Calibration System for Spectrographic Analyzing Instruments

This invention relates to a system for analyzing materials and, more particularly, to a system for analyzing materials by infrared analysis.

Near infrared spectrographic instruments are used to provide accurate analysis of materials such as to determine measurable characteristics of materials, such as the concentrations of constituents of the materials or characteristics of the materials. For example, near infrared spectrographic instruments are used in agriculture to determine the oil, protein and moisture content of grain, the fat content of meat, the fat, protein and lactose content of milk and the urea content of milk. In addition, the near infrared spectrophotometers are used to analyze blood samples, and to analyze pharmaceutical samples. The instruments also have been used to measure physical properties or physical characteristics of materials. For example, the instruments have been successfully used to measure the hardness of wheat.

In typical systems of the prior art, a measurable characteristic is expected to correlate with absorbance at selected wavelengths in the near infrared spectrum. The measurable characteristic of a material can be represented in an absorbance equation summing products of values from an absorbance spectrum and weighting coefficients or summing products of derivatives of the absorbance spectrum and weighting coefficients. To measure the concentrations of constituents of an unknown sample, the absorbances of a multiplicity of sample materials similar to the unknown material are measured by the spectrographic instrument. The concentrations of the constituents of the sample materials are known. When the unknown material has a property to be measured, such as hardness of wheat, then this property will be known for each of the sample materials. From the absorbance measurements made on the multiplicity of sample materials, the weighting coefficients of the equations relating to the measurable characteristics to the absorbance measurements can be determined by multiple regression or by partial least squares regression. The process of determining the values of the weighting coefficients is called calibration. After the coefficients have been determined, the unknown material can be analyzed by the spectrographic instrument using the coefficients that have been determined from the sample materials.

Instead of measuring the absorbances at selected specific wavelengths which are known or presumed to correlate with the measurable characteristics, the absorbance of the sample materials can be measured at wavelengths distributed throughout the near infrared spectrum and coefficients in equations relating the measurable characteristics to the absorbance measurements can be developed by partial least squares regression. The measurable characteristics of the unknown material can then be determined by the spectrographic instrument by measuring the absorbances of the unknown material and then calculating the measurable characteristics from the measured absorbance values in accordance with the equations.

The above methods of analyzing material, to be accurate, require the sample materials to be similar to the unknown material being measured. However, because the unknown material is in fact unknown, it is sometimes difficult to obtain samples which closely resemble the known material and, as a result, the accuracy of the measurement suffers.

The present invention provides a new, improved method of calibrating an instrument to determine the coefficients to use in determining measurable characteristics of the unknown material. In accordance with the present invention, a library of near infrared spectra of a large number of sample materials is maintained in computer storage. For example, the spectra from a thousand different sample materials could be maintained in the instrument library. For each of these spectra, the measurable characteristics of the sample materials, to be determined in the unknown material, are known and are stored in computer storage. To provide a set of coefficients to analyze an unknown material, first the near infrared absorbance spectrum of the unknown material is measured and then is compared with the library of spectra of the sample materials to select a subset of spectra which most closely resemble the spectrum measured from unknown material. From this subset of spectra of sample materials, the weighting coefficients of the equations relating the measurable characteristics to the absorbance values are determined. In the preferred embodiment, this determination is carried out by partial least squares regression, but coefficients could also be determined by multiple regression. Once the coefficients have been determined, the measured absorbance spectrum of the unknown material and the equations yield the determinations of the measurable characteristics.

FIG. 1 is a block diagram illustrating the spectrographic instrument employed in the system of the invention; and

FIG. 2 is a flow chart representing the process of the invention.

