CN108152239A - The sample composition content assaying method of feature based migration - Google Patents
The sample composition content assaying method of feature based migration Download PDFInfo
- Publication number
- CN108152239A CN108152239A CN201711334521.0A CN201711334521A CN108152239A CN 108152239 A CN108152239 A CN 108152239A CN 201711334521 A CN201711334521 A CN 201711334521A CN 108152239 A CN108152239 A CN 108152239A
- Authority
- CN
- China
- Prior art keywords
- elm
- models
- source domain
- feature
- sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 62
- 238000013508 migration Methods 0.000 title claims abstract description 41
- 230000005012 migration Effects 0.000 title claims abstract description 41
- 235000002017 Zea mays subsp mays Nutrition 0.000 claims abstract description 44
- 235000005824 Zea mays ssp. parviglumis Nutrition 0.000 claims abstract description 43
- 235000005822 corn Nutrition 0.000 claims abstract description 43
- 238000013526 transfer learning Methods 0.000 claims abstract description 9
- 238000013507 mapping Methods 0.000 claims abstract description 5
- 240000008042 Zea mays Species 0.000 claims description 42
- 239000011159 matrix material Substances 0.000 claims description 20
- 238000012549 training Methods 0.000 claims description 20
- 238000002790 cross-validation Methods 0.000 claims description 16
- 230000006870 function Effects 0.000 claims description 14
- 238000012546 transfer Methods 0.000 claims description 14
- 238000000605 extraction Methods 0.000 claims description 12
- 239000004576 sand Substances 0.000 claims description 11
- 230000004913 activation Effects 0.000 claims description 9
- 238000012512 characterization method Methods 0.000 claims description 6
- 238000002329 infrared spectrum Methods 0.000 claims description 6
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 241000196324 Embryophyta Species 0.000 claims description 3
- 239000003814 drug Substances 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 238000005096 rolling process Methods 0.000 claims 1
- 230000000694 effects Effects 0.000 abstract description 22
- 241000209149 Zea Species 0.000 abstract 2
- 238000001228 spectrum Methods 0.000 description 21
- 238000012360 testing method Methods 0.000 description 18
- 239000000470 constituent Substances 0.000 description 14
- 238000010586 diagram Methods 0.000 description 13
- 238000002474 experimental method Methods 0.000 description 10
- 229920002472 Starch Polymers 0.000 description 9
- 235000019698 starch Nutrition 0.000 description 9
- 239000008107 starch Substances 0.000 description 9
- 239000003921 oil Substances 0.000 description 8
- 235000019198 oils Nutrition 0.000 description 8
- 230000008859 change Effects 0.000 description 5
- 238000004611 spectroscopical analysis Methods 0.000 description 5
- 230000009466 transformation Effects 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 4
- 230000003595 spectral effect Effects 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000012937 correction Methods 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- 235000018102 proteins Nutrition 0.000 description 3
- 102000004169 proteins and genes Human genes 0.000 description 3
- 108090000623 proteins and genes Proteins 0.000 description 3
- 239000000126 substance Substances 0.000 description 3
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 3
- 238000010987 Kennard-Stone algorithm Methods 0.000 description 2
- 238000013480 data collection Methods 0.000 description 2
- 230000014509 gene expression Effects 0.000 description 2
- 238000012417 linear regression Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 230000017105 transposition Effects 0.000 description 2
- 240000007594 Oryza sativa Species 0.000 description 1
- 235000007164 Oryza sativa Nutrition 0.000 description 1
- 235000016383 Zea mays subsp huehuetenangensis Nutrition 0.000 description 1
- 239000004480 active ingredient Substances 0.000 description 1
- 238000003556 assay Methods 0.000 description 1
- 239000002285 corn oil Substances 0.000 description 1
- 235000005687 corn oil Nutrition 0.000 description 1
- 238000013501 data transformation Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 239000004615 ingredient Substances 0.000 description 1
- 239000010977 jade Substances 0.000 description 1
- 235000009973 maize Nutrition 0.000 description 1
- 230000015654 memory Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 235000004252 protein component Nutrition 0.000 description 1
- 235000019624 protein content Nutrition 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 235000009566 rice Nutrition 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
- G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
- G01N21/31—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry
- G01N21/35—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
- G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
- G01N21/27—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands using photo-electric detection ; circuits for computing concentration
- G01N21/274—Calibration, base line adjustment, drift correction
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
- G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
- G01N21/31—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry
- G01N21/35—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light
- G01N21/3563—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light for analysing solids; Preparation of samples therefor
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
- G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
- G01N21/31—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry
- G01N21/35—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light
- G01N21/359—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light using near infrared light
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N2201/00—Features of devices classified in G01N21/00
- G01N2201/12—Circuits of general importance; Signal processing
- G01N2201/129—Using chemometrical methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
Landscapes
- Physics & Mathematics (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Immunology (AREA)
- Pathology (AREA)
- Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Theoretical Computer Science (AREA)
- Investigating Or Analysing Materials By Optical Means (AREA)
Abstract
The invention discloses a kind of sample composition content assaying methods of feature based migration, using the calibration transfer learning method based on ELM AE models, by in the ir data Feature Mapping to the feature space of source domain of aiming field sample to be tested, then the component content of the aiming field sample to be tested is predicted using the sample composition content prediction model of source domain.Sample composition content assaying method using the present invention, precision of prediction higher.In order to verify the effect of the present invention, inventor verifies the calibration transfer learning method proposed by the present invention based on ELM AE models using corn and tablet data set, and simultaneously with PDS models, the prediction effect of SBC migration models compares.The result shows that:For corn data set and tablet data set, generally speaking, the estimated performance of TL ELM AE models of the invention is superior to PDS models and SBC migration models.
Description
Technical field
The present invention relates to a kind of sample composition content assaying methods of feature based migration, belong to sample composition assay
Technical field.
Background technology
Near-infrared spectrum technique is a kind of indirect analysis technology of quick nondestructive low cost, can be fast using infrared spectrometer
The near infrared spectrum of sample is measured fastly, in conjunction with the method for Chemical Measurement, it is established that the volume infrared spectrum of sample with effectively
Polynary peg model between constituent content, and then the response component of unknown sample can be predicted.And near infrared light
In the measurement process of spectrum, due to the change of the difference or measuring condition of measuring instrument, original polynary peg model can be caused
Effect is lost, and re-establishing model is the something taken time and effort or even is modeled again without feasibility sometimes.
