CN111220565B - CPLS-based infrared spectrum measuring instrument calibration migration method - Google Patents

Abstract

The invention relates to the technical field of migration learning under a machine learning module, and provides a calibration and migration method of an infrared spectrum measuring instrument based on CPLS. First, the source domain dataset {X _m , Y} and the target domain dataset {X _s , Y} are collected, and they are centrally processed to obtain the centrally processed source domain dataset {X _{m_center} , Y _center } and Target domain data set {X _{s_center} , Y _center }; Then, based on the CPLS algorithm, the matrix X _{m_center} and Y _center are subjected to principal component analysis, and the matrix X _{s_center} is subjected to principal component analysis; The transition matrix M _{trans_pre} and the transition matrix M _trans are calculated again; Finally, predict the substance concentration variable of the measured object. The invention can clear the random noise measured by the main instrument, improve the data utilization rate and modeling accuracy, and reduce the time complexity.

Description

A calibration migration method for infrared spectroscopy measuring instruments based on CPLS

technical field

The invention relates to the technical field of migration learning under a machine learning module, in particular to a calibration and migration method of an infrared spectrum measuring instrument based on CPLS.

Background technique

Near-infrared spectroscopy (NIRS) analysis technology has the advantages of simple instrument operation, fast data analysis, low cost, and no contamination of samples, and has been widely used in various fields. In the production process, the near-infrared spectroscopy analysis technology is used for modeling, because the measurement conditions and instrument hardware performance are often unstable, which will lead to the failure of the existing calibration model.

The main purpose of transfer learning is to extract classification or regression knowledge from one or more tasks in the source domain and apply this knowledge to the target domain task, if the knowledge of one task is successfully transferred to another task, then the new A model for the task can be obtained without too many new samples. Using the knowledge learned in one or more source domains to improve the learning performance of the target domain, it solves the problems of missing labels in the target domain, high label cost, and time-consuming learning process, and achieves the purpose of improving the learning performance.

The calibration migration method refers to the migration of multivariate calibration models under different measuring instruments or measurement states. This method uses the linear relationship between spectral data from different sources to convert spectral samples measured in new instruments or in new states, and then directly use the original model to predict new samples. Migration research can be applied to related fields instead of the same field, to achieve useful information for migration and inter-domain conversion, so that the validity of the original model can be maintained or the original information can be used to speed up the modeling speed and avoid using a large number of targets. Domain samples or models again sample or model the target domain, thereby increasing the effectiveness of the model, reducing costs to a large extent and speeding up modeling.

The existing calibration migration methods have problems such as low prediction accuracy and limited application occasions. For example, in the calibration migration method based on PLS, partial least-regression (PLS) is one of the commonly used algorithms in data information extraction and process monitoring. The process variables are divided, and the process variables and quality variables are converted into the main subspace and the residual subspace, which realizes the compression and extraction of data. However, the PLS algorithm first uses the principal component analysis method to extract the principal components of the process variable and the quality variable respectively, and the two principal components are not related. It defaults to all process variables acting on quality variables, ignoring state information for internal variables. In many cases, due to the lack of excitation of the process data, there are a large number of unmeasured process and quality disturbances, when the remaining information of the quality variable changes, the alarm failure phenomenon occurs, resulting in poor PLS prediction output. In fact, the monitoring of changes in quality variable information is more important than process variables. On the other hand, the optimization goal involved in building a PLS model is to maximize the principal component correlation between the process variable and the quality variable without being constrained by the residual, so that the residual variance between the process variable and the quality variable to reach maximum. The variables are not guaranteed to be minimal, which may result in a large amount of information remaining on process variables and quality variables. Furthermore, the current near-infrared spectrum modeling processing data volume is large, the time complexity of the serial partial least squares algorithm is high, and the training and testing process is long.

SUMMARY OF THE INVENTION

Aiming at the problems existing in the prior art, the present invention provides a calibration and migration method for an infrared spectrum measuring instrument based on CPLS, which can remove random noise measured by the main instrument, improve data utilization and modeling accuracy, and reduce time complexity.