The apparatus employed in the system of the present invention comprises a near infrared spectrometer 11 having an oscillating grating 13 on which the spectrometer directs light. The grating 13 reflects light with a narrow wavelength band through exit slit optics 15 to a sample 17. As the grating oscillates, the center wavelength of the light that irradiates the sample is swept through the near infrared spectrum. Light from the diffraction grating that is reflected by the sample is detected by near infrared photodetectors 19. The photodetectors generate a signal that is transmitted to an analog-to-digital converter 22 by amplifier 20. An indexing system 23 generates pulses as the grating 13 oscillates and applies these pulses to a computer 21 and to the analog-to-digital converter 22. In response to the pulses from the indexing system 23, the analog-to-digital converter converts successive samples of the output signal of the amplifier 20 to digital values. Each digital value thus corresponds to the reflectivity of the sample at a specific wavelength in the near infrared range. The computer 21 monitors the angular position of the grating 13 and accordingly monitors the wavelength irradiating the sample as the grating oscillates, by counting the pulses produced by the indexing system 23. The pulses produced by the indexing system 23 define incremental index points at which values of the output signal of the amplifier are converted to digital values. The index points are distributed incrementally throughout the near infrared spectrum and each correspond to a different wavelength at which the sample is irradiated. The computer 21 converts each reflectivity value to an absorbance of the material at the corresponding wavelength. The structure and operation of a suitable spectrometer is described in greater detail in U.S. Pat. No. 4,969,739.

In accordance with the present invention, the instrument shown in FIG. 1 is used to measure the absorbance spectra from a large number of sample materials and stores a library of these spectra in the memory of the computer 21. For example, the spectra of known materials may be from a thousand different known products. Associated with each spectrum of a sample material in the computer memory is the concentrations of the constituents of the material and/or the quantification of the property or properties to be measured in the unknown material.

In the preferred embodiment, the absorbance data in the library is compressed by averaging the absorbance values over eight nanometers or in other words by averaging successive groups of four of the measurements taken at the two nanometer increments. Assuming the spectrographic instrument measures the near infrared spectrum from 1100 nanometers to 2498 nanometers, the averaging step reduces the number of data points from 700 to 175.

The flow chart in FIG. 2 illustrates the process employed by the system to analyze an unknown material making use of the library of spectra stored in the computer 21 of the system of FIG. 1.

As shown in FIG. 2, the first step of the process is to measure the absorbance spectrum of the unknown material. This step creates a spectrum of absorbance values distributed throughout the near infrared spectrum at every two nanometers. This data is then compressed by averaging successive sets of four measurements to conform with the compressed data in the library representing the sample materials. Accordingly, the spectral data of the unknown material reduces to 175 data points. This compressed absorbance spectrum is called the target spectrum. To select the spectra in the library most closely resembling the target spectrum, the target spectrum is correlated with each compressed absorbance spectrum of the sample products. To carry out this correlation, the target spectrum is broken into peak regions by identifying local minima in the curve represented by the target spectrum. The computer program searches for all spectral values lower than two of the neighboring spectral values to find each minimum. A peak region is defined as the region from one local minimum up to, but not including, the next local minimum. To carry out the correlation for each peak region, the target spectrum is mean centered, that is, the average of the data points is found and then each spectral data point is represented by the difference between this average and the value of each data point. The data in each of the sample material spectra are also mean centered in the same manner. The mean centered data are represented as vectors x and y. The squared correlation between the mean centered vector x representing the unknown material and the mean centered vector y representing a sample material is defined as (.SIGMA.xy).multidot.(.SIGMA.xy)/›(.SIGMA.xx).multidot.(.SIGMA.yy)!. In the computer program, the product sums .SIGMA.xy, .SIGMA.xx and .SIGMA.yy are computed for each peak region. These sums are then pooled into three grand sums, .SIGMA..SIGMA.xy, .SIGMA..SIGMA.xx, .SIGMA..SIGMA.yy. When the sums from all the peak regions have been summed into grand sums, the pooled correlation is computed as:

This pooled correlation is computed between the target spectrum and each library spectrum. Those library spectra which have the highest correlation with the target spectrum are selected as the library spectra to be used to compute the weighting coefficients to carry out the analysis of the unknown material. In the preferred embodiment, any spectrum of a sample material which perfectly correlates with the unknown sample is not used in the calibration. This allows testing the program with spectra contained in the library. A minimum acceptable correlation may be specified to prevent sample materials with low correlations from being used in the calibration process. In addition, a minimum number of known materials having acceptable correlation can be specified. If this minimum number is not achieved, the analysis is not performed. In the appendix hereto, the program listing entitled SIMILAR.C is the source code in C of the program for correlating the target spectrum with the library spectra to select those spectra in the library which most closely, but not perfectly, match the target spectrum.