More acceptable mode is to do calibration migration, for correcting the spectroscopic data of main instrument and another instrument (sub- instrument).Essence
On, it is exactly to convert the spectrum of sub- instrument, is allowed to appear more like the data of key light spectrometer, then can use key light spectrometer
Model it handle.
In the past few years, different calibration migrating technologies is developed, and can be mainly divided into three classes.The first kind is to work as
A small amount of new samples add in old sample set, and model is updated.But add in sample must be enough variation information.The
Two classes calibration moving method is the difference for reducing different measurement data, including multiplicative scatter correction (MSC), there is limit for length's unit impulse
Response filter (FIR), Orthogonal Signal Correction Analyze (OSC), Generalized Least Square (GLS) etc..Third class is standardized method.Pass through
Slope and bias correction (SBC), predicted value can be standardized, this is based on a hypothesis:Prediction between two spectrometers
It is worth there is linear correlation, the predicted value of new samples can be corrected by the slope biased with regression equation.In addition, sub- instrument
The spectrum of device is by correcting the spectrum become close to the standardization of main instrument, it is possible to original model to sub- instrument spectral
It is handled.Such as directly standardization (PS) and the piecewise direct standardization (PDS) proposed by Wang et al..Usually used is
Three classes demarcate migrating technology.But existing above-mentioned calibration moving method still has that component content precision of prediction is relatively low to ask
Topic.
Invention content
The object of the present invention is to provide a kind of sample composition content assaying methods of feature based migration, it can have
Effect solves problems of the prior art, improves the whole precision of prediction of sample composition content.
In order to solve the above technical problems, the present invention adopts the following technical scheme that:The sample composition of feature based migration contains
Quantity measuring method, using the calibration transfer learning method based on ELM-AE models, by the infrared spectrum number of aiming field sample to be tested
In feature space according to Feature Mapping to source domain, then the aiming field is treated using the sample composition content prediction model of source domain
The component content of sample is predicted.
Preferably, following steps are specifically included:
S1, using ELM-AE models (i.e. extreme learning machine own coding algorithm) to the ir data of source domain and aiming field
Carry out feature extraction;
S2 establishes the polynary calibration migration mould between polynary peg model and source domain and aiming field based on source domain feature
Type;
S3 acquires the ir data of aiming field sample to be tested, after carrying out feature extraction conversion, utilizes the polynary of source domain
Peg model predicts the component content of the aiming field sample to be tested.
The present invention is by using ELM-AE models (i.e. extreme learning machine own coding algorithm) to the infrared of source domain and aiming field
Spectroscopic data carries out feature extraction, and not only extraction rate is fast, joins without adjusting, global optimum, and can be used for Feature Dimension Reduction.
In the sample composition content assaying method of aforementioned feature based migration, ELM-AE moulds are used described in step S1
Type (i.e. extreme learning machine own coding algorithm) carries out feature extraction to the ir data of source domain and aiming field, especially by
In the following manner obtains:
HS=XS·WS
HT=XT·WT
Wherein, HSFor the feature of source domain, HTFor the feature of aiming field, XsExpression is measured infrared by the key light spectrometer of source domain
Spectroscopic data, XTRepresent the ir data by being measured from spectrometer;WSRepresent the projector space of source domain, WTRepresent aiming field
Projector space, WSAnd WTRespectively by { XS,XSAnd { XT,XTELM-AE models acquire (be specifically according to own coding algorithm, into
One hidden layer transformation of row, obtains transformation and the linear regression coeffficient of itself, finally obtains projection;The mode of own coding can be in nothing
It needs to obtain the feature (to itself being trained study) of itself in the case of y).
The polynary calibration migration models established between source domain and aiming field described in the step S2 of the present invention, that is, be to solve for
The following formula:
min||yT-XTWTMβ||2+||yS-XSWSβ||2
s.t.||wS||2=1, | | wT||2=1
Optimization problem (solve obtain feature transfer matrix M and regression coefficient β), wherein, { XS,ysRepresent source domain
Data set, { XT,yTRepresent aiming field data set, wherein XsIt is measured by key light spectrometer, XTBy being measured from spectrometer;WSRepresent source domain
Projector space, WTRepresent the projector space of aiming field, WSAnd WTRespectively by { XS,ysAnd { XT,yTELM- AE models ask
, M represents target domain characterization to the transfer matrix of source domain feature, the recurrence of polynary peg model of the β expressions based on source domain feature
Coefficient.
In the sample composition content assaying method of aforementioned feature based migration, described in step S2 based on source domain feature
Polynary peg model using pls models, utilize source domain characteristic { HS,ysEstablish polynary peg model acquisition regression coefficient
β, wherein, HSFor the feature of source domain, ysFor the sample composition content of source domain, so that content prediction accuracy higher.
In the present invention, the ELM-AE models (a kind of special circumstances for ELM models), by by the hidden of ELM models
The random weight w of node layeriWith biasing biAll orthogonalization, and cause ELM models input data and output data it is equal to get;
I.e.:
H=g (wsx+b),
H η=t=x,
wTW=I, bTB=I,
Wherein, w=[w1,...,wL] be orthogonalization random weight matrix, b=[b1,...,bL] it is the random of orthogonalization
Bias matrix.
Preferably, the output weight η of ELM-AE models is obtained by feature space and input data training:
If the number of input sample data is more than or equal to the number of hidden node, the output weight η of ELM-AE is:
If the number of input sample data is less than or equal to the number of hidden node, the output weight η of ELM-AE is:
For the ELM-AE (i.e. the number of hidden node and input number are the same) of countertype, then exporting weight η is:
η=H-1X
ηTη=I;
Wherein, H=[h1,h2,…hN] be ELM-AE hidden node output, X=[x1,x2,…xN] it is the defeated of ELM-AE
Enter data;C is regularization factors, and N is the number of input sample data.
The output weight η is the regression coefficient trained in own coding training process, and feature is equal to regression coefficient
Transposition.
In the above method, by adding regularization factors, so as to effectively promote the Generalization Capability of migration models, prevent
Over-fitting so that prediction result has more robustness.
H=[the h1,h2,…hN], according to formula h=g (ws), x+b by random initializtion input weight w and
B is biased, then carries out conversion acquisition.
It is furthermore preferred that the corresponding hidden node of RMSECV minimum values is chosen using the method for 10 folding cross validations in the present invention
Number be the best hidden node number of ELM-AE models so that the present invention the calibration based on ELM-AE models move
Learning method (i.e. TL-ELM-AE models) is moved with more robustness.