The technical scheme of the present invention is:

A kind of infrared spectroscopic measuring instrument calibration migration method based on CPLS, is characterized in that, comprises the following steps:

Step 1: Correspond the main infrared spectrum measurement instrument to the source domain and the infrared spectrum measurement slave instrument to the target domain. Use the infrared spectrum measurement master instrument and the infrared spectrum measurement slave instrument to collect the spectrum of each sample, and obtain the main spectrum and the slave spectrum respectively. spectrum, extract spectral data at intervals of anm in the wavelength range for the main spectrum and the secondary spectrum, and collect the variable value of the substance concentration of each sample to obtain the source domain data set {X _m , Y} and the target domain data set {X _s , Y};

Wherein, X _m =(X _m1 ,X _m2 ,...,X _mi ,...,X _mI ) ^T , X _mi =(x _mi1 ,x _mi2 ,...,x _mij ,...,x _miJ ), X _s =(X _s1 ,X _s2 ,...,X _si ,...,X _sI ) ^T ,X _si =(x _si1 ,x _si2 ,...,x _sij ,..., x _siJ ), x _mij , x _sij are the j-th master spectral data and slave spectral data of the ith sample, respectively, i=1,2,...,I, j=1,2,...,J , I is the total number of samples, J is the total number of spectral data points extracted; Y=(Y ₁ ,Y ₂ ,...,Y _i ,...,Y _I ) ^T ,Y _i =(y _i1 ,y _i2 , ...,y _ik ,...,y _iK ), y _ik is the value of the kth substance concentration variable of the ith sample, k=1,2,...,K, K is the total number of substance concentration variables ;

Step 2: Centralize the source domain data set and the target domain data set, and obtain the centrally processed source domain data set {X _{m_center} , Y _center } and target domain data set {X _{s_center} , Y _center };

Step 3: Perform principal component analysis on the matrices X _{m_center} and Y _center based on the CPLS algorithm:

Step 3.1: Based on the PLS algorithm, establish a calibration model Y _center = X _{m_center} B for the data set {X _{m_center} , Y _center }, and calculate the coefficient matrix B, the score matrix T of X _{m_center} , the load matrix P of X _{m_center} , and the score of Y _center Load matrix Q of matrix U, Y _center , introduce matrix R to make T=X _{m_center} R, and determine the number of latent variables l;

Step 3.2: Calculate the predictable substance concentration variable as

Singular value decomposition of the predictable species concentration variable yields

Among them, U _c is a left singular matrix, D _c is a singular value diagonal matrix, V _c is a right singular matrix, and V _c is an orthogonal matrix; Q _c =V _c D _c ^T , including _lc non-zero singularities in descending order value and the corresponding right singular vector;

It can be obtained by formula (2)

get

R _c =RQ ^T V _c D _c ^-1 (4)

Step 3.3: Calculate the unpredictable substance concentration variable as

Principal component extraction is performed on the unpredictable substance concentration variable, and _ly principal components are obtained as

为

的输出残差矩阵；in,

for

The output residual matrix of ;

The matrix is obtained by formula (6)

Step 3.4: Through the projection of R _c in space, the input variable independent of the substance concentration variable is obtained as

Wherein, R _c ^* = (R _c ^T R _c ) ^-1 R _c ^T ;

Principal component extraction is performed on the input variables unrelated to the substance concentration variable, and the number of l _x principal components is

为

的输入残差矩阵；in,

for

The input residual matrix of ;

The matrix is obtained by formula (8)

Step 3.5: From step 3.1 to step 3.4, the principal components extracted by the PLS algorithm of X _{m_center} and Y _center are respectively X _{m_pre} =TP ^T , Y _pre =UQ ^T , and the residuals of X _{m_center} and Y _center are respectively X _{m_res_c} =X _{m_center} -X _{m_pre} , Y _{res_c} =Y _center -Y _pre , that is, to get

Step 4: adopt the same method as in step 3 to carry out principal component analysis on the matrix X _{s_center} , and obtain the residual of X _{s_center} as X _{s_res_c} ;

Step 5: Calculate the score T m_pre =X _{m_center} R of the source domain data set after the principal components are extracted from the main spectrum by the PLS algorithm, calculate the score T _{s_pre} =X _{s_center} R of the target domain data set _after the principal components are extracted from the spectrum by the PLS algorithm, according to T _{m_pre} , T _{s_pre} calculate the transition matrix M _{trans_pre} based on the least squares method; calculate the score T _m =X _{m_res_c} P of the source domain data set after the main spectral pair residuals extract the principal components, calculate the target domain after extracting the principal components from the spectral pair residuals The score of the dataset T _s =X _{s_res_c} P, and the transition matrix M _trans is calculated based on the least squares method according to T _m and T _s ;

