CN106680238B - A Method for Analyzing Substance Content Based on Infrared Spectroscopy - Google Patents

Abstract

The invention relates to a method for analyzing the content of material components based on infrared spectroscopy, which includes establishing a first regression model based on source domain infrared spectrum data and source domain material component content corresponding to the source domain infrared spectrum data, and calculating the first Parameters in the regression model; obtain target domain infrared spectrum data, set up a transfer model between target domain infrared spectrum data and source domain infrared spectrum data, and obtain parameters in the transfer model; according to the target domain infrared spectrum data, The transfer model uses the first regression model to obtain the content of the substance components in the target domain corresponding to the infrared spectrum data of the target domain. The analysis method of the present invention combines the feature transfer method and the partial least squares algorithm in transfer learning to realize the transformation from the target domain to the source domain spectral feature space, which can not only remove redundant information, improve the accuracy of transfer, and can be used in a large To a certain extent, the calculation amount of the transfer process is reduced.

Description

A Method for Analyzing Substance Content Based on Infrared Spectroscopy

technical field

The invention relates to the field of infrared spectrum analysis, in particular to a method for analyzing the content of material components based on infrared spectrum.

Background technique

The content of the material components can be known by infrared spectrum analysis. By measuring the infrared spectrum and analyzing it, we can know the content of the material composition, which can not only be analyzed qualitatively but also quantitatively. However, in the existing infrared spectrum measurement process, changes in measuring instruments or measurement conditions will cause the original calibration model to fail, and re-establishing the model will waste a lot of time and cost, resulting in inaccurate analysis results and low analysis efficiency. .

Contents of the invention

In order to solve the problem of low efficiency of existing remodeling, the present invention proposes a method for analyzing the content of material components based on infrared spectroscopy, including the following steps:

S1. Establishing a first regression model according to the source-domain infrared spectrum data and the content of source-domain material components corresponding to the source-domain infrared spectrum data, and obtaining parameters in the first regression model;

S2. Acquire target domain infrared spectral data, establish a transfer model between target domain infrared spectral data and source domain infrared spectral data, and obtain parameters in the transfer model;

S3. According to the target domain infrared spectrum data and the transfer model, use the first regression model to acquire the target domain material component content corresponding to the target domain infrared spectrum data.

Further, the first regression model is a partial least squares regression model, and the step S1 includes performing feature extraction on the source domain infrared spectral data to obtain a first spectral feature, and according to the first spectral feature and the source domain The partial least squares regression model is established according to the content of the material components, and the regression coefficient is obtained.

Further, the target domain infrared spectrum data includes target domain infrared spectrum standard data and target domain infrared spectrum test data, and the step S2 includes performing feature extraction according to the target domain infrared spectrum standard data to obtain second standard spectral features; The first spectral feature and the second standard spectral feature establish the transfer model to obtain a transfer matrix.

Further, the step S3 includes acquiring a third spectral feature according to the infrared spectrum test data in the target domain, bringing the third spectral feature and the transfer model into the least partial squares regression model to obtain the Describe the content of the substance in the target domain.

Further, the step of performing feature extraction on the source domain infrared spectral data to obtain the first spectral feature includes performing centralized processing on the source domain infrared spectral data and source domain material composition content, and according to the centralized processing The source domain infrared spectrum data and the content of source domain material components are used to establish a least squares regression model to obtain the first spectral feature.

Further, the content of the standard substance in the target domain is also acquired, and the step of performing feature extraction according to the infrared spectrum standard data in the target domain to obtain the second standard spectral features includes: analyzing the infrared spectrum standard data in the target domain and the target Centralized processing is performed on the component content of the standard substance in the target domain, and a partial least squares regression model is established according to the centralized infrared spectrum standard data of the target domain and the component content of the standard substance in the target domain to obtain the spectral characteristics of the second standard.

Further, while the second standard spectral feature is obtained in the step S2, the second standard projection data and the second standard load data are also obtained; in the step S3, the third spectral feature is obtained according to the target domain infrared spectrum test data The step includes, using the mean value of the infrared spectrum standard data in the target domain to perform centralized processing on the infrared spectrum test data in the target domain, and using the centrally processed infrared spectrum test data in the target domain to recursively obtain the third Spectral features: Among them, i is greater than or equal to 1 and less than or equal to k, T _{T_test} is the third spectral feature, k is the number of the third spectral feature, is the i-th component of the second standard projection data, is the i-th residual item of the infrared spectrum test data in the target domain after centralized processing, is the i-th component of the second standard load data.

Further, by solving the optimization problem of the following formula, Among them, B represents the coefficient of the regression model based on the source domain features, M represents the transfer matrix from the target domain features to the source domain features, _{WS and W T represent} the projection matrices of the source domain and the target domain respectively; by T _S =X _S *W _S solves for the first spectral feature, where the first spectral feature is i is greater than or equal to 1 and less than or equal to k, k is the number of the first spectral feature; by Calculate the regression coefficient Β ^T =[b ₁ ,b ₂ ,...,b _k ], and y represents the material composition content in the source domain.

Further, the second standard spectral feature is obtained by the following formula, T _T =X _T *W _T , wherein the second standard spectral feature is i is greater than or equal to 1 and less than or equal to k, where k is the number of second spectral features.

Further, using the second standard spectral feature and the first spectral feature by the following formula Acquire transfer matrix M=[m ₁ ,m ₂ ,...,m _k ], i is greater than or equal to 1 and less than or equal to k, k is the number of second standard spectral features, where from to choose from.