It is furthermore preferred that the present invention selects the regularization factors C using the method for 10 folding cross validations, choose
The corresponding regularization factors of RMSECV minimum values are optimum factor, so that the calibration based on ELM-AE models of the present invention
Transfer learning method (i.e. TL-ELM-AE models) is with more robustness.
Preferably, the activation primitive of heretofore described ELM-AE models uses tanh functions, so that of the invention
The calibration transfer learning method (i.e. TL-ELM-AE models) based on ELM-AE models with more robustness.
The sample composition content assaying method of above-mentioned feature based migration, the sample to be tested are plant or Chinese medicine.
Preferably, the sample to be tested be corn or tablet, content prediction accuracy higher.
In aforementioned method, in step S3, in order to realize effective prediction of the source domain to target numeric field data, need to Spectral Properties
Sign is converted:
HS'=HT·M
It establishes after the transformation model between the peg model of source domain and source domain and aiming field, it is possible to realize to aiming field
Effective prediction of sample, equation below:
yT=XTWTMβ。
Compared with prior art, the present invention uses calibration transfer learning method (the i.e. TL-ELM-AE based on ELM-AE models
Model), by the ir data Feature Mapping to the feature space of source domain of aiming field sample to be tested, then utilize source domain
Sample composition content prediction model predicts the component content of the aiming field sample to be tested, so as to avoid to aiming field
Sample models again, reduces modeling cost;Sample composition content assaying method using the present invention simultaneously, whole prediction essence
Spend higher.
In order to verify the effect of the present invention, inventor is based on ELM- using corn and tablet data set to proposed by the present invention
The calibration transfer learning method of AE models verified, and simultaneously with PDS models, the estimated performance of SBC migration models carries out
Compare.The result shows that:For corn data set and tablet data set, generally speaking, mould is migrated relative to PDS models and SBC
Type, the prediction effect of TL-ELM-AE models of the invention is more preferable, precision of prediction higher.
Description of the drawings
Fig. 1 is ELM-AE schematic network structures;
Fig. 2 is the realization procedure chart schematic diagram of TL-ELM-AE models;
Fig. 3 is the principal spectrum (A) of corn data set and from spectrum (B) and deviation spectrum (C);
Fig. 4 is the principal spectrum (A) of tablet data set and from spectrum (B) and deviation spectrum (C);
Fig. 5 is the hidden node number selection course schematic diagram of the TL-ELM-AE models of corn;
Fig. 6 is the RMSECP change schematic diagrams of regular terms factor C on corn data set;
Fig. 7 is sigmoid functional arrangements and tanh functional arrangements;
Fig. 8 is the comparison signal of water content actual value and predicted value under SBC, PDS, TL-ELM-AE model in corn
Figure;
Fig. 9 is the comparison signal of oil content content actual value and predicted value under SBC, PDS, TL-ELM-AE model in corn
Figure;
Figure 10 is the comparison signal of content of starch actual value and predicted value under SBC, PDS, TL-ELM-AE model in corn
Figure;
Figure 11 is that protein content comparison of actual value and predicted value under SBC, PDS, TL-ELM-AE model is shown in corn
It is intended to;
Figure 12 is the comparison schematic diagram of predicted value and actual value of the water content before and after calibration migration in corn;
Figure 13 is the comparison schematic diagram of predicted value and actual value of the oil content content before and after calibration migration in corn;
Figure 14 is the comparison schematic diagram of predicted value and actual value of the content of starch before and after calibration migration in corn;
Figure 15 is the comparison schematic diagram of predicted value and actual value of the protein content before and after calibration migration in corn;
Figure 16 is 1 content of active constituent actual value and ratio of predicted value under SBC, PDS, TL_ELM-AE model in tablet
Compared with schematic diagram;
Figure 17 is 2 content of active constituent actual value and ratio of predicted value under SBC, PDS, TL_ELM-AE model in tablet
Compared with schematic diagram;
Figure 18 is 3 content of active constituent actual value and ratio of predicted value under SBC, PDS, TL_ELM-AE model in tablet
Compared with schematic diagram;
Figure 19 is the comparison schematic diagram of predicted value and actual value of 1 content of active constituent before and after calibration migration in tablet;
Figure 20 is the comparison schematic diagram of predicted value and actual value of 2 content of active constituent before and after calibration migration in tablet;
Figure 21 is the comparison schematic diagram of predicted value and actual value of 3 content of active constituent before and after calibration migration in tablet.
The present invention is further illustrated with reference to the accompanying drawings and detailed description.
Specific embodiment
The embodiment of the present invention:The sample composition content assaying method of feature based migration, using based on ELM-AE models
Calibration transfer learning method, by the ir data Feature Mapping to the feature space of source domain of aiming field sample to be tested,
Then the component content of the aiming field sample to be tested is predicted using the sample composition content prediction model of source domain.
Optionally, the present invention specifically may comprise steps of:
S1, using ELM-AE models (i.e. extreme learning machine own coding algorithm) to the ir data of source domain and aiming field
Carry out feature extraction;
S2 establishes the polynary calibration migration mould between polynary peg model and source domain and aiming field based on source domain feature
Type;
S3 acquires the ir data of aiming field sample to be tested, after carrying out feature extraction conversion, utilizes the polynary of source domain
Peg model predicts the component content of the aiming field sample to be tested.
Optionally, ELM-AE models (i.e. extreme learning machine own coding algorithm) are used to source domain and mesh described in step S1
The ir data for marking domain carries out feature extraction, is obtained especially by the following manner:
HS=XS·WS
HT=XT·WT
Wherein, HSFor the feature of source domain, HTFor the feature of aiming field, XsExpression is measured infrared by the key light spectrometer of source domain
Spectroscopic data, XTRepresent the ir data by being measured from spectrometer;WSRepresent the projector space of source domain, WTRepresent aiming field
Projector space, WSAnd WTRespectively by { XS,XSAnd { XT,XTELM-AE models acquire (be specifically according to own coding algorithm, into
One hidden layer transformation of row, obtains transformation and the linear regression coeffficient of itself, finally obtains projection).
Optionally, the polynary calibration migration models established between source domain and aiming field described in step S2, that is, be to solve for
The following formula:
min||yT-XTWTMβ||2+||yS-XSWSβ||2
s.t.||wS||2=1, | | wT| | 2=1
Optimization problem, wherein, { XS,ysRepresent source domain data set, { XT,yTRepresent aiming field data set, wherein Xs
It is measured by key light spectrometer, XTBy being measured from spectrometer;WSRepresent the projector space of source domain, WTRepresent the projector space of aiming field, WS
And WTRespectively by { XS,ysAnd { XT,yTELM-AE models acquire, M represent target domain characterization to source domain feature transfer matrix,
β represents the regression coefficient of the polynary peg model based on source domain feature.
