CN105842190A - Near-infrared model transfer method based on spectral regression - Google Patents

Abstract

The invention discloses a near-infrared model transfer method based on spectral regression. The method comprises the following steps: step 1, acquiring the near-infrared spectra of a plurality of tobacco calibration samples by using a host computer and a slave computer separately so as to obtain host-computer near-infrared spectra Xm and slave computer near-infrared spectra Xs; step 2, separately subjecting Xm and Xs to spectral pretreatment; step 3, respectively calculating expressions Zm and Zs of Xm and Xs at a low dimension by using a spectral regression method; step 4, calculating transform in virtue of Zm and Zs and transferring the host-computer near-infrared spectrum Xtrn of a tobacco sample in a training set into a slave computer near-infrared spectrum Xt; step 5, carrying out modeling by using the slave computer near-infrared spectrum Xt and the chemical value of the tobacco sample; and step 6, acquiring the near-infrared spectrum of a to-be-detected sample by using the slave computer and calculating the contents of chemical components of tobacco by using a model established in the step 5. The method provided by the invention can reveal the internal structure of data, inhibit noise and redundancy characteristics in data and improve the success probability of transfer of the infrared spectral model.

Description

A kind of method for transferring near infrared model returned based on spectrum

Technical field

The present invention relates to Infrared Spectrum Technology field, be specifically related to a kind of near-infrared model transfer returned based on spectrum Method.

Background technology

Infrared spectrum, because having quick, accurate and lossless advantage, is widely used in industrial circle.Spectrum Multivariate Correction Technology can be efficiently used for material component content detection and online process monitoring, but multivariate calibration techniques is in actual applications Usually can be limited to, this cannot be effectively applied to new environmental condition mainly due to the multivariate calibration model having built up The spectrogram that the infrared spectrum of lower observation or different instrument gather.

The method of re-calibrating can overcome this to limit to, but need to re-establish model every time, and not only cost is suitable Greatly, and waste time and energy.Model transfer is to solve above-mentioned model to promote a kind of effective ways of problem, and it will be on an instrument The qualitative or quantitative calibration model set up reliably is transplanted on other same or similar instrument use, or will be in a certain condition The spectrogram that the model of lower foundation gathers under the conditions of another for same instrument, its essence is to eliminate to measure differing between signal Cause property so that it is be suitable for same model.

Model transfer is a kind of method of spectrum transform, finds a transformation matrix and will be mapped to master from instrument spectral response Instrument, it is achieved model sharing, representational method has sub-space learning, direct standardization (DirectStandard, DS) and divides Section directly standardization (PiecewiseDirectStandard, PDS).

Common sub-space learning method has principal component analysis (PCA), locality preserving projections (LPP) and neighborhood to keep embedding (NPE) under etc., these sub-space learning methods can bring the Unified frame that figure embeds into.Sub-space learning method relates to Relatively big to the feature decomposition of dense matrix, amount of calculation and amount of storage, and when data dimension exceedes number of samples, algorithm pole is not Stable.

Directly the transformation matrix between principal and subordinate's instrument spectral is directly sought in standardized method.Piecewise direct standardization method In, certain wavelength of main instrument sets up transformation relation with from instrument corresponding local spectrum range.In existing spectrum transform method In, PDS method is most widely used algorithm, and this local regression model being primarily due in PDS method can reflect principal and subordinate Instrument spectral change at corresponding wavelength, PDS method is directly corrected based on spectral signal simultaneously, easy to use.

But PDS method needs to be determined in advance the size of mapping window, transfer easily occurs when window selection is incorrect Failure.

Summary of the invention

The invention provides a kind of method for transferring near infrared model returned based on spectrum, it is possible to disclose the internal junction of data Structure, the noise in suppression data and redundancy feature, improve the probability of success of infrared spectrum Model transfer.

