A screening method for near-infrared model calibration sample sets

Abstract

The application discloses a kind of near-infrared model correction sample set screening method, comprising: determining basic sample set and candidate sample set based on each to be tested sample;According to the near-infrared spectrum and concentration of the component to be tested of each sample in basic sample set, the near-infrared spectrum matrix and the first concentration vector of basic sample set are constructed;Based on the near-infrared spectrum matrix and the first concentration vector, the pure spectrum signal of each sample in candidate sample set is determined;The Euclidean distance between the pure spectrum signal of each sample in candidate sample set is calculated respectively, and each sample in candidate sample set is sorted based on each Euclidean distance;Based on basic sample set and sorted candidate sample set, correction sample set is constructed.This scheme considers the spectrum information and the concentration information of component to be tested of sample simultaneously, improves the accuracy of sample selection, and in the process of selecting correction sample, only the selected part of sample is determined the concentration value of component to be tested, greatly reduces the workload of component to be tested chemical value detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of near infrared spectrum analysis, in particular to a screening method of near infrared model calibration sample set. BACKGROUND

[0002] The near infrared spectrum analysis technology is a secondary analysis technology (also known as indirect analysis technology), which needs a large number of samples to construct a near infrared spectrum quantitative analysis model by means of multivariate correction technology. The selection of calibration samples is of great significance to the establishment and application of the near infrared spectrum quantitative analysis model. Generally speaking, selecting representative samples for modeling can not only reduce the workload of chemical analysis, but also directly affect the applicability and accuracy of the established model, which is an important factor for improving efficiency and reducing cost in near infrared analysis technology.

[0003] The current common sample selection methods include random method, Kennard-Stone algorithm (abbreviated as KS algorithm), sampleset partitioning based on joint x-y distance algorithm (abbreviated as SPXY algorithm) and the like. However, the above methods have certain limitations, for example, the KS algorithm only considers the spectral difference of the samples, without involving other information of the samples, which affects the accuracy of sample selection to some extent, and the SPXY algorithm considers the spectral and concentration information, but the algorithm needs to obtain the concentration information of all samples, which is large in calculation amount and long in modeling time. SUMMARY

[0004] The main purpose of the present application is to solve the problem of large calculation amount or low accuracy of the existing calibration sample selection method. To achieve the above purpose, the present application provides a screening method of near infrared model calibration sample set, which can balance the accuracy and calculation amount of calibration sample set selection, and further improve the prediction accuracy and modeling efficiency of the model.

[0005] An embodiment of the present application provides a screening method of near infrared model calibration sample set, comprising:

[0006] A sample setting step: determining a basic sample set and a candidate sample set based on each to-be-tested sample;

[0007] A construction step: constructing a near infrared spectrum matrix and a first concentration vector of the basic sample set according to the near infrared spectrum and the concentration of the to-be-tested component of each sample in the basic sample set;

[0008] A pure spectrum calculation step: determining the pure spectrum signal of each sample in the candidate sample set based on the near infrared spectrum matrix and the first concentration vector;

[0009] Sorting steps: Calculate the Euclidean distance between the pure spectral signals of each sample in the candidate sample set, and sort the samples in the candidate sample set based on each Euclidean distance;

[0010] The steps for determining the calibration sample are as follows: Construct a calibration sample set based on the base sample set and the sorted candidate sample set.

[0011] Specifically, the analyte in this method can be a chemical component of tobacco or tobacco products, such as total alkaloids, total sugars, and total nitrogen. The pure spectral signal of the sample can be determined based on the near-infrared spectrum and concentration value of the analyte in the sample.

[0012] Specifically, a predetermined number of samples can be selected from each test sample as the base samples, and the remaining test samples can be used as candidate samples. Furthermore, the selection of the base samples can be based on the KS algorithm to select no fewer than 30 samples from each test sample as the aforementioned base samples.

[0013] This method selects calibration samples based on the pure spectral signal of the analyte, while simultaneously considering both the spectral information of the sample and the concentration information of the analyte, thus improving the accuracy of sample selection. Furthermore, during the selection of calibration samples, this method only measures the concentration of the analyte in a selected subset of samples, significantly reducing the workload of detecting the chemical values ​​of the analyte. Using this screening method, the distribution of calibration samples can be more uniform and representative, thereby enabling the calibration samples to cover as much sample and concentration information as possible, contributing to improved robustness and accuracy of the calibration model.

