CN107655850A - Non-linear modeling method and system based near infrared spectrum - Google Patents
Non-linear modeling method and system based near infrared spectrum Download PDFInfo
- Publication number
- CN107655850A CN107655850A CN201610589280.3A CN201610589280A CN107655850A CN 107655850 A CN107655850 A CN 107655850A CN 201610589280 A CN201610589280 A CN 201610589280A CN 107655850 A CN107655850 A CN 107655850A
- Authority
- CN
- China
- Prior art keywords
- sample
- msub
- near infrared
- infrared spectrum
- checking
- 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 53
- 238000002329 infrared spectrum Methods 0.000 title claims abstract description 40
- 238000012937 correction Methods 0.000 claims abstract description 41
- 230000003595 spectral effect Effects 0.000 claims abstract description 31
- 238000013528 artificial neural network Methods 0.000 claims abstract description 28
- 238000002474 experimental method Methods 0.000 claims abstract description 14
- 238000000513 principal component analysis Methods 0.000 claims abstract description 6
- 239000011159 matrix material Substances 0.000 claims description 58
- 238000004364 calculation method Methods 0.000 claims description 7
- 239000012141 concentrate Substances 0.000 claims description 6
- 238000010200 validation analysis Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 abstract description 4
- 230000003044 adaptive effect Effects 0.000 abstract description 2
- 238000001228 spectrum Methods 0.000 description 27
- 238000012360 testing method Methods 0.000 description 27
- 238000012549 training Methods 0.000 description 12
- 239000000126 substance Substances 0.000 description 11
- 241000208125 Nicotiana Species 0.000 description 7
- 235000002637 Nicotiana tabacum Nutrition 0.000 description 7
- 239000000203 mixture Substances 0.000 description 7
- SNICXCGAKADSCV-JTQLQIEISA-N (-)-Nicotine Chemical compound CN1CCC[C@H]1C1=CC=CN=C1 SNICXCGAKADSCV-JTQLQIEISA-N 0.000 description 6
- 230000000052 comparative effect Effects 0.000 description 6
- 230000006870 function Effects 0.000 description 6
- 229960002715 nicotine Drugs 0.000 description 6
- SNICXCGAKADSCV-UHFFFAOYSA-N nicotine Natural products CN1CCCC1C1=CC=CN=C1 SNICXCGAKADSCV-UHFFFAOYSA-N 0.000 description 6
- 239000013598 vector Substances 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 5
- 108090000623 proteins and genes Proteins 0.000 description 5
- 230000005540 biological transmission Effects 0.000 description 4
- 238000012417 linear regression Methods 0.000 description 4
- 241001269238 Data Species 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 230000004044 response Effects 0.000 description 3
- 238000002835 absorbance Methods 0.000 description 2
- 238000010009 beating Methods 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000001035 drying Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000003801 milling Methods 0.000 description 2
- 238000004611 spectroscopical analysis Methods 0.000 description 2
- 102100036962 5'-3' exoribonuclease 1 Human genes 0.000 description 1
- 101000804879 Homo sapiens 5'-3' exoribonuclease 1 Proteins 0.000 description 1
- 238000010521 absorption reaction Methods 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 235000019504 cigarettes Nutrition 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 238000005206 flow analysis Methods 0.000 description 1
- 230000008676 import Effects 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 239000003041 laboratory chemical Substances 0.000 description 1
- 238000002156 mixing Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000010238 partial least squares regression Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000000611 regression analysis Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 208000011580 syndromic disease Diseases 0.000 description 1
- 238000012795 verification Methods 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
- 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
- G01N2201/1296—Using chemometrical methods using neural networks
Abstract
The present invention provides non-linear modeling method and system based near infrared spectrum, gathers the spectrogram of each experiment sample, and is converted near infrared spectrum data;Choose a part at random from each near infrared spectrum data and be used as calibration set, choose a part as checking collection;The calibration set and the checking collection are obtained into Spectral feature scale by principal component analysis;In the Spectral feature scale, sample closest with each sample of the checking collection in the calibration set is chosen by mahalanobis distance method and is used as correction subsets;Number of principal components is extracted from the correction subsets, input layer as BP neural network establishes regression model, can not only solve the problems, such as multiple correlation between each factor, it also avoid substantial amounts of noise and some useless information, reduce dimension, on the basis of the non-linear mapping capability and adaptive learning ability of BP neural network, the prediction stability and precision of model are improved.
Description
Technical field
The present invention relates to Nonlinear Modeling field, more particularly to non-linear modeling method and system based near infrared spectrum.
Background technology
Near infrared spectrum is mainly the absorption of intramolecular frequency multiplication and sum of fundamental frequencies, and spectral intensity is weaker, and bands of a spectrum are wider, overlapping serious, closely
The analysis of infrared spectrum is often required to combine chemometrics method.Conventional chemometrics method has multiple linear regression, master
The homing methods such as components regression, PLS, these regression analyses are all based on spectral response and corresponding chemical score
Between linear relationship come what is established, do not have the ability of gamma correction.
However, it is not there is strict linear relationship, especially for complicated component between spectral response and chemical score
Sample.In regression analysis, the selection of calibration set sample influences the precision of prediction and applicability of whole model, so will during modeling
The spectral signature of modeling collection sample and the spectral signature and property ranges of the whole checking sum of chemical score scope covering are asked, this is just often
Cause modeling collection sample size very big.But substantial amounts of sample not only brings the interference between a large amount of uncorrelated factors, due to not
With sample room can also have larger difference, especially for nonlinear spectral response, have a strong impact on the applicability of model
With the degree of accuracy of prediction.Thus it is clear that it is a kind of simple, conveniently, calibration set sample size it is few, and can covering checking collect all spectral signatures and
The non-linear modeling method of chemical property scope is highly desirable.
The content of the invention
In view of the above the shortcomings that prior art, it is an object of the invention to provide the Nonlinear Modeling based near infrared spectrum
Method and system, for solving problem above present in prior art.
In order to achieve the above objects and other related objects, the present invention provides the non-linear modeling method based near infrared spectrum, bag
Include:Prepare multiple experiment samples, including:The spectrogram of each experiment sample is gathered, and is converted near infrared spectrum data;From
Each near infrared spectrum data chooses a part and is used as calibration set at random, chooses a part as checking collection;By the correction
Collection and the checking collection obtain Spectral feature scale by principal component analysis;In the Spectral feature scale, by geneva away from
Sample closest with each sample of the checking collection in the calibration set is chosen as correction subsets from method;From the correction
Number of principal components is extracted in subset, the input layer as BP neural network establishes regression model.
