical least squares multivariate spectral analysis

• US 6,842,702 B2
• Filed: 09/11/2003
• Issued: 01/11/2005
• 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:
  US10/661,968
• Publication Number:
  US 20040064259A1
• Time in Patent Office: 488 Days
• Field of Search: 702 22- 26, 702/27, 702/28, 702/30, 702/181, 702189-190, 702196-199, 250/339.08, 250/339.09, 250/339.12, 250/559.1, 250/573
• US Class Current: 702/18
• 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 of multivariate spectral analysis, comprising the steps of:

a) obtaining an estimate of spectral error covariance E_A for measured set of multivariate spectral data A;

b) decomposing the spectral error covariance 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 error covariance E_A, and E is a set of n×

p random errors and spectral variations not useful for prediction;

c) guessing pure-component spectra K for the set of multivariate spectral data A;

d) predicting a set of component values Ĉ

according to

Ĉ

AK^T(KK^T)<sup>−

1</sup>=A(K^T)⁺;

e) augmenting the set of predicted component values Ĉ

with at least one vector of the T scores to obtain a first set of augmented component values $\hat{\tilde{C}};<br><br>$

f) estimating augmented pure-component spectra $\hat{\tilde{K}}$

according to $\hat{\tilde{K}}={\left({\tilde{C}}^T<br><br>\tilde{C}\right)}^{-1}<br><br>{\tilde{C}}^T<br><br>A=\hat{\tilde{C}}<br><br>{\hspace{0.17em}<br><br>}^{+}<br><br>A;<br><br>$

g) testing for convergence according to $<br><br>A-\hat{\tilde{C}}<br><br>\hat{\tilde{K}}<br><br><br><br>{\hspace{0.17em}<br><br>}^2<br><br>\text{\hspace{1em}<br><br>};<br><br>$

h) predicting a second set of augmented component values $\hat{\tilde{C}}$

 

according to $\hat{\tilde{C}}=A<br><br>{{\hat{\tilde{K}}}^T<br><br>\left(\hat{\tilde{K}}<br><br>{\hat{\tilde{K}}}^T\right)}^{-1}={A<br><br>\left({\hat{\tilde{K}}}^T\right)}^{+};<br><br>$

i) replacing the augmented portion of the second set of augmented component values $\hat{\tilde{C}}$

 

with the at least one vector of the T scores to obtain a third set of augmented component values $\widehat{\tilde{C};<br><br>}$

 

and

j) repeating steps f) through i) at least once.

2. The method of claim 1, wherein the steps f) through i) are repeated until the test of step g) converges to obtain an alternating classical least squares solution for $\hat{\tilde{K}}$

and $\hat{\tilde{C}}.$

3. The method of claim 1, further comprising replacing the augmented portion of the augmented pure-component spectra $\hat{\tilde{K}}$

with at least one vector of the P loading vectors prior to step h).

4. The method of claim 1, further comprising augmenting $\hat{\tilde{K}}$

with at least one vector representing a spectral shape that is representative of at least one additional source of spectral variation prior to step h).

5. The method of claim 1, further comprising applying at least one constraint to the non-augmented portion of $\hat{\tilde{K}}$

at step f).

6. The method of claim 5, wherein the at least one constraint is selected from the group consisting of non-negativity, equality, closure, monotonic constraint, unimodality, and selectivity.

7. The method of claim 1, further comprising applying at least one constraint to the non-augmented portion of $\hat{\tilde{C}}$

at step h).

8. The method of claim 7, wherein the at least one constraint is selected from the group consisting of non-negativity, equality, closure, monotonic constraint, unimodality, and selectivity.

9. The method of claim 1, wherein the guessed pure-component spectra K comprises random numbers.

10. The method of claim wherein the measured set of multivariate spectral data A comprises image data.

11. The method of claim 10, wherein the spectral error covariance E^A is obtained from a shift difference generated from a single image.

12. The method of claim 10, wherein the spectral error covariance E^A is obtained from repeat image spectra.

13. A method of multivariate spectral analysis, comprising the steps of:

a) obtaining an estimate of spectral error covariance E_A for measured set of multivariate spectral data A;

b) decomposing the spectral error covariance 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 error covariance E_A, and E is a set of n×

p random errors and spectral variations not useful for prediction;

c) guessing pure-component spectra K for the set of multivariate spectral data A;

d) augmenting the pure-component spectra K with at least one vector of the P loading vectors to obtain first augmented pure-component spectra {tilde over (K)};

e) predicting a first set of augmented component values $\hat{\tilde{C}}$

 

according to $\hat{\tilde{C}}=A<br><br>{{\tilde{K}}^T<br><br>\left(\tilde{K}<br><br>{\tilde{K}}^T\right)}^{-1}+{A<br><br>\left({\tilde{K}}^T\right)}^{+};<br><br>$

