JP2006112996A - Near-infrared spectroscopic analyzer - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a near-infrared spectroscopic analyzer that is enhanced in the accuracy of the calibration curve. <P>SOLUTION: The near-infrared spectroscopic analyzer is constituted so as to input the sample spectrum measured by a spectrum measuring means, to apply it to the calibration curve for calculating property values and equipped with the spectrum-measuring means for measuring the spectrum, a specific value resolving means for inputting the measured spectrum to resolve specific values, a spectrum deciding means for deciding whether the resolved spectrum enters a predetermined range, a spectrum conversion means for inputting the spectrum decided as being within the predetermined range to perform spectral conversion and a calibration curve calculating means for inputting the spectrum decided as being out of the predetermined range, and the spectrum subjected to spectral conversion for calculating the calibration curve. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a near-infrared spectroscopic analyzer suitable for online measurement of component concentrations and physical characteristics of heavy chemical industrial processes, and in particular, calibration by appropriately converting a specific sample spectrum in the measured spectrum. The present invention relates to a near-infrared spectroscopic analyzer that improves the output accuracy of a line.

  For example, ethylene vinyl acetate copolymer (EVA) is a general plastic, but its physical properties greatly change depending on the content of vinyl acetate (VA). Therefore, many products with different VA contents are manufactured. Measuring the VA content quickly has a great effect from the viewpoint of reducing manufacturing costs and quality control.

  As a prior art for detecting that an outlier state has occurred due to a process abnormality or the like, for example, there are the following.

JP-A-9-281042

FIG. 12 is a block diagram showing the configuration of the near-infrared spectrometer described in the above publication.
In the figure, a near-infrared (NIR) spectrometer 10 measures the absorbance of near-infrared light having a wavelength band of 0.9 to 2.5 μm with respect to a sample, and outputs a measurement spectrum.
The applied calibration curve storage unit 20 stores parameters representing the statistical characteristics of the sample spectrum group when creating the calibration curve. For example, parameters representing the statistical characteristics of PLS (partial least squares) regression analysis include the following.
(1) Number of factors (2) Average absorbance spectrum at the time of calibration curve creation (3) Loading weight for calibration curve factor (4) Loading for calibration curve factor (5) Standard deviation of calibration curve score (6) Threshold value for outlier range

Here, the number of factors is a kind of mathematical parameter determined so that the prediction accuracy of PLS regression analysis is high, and is also called the number of potential factors. The loading weight is a kind of mathematical parameter calculated on the PLS regression analysis algorithm, and is also called a load vector. Loading is a type of mathematical parameter calculated on the PLS regression analysis algorithm. The score of the calibration curve is a kind of moment representing the degree to which the measurement spectrum varies with respect to the loading.
The measurement spectrum is the i-th measurement spectrum measured on-line by the near-infrared spectrometer, and the total number of measurement spectra is N.

  The outlier determination unit 30 determines whether a sample spectrum is included in the sample spectrum group of the calibration curve stored in the applied calibration curve storage unit 20. There are several outlier criteria, such as Mahalanobis distance, RMSSR (Root Mean Square Spectral Residual), NND (Nearest Neighbor Distance), etc. Calculate the Mahalanobis general distance (a measure of the distance in a space defined by two or more correlated variables) with the sample spectrum, and if this Mahalanobis general distance is within the threshold, the sample spectrum Is included in the sample spectrum group of the calibration curve, and if the Mahalanobis's pan-distance exceeds a threshold value, it is determined that the sample spectrum is an outlier.

  The property value calculation unit 40 calculates the property value by applying the calibration curve to the sample spectra that are recognized by the outlier determination unit 30 as being included in the sample spectrum group of the calibration curve. The alarm information adding unit 50 attaches a signal indicating that the sampler is outlier to the sample spectrum recognized by the outlier determination unit 30 as not being included in the sample spectrum group of the calibration curve.

FIG. 13 is an explanatory diagram of an outlier.
In the figure, the horizontal axis is the true property value, the lower limit value of the range in which the calibration curve was obtained is YL, the upper limit value is YH, and this intermediate value is YM. The vertical axis is the predicted property value obtained by applying the sample spectrum to the calibration curve. The lower limit value of the range corresponding to the calibration curve is YL *, the upper limit value is YH *, and the intermediate value is YM *. The origin is represented by Y0 and Y0 *. In the figure, the 45-degree line is a calibration curve. The broken lines drawn on the upper and lower sides of the calibration curve are boundary lines for distinguishing between normal samples and outliers, and the width is determined by the magnitude of the threshold value.

