METHOD FOR CALIBRATING A PLURALITY OF IDENTICAL SPECTROMETERS
Patent Information
- Application Number
- DE502021008243
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2020-06-18
- Filing Date
- 2021-06-08
- Publication Date
- 2025-08-28
- Estimated Expiration
- 2041-06-08
AI Technical Summary
Calibrating a large number of identical spectrometers for ingredient analysis is complex due to manufacturing variations, requiring extensive effort and chemical analysis, which is impractical for miniaturized sensors used in process engineering applications.
A method involving a regression model is developed to calibrate multiple identical spectrometers by simulating error spectra based on a mathematical model of their components, improving robustness against manufacturing tolerances without individual measurements.
The method enhances Inter-Instrument Agreement (IIA) by up to 30% with reduced effort, enabling efficient calibration of large numbers of spectrometers without additional samples or adjustments.
Description
[0001] The present invention relates to a method for calibrating a plurality of identical spectrometers for ingredient analysis. Calibration is intended to enable the concentration of a specific ingredient in samples to be measured to be accurately measured with the individual spectrometers. The calibration process includes at least the provision of a regression model, which is to be used to determine the concentrations of the ingredient based on the sample spectra measured with the spectrometers.
[0002] DE 601 14 036 T2 discloses a method for characterizing spectrometer instruments according to the instrumental variation present between the instruments or according to the variation over time within the same instrument. A plurality of spectra of known standards from at least one spectrometer instrument is provided. The at least one spectrometer instrument is classified into at least one of a plurality of predefined clusters based on spectral features extracted from the at least one spectrum. At least one calibration model is provided for each of the predefined clusters. Each calibration model compensates for the instrumental variation of instruments classified into the respective cluster.
[0003] In the article by Workman JR, JJ: "A Review of Calibration Transfer Practices and Instrument Differences in Spectroscopy" in Applied Spectroscopy, Vol. 72(3), 2018, pages 340 to 365, DOI: 10.1177 / 0003702817736064, an overview of procedures for transferring calibration models between spectrometers is given.
[0004] DE 696 08 252 T2 shows a method for standardizing a large number of spectrometers.
[0005] From EP 1 998 155 A1 a method for wavelength calibration of a spectrometer is known, which is based on the principle of a step-like relative shift of corresponding measured value blocks of a model and calibration spectrum.
[0006] DE 101 52 679 A1 teaches a method for the fully automatic transfer of calibrations of optical emission spectrometers.
[0007] DE 10 2004 061 178 A1 describes a method for the fully automated transfer of calibrations between spectrometers. The spectrometers include spectrometer optics with positionable slits.
[0008] DE 692 11 163 T2 discloses a method for calibrating a spectral instrument for determining a value of a property of an unknown sample by using constants in a second calibration equation for the instrument by reference to a spectral instrument having a first calibration equation. From the first calibration equation, a value of a dependent first variable corresponding to each term of a set of
[0009] Calibration transfer standards are determined based on first spectral data for each member of the set of calibration transfer standards. Second constants in the second calibration equation are determined such that the sum of the absolute differences between each stated value of the independent second variable for each member of the set of calibration transfer standards using second spectral data measured by the instrument and the corresponding value of the dependent first variable determined for each member of the set of calibration transfer standards determined by the first calibration equation is minimized. The values of the stated property in the unknown sample are determined using the instrument.
[0010] Calibrating a spectrometer for ingredient analysis is complex due to the use of a reference measurement method to determine the concentration of the respective ingredient, as the reference measurement method usually requires chemical analysis. Therefore, it is desirable not to have to calibrate each individual example of a mass-produced spectrometer type for ingredient analysis. To achieve this, a calibration performed on a sample basis must be designed to be robust against small differences in spectrometer manufacturing. This issue is becoming increasingly important for manufacturers and calibrators, particularly given that miniaturized spectrometers are increasingly being used as sensors. With large quantities of spectrometers used as measurement sensors, it is not possible to calibrate each individual product on each individual device against the reference measurement method performed in a laboratory.According to the state of the art, a large number of spectrometers are used for calibrations to minimize the influence of individual instruments. To achieve a high inter-instrument agreement (IIA), a selection of spectrometers must be available for calibration work, which involves considerable effort. However, in process engineering applications where samples are taken from the process after spectrum acquisition, this is virtually impossible. Furthermore, it is not possible to subsequently optimize calibrations with a view to improving the IIA, as this would require new samples and a large number of spectrometers.