In one embodiment of the invention, the one or more during methods described is further comprising the steps of:Step A, with the recurrence
The sample that the model checking checking is concentrated;Step B, the sample of concentration is verified described in global modeling method validation, and will be tested
Card result compares with the obtained the results of the step A.
In one embodiment of the invention, the calculation formula of the mahalanobis distance method is:
Wherein, MDiI-th of sample is concentrated to the mahalanobis distance of calibration set, S for checkingiFor checking concentrate i-th sample it is main into
Get sub-matrix, SjFor the principal component scores matrix of j-th of sample in calibration set, V is the covariance of principal component scores matrix.
It is described to extract number of principal components from the correction subsets in one embodiment of the invention, the input as BP neural network
Layer establishes regression model, including:With PLS to the correction subsets by Cross gain modulation analysis determine it is main into
Fraction, nonlinear model is established using the number of principal components as the input layer of BP neural network.
In one embodiment of the invention, after the spectrogram for acquiring each experiment sample, in addition to:By each spectrum
Figure independently saves as the file of each .dx forms.
In order to achieve the above objects and other related objects, the present invention provides the Nonlinear Modeling system based near infrared spectrum, bag
Include:First module to the 5th module.First module is used for the spectrogram for gathering each experiment sample, and is converted near infrared light
Modal data;Second module is used as calibration set for choosing a part at random from each near infrared spectrum data, chooses a part
Collect as checking;3rd module is used to the calibration set and the checking collection obtaining spectral signature sky by principal component analysis
Between;4th module is used in the Spectral feature scale, by mahalanobis distance method choose in the calibration set with the checking
The closest sample of each sample of collection is as correction subsets;5th module is used to extract principal component from the correction subsets
Number, the input layer as BP neural network establish regression model.
In one embodiment of the invention, the system also includes:6th module, or, the combination of the 6th module and the 7th module.
6th module is used for the sample concentrated according to the regression model checking checking;7th module is used for according to global modeling side
The sample that the method checking checking is concentrated, and the result that the result is obtained with the regression model compares.
In one embodiment of the invention, the calculation formula of the mahalanobis distance method is:
Wherein, MDiI-th of sample is concentrated to the mahalanobis distance of calibration set, S for checkingiFor checking concentrate i-th sample it is main into
Get sub-matrix, SjFor the principal component scores matrix of j-th of sample in calibration set, V is the covariance of principal component scores matrix.
It is described to extract number of principal components from the correction subsets in one embodiment of the invention, the input as BP neural network
Layer establishes regression model, including:With PLS to the correction subsets by Cross gain modulation analysis determine it is main into
Fraction, nonlinear model is established using the number of principal components as the input layer of BP neural network.
In one embodiment of the invention, first module is additionally operable to:After the spectrogram for acquiring each experiment sample,
Each spectrogram is independently saved as to the file of each .dx forms.
As described above, the non-linear modeling method and system based near infrared spectrum of the present invention, it is proposed that on principal component point
Analysis, mahalanobis distance method, least square method, the built-up pattern of BP neural network method.Pass through PCA and mahalanobis distance method
Combination, choose and collect most like calibration set sample as correction subsets with checking, with PLS to correction subsets
Analyzed by Cross gain modulation and determine number of principal components, nonlinear model is established using number of principal components as BP neural network input layer.
Can not only solve the problems, such as multiple correlation between each factor using this method, it is thus also avoided that substantial amounts of noise and some useless letters
Breath, reduces dimension, on the basis of the non-linear mapping capability and adaptive learning ability of BP neural network, improves mould
The prediction stability and precision of type.
Brief description of the drawings
Fig. 1 is shown as the non-linear modeling method flow chart based near infrared spectrum of one embodiment of the invention.
Fig. 2 is shown as whole sample light spectrograms of one embodiment of the invention.
Fig. 3 is shown as the distribution map of the original calibration set and checking collection of one embodiment of the invention in principal component.
Fig. 4 is shown as the distribution map of the calibration set after the selection of one embodiment of the invention and checking collection in principal component.
Fig. 5 is shown as the main cause subnumber distribution map of the Nonlinear Modeling offset minimum binary of one embodiment of the invention.
Fig. 6 is shown as the main cause subnumber distribution map of the global Nonlinear Modeling offset minimum binary of one embodiment of the invention.
Fig. 7 is shown as the Nonlinear Modeling test nicotine value content prediction comparative result figure of one embodiment of the invention.
Fig. 8 is shown as the global modeling test nicotine value content prediction comparative result figure of one embodiment of the invention.
Fig. 9 is shown as the Nonlinear Modeling system module figure based near infrared spectrum of one embodiment of the invention.
Embodiment
Illustrate embodiments of the present invention below by way of specific instantiation, those skilled in the art can be taken off by this specification
The content of dew understands other advantages and effect of the present invention easily.The present invention can also pass through specific embodiment parties different in addition
Formula is embodied or practiced, and the various details in this specification can also be based on different viewpoints and application, without departing from this hair
Various modifications or alterations are carried out under bright spirit.It should be noted that in the case where not conflicting, in following examples and embodiment
Feature can be mutually combined.
It should be noted that the diagram provided in following examples only illustrates the basic conception of the present invention in a schematic way, then
Component count, shape and size drafting when the component relevant with the present invention is only shown in schema rather than being implemented according to reality,
Kenel, quantity and the ratio of each component can be a kind of random change during its actual implementation, and its assembly layout kenel may also
It is increasingly complex.
Referring to Fig. 1, the present invention provides the non-linear modeling method based near infrared spectrum, comprise the following steps:
Step S101:Experiment sample is prepared, for example, the tobacco leaf that Honghe, Yunnan Redrying Factory passes through after beating and double roasting is taken, tobacco leaf warp
After the On-line NIR instrument for crossing Zeiss, Germany is scanned, spectrometer every five seconds for example run-down, while crawl one is scanned through
Tobacco leaf, about 30g, 10 formation, one aggregate sample are captured altogether, the corresponding scanning of spectrometer 10 times, spectrometer calculates automatically
The averaged spectrum that scanning is 10 times, this averaged spectrum is exactly the spectrum of corresponding aggregate sample, takes 509 samples altogether, and stick
Corresponding 1/2/3 ... labels, their basic data will be measured with flow analysis instrument after each sample drying milling.It is optional
, each spectrogram can also independently be saved as to the file of .dx forms, data processing is carried out for importing MATLAB.
Step S102:Choose a part at random from each near infrared spectrum data and be used as calibration set, choose a part as checking collection.
For example, selecting 409 calibration sets as model from 509 sample near infrared spectrum datas by the use of K-S methods, 100 are selected
Collect as checking.