Through the technical solutions of the above embodiments, the method of the present invention based on infrared spectrum analysis of material composition content establishes the transfer relationship between the sample characteristics of the source domain and the target domain, on the one hand, redundant information can be removed, and a more accurate and simple transfer relationship can be obtained. In turn, better prediction results can be obtained. On the other hand, for high-dimensional small sample data sets, the amount of calculation can be greatly reduced.

Description of drawings

The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the accompanying drawings:

Fig. 1 is a schematic flow chart of a method for analyzing material component content based on infrared spectroscopy according to an embodiment of the present invention;

Fig. 2 is a schematic flow chart of a method for analyzing material component content based on infrared spectroscopy according to an embodiment of the present invention;

Fig. 3 is the master-slave spectrum and the deviation spectrum of the corn data set that the analysis method of the present invention is verified;

Fig. 4 is the master-slave spectrum and the deviation spectrum of the tablet dataset verified by the analytical method of the present invention;

Fig. 5 is the selection process of the principal component number of the PLS model of the corn that analytical method of the present invention verifies;

Fig. 6 is the window size selection process of the PDS model of the corn that analytical method of the present invention is verified;

Fig. 7 is the comparison schematic diagram of the actual value and the predicted value of the moisture content in the corn under each model that the analytical method of the present invention is verified;

Fig. 8 is the comparison schematic diagram of the actual value and the predicted value of the oil content in corn verified by the analytical method of the present invention under each model;

Fig. 9 is the comparative schematic diagram of the protein content in corn under each model that the analytical method of the present invention is verified;

Fig. 10 is a comparison schematic diagram of the actual value and the predicted value of the starch content in corn under each model to verify the analytical method of the present invention;

Fig. 11 is the comparison schematic diagram of the predicted value and the real value of the moisture content in the corn before and after calibration migration to the analytical method of the present invention verified;

Fig. 12 is a comparison schematic diagram of the predicted value and the real value of the oil content in corn before and after calibration migration to verify the analytical method of the present invention;

Fig. 13 is a comparison schematic diagram of the predicted value and the real value of the protein content in corn before and after calibration migration to verify the analytical method of the present invention;

Fig. 14 is a comparison schematic diagram of the predicted value and the real value of the starch content in corn before and after the calibration migration to verify the analytical method of the present invention;

Fig. 15 is the schematic diagram of the selection process of the principal component number of the PLS model of the corn that the analytical method of the present invention is verified;

Fig. 16 is a schematic diagram of the window size selection process of the PDS model of corn verified by the analytical method of the present invention;

Figure 17 is a schematic diagram of the comparison between the predicted value and the real value of the first active ingredient in the tablet under different models for verifying the analytical method of the present invention;

Figure 18 is a schematic diagram of the comparison between the predicted value and the real value of the second active ingredient in the tablet under different models for verifying the analytical method of the present invention;

Fig. 19 is a schematic diagram of the comparison between the predicted value and the real value of the third active ingredient in the tablet under different models for verifying the analytical method of the present invention;

Figure 20 is a schematic diagram of the comparison of the predicted value and the real value of the content of active ingredient 1 in the tablet before and after calibration migration for verifying the analytical method of the present invention;

Fig. 21 is a comparison schematic diagram of the predicted value and the real value of the content of active ingredient 2 in the tablet before and after calibration migration for verifying the analytical method of the present invention;

Fig. 22 is a schematic diagram comparing the predicted value and the real value of the content of active ingredient 3 in the tablet before and after calibration migration for verification of the analytical method of the present invention.

Detailed ways

In order to understand the above-mentioned purpose, features and advantages of the present invention more clearly, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described here. Therefore, the protection scope of the present invention is not limited by the specific details disclosed below. EXAMPLE LIMITATIONS.

Embodiment one

As shown in Figure 1, the present invention provides a kind of method based on infrared spectrum analysis material component content, comprises the following steps:

S101. Establish a first regression model according to the source-domain infrared spectrum data and the content of source-domain material components corresponding to the source-domain infrared spectrum data, and obtain parameters in the first regression model; the first regression model is, for example, A partial least squares regression model, performing feature extraction on the source domain infrared spectral data to obtain a first spectral feature, establishing the partial least squares regression model according to the first spectral feature and source domain material component content, and calculating the regression coefficient; specifically, the step of performing feature extraction on the source domain infrared spectral data to obtain the first spectral feature includes performing centralized processing on the source domain infrared spectral data and source domain material composition content, according to the centralized processing The first spectral feature is obtained by establishing a least squares regression model based on the source domain infrared spectral data and the content of source domain material components. The operation of centralized processing is to subtract the mean value of the source domain infrared spectrum data from the source domain infrared spectrum data, and subtract the mean value of the source domain material component content from the source domain material composition content, so as to reduce the impact of deviation on model building.

Specifically, by solving the optimization problem of the following formula, Among them, B represents the coefficient of the regression model based on the source domain features, M represents the transfer matrix from the target domain features to the source domain features, and _{WS and W T represent} the projection matrices of the source domain and the target domain, respectively. Solving the first spectral feature by T _S =X _S * _WS , wherein the first spectral feature is i is greater than or equal to 1 and less than or equal to k, k is the number of the first spectral feature; by Calculate the regression coefficient Β ^T =[b ₁ ,b ₂ ,...,b _k ], and y represents the material composition content in the source domain.

