CN107918718A - Sample composition content assaying method based on online order limit learning machine - Google Patents
Sample composition content assaying method based on online order limit learning machine Download PDFInfo
- Publication number
- CN107918718A CN107918718A CN201711068234.XA CN201711068234A CN107918718A CN 107918718 A CN107918718 A CN 107918718A CN 201711068234 A CN201711068234 A CN 201711068234A CN 107918718 A CN107918718 A CN 107918718A
- Authority
- CN
- China
- Prior art keywords
- mrow
- msub
- mtd
- sample
- msup
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16Z—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS, NOT OTHERWISE PROVIDED FOR
- G16Z99/00—Subject matter not provided for in other main groups of this subclass
Landscapes
- Investigating Or Analysing Materials By Optical Means (AREA)
- Other Investigation Or Analysis Of Materials By Electrical Means (AREA)
Abstract
The invention discloses a kind of sample composition content assaying method based on online order limit learning machine, including:The spectroscopic data sample of sample is gathered, and is modeled using online order limit learning machine algorithm;Established model is utilized to be measured the component content of sample.The present invention is modeled by using online order limit learning machine algorithm, only remains the knowledge above learnt for future use without retaining the data of mistake used above;When having new spectroscopic data arrival every time, it is only necessary to calculate the new hidden layer to data and export, then utilize the knowledge above learnt dynamically to update middle hidden layer to the output weight between output layer, you can to carry out rapid modeling.Compared to traditional modeling method, improve modeling speed, reduce the unnecessary amount of computing repeatedly and the consumption to data space, improve the precision and Generalization Capability of model, and the data to arrive one by one every time can be handled or handle the data of one piece of one piece of arrival.
Description
Technical field
The present invention relates to a kind of sample composition content assaying method based on online order limit learning machine, belong to sample into
Divide 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 near infrared spectrum of sample with effectively
Polynary peg model between constituent content, and then response component that can be to unknown sample is predicted.However, actually make
Near infrared spectrum data is not once to produce during, but what streaming produced.If in existing data sample
Model is established on this, and is produced as the change of time might have new data sample, in order to improve the generalization of model
Energy and precision of prediction, are necessarily required to newly generated data and pervious data being modeled together.Most simple direct method
Existing all data are exactly reruned into an original algorithm, but this method be when data volume very little can
With receiving, and if data are weighed in units of GB, the data sample newly to arrive may up to count MB, then so
By original data, new data establish model together in addition, are very time-consuming effort, sometimes may above new arrive
Data also without processing complete, and have renewal data arrive, it is clear that in this case, completely again modeling be impossible
Accomplish.Online streaming algorithm is also increasingly suitable for having radial basis function (radial basis function, RBF) section
The feedforward neural network of point.There are many algorithms in online streaming learning algorithm evolution is handled, wherein than more typical
Algorithm have GAP-PBF algorithms and GGAP-RBF algorithms.It can simplify learning process it is desirable to these algorithms and improve study
Speed, these algorithms need the distribution of input sample or the information of input sample order.But the modeling speed of these algorithms
It is still very slow, and Generalization Capability is also general.And can only be (one by one by one during these algorithm process newcomer's data
One) and cannot be one piece one piece (chunk by chunk).
In addition, in the measurement process of near infrared spectrum, due to the change of the difference or measuring condition of measuring instrument, meeting
Cause original polynary peg model to lose effect, and it is the something taken time and effort to re-establish model, in addition have when
Wait modeling again and do not have feasibility.More acceptable mode is to do calibration migration, for correcting main instrument and another instrument
The spectroscopic data of (sub- instrument).Substantially, it is exactly to change the spectrum of sub- instrument, is allowed to appear more like the number of key light spectrometer
According to, then can use key light spectrometer model it handle.In the past few years, different calibration migration skill
Art is developed, and common calibration moving method includes:Multiplicative scatter correction method (being abbreviated as MSC), direct standardized method (letter
Be written as DS), indirect standardization method (being abbreviated as PDS), canonical correlation analysis method (being abbreviated as CCA) etc..But it is existing on
State the problem of calibration moving method still has component content precision of prediction and poor stability;And the polynary scattering school
Execute, it is necessary to measure the preferable spectrum of a sample to be tested, then other spectrum measured are repaiied using the ideal spectrum
Just, so-called preferable spectrum but is hardly resulted in practical applications.