[0014] As a specific embodiment of the present invention, the step of determining the calibration sample includes:

[0015] A first predetermined number of samples are selected sequentially from the sorted candidate sample set;

[0016] The concentration of the first set number of selected samples is unknown. A calibration sample set is constructed based on the first set number of selected samples and the base sample set.

[0017] In one specific embodiment of the present invention, the samples in the sorted candidate sample set are sorted from largest to smallest according to their Euclidean distance; the sum of the first set number and the number of samples in the basic sample set is equal to 1 / 2 to 2 / 3 of the total number of samples to be tested.

[0018] As a specific embodiment of the present invention, the calibration sample determination step includes:

[0019] A second predetermined number of samples are selected sequentially from the sorted candidate sample set;

[0020] The concentrations of the selected second set number of samples are determined, and the basic sample set and the candidate sample set are updated based on the second set number of samples;

[0021] The steps of constructing, determining and updating are repeated in sequence until the total number of samples in the basic sample set reaches a set value, and then the basic sample set at this time is determined as the calibration sample set.

[0022] As a specific embodiment of the present application, the second set number is greater than or equal to 10, and the set value is equal to 1 / 2-2 / 3 of the total number of samples to be measured.

[0023] As a specific embodiment of the present application, updating the basic sample set and the candidate sample set based on the second set number of samples comprises:

[0024] The selected second set number of samples are added to the basic sample set as an updated basic sample set;

[0025] The selected second set number of samples are deleted from the candidate sample set as an updated candidate sample set.

[0026] As a specific embodiment of the present application, the pure spectrum calculation step comprises:

[0027] The first spectrum matrix is reconstructed based on the near-infrared spectrum matrix, wherein the elements in the same row in the first spectrum matrix represent the spectral data of the same sample;

[0028] A second concentration vector is determined according to the first spectrum matrix and the first concentration vector;

[0029] A second spectrum matrix is determined according to the first spectrum matrix and the second concentration vector, wherein the second spectrum matrix is used to represent a space composed of other information orthogonal to the subspace of the component to be measured;

[0030] A projection matrix is constructed based on the second spectrum matrix and the generalized inverse matrix of the second spectrum matrix;

[0031] The near-infrared spectrum of each sample in the candidate sample set is projected according to the projection matrix to obtain the pure spectrum signal of each sample in the candidate sample set.

[0032] As a specific embodiment of the present application, the first spectrum matrix is reconstructed based on the near-infrared spectrum matrix, comprising:

[0033] The near-infrared spectrum matrix is singular value decomposed, and the first spectrum matrix is obtained by using the first p principal components obtained by decomposition to reconstruct the near-infrared spectrum matrix.

[0034] As a specific embodiment of the present application, the expression of the second concentration vector is:

[0035]

[0036] wherein, represents a second concentration vector, X SR represents a first spectral matrix, represents a generalized inverse matrix of the first spectral matrix, represents a first concentration vector.

[0037] As a specific embodiment of the present application, the expression of the second spectral matrix is:

[0038]

[0039] wherein, X SR，-k represents a second spectral matrix, represents an average spectrum of the first spectral matrix, represents a second concentration vector, and a is a scalar value related to the first concentration vector and the first spectral matrix.

[0040] As a specific embodiment of the present application, the expression of the projection matrix is:

[0041] H = I - X SR,-k (X SR,-k ) +

[0042] wherein, H represents a projection matrix, I represents a unit matrix, X SR,-k represents a second spectral matrix, (X SR,-k ) + represents a generalized inverse matrix of the second spectral matrix.