Step 6: Predict the substance concentration variable of the measured object:

Step 6.1: use infrared spectrum measurement to collect the spectrum of the measured object from the instrument, extract the spectral data using the same method as in step 1, and obtain J matrix X _{s_test} formed from the spectral data of the measured object;

Step 6.2: perform principal component analysis on X _{s_test} based on the CPLS algorithm, and obtain the residual of X _{s_test} as X _{s_res_c_test} ;

Step 6.3: The matrix formed by predicting the substance concentration variables of the tested object is Y _{test_predict} =(X _{s_test} *R*M _{trans_pre} *P ^T +X _{s_res_c_test} *R*M _trans *P ^T )*B.

Further, in the step 1, the sample is grain, the spectral data is absorbance, and the substance concentration variable includes the moisture content, oil content, protein content, and starch content of the grain.

The beneficial effects of the present invention are:

Based on the CPLS algorithm, the invention extracts the principal components of the source domain data set and the target domain data set once, and then extracts the principal components for the residuals again. On the basis of the two principal component extractions, the transfer matrix is calculated, and the measurement of the main instrument is eliminated. The random noise improves the data utilization and modeling accuracy, reduces the time complexity, and improves the speed of training and testing.

Description of drawings

FIG. 1 is a flow chart of a method for calibrating migration of an infrared spectrometer measuring instrument based on CPLS of the present invention.

FIG. 2 is a flow chart of performing principal component analysis on a source domain data set based on CPLS in the method for calibrating and migrating an infrared spectrometer measuring instrument based on CPLS of the present invention.

FIG. 3 is a flow chart of finding a transfer matrix in the method for calibrating and transferring an infrared spectrometer measuring instrument based on CPLS of the present invention.

FIG. 4 is a flow chart of predicting the substance concentration variable of the measured object in the method for calibrating migration of an infrared spectrometer measuring instrument based on CPLS of the present invention.

FIG. 5 is a schematic diagram of the cross-validation error of the oil content on the corn dataset in the specific embodiment as a function of the principal component fraction.

FIG. 6 is a fitting result diagram of mp6spec-mp5spec in the specific embodiment.

FIG. 7 is a graph of the fitting result of m5spec-mp5spec in a specific embodiment.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

The invention proposes a calibration migration method of an infrared spectrum measuring instrument based on CPLS. Because in data processing, PLS simply performs principal component extraction on X and Y, but usually the residuals of X and Y also contain valid information. Due to insufficient extraction, the established model has a large error. Therefore, a parallel method is proposed. Partial least squares (Concurrent PLS, CPLS) algorithm, on the basis of PLS, extracts the principal components of the residual again, so that the model error is smaller and the linear relationship is closer to the real situation. However, in reality, the collection of samples is very expensive and time-consuming, so transfer learning is proposed on the basis of CPLS, and the prediction of the test set of the target domain is completed by establishing a mapping relationship between the standard set of the source domain and the target domain.

The CPLS algorithm adopted by the present invention further improves the PLS algorithm, and performs principal component analysis on the process variable information that is not related to the quality variable and the quality of the information that cannot be predicted separately, and is divided into 5 subspaces: the subspace of the information related to the process variable and the quality variable. (Relevant Pivot Subspace), Process Variable Pivot Space, Process Variable Residual Space, Quality Variable Pivot Space, Quality Variable Residual Subspace.

The CPLS model achieves three goals: (1) extract scores directly related to predictable changes in the output from standard PLS projections, and these score vectors constitute the covariant subspace (CVS); (2) further convert the unpredicted The output changes are projected to the output principal component subspace (OPS) and the output residual subspace (ORS) to monitor abnormal changes in these subspaces; (3) input changes unrelated to the predicted output are further projected to the input principal component subspace (IPS) and Input Residual Subspaces (IRS) to monitor anomalous changes in these subspaces.

The CPLS algorithm sets the process variable data into two main parts, one of which is the information related to the quality variable, and the other part is the information not related to the quality variable. The quality variable data is also divided into two main parts, one is the information that can be predicted by the process variable, and the other is the information that cannot be predicted by the process variable. Therefore, the CPLS-based monitoring method provides a complete monitoring framework capable of monitoring process variables and quality variables as well as other parts of the information.