Step S103:Calibration set and checking collection are subjected to principal component projection to choose principal component as Spectral feature scale, preferably
, the first two principal component is chosen as Spectral feature scale.The calculation formula of the principal component is:
Wherein, n is number of samples, and p counts for wavelength, xnpIn wavelength points it is spectroscopic data value at p for n-th of sample, then
Establish the correlation matrix R of variableij=(rij)p×p, wherein,Obtain Rij
Characteristic root:T1≥T2≥…≥Tp> 0, and corresponding characteristic vector:Push away
Export principal component Fi=a1iX1+a2iX2+…+apiXp(i=1 ..., p), further calculate the contribution rate of accumulative total of each principal componentContribution rate of accumulative total reaches preset value (such as:85%~95%) characteristic value T1,T2,…,Tm
Corresponding 1st, the 2nd ..., m (m≤p) individual principal component, m principal component contains that to all refer to target specific substantially before explanation
Information, m characteristic value before taking, and corresponding characteristic vector is calculated, form Spectral feature scale.
Step S104:In the Spectral feature scale, choose what is collected in the calibration set with the checking using mahalanobis distance method
The most like sample of each sample is as correction subsets, it is preferred that chooses each sample collected in the calibration set with the checking most
Two similar samples are as correction subsets.The calculation formula of the mahalanobis distance method is:
Wherein, MDiI-th of sample is concentrated to the mahalanobis distance of calibration set, S for checkingiFor checking concentrate i-th sample it is main into
Get sub-matrix, SjFor the principal component scores matrix of j-th of sample in calibration set, V is the covariance of principal component scores matrix.Need
It is noted that the threshold value setting procedure of mahalanobis distance includes:First, the checking collection is calculated in the Spectral feature scale
In each spectrum samples each spectrum samples into the calibration set mahalanobis distance;Secondly, these mahalanobis distances are pressed
Descending arranges, and chooses correction subsets of the spectrum samples of above two beelines as checking collection spectrum;Finally, described in calculating
The correction subsets of each spectrum are concentrated in checking, so as to establish new correction subsets.New correction subsets selection is all with testing
Card concentrates the most similar sample of each sample distance, will not only cause to bring sample that is a large amount of uncorrelated and differing greatly into, and
And the spectral signature and underlying data range of sample, the sample of calibration set when greatly reducing modeling are concentrated in and can covering checking
This number.
Step S105:Number of principal components is extracted from above-mentioned new correction subsets, the input layer as BP neural network, which is established, to be returned
Model, it is preferred that by the correction subsets by PLS dimensionality reduction, be mutually authenticated by intersecting to determine number of principal components
Correlation model is established according to the input layer as BP neural network.
When with multiple linear regression, serious multiple correlation is there may be between independent variable, causes multiple linear regression
It can fail, so as to the estimation of damage parameter, expand the error of near-infrared model so that model scatters and disappears stability, can also go out sometimes
The result now disagreed with real-life general knowledge.However, PLS can but solve this well in near-infrared
One problem, it can well play a role when independent variable is more than and multicollinearity be present between sample number and independent variable, has
Beneficial to the noise information and signal message distinguished near infrared spectrum.The central idea of PLS is dimensionality reduction, respectively from
Extract component R, U successively in spectroscopic data and chemical composition data, order:
In formula, X is the spectrum data matrix of correcting sample composition, is to be detected to obtain by karr Zeiss near infrared device;Y is correction
The concentration matrix of basic data composition corresponding to collection sample, it is to be measured to obtain by Flow Analyzer, R and U are respectively obtaining for X and Y
Sub-matrix, P and Q are respectively X and Y loading matrix, EXAnd EYRespectively X and Y partial least-squares regression method (referred to as PLS)
Regression criterion matrix.Wherein:rk(n × 1) be X k-th of main gene score matrix, pk(1 × m) is k-th of X matrix
The loading matrix of main gene, uk(n × 1) is the score matrix of Y k-th of main gene;qk(1 × m) is k-th of main cause of Y matrixes
The loading matrix of son, n are the line number of score matrix, and m is the columns of loading matrix, and f is main gene number.Then, according to partially minimum
The principle of square law, R and U are done into linear regression:U=RB, B=(RTR)-1RTY, in prediction, obtained according to P treat test sample first
The score R of product score matrixPrediction, then as corresponding to following formula obtains this sample concentration matrix predicted value YPrediction=UQT=RPrediction
BQ。
It is initially to ignore residual matrix E it should be noted that in model process is builtXAnd EY, after obtaining middle parameter,
Return again to and seek residual matrix.Detailed process is as follows, starts to have when taking main cause subnumber f=1:
To X=RPT, the left side multiplies RT, then the right side multiply P and obtain:To Y=UQT, the left side multiplies UTThen both sides with divided by QT:Initial iteration values of the concentration matrix Y as U is taken, R is replaced with U, according to equation:X=UWTW is calculated, its solution is:W is X weight vectors;X score matrix R is sought after being normalized to weight, equation is:X=RWT, its solution is:U calculating Y loading matrix Q is replaced with R, its equation is:Y=RQT, its solution is:To load moment
Y score matrix U is sought after battle array Q normalization, equation is:Y=UQT, its solution is:Replace R to return with this U again most to open
Begin to calculate WT, by WTCalculate R1, so iterate, if R has restrained, even | | R-R1| | < 10-6| | R | |, stop iteration;It is no
Then, return and continue the weight vectors W for seeking X until the X obtained score matrix convergence;X loading matrix is sought according to the R after convergence
P, its equation are:X=RPT, its solution is:X score matrix R=R is sought after being normalized to loading matrix P | | P | |;
Standardized weight vector W=W | | P | |;Calculate the internal relation between R and UResidual matrix E is calculated againX=X-
RPT、EY=Y-UQT=Y-BRQT;Finally, with EXInstead of X, EYInstead of Y, the weight vectors W that the step of most starting seeks X is returned.
By that analogy, X, Y main gene W, R, P, U, Q, B are obtained, passes through cross-verification method(preferred value) determines optimal
Stop iteration during main cause subnumber f.
It should be noted that BP neural network is Backward error propagation artificial neural network, it is possible to achieve between input and output
Arbitrary nonlinear mapping, there is very strong Nonlinear Mapping approximation capability and predictive ability.PLS is chosen
The data of number of principal components import the input layer of BP neural network, by standardization, give weight and are input to hidden layer, imply
Layer passes to output layer, output layer provides the predicted value of sample, the phase with sample after weight, threshold value and excitation function computing
Prestige value is compared, and if there is error, then reversely passes the error back since output layer, constantly carries out weights and threshold value
Adjustment so that the result of prediction tends to consistent with desired output.
For example, BP god can realize network algorithm by following steps:
1) the number of principal components evidence of the sample of selection is input to input layer.