A kind of method for transferring near infrared model returned based on spectrum, including:

Step 1, demarcates sample for multiple Nicotiana tabacum L.s, is utilized respectively main frame and carries out near infrared spectra collection from machine, obtaining Main frame near infrared spectrum X_mWith from machine near infrared spectrum X_s；

Step 2, to main frame near infrared spectrum X_mWith from machine near infrared spectrum X_sCarry out Pretreated spectra respectively；

Step 3, utilizes spectrum homing method, calculates main frame near infrared spectrum X respectively_mWith from machine near infrared spectrum X_sIn low-dimensional The expression Z of degree_mAnd Z_s；

Step 4, utilizes Z_mAnd Z_sCalculate conversion, by the main frame near infrared spectrum X of training set tobacco sample_trnTransfer is for from machine Near infrared spectrum X_t；

Step 5, utilizes from machine near infrared spectrum X_tAnd the chemical score of tobacco sample is modeled；

Step 6, utilizes, from machine, testing sample is carried out near infrared spectra collection, and the model then utilizing step 5 to set up enters The calculating of row tobacco components content.

The method for transferring near infrared model that the present invention provides shares problem for solving the model between different nir instrument, Return embedding by main frame and the near infrared spectrum from machine are carried out spectrum respectively, then look for converting the relation between spectrogram.Adopt The immanent structure of data, the noise in suppression data and redundancy feature can be disclosed with spectrum homing method, improve Model transfer The probability of success, reduces the forecast error of the model after transfer.

Use the inventive method by main engine modeling spectral translation for being modeled after machine modeling spectrum, it is to avoid to use from machine After sample is re-started measurement, then it is modeled, reduces the time cost and Financial cost analyzed.

Described main frame and from the near infrared spectrometer that machine is two same model, main frame is the near infrared spectrum having model Instrument, is the near infrared spectrometer not having model from machine.In step 1 when carrying out near-infrared spectral measurement, in identical test strip Carry out under part.

Near infrared spectrum is down to low-dimensional by the method that the present invention utilizes spectrum to return, it is to avoid dimension occurs more than during sample number Calculate unstable problem, additionally, present invention, avoiding existing department pattern transfer method to need application model information, directly Realize by main frame to from the transfer of machine.

Main frame near infrared spectrum is converted into after machine near infrared spectrum, utilizes chemical score to be modeled, then to from machine Sample utilizes the model set up to carry out the prediction of tobacco components content.

As preferably, step 3 comprises the following steps:

Step 3-1, builds main frame near infrared spectrum X respectively_mWith from machine near infrared spectrum X_sCorrelation matrix；

Step 3-2, is utilized respectively spectrum Regressive Solution main frame near infrared spectrum X_mWith from machine near infrared spectrum X_sFeature to Amount, finds h characteristic vector, and the maximum of h is the quantity of transfer correction spectrum；

Step 3-3, utilizes h characteristic vector of step 3 gained, solves main frame near infrared spectrum X respectively_mThe reddest with from machine External spectrum X_sProjection vector collection；

Step 3-4, utilizes following formula respectively to main frame near infrared spectrum X_mWith from machine near infrared spectrum X_sCarry out spectrum and return embedding Enter:

$Z_m=A_m^T{X}_m;$

$Z_s=A_s^T{X}_s;$

In formula, A_mProjection vector collection for main frame near infrared spectrum；

A_sFor from the projection vector collection of machine near infrared spectrum.

In step 3-1～step 3-4, with from machine near infrared spectrum, identical process side is used for main frame near infrared spectrum Formula, is respectively processed.

The chemical score of the tobacco sample in step 5 according to People's Republic of China's tobacco business standard YC/T32-1996, YC/T161-2002, YC/T160-2002, YC/T173-2003, YC/T162-2002, YC/T202-2006, detect the total of Nicotiana tabacum L. Sugar, reducing sugar, total nitrogen, nicotine, potassium, chlorine, polyphenol content.