As shown in Figure 1, the CPLS-based infrared spectroscopy measuring instrument calibration migration method of the present invention comprises the following steps:

Step 1: Correspond the main infrared spectrum measurement instrument to the source domain and the infrared spectrum measurement slave instrument to the target domain. Use the infrared spectrum measurement master instrument and the infrared spectrum measurement slave instrument to collect the spectrum of each sample, and obtain the main spectrum and the slave spectrum respectively. spectrum, extract spectral data at intervals of anm in the wavelength range for the main spectrum and the secondary spectrum, and collect the variable value of the substance concentration of each sample to obtain the source domain data set {X _m , Y} and the target domain data set {X _s , Y};

Wherein, X _m =(X _m1 ,X _m2 ,...,X _mi ,...,X _mI ) ^T , X _mi =(x _mi1 ,x _mi2 ,...,x _mij ,...,x _miJ ), X _s =(X _s1 ,X _s2 ,...,X _si ,...,X _sI ) ^T ,X _si =(x _si1 ,x _si2 ,...,x _sij ,..., x _siJ ), x _mij , x _sij are the j-th master spectral data and slave spectral data of the ith sample, respectively, i=1,2,...,I, j=1,2,...,J , I is the total number of samples, J is the total number of spectral data points extracted; Y=(Y ₁ ,Y ₂ ,…,Y _i ,…,Y _I ) ^T ,Y _i =(y _i1 ,y _i2 ,…,y _ik ,...,y _iK ), y _ik is the value of the kth substance concentration variable of the ith sample, k=1,2,...,K, K is the total number of substance concentration variables.

In this embodiment, the sample is corn in cereals, the spectral data is the absorbance, and the substance concentration variables include the moisture content, oil content, protein content, and starch content of the corn. The data measured by the three spectrometers on the same I=80 samples constituted the corn dataset. The infrared spectrum is measured every a=2nm in the wavelength range of 1100-2498nm with the infrared spectrum measuring instruments m5, mp5, and mp6, with a total of J=700 attributes. The main spectrum of the first experiment - the secondary spectrum is m5spec-mp6spec, that is, the spectrum measured by m5 is used as the main spectrum, and the corresponding spectral data set is used as the initial source domain data set; due to the difference between the spectrum measured by mp6 and that measured by m5 The larger one is selected from the spectrum, and the corresponding spectral dataset is used as the initial target domain dataset. Then, five more experiments were performed in sequence on mp5spec-mp6spec, mp6spec-mp5spec, m5spec-mp5spec, mp5spec-m5spec, and mp6spec-m5spec.

In this embodiment, the Kennard-Stone (KS) algorithm is used to segment the corn dataset. First, 20% of the data in the initial source domain dataset and the initial target domain dataset are extracted as test samples, which are data of 16 samples respectively. The calibration transfer model is tested with test samples from the target domain. Then, the remaining 80% of the data in the initial source domain dataset and the initial target domain dataset are extracted as training samples, which are data of 64 samples respectively. Use the training samples of the source domain to establish a reference model to predict the migration samples of the target domain; and use the training samples of the target domain to establish a standard model of the target domain, so as to compare the performance of other migration models. Next, from the training samples of the source domain and the training samples of the target domain, KS algorithm is used to extract 20% of the data to form the standard sample set of the source domain and the standard sample set of the target domain, respectively, as the source used in the method of the present invention. Domain dataset {X _m , Y} and target domain dataset {X _s , Y} to establish the transfer relationship between source domain samples and target domain samples.

Step 2: Centralize the source domain data set and the target domain data set, that is, calculate the mean value of each column of data, and then subtract the mean value of the column from the original data of each column to obtain the centrally processed source domain data Set {X _{m_center} , Y _center } and target domain dataset {X _{s_center} , Y _center }, which can effectively avoid the deviation caused by large numerical differences.