2) random number gives the initial weight in the range of (0,1), by the output of hidden layerThe output of output layerError(n is sample number) calculates the transmission of positive information, wherein:M is the nodes of input layer, and j=1,2 ..., h, h is the nodes of hidden layer,
wijFor the connection weight between input layer i and hidden layer node j;p
For the nodes of output layer, vjkFor the connection weight between hidden layer node j and output node layer k;bjAt hidden layer node j
Threshold value, bkIt for the threshold value at output node layer k, can move left and right transmission function below;oijAt hidden layer node j
Output valve;yijFor its corresponding desired output.
3) the error parameter u of output layerk=(yk-ok)f'(netk), wherein:
If transmission function uses logarithm Sigmoid functions, f'(netk)=f (netk)[1-f(netk)], then uk=(yk-ok)ok
(1-ok), k=1,2 ..., p, p be output layer nodes, ykFor the output of output layer, okFor corresponding desired output.
4) the error parameter u of hidden layerj=(∑kukvkj)f'(netj), wherein:
If transmission function uses logarithm Sigmoid functions, f'(netj)=f (netj)[1-f(netj)], then uj=(∑kukvkj)
gj(1-gj), j=1,2 ..., h, h be hidden layer nodes, k=1,2 ..., p, p be output layer nodes, vkjFor output
Connection weight at node layer k and hidden layer node j.
5) the error parameter u of output layer and hidden layer is calculated:By the connection weight between hidden layer node j and output node layer k
vjk(L+1)=vjk(L)+tukgjConnection weight w between input layer i and hidden layer node jij(L+1)=wij(L)+
tujxiThe adjustment of weight is constantly carried out, wherein:T is learning rate, i.e. step-length, determines the speed of training (iteration);L+1 is instruction
Iterations in white silk.
6) repeat the above steps and calculate next training sample.
For training sample used, then stop iteration when error reaches previously given value, all samples in training set are carried out
The training of weight is referred to as an iteration, typically to pass through the iteration of up to a hundred times (such as 100-5000 times) and can just make error
Reach minimum, and iterate to calculate preferably randomly select training sample each time.In order to accelerate iterative process and prevent iteration
The vibration of process, the learning algorithm for introducing factor of momentum can be used, the next item up " momentum " item is added in weight correction value, i.e.,:△w
(L+1)=tuo+ a △ w (L), wherein, uoThe error parameter of original state output layer and hidden layer is represented, a △ w (L) are momentum
, factor of momentum a initial values are typically set at 0.9.
Illustrate the superiority of this method below with reference to specific experiment and experimental result:
1) On-line NIR and laboratory chemical value of sample are gathered, tobacco leaf passes through after beating and double roasting, by online
Near infrared spectrometer scans, every five seconds for example run-down, scans 10 times seek its averaged spectrum altogether, while according to the side of uniform sampling
Formula, every five seconds for example capture once, grab 10 samples for forming a mixing altogether, and averaged spectrum is exactly the spectrum of this aggregate sample, and
Enter line label to sample, altogether 509 piece cigarette samples.As shown in Figure 2.Tobacco leaf sample is by steps such as drying, millings, with flowing
Analyzer measures its corresponding basic data, due to using MATLAB to carry out data simulation in this example, so by each light
Spectrogram independently saves as the file of .dx forms.Certainly, saving as what type of file can be according to the work of actual emulation
Tool is selected.
2) 409 calibration sets as model are selected from 509 sample near infrared spectrum datas by the use of K-S methods, 100 are to test
Card collection.Realized by following MATLAB program in machine codes, obtain Fig. 3:
clear all;closeall;Clc (empties all variables).
Pname=[' E:The line spectrum of work patents in March Red River two in 2016 '];(file that spectrum preserves is read, is read
The data taken are a structures)
A=dir (fullfile (pname, ' * .dx'));(by .dx type files list in file, reading is a knot
Structure array)
A=struct2cell (A);(structural array is converted into cell array)
B=A (1,:);(reading all spectrum filenames)
X=[];(empty matrix is set)
For i=1:Length (B) (reads all spectrum numbers)
D=strcat (pname, ' ', B (i));(reading each spectral information)
[W, H]=textread (char (D), ' %f%f', 256, ' headerlines', 18);(read the ripple of each spectrum
Long and absorbance)
X=[X, H];(all spectral absorbances are formed into spectrum matrix)
End (end loop)
X=X';(conversion of spectrum matrix column is embarked on journey)
Y=xlsread (' E:Work San Yue Zhuan Li Red River two wires value .xlsx' in 2016, ' B1:B509');(read phase
Corresponding chemical score)
Ind1=randperm (609,429);(matrix for randomly selecting 509 rows 429 row)
Xcal=X (ind1,:);(choosing 429 spectral compositions, one spectrum matrix, that is, calibration set in X-ray composes battle array)
Ycal=Y (ind1);(chemical score corresponding with Xcal calibration sets is chosen in Y chemical scores)
X(ind1,:)=[];(X-ray is composed into 429 spectrum in battle array to empty, that is, forms the spectrum square of 429 spectrum of a rejecting
Battle array)
Y(ind1,:)=[];(chemical score corresponding with X rejectings concentration is chosen in Y chemical scores)
Xtest=T (ind1,:);(choosing 100 spectral compositions, one spectrum matrix, that is, checking collection in T spectrum battle array)
Ytest=Y1 (ind1,:);(100 concentration datas corresponding with checking collection are chosen in Y1 chemical concentrations data)
XX=[Xcal;Xtest];(calibration set and checking collection are merged into a big matrix XX)
[coeff, score, latent, tsquare]=princomp (XX);(principal component projection is carried out to XX)
P1=score (:,1);P2=score (:,2);(extraction XX the first two number of principal components)
figure
plot(p1(1:429),p2(1:429),'ko',p1(430:end),p2(430:), end ' k*') (the first two it is main into
Divide in space, calibration set and the two-dimensional space distribution for verifying collection, respectively with the * marks for enclosing circle, black of black)
Legend (' calibration set ', ' checking collection ') (title in figure)
Xlabel (' PC1') (title of X-axis, PC1 are first principal component number)
Ylabel (' PC2') (title of Y-axis, PC2 are Second principal component, number)
Title (' original calibration set and checking collection distribution map ' in principal component) (title)
3) calibration set and checking collection are subjected to principal component projection, it is Spectral feature scale to choose the first two principal component, in Spectral Properties
Space is levied, calibration set is chosen by the use of mahalanobis distance and verifies that collecting two most like calibration samples of sample is used as syndrome with each
Collection.Realized by following MATLAB program in machine codes, obtain Fig. 4:
Xcall=[];Ycall=[];(setting empty matrix)
For i=1:Size (Xtest, 1) (circulation, from 1 to 429)
Xcal=[Xcal;Xtest(i,:)];