Modeling method in step 5 is partial least squares algorithm (Partial Least Squares, PLS), latent variable number Purpose is chosen for 5 folding cross validations.

As preferably, in step 4, following formula is utilized to calculate transformation matrix F₁With transformation matrix F₂:

$F_1=Z_m^{+}{Z}_s;$

$F_2=Z_s^{+}{X}_s;$

In formula: the generalized inverse computing of+representing matrix.

As preferably, in step 4, utilize following formula by the main frame near infrared spectrum X of training set tobacco sample_trnTransfer for from Machine near infrared spectrum X_t:

$X_t=A_m^T{X}_{trn}{F}_1{F}_2.$

As preferably, in step 2, to main frame near infrared spectrum X_mWith from machine near infrared spectrum X_sMake identical spectrum to locate in advance Reason, Pretreated spectra is smooth, seeks first derivative, asks second dervative, at least one in standard normal correction.

As preferably, the element w of correlation matrix in step 3-1_ijComputing formula as follows:

$w_{ij}=\left\{\begin{array}{cc}{e}^{-\frac{\left|\right|x_i-x_j|{|}^2}{2{\mathrm{\σ}}^2}}& x_i\∈N\left(x_j\right)or{x}_j\∈N\left(x_i\right)\\ 0& otherwise\end{array}\right.;$

In formula: x_iNear infrared spectrum for tobacco sample i；

x_jNear infrared spectrum for tobacco sample j；

N(x_i) represent tobacco sample x_iNeighbour territory；

N(x_j) represent tobacco sample x_jNeighbour territory；

σ takes 0.1.

As preferably, in step 3-2, solve Wy=λ Dy, find h characteristic vector of maximum；

In formula, W is correlation matrix；

λ is regularization parameter:

D is a diagonal matrix, and its diagonal element is

As preferably, in step 3-3, utilize regularization least square method to solve projection vector, make projection vector meet bar Part X^TY=a, in formula, a is projection vector；X is that Nicotiana tabacum L. demarcates the machine near infrared spectrum of sample or from machine near infrared spectrum；Y is With X characteristic of correspondence vector.

As preferably, in step 3-3, computing formula when utilizing regularization least square method to solve projection vector is as follows:

$a=\underset{a}{\arg \min }\left|\right|X^{T}y-a|{|}^2+\mathrm{\γ}|\left|a\right|{|}^2$

In formula, a is projection vector；

X is that Nicotiana tabacum L. demarcates the machine near infrared spectrum of sample or from machine near infrared spectrum；

Y is and X characteristic of correspondence vector；

γ is regularization parameter.

The method for transferring near infrared model returned based on spectrum that the present invention provides, it is possible to disclose the internal structure of data, press down Noise in data processed and redundancy feature, improve the probability of success of infrared spectrum Model transfer.

Accompanying drawing explanation

Fig. 1 is the flow chart of the method for transferring near infrared model that the present invention returns based on spectrum；

Fig. 2 a is the main frame near infrared spectrum that Nicotiana tabacum L. demarcates sample；

Fig. 2 b be Nicotiana tabacum L. demarcate sample from machine near infrared spectrum；

Fig. 3 a is main frame near infrared spectrum and the difference from machine near infrared spectrum；

Fig. 3 b is to use PDS method to process aft engine near infrared spectrum and the difference from machine near infrared spectrum；

Fig. 3 c is for the main frame near infrared spectrum after using the method for transferring near infrared model of present invention offer to process with from machine The difference of near infrared spectrum.

Detailed description of the invention

Below in conjunction with the accompanying drawings, the method for transferring near infrared model returned the present invention based on spectrum is described in detail.

As it is shown in figure 1, the method for transferring near infrared model returned based on spectrum, including:

Step 1, selects 33 Nicotiana tabacum L. samples to demarcate sample as Nicotiana tabacum L., for being respectively utilized respectively main frame and carrying out closely from machine Infrared spectrum gathers, and obtains main frame near infrared spectrum X_mWith from machine near infrared spectrum X_s, main frame and the instrument type from machine and adopt Sample parameter is consistent.Main frame near infrared spectrum as shown in Figure 2 a, from machine near infrared spectrum as shown in Figure 2 b.