[0043] As a specific embodiment of the present application, it further includes: pre-processing the near-infrared spectral matrix, and the pre-processing includes standard normal variable transformation or multivariate scatter correction. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 shows the flow of the screening method of the near-infrared model correction sample set provided by the embodiment of the present application Figure 1 ;

[0045] Figure 2 shows the flow of the screening method of the near-infrared model correction sample set provided by the embodiment of the present application Figure 2 ;

[0046] Figure 3 shows a schematic diagram of the near-infrared spectrum of the tobacco powder sample processed by the SNV method provided by the embodiment of the present application;

[0047] Figure 4A principal component score diagram of the tobacco candidate sample spectrum and the tobacco base sample spectrum provided by the embodiment of the present application is shown;

[0048] Figure 5 A pure spectrum signal of the tobacco candidate sample when the component to be measured is total sugar provided by the embodiment of the present application is shown;

[0049] Figure 6 A distribution diagram of the total sugar content of the calibration samples selected based on the random method and the pure spectrum signal method respectively provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0050] The following describes the embodiments of the present application by specific examples, and other advantages and effects of the present application can be easily understood by those skilled in the art according to the disclosure. Although the description of the present application will be introduced in combination with the preferred embodiments, this does not mean that the features of the present application are limited to the embodiments. On the contrary, the purpose of introducing the present application in combination with the embodiments is to cover other options or modifications that can be extended based on the claims of the present application. In order to provide a deep understanding of the present application, many specific details will be included in the following description. The present application can also be implemented without using these details. In addition, in order to avoid confusion or obscure the focus of the present application, some specific details will be omitted in the description. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0051] It should be noted that in the present specification, similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0052] In order to make the purpose, technical solutions and advantages of the present application more clear, the embodiments of the present application will be further described in detail below in combination with the drawings.

[0053] As Figure 1 shown, a method for screening a near-infrared model calibration sample set provided by an embodiment of the present application includes the following steps:

[0054] A sample setting step S101: determining a base sample set and a candidate sample set based on each sample to be measured.

[0055] Specifically, first, the near-infrared spectrum of all samples to be measured is obtained, and then the near-infrared spectrum of all samples to be measured is analyzed, and a certain number (generally not less than 30 samples) of samples are selected as base samples by the KS method, and the other samples to be measured are selected as candidate samples. The base sample set is composed of each base sample, and the candidate sample set is composed of each candidate sample.

[0056] Specifically, the KS method is a relatively mature technology for screening samples, and thus will not be described herein.

[0057] The constructing step S102: constructing a near-infrared spectrum matrix of the basic sample set and a first concentration vector according to the near-infrared spectrum and the concentration of the to-be-tested component of each sample in the basic sample set.

[0058] Specifically, the concentration value of the to-be-tested component of each basic sample can be determined by using a conventional chemical component detection method. Taking the basic sample set with m basic samples as an example, the first concentration vector can be expressed as wherein, y i (i = 1, 2, ……m) represents the concentration value of the to-be-tested component of the ith basic sample in the basic sample set.

[0059] In this step, after obtaining the near-infrared spectrum data of each basic sample, a near-infrared spectrum matrix X of the basic sample set is constructed according to the obtained near-infrared spectrum data. In the near-infrared spectrum matrix X, the elements in the same row represent the spectrum data of the same basic sample at different wave number points, and the elements in the same column represent the spectrum data of each basic sample at the same wave number point. Assuming that the number of basic samples is m, and the spectrum of each basic sample is composed of n variables, the near-infrared spectrum matrix corresponding to the basic sample set is an m x n matrix, that is, the matrix has m rows and n columns, and n corresponds to the wave number point, that is, the elements in the same row of the matrix correspond to the spectrum information of n wave number points of the same sample.

[0060] The pure spectrum calculating step S103: determining the pure spectrum signal of each sample in the candidate sample set based on the near-infrared spectrum matrix and the first concentration vector.

[0061] This step is based on the near-infrared spectrum matrix Xs of the basic sample and the first concentration vector The pure spectrum signal of each candidate sample is calculated one by one This step can further include the following processes:

[0062] The step S1031: reconstructing a first spectrum matrix based on the near-infrared spectrum matrix, wherein the elements in the same row of the first spectrum matrix represent the spectrum data of the same sample.

[0063] Specifically, after the near-infrared spectrum matrix X is constructed, the near-infrared spectrum matrix X can also be subjected to a standard normal variate (SNV) or multiplicative scatter correction (MSC) pretreatment to obtain a processed spectrum X S The function of the spectrum pretreatment is mainly to eliminate the influence of the sample solid particle size, surface scattering and optical path variation on the spectrum.