S102. Acquire target domain infrared spectrum data, establish a transfer model between target domain infrared spectrum data and source domain infrared spectrum data, and obtain parameters in the transfer model; the target domain infrared spectrum data includes target domain infrared spectrum standards Data and target domain infrared spectrum test data, perform feature extraction according to the target domain infrared spectrum standard data to obtain a second standard spectral feature; establish the transfer model according to the first spectral feature and the second standard spectral feature, and obtain To obtain the transfer matrix, in order to improve the accuracy, some spectral features can be selected from the first spectral feature and the second standard spectral feature to establish a transfer model, and the selection is corresponding to the corresponding selection according to the concentration of the substance. If it can be adopted, the source domain The data set with the same substance composition content as the standard substance concentration in the target domain is used for calculation.

Specifically, the second standard spectral feature is obtained by the following formula, T _T =X _T *W _T , wherein the second standard spectral feature is i is greater than or equal to 1 and less than or equal to k, where k is the number of second spectral features.

Using the second standard spectral feature and the first spectral feature by the following formula Acquire transfer matrix M=[m ₁ ,m ₂ ,...,m _k ], i is greater than or equal to 1 and less than or equal to k, k is the number of second standard spectral features, where from to choose from.

S103, according to the target domain infrared spectrum data and the transfer model, use the first regression model to acquire the target domain material component content corresponding to the target domain infrared spectrum data; acquire according to the target domain infrared spectrum test data The third spectral feature, bringing the third spectral feature and the transfer model into the least partial squares regression model to obtain the content of the substance in the target domain.

The embodiment of the present invention also includes obtaining the target domain standard material component content, and the step of performing feature extraction according to the target domain infrared spectrum standard data to obtain the second standard spectral feature includes: analyzing the target domain infrared spectrum standard data and the The content of standard substance components in the target domain is centrally processed, and a partial least squares regression model is established based on the centralized infrared spectrum standard data of the target domain and the content of standard substance components in the target domain to obtain the second standard spectral features. The steps of the centralized processing are similar to the above-mentioned processing steps of the data originating from the infrared spectrum.

In the embodiment of the present invention, step S102 obtains the second standard spectral feature while obtaining the second standard projection data and the second standard load data; in the step S103, the third The step of spectral characteristics includes, using the mean value of the infrared spectrum standard data in the target domain to perform centralized processing on the infrared spectrum test data in the target domain, and using the centrally processed infrared spectrum test data in the target domain to obtain recursively according to the following formula The third spectral feature: Among them, i is greater than or equal to 1 and less than or equal to k, T _{T_test} is the third spectral feature, k is the number of the third spectral feature, is the i-th component of the second standard projection data, is the i-th residual item of the infrared spectrum test data in the target domain after centralized processing, is the i-th component of the second standard load data.

The method of the present invention based on infrared spectrum analysis of material component content establishes the transfer relationship between the source domain and the target domain sample characteristics, on the one hand, it can remove redundant information and obtain a more accurate and simple transfer relationship, so better prediction results can be obtained , on the other hand, for high-dimensional small-sample data sets, it can greatly reduce the amount of computation. In addition, only one parameter of the latent variable of the partial least squares algorithm (PLS algorithm) needs to be set, and the implementation process is very simple. It should be noted that the term "infrared spectrum" used in the present invention can be understood as including near-infrared spectrum, mid-infrared spectrum and far-infrared spectrum.

Embodiment two

The method for analyzing the content of material components based on infrared spectroscopy in the present invention combines transfer learning and PLS algorithm to form a transfer calibration algorithm (CT_pls algorithm). Domain feature space, and then the model of the source domain can be used to process the data of the target domain. This method first uses the PLS algorithm to extract the features of the source domain samples and the target samples, then establishes a multivariate calibration model based on the source domain features and a linear transfer model between the source domain and target domain features, and finally uses the same method for the unknown After feature extraction and transfer of target domain samples, the source domain calibration model is used to predict the transferred features.

Assuming that there are source domain datasets {X _S ,y} and target domain datasets {X _T ,y} respectively, where X _S and X _T are measured by the master spectrometer and slave spectrometer respectively, establish the relationship between the source domain and the target domain Calibrating the migration model is actually an optimization problem for solving formula (3.1).

In formula (3.1), B represents the coefficients of the regression model based on source domain features, M represents the transfer matrix from target domain features to source domain features, and W _S and W _T represent the projection matrices of the source and target domains, respectively. In this paper, the partial least squares algorithm is selected as the main algorithm. W _S and W _T are obtained by establishing PLS models of {X _S ,y} and {X _S ,y} respectively. The feature T _S of the source domain and the feature T of the target domain _T is obtained by formula (3.2).

After obtaining the source domain characteristics T _S , use the source domain characteristic data {T _S , y _S } to establish a multivariate calibration model, where Calculate the regression coefficient Β ^T =[b ₁ ,b ₂ ,...,b _k ], where k represents the number of extracted main features.

In order to realize the effective prediction of the target domain data by the source domain model, it is necessary to use the standard set to transform the spectral space. Formulas (3.4) (3.5) indicate the realization method of transforming the spectral features from the target domain to the source domain.

Τ' _S ←Τ _T Μ (3.4)

in, Τ' _S and T _T are the features of the source domain and target domain sample sets respectively, from which Τ' _S obtains T _S , which is used to calculate the transfer matrix M=[m ₁ ,m ₂ ,...,m _k ].