The content of the invention
It is an object of the present invention to provide a kind of sample composition assay side based on online order limit learning machine
Method, it can effectively solve problems of the prior art, and especially existing algorithm modeling speed is slow, Generalization Capability one
As, and the problem of can only one by one handle new data and cannot be handled block by block.
In order to solve the above technical problems, the present invention adopts the following technical scheme that:One kind is learnt based on online order limit
The sample composition content assaying method of machine, gathers the spectroscopic data sample of sample, and utilizes online order limit learning machine algorithm
It is modeled;Established model is utilized to be measured the component content of sample.
The foregoing sample composition content assaying method based on online order limit learning machine, specifically includes following steps:
S1, according to initial principal spectrum SPmaster(0)And corresponding sample composition content y0With hidden node number L, calculate
Initial weight matrix α of the hidden layer to output layer(0), wherein, SPmaster(0)And y0Include M0A sample;
S2, when there is new principal spectrum SPmaster(k+1)And corresponding sample composition content yk+1During arrival, according to online order
Extreme learning machine algorithm calculates hidden layer to the weight matrix α of output layer(k+1);Wherein, the data of+1 arrival of kth
SPmaster(k+1)And yk+1Include Mk+1A sample;k≥0;
S3, if also there are new principal spectrum SPmaster(k+1)With corresponding sample composition content yk+1, then k=k+1 is made, and
S2 is gone to, otherwise goes to S4;
S4, the spectroscopic data sp for obtaining sample is calculated according to the following formulamasterPreCorresponding component content predicted value
pre_y:
Pre_y=Hpreα(new);
Wherein,
α(new)For current newest hidden layer to the weight matrix of output layer, spmasterPreInclude N number of sample;W and b are respectively
The orthogonal input weight matrix generated at random and biasing;G(w,spmasterPre, b) and it is activation primitive.
Preferably, further include:Using online order limit learning machine algorithm, gathered to key light spectrometer and from spectrometer
The spectroscopic data sample of sample is modeled, and realizing will migrate to the spectroscopic data sky of key light spectrometer from the spectroscopic data of spectrometer
Between;Then the content prediction model established using key light spectrometer carries out the measure of sample composition content.
It is furthermore preferred that the online order limit learning machine algorithm of the utilization, gathers to key light spectrometer and from spectrometer
The spectroscopic data sample of sample be modeled, realizing will migrate to the spectroscopic data of key light spectrometer from the spectroscopic data of spectrometer
Space comprises the following steps:
S01, according to initial principal spectrum SPmaster0With from spectrum spslave0And the number of hidden nodes L, generation hidden layer to output layer
Weight matrix β(0), wherein SPmaster0In the number of samples that includes be M0;
S02, new M is included when havingk+1The SP of a samplemaster(k+1)And spslave(k+1)During arrival, according to online order pole
Limit learning machine algorithm calculates hidden layer to the weight matrix β of output layer(k+1);Wherein, k >=0;
S03, if the SP of also new samplemaster(k+1)And spslave(k+1)Arrive, then make k=k+1, and go to S02,
Otherwise S04 is gone to;S04, according to the following formula to the test data sp comprising N number of sampleslaveTestMigrated:
spslaveTomaster=H'testβnew
Wherein, spslaveTomasterRepresent the spectroscopic data after migration;βnewRepresent newest hidden layer to the power of output interlayer
Weight matrix;
W and b be respectively with
The orthogonal input weight matrix of machine generation and biasing;G(w,spslaveTest, b) and it is activation primitive.
In the foregoing sample composition content assaying method based on online order limit learning machine, the hidden node
Number L is less than or equal to initial number of samples M0.So that the sample composition content prediction model based on OSELM has meter faster
Speed is calculated, and the consuming to system resource is less.
Preferably, in step S1,
α(0)=(H0 TH0)-1H0 Ty0;
Wherein,
Preferably, in step S2,
Wherein,
Preferably, in step S01,
β(0)=(H'0 TH'0)-1H'0 Tspmaster0;
Wherein,
Preferably, in step S02,
Wherein,
In the present invention, the method that cross validation is rolled over by k determines optimal the number of hidden nodes L;The activation primitive uses
Sigmoid functions, so as to improve the precision of prediction of sample composition content.