[0064] Taking m samples and n variables as an example, the preprocessed near-infrared spectral matrix X S It can be represented as:

[0065]

[0066] The data element x located in the i-th row and j-th column of the matrix Xs above ij This represents the preprocessed spectral data of the i-th sample in the sample set at the j-th wave point.

[0067] In this step, the first spectral matrix X is reconstructed based on the processed spectral matrix Xs. SR In the first spectral matrix X SR In this context, elements in the same row represent the spectral data of the same base sample.

[0068] Specifically, for the preprocessed near-infrared spectral matrix X S Perform singular value decomposition [u,s,v]=svd(X) S The first p principal components obtained from the decomposition were used to analyze the preprocessed near-infrared spectral matrix X. S Perform spectral reconstruction to obtain the first spectral matrix X SR =u p *s p *v p Reconstruction using principal component spectra can reduce the influence of errors such as background and noise in the measured spectrum. It is a relatively mature technique and will not be elaborated on here.

[0069] Step S1032: Based on the first spectral matrix X SR and the first concentration vector Determine the second concentration vector

[0070] Specifically, the second concentration vector The expression is:

[0071]

[0072] in, Denotes the generalized inverse matrix of the first spectral matrix. This represents the first concentration vector composed of the concentration values ​​of the analyte in each base sample.

[0073] Step S1033: Based on the first spectral matrix X SR Second concentration vector Determine the second spectral matrix X SR,-k , where the second spectral matrix X SR,-k It can be used to characterize a space composed of other information that is orthogonal to the subspace of the component to be measured.

[0074] Specifically, the second spectral matrix X SR,-k is expressed as:

[0075]

[0076] wherein X S is the pre-processed near-infrared spectral matrix, represents the spectral information of the component to be measured, and can be replaced by the average spectrum table of the first spectral matrix X SR Further, the average spectrum of the first spectral matrix X SR is obtained by adding all column elements of the matrix X SR and dividing the sum by the number of samples. Alpha is a scalar, and its value is

[0077] Step S1034: constructing a projection matrix H based on the second spectral matrix X SR,-k and the generalized inverse matrix of the second spectral matrix (X SR,-k ) + .

[0078] Specifically, the projection matrix H is expressed as:

[0079] H = I - X SR,-k (X SR,-k ) +

[0080] wherein I represents an identity matrix.

[0081] Step S1035: projecting the near-infrared spectrum of each sample in the candidate sample set according to the projection matrix H to obtain the pure spectral signal of each sample in the candidate sample set.

[0082] Specifically, the expression of the pure spectral signal of the i-th sample in the candidate sample set is:

[0083] wherein represents a vector composed of the near-infrared spectral data of the i-th sample in the candidate sample set.

[0084] Step S104: calculating the Euclidean distance between the pure spectral signals of each sample in the candidate sample set, and sorting the samples in the candidate sample set based on the Euclidean distances.

[0085] Specifically, the Euclidean distance between the pure spectral signals of each sample in the candidate sample set is calculated, and the candidate samples with larger pure spectral differences are sorted.

[0086] The correction sample determination step S105: screening part of the candidate samples in the ordered candidate sample set, and constructing the correction sample set based on at least the screened part of the candidate samples and the base sample set.

[0087] That is, in all candidate samples, the candidate samples are sorted according to the size of the Euclidean distance. For example, in all candidate samples, the two samples with the farthest Euclidean distance are selected into the training set, and then the Euclidean distance of each remaining candidate sample to each sample in the training set is calculated to find the candidate sample with the largest Euclidean distance to be put into the training set, and so on, until the required number of samples is reached.

[0088] Specifically, if the concentration of the candidate sample is unknown, the required number (i.e. the first set number) of candidate samples are selected in turn according to the sorting of step S104, and the correction sample set is constructed together with each base sample. The number of samples required for the correction model is determined according to the complexity of the sample to be measured. Generally, the number of correction samples (i.e. the sum of the first set number and the number of base samples) is about 1 / 2 to 2 / 3 of the total number of samples to be measured. Then the candidate samples in the correction sample set are used to obtain the concentration value of the detection component by the traditional method, so as to establish the near-infrared correction model.