Optionally, the polynary peg model based on source domain feature described in step S2 uses pls models, utilizes source domain spy
Levy data { HS,ysPolynary peg model acquisition regression coefficient β is established, wherein, HSFor the feature of source domain, ysSample for source domain
Component content.
Optionally, the ELM-AE models (a kind of special circumstances for ELM models), by by the hidden layer of ELM models
The random weight w of nodeiWith biasing biAll orthogonalization, and cause ELM models input data and output data it is equal to get;
I.e.:
H=g (wsx+b),
T=x,
wTW=I, bTB=I,
Wherein, w=[w1,...,wL] be orthogonalization random weight matrix, b=[b1,...,bL] it is the random of orthogonalization
Bias matrix.
Optionally, the output weight η of ELM-AE models can be obtained by feature space and input data training:
If the number of input sample data is more than or equal to the number of hidden node, the output weight η of ELM-AE is:
If the number of input sample data is less than or equal to the number of hidden node, the output weight η of ELM-AE is:
For the ELM-AE of countertype, then exporting weight η is:
η=H-1X
ηTη=I;
Wherein, H=[h1,h2,…hN] be ELM-AE hidden node output, X=[x1,x2,…xN] it is the defeated of ELM-AE
Enter data;C is regularization factors, and N is the number of input sample data.
Optionally, the method that the present invention can utilize 10 folding cross validations chooses the corresponding hidden node of RMSECV minimum values
Number be the best hidden node number of ELM-AE models.
Optionally, the present invention can select the regularization factors C using the method for 10 folding cross validations, select
It is optimum factor to take the corresponding regularization factors of RMSECV minimum values.
Optionally, tanh functions may be used in the activation primitive of heretofore described ELM-AE models.
Optionally, the sample to be tested can be plant or Chinese medicine, it is particularly possible to be corn or tablet.
In order to verify the effect of the present invention, inventor has also carried out tests below research:
First, experimental situation and data
1st, experimental situation
All experimental arrangements can all be realized that software version is by python language on PC in the present invention
Python2.7, IDE eclipse, computer are configured to the Intel processor of 2.50GHZ, 4GB memories, Windows8 operations system
System.
2nd, experimental data
(1) corn data set
The data set is from http://www.eigenvector.com/Data/Data_sets.html is obtained.Corn data
Collection has 80 samples, corresponds to the content of four kinds of moisture, oil, starch, protein substances.Ir data collection respectively from
Tri- spectrometers of M5spec, MP5spec, MP6spec are the data that 1100-2498nm is measured using 2nm as interval in wave-length coverage.
Using the spectrum that M5spec is measured as principal spectrum, the spectrum that MP5spec, MP6spec are measured is from spectrum for this experiment.Spectrogram is such as
Shown in Fig. 3, wherein subgraph (A) represents key light spectrogram, and subgraph (B) is represented from spectrogram, subgraph (C) represent key light spectrogram with from light
The difference of spectrogram.Principal spectrum is can be seen that from the subgraph (C) of Fig. 3 and there are larger differences from spectrum.
Data set is divided into 64 training samples and 16 test samples by experiment using Kennard-Stone (KS) algorithms.
Wherein training set equally uses KS to extract 40 and is used as initial training collection, and 14 as checksum set training optimal parameter.
(2) tablet data set
International diffusing reflection meeting (IDRC) is in the Shootout data sets of publication in 2002, comprising two spectrometers in wave
The near infrared spectrum data of tablet sample that long 600-1898nm is measured with the interval of 2nm, comprising 155 calibration collection samples and
460 test set samples, each 650 variables of sample.The infrared spectrogram of tablet such as Fig. 4, wherein subgraph (A), (B), (C)
Principal spectrum is represented respectively, from spectrum and principal spectrum and from the SPECTRAL DIVERSITY figure between spectrum.
3rd, experimental method
1st, data prediction
Before training pattern, to data carry out centralization processing, avoid due to numerical value difference greatly caused by deviation.
2nd, parameter selection
The parameter of model can cause large effect for the performance of model, parameter selection it is good, the general of model can be made
It is more preferable to change performance.Cross validation is a kind of method for being used for Selecting All Parameters well.
In the present invention, ELM-AE choose best hidden node number when, using 10 folding cross validations method to it into
Row is chosen.Meanwhile in order to enable the Generalization Capability of model is more preferable, the selection of regularization factors C, C is added to, the present invention selects
The method of 10 folding cross validations selects the regularization factors of ELM-AE, chooses the canonical corresponding to minimum RMSECV
Change the factor as optimum factor.
2nd, model parameter selects
(1) selection of TL-ELM-AE hidden nodes number
Influence of the hidden node number of TL-ELM-AE for the prediction effect of model is larger, so hidden choosing
During the number of node layer, data set is divided into three parts by this experiment by taking corn data set moisture as an example, first with KS algorithms, instruction
Practice collection, checksum set and test set, wherein carry out 10 folding cross validations to checksum set, respectively to 10,50,100,150,200,
250,300,500,700,1000 nodes carry out cross validation, last result RMSECV such as Fig. 5 to cross validation.
As shown in figure 5, when hidden node is 300, the result of RMSECV is minimum, and prediction effect reaches best, institute
To set the number of hidden node on corn data set as 300.Same experiment is done on tablet data set, obtains hidden layer section
The number of point is 500.
(2) selection of the TL-ELM-AE regular terms factor
In TL-ELM-AE, the regular terms factor can effectively promote the Generalization Capability of migration models so that prediction result is more
With robustness.The optimal selection of the regular terms factor is also to carry out cross validation acquisition by checksum set.In order to optimize regular terms
Factor C, inventor considers following set of parameter, and carries out cross validation respectively:0、100、200、500、1000、 2000、
3000、4000、5000、6000、7000、8000、9000、10000、20000、30000、 50000、80000.Fig. 6 shows jade
Rice data set RMSECV is with the figure of regular terms factor variations.
As shown in Figure 6, when regular terms factor C is set as 20000, RMSECV is minimum, the table on corn data set
It is now most good.It is same to test on tablet data set (this experiment is using the active constituent 1 on tablet data set), work as canonical
When item factor C is set as 1, RMSECV is minimum, so regular terms factor C optimal on setting tablet data set is 1.Table 1 is shown
There are regularization factors and the error of the model prediction without regularization factors in two datasets.