[0011] The article by Despagne, Frédéric et al., "Transfer of Calibrations of Near-Infrared Spectra Using Neural Networks," in Applied Spectroscopy, The Society for Applied Spectroscopy, Baltimore, USA, Vol. 52, No. 5, 1998, pages 732 to 745, describes an approach for multivariate instrument standardization based on neural networks. According to this approach, spectral differences between two instruments are mathematically modeled.
[0012] The object of the present invention is, based on the state of the art, to be able to improve the Inter-Instrument Agreement (IIA) of spectrometers for ingredient analysis with little effort.
[0013] The stated object is achieved by a method according to the appended claim 1.
[0014] The method according to the invention serves to calibrate a plurality of identical spectrometers. The identical spectrometers are used for ingredient analysis. Calibration is carried out with regard to ingredient analysis, i.e., the determination of the concentration of a constituent of a sample based on a spectrum measured on the sample. The spectrometers are identical at least with regard to measurement signal acquisition and processing, although the spectrometers may differ with regard to properties that do not influence measurement signal acquisition and processing, such as the design of a housing or a connector. The identical spectrometers are preferably of the same type.
[0015] Examples of spectrometers include NIR spectrometers, VIS / NIR spectrometers, or full-range spectrometers. These spectrometers can have a transmission, transflectance, or reflection design. Examples of spectrometers include monochromator-based spectrometers, interferometer-based spectrometers, filter-based spectrometers, or MEMS spectrometers.
[0016] The calibration process includes at least providing a regression model to be used to determine the concentrations of the constituent based on the sample spectra measured with the spectrometers. In this respect, the method serves at least to provide a regression model for calibrating a plurality of identical spectrometers for constituent analysis.
[0017] In one step of the method, a large number of samples are provided. The samples contain an ingredient in different concentrations. This is the ingredient for whose measurement the spectrometers are to be calibrated. Preferably, at least ten of the samples are provided and more preferably at least 100 of the samples are provided. Preferably between 50 and 500 of the samples are provided. The samples were taken from a material or product whose contents are to be analyzed by the spectrometers. The material or product is preferably an agricultural product, a foodstuff, or a foodstuff. The agricultural product is preferably a harvested crop.The spectrometers are preferably designed to measure the concentration of at least one ingredient of the agricultural product or food on an agricultural vehicle, in an agricultural machine, or in a food production process. The ingredient is preferably water, a protein, an oil, sugar, starch, or crude fiber. Measuring the water concentration represents a moisture measurement.
[0018] In a further step of the process, the concentrations of the constituents in the individual samples are measured using a reference measurement method, thus obtaining reference measurement values for the concentrations of the constituents. The reference measurement method leads to precise measured values of the concentrations of the constituents in the individual samples. The reference measurement method is preferably formed by a chemical analysis method. The chemical analysis method is preferably a wet chemical analysis method.
[0019] In a further step of the process, spectra of the individual samples are measured using a spectrometer selected from a group of identical spectrometers. The spectra are preferably absorption, transmission, and / or reflection spectra. The spectra are measured for each sample using the same measurement principle.
[0020] The spectra are preferably measured in the infrared range, the visible range, and / or the ultraviolet range of the electromagnetic spectrum. The spectra are preferably measured in the near infrared range of the electromagnetic spectrum. The spectra are measured over a wavelength range that is preferably at least 100 nm and more preferably at least 300 nm. The spectra are preferably measured by measuring amplitudes at selected wavelengths. Preferably, the amplitudes are measured at at least 10 selected wavelengths. More preferably, the amplitudes are measured at at least 100 selected wavelengths. More preferably, the amplitudes are measured at at least 300 selected wavelengths.
[0021] The spectrometer selected as an example is identical to the other spectrometers of the same design. The spectrometer selected as an example preferably has a medium error.
[0022] In a further step of the method, a preliminary regression model is determined, in which the amplitude values of the measured spectra form the independent variables and the reference measured values of the ingredient concentrations form the dependent variable. The result is a mathematical function that describes the dependence of the ingredient concentration on the amplitude values of the spectrum. The preliminary regression model preferably has at least one additional dependent variable that describes a physical or chemical property of the samples.