[coeff, score, latent, tsquare]=princomp (xcal);(checking collection carries out principal component projection)
T=score (:,1:10);(choosing 10 principal components)
Ms=[];(empty matrix is set)
For i=1:size(Xcal,1)
Ma=sqrt ((T (j,:)-T(size(Xcal,1)+1))*inv(cov(T))*(T(j,:)-T(size(Xcal,1)+
1))');(calculating each checking sample to the distance of calibration set)
Ms=[ms, ma];(the geneva matrix of all checking collection to calibration set samples is formed into a row matrix)
end
[MA, index]=sort (ms);(mahalanobis distance is arranged from big to small)
Ind=index (1:2);(choosing two above closest data numbers)
X1=Xcal (ind,:);(choose and collect two most like composition correction subsets in calibration set with each checking)
Y1=Ycal (ind,:);(choosing corresponding concentration data)
Xcall=[Xcall;X1];(all correction subsets are formed into a matrix)
Ycall=[Ycall;Y1];(corresponding concentration data is formed into a column matrix)
Xcal(ind,:)=[];
Ycal(ind,:)=[];
End
DD=[ZJ;N];(by one matrix of newly-established correction subsets and checking collection composition)
[coeff, score, latent, tsquare]=princomp (DD);(by newly-established correction subsets and checking collect into
Row principal component projection)
P1=score (:,1);P2=score (:,2);(choosing the first two principal component)
plot(p1(1:100),p2(1:100),'ro',p1(101:end),p2(101:End), ' b*') (draw in two masters
The two-dimensional space distribution map of component space calibration set and checking collection)
Legend (' calibration set ', ' checking collection ') (title in figure)
Xlabel (' PC1') (X-axis title, PC1 are first principal component number)
Ylabel (' PC2') (Y-axis title, PC2 are Second principal component, number)
Title (' choose after distribution map in principal component of calibration set and checking collection ') (title of picture)
4) by by the correction subsets that mahalanobis distance chooses by PLS dimensionality reduction, it is mutually authenticated really with intersection
Fixed optimal number of principal components.Realized by following MATLAB program in machine codes, obtain Fig. 5-6:
[xl, yl, xs, ys, beta, pctvar, mse]=plsregress (Xcall, Ycall, 20, ' CV', 20);(to extraction
Calibration set subset and corresponding concentration data carry out offset minimum binary, MSE is mean square error, refers to estimates of parameters and parameter
The desired value of the difference square of true value, when MSE reaches minimum, illustrate that true value is closest with estimate, select the master of this when
Factor number is best suitable for)
figure
plot(0:20,mse(2,:),'ko');(picture)
hold on
plot(0:20,mse(2,:),'k');
XPC=xs (:,1:15);(choosing 1 to 15 principal component numbers)
[x2, y2, xs1, ys1, beta1, pctvar1, mse1]=plsregress (Xtest, Ytest, 20, ' CV', 20);
(PLS is carried out to checking collection and corresponding concentration data)
XPC1=xs1 (:,1:15);(choosing 1 to 15 principal component numbers)
Xlabel (' main cause subnumber ')
ylabel('MSE')
The title main cause subnumber of Nonlinear Modeling offset minimum binary (' ')
%% is to original modeling offset minimum binary
[x3, y3, xs3, ys3, beta3, pctvar3, mse3]=plsregress (Xcal, Ycal, 20, ' CV', 20);It is (right
Calibration set and corresponding concentration data originally carries out offset minimum binary)
figure
plot(0:20,mse3(2,:),'ko');
hold on
plot(0:20,mse3(2,:),'k');
XPC2=xs3 (:,1:12);
[x4, y4, xs4, ys4, beta4, pctvar4, mse4]=plsregress (Xtest, Ytest, 20, ' CV', 20);
XPC4=xs4 (:,1:12);
Xlabel (' main cause subnumber ')
ylabel('MSE')
Title (the main cause subnumber of Nonlinear Modeling offset minimum binary ' global ')
5) number of principal components is established into correlation model according to the input layer as BP neural network.It is real by following MATLAB program in machine codes
It is existing:
P_train=XPC';
T_train=Ycall';
P_test=XPC1';
T_test=Ytest';
N=size (P_test, 2);
Net=newff (P_train, T_train, 8);(BP neural network establishment, training and emulation testing)
% sets training parameter
Net.trainParam.epochs=1000;
Net.trainParam.goal=1e-3;
Net.trainParam.lr=0.01;
% training networks
Net=train (net, P_train, T_train);
6) with the model authentication collection built up, realized by following MATLAB program in machine codes, obtain Fig. 7:
% emulation testings
T_sim_bp=sim (net, P_test);
%% performance evaluations
% coefficient correlations
Mx=mean (Ytest);
Rmx=repmat (mx, 100,1);
B1=sum ((Ytest-rmx) .^2);
B2=sum ((Ytest-T_sim_bp') .^2);
R=sqrt (1-b2/b1);
% prediction standard deviations
SEP=sqrt (b2/ (N-1));
% relative errors error
Error_bp=abs (T_sim_bp-T_test) ./T_test;
Error_bpp1=mean (error_bp);
R2_bp=r;
% Comparative results
Result_bp=[T_test'T_sim_bp'error_bp'];
%xlsxwrite (' E:Work San Yue Zhuan Li abc.xlsx', result_bp in 2016)
Figure (drawing)
plot(1:N,T_test,'b:*',1:N,T_sim_bp,'r-o')
Legend (' actual value ', ' BP predicted values ')
Xlabel (' forecast sample ')
Ylabel (' tobacco leaf nicotine value ')
String=' test nicotine value content prediction Comparative result ';[' R='num2str (R2_bp) ' (BP) '];
title(string)
7) contrasted with global modeling and with the model built up, realized by following MATLAB program in machine codes, obtain Fig. 8:%%
Global modeling BP neural network
P1_train=XPC2';
T1_train=Ycal';
P1_test=XPC4';
T1_test=Ytest';
N=size (P1_test, 2);
Net=newff (P1_train, T1_train, 8);(BP neural network establishment, training and emulation testing)
% sets training parameter
Net.trainParam.epochs=1000;
Net.trainParam.goal=1e-3;
Net.trainParam.lr=0.01;
Net=train (net, P1_train, T1_train);(training network)
% emulation testings
T1_sim_bp=sim (net, P1_test);
%% performance evaluations
Mx1=mean (Ytest);
Rmx1=repmat (mx1,100,1);
B11=sum ((Ytest-rmx1) .^2);
B21=sum ((Ytest-T1_sim_bp') .^2);
R1=sqrt (1-b21/b11);(coefficient correlation)
SEP1=sqrt (b21/ (N-1));(prediction standard deviation)
% relative errors error
Error_bp1=abs (T1_sim_bp-T1_test) ./T1_test;
Error_bpp1=mean (error_bp1);
R2_bp1=r1;
% Comparative results
Result_bp1=[T1_test'T1_sim_bp'error_bp1'];
figure
plot(1:N,T1_test,'b:*',1:N, T1_sim_bp, ' r-o') (picture)
Legend (' actual value ', ' BP predicted values ')
Xlabel (' forecast sample ') (X-axis title)
Ylabel (' tobacco leaf nicotine value ') (Y-axis title)
String=' global modeling test nicotine value content prediction Comparative result ';[' R='num2str (R2_bp1) '
(BP)']};
Title (string) (graphical banner)
xlswrite('E:Work San Yue Zhuan Lis in 2016 it is global ', result_bp1);(preserve result data to arrive
File)
The result data table collected below for Nonlinear Modeling and global modeling to checking:
The Nonlinear Modeling of table 1 checking collection result
The global modeling of table 2 checking collection result
The actual value of table 3 and two kinds of model predication value comparison sheets
Title | True value | World model's predicted value | Non-linear mould predictive value |
Maximum | 3.94 | 4.63 | 3.70 |
Minimum value | 1.32 | 0.96 | 1.46 |
Average value | 2.80 | 2.84 | 2.75 |
From table 3 it is observed that global its predicted value bound of model founded a capital verifies the scope of collection model beyond us, and
The non-linear modeling method designed with us, its bound is all within the scope of our institute's established models.