Instead of using the above method to select the spectra representing the known materials which most closely match the target spectrum, the method disclosed in application Ser. No. 08/645,104, filed May 13, 1996, by Xiaolin Chen and Stephen L. Monfre, may be used to select the spectra of the sample materials which most closely match the unknown material. Other methods of selecting the spectra of the sample materials which closely match the unknown material spectrum may also be used.

Assuming that the analysis of the unknown material is to measure concentrations of constituents in the unknown material, the analysis of the material may be represented by a set of equations or models, such as: ##EQU1## In these equations, C.sub.1 through C.sub.m are the estimated percentage concentrations of the constituents being measured. A.sub.1 through A.sub.n are the values of the target spectrum and k.sub.11 through k.sub.mn are weighting coefficients to be determined by the system of the present invention. In accordance with the present invention, the absorbance spectra representing those sample materials in the library of spectra of sample materials which are selected as those most closely resembling the target spectrum of the unknown material are used to determine the coefficients k.sub.11 through k.sub.mn by partial least squares regression (PLS). In the preferred embodiment, models are derived sequentially with from 1 to a specified maximum number of factors for each constituent being measured. Each model can be used to predict the constituent value of the unknown sample material. In the preferred embodiment a minimum number of factors is specified. The final predicted constituent value of the unknown material is a weighted average of the individual model predicted values with from the minimum to the maximum number of factors Y=.SIGMA.(y.sub.i w.sub.i)/.SIGMA.w.sub.i), where Y is the final predicted constituent value, the y.sub.i values are the predicted values from the PLS models, and the w.sub.i are the inverse of the sum of squared target sample spectrum residuals. Large coefficients are associated with overfitting, and should be avoided. Small residuals mean that more of the target spectrum variation has been modeled by PLS. These residuals are natural by-product of the PLS algorithm. In the appendix, the source code listing (in C) entitled MY.sub.-- PLS.C represents the program carrying out the partial least squares determination of the weighting coefficients and final constituent values. As described above, the unknown material are then calculated making use of these coefficients and the values of the target spectrum in the equations for C.sub.1 through C.sub.m. In this manner, the unknown material is analyzed to determine the percentages of its constituents.

Instead of analyzing the material by partial least squares, the coefficients equations relating the percentage constituents to absorbance spectrum values can be determined by multiple regression. When multiple regression is used, the absorbance values at specific selected wavelengths known to correlate with the constituents being measured are selected and these specific wavelengths are used in multiple regression equations.

In a similar manner, a property or properties of the unknown material can be determined. In order to determine physical properties of the material, the properties of the sample materials corresponding to the spectra in the library must be known and are stored in the computer memory. From these quantified physical properties of the sample materials, the coefficients of equations relating the properties to the spectrum values can be determined in the same manner as for the constituent percentages as described above.

If the target spectrum of the unknown material does not closely resemble or is not well represented in the library of spectra, then the target spectrum is added to the library with the measurable characteristics determined by the analysis so that the new spectrum can be used in measuring of a future material.

In the above description, the library spectra and the target spectra are described as absorbance spectra, which are determined as log 1/R of the reflectance measurements R. Instead of operating on the undifferentiated absorbances values, the first derivative of the spectrum of absorbance values may be determined and the process applied to the set of first derivative values, which is also called an absorbance spectrum. When a target spectrum in the form of first derivative values is correlated with the library spectra, the target spectrum is divided into peak regions by searching for all spectra locations with a positive first derivative followed immediately by a negative first derivative value. This sequence indicates a local minimum in the non-differentiated absorbance spectrum of log (1/R) values. The peak regions in the first derivative spectrum extending between these local minima in the undifferential target spectrum are correlated with first derivatives of the sample spectra as described above in connection with the log (1/R) absorbance spectra.

The instrument of the invention is described above as making reflectance measurements to determine the absorbance, log (1/R), spectra. The absorbance spectra may also be determined from transmittance measurements of T wherein the absorbance is log (1/T).

As described above, the system of the invention analyzes an unknown material from its absorbance spectrum. Absorbance measurements are used because they generally are proportional to concentrations of constituents or properties of the material. It will be apparent that the system is applicable to representations of the reflectance or transmission measurements in other forms. These and many other modifications may be made to the above-described specific embodiments of the invention, without departing from the spirit and scope of the invention, which is defined in the appended claims. ##SPC1##