After establishing the calibration model of the source domain and the transfer model between the source domain and the target domain, the effective prediction of the target domain samples can be realized, as shown in formula (3.6).

y _T =T _T *M*B (3.6)

Specifically, as shown in Figure 2, the method for analyzing the content of material components based on infrared spectroscopy in the present invention includes obtaining the source domain training set, that is, obtaining the source domain infrared spectrum data and the source domain material component content; obtaining the target domain standard set, that is, obtaining Infrared spectrum standard data in the target domain and the content of standard substances in the target domain; obtain the test set in the target domain, that is, obtain the infrared spectrum test data in the target domain and the content of the test material in the target domain; centralize the source domain data and process the target domain data Centralized processing; use the pls model to extract the first spectral feature of the source domain data to form a combined feature data set, from which the features corresponding to the standard set (that is, corresponding to the content of the material composition) are extracted, and the combined feature data set and the pls algorithm are used to establish the first A regression model, the target domain standard set uses pls to perform feature extraction to obtain the second standard spectral feature, and the transfer matrix between the selected first spectral feature and the second standard spectral feature is obtained through the pls model, and the target domain test data is used The pls model obtains the third spectral feature, and brings the third spectral feature and the transfer matrix into the first regression model, so as to obtain the content of the material composition corresponding to the test data in the target domain. The specific implementation process includes steps such as data preprocessing, feature extraction, establishment of a source domain calibration model, calculation of transfer relationships, and prediction of unknown target domain data.

Specifically, it can be realized by a processor circuit carrying a computer program, and the flow of the computer program is as follows:

The method for analyzing the content of material components based on infrared spectroscopy in the embodiment of the present invention adopts partial least squares regression analysis, and partial least squares regression analysis (PLS) provides a method for many-to-many linear regression modeling, especially when two groups of variables There are many, and there are multiple correlations, and when the number of observed data (sample size) is small, the model established by partial least squares regression analysis has advantages that traditional methods such as classical regression analysis do not have. When two sets of measurement samples of the same item come from different measuring instruments or measurement states, the two sets of samples are different but related, so the samples from the new space can be migrated to the reference space, and then the model of the reference space can be directly used for the new samples. predict. The original model is reused to reduce the cost of modeling.

1. Establish a PLS regression model based on spectral features

First, a partial least squares regression model was established for the infrared spectral data and its corresponding component concentrations to obtain spectral features, and the number of spectral features was selected by cross-validation method. Then, the PLS model is re-established for the spectral features and their corresponding component concentrations to calculate the regression coefficient of the model. At this time, the number of main features (spectral features) is still selected through the cross-validation method. Establishing the PLS model twice and directly establishing the PLS model for the infrared spectral data basically has no effect on the prediction accuracy, and the regression coefficient calculated by using the spectral characteristics can directly predict the spectral characteristics of the target domain after transfer.

2. Realize transfer learning between spectral features

The conditional probability or marginal probability distribution of infrared spectral data measured by different spectrometers may be different, so that the original multivariate calibration model cannot accurately predict the infrared spectral data of the target domain, and there will often be a large prediction deviation. Due to the remodeling The cost is high, so it is necessary to migrate the spectral features of the target domain to the source domain, thereby reducing the distribution difference between the source domain and the target domain. Firstly, features are extracted from the standard spectral samples in the source and target domains, and then a feature-to-feature PLS model is established to calculate the transfer matrix. The transfer of features can be achieved by multiplying the target domain features with the transfer matrix.

3. Predict the target domain spectral data

After the features of the target domain are migrated to the feature space of the source domain, the feature-based regression model of the source domain can be directly used to predict the features of the target domain. This avoids re-establishing the model for the target domain samples and greatly reduces the modeling cost.

Corn and tablet data have been analyzed respectively at the analytical method in the present invention, specifically as follows:

1. Corn Dataset

The corn data set has 80 samples, corresponding to the contents of water, oil, protein, and starch, which can be obtained from (http://www.eigenvector.com/Data/Data_sets.html). The infrared spectrum data sets are measured by three different instruments, m5, mp5, and mp6, in the wavelength range of 1100–2498nm at intervals of 2nm, with a total of 700 channels. In this experiment, the spectrum measured by m5 is used as the main spectrum, and the spectral data is used as the source domain data set X _S . Since the spectrum measured by mp6 is quite different from that measured by m5, it is selected as the secondary spectrum, and the corresponding data set is used as the target. Domain dataset X _T . The spectrogram is shown in Figure 3, where the sub-figures (A), (B), and (C) respectively represent the master spectrogram, the slave spectrogram, and the spectral difference map between the master spectrum and the slave spectrum.

In the experiment, the Kennard-Stone (KS) algorithm was used to divide the data set. First, 20% of the data were extracted from the source domain and the target domain data set as test samples, respectively 16, and the test samples of the target domain were used for testing Calibrate the transfer model. The remaining 80% samples are used as training samples, which are 64 respectively. The training samples of the source domain are used to establish a reference model, which can predict the migration samples of the target domain, and the target domain is used to establish a standard model of the target domain, so that Compare the performance of other transfer models. Then, several samples are extracted from the training samples of the source domain and the target domain through the KS algorithm as standard sample sets, which are used to establish the transfer relationship between the source domain samples and the target domain samples. The number of standard samples has a great influence on the transfer relationship. The number of standard samples is too small to obtain sufficient sample information, and the number of standard samples is too large to easily introduce redundant information. In both cases, accurate transfer relationships cannot be obtained. In order to take both into account, this experiment uses the KS algorithm to extract 50% of the samples from the training samples of the source domain and the target domain as standard sample sets, 32 respectively.