The actual value of table 4 and two kinds of model predication value relative error tables
Title | World model's predicted value | Non-linear mould predictive value |
Relative error | 0.14 | 0.10 |
Prediction standard deviation | 0.45 | 0.35 |
Coefficient correlation | 0.73 | 0.84 |
As can be seen from Table 4 nonlinear model either prediction deviation should coefficient correlation will be than the model that build of the overall situation.
Referring to Fig. 9, with above method embodiment principle similarly, the present invention also provides the Nonlinear Modeling of near infrared spectrum
System 9.Because the technical characteristic in above method embodiment can be used for the system embodiment, thus it is no longer repeated.
System 9 includes the module 905 of the first module 901 to the 5th.First module 901 is used to prepare multiple experiment samples, including:Adopt
Collect the spectrogram of each experiment sample, and be converted near infrared spectrum data, optionally, also independently protect each spectrogram
Save as the file of each .dx forms.Second module 902 chooses a part as correction at random from each near infrared spectrum data
Collection, choose a part as checking collection.3rd module 903 obtains the calibration set and the checking collection by principal component analysis
Spectral feature scale.4th module 904 in the Spectral feature scale, by mahalanobis distance method choose in the calibration set with
The sample for verifying that each sample collected is closest is as correction subsets, for example, the calculation formula of the mahalanobis distance method is:Wherein, MDiI-th of sample is concentrated to the mahalanobis distance of calibration set, S for checkingiFor
The principal component scores matrix of i-th of sample, S are concentrated in checkingjFor the principal component scores matrix of j-th of sample in calibration set, based on V
The covariance of component score matrix.5th module 905 extracts number of principal components from the correction subsets, as BP neural network
Input layer establishes regression model, it is preferred that the correction subsets is analyzed by Cross gain modulation with PLS true
Determine number of principal components, nonlinear model is established using the number of principal components as the input layer of BP neural network.
In another embodiment, the system 9 can also include the 6th module, for being tested according to regression model checking
Demonstrate,prove the sample concentrated.On this basis, the system 9 can also include the 7th module, for being tested according to global modeling method
The sample that the checking is concentrated is demonstrate,proved, and by the result that the result and the regression model obtain than equity.
In summary, non-linear modeling method and system of the invention based near infrared spectrum, effectively overcome prior art
In various shortcoming and have high industrial utilization.
The above-described embodiments merely illustrate the principles and effects of the present invention, not for the limitation present invention.It is any to be familiar with this
The personage of technology all can carry out modifications and changes under the spirit and scope without prejudice to the present invention to above-described embodiment.Therefore, lift
What all those of ordinary skill in the art were completed under without departing from disclosed spirit and technological thought
All equivalent modifications or change, it should be covered by the claim of the present invention.
Claims (10)
- A kind of 1. non-linear modeling method based near infrared spectrum, it is characterised in that including:The spectrogram of multiple experiment samples is gathered, and is converted near infrared spectrum data;Choose a part at random from each near infrared spectrum data and be used as calibration set, choose a part as checking collection;The calibration set and the checking collection are obtained into Spectral feature scale by principal component analysis;In the Spectral feature scale, each sample in the calibration set with the checking collection is chosen by mahalanobis distance method Closest sample is as correction subsets;Number of principal components is extracted from the correction subsets, the input layer as BP neural network establishes regression model.
- 2. the non-linear modeling method according to claim 1 based near infrared spectrum, it is characterised in that also include:Step Rapid A, or, step A and step B combination:Step A, the sample concentrated with the regression model checking checking;Step B, the sample of concentration is verified described in global modeling method validation, and the result and the step A are obtained The result compares.
- 3. the non-linear modeling method according to claim 1 based near infrared spectrum, it is characterised in that the geneva away from Calculation formula from method is:<mrow> <msub> <mi>MD</mi> <mi>i</mi> </msub> <mo>=</mo> <msqrt> <mrow> <mo>(</mo> <msub> <mi>S</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>S</mi> <mi>j</mi> </msub> <mo>)</mo> <mo>&times;</mo> <msup> <mi>V</mi> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </msup> <msup> <mrow> <mo>(</mo> <msub> <mi>S</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>S</mi> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mi>T</mi> </msup> </mrow> </msqrt> <mo>,</mo> </mrow>Wherein, MDiI-th of sample is concentrated to the mahalanobis distance of calibration set, S for checkingiFor checking concentrate i-th sample it is main into Get sub-matrix, SjFor the principal component scores matrix of j-th of sample in calibration set, V is the covariance of principal component scores matrix.
- 4. the non-linear modeling method according to claim 1 based near infrared spectrum, it is characterised in that described from described Number of principal components is extracted in correction subsets, the input layer as BP neural network establishes regression model, including:With offset minimum binary Method is analyzed the correction subsets by Cross gain modulation and determines number of principal components, using the number of principal components as BP neural network Input layer establishes nonlinear model.
- 5. the non-linear modeling method according to claim 1 based near infrared spectrum, it is characterised in that in the collection After the spectrogram of each experiment sample, in addition to:Each spectrogram is independently saved as to the file of each .dx forms.