Sample composition content assaying method of the present invention, the spectrum samples of all kinds suitable for on-line checking,
There is more preferable effect especially for the component content of tablet and corn measure.
Preferably, the spectroscopic data space to key light spectrometer is migrated as new principal spectrum using from the spectroscopic data of spectrometer
SPmaster(k+1), corresponding sample is obtained after the component content measure of sample is carried out using the content prediction model of key light spectrometer foundation
Product component content yk+1, k=k+1 is made, and go to S2.So as to further improve the precision of prediction of model.
Compared with prior art, the present invention is modeled by using online order limit learning machine algorithm, without
Retain the data of mistake used above and only remain the knowledge above learnt for future use;There is new spectroscopic data every time
During arrival, it is only necessary to calculate the new hidden layer to data and export, then update middle hidden layer using the knowledge dynamic above learnt
To the output weight between output layer, you can carry out rapid modeling.The present invention improves modeling compared to traditional modeling method
Speed, reduces the unnecessary amount of computing repeatedly and the consumption to data space, improves the precision of model and extensive
Performance, and the data to arrive one by one every time can be handled or handle the data of one piece of one piece of arrival.In addition, this
Invention is by using online order limit learning machine algorithm, the spectroscopic data of the sample gathered to key light spectrometer and from spectrometer
Sample is modeled, and realizing will migrate to the spectroscopic data space of key light spectrometer from the spectroscopic data of spectrometer;Then master is utilized
The content prediction model that spectrometer is established carries out the measure of sample composition content, so as to improve the essence of sample composition content prediction
Degree and stability.Test result indicates that:The calibration moving method based on online order limit learning machine algorithm of the present invention, in medicine
The spectrum migration algorithm better performance compared to PDS and based on CCA is shown on sheet data collection and corn data set.
Brief description of the drawings
Fig. 1 is that the spectrum (a) in corn data set is MP5, and (b) is M5, and (c) is MP6 spectrum;
Fig. 2 is tablet data set owner spectrum (a) from spectrum (b), the two deviation spectrum (c);
Fig. 3 is change schematic diagrams of the RMSEP with the number of hidden nodes;
Fig. 4 is M5 the and MP5 deviations (a) after corn data set migration and M5 the and MP5 deviations (b) without migration;
Fig. 5 is influence schematic diagram of the modeling sample number to RMSEP;
Fig. 6 is the RMSEP of physics and chemistry sex character and the relation schematic diagram of the number of hidden nodes;
Fig. 7 is predicted value schematic diagram of all kinds of spectrum on water;
Fig. 8 be on protein (a), starch (b), oily (c), water (d) predicted value situation schematic diagram;
Fig. 9 is the migration algorithm based on ELM and the migration algorithm modeling run time contrast schematic diagram based on OSELM;
Figure 10 is the relation schematic diagram of hidden node and spectrum RMSEP in tablet data set;
Figure 11 is the schematic diagram for not migrating residual error based on tablet data (a) spectrum migration residual error (b);
Figure 12 is the predicted value of the first active ingredient and the comparison schematic diagram of actual value in tablet;
Figure 13 is the comparison of (a) second active ingredient and (b) the third active ingredient predicted value and actual value in tablet
Schematic diagram;
Figure 14 is the prediction result contrast schematic diagram of PDS, CCA, TLOSELM based on corn data water content;
Figure 15 is the prediction result contrast schematic diagram of PDS, CCA, TLOSELM based on the third active ingredient of tablet;
Figure 16 is the method flow schematic diagram of the present invention.
The present invention is further illustrated with reference to the accompanying drawings and detailed description.
Embodiment
The embodiment of the present invention 1:A kind of sample composition content assaying method based on online order limit learning machine, such as schemes
Shown in 16, the spectroscopic data sample of sample is gathered, and be modeled using online order limit learning machine algorithm;Using being established
Model the component content of sample is measured.