The RMSEP of different regularization factors in 1 TL-ELM-AE two datasets of table
(3) selection of activation primitive
Existing activation primitive includes sigmoid functions and tanh functions, their mathematic(al) representation is respectively such as following public affairs
Shown in formula.
Sigmoid functions:
Tanh functions:
Wherein, tanh functions have stronger gradient, can be to avoid biasing gradient disappearance problem and very symmetrical, tanh
Function ratio sigmoid functions perform better.Table 2 illustrates RMSEP of the both the above function as activation primitive, corn data set
By taking moisture as an example, tablet data set is by taking the first active constituent as an example.
The RMSEP of different activation primitives in 2 TL-ELM-AE two datasets of table
Performance of the tanh functions on corn data set and tablet data set is all than sigmoid function as can be seen from Table 2
More preferably.Therefore, in the present invention, tanh functions are chosen for the activation primitive of TL-ELM models.
3rd, experimental result is to when analyzing
Present invention selection tablet data set and corn data set are tested.For the two data sets, ELM-AE extractions
Feature:Input hidden layer weight wiWith biasing bi- 0.1 to 0.1 number is all randomly generated, as described in foregoing teachings, corn and tablet
Hidden node number choose more excellent node 300 and 500 respectively by cross validation, activation primitive all selects tanh, regularization
Coefficient selects 20000 and 1 respectively after checksum set cross validation.
1st, corn data set analysis of experimental results
The present invention is tested using corn data set.3~table of table 6 respectively shows 3 kinds of methods on corn data set
As a result the RMSEP data of prediction clearly show that TL-ELM-AE (the calibration migrations i.e. based on ELM-AE models of the present invention
Learning method) there is predictablity rate more higher than PDS and SBC.As shown in the table, W represents the big of the window of PDS methods setting
Small, N represents to participate in the number of samples of training.For PDS algorithms, estimated performance is influenced by the number of master sample, so being
Preferably performance participates in influence of the number of training sample to model in various methods, inventor is provided with 10,15,20,
25th, 30,35 and 40 samples are compared.In the case that 3~table of table 6 respectively shows master sample number difference, maize seed water
Point, the prediction error of the content of oil, protein and four kinds of substances of starch in different models.It, can be with by 3~table of table 6
It was found that influence of the different size of window for model is not very big in PDS methods.With the increase of number of samples, PDS's
Precision of prediction is more stable, and the precision of prediction of TL-ELM-AE and SBC are higher and higher.But in general, TL- of the invention
The precision of prediction of ELM-AE models is than remaining two methods higher.
The prediction error of moisture in 3 corn of table
Table 3 shows the prediction error of three kinds of models on corn data set for moisture, as can be seen from Table 3 PDS
Algorithm stability is fairly good, and prediction error value is almost near 0.170, however compared with TL-ELM-AE, the prediction error of PDS
It is worth bigger than normal;The prediction error of SBC reduces with the increase of training sample number, but compared to TL-ELM-AE, prediction misses
Difference remains on larger.The error of TL-ELM-AE maintains a relatively low state as can be seen from Table 3, even if instruction
In the case of practicing sample less, estimated performance is still good, and (during such as N=15, prediction error is 0.072), this demonstrate that TL-
ELM-AE's has robustness.
The prediction error of oil content in 4 corn of table
Table 4 shows the prediction error of three kinds of models on corn data set for oil, as can be seen from Table 4 PDS
Algorithm stability is still fairly good, and lower than the prediction error of SBC algorithm, this demonstrate PDS algorithms on corn oil
Estimated performance be better than SBC, however compared with TL-ELM-AE, the prediction error value of PDS is still bigger than normal, and with training sample
The increase effect of this number is more obvious.
The prediction error of content of starch in 5 corn of table
Table 5 shows the error information that three kinds of methods predict starch on corn data set.This time experiment can be seen
Go out window size still has what is centainly influenced for PDS, and when window size is set as 1, the estimated performance of PDS is significantly than remaining
Window size estimated performance it is more preferable.The estimated performance of SBC algorithms is general, although prediction error is with the increase of training sample
And reduce, but fluctuation is very big, illustrates that SBC is not sufficiently stable, robust performance is poor.TL-ELM-AE models are relatively stablized,
Between prediction error maintains 0.300-0.650, and with the increase of training sample, prediction error is gradually reduced.
The prediction error of protein content in 6 corn of table
Table 6 shows that three kinds of methods are for the difference data of the prediction of protein content on corn data set.SBC algorithm tables
It is now still worst, PDS algorithms and TL-ELM-AE models still quite stable, but showing for PDS algorithms will be slightly in this time experiment
It is better than TL-ELM-AE models.
In general, for TL-ELM-AE models, with the increase of master sample number, RMSEP is gradually reduced, TL-
The precision of prediction of ELM-AE models is higher, and stability is preferable.And under 40 master samples, the RMSEP of three kinds of methods is relatively low.
So in order to it is more fair, more intuitively show prediction effect between these methods, this experiment has selected 40 standards
The Sample Establishing transformational relation between source domain and aiming field, the window size that wherein PDS is selected are 1.Fig. 8 to Figure 11 is provided
Corresponding corn Zhong Shui, oil, starch and protein component each algorithm predicted value and actual value between transformational relation, prediction
Value is closer to actual value, then the point of corresponding standard is just closer to this straight line of x=y, therefore can be corresponded to according to various algorithms
Mark intensity of the point near straight line x=y judge the performance of the prediction of algorithm, and then can more intuitively observe
Go out their prediction effect.Again because the standard point of three kinds of algorithmic notations on picture compares set, although the intuitive amount of being difficult to
Change the difference between algorithm.Therefore the prediction error of predicted value in being illustrated in table 7 corresponding to Fig. 8 to Figure 11.
Prediction error of the 7 each component content of corn data set of table under different models
From Fig. 8~Figure 11 it can be found that generally speaking, prediction effect of the prediction error than SBC of TL-ELM-AE and PDS
More preferably.From table 7 it is also seen that coming, the estimated performance of TL-ELM-AE models is more preferable.
Finally, prediction is directly carried out to the spectroscopic data do not converted using source domain model with utilizing TL-ELM-AE moulds
The algorithm that type is predicted is compared.And then it can intuitively observe the effect after migration.Figure 12~Figure 15 gives pair
It should be in the predicted value and actual value of data that the moisture of corn data set, oil, starch and protein content are not migrated
The predicted value and the comparison figure of actual value for comparing figure and being migrated using TL_ELM-AE models.Predicted value is closer to true
Value, then the point of corresponding standard is just closer to this straight line of x=y.In Figure 12~Figure 15, cross star representation is not migrated
When, the relationship between aiming field test sample actual value and predicted value, triangle expression has used TL_ELM-AE models to be moved
The relationship between aiming field predicted value and actual value after shifting.