2. Pill Dataset

In 2002, the "Shootout" data set released at the International Diffuse Reflectance Conference (IDRC) contained infrared spectral data of tablet samples measured by two spectrometers in the wavelength range of 600-1898nm at 2nm intervals, respectively as source domain data and The target domain data, each containing 650 variables, was used to analyze the content of three active ingredients in the tablet. These samples are divided into source domain calibration sample set and target domain calibration sample set, each containing 155 samples, source domain test set and target domain test set, each containing 460 samples. 50% of the samples are extracted from the calibration set of the source domain and the target domain as the standard set through the KS algorithm, 78 samples respectively. The infrared spectrum of the tablet is given in Figure 4, where Figure 4(A) represents the master spectrum, Figure 4(B) represents the slave spectrum, and Figure 4(C) represents the spectral difference between the master spectrum and the slave spectrum. It can be seen from Figure 4(C) that the wavenumber and In the range of , there are differences between the main spectrum and the secondary spectrum, and there are large fluctuations in the difference at the front end, while in other wavenumber ranges, there are small differences. It shows that it is easier to introduce noise at both ends of the spectrum. Since the difference between master and slave spectra is not large, it can be guessed that the predicted performance does not shift much before and after model migration.

The specific process is as follows:

1. Data preprocessing method

Before training the model, choose a centralized method to preprocess the data to avoid deviations caused by large numerical differences.

2. Parameter selection

The parameter selection of the model can have a great impact on the performance of the model, and choosing an optimal parameter can make the model obtain the optimal performance. For example, for the PLS algorithm, choosing the best number of principal components can make the model obtain the best prediction effect. In the experiment of the present invention, SBC (slope and deviation correction method), MSC (multivariate scattering correction), PDS (subsection direct standard), CT_pls all adopts the multivariate calibration model of PLS algorithm to establish master spectrum data, so after determining the standard sample quantity , the SBC and CT_pls algorithms only need to set a parameter of the principal component score, and the PDS algorithm needs to set the window size in addition to the principal component score. In the present invention, the method of selecting 10-fold cross-validation selects the number of principal components of the PLS algorithm, sets the number of principal components from 1 to 5, and the interval is 1, calculates its corresponding cross-validation error (RMSECV) respectively, and selects the smallest The number of principal components corresponding to RMSECV is the best number of principal components. For the PDS algorithm, due to the small number of samples in the standard data set, when establishing the PLS sub-model for each window, use 5-fold cross-validation, set the window size from 3 to 20, the interval is 2, and the window size should be an odd number not less than 3 , calculate the RMSECV corresponding to each window size, and the sampling corresponding to the minimum RMSECV is the best window. model evaluation

In the experiment of the present invention, root mean square error (RMSE) is used as an index for parameter selection and model evaluation. The calculation method of RMSE is as formula (3.11).

in, is the predicted value, as the reference value (true value or comparative value), is the number of test samples.

RMSEC represents the training error of the calibration set, RMSEP represents the prediction error of the test set, and RMSECV represents the cross-validation error. For the cross-validation error of the PLS algorithm, represents the real value. For the PDS algorithm, choose the cross-validation error of the window size, Indicates the predicted value of the main spectral standard set.

In order to more intuitively compare the CT_pls model proposed by the present invention with other classical models and the PLS benchmark model in the degree of difference in predictive performance, use formula (3.12) to calculate the improvement rate of CT_pls algorithm relative to other algorithmic performance or drop rate.

In formula (3.12), RMSEP _{CT_pls} represents the prediction error of the CT_pls algorithm, Indicates the prediction error of the other comparison algorithms.

In addition, the present invention uses the rank sum test method to check whether there is a significant difference between the CT_pls method and other algorithms, and uses the wilcoxon function in the scipy package in python to directly calculate the p value between the predicted values. If p>0.05, it means There is no significant difference between the two algorithms, otherwise there is a significant difference.

The present invention selects corn data set, tablet data set to carry out experiment. For the SBC, PDS, and CT_pls algorithms, the PLS algorithm is used as the main algorithm, and the source domain data is used to establish a multivariate calibration model as a reference model, which is used to predict the migrated target domain prediction samples. At the same time, the PLS algorithm is used to establish a multivariate calibration model for training samples in the target domain, which is used to compare the prediction performance of the calibration transfer model, which facilitates a more comprehensive and accurate evaluation of the SBC, PDS, and CT_pls calibration transfer methods. The experimental results mainly include the following parts:

(1) The selection process of the principal components of the PLS algorithm and the display of the results of RMSEC, RMSEP, and RMSECV.

(2) The selection process of the window size of the PDS algorithm.

(3) Under different standard sample numbers, the change of RMSEP of the three migration algorithms SBC, PDS, and CT_pls.

(4) Set a fixed number of standard samples, and compare the predictive capabilities of the three migration algorithms, SBC, PDS, and CT_pls.

(5) Comparison of model prediction ability and parameter setting before and after calibration migration.

Experiments were performed using the corn dataset. Table 3.1 shows the training error, cross-validation error, prediction error, and principal component number of the PLS model corresponding to moisture, oil, protein, and starch content directly using the target domain training set of corn.

Table 3.1 Errors and parameters of the PLS model of the corn target domain dataset

It can be seen from Table 3.1 that the RMSEC, RMSECV, and RMSEP of each component in corn are not very different, indicating that there is no over-fitting phenomenon, and the RMSEP is small, indicating that there is no under-fitting phenomenon, which can further explain The selection of principal components is reasonable. The present invention adopts 10-fold cross-validation method to select the principal components of the PLS algorithm, and Fig. 5 (A) (B) (C) (D) provides respectively about the PLS of moisture content, oil content, protein, starch content in corn The change process of RMSECV of the model with the number of principal components, respectively, when the scores of principal components are 5, 5, 5, and 5, the minimum value of RMSECV is obtained, so the optimal principal components of the PLS model about the content of each component in corn are respectively For 5, 5, 5, 5. Although the maximum principal component score is set to 5, the RMSECV of each component of the corn data set does not converge with the change of the principal component score, and the global minimum value cannot be obtained. However, if the principal component score is selected too large, overfitting will occur, and To increase the complexity of the PLS model, through multiple experimental analysis, selecting the maximum principal component to be 5 can obtain a more satisfactory effect.