Specifically it may include following steps:
S1, according to initial principal spectrum SPmaster(0)And corresponding sample composition content y0With hidden node number L, calculate
Initial weight matrix α of the hidden layer to output layer(0), wherein, SPmaster(0)And y0Include M0A sample;
S2, when there is new principal spectrum SPmaster(k+1)And corresponding sample composition content yk+1During arrival, according to online order
Extreme learning machine algorithm calculates hidden layer to the weight matrix α of output layer(k+1);Wherein, the data of+1 arrival of kth
SPmaster(k+1)And yk+1Include Mk+1A sample;k≥0;
S3, if also there are new principal spectrum SPmaster(k+1)With corresponding sample composition content yk+1, then k=k+1 is made, and
S2 is gone to, otherwise goes to S4;
S4, the spectroscopic data sp for obtaining sample is calculated according to the following formulamasterPreCorresponding component content predicted value
pre_y:
Pre_y=Hpreα(new)
Wherein,
α(new)For current newest hidden layer to the weight matrix of output layer, spmasterPreInclude N number of sample;W and b are respectively
The orthogonal input weight matrix generated at random and biasing;G(w,spmasterPre, b) and it is activation primitive.
In order to enable the sample composition content prediction model based on OSELM has faster calculating speed, and system is provided
The consuming in source is less, and the hidden node number L is less than or equal to initial number of samples M0。
In step S1,
α(0)=(H0 TH0)-1H0 Ty0;
Wherein,
In step S2,
Wherein,
In order to carry out accurate content prediction to the spectroscopic data gathered from spectrometer, further include:Utilize online order
Extreme learning machine algorithm, the spectroscopic data sample of the sample gathered to key light spectrometer and from spectrometer are modeled, and realizing will
Migrated from the spectroscopic data of spectrometer to the spectroscopic data space of key light spectrometer;Then the content prediction established using key light spectrometer
Model carries out the measure of sample composition content.
Wherein, the online order limit learning machine algorithm of the utilization, the sample gathered to key light spectrometer and from spectrometer
The spectroscopic data sample of product is modeled, and realizing will migrate to the spectroscopic data space of key light spectrometer from the spectroscopic data of spectrometer
It may include following steps:
S01, according to initial principal spectrum SPmaster0With from spectrum spslave0And the number of hidden nodes L, generation hidden layer to output layer
Weight matrix β(0), wherein SPmaster0In the number of samples that includes be M0;
S02, new M is included when havingk+1The SP of a samplemaster(k+1)And spslave(k+1)During arrival, according to online order pole
Limit learning machine algorithm calculates hidden layer to the weight matrix β of output layer(k+1);Wherein, k >=0;
S03, if the SP of also new samplemaster(k+1)And spslave(k+1)Arrive, then make k=k+1, and go to S02,
Otherwise S04 is gone to;S04, according to the following formula to the test data sp comprising N number of sampleslaveTestMigrated:
spslaveTomaster=H'testβnew
Wherein, spslaveTomasterRepresent the spectroscopic data after migration;βnewRepresent newest hidden layer to the power of output interlayer
Weight matrix;
W and b be respectively with
The orthogonal input weight matrix of machine generation and biasing;G(w,spslaveTest, b) and it is activation primitive.
In order to enable the sample composition content prediction model based on OSELM has faster calculating speed, and system is provided
The consuming in source is less, and the hidden node number L is less than or equal to initial number of samples M0。
Specifically, in step S01,
β(0)=(H'0 TH'0)-1H'0 Tspmaster0;
Wherein,
In step S02,
Wherein,
In above method, the method that cross validation is rolled over by k determines optimal the number of hidden nodes L;The activation primitive is adopted
With sigmoid functions.
Embodiment 2:A kind of sample composition content assaying method based on online order limit learning machine, gathers the light of sample
Modal data sample, and be modeled using online order limit learning machine algorithm;Utilize component of the established model to sample
Content is measured.
Specifically it may include following steps:
S1, according to initial principal spectrum SPmaster(0)And corresponding sample composition content y0With hidden node number L, calculate
Initial weight matrix α of the hidden layer to output layer(0), wherein, SPmaster(0)And y0Include M0A sample;
S2, when there is new principal spectrum SPmaster(k+1)And corresponding sample composition content yk+1During arrival, according to online order
Extreme learning machine algorithm calculates hidden layer to the weight matrix α of output layer(k+1);Wherein, the data of+1 arrival of kth
SPmaster(k+1)And yk+1Include Mk+1A sample;k≥0;
S3, if also there are new principal spectrum SPmaster(k+1)With corresponding sample composition content yk+1, then k=k+1 is made, and
S2 is gone to, otherwise goes to S4;
S4, the spectroscopic data sp for obtaining sample is calculated according to the following formulamasterPreCorresponding component content predicted value
pre_y:
Pre_y=Hpreα(new)
Wherein,
α(new)For current newest hidden layer to the weight matrix of output layer, spmasterPreInclude N number of sample;W and b are respectively
The orthogonal input weight matrix generated at random and biasing;G(w,spmasterPre, b) and it is activation primitive.