Step 2, utilizes standard normal to correct main frame near infrared spectrum X_mWith from machine near infrared spectrum X_sCarry out identical pre- Process.

Step 3, utilizes spectrum homing method, calculates main frame near infrared spectrum X respectively_mWith from machine near infrared spectrum X_sIn low-dimensional The expression Z of degree_mAnd Z_s, concrete operations are as follows:

Step 3-1, builds main frame near infrared spectrum X respectively_mWith from machine near infrared spectrum X_sCorrelation matrix, wherein, Main frame near infrared spectrum X_mCorresponding correlation matrix is W_m, from machine near infrared spectrum X_sCorresponding correlation matrix is W_s。

If there being K Nicotiana tabacum L. to demarcate sample, then correlation matrix is the symmetrical matrix of a K × K, correlation matrix Element w_ijRepresent the dependency of tobacco sample i and tobacco sample j, the element w of correlation matrix_ijComputing formula as follows:

$w_{ij}=\left\{\begin{array}{cc}{e}^{-\frac{\left|\right|x_i-x_j|{|}^2}{2{\mathrm{\σ}}^2}}& x_i\∈N\left(x_j\right)or{x}_j\∈N\left(x_i\right)\\ 0& otherwise\end{array}\right.;$

In formula: x_iNear infrared spectrum for tobacco sample i；

x_jNear infrared spectrum for tobacco sample j；

N(x_i) represent tobacco sample x_iNeighbour territory, the present embodiment take 3 neighborhoods (i.e. distance tobacco sample x_iNearest 3 Tobacco sample)；

N(x_j) represent tobacco sample x_jNeighbour territory, the present embodiment takes 3 neighborhoods；

σ takes 0.1.

Step 3-2, solves main frame near infrared spectrum X respectively_mWith from machine near infrared spectrum X_sCharacteristic vector, find maximum H characteristic vector be respectively y₁y₂...y_h。

In this step, solve Wy=λ Dy, the quantity finding h characteristic vector, h maximum to be transfer calibration samples, this reality Execute example and take 30.H is the least, and dimensionality reduction have lost too much information, and typically when sample number is little, h value is more lower slightly than sample number is Can.

In formula, W is correlation matrix；

λ is regularization parameter, and the present embodiment takes 0.001；

D is a diagonal matrix, and its diagonal element is D_ii=Σ_jw_ji。

Step 3-3, utilizes step 3 gained characteristic vector, solves main frame near infrared spectrum X respectively_mWith from machine near infrared light Spectrum X_sProjection vector.

In this step, utilize regularization least square method to solve projection vector, make projection vector meet condition X^TY=a, formula In, a is projection vector；X is that Nicotiana tabacum L. demarcates the machine near infrared spectrum of sample or from machine near infrared spectrum；Y is the spy corresponding with X Levy vector.

Such as, X_mProjection vector be a₁a₂…a_h, X_mProjection vector collection A_m=(a₁,a₂,…,a_h), meet condition X_m ^Ty_i =a_i, the span of i is 1～h.

Computing formula when utilizing regularization least square method to solve projection vector is as follows:

$a=\underset{a}{\arg \min }\left|\right|X^{T}y-a|{|}^2+\mathrm{\γ}|\left|a\right|{|}^2$

In formula, a is projection vector；

X is that Nicotiana tabacum L. demarcates the machine near infrared spectrum of sample or from machine near infrared spectrum；

Y is and X characteristic of correspondence vector；

γ is regularization parameter, takes 0.001 in the present embodiment.