[0023] In a further step of the method, at least one spectrum is selected from a set of spectra, this set comprising the measured spectra and a mean spectrum formed from the measured spectra. The selection is made according to their suitability for predicting the concentration of the constituent, particularly in the respective sample. Preferably, one to ten spectra are selected from the measured spectra and / or the mean spectrum formed from the measured spectra.
[0024] The step of selecting at least one spectrum from the measured spectra and / or a mean spectrum formed from the measured spectra preferably comprises one or both of the two sub-steps described below. According to one sub-step, the preliminary regression model is applied to the measured spectra in order to obtain a predicted value for the concentration of the ingredient in the respective sample, wherein the measured spectrum or those measured spectra are selected for which the predicted value is closest to the reference measured value of the respective sample. Thus, those measured spectra are selected which lie as close as possible to the regression line of the preliminary regression model. According to another sub-step, the mean spectrum is selected.The mean spectrum is preferably formed from the measured spectra by averaging the amplitude values of the measured spectra at each of the selected wavelengths, for example, by arithmetic averaging. A reference measurement value for the concentration of the constituent, determined by applying the preliminary regression model to the mean spectrum, is assigned to the mean spectrum.
[0025] In a further step of the process, tolerances of components of identical spectrometers are simulated multiple times using a mathematical model of the identical spectrometers to obtain a variety of error spectra. The mathematical model of the identical spectrometers describes the components of the identical spectrometers, particularly the optical and electronic components. The mathematical model describes error properties, namely the tolerances that occur in the components, so that the mathematical model is suitable for calculating a maximum error for the accuracy of each wavelength and the associated amplitude of transmission, reflection, and / or absorption. Based on these maximum errors for the individually selected wavelengths, the possible error spectra for the ideal spectrometer are simulated. The error spectra represent difference spectra.Preferably, at least 100 simulations are performed, so that at least 100 error spectra are obtained. More preferably, at least 500 simulations are performed, so that at least 500 error spectra are obtained.
[0026] In a further step of the method, the individual error spectra are each added to the selected spectrum or to the individually selected spectra in order to obtain simulated spectra. The simulated spectra are spectra that could actually be measured with identical spectrometers, since they are based on spectra actually measured with identical spectrometers and also contain an error component that was determined by simulation using the mathematical model of the identical spectrometers. The addition of the individual error spectra to the selected spectrum or to the individually selected spectra is preferably carried out for the amplitude values of the selected wavelengths. If the number of error spectra is, for example, 1,000 and the number of selected spectra is, for example, one, the number of simulated spectra obtained is 1 x 1,000 = 1,000.For example, if the number of error spectra is 1,000 and the number of selected spectra is five, the number of simulated spectra obtained is 5 * 1,000 = 5,000.
[0027] In a further step of the process, the preliminary regression model is applied to the simulated spectra to obtain a predicted value for the concentration of the ingredient. This results in a predicted value for the concentration of the ingredient for each of the simulated spectra. The simulated spectra are each assigned the reference measured value for the concentration of the ingredient that was determined for the selected spectrum underlying the respective simulated spectrum.
[0028] In a further step of the method, a number of simulated spectra are selected, with the prediction values obtained for the selected simulated spectra reflecting a spread of the prediction values obtained for the simulated spectra. Thus, those simulated spectra are selected that are suitable for contributing to improving the preliminary regression model. The number of selected simulated spectra is preferably between 50 and 1,000.
[0029] In a further step of the process, a resulting regression model is determined, in which the amplitude values of the measured spectra and the selected simulated spectra form the independent variables, and in which the reference measured values of the ingredient concentrations form the dependent variable. The determination of the resulting regression model differs from the determination of the preliminary regression model only in that the amplitude values of the selected simulated spectra are also taken into account. The resulting regression model is therefore an improvement of the preliminary regression model, so that the resulting regression model represents a calibration that is more robust against errors, especially against differences due to manufacturing tolerances of identical spectrometers.
[0030] The independent variables of the preliminary regression model and the independent variables of the resulting regression model are preferably the amplitude values of the measured or selected simulated spectra at selected wavelengths. The independent variables of the preliminary regression model and the independent variables of the resulting regression model are preferably the amplitude values of the measured or selected simulated spectra of the same selected wavelengths. The preliminary regression model and the resulting regression model are preferably each formed by a multilinear regression model. The preliminary regression model and the resulting regression model are preferably each determined by partial least squares regression.