[0089] If the concentration of the candidate sample is known (for example, 10 can be selected and then measured), the second set number (generally not less than 10) of candidate samples can be selected in turn in the ordered candidate sample set, and then the selected second set number of candidate samples are added to the base sample set to update the base sample set, and the selected second set number of candidate samples are removed from the candidate sample set to update the candidate sample set. Then repeat the above steps S102-S104, and recalculate the new projection matrix, the new candidate sample, and continue to update the base sample set and the candidate sample set, until the total number of samples in the updated base sample set reaches the set value (generally 1 / 2 to 2 / 3 of the total number of samples to be measured), then the base sample set at this time is determined as the correction sample set to establish the near-infrared quantitative analysis correction model.

[0090] It should be noted that the above two correction sample set construction methods can also be used together. For example, when there are candidate samples with known concentrations and candidate samples with unknown concentrations in the candidate sample set, the two screening schemes can be used comprehensively according to the situation.

[0091] The screening method of the near-infrared model correction sample set provided by the embodiment of the application obtains a candidate sample pure spectrum signal according to concentration information of a to-be-measured component, and then selects a representative correction sample to establish a near-infrared quantitative model. When the correction sample is selected, not only the spectral information of the sample is referred to, but also the concentration information of the to-be-measured component in the sample is referred to, so that the correction sample covers as much sample information and concentration information as possible, which helps to improve the robustness and accuracy of the correction model. In addition, in the process of selecting the correction sample, the concentration value of the to-be-measured component is only determined for the selected samples, so that the workload of the to-be-measured component chemical value detection is greatly reduced. By using the screening method of the application, the correction sample distribution is more uniform and more representative, and the robustness and accuracy of the correction model can be improved.

[0092] Embodiment

[0093] Step 1) Tobacco sample processing

[0094] 207 tobacco samples were provided by Guizhou China Tobacco Industry Co., Ltd. and came from production areas in Guangdong, Henan, Heilongjiang, Hunan, Liaoning, Shaanxi, Sichuan and Yunnan. Before collecting the sample spectrum, the tobacco sample was placed in a 40℃ oven according to YCT 31-1996 Tobacco and Tobacco Products Preparation and Moisture Determination Oven Method, and was baked for two hours; then the sample was taken out and cooled to room temperature. The sample was poured into a plant crusher and crushed, and then sieved through a 40-mesh sieve to separate the sample with a particle size of less than 40 mesh (≤0.45 mm). After the sample was cooled to room temperature, it was loaded into a disposable sealed bag for low-temperature and light-proof storage.

[0095] Step 2) Experimental instruments and spectrum collection

[0096] The laboratory temperature was controlled between 22±2℃, and the relative humidity was controlled between 40%±10%. The near-infrared instrument was turned on and preheated for not less than 1 hour, and then was used after passing the ValPro program check. An appropriate amount of prepared tobacco powder was loaded into a sample cup for scanning, and the scanning range was 4000-10000 cm -1 , the resolution was 8 cm -1 , and the scanning number was 64.

[0097] Step 3) Determination of chemical values

[0098] The contents of total plant alkaloids and total sugar chemical components in tobacco were determined according to the tobacco industry standards, YC / T 468-2013 Tobacco and Tobacco Products Determination of Total Plant Alkaloids Continuous Flow (Potassium Thiocyanate) Method and YC / T 159-2002 Tobacco and Tobacco Products Determination of Water-soluble Sugar Continuous Flow Method.

[0099] Step 4) Establishment of a correction sample set

[0100] In the process of near infrared spectrum data analysis modeling, the most critical step is the spectrum data pretreatment. The spectrum data contains baseline drift, noise and other interference signals in addition to pure spectrum. The near infrared spectrum of the tobacco powder sample has a large drift, so the standard state transformation (SNV) is used to eliminate the influence of solid particle size, surface scattering and optical path variation on NIR. The processed near infrared spectrum is shown in FIG. 1. Figure 3

[0101] According to the processed near infrared spectrum of the sample, the Euclidean distance between each spectrum in the spectrum set is calculated by using the KS algorithm. The 50 samples with large spectrum difference are selected as the basic samples, and the reference values of the components to be measured in these samples are obtained, mainly the content information of total sugar and total alkaloid in the tobacco powder sample. Figure 4 The principal component score diagram of the basic sample and the candidate sample is shown in FIG. 2. Figure 4 It can be seen that the samples selected by the KS method are uniformly distributed in the spectrum space. According to the foregoing method, the projection matrix is established by using the spectrum and reference value of the basic sample, and then the pure spectrum signal of the candidate sample is obtained, as shown in FIG. 3. Figure 5