f) estimating second augmented pure-component spectra $\hat{\tilde{K}}$

 

according to $\hat{\tilde{K}}={\left({\tilde{C}}^T<br><br>\hat{\tilde{C}}\right)}^{-1}<br><br>{\hat{\tilde{C}}}^T<br><br>A=\hat{\tilde{C}}<br><br>{\hspace{0.17em}<br><br>}^{+}<br><br>A;<br><br>$

g) testing for convergence according to $<br><br>A-\hat{\tilde{C}}<br><br>\hat{\tilde{K}}<br><br><br><br>{\hspace{0.17em}<br><br>}^2<br><br>\text{\hspace{1em}<br><br>};<br><br>$

h) replacing the augmented portion of the second augmented pure-component spectra $\hat{\tilde{K}}$

 

with the at least one vector of the P loading vectors to obtain third augmented pure-component spectra $\hat{\tilde{K}};<br><br>$

 

and

i) predicting a second set of augmented component values $\hat{\tilde{C}}$

 

according to $\hat{\tilde{C}}=A<br><br>{{\hat{\tilde{K}}}^T<br><br>\left(\hat{\tilde{K}}<br><br>\text{\hspace{1em}<br><br>}<br><br>{\hat{\tilde{K}}}^T\right)}^{-1}={A<br><br>\left({\hat{\tilde{K}}}^T\right)}^{+};<br><br>$

j) repeating steps f) through i) at least once.

14. The method of claim 13, wherein the steps f) through i) are repeated until the test of step g) converges to obtain an alternating classical least squares solution for $\hat{\tilde{K}}$

and $\widehat{\tilde{C}.}$

15. The method of claim 14, wherein the measured set of multivariate spectral data A comprises image data.

16. The method of claim 13, further comprising replacing the augmented portion of the set of augmented component values $\hat{\tilde{C}}$

with at least one vector of the T scores prior to step h).

17. The method of claim 13, further comprising augmenting $\hat{\tilde{K}}$

with at least one vector representing a spectral shape that is representative of at least one additional source of spectral variation prior to step h).

18. The method of claim 13, further comprising applying at least one constraint to the non-augmented portion of $\hat{\tilde{K}}$

at step f).

19. The method of claim 18, wherein the at least one constraint is selected from the group consisting of non-negativity, equality, closure, monotonic constraint, unimodality, and selectivity.

20. The method of claim 13, further comprising applying at least one constraint to the non-augmented portion of $\hat{\tilde{C}}$

at step h).

21. The method of claim 20, wherein the at least one constraint is selected from the group consisting of non-negativity, equality, closure, monotonic constraint unimodality, and selectivity.

22. The method of claim 13, wherein the guessed pure-component spectra K comprises random numbers.

23. The method of claim 13, wherein the measured set of multivariate spectral data A comprises image data.

24. The method of claim 23, wherein the estimate of the error covariance E_A is obtained from a shift difference generated from a single image.

25. The method of claim 23, wherein the estimate of the error covariance E_A is obtained from repeat image spectra.

26. The method of claim 25, wherein the spectral error covariance E_A is obtained from a shift difference generated from a single image.

27. The method of claim 25, wherein the spectral error covariance E_A is obtained from repeat image spectra.

28. A method of multivariate spectral analysis, comprising the steps of:

a) obtaining an estimate of the spectral error covariance E_A for measured set of multivariate spectral data A;

b) decomposing the spectral error covariance 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 error covariance E_A, and E is a set of n×

p random errors and spectral variations not useful for prediction;

c) guessing a set of component values C for the set of multivariate spectral data A;

d) estimating pure-component spectra {circumflex over (K)} according to {circumflex over (K)}=(C^TC)<sup>−

1</sup>C^TA=C⁺A;

e) augmenting the pure-component spectra {circumflex over (K)} with at least one vector of the P loading vectors to obtain first augmented pure-component spectra $\hat{\tilde{K}};<br><br>$

f) predicting a first set of augmented component values $\hat{\tilde{C}}$

 

according to $\hat{\tilde{C}}=A<br><br>{{\hat{\tilde{K}}}^T<br><br>\left(\hat{\tilde{K}}<br><br>{\hat{\tilde{K}}}^T\right)}^{-1}=A<br><br>\left({\hat{\tilde{K}}}^T\right);<br><br>$

g) testing for convergence according to ${<br><br>A-\hat{\tilde{C}}<br><br>\hat{\tilde{K}}<br><br>}^2;<br><br>$

h) estimating second augmented pure-component spectra $\hat{\tilde{K}}$

 

according to $\hat{\tilde{K}}={\left({\hat{\tilde{C}}}^T<br><br>\hat{\tilde{C}}\right)}^{-1}<br><br>{\hat{\tilde{C}}}^T<br><br>A=\widehat{\tilde{C^{+}}}<br><br>A;<br><br>$

i) replacing the augmented portion of the second augmented pure-component spectra $\hat{\tilde{K}}$

 

with the at least one vector of the P loading vectors to obtain a third augmented pure-component spectra $\hat{\tilde{K}}$

 

and

j) repeating steps f) through i) at least once.