  This drawing is divided into three regions. The first is an interpolation area, which is between the lower limit value YL and the upper limit value YH of the range in which the calibration curve is obtained, and is surrounded by upper and lower broken line boundary lines. The second is an extrapolation region, which is a lower limit value YL or more than the upper limit value YH of the range in which the calibration curve is obtained, and is a range surrounded by upper and lower dashed boundary lines along the calibration curve. The third is the outlier region, which is a range that deviates from the upper and lower broken line boundaries. Here, samples 1 to 14 are plotted, samples 7, 10, 13, and 14 exist in the outlier region, samples 4 and 6 exist in the extrapolation region, and other samples exist in the interpolation region. Existing.

  Such a conventional calibration curve creation method can be schematically represented as shown in FIGS. That is, the conventional example of FIG. 14 is a method of creating a calibration curve by excluding the spectrum determined to be an outlier, and the conventional example shown in FIG. 15 is a method of calculating a wide calibration curve for each range.

The outlier discriminating method as described above is a method for discriminating whether or not an outlier is present, and it cannot be discriminated to which sample an unknown sample applies. In addition, a discrimination method such as SIMCA using multivariate analysis has many functions and excellent discrimination capability, but a simple discrimination method has been desired for online use.
An object of the present invention is to provide a near-infrared spectroscopic analyzer with online discrimination having a spectrum conversion function.

In order to achieve such a problem, the invention according to claim 1 of the present invention is:
In the near-infrared spectroscopic analyzer that inputs the sample spectrum measured by the spectrum measuring means and obtains the property value by applying it to the calibration curve,
Spectrum measuring means for measuring the spectrum, singular value decomposition means for inputting the measured spectrum and performing singular value decomposition, and spectrum determining means for determining whether the decomposed spectrum is within a predetermined range or outside the predetermined range A spectrum conversion means for inputting a spectrum determined to be within a predetermined range and converting the spectrum; a spectrum determined to be out of the predetermined range; A calibration curve calculating means for calculating is provided.

The invention described in claim 2 is the near-infrared spectrometer according to claim 1,
In the spectrum conversion, a predetermined process is performed on all of the spectra determined to be within a predetermined range.

The invention described in claim 3 is the near-infrared spectroscopic analyzer according to claim 1 or 2,
The method of changing the spectrum is characterized by using one or more of the spectral pre-processing for a particular sample spectrum, including wavelength shift, scale conversion, offset, differentiation.

The invention according to claim 4 is the near infrared spectroscopic analyzer according to claims 1 to 3,
An ethylene vinyl acetate copolymer is used as a sample.

The present invention has the following effects.
According to invention of Claims 1-4,
In the near-infrared spectroscopic analyzer that inputs the sample spectrum measured by the near-infrared spectroscopic unit and obtains the property value by applying it to the calibration curve.
A spectrum determining unit that performs singular value decomposition on the spectrum and determines whether the score having the calculated number of specific factors is within a predetermined range or outside the predetermined range; and a spectrum determined to be within the predetermined range Calibration is provided by including a spectrum converter that inputs and converts a spectrum, and a calibration curve generator that generates a calibration curve by inputting the spectrum determined to be out of the predetermined range and the spectrum that has been subjected to the spectrum conversion. A near-infrared spectroscopic analysis apparatus with improved line accuracy can be realized.

  FIG. 11 is a chemical formula of EVA in which polyethylene and polyvinyl acetate used in the present invention are randomly bonded. The VA content of EVA is an important index for managing its mechanical strength.

FIG. 1 is a block diagram showing an embodiment of the present invention.
The spectrum measured by the spectrum measuring unit 130 is input to the singular value decomposing unit 131 and subjected to singular value decomposition and then input to the spectrum determining unit 132. The spectrum determination unit 132 makes a determination based on a score plot based on the input spectrum, and determines whether the score plot is a target sample within a predetermined range or a non-target sample outside a predetermined range.

  The spectrum determined by the spectrum determination unit 132 to be within a predetermined range is input to the spectrum conversion unit 133. The spectrum conversion unit 133 performs a predetermined process (for example, shifts the wavelength) on the spectrum for which the score plot is determined to be within a predetermined range.

  The calibration curve calculation means 134 creates a calibration curve by inputting the spectrum determined to be outside the predetermined range and the spectrum subjected to spectrum conversion.