It can be seen that in moisture, starch and protein from Figure 12~Figure 15, cross star all substantial deviation straight line x=y,
And triangle is all concentrated near straight line x=y, this illustrates directly to predict target numeric field data using source domain model,
It will appear bigger deviation, this deviation is introduced by different measuring instruments, and TL-ELM-AE models is being used to be migrated
Afterwards, the deviation between source domain data and target numeric field data is largely reduced as seen from the figure, and then can be direct
Transformed target numeric field data is predicted, and obtain good prediction effect using source domain model.
2nd, tablet data set analysis of experimental results
The present invention is tested using tablet data set.8~table of table 10 respectively shows SBC, PDS and TL-ELM-AE tri-
The RMSEP data that kind algorithm is predicted on first, second, third active constituent of tablet data set.
It is as shown in the table, and W represents the size of the window of PDS methods setting, and N represents to participate in the number of samples of training.Inventor
There is provided 10,15,20,25,30,35,40,45,50,55 and 60 samples to be compared.
From 8~table of table 10 as can be seen that the prediction error of TL-ELM-AE will be less than PDS, and the prediction error of PDS is less than
SBC.This illustrates that the estimated performance of the TL-ELM-AE models of the present invention is better than PDS, and the estimated performance of PDS is better than SBC.
In addition to this, for SBC with the increase of training sample, prediction error is higher and higher instead, illustrates that SBC algorithms are not sufficiently stable, does not have
There is robustness.In table 10, for the third ingredient of tablet data set, PDS's puts up the best performance, but with the increasing of training sample
Add, the prediction error of TL-ELM-AE is become better and better, even better than the prediction error of PDS and SBC algorithms.Therefore, in general,
The prediction effect of TL-ELM-AE is more preferable.
It can be seen that the size of PDS windows is not very big, and PDS to its impact effect by data in 8~table of table 10
The experimental result RMSEP of acquisition relatively stablizes, and the effect fluctuation of SBC is larger, and in 8~table of table 9, its prediction error is very
To increasing with the increase of training data, this illustrates that the stability of its algorithm is not fine.And TL-ELM-AE models is pre-
It is smaller to survey error, it is whole more excellent.
The prediction error of the first active constituent in 8 tablet of table
The prediction error of second active ingredient in 9 tablet of table
The prediction error of the third active constituent in 10 tablet of table
Figure 16, Figure 17, Figure 18 are respectively shown in N=60, SBC corresponding to three kinds of active constituents of tablet data set,
Actual value and the comparison figure of predicted value under tri- kinds of models of PDS, TL_ELM-AE.
Since the sample number of tablet data integrated test is more, most of sample in the comparison figure of predicted value and actual value
The phenomenon that being overlapped is difficult to be contrasted according to the intensity of mark point to ocular and clear very much by Figure 16, Figure 17, Figure 18
The quality of three kinds of algorithms.Therefore the prediction error of predicted value in being illustrated in table 11 corresponding to Figure 16, Figure 17, Figure 18.
As can be seen from Table 11, for the prediction of three kinds of active component contents of tablet data, TL_ELM-AE models it is pre-
Better performances are surveyed, on the whole, the estimated performance of TL_ELM-AE models is better than PDS algorithms and SBC algorithms.
Prediction error of the 11 each active component content of tablet data set of table under different models
The pre- of the model that three kinds of active constituents are not migrated in corn data set is set forth in Figure 19, Figure 20, Figure 21
The comparison figure of measured value and actual value and comparison between the predicted value and actual value that TL_ELM-AE models are migrated is used
Figure.Wherein, predicted value is closer to actual value, then the point of corresponding standard is just closer to this straight line of x=y.Figure 19, Figure 20,
In Figure 21, when cross star representation is not migrated, the relationship between aiming field test sample actual value and predicted value, triangle table
Show used TL_ELM-AE models migrated after aiming field predicted value and actual value between relationship.
As can be seen that the point of triangle mark is more nearly and collects than the mark point of cross star from Figure 19, Figure 20, Figure 21
In near straight line y=x, after this explanation migrates tablet data set using TL_ELM-AE models, obtain preferably pre-
Survey effect.However, compared with corn data set experiment effect above, the migration effect of tablet data set is not fairly obvious, this
It is primarily due to the principal spectrum of tablet data set and is not too big from SPECTRAL DIVERSITY, this point can be with from the subgraph (C) of Fig. 4
Find out.
The realization algorithm of the TL-ELM-AE models of the present invention
Input:Source domain data set:{xS_cal,yS_calCalibration collection, { xS_test,yS_testTest set;
Aiming field data set:{xT_cal,yT_calCalibration collection, { xT_test,yT_testTest set;
Output:Source domain regression coefficient β, transfer matrix M, the predicted value of aiming field
Algorithm steps:
1st, source domain peg model is established.
(1) collection data are demarcated to source domain and carries out centralization pretreatment:
xS_cal_center=xS_cal-mean(xS_cal)
(2) collection label is demarcated to source domain and carries out centralization pretreatment:
yS_cal_center=yS_cal-mean(yS_cal)
(3) the feature H of source domain is extracted using ELM-AES_calAnd data are to the transfer matrix W of featureS_cal:
(4) to source domain feature HS_calPls modelings are carried out with the label after centralization, obtain regression coefficient β:
The transfer matrix WS_cal, obtain in the following manner:
A. random initializtion weight w and biasing b;
B. according to formula h=g (ws), x+b hidden node h is obtained;
C. according to h and x, regression coefficient η is obtained;
D. transfer matrix WS_calThe as transposition of η.
The transfer matrix is obtained by data training, can be used for carrying out new data transformation acquisition feature.
(1) transfer matrix M of the target domain characterization to source domain feature is calculated using standard set sample.
(1) collection data are demarcated to aiming field and carries out centralization pretreatment:
xT_cal_center=xT_cal-mean(xT_cal)
(2) collection label is demarcated to aiming field and carries out centralization pretreatment:
yT_cal_center=yT_cal-mean(yT_cal)
(3) the feature H of aiming field is extracted using ELM-AET_calData are to the transfer matrix W of featureT_cal:
(4) target domain characterization H is obtained using ELM-AET_calTo source domain feature HS_calTransfer matrix M:
(2) aiming field test set is predicted
(1) pretreatment centralization is carried out to aiming field test set:
xT_test_center=xT_test-mean(xT_cal)
(2) the transfer matrix W of target numeric field data to feature is utilizedT_calObtain the feature H of aiming field test setT_test:
(3) the regression coefficient β of the transfer matrix M of target domain characterization to source domain feature and source domain data to label is utilized:
Wherein, mean () represents mean value,Represent predicted value.