Step 3-4, utilizes following formula respectively to main frame near infrared spectrum X_mWith from machine near infrared spectrum X_sCarry out spectrum and return embedding Enter:

$Z_m=A_m^T{X}_m;$

$Z_s=A_s^T{X}_s;$

In formula, A_mProjection vector collection for main frame near infrared spectrum；

A_sFor from the projection vector collection of machine near infrared spectrum.

Step 4, utilizes Z_mAnd Z_sCalculate transformation matrix, by the main frame near infrared spectrum X of training set tobacco sample_trnTransfer is From machine near infrared spectrum X_t。

In this step, following formula is utilized to calculate transformation matrix F₁With transformation matrix F₂:

$F_1=Z_m^{+}{Z}_s;$

$F_2=Z_s^{+}{X}_s;$

In formula: the generalized inverse computing of+representing matrix.

In this step, utilize following formula by the main frame near infrared spectrum X of training set tobacco sample_trnTransfer is for from machine near-infrared Spectrum X_t:

$X_t=A_m^T{X}_{trn}{F}_1{F}_2.$

Step 5, utilizes from machine near infrared spectrum X_tAnd the chemical score of tobacco sample is modeled.

The chemical score of the tobacco sample in this step according to People's Republic of China's tobacco business standard YC/T32-1996, YC/T161-2002, YC/T160-2002, YC/T173-2003, YC/T162-2002, YC/T202-2006, detect the total of Nicotiana tabacum L. Sugar, reducing sugar, total nitrogen, nicotine, potassium, chlorine, polyphenol content.

Modeling method in this step is partial least squares algorithm (Partial Least Squares, PLS), latent variable Number be chosen for 5 folding cross validations.

Step 6, utilizes, from machine, testing sample is carried out near infrared spectra collection, and the model then utilizing step 5 to set up enters The calculating of row tobacco components content.

Fig. 3 a～Fig. 3 b represents that Nicotiana tabacum L. is demarcated sample and composes at main frame and the difference from machine respectively, Fig. 3 a～3b can see Going out, the transfer method using the present invention to provide can reduce Nicotiana tabacum L. and demarcate sample in main frame and the difference from machine.

Under different Model transfer methods, after transfer, spectrum forecast error in a model is as shown in table 1.

Table 1

In table 1, RMSEC: training set root mean square error；RMSECV: training set cross validation root mean square error；RMSEP: survey Examination collection root mean square error.

The method that the present invention provides compares traditional Model transfer method, utilizes spectrum to return, at low dimensional implementation model Transfer, it is to avoid because of the singularity in the calculating that dimension causes more than sample number, it is ensured that the model after transfer has good Application effect.

Claims (9)

1. the method for transferring near infrared model returned based on spectrum, it is characterised in that including:

Step 1, demarcates sample for multiple Nicotiana tabacum L.s, is utilized respectively main frame and carries out near infrared spectra collection from machine, obtaining main frame Near infrared spectrum X_mWith from machine near infrared spectrum X_s；

Step 2, to main frame near infrared spectrum X_mWith from machine near infrared spectrum X_sCarry out Pretreated spectra respectively；

Step 3, utilizes spectrum homing method, calculates main frame near infrared spectrum X respectively_mWith from machine near infrared spectrum X_sIn low dimensional Represent Z_mAnd Z_s；

Step 4, utilizes Z_mAnd Z_sCalculate conversion, by the main frame near infrared spectrum X of training set tobacco sample_trnTransfer is for the reddest from machine External spectrum X_t；

Step 5, utilizes from machine near infrared spectrum X_tAnd the chemical score of tobacco sample is modeled；

Step 6, utilizes, from machine, testing sample is carried out near infrared spectra collection, and the model then utilizing step 5 to set up carries out cigarette The calculating of grass chemical composition content.