Step 3: As shown in Figure 2, perform principal component analysis on the matrices X _{m_center} and Y _center based on the CPLS algorithm:

Step 3.1: Based on the PLS algorithm, establish a calibration model Y _center = X _{m_center} B for the data set {X _{m_center} , Y _center }, and calculate the coefficient matrix B, the score matrix T of X _{m_center} , the load matrix P of X _{m_center} , and the score of Y _center Load matrix Q of matrix U, Y _center , introduce matrix R to make T=X _{m_center} R, and determine the number of latent variables l;

Step 3.2: Calculate the predictable substance concentration variable as

The Singular Value Decomposition (SVD, Singular Value Decomposition) decomposition of the predictable substance concentration variable yields

Among them, U _c is a left singular matrix, D _c is a singular value diagonal matrix, V _c is a right singular matrix, and V _c is an orthogonal matrix; Q _c =V _c D _c ^T , including _lc non-zero singularities in descending order value and the corresponding right singular vector;

It can be obtained by formula (2)

get

R _c =RQ ^T V _c D _c ^-1 (4)

Step 3.3: Calculate the unpredictable substance concentration variable as

Principal component extraction (PCA) is performed on the unpredictable _substance concentration variable, and the number of principal components is obtained as

为

的输出残差矩阵；in,

for

The output residual matrix of ;

The matrix is obtained by formula (6)

Step 3.4: Through the projection of R _c in space, the input variable independent of the substance concentration variable is obtained as

Wherein, R _c ^* = (R _c ^T R _c ) ^-1 R _c ^T ;

Principal component extraction is performed on the input variables unrelated to the substance concentration variable, and the number of l _x principal components is

为

的输入残差矩阵；in,

for

The input residual matrix of ;

The matrix is obtained by formula (8)

Step 3.5: From step 3.1 to step 3.4, the principal components extracted by the PLS algorithm of X _{m_center} and Y _center are respectively X _{m_pre} =TP ^T , Y _pre =UQ ^T , and the residuals of X _{m_center} and Y _center are respectively X _{m_res_c} =X _{m_center} -X _{m_pre} , Y _{res_c} =Y _center -Y _pre , that is, to get

According to the algorithm flow of CPLS, it can be clearly seen that X _{m_center} and Y _center are divided into three parts: principal components extracted by the PLS algorithm, principal components extracted from residuals, and unpredictable errors. Compared with the PLS algorithm, the CPLS algorithm has the advantage of more processing of the principal components of the residual error extraction, which improves the data utilization rate.

Step 4: Perform principal component analysis on the matrix X _{s_center} using the same method as in Step 3, and obtain the residual of X _{s_center} as X _{s_res_c} .

In this embodiment, the analysis of the selection result of the optimal number of principal components of the PLS algorithm is as follows: the principal component number of the PLS method is selected by the 10-fold cross-validation method, taking the component of oil as an example, the corn data caused by the change of the number of principal components The variation of the cross-validation error of the oil content model in the training set of the centralized target domain is shown in Figure 5. As can be seen from Figure 5, the cross-validation error of oil on the corn set reaches the global minimum when the number of principal components is 12, so we set the optimal number of principal components for oil as 12. The optimal principal component fraction selection method for the other three components is the same as this method.

Step 5: As shown in Figure 3, use the least squares algorithm to establish a transition matrix that maps the latent structure of the target domain to the latent structure of the source domain: Calculate the score of the source domain dataset after the principal spectrum is extracted by the PLS algorithm T _{m_pre} =X _{m_center} R, calculate the score T _{s_pre} =X _{s_center} R of the target domain data set after extracting the principal components from the spectrum through the PLS algorithm, calculate the transition matrix M _{trans_pre} based on the least squares method according to T _{m_pre} , T _{s_pre} ; The score T _m =X _{m_res_c} P of the source domain dataset after the composition, calculate the score T _s =X _{s_res_c} P of the target domain dataset after extracting the principal components from the spectral pair residuals, calculate the transfer based on T _m , T _s based on the least squares method matrix M _trans .

Step 6: As shown in Figure 4, predict the substance concentration variable of the measured object:

Step 6.1: use infrared spectrum measurement to collect the spectrum of the measured object from the instrument, extract the spectral data using the same method as in step 1, and obtain J matrix X _{s_test} formed from the spectral data of the measured object;

Step 6.2: perform principal component analysis on X _{s_test} based on the CPLS algorithm, and obtain the residual of X _{s_test} as X _{s_res_c_test} ;

Step 6.3: The matrix formed by predicting the substance concentration variables of the tested object is Y _{test_predict} =(X _{s_test} *R*M _{trans_pre} *P ^T +X _{s_res_c_test} *R*M _trans *P ^T )*B.