Claims (10)
1. the sample composition content assaying method of feature based migration, which is characterized in that use the calibration based on ELM-AE models
Transfer learning method, by the ir data Feature Mapping to the feature space of source domain of aiming field sample to be tested, Ran Houli
The component content of the aiming field sample to be tested is predicted with the sample composition content prediction model of source domain.
2. the sample composition content assaying method of feature based migration according to claim 1, which is characterized in that specific packet
Include following steps:
S1 carries out feature extraction using ELM-AE models to the ir data of source domain and aiming field;
S2 establishes the polynary calibration migration models between polynary peg model and source domain and aiming field based on source domain feature;
S3 acquires the ir data of aiming field sample to be tested, after carrying out feature extraction conversion, utilizes the polynary calibration of source domain
Model predicts the component content of the aiming field sample to be tested.
3. the sample composition content assaying method of feature based migration according to claim 2, which is characterized in that step S1
Described in feature extraction is carried out to the ir data of source domain and aiming field using ELM-AE models, especially by with lower section
Formula obtains:
HS=XS·WS
HT=XT·WT
Wherein, HSFor the feature of source domain, HTFor the feature of aiming field, XsRepresent the infrared spectrum measured by the key light spectrometer of source domain
Data, XTRepresent the ir data by being measured from spectrometer;WSRepresent the projector space of source domain, WTRepresent the throwing of aiming field
Shadow space, WSAnd WTRespectively by { XS,XSAnd { XT,XTELM-AE models acquire.
4. the sample composition content assaying method of feature based migration according to claim 2, which is characterized in that step S2
Described in the polynary calibration migration models established between source domain and aiming field, that is, be to solve for the following formula:
min||yT-XTWTMβ||2+||yS-XSWSβ||2
s.t.||wS||2=1, | | wT||2=1
Optimization problem, wherein, { XS,ysRepresent source domain data set, { XT,yTRepresent aiming field data set, wherein XsBy leading
Spectrometer measures, XTBy being measured from spectrometer;WSRepresent the projector space of source domain, WTRepresent the projector space of aiming field, WSAnd WT
Respectively by { XS,ysAnd { XT,yTELM-AE models acquire, M represents target domain characterization to the transfer matrix of source domain feature, β tables
Show the regression coefficient of the polynary peg model based on source domain feature.
5. the sample composition content assaying method of the feature based migration according to claim 2 or 4, which is characterized in that step
The polynary peg model based on source domain feature described in rapid S2 utilizes source domain characteristic { H using pls modelsS,ysEstablish
Polynary peg model obtains regression coefficient β, wherein, HSFor the feature of source domain, ysSample composition content for source domain.
6. according to the sample composition content assaying method that any feature based of claim 2~4 migrates, feature exists
In, ELM-AE models, by by the random weight w of the hidden node of ELM modelsiWith biasing biAll orthogonalization, and cause
The input data and output data of ELM models it is equal to get;I.e.:
H=g (wsx+b),
T=x,
wTW=I, bTB=I,
Wherein, w=[w1,...,wL] be orthogonalization random weight matrix, b=[b1,...,bL] be orthogonalization random bias
Matrix.
7. the sample composition content assaying method of feature based migration according to claim 6, which is characterized in that pass through spy
Levy the output weight η that space obtains ELM-AE models with input data training:
If the number of input sample data is more than or equal to the number of hidden node, the output weight η of ELM-AE is:
If the number of input sample data is less than or equal to the number of hidden node, the output weight η of ELM-AE is:
For the ELM-AE of countertype, then exporting weight η is:
η=H-1X
ηTη=I;
Wherein, H=[h1,h2,…hN] be ELM-AE hidden node output, X=[x1,x2,…xN] be ELM-AE input number
According to;C is regularization factors, and N is the number of input sample data.
8. the sample composition content assaying method of feature based migration according to claim 7, which is characterized in that utilize 10
The number for rolling over the corresponding hidden node of method selection RMSECV minimum values of cross validation is the best hidden layer of ELM-AE models
Node number;The regularization factors C is selected using the method for 10 folding cross validations, chooses RMSECV minimum values pair
The regularization factors answered are optimum factor.
9. the sample composition content assaying method of feature based migration according to claim 6, which is characterized in that described
The activation primitive of ELM-AE models uses tanh functions.