For the PDS algorithm, the window size needs to be reasonably selected, and the present invention selects the window size by the method of 5-fold cross-validation, and Fig. 6 (A) (B) (C) (D) provides respectively about moisture content in corn , oil, protein, starch content of the PDS model window size selection process, select the window size corresponding to the minimum RMSECV as the best window of the PDS model. From Fig. 6, for the PDS model of moisture content, the optimal window size is 13, while for the PDS models of the other three components, the optimal window size is 3.

For the SBC, PDS, and CT_pls algorithms, their predictive performance is affected by the standard sample size. Because the number of standard samples affects the transfer relationship, and the transfer relationship directly affects the prediction accuracy, the number of standard samples affects the prediction performance of the calibration transfer model. Table 3.2-Table 3.5 shows the prediction errors of different models for the contents of moisture, oil, protein, and starch in corn when the number of standard samples is different, where N in the first row represents the number of standard samples. The PLS model here represents the benchmark model established directly using the training data of the target domain, so when predicting the test samples of the target domain, the samples do not need to be migrated, so the prediction error has nothing to do with the number of standard samples.

Table 3.2 Prediction errors for moisture content in corn

It can be seen from Table 3.2 that for the SBC algorithm, the minimum prediction error is 0.3081, which is quite different from that of the PLS method at 0.1916. Since SBC is only applicable to systematic errors, it shows that the SBC method is not suitable for moisture prediction. For the PDS algorithm, the minimum prediction error is obtained at N=45, RMSECP=0.1767, for the CT_pls algorithm, the minimum prediction error is obtained at N=13, RMSEP=0.1678, the smaller prediction errors of the two are all at N= 32 places were obtained, respectively 0.1860 and 0.1831. It can be seen that the optimal transfer relationship cannot be obtained if the number of standard samples is too large or too small.

It can be seen from Table 3.3 that the SBC method obtains the smallest prediction error of 0.0668 when N=52, but the prediction errors under other standard sample numbers except N=26 are close to it, and they are all close to the prediction error of PLS 0.0624, indicating that the SBC method is suitable for the prediction of oil content. The minimum prediction error of the PDS algorithm is obtained at N=52, RMSEP=0.0787, the minimum prediction error of the CT_pls algorithm is obtained at N=45, RMSEP=0.0723, and the smaller values are obtained at N=32, respectively 0.0832 and 0.0740 , and since N=32, the RMSEP of PDS and CT_pls have little change.

Table 3.3 Prediction error of oil content in corn

Table 3.4 Prediction errors for protein content in corn

It can be seen from Table 3.4 that the SBC method achieves the minimum prediction error at N=39, RMSEP=0.2552, and the change of RMSEP is not large during the whole process of changing the standard sample size. The PDS algorithm achieves the minimum value of 0.2296 at N=45, and since N=32, RMSEP is relatively stable. The CT_pls algorithm achieves the minimum prediction error of 0.2093 at N=45, and since N=26, RMSEP is relatively stable.

Table 3.5 Prediction Errors for Starch Content in Corn

It can be seen from Table 3.5 that the minimum prediction error of the SBC algorithm is obtained at N=39, RMSEP=0.5775, and RMSEP is relatively stable. The PDS algorithm is obtained at N=32, RMSEP=0.4964, and the smaller values are obtained at N=26 and N=39, which are 0.5101 and 0.5270 respectively. The CT_pls algorithm obtains the minimum prediction error value of 0.4592 at N=52, and obtains smaller values at N=(26, 32, 39, 45).

Through the analysis of Table 3.2-Table 3.5, the following conclusions can be drawn: First, the change of the standard sample size has little predictive ability for the SBC algorithm, and the predictive ability of the SBC algorithm is not stable. For example, the prediction of oil content is very good, slightly better than the PDS and CT_pls algorithm, and close to the PLS algorithm, but the effect of the water prediction is very poor, far inferior to the PDS and CT_pls algorithm, and it is similar to the PLS prediction The error is quite different. Second, the prediction error of the PDS and CT_pls algorithms is greatly affected by the number of standard samples. Generally speaking, when N<32, the prediction error is large, and as the number of samples increases, the RMSEP will decrease. At N=32 The minimum value or smaller value is obtained. After that, as the number of samples increases, the RMSEP does not change much or decreases. Therefore, choosing 32 standard samples (ie 50% of the training samples) can obtain a better migration effect. Third, comprehensively compare the predictive performance of SBC, PDS, and CT_pls algorithms. CT_pls has the best predictive performance, followed by PDS algorithm, and then SBC algorithm.

In order to compare the prediction effect of the calibration migration algorithm more fairly and intuitively, the present invention selects 32 standard samples to establish the transfer relationship between the source domain and the target domain, and Fig. 7-Fig. The comparison graph between the predicted value and the real value of each algorithm. The closer the predicted value is to the real value, the closer the corresponding marked point is to the line y=x. The degree of concentration is used to judge the predictive performance of the algorithm, and then to observe their predictive effect more intuitively.

