Hybrid least squares multivariate spectral analysis methods

• US 6,341,257 B1
• Filed: 03/03/2000
• Issued: 01/22/2002
• 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): WACHSMAN, HAL D
• Application Number:
  US09/518,773
• Time in Patent Office: 690 Days
• Field of Search: 702/191, 702/19, 702/21-23, 702/27, 702/28, 702/30-32, 702/75-78, 702/85, 702/86, 702/179-181, 702/183, 702/189, 702/194, 702/196, 702/FOR.103, 702/FOR.104, 702/FOR.107, 702/FOR.108, 702/FOR-115, 702/FOR.131, 702/FOR.134, 702/FOR-156, 702/FOR.165, 702/FOR.166, 702/FOR-168, 356/346, 356/300-303, 356/306, 356/326, 356/36, 700/266-269, 703/11, 703/12, 706/924, 436/8, 436/43, 436/149, 436/150, 436/805, 436/806, 436/808, 436/171, 436/174, 422/50, 422/62, 422/68.1, 422/70, 422/82.01, 422/82.05, 422/82.09, 422/82, 422/89, 250/339.08, 250/339.09, 250/338.5, 250/338.1, 250/339.07, 250/341.1, 250/339-1-3, 250/341.5, 250/341.8, 73/233.6, 73/233.7, 73/234.1, 73/240.2
• US Class Current: 702/27
• CPC Class Codes: G01J 2003/2866 Markers; Calibrating of scan
• CPC Class Codes: G01J 3/28 Investigating the spectrum using colour filters G01J3/51

Abstract

A set of hybrid least squares multivariate spectral analysis methods in which spectral shapes of components or effects not present in the original calibration step are added in a following estimation or calibration step to improve the accuracy of the estimation of the amount of the original components in the sampled mixture. The “hybrid” method herein means a combination of an initial classical least squares analysis calibration step with subsequent analysis by an inverse multivariate analysis method. A “spectral shape” herein means normally the spectral shape of a non-calibrated chemical component in the sample mixture but can also mean the spectral shapes of other sources of spectral variation, including temperature drift, shifts between spectrometers, spectrometer drift, etc. The “shape” can be continuous, discontinuous, or even discrete points illustrative of the particular effect.

Claims

1. A method of forming a hybrid model of at least one known constituent or property in a set of samples comprising:

(a) forming a classical least squares (CLS) calibration model of the at least one known constituent or property in the set of samples from reference values and measured responses to a stimulus of individual samples in the set of samples;

(b) estimating a CLS prediction value of the at least one known constituent or property in the set of samples from the CLS calibration model by a CLS prediction model, wherein the CLS prediction model produces residual errors;

(c) adding, as needed, spectral shapes representative of sources of signal variation not specifically modeled in step (a) to the CLS prediction model; and

(d) passing the residual errors to an inverse analysis algorithm, to provide a hybrid model of the at least one known constituent or property in the set of samples.

2. The method of claim 1 wherein the inverse analysis algorithm is partial least squares, partial least squares 2, principle component regression, multiple linear regression, or continuum regression.

3. The method of claim 1 for additionally estimating a value of at least one known constituent or property of a sample further comprising:

(e) measuring responses of the sample to the stimulus, using the CLS calibration model and the CLS prediction model and adding at least one spectral shape, as needed, to the CLS prediction model to estimate a CLS prediction value of the at least one known constituent or property of the sample and producing residual errors;

(f) taking the residual errors produced in step (e) and inserting them into the inverse analysis algorithm of the hybrid model to estimate an inverse portion of the residual errors of the CLS prediction value; and

(g) combining the CLS prediction value from step (e) and the inverse portion of the residual errors of the CLS prediction value from step (f) to form an estimate of the value of the at least one known constituent or property of the sample.

4. The method of claim 3 wherein the applicable steps are repeated for additional constituents or properties from the CLS calibration model that are present in the sample.

5. The method of claim 3 further comprising factor analyzing the residual errors and wherein the inverse analysis algorithm uses only those factor-analyzed residual errors that are most effective in estimating the value of the at least one known constituent or property in the sample.

6. The method of claim 3 further including identifying outliers having values outside an allowed statistical range of the hybrid model.

7. The method of claim 6 wherein the identifying is done by spectral F ratios, and Mahalonobis distances, or examination of spectral residuals.

8. The method of claim 3 wherein the estimate of the value represents a first class of the at least one known constituent or property in the set of samples, so long as the estimate of the value is within acceptable statistical limits.

9. The method of claim 3 wherein the inverse analysis algorithm is partial least squares, partial least squares 2, principal components regression, inverse lease squares, multiple linear regression, or continuum regression.

10. The method of claim 1 further including performing a cross-validation to select a preferred number of residual errors for passing to the inverse analysis algorithm of the hybrid model.

11. The method of claim 1 wherein the reference values are selected to distinguish classes of samples in the set of samples and the method of forming the hybrid model includes a further step of classifying whether or not the at least one known constituent or property in the set of samples is within a first class.

12. The method of claim 1 further including factor analyzing the residual errors from step (b) to select a preferred number of residual errors for passing to the inverse analysis algorithm in step (d).

13. The method of claim 1 further including identifying outliers having values outside an allowed range of the hybrid model.

14. The method of claim 13 wherein the identifying is done by spectral F ratios, Mahalonobis distances, or examination of concentration and spectral residuals.

15. A method for estimating a value of at least one known constituent or property of a sample utilizing a hybrid model comprising:

(a) forming a classical least squares (CLS) calibration model of the at least one known constituent or property in a set of samples from reference values and measured responses to a stimulus of individual samples in the set of samples;

(b) estimating a CLS prediction value of the at least one known constituent or property in the set of samples from the CLS calibration model by a CLS prediction model, wherein the CLS prediction model produces residual errors;

(c) adding, as needed, spectral shapes representative of sources of signal variation not explicitly modeled in step (a) to the CLS prediction model;

(d) passing the residual errors for step (b) to an inverse analysis algorithm, to provide a hybrid model of the at least one known constituent or property in the set of samples;

(e) measuring responses of a sample to the stimulus, using the CLS calibration model and the CLS prediction model and adding at least one spectral shape, as needed, to the CLS prediction model of the hybrid model to estimate a CLS prediction value of the at least one known constituent or property of the sample and producing residual errors;

(f) taking the residual errors produced in step (e) and inserting them into the inverse analysis algorithm to estimate an inverse portion of the residual errors of the CLS prediction value; and

(g) combining the CLS prediction value from step (e) and the inverse portion of the residual errors of the CLS prediction value from step (f) to form an estimate of the value of the at least one known constituent or property of the sample.

16. The method of claim 15 further including performing a cross-validation to select a preferred number of residual errors for passing to the inverse analysis algorithm.

17. The method of claim 15 further including identifying outliers having values outside an allowed range of the hybrid model.

18. The method of claim 17 wherein the identifying is done by spectral F ratios, concentration F ratios, Mahalonobis distances, or examination of concentration and spectral residuals.

19. The method of claim 15 wherein the reference values are selected to distinguish classes of samples in the set of samples and the method includes a further step of classifying whether or not the at least one known constituent or property in the set of samples is within a first class.