- A kind of 6. Nonlinear Modeling system based near infrared spectrum, it is characterised in that including:First module, for gathering the spectrogram of each experiment sample, and it is converted near infrared spectrum data;Second module, calibration set is used as choosing a part at random from each near infrared spectrum data, chooses a part of work Collect for checking;3rd module, for the calibration set and the checking collection to be obtained into Spectral feature scale by principal component analysis;4th module, in the Spectral feature scale, being chosen by mahalanobis distance method and being tested in the calibration set with described The closest sample of each sample collected is demonstrate,proved as correction subsets;5th module, for extracting number of principal components from the correction subsets, the input layer as BP neural network, which is established, to be returned Model.
- 7. the Nonlinear Modeling system according to claim 6 based near infrared spectrum, it is characterised in that also include:The Six modules, or, the combination of the 6th module and the 7th module:6th module, for the sample concentrated according to the regression model checking checking;7th module, for verifying the sample of concentration according to global modeling method validation, and by the result and described time The result for returning model to obtain compares.
- 8. the Nonlinear Modeling system according to claim 6 based near infrared spectrum, it is characterised in that the geneva away from Calculation formula from method is:<mrow> <msub> <mi>MD</mi> <mi>i</mi> </msub> <mo>=</mo> <msqrt> <mrow> <mo>(</mo> <msub> <mi>S</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>S</mi> <mi>j</mi> </msub> <mo>)</mo> <mo>&times;</mo> <msup> <mi>V</mi> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </msup> <msup> <mrow> <mo>(</mo> <msub> <mi>S</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>S</mi> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mi>T</mi> </msup> </mrow> </msqrt> <mo>,</mo> </mrow>Wherein, MDiI-th of sample is concentrated to the mahalanobis distance of calibration set, S for checkingiFor checking concentrate i-th sample it is main into Get sub-matrix, SjFor the principal component scores matrix of j-th of sample in calibration set, V is the covariance of principal component scores matrix.
- 9. the Nonlinear Modeling system according to claim 6 based near infrared spectrum, it is characterised in that described from described Number of principal components is extracted in correction subsets, the input layer as BP neural network establishes regression model, including:With offset minimum binary Method is analyzed the correction subsets by Cross gain modulation and determines number of principal components, using the number of principal components as BP neural network Input layer establishes nonlinear model.
- 10. the Nonlinear Modeling system according to claim 6 based near infrared spectrum, it is characterised in that described first Module is additionally operable to:After the spectrogram for acquiring each experiment sample, each spectrogram is independently saved as into each .dx lattice The file of formula.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610589280.3A CN107655850A (en) | 2016-07-25 | 2016-07-25 | Non-linear modeling method and system based near infrared spectrum |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610589280.3A CN107655850A (en) | 2016-07-25 | 2016-07-25 | Non-linear modeling method and system based near infrared spectrum |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107655850A true CN107655850A (en) | 2018-02-02 |
Family
ID=61126508
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610589280.3A Pending CN107655850A (en) | 2016-07-25 | 2016-07-25 | Non-linear modeling method and system based near infrared spectrum |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107655850A (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108710920A (en) * | 2018-06-05 | 2018-10-26 | 北京中油瑞飞信息技术有限责任公司 | Indicator card recognition methods and device |
CN108827907A (en) * | 2018-04-26 | 2018-11-16 | 新疆维吾尔自治区分析测试研究院 | It is a kind of based near infrared spectrum to the rapid assay methods of color cotton coloration |
EP3620986A1 (en) * | 2018-09-05 | 2020-03-11 | Sartorius Stedim Data Analytics AB | Computer-implemented method, computer program product and system for analysis of cell images |
CN112858208A (en) * | 2019-11-27 | 2021-05-28 | 国能生物发电集团有限公司 | Biomass potassium content measurement and modeling method based on infrared spectrum principal component and neural network |
CN112861414A (en) * | 2019-11-27 | 2021-05-28 | 国能生物发电集团有限公司 | Biomass carbon content measurement and modeling method based on infrared spectrum principal component and neural network |
CN112861299A (en) * | 2019-11-27 | 2021-05-28 | 国能生物发电集团有限公司 | Biomass chlorine content measurement and modeling method based on infrared spectrum principal component and neural network |
CN112858205A (en) * | 2019-11-27 | 2021-05-28 | 国能生物发电集团有限公司 | Biomass hydrogen content measurement and modeling method based on infrared spectrum principal component and neural network |
CN112858209A (en) * | 2019-11-27 | 2021-05-28 | 国能生物发电集团有限公司 | Biomass nitrogen content measurement and modeling method based on infrared spectrum principal component and neural network |
CN114354534A (en) * | 2021-12-30 | 2022-04-15 | 中国航空油料有限责任公司 | Method for establishing aviation kerosene property prediction model by utilizing binary linear classifier |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2005045391A2 (en) * | 2003-11-06 | 2005-05-19 | Elan Group Ltd | System and process for detecting substances |
CN101021470A (en) * | 2007-03-13 | 2007-08-22 | 山东医学高等专科学校 | Method for Chinese patent drug fast qualitative analysis utilizing acousto-optic filter near infrared spectral technique |
CN104897709A (en) * | 2015-06-15 | 2015-09-09 | 江苏大学 | Agricultural product element quantitative detection model building method based on X-ray fluorescence analysis |
CN105334185A (en) * | 2015-09-14 | 2016-02-17 | 上海创和亿电子科技发展有限公司 | Spectrum projection discrimination-based near infrared model maintenance method |
-
2016
- 2016-07-25 CN CN201610589280.3A patent/CN107655850A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2005045391A2 (en) * | 2003-11-06 | 2005-05-19 | Elan Group Ltd | System and process for detecting substances |
CN101021470A (en) * | 2007-03-13 | 2007-08-22 | 山东医学高等专科学校 | Method for Chinese patent drug fast qualitative analysis utilizing acousto-optic filter near infrared spectral technique |
CN104897709A (en) * | 2015-06-15 | 2015-09-09 | 江苏大学 | Agricultural product element quantitative detection model building method based on X-ray fluorescence analysis |
CN105334185A (en) * | 2015-09-14 | 2016-02-17 | 上海创和亿电子科技发展有限公司 | Spectrum projection discrimination-based near infrared model maintenance method |
Non-Patent Citations (1)
Title |
---|
石雪 等: "局部建模方法用于烟草样品的近红外光谱定量分析", 《光谱学与光谱分析》 * |