[0031] Preferably, the number of selected simulated spectra is at most 50% of the number of measured spectra. For example, the number of selected simulated spectra is approximately 75, while the number of measured spectra is approximately 150. This achieves significantly increased robustness of the resulting regression model compared to the preliminary regression model.
[0032] A particular advantage of the described method is that the calibration of identical spectrometers for ingredient analysis can be improved through a comparatively low-effort simulation. No measurements are required with individual identical spectrometers to calibrate them individually for ingredient analysis. Using the mathematical model takes the error characteristics of identical spectrometers into account. The spread of identical spectrometers is reduced by up to 30%, depending on the product and ingredient, when using the same model without further adjustments.
[0033] In preferred embodiments, the method is further designed to utilize the calibration or the resulting regression model. For this purpose, it comprises a further step in which the resulting regression model is used in the individual identical spectrometers in order to determine the concentration of the ingredient in a sample based on the spectrum of the sample measured with the respective spectrometer. For this purpose, the resulting regression model is to be loaded into the individual spectrometers in the form of software that describes the resulting regression model as a mathematical relationship between the concentration of the ingredient and the measured spectrum. Alternatively, the resulting regression model is preferably used in a network. The network comprises the individual identical spectrometers and at least one computing unit. The network represents a data network.The spectrometers and the at least one computing unit are interconnected via data links. The at least one computing unit is preferably formed by a computer, in particular by a server. The network preferably comprises several of the computers. The at least one computing unit is used to determine the concentrations of the constituents in the samples based on the spectra of the samples measured with the spectrometers. Thus, the concentrations are preferably determined using cloud computing. For this purpose, the resulting regression model is transferred to the at least one computing unit in the form of software that describes the resulting regression model as a mathematical relationship between the concentration of the constituent and the measured spectrum.
[0034] Outside the scope of the claims, a plurality of identical spectrometers for ingredient analysis are further disclosed. The individual spectrometers are each configured to determine the concentration of an ingredient in a sample based on a spectrum measured with the respective spectrometer. The individual spectrometers are calibrated according to the method described above by defining a relationship between a measured value to be determined for the concentration of the ingredient and the measured spectrum using the resulting regression model. Preferably, the resulting regression model was determined using one of the preferred embodiments of the method described above.
[0035] The plurality of identical spectrometers preferably comprises at least 100 of the identical spectrometers and more preferably at least 1,000 of the identical spectrometers.
[0036] The spectrometers preferably also have features described above in connection with the method.
[0037] Further details and developments of the invention will become apparent from the following description of preferred embodiments of the invention, with reference to the drawings. They show: Fig. 1: a diagram illustrating error spectra generated according to a preferred embodiment of a method according to the invention; and Fig. 2: a diagram illustrating predicted values determined according to the prior art and according to a preferred embodiment of the method according to the invention.
[0038] Fig. 1shows a diagram illustrating error spectra O1 that were generated according to a preferred embodiment of a method according to the invention. The error spectra O1 were each generated by simulating component tolerances of identical spectrometers to be calibrated for ingredient analysis using a mathematical model, wherein the mathematical model depicts one of the many identical spectrometers. The x-axis of the diagram represents the wavelength λ. The y-axis of the diagram shows the deviation, i.e. the error, that results from the respective application of the mathematical model in comparison to a predetermined wavelength value. An ideal spectrometer would have an error of zero. The diagram also shows a 3a deviation O2 resulting from the mathematical model.The error spectra 01 are used to develop a regression model for calibrating the many identical spectrometers to be calibrated for ingredient analysis that is more robust against spectrometer errors.