Since the prediction errors of the PDS and CT_pls models are not very different, it is impossible to compare the advantages and disadvantages of the two algorithms according to the concentration of the marked points through Figure 7-Figure 10, so Table 3.6 shows the prediction values corresponding to Figure 7-Figure 10 forecast error. At the same time, Table 3.7-Table 3.10 gives the prediction error improvement rate or decrease rate of CT_pls for PLS, SBC, and PDS algorithms and the p-value of the rank sum test between them.

Table 3.6 Prediction error of each component concentration of corn data set under different models

Table 3.7 The improvement rate of moisture content in corn CT_pls algorithm to other algorithms and the p-value of rank sum test

Table 3.8 The improvement rate of CT_pls algorithm for oil content in corn to other algorithms and the p-value of rank sum test

Table 3.9 The improvement rate of CT_pls algorithm for protein content in corn to other algorithms and the p-value of rank sum test

Table 3.10 The improvement rate of CT_pls algorithm for starch content in corn to other algorithms and the p-value of rank sum test

From Table 3.6 to Table 3.10, it is further explained that among the three migration algorithms of SBC, PDS and CT_pls, CT_pls algorithm has the best prediction performance, followed by PDS algorithm and SBC algorithm is the worst. Moreover, since the p values in Table 3.7-3.10 are all greater than 0.05, it shows that there is no significant difference between the CT_pls algorithm and other algorithms.

Finally, by using the source domain model directly to predict the target domain test samples that have not been transferred, and comparing it with the prediction using the CT_pls algorithm, the migration ability of the CT_pls model can be evaluated intuitively. Figures 11-14 show the comparison charts of the predicted value and the real value of the model without calibration migration and the comparison charts of the predicted value and the real value of the calibration migration using the CT_pls algorithm.

In Figures 11-14, the dots represent the relationship points between the real and predicted values of the test samples in the target domain when calibration migration is not performed, and the five-pointed star represents the predicted and real values of the target domain after calibration and migration using the CT_pls algorithm relationship between points. From Figures 11 to 14, it can be seen that the dark dots deviate seriously from the straight line y=x, while the five-pointed stars are all concentrated near the straight line y=x, indicating that direct use of the source domain model to predict the target domain data will appear very Large deviation, which is introduced by different measuring instruments, and after using the CT_pls algorithm for calibration migration, the deviation between the source domain data and the target domain data can be reduced to a large extent, and the source domain model can be directly used Predict the transferred target and data, and obtain good prediction results.

Experiments were performed using the pill dataset. Table 11 shows the training error, cross-validation error, prediction error and the number of principal components of the PLS model corresponding to the contents of the three active ingredients directly using the target domain training set of the tablet.

Table 3.11 Errors and parameters of the PLS model for the target domain dataset of tablets

It can be seen from Table 3.11 that the RMSEC, RMSECV, and RMSEP of each component in the tablet are on the same order of magnitude, indicating that there is no over-fitting phenomenon, and the RMSEP is small, indicating that there is no under-fitting phenomenon. It shows that the selection of principal components is reasonable. Embodiments of the present invention adopt the 10-fold cross-validation method to select the principal components of the PLS algorithm, and Fig. 15 (A) (B) (C) respectively provides the RMSECV of the PLS model about three kinds of active ingredient contents in the tablet. In the change process of the number of components, the minimum value of RMSECV is obtained when the number of principal components is 3, 2, and 5 respectively. Therefore, the optimal number of principal components of the PLS model about the content of each component in the tablet is 3, 2, and 5 respectively.

For the PDS algorithm, the embodiment of the present invention selects the window size through a 5-fold cross-validation method, and Fig. 16 (A) (B) (C) respectively provides the window size of the PDS model about the content of the three active ingredients of the tablet selection process. It can be seen from Figure 16 that for the PDS model corresponding to the first active ingredient, the optimal window size is 19, while for the PDS models of the two active ingredients, the optimal window sizes are 3 and 13, respectively.

For the SBC, PDS, and CT_pls algorithms, Table 12-Table 14 shows the prediction errors of the contents of the three active ingredients in the tablet under different models when the number of standard samples is different, where N in the first row represents the number of standard samples, The PLS model is a model built for the target domain training data.

Table 3.12 Prediction errors for the content of the first active ingredient in tablets

Table 3.13 Prediction error for the content of the second active ingredient in the tablet

Table 3.14 Prediction error for third active ingredient content in tablets

It can be seen from Table 3.12-Table 3.14 that in the process of changing the number of standard samples, the prediction error of the CT_pls algorithm is basically slightly lower than that of the PDS algorithm, and the prediction error of the SBC algorithm is often higher than that of the PDS algorithm. error. It shows that the prediction performance of the CT_pls algorithm is better than that of the PDS algorithm, and the prediction performance of the PDS algorithm is better than that of the SBC algorithm, and the prediction errors of the CT_pls and PDS algorithms are close to the prediction errors of the PLS algorithm, indicating that both have good calibration migration capabilities. In addition, the prediction error of the SBC algorithm for the second active ingredient is also close to that of the PLS algorithm, but the prediction error for the first active ingredient is quite different from that of the PLS algorithm, which further shows that the application of the SBC algorithm is not widespread.

Figure 17, Figure 18, and Figure 19 respectively show the true and predicted values of the four models of PLS, SBC, PDS, and CT_pls corresponding to the three active ingredients when N=78 (i.e., 50% of the samples in the training set) comparison chart.

Through Figure 17, Figure 18, and Figure 19, it is impossible to clearly compare the advantages and disadvantages of the two algorithms according to the concentration of the marked points. Therefore, Table 3.14 shows the prediction errors corresponding to the predicted values in Figure 17, Figure 18, and Figure 19. At the same time, Table 3.16, Table 3.17, and Table 3.18 give the prediction error improvement rate or decrease rate of CT_pls for PLS, SBC, and PDS algorithms and the p-values of the rank sum test between them.