In this embodiment, the model is used to predict the data, and the prediction error RMSEP results under different master-slave-instrument combinations in the corn data set are shown in Table 1 below:

Table 1

Analysis of Table 1 shows that: in general, the operation effect of the present invention between the spectrum mp5spec and the spectrum mp6spec is generally better than that of the other two groups. This is because the similarity between mp5spec and mp6spec is relatively high. The difference between the group and the spectral m5spec is relatively large, so it makes more sense to transfer learning between the two, so the error in the results is relatively small. It is not difficult to see that with mp6spec as the main spectrum and mp5spec as the secondary spectrum, the measurement errors of moisture, oil, protein and starch are basically the smallest among these six groups of experiments, while the migration results between m5spec and mp5spec and mp6spec are It is the largest error among the six groups.

As shown in FIG. 6 and FIG. 7 , the fitting result diagrams of mp6spec-mp5spec and m5spec-mp5spec in this embodiment are respectively. Comparing Figure 6 and Figure 7, it can be clearly seen that the fitting effect of the two groups is good or bad. Between spectral mp6spec and spectral mp5spec, the similarity is high and the fitting degree is good. Comparing the transfer learning between spectral m5spec and spectral mp5spec, it can be seen that most points of the former fall near the fitting line or on the fitting line. All the points of the latter fall below the fitted straight line, indicating that the effect of the former transfer learning is significantly better than that of the latter, and there is no need to transfer between the two spectra of the latter, because the prediction effect is not good at all.

Since the migration effect between the spectra mp6spec-mp5spec is the best, this group of spectra is selected for experiments and compared with other algorithms. The other algorithms described here are: Multiplicative Scatter/Signal Correction (MSC), typical Correlation analysis (Canonical Correlation Analysis, CCA), Slope and Bias Correction (Slope and Bias Correction, SBC), Piecewise Direct Standardization (Piecewise Direct Standardization, PDS). As shown in Table 2, it is the RMSEP comparison result of mp6spec-m5spec in the corn dataset under each algorithm. As can be seen from Table 2, in general, the migration effect of the CPLS-based infrared spectroscopy measuring instrument calibration migration method of the present invention is very good: compared with MSC, CCA and PDS algorithms, the present invention predicts four components. They are far superior to these three algorithms; compared with the SBC algorithm, the prediction effect of moisture and oil content is better, but the prediction effect of protein and starch is not much different.

Table 2

In a word, through the six groups of experiments on the corn data set, according to the obtained experimental results, and comparing with the MSC algorithm, the CCA algorithm, the SBC algorithm and the PDS algorithm, it can be seen that the CPLS algorithm of the present invention is combined with the transfer learning. The prediction effect is similar to that of SBC, but far superior to MSC algorithm, CCA algorithm and PDS algorithm. It can be seen that the present invention removes the random noise measured by the main instrument, and improves the data utilization rate and modeling accuracy.

Obviously, the above-mentioned embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. The above embodiments are only used to explain the present invention, and do not constitute a limitation on the protection scope of the present invention. Based on the above-mentioned embodiments, all other embodiments obtained by those skilled in the art without creative work, that is, all modifications, equivalent replacements and improvements made within the spirit and principle of the present application, are fall within the scope of protection required by the present invention.

Claims (2)

1. A CPLS-based infrared spectrum measuring instrument calibration migration method is characterized by comprising the following steps:

step 1: the method comprises the steps of enabling an infrared spectrum measurement master instrument to correspond to a source domain, enabling an infrared spectrum measurement slave instrument to correspond to a target domain, collecting the spectrum of each sample by using the infrared spectrum measurement master instrument and the infrared spectrum measurement slave instrument, respectively obtaining a master spectrum and a slave spectrum, respectively extracting spectral data of the master spectrum and the slave spectrum at intervals anm within a wavelength range, collecting material concentration variable values of each sample, and obtaining a source domain data set { X _m Y and target Domain data set { X _s ,Y}；

Wherein, X _m ＝(X _m1 ,X _m2 ,...,X _mi ,...,X _mI ) ^T ，X _mi ＝(x _mi1 ,x _mi2 ,...,x _mij ,...,x _miJ )，X _s ＝(X _s1 ,X _s2 ,...,X _si ,...,X _sI ) ^T ，X _si ＝(x _si1 ,x _si2 ,...,x _sij ,...,x _siJ )，x _mij 、x _sij J, I is the total number of samples, and J is the total number of extracted spectral data points; y ═ Y ₁ ,Y ₂ ,...,Y _i ,...,Y _I ) ^T ，Y _i ＝(y _i1 ,y _i2 ,...,y _ik ,...,y _iK )，y _ik Is the value of the kth species concentration variable for the ith sample, K being 1,2The total number of concentration variables;

step 2: the source domain data set and the target domain data set are subjected to centralized processing to obtain a centralized source domain data set { X _{m_center} ,Y _center And a target domain data set { X } _{s_center} ,Y _center }；