29. The method of claim 28, wherein the steps f) through i) are repeated until the test of step g) converges to obtain an alternating classical least squares solution for $\hat{\tilde{K}}$

and $\hat{\tilde{C}}.$

30. The method of claim 28, further comprising replacing the augmented portion of the set of augmented component values $\hat{\tilde{C}}$

with at least one vector of the T scores prior to step h).

31. The method of claim 28, further comprising augmenting $\hat{\tilde{K}}$

with at least one vector representing a spectral shape that is representative of at least one additional source of spectral variation prior to step h).

32. The method of claim 28, further comprising applying at least one constraint to the non-augmented portion of $\hat{\tilde{K}}$

at step h).

33. The method of claim 32, wherein the at least one constraint is selected from the group consisting of non-negativity, equality, closure, monotonic constraint, unimodality, and selectivity.

34. The method of claim 28, further comprising applying at least one constraint to the non-augmented portion of $\hat{\tilde{C}}$

at step f).

35. The method of claim 34, wherein the at least one constraint is selected from the group consisting of non-negativity, equality, closure, monotonic constraint unimodality, and selectivity.

36. The method of claim 28, wherein the guessed set of component values C comprises random numbers.

37. A method of multivariate spectral analysis, comprising the steps of:

a) obtaining an estimate of the spectral error covariance E_A for measured set of multivariate spectral data A;

b) decomposing the spectral error covariance 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 error covariance E_A, and E is a set of n×

p random errors and spectral variations not useful for prediction;

c) guessing a set of component values C for the set of multivariate spectral data A;

d) augmenting the set of component values C with at least one vector of the T scores to obtain a first set of augmented component values {tilde over (C)};

e) estimating augmented pure-component spectra $\hat{\tilde{K}}$

 

according to $\hat{\tilde{K}}={\left({\tilde{C}}^T<br><br>\tilde{C}\right)}^{-1}<br><br>{\tilde{C}}^T<br><br>A={\tilde{C}}^{+}<br><br>A;<br><br>$

f) testing for convergence according to ${<br><br>A-\tilde{C}<br><br>\hat{\tilde{K}}<br><br>}^2;<br><br>$

g) predicting a second set of augmented component values $\hat{\tilde{C}}$

 

according to $\hat{\tilde{C}}=A<br><br>\text{\hspace{1em}<br><br>}<br><br>{{\tilde{K}}^T<br><br>\left(\hat{\tilde{K}}<br><br>\text{\hspace{1em}<br><br>}<br><br>{\hat{\tilde{K}}}^T\right)}^{-1}<br><br>{A<br><br>\left({\hat{\tilde{K}}}^T\right)}^{+};<br><br>$

h) replacing the augmented portion of the second set of augmented component values $\hat{\tilde{C}}$

 

with the at least one vector of the T scores to obtain a third set of augmented component values $\hat{\tilde{C}}$

 

and

i) repeating steps e) through h) at least once, using the augmented component values $\hat{\tilde{C}}$

 

in step f).

38. The method of claim 37, wherein the steps e) through h) are repeated until the test of step g) converges to obtain an alternating classical least squares solution for $\hat{\tilde{K}}$

and $\hat{\tilde{C}}$

39. The method of claim 37, further comprising replacing the augmented portion of the augmented pure-component spectra $\hat{\tilde{K}}$

with at least one vector of the P loading vectors prior to step e).

40. The method of claim 37, further comprising augmenting $\hat{\tilde{K}}$

with at least one vector representing a spectral shape that is representative of at least one additional source of spectral variation prior to step e).

41. The method of claim 37, further comprising applying at least one constraint to the non-augmented portion of $\hat{\tilde{K}}$

at step e).

42. The method of claim 41, wherein the at least one constraint is selected from the group consisting of non-negativity, equality, closure, monotonic constraint, unimodality, and selectivity.

43. The method of claim 37, further comprising applying at least one constraint to the non-augmented portion of $\hat{\tilde{C}}$

at step g).

44. The method of claim 43, wherein the at least one constraint is selected from the group consisting of non-negativity, equality, closure, monotonic constraint unimodality, and selectivity.

45. The method of claim 37, wherein the guessed set of component values C comprises random numbers.

46. The method of claim 37, wherein the measured set of multivariate spectral data A comprises image data.

47. The method of claim 46, wherein the spectral error covariance E_A is obtained from a shift difference generated from a single image.

48. The method of claim 46, wherein the spectral error covariance E_A is obtained from repeat image spectra.