[0102] The KS method is used to calculate the Euclidean distance between each pure spectrum signal in the candidate sample. The candidate samples with large pure spectrum difference are sorted. If the concentration of the candidate sample is unknown, the required number of samples can be selected in order, for example, 70 samples are selected at a time, and the reference values of total sugar and total alkaloid of the 70 candidate samples are obtained by using the traditional method. The first calibration sample set of 120 samples is obtained by combining the 50 basic samples and the 70 candidate samples. The partial least squares method is used to establish the calibration model of the components. If the concentration of the candidate sample is known, 35 candidate samples can be selected to obtain the concentration value of the component, and then the first calibration sample set is constructed by combining the 50 basic samples and the 35 candidate samples. The new projection matrix is calculated by the first calibration sample set, and the spectrum pure signal of the remaining 122 samples is obtained. The KS method is used to calculate the spectrum distance of the 122 samples, and the 35 candidate samples are selected again according to the difference size to construct the second sample set. The partial least squares method is used to establish the calibration model.

[0103] The calibration model parameters established by different calibration sample selection methods are shown in Tables 1 and 2. Table 1 is a correlation table of the calibration model parameters established based on different calibration sample selection methods when the component to be measured is total sugar; Table 2 is a correlation table of the calibration model parameters established based on different calibration sample selection methods when the component to be measured is total alkaloid.

[0104] Table 1 Calibration model parameters of the component to be measured total sugar

[0105] ​​

[0106] Table 2. Calibration model parameters for total alkaloids of the analyte.

[0107]

[0108]

[0109] Specifically, the performance of near-infrared spectroscopy calibration models is evaluated from three aspects: robustness, accuracy, and resolution. SEC (Standard Error of Prediction) and SEP (Standard Error of Calibration) represent the prediction standard deviation and calibration standard deviation, respectively. SEP / SEC indicates the model's fit range and its coverage of the tested samples, used to evaluate the model's robustness. Tables 1 and 2 show that the SEP / SEC values ​​of models built using calibration samples selected by the KS method and the pure spectral signal method are less than 1.2, indicating that the models have a strong ability to cope with sample spectral variations. Conversely, the SEP / SEC values ​​of models built using calibration samples selected by the concentration ranking method and the random method are greater than 1.2, indicating that the models have a poor ability to cope with sample spectral variations and insufficient robustness. For total sugar, the SEP / SEC value selected by the concentration ranking method is 2.39, indicating that the model's concentration coverage of the components is too narrow and it is not well-suited to the predicted samples.

[0110] Furthermore, the relative standard error of prediction (RSEP) is an indicator parameter for evaluating the accuracy of the model analysis, expressed as the ratio of SEP to the mean of the reference values ​​in the calibration sample set. As shown in Tables 1 and 2, the RSEP values ​​of the models built using calibration samples selected according to the concentration ranking method and the random method are relatively large, indicating that the model's tolerance cannot cover the test samples (this comparison result can also be found in...). Figure 6 As shown in the figure, the accuracy of the calibration model analysis is poor, while the RSEP of the models built from the samples selected by the other three methods is small and close, and the accuracy of the models is high.

[0111] The relative prediction deviation (RPD) represents the ratio of the standard deviation (SD) of the corrected sample data distribution to the standard deviation (SEP). This value is an important evaluation parameter that is independent of model robustness and is used to assess the model's resolving power.

[0112] The standard deviation SD of the calibration set modeling samples represents the range of the sample concentration reference value, and PRD = SD / SEP (the ratio of the standard deviation SD of the modeling data distribution and the SEP value) indicates the number of samples in the SD range that can be classified by the near infrared analysis. Therefore, RPD is an important evaluation parameter for evaluating the resolution of the model independently of the model file. As shown in the table, the component content of the samples selected by the spectral pure signal method tends to be normally distributed (e.g. Figure 6 The chemical composition concentration of the calibration samples covers a wide range to ensure the prediction range of the unknown samples. As discussed above, the selection of the calibration set samples is very important. Only by considering the spectrum and the concentration information of the components of the samples can representative samples be selected to establish the model, and the robustness, accuracy and resolution of the model can be improved.