And 3, step 3: CPLS algorithm based matrix X _{m_center} 、Y _center Performing principal component analysis:

step 3.1: data set { X) based on PLS algorithm _{m_center} ,Y _center Establishment of calibration model Y _center ＝X _{m_center} B, calculating to obtain a coefficient matrix B, X _{m_center} Score matrix T, X _{m_center} Load matrix P, Y _center Score matrix U, Y _center The matrix R is introduced so that T is X _{m_center} R, and determining the number l of the latent variables;

step 3.2: calculating a predictable substance concentration variable of

Performing singular value decomposition on predictable substance concentration variables to obtain

Wherein, U _c As a left singular matrix, D _c As diagonal matrix of singular values, V _c As a right singular matrix, V _c Is an orthogonal matrix; q _c ＝V _c D _c ^T Including l in descending order _c A plurality of non-zero singular values and corresponding right singular vectors;

obtained by the formula (2)

To obtain

R _c ＝RQ ^T V _c D _c ^-1 (4)

Step 3.3: calculating an unpredictable substance concentration variable as

Extracting main components from unpredictable substance concentration variables to obtain l _y The main component number is

Wherein,

is composed of

The output residual matrix of (3);

The matrix is obtained by equation (6)

Step 3.4: by spatially R _c Projection of an input variable independent of the material concentration variable as

Wherein R is _c ^* ＝(R _c ^T R _c ) ^-1 R _c ^T ；

Subjecting the input variable independent of the concentration variable of the substance to principal component extraction to obtain l _x The main component number is

Wherein,

is composed of

The input residual matrix of (3);

obtaining a matrix by equation (8)

Step 3.5: from step 3.1 to step 3.4, X is obtained _{m_center} 、Y _center The main components extracted by the PLS algorithm are respectively X _{m_pre} ＝TP ^T 、Y _pre ＝UQ ^T ，X _{m_center} 、Y _center Respectively have a residual error of X _{m_res_c} ＝X _{m_center} -X _{m_pre} 、Y _{res_c} ＝Y _center -Y _pre That is to obtain

And 4, step 4: applying the same method as in step 3 to the matrix X _{s_center} Performing principal component analysis to obtain X _{s_center} Has a residual error of X _{s_res_c} ；

And 5: calculating the score T of the source domain data set after the principal spectrum is extracted by the PLS algorithm _{m_pre} ＝X _{m_center} R, calculating the score T of the target domain data set after extracting the principal components from the spectrum by a PLS algorithm _{s_pre} ＝X _{s_center} R, according to T _{m_pre} 、T _{s_pre} Calculating transfer matrix M based on least square method _{trans_pre} (ii) a Calculating the score T of the data set of the source domain after extracting principal components from the residual error of the principal spectrum _m ＝X _{m_res_c} P, calculating the score T of the target domain data set after extracting the principal component from the spectrum pair residual error _s ＝X _{s_res_c} P, according to T _m 、T _s Calculating transfer matrix M based on least square method _trans ；

Step 6: predicting the substance concentration variable of the measured object:

step 6.1: collecting the spectrum of the measured object from the instrument by infrared spectrometry, and extracting the spectrum data by the same method as step 1 to obtain J matrixes X formed by the spectrum data of the measured object _{s_test} ；

Step 6.2: x pair based on CPLS algorithm _{s_test} Performing principal component analysis to obtain X _{s_test} Has a residual error of X _{s_res_c_test} ；

Step 6.3: the matrix formed by predicting the material concentration variable of the measured object is Y _{test_predict} ＝(X _{s_test} *R*M _{trans_pre} *P ^T +X _{s_res_c_test} *R*M _trans *P ^T )*B。

2. The CPLS-based Infrared Spectroscopy measurement instrument calibration migration method according to claim 1, wherein in the step 1, the sample is grain, the spectral data is absorbance, and the substance concentration variables comprise moisture content, oil content, protein content and starch content of grain.

Images