Cited By (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108827907A (en) * | 2018-04-26 | 2018-11-16 | 新疆维吾尔自治区分析测试研究院 | It is a kind of based near infrared spectrum to the rapid assay methods of color cotton coloration |
CN108710920A (en) * | 2018-06-05 | 2018-10-26 | 北京中油瑞飞信息技术有限责任公司 | Indicator card recognition methods and device |
CN108710920B (en) * | 2018-06-05 | 2021-05-14 | 北京中油瑞飞信息技术有限责任公司 | Indicator diagram identification method and device |
WO2020049094A1 (en) * | 2018-09-05 | 2020-03-12 | Sartorius Stedim Data Analytics Ab | Computer-implemented method, computer program product and system for data analysis |
JP2022500744A (en) * | 2018-09-05 | 2022-01-04 | ザルトリウス ステディム データ アナリティクス アーベー | Computer implementation methods, computer program products and systems for analysis of cell images |
EP3620983A1 (en) * | 2018-09-05 | 2020-03-11 | Sartorius Stedim Data Analytics AB | Computer-implemented method, computer program product and system for data analysis |
WO2020049098A1 (en) * | 2018-09-05 | 2020-03-12 | Sartorius Stedim Data Analytics Ab | Computer-implemented method, computer program product and system for analysis of cell images |
CN112639834A (en) * | 2018-09-05 | 2021-04-09 | 赛多利斯司特蒂姆数据分析公司 | Computer-implemented method, computer program product, and system for data analysis |
CN112655004A (en) * | 2018-09-05 | 2021-04-13 | 赛多利斯司特蒂姆数据分析公司 | Computer-implemented method, computer program product, and system for anomaly detection and/or predictive maintenance |
CN112655003A (en) * | 2018-09-05 | 2021-04-13 | 赛多利斯司特蒂姆数据分析公司 | Computer-implemented method, computer program product and system for analysis of cellular images |
EP3620986A1 (en) * | 2018-09-05 | 2020-03-11 | Sartorius Stedim Data Analytics AB | Computer-implemented method, computer program product and system for analysis of cell images |
CN112655004B (en) * | 2018-09-05 | 2024-03-26 | 赛多利斯司特蒂姆数据分析公司 | Computer-implemented method for anomaly detection and/or predictive maintenance |
EP4290412A3 (en) * | 2018-09-05 | 2024-01-03 | Sartorius Stedim Data Analytics AB | Computer-implemented method, computer program product and system for data analysis |
JP7223839B2 (en) | 2018-09-05 | 2023-02-16 | ザルトリウス ステディム データ アナリティクス アーベー | Computer-implemented methods, computer program products and systems for anomaly detection and/or predictive maintenance |
JP7187681B2 (en) | 2018-09-05 | 2022-12-12 | ザルトリウス ステディム データ アナリティクス アーベー | Computer-implemented method, computer program product and system for analysis of cell images |
JP2022500745A (en) * | 2018-09-05 | 2022-01-04 | ザルトリウス ステディム データ アナリティクス アーベー | Computer implementation methods, computer program products and systems for anomaly detection and / or predictive maintenance |
WO2020049087A1 (en) * | 2018-09-05 | 2020-03-12 | Sartorius Stedim Data Analytics Ab | Computer-implemented method, computer program product and system for anomaly detection and/or predictive maintenance |
CN112858209A (en) * | 2019-11-27 | 2021-05-28 | 国能生物发电集团有限公司 | Biomass nitrogen content measurement and modeling method based on infrared spectrum principal component and neural network |
CN112858205A (en) * | 2019-11-27 | 2021-05-28 | 国能生物发电集团有限公司 | Biomass hydrogen content measurement and modeling method based on infrared spectrum principal component and neural network |
CN112861299A (en) * | 2019-11-27 | 2021-05-28 | 国能生物发电集团有限公司 | Biomass chlorine content measurement and modeling method based on infrared spectrum principal component and neural network |
CN112861414A (en) * | 2019-11-27 | 2021-05-28 | 国能生物发电集团有限公司 | Biomass carbon content measurement and modeling method based on infrared spectrum principal component and neural network |
CN112858208A (en) * | 2019-11-27 | 2021-05-28 | 国能生物发电集团有限公司 | Biomass potassium content measurement and modeling method based on infrared spectrum principal component and neural network |
CN114354534A (en) * | 2021-12-30 | 2022-04-15 | 中国航空油料有限责任公司 | Method for establishing aviation kerosene property prediction model by utilizing binary linear classifier |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107655850A (en) | Non-linear modeling method and system based near infrared spectrum | |
CN105046277B (en) | Robust mechanism study method of the feature significance in image quality evaluation | |
Feilhauer et al. | Multi-method ensemble selection of spectral bands related to leaf biochemistry | |
Suryawati et al. | Deep structured convolutional neural network for tomato diseases detection | |
CN107316066A (en) | Image classification method and system based on multi-path convolutional neural networks | |
WO2017143921A1 (en) | Multi-sampling model training method and device | |
CN104732240B (en) | A kind of Hyperspectral imaging band selection method using neural network sensitivity analysis | |
CN104020135B (en) | Calibration model modeling method based near infrared spectrum | |
CN107609399A (en) | Malicious code mutation detection method based on NIN neutral nets | |
CN109727246A (en) | Comparative learning image quality evaluation method based on twin network | |
CN107644415A (en) | A kind of text image method for evaluating quality and equipment | |
CN107145830A (en) | Hyperspectral image classification method with depth belief network is strengthened based on spatial information | |
CN106570516A (en) | Obstacle recognition method using convolution neural network | |
CN107590515A (en) | The hyperspectral image classification method of self-encoding encoder based on entropy rate super-pixel segmentation | |
CN104298999B (en) | EO-1 hyperion feature learning method based on recurrence autocoding | |
CN110457516A (en) | A kind of cross-module state picture and text search method | |
CN104751179A (en) | Multi-target high spectral remote sensing image wave band selection method based on game theory | |
CN107609573A (en) | High spectrum image time varying characteristic extracting method based on low-rank decomposition and empty spectrum constraint | |
CN112270441A (en) | Method for establishing autism child rehabilitation effect prediction model and method and system for predicting autism child rehabilitation effect | |
CN105631474A (en) | Hyperspectral data multi-class method based on Jeffries-Matusita distance and class pair decision tree | |
CN106706552A (en) | Method for detecting wheat quality based on multiple-information integration | |
CN107153046A (en) | Near infrared spectrum based on concentration residual information detects the method and system of fruit quality | |
US11525774B2 (en) | Sensory evaluation method for spectral data of mainstream smoke | |
CN111398238A (en) | Laser-induced fluorescence spectrum identification method for edible oil doped with castor oil | |
CN106485049A (en) | A kind of detection method of the NIRS exceptional sample based on Monte Carlo cross validation |
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 |
Application publication date: 20180202 |
|
RJ01 | Rejection of invention patent application after publication |