10. the sample composition content assaying method of feature based migration according to claim 1, which is characterized in that described
Sample to be tested be plant or Chinese medicine;Preferably corn or tablet.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711334521.0A CN108152239A (en) | 2017-12-13 | 2017-12-13 | The sample composition content assaying method of feature based migration |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711334521.0A CN108152239A (en) | 2017-12-13 | 2017-12-13 | The sample composition content assaying method of feature based migration |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108152239A true CN108152239A (en) | 2018-06-12 |
Family
ID=62467139
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711334521.0A Pending CN108152239A (en) | 2017-12-13 | 2017-12-13 | The sample composition content assaying method of feature based migration |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108152239A (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108960193A (en) * | 2018-07-24 | 2018-12-07 | 中北大学 | A kind of across component infrared spectroscopy model transplantations method based on transfer learning |
CN110728328A (en) * | 2019-10-22 | 2020-01-24 | 支付宝(杭州)信息技术有限公司 | Training method and device for classification model |
CN111220566A (en) * | 2020-01-16 | 2020-06-02 | 东北大学秦皇岛分校 | Infrared spectrum measuring instrument calibration migration method based on OPLS and PDS |
CN111220565A (en) * | 2020-01-16 | 2020-06-02 | 东北大学秦皇岛分校 | CPLS-based infrared spectrum measuring instrument calibration migration method |
CN111415032A (en) * | 2020-03-03 | 2020-07-14 | 东华大学 | Method for predicting production performance of polyester fiber precursor based on E L M-AE of transfer learning |
CN111563436A (en) * | 2020-04-28 | 2020-08-21 | 东北大学秦皇岛分校 | Infrared spectrum measuring instrument calibration migration method based on CT-CDD |
CN113160903A (en) * | 2021-04-23 | 2021-07-23 | 中国科学院西安光学精密机械研究所 | Nitrate concentration prediction model generalization method based on transfer learning |
CN113552091A (en) * | 2021-08-24 | 2021-10-26 | 江南大学 | High-purity 2, 6-xylenol near infrared spectrum on-line detection method |
CN113592020A (en) * | 2021-08-10 | 2021-11-02 | 山东大学 | Spectrum transfer method, system, medium and electronic equipment |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106680238A (en) * | 2017-01-06 | 2017-05-17 | 东北大学秦皇岛分校 | Method for analyzing material composition content on basis of infrared spectroscopy |
-
2017
- 2017-12-13 CN CN201711334521.0A patent/CN108152239A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106680238A (en) * | 2017-01-06 | 2017-05-17 | 东北大学秦皇岛分校 | Method for analyzing material composition content on basis of infrared spectroscopy |
Non-Patent Citations (2)
Title |
---|
MUHAMMAD UZAIR,AJMAL MIAN: "Blind Domain Adaptation With Augmented Extreme Learning Machine Features", 《IEEE TRANSACTIONS ON CYBERNETICS》 * |
WO-RUO CHEN,JUN BIN ET AL.: "Calibration transfer via an extreme learning machine auto-encoder", 《THE ANALYST》 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108960193A (en) * | 2018-07-24 | 2018-12-07 | 中北大学 | A kind of across component infrared spectroscopy model transplantations method based on transfer learning |
CN108960193B (en) * | 2018-07-24 | 2021-09-14 | 中北大学 | Cross-component infrared spectrum model transplanting method based on transfer learning |
CN110728328A (en) * | 2019-10-22 | 2020-01-24 | 支付宝(杭州)信息技术有限公司 | Training method and device for classification model |
CN110728328B (en) * | 2019-10-22 | 2022-03-01 | 支付宝(杭州)信息技术有限公司 | Training method and device for classification model |
CN111220566A (en) * | 2020-01-16 | 2020-06-02 | 东北大学秦皇岛分校 | Infrared spectrum measuring instrument calibration migration method based on OPLS and PDS |
CN111220565A (en) * | 2020-01-16 | 2020-06-02 | 东北大学秦皇岛分校 | CPLS-based infrared spectrum measuring instrument calibration migration method |
CN111220565B (en) * | 2020-01-16 | 2022-07-29 | 东北大学秦皇岛分校 | CPLS-based infrared spectrum measuring instrument calibration migration method |
CN111415032A (en) * | 2020-03-03 | 2020-07-14 | 东华大学 | Method for predicting production performance of polyester fiber precursor based on E L M-AE of transfer learning |
CN111415032B (en) * | 2020-03-03 | 2022-04-29 | 东华大学 | Method for predicting production performance of polyester fiber protofilament based on ELM-AE of transfer learning |
CN111563436B (en) * | 2020-04-28 | 2022-04-08 | 东北大学秦皇岛分校 | Infrared spectrum measuring instrument calibration migration method based on CT-CDD |
CN111563436A (en) * | 2020-04-28 | 2020-08-21 | 东北大学秦皇岛分校 | Infrared spectrum measuring instrument calibration migration method based on CT-CDD |
CN113160903A (en) * | 2021-04-23 | 2021-07-23 | 中国科学院西安光学精密机械研究所 | Nitrate concentration prediction model generalization method based on transfer learning |
CN113592020A (en) * | 2021-08-10 | 2021-11-02 | 山东大学 | Spectrum transfer method, system, medium and electronic equipment |
CN113592020B (en) * | 2021-08-10 | 2023-08-22 | 山东大学 | Spectrum transfer method, system, medium and electronic equipment |
CN113552091A (en) * | 2021-08-24 | 2021-10-26 | 江南大学 | High-purity 2, 6-xylenol near infrared spectrum on-line detection method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108152239A (en) | The sample composition content assaying method of feature based migration | |
Liu et al. | Estimating forest soil organic carbon content using vis-NIR spectroscopy: Implications for large-scale soil carbon spectroscopic assessment | |
Shi et al. | Prediction of soil organic matter using a spatially constrained local partial least squares regression and the C hinese vis–NIR spectral library | |
Jaconi et al. | Near infrared spectroscopy as an easy and precise method to estimate soil texture | |
CN104897607B (en) | Portable near infrared spectrum food modeling and quick detection integral method and system | |
CN106815643B (en) | Infrared spectroscopy Model Transfer method based on random forest transfer learning | |
Guerrero et al. | Assessment of soil organic carbon at local scale with spiked NIR calibrations: effects of selection and extra‐weighting on the spiking subset | |
CN106680238B (en) | Method based on infrared spectrum analysis material component content | |
Sila et al. | Evaluating the utility of mid-infrared spectral subspaces for predicting soil properties | |
Mishra et al. | Partial least square regression versus domain invariant partial least square regression with application to near-infrared spectroscopy of fresh fruit | |
JP7123866B2 (en) | Cross-validation-based calibration of spectral models | |
Panchuk et al. | Calibration transfer between different analytical methods | |
CN106596450A (en) | Incremental method for analysis of material component content based on infrared spectroscopy | |
Mishra et al. | Deep chemometrics: Validation and transfer of a global deep near‐infrared fruit model to use it on a new portable instrument | |
Kothari et al. | Predicting leaf traits across functional groups using reflectance spectroscopy | |
WO2020130947A1 (en) | Method and system for predicting quantitative measures of oil adulteration of an edible oil sample | |
Zhang et al. | Sampling Error Profile Analysis for calibration transfer in multivariate calibration | |
Hemmateenejad et al. | Net analyte signal-based simultaneous determination of antazoline and naphazoline using wavelength region selection by experimental design-neural networks | |
Mas et al. | Screening and quantification of proteinaceous binders in medieval paints based on μ-Fourier transform infrared spectroscopy and multivariate curve resolution alternating least squares | |
CN107918718B (en) | Sample component content determination method based on online sequential extreme learning machine | |
dos Santos et al. | Improving the generalization error and transparency of regression models to estimate soil organic carbon using soil reflectance data | |
Wang et al. | Estimation of soil organic matter by in situ Vis-NIR spectroscopy using an automatically optimized hybrid model of convolutional neural network and long short-term memory network | |
Ferrer et al. | Artificial neural networks (ANNs) in the analysis of polycyclic aromatic hydrocarbons in water samples by synchronous fluorescence | |
Pierna et al. | A methodology to detect outliers/inliers in prediction with PLS | |
CN111220565B (en) | CPLS-based infrared spectrum measuring instrument calibration migration method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180612 |