[0039] Fig. 2shows a diagram illustrating predicted values 03, which were determined according to the prior art, and predicted values 04, which were determined according to a preferred embodiment of the method according to the invention. The plurality of spectrometers is plotted on the x-axis. The magnitude of the predicted value of the concentration of an ingredient, which was measured with the respective spectrometer, is plotted on the y-axis. The predicted values 03, 04 are each sorted according to the magnitude of their deviation. The predicted values 04, which were determined according to the preferred embodiment of the method according to the invention, illustrate that the Inter-Instrument Agreement (IIA) has been significantly improved compared to the prior art. List of reference symbols
[0040] 01Error spectra 023σ-deviation 03Prediction values according to the prior art 04Prediction values according to the invention
Claims
1. A method for providing a regression model for calibrating a plurality of structurally identical spectrometers for ingredient analysis; comprising the following steps: - providing a multiplicity of samples which contain an ingredient in different concentrations; - measuring the concentrations of the ingredient in the individual samples using a reference measurement method to obtain reference measurement values for the concentrations of the ingredient; - measuring spectra of the individual samples with a spectrometer selected by way of example from the structurally identical spectrometers; - determining a preliminary regression model in which amplitude values of the measured spectra form the independent variables and in which the reference measurement values of the concentrations of the ingredient form the dependent variable; - selecting at least one spectrum from the measured spectra and / or an average spectrum formed from the measured spectra according to a suitability for predicting the concentration of the ingredient; - simulating tolerances of components of structurally identical spectrometers multiple times with a mathematical model of structurally identical spectrometers to obtain a multiplicity of error spectra (01); - adding the individual error spectra (01) respectively to the selected spectrum or to the individual selected spectra to obtain simulated spectra; - applying the preliminary regression model to the simulated spectra to obtain in each case a prediction value for the concentration of the ingredient; - selecting a number of the simulated spectra, wherein the prediction values obtained for the selected simulated spectra depict a variation of the prediction values obtained for the simulated spectra; and - determining a resulting regression model in which amplitude values of the measured spectra and the selected simulated spectra form the independent variables and in which the reference measurement values of the concentrations of the ingredient form the dependent variable.
2. The method as claimed in claim 1, characterized in that the reference measurement method is a chemical analysis method.
3. The method as claimed in claim 1 or 2, characterized in that the ingredient is water, a protein, an oil, sugar, starch, or a crude fibre.
4. The method as claimed in any of claims 1 to 3, characterized in that the spectra are measured in the infrared range, in the visible range, and / or in the ultraviolet range of the electromagnetic spectrum.
5. The method as claimed in any of claims 1 to 4, characterized in that the spectra of the individual samples are measured with the spectrometer selected by way of example from the structurally identical spectrometers over a wavelength range which is at least 300 nm in size, wherein the spectra are measured by measuring amplitudes at at least 10 selected wavelengths.
6. The method as claimed in claim 5, characterized in that the independent variables of the preliminary regression model are each formed by the amplitude values of the measured spectra at the selected wavelengths, and in that the independent variables of the resulting regression model are each formed by the amplitude values of the selected simulated spectra at the selected wavelengths.
7. The method as claimed in any of claims 1 to 6, characterized in that the independent variables of the preliminary regression model and the independent variables of the resulting regression model are formed by amplitude values of the spectra at selected wavelengths.
8. The method as claimed in any of claims 1 to 7, characterized in that the preliminary regression model and the resulting regression model are each formed by a multilinear regression model.
9. The method as claimed in any of claims 1 to 8, characterized in that the step of selecting at least one spectrum from the measured spectra and / or an average spectrum formed from the measured spectra comprises one or both of the following two sub-steps: - applying the preliminary regression model to the measured spectra in order to obtain in each case a prediction value for the concentration of the ingredient in the respective sample, wherein that measured spectrum or those measured spectra for which the prediction value comes closest to the reference measurement value of the respective sample is / are selected; - selecting the average spectrum.
10. The method as claimed in any of claims 1 to 9, characterized in that one to ten spectra are selected from the measured spectra and / or the average spectrum formed from the measured spectra.
11. The method as claimed in any of claims 1 to 10, characterized in that at least 100 simulations for simulating tolerances of components of the structurally identical spectrometers are carried out using the mathematical model of the structurally identical spectrometers, with the result that at least 100 of the error spectra (01) are obtained.
12. The method as claimed in any of claims 1 to 11, characterized in that the number of the simulated spectra selected is between 50 and 1,000.
13. The method as claimed in any of claims 1 to 12, characterized in that it comprises the following further step: - using the resulting regression model in the individual structurally identical spectrometers in order to determine the concentration of the ingredient of a sample with the spectrometers based on the spectrum of the sample measured with the respective spectrometer.
14. The method as claimed in any of claims 1 to 12, characterized in that it comprises the following further step: - using the resulting regression model in a network which comprises the individual structurally identical spectrometers and at least one computing unit, wherein the at least one computing unit is used to determine the concentrations of the ingredient of the samples based on the spectra of the samples measured with the spectrometers.