Next, examples carried out using samples will be described.
FIG. 2 shows EVA samples (Sample Nos. S1 to S9) used in the experiments of the present invention, and VA (w / w%) shows the weight ratio of vinyl acetate to polyethylene. Here, nine types of samples having a minimum w / w% = 0 as sample 1 and a maximum w / w% = 41.1 as sample 9 were used.

FIG. 3 shows near-infrared spectra measured by melting nine types of EVA samples shown in FIG. 2 at 160 ° C., respectively. The horizontal axis represents the wave number (cm ⁻¹ ), and the vertical axis represents the absorbance.
FIG. 4 shows a partially enlarged view of the spectrum of FIG. 3 and shows that the peak height of the spectrum varies depending on the VA content in EVA. The peak at 4679 cm ⁻¹ is smaller as the VA content in EVA is lower, and the peak is larger as the VA content is higher.

FIG. 5 summarizes such changes in a diagram. In addition, since there may be no peak at a VA content of zero, VA = 6.3% indicated by (D) in the figure is used as a reference.
Each spectrum peak is attributed to a chemical functional group. The mark ● represents a peak attributed to CH ₂ , the mark ◯ represents CH ₃ , C═O, CH and C═O bond sound, and the like. According to the figure, it can be seen that the shift in wavelength changes depending on the attribution of the peak.

  FIG. 6A shows the SECV (Standard Error of Cross Validation) and SEC (Standard Error of Calibration) of each calibration curve in which the spectrum of the S1 sample farthest from the population is wavelength-shifted and the spectrum is used instead of the S1 sample spectrum. ).

In FIG. 6A, the vertical axis represents SECV / SEC (%), and the horizontal axis represents the wavelength shift amount. The point indicated by (A) in the figure indicates the error in the calibration curve creation when the sample spectrum of S1 is shifted by 10 cm ^{−1 and} the same one of the remaining sample spectra S1 to S9 is used. It shows the error of the calibration curve when shifting back and forth.

In FIG. 6B, when the wave number shift is 0, the SECV is 0.521%, the SEC is 0.423%, and the correlation function (Correlation Coefficient) is 0.999475. When the wave number is shifted by 0 ± 12 cm ⁻¹ (2, 4, 6... 12 on the + side and 3, 5, 6... 12 on the − side), the best wave number is −8 cm ^{−. In the case of 1} shift, the SECV is 0.35% and the SEC is 0.323%, and the correlation function is improved.
FIG. 7 shows a part (* part) of the spectrum 5500 to 4500 cm ⁻¹ before and after the wavelength shift enlarged in the lower right, and the spectrum indicated by (R) is the original spectrum indicated by (H). The state shifted by 8 cm ⁻¹ is shown.

  FIG. 8 shows the score analysis result of the near-infrared spectrum of EVA measured continuously. It can be seen that an aggregate (cluster) is formed for each of the samples S1 to S9. The ◆ marks scattered in the middle are the scores in the middle of converting the brand into another sample. S1 having zero VA amount (polyethylene) is greatly deviated from the group.

FIG. 9 shows the score analysis results before and after shifting the S1 sample from the position (f) to the position (g) by −8 cm ⁻¹ wavelength.
It can be seen that the position of the aggregate (cluster) was changed by the wavelength shift.

  FIG. 10 shows a real-time calibration curve output result when the analysis system shown in FIG. 1 is used. The ● mark indicates the result of the calibration curve calculation using an analysis system without a spectrum judgment / conversion function. A circle indicates the result of the calibration curve calculation by the system of the present invention having the spectrum determination / conversion function, and the measured sample spectrum is converted when the reference value of the score plot is exceeded.

In this example, the sample S1 is shifted by −8 cm ^{−1 in} accordance with the result of FIG. 6 determined by the spectrum determining means 132 (see FIG. 1), and then the calibration curve calculation is performed. Since the spectrum of sample S1 shifted by −8 cm ⁻¹ is used when preparing the calibration curve, the SEC and SECV values of the entire calibration curve are improved as shown in FIG. 6, and the predicted values of S3, S2 and S1 are also obtained. It is better than the mark ● and closer to the value in Fig. 2.

That is, paying attention to the sample of S1, the result of the calibration curve calculation by the analysis system without the spectrum determination / conversion function (● mark) shows that the VA value has an error of about 0.8%, but the spectrum determination / conversion The result of the calibration curve calculation (○ mark) in the system having a function is about 0%, and there is almost no error with respect to 0% of the value in FIG.
Focusing on the sample of S2, the result of the calibration curve calculation by the analysis system without the spectrum judgment / conversion function (marked with ●) shows that an error of 0.2% occurs when the VA value is about 6.1%. On the other hand, the result of the calibration curve calculation by the system having the spectrum determination / conversion function (◯ mark) is about 6.3%, and there is almost no error with respect to 6.3% of the value of FIG.