[0113] Although the present application has been illustrated and described with reference to the embodiments thereof, it is to be understood that the above description is a further illustration of the application and is not intended to limit the present application thereto. Various modifications in form and details of the application can be made by those skilled in the art without departing from the spirit and scope of the application.

Claims

1. A method for screening a near-infrared model calibration sample set, characterized in that, include: Sample setup steps: Determine the basic sample set and candidate sample set based on each sample to be tested; Construction steps: Based on the near-infrared spectra and concentrations of the analytes in each sample of the basic sample set, construct the near-infrared spectral matrix and the first concentration vector of the basic sample set; Pure spectral calculation steps: Based on the near-infrared spectral matrix and the first concentration vector, determine the pure spectral signal of each sample in the candidate sample set; Sorting steps: Calculate the Euclidean distance between the pure spectral signals of each sample in the candidate sample set, and sort the samples in the candidate sample set based on each Euclidean distance; The steps for determining the corrected samples are as follows: a portion of the candidate samples are selected from the sorted candidate sample set, and a corrected sample set is constructed based on at least the selected portion of the candidate samples and the base sample set; The pure spectral calculation step includes: The first spectral matrix is ​​reconstructed based on the near-infrared spectral matrix. In the first spectral matrix, elements in the same row represent the spectral data of the same sample. The second concentration vector is determined based on the first spectral matrix and the first concentration vector; A second spectral matrix is ​​determined based on the first spectral matrix and the second concentration vector, wherein the second spectral matrix is ​​used to characterize a space composed of other information orthogonal to the subspace of the analyte; Construct a projection matrix based on the second spectral matrix and the generalized inverse of the second spectral matrix; The near-infrared spectra of each sample in the candidate sample set are projected according to the projection matrix to obtain the pure spectral signals of each sample in the candidate sample set.

2. The method as described in claim 1, characterized in that, The step of determining the calibration sample includes: A first predetermined number of samples are selected sequentially from the sorted candidate sample set; Since the concentration of the first set number of selected samples is unknown, the calibration sample set is constructed based on the first set number of selected samples and the base sample set.

3. The method as described in claim 2, characterized in that, The samples in the sorted candidate sample set are sorted from largest to smallest according to their Euclidean distance; the sum of the first set number and the number of samples in the basic sample set is equal to 1 / 2 to 2 / 3 of the total number of samples to be tested.

4. The method as described in claim 1, characterized in that, The calibration sample determination step includes: A second predetermined number of samples are selected sequentially from the sorted candidate sample set; Based on the known concentration of the second set number of samples, the base sample set and the candidate sample set are updated based on the second set number of samples; The construction step is repeated sequentially up to the correction sample determination step until the total number of samples in the basic sample set reaches a set value. Then, the basic sample set at this time is determined to be the correction sample set.

5. The method as described in claim 4, characterized in that, The second set quantity is greater than or equal to 10, and the set value is equal to 1 / 2 to 2 / 3 of the total number of samples to be tested.

6. The method as described in claim 4, characterized in that, Updating the base sample set and the candidate sample set based on the second predetermined number of samples includes: Add the selected second predetermined number of samples to the base sample set to form the updated base sample set; The second set number of samples selected are deleted from the candidate sample set to form the updated candidate sample set.

7. The method as described in claim 1, characterized in that, Reconstructing the first spectral matrix based on the near-infrared spectral matrix includes: The near-infrared spectral matrix is ​​subjected to singular value decomposition, and the first p principal components obtained from the decomposition are used to reconstruct the spectrum of the near-infrared spectral matrix to obtain the first spectral matrix.

8. The method as described in claim 1, characterized in that, The expression for the second concentration vector is: in, Represents the second concentration vector. Represents the first spectral matrix, Denotes the generalized inverse matrix of the first spectral matrix. This represents the first concentration vector.

9. The method as described in claim 1, characterized in that, The expression for the second spectral matrix is: in, Represents the second spectral matrix, The average spectrum of the first spectral matrix, Let α represent the second concentration vector, and let α be a scalar whose value is related to the first concentration vector and the first spectral matrix.

Images