2. the method for transferring near infrared model returned based on spectrum as claimed in claim 1, it is characterised in that step 3 include with Lower step:

Step 3-1, builds main frame near infrared spectrum X respectively_mWith from machine near infrared spectrum X_sCorrelation matrix；

Step 3-2, is utilized respectively spectrum Regressive Solution main frame near infrared spectrum X_mWith from machine near infrared spectrum X_sCharacteristic vector, look for To h characteristic vector, the maximum of h is the quantity of transfer correction spectrum；

Step 3-3, utilizes h characteristic vector of step 3 gained, solves main frame near infrared spectrum X respectively_mWith from machine near infrared light Spectrum X_sProjection vector collection；

Step 3-4, utilizes following formula respectively to main frame near infrared spectrum X_mWith from machine near infrared spectrum X_sCarry out spectrum recurrence to embed:

$Z_m=A_m^T{X}_m;$

$Z_s=A_s^T{X}_s;$

In formula, A_mProjection vector collection for main frame near infrared spectrum；

A_sFor from the projection vector collection of machine near infrared spectrum.

3. the method for transferring near infrared model returned based on spectrum as claimed in claim 2, it is characterised in that in step 4, utilizes Following formula calculates transformation matrix F₁With transformation matrix F₂:

$F_1=Z_m^{+}{Z}_s;$

$F_2=Z_s^{+}{X}_s;$

In formula: the generalized inverse computing of+representing matrix.

4. the method for transferring near infrared model returned based on spectrum as claimed in claim 3, it is characterised in that in step 4, utilizes Following formula is by the main frame near infrared spectrum X of training set tobacco sample_trnTransfer is for from machine near infrared spectrum X_t:

$X_t=A_m^T{X}_{trn}{F}_1{F}_2.$

5. the method for transferring near infrared model returned based on spectrum as claimed in claim 4, it is characterised in that in step 2, to master Machine near infrared spectrum X_mWith from machine near infrared spectrum X_sMaking identical Pretreated spectra, Pretreated spectra is smooth, asks single order to lead Number, asks second dervative, at least one in standard normal correction.

6. the method for transferring near infrared model returned based on spectrum as claimed in claim 5, it is characterised in that phase in step 3-1 Close the element w of coefficient matrix_ijComputing formula as follows:

$w_{ij}=\left\{\begin{array}{cc}{e}^{-\frac{\left|\right|x_i-x_j|{|}^2}{2{\mathrm{\σ}}^2}}& x_i\∈N\left(x_j\right)or{x}_j\∈N\left(x_i\right)\\ 0& \mathrm{\ω}herwise\end{array}\right.;$

In formula: x_iNear infrared spectrum for tobacco sample i；

x_jNear infrared spectrum for tobacco sample j；

N(x_i) represent tobacco sample x_iNeighbour territory；

N(x_j) represent tobacco sample x_jNeighbour territory；

σ takes 0.1.

7. the method for transferring near infrared model returned based on spectrum as claimed in claim 6, it is characterised in that in step 3-2, ask Solve Wy=λ Dy, find h characteristic vector of maximum；

In formula, W is correlation matrix；

λ is regularization parameter；

D is a diagonal matrix, and its diagonal element is D_ii=∑_jw_ji。

8. the method for transferring near infrared model returned based on spectrum as claimed in claim 7, it is characterised in that in step 3-3, profit Solve projection vector by regularization least square method, make projection vector meet condition X^TY=a, in formula, a is projection vector；X is Nicotiana tabacum L. demarcates the machine near infrared spectrum of sample or from machine near infrared spectrum；Y is and X characteristic of correspondence vector.

9. the method for transferring near infrared model returned based on spectrum as claimed in claim 8, it is characterised in that in step 3-3, profit The computing formula solved during projection vector by regularization least square method is as follows:

$a=\underset{a}{\arg \min }\left|\right|X^{T}y-a|{|}^2+\mathrm{\γ}|\left|a\right|{|}^2$

In formula, a is projection vector；

X is that Nicotiana tabacum L. demarcates the machine near infrared spectrum of sample or from machine near infrared spectrum；

Y is and X characteristic of correspondence vector；

γ is regularization parameter.