  Focusing on the sample of S3, the result of the calibration curve calculation by the analytical system without the spectrum determination / conversion function (marked with ●) is that the VA value is about 9.6%, and an error of 4% occurs. The result of the calibration curve calculation by the system having the spectrum determination / conversion function (marked by ○) is about 9.8%, and the error is as small as 0.2% with respect to 10% of the value of FIG.

As described above, according to the present invention, the accuracy of the calibration curve can be improved by finding a sample affecting the calibration curve in the sample set for creating the calibration curve and shifting the wavelength.
The above description merely shows a specific preferred embodiment for the purpose of explanation and illustration of the present invention. In this embodiment, wavelength shift is used as a means for transforming the spectrum. For example, one or more kinds of spectrum preprocessing such as multiplication, addition, baseline correction, scale conversion, offset, and differentiation are used for the specific sample spectrum. Can do.

In the present invention, a PLS score plot is used as a sample discrimination method. However, if there is a factor representing a feature, a score of principal component analysis may be used.
In the present invention, factor 1 and factor 2 are used for the score, but other factors may be used as long as there is a factor representing the feature.

In addition, three or more factors are used when distinguishing samples.
Factor> 1
Factor> 0.3
Factor> 5
It is also possible to target a thing that satisfies or two or more.
It is also possible to convert multiple samples into different forms.
Therefore, the present invention is not limited to the above-described embodiments, and includes many changes and modifications without departing from the essence thereof.

It is a figure which shows the structure of the near-infrared spectroscopy analyzer of this invention. It is a figure which shows the sample used for experiment of this invention. It is a figure which shows the near-infrared spectrum which melted and measured the sample of EVA. It is a figure which shows the elements on larger scale of the spectrum of FIG. It is a figure which shows the state which the peak position is changing with EVA content. It is a figure which shows the influence of SEC and SECV by the wavelength shift of S1 sample. It is a figure which shows the state which expanded a part of spectrum before and behind wavelength shift. It is a figure which shows the score analysis result of the near-infrared spectrum of EVA measured continuously. It is a figure which shows the score analysis result before and behind shifting the wavelength of S1 sample. It is a figure which compares and shows the real-time calibration curve output result at the time of using the near-infrared spectroscopy analyzer of this invention, and the conventional apparatus. It is a figure which shows the chemical formula of EVA which polyethylene and polyvinyl acetate couple | bonded at random. It is a figure which shows the structure of the conventional near-infrared spectrometer. It is explanatory drawing of an outlier. It is explanatory drawing which shows the analytical curve preparation method using the conventional near-infrared spectrometer. It is explanatory drawing which shows the analytical curve preparation method using the conventional near-infrared spectrometer.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Near-infrared analyzer 20 Applicable calibration curve memory | storage part 30 Outlier determination part 40 Property value calculation part 50 Alarm information addition part 130 Spectrum measurement means 131 Singular value decomposition means 132 Spectrum determination means 133 Spectrum conversion means 134 Calibration curve calculation means

Claims (4)

In the near-infrared spectroscopic analyzer that inputs the sample spectrum measured by the spectrum measuring means and obtains the property value by applying it to the calibration curve,
Spectrum measuring means for measuring the spectrum, singular value decomposition means for inputting the measured spectrum and performing singular value decomposition, and spectrum determining means for determining whether the decomposed spectrum is within a predetermined range or outside the predetermined range A spectrum conversion means for inputting a spectrum determined to be within a predetermined range and converting the spectrum; a spectrum determined to be out of the predetermined range; A near-infrared spectroscopic analyzer comprising a calibration curve calculation means for calculating.   The near-infrared spectroscopic analysis apparatus according to claim 1, wherein the spectrum conversion performs a predetermined process on all of the spectra determined to be within a predetermined range.   3. The near-infrared spectroscopy according to claim 2, wherein the spectrum conversion method includes wavelength shift, difference spectrum, scale conversion, offset, and differentiation, and uses one or more of spectrum pretreatments for a specific sample spectrum. Analysis equipment. 4. The near-infrared spectroscopic analyzer according to claim 1, wherein an ethylene vinyl acetate copolymer is used as a sample.

Images