Table 3.15 Prediction errors of the content of each active ingredient in the tablet data set under different models

Table 3.16 The improvement rate of the CT_pls model for the content of active ingredient 1 in tablets to other models and the p-value of the rank sum test

Table 3.17 The improvement rate of the CT_pls model for the content of active ingredient 2 in tablets to other models and the p-value of the rank sum test

Table 3.18 The improvement rate of the CT_pls model for the content of active ingredient 3 in tablets to other models and the p-value of the rank sum test

It can be seen from Table 3.15-Table 3.17 that for the prediction of the content of active ingredients in tablet data, the prediction performance of the CT_pls algorithm is the best, even better than the PLS model established directly using the target domain data, and the prediction performance of the PDS algorithm is very close to that of PLS model, the SBC model has the worst predictive performance. And the p value of each group is less than 0.05, indicating that there is a significant difference between the CT_pls algorithm and other algorithms.

Figure 20, Figure 21, and Figure 22 show the comparison charts of the predicted value and the real value of the model without calibration migration and the comparison charts of the predicted value and the real value of the calibration migration using the CT_pls algorithm.

It can be seen from Figure 20, Figure 21, and Figure 22 that the five-pointed star-shaped marking points are closer than the dot-shaped marking points and are concentrated near the straight line y=x, indicating that after using the CT_pls algorithm for calibration migration, more good predictive effect. However, compared with the corn dataset, the migration effect of the pill dataset is not obvious, because the difference between the master and slave spectra of the pill dataset is not too large, which can be seen from Figure 4.

The analysis method of the present invention uses target domain training samples to establish a PLS model as a benchmark model, which is used to compare the transfer capabilities of the three calibration transfer models of SBC, PDS and CT_pls. The experimental results show that the prediction errors of the PDS and CT_pls models are close to those of the PLS, indicating that both have good transferability, and the prediction errors of the CT_pls model are smaller than those of the PDS model. However, the SBC model does not always achieve good prediction results, indicating that its stability and predictive ability are far inferior to those of the PDS and CT_pls models. Therefore, in general, among the three migration models, CT_pls model has the best prediction performance, followed by PDS and SBC is the worst.

In the present invention, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance. The term "plurality" means two or more, unless otherwise clearly defined.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (5)

1. A method for analyzing the content of substance components based on infrared spectroscopy is characterized by comprising the following steps:

s1, establishing a first regression model according to source domain infrared spectrum data and source domain substance component content corresponding to the source domain infrared spectrum data, and solving parameters in the first regression model;

s2, acquiring target domain infrared spectrum data, establishing a transfer model between the target domain infrared spectrum data and the source domain infrared spectrum data, and solving parameters in the transfer model;

s3, obtaining the content of the target domain substance component corresponding to the target domain infrared spectrum data by using the first regression model according to the target domain infrared spectrum data and the transfer model;

the first regression model is a partial least squares regression model, and the step S1 includes performing feature extraction on the source domain infrared spectrum data to obtain a first spectral feature, establishing the partial least squares regression model according to the first spectral feature and the source domain substance component content, and calculating a regression coefficient;

the target domain infrared spectrum data comprises target domain infrared spectrum standard data and target domain infrared spectrum test data, and the step S2 comprises the steps of extracting features according to the target domain infrared spectrum standard data to obtain second standard spectral features; and establishing the transfer model according to the first spectral feature and the second standard spectral feature to calculate a transfer matrix.

2. The method for analyzing the contents of the components of the substance according to claim 1, wherein the step S3 comprises obtaining a third spectral feature according to the target domain infrared spectrum test data, and obtaining the contents of the components of the target domain substance by fitting the third spectral feature and the transfer model into the least-squares regression model.

3. The method according to claim 1, wherein the step of performing feature extraction on the source region infrared spectral data to obtain a first spectral feature comprises performing a centering process on the source region infrared spectral data and the source region material component content, and establishing a least squares regression model based on the centered source region infrared spectral data and the source region material component content to obtain the first spectral feature.

4. The method for analyzing the contents of the components of the substance based on the infrared spectrum as claimed in claim 1, further obtaining the contents of the components of the standard substance including a target domain, wherein the step of performing the feature extraction according to the standard data of the infrared spectrum of the target domain to obtain the second standard spectral feature comprises: and performing centralization treatment on the target domain infrared spectrum standard data and the component content of the target domain standard substance, and establishing a partial least square regression model according to the centralization treated target domain infrared spectrum standard data and the component content of the target domain standard substance to obtain a second standard spectral feature.

5. The method for analyzing the contents of the components of a substance based on the infrared spectroscopy as set forth in claim 2, wherein in the step S2, the second standard projection data and the second standard loading data are acquired simultaneously with the acquisition of the second standard spectral feature; the step of obtaining the third spectral characteristic according to the target domain infrared spectrum test data in the step S3 includes performing centering processing on the target domain infrared spectrum test data by using the mean value of the target domain infrared spectrum standard data, and sequentially recurrently obtaining the third spectral characteristic by using the centered target domain infrared spectrum test data according to the following formula:wherein i is not less than 1 and not more than k, T_{T_test}Is the third spectral feature, k is the number of the third spectral features,for the ith component of the second standard projection data,for the ith residual error of the centered target domain infrared spectroscopy test data,is the transpose of the i-th component of the second standard payload data.