In order to enable the sample composition content prediction model based on OSELM has faster calculating speed, and system is provided
The consuming in source is less, and the hidden node number L is less than or equal to initial number of samples M0。
Specifically, in step S1,
α(0)=(H0 TH0)-1H0 Ty0;
Wherein,
Specifically, in step S2,
Wherein,
In order to verify the effect of the present invention, inventor is using in PDS algorithms and the migration algorithm based on CCA and literary invention
Migration algorithm based on OSELM (i.e. online order limit learning machine) is contrasted.During experiment, it is necessary to by slave
Algorithm model of the spectrum using PDS, CCA and based on OSELM is migrated to master spaces.The spectroscopic data after migration is brought into again
Component content prediction is carried out in the spectrum that master is spectrally established prediction model corresponding with component content.
1.1 experimental situation
This experiment is completed based on python2.7.The operating system of computer is win8.1,64 bit manipulation systems, and CPU is
The A84500 of AMD, inside saves as 8GB.Experiment used python some it is common bag such as:Numpy, matplotpy and sklearn
Bag.The program used in experiment is all to develop to complete on Integrated Development Environment Eclipse.
1.2 experimental data
Used corn NIR light spectrum data set in this experiment, this NIR light spectrum data set include 80 it is different
NIR data samples.M5 is included in data set, MP5, MP6 tri- opens NIR spectra tables of data, and (this three spectroscopic data tables are from not
NIR light spectrometry with spectrometer to same substance, that is, corn), also have in this data set and this 80 spectroscopic data samples
Corresponding physics and chemistry sex character table, such as:Water content (water), oily (soil), protein content (protein), content of starch
(starch).The wave-length coverage of spectroscopic data is from 1100nm to 2498nm, at intervals of 2nm (including 700 passages).MP5 this
Spectroscopic data table comes from the FOSS NIRSystem5000 as main (master) instrument, and M5, MP6 are then to come to be used as subordinate
(slave) spectrum is measured from FOSS NIRSystem 6000 and FOSS NIRSystem 5000 respectively.
(a) subgraph is the spectrogram of MP5 in Fig. 1, and (b) subgraph is the spectrogram of M5, and (c) subgraph is the spectrogram of MP6, from
It can be seen that the image of MP5 and MP6 is quite similar in Fig. 1, because they are come out with the spectrometer measurement of same model.
And M5 is then larger with MP5 image differences, M5 has compared to MP5 significantly moves up trend.
Tablet spectroscopic data collection is also used in this experiment, which is 2002, a NIR spectra that IDRC is delivered
Data, including 654 kinds of tablets from two spectrometers.This two spectrometers be respectively FOSS NIRSystems and
Silver-Spring.Two parts of NIR light modal datas from different spectrometers of this part are divided into calibration set (calibration
Set includes 155 spectroscopic data samples) and test set (test set include 460 spectroscopic data samples) also verification collection
(validation set include 40 spectroscopic data samples).Spectral wavelength in data set concentrate on 1100nm~1750nm it
Between.
What subgraph (a) and subgraph (b) were drawn respectively in Fig. 2 is verification collection in tablet data set in two spectrometers
SPEC1With spectrometer SPEC2On to the collection result of same substance.What subgraph (c) represented in Fig. 2 is the spectrum measured on SPEC1
As a result with measuring the difference of result on SPCE2.It can be seen that the spectral differences drawn on two spectrometers from the result of subgraph (c)
The opposite sex is not apparent in the difference of most of wave band.The result of subgraph (c) is even it may be said that the result extremely phase that both measure
Seemingly, because difference tends to 0 substantially, the wave band change only after 1700nm just starts bigger.
1.3 experimental design
1.3.1 the experiment based on corn
The NIR light modal data of corn is divided into three parts, by data division calibration set (calibration set), verification
Collect (validation set) and test set (test set).Calibration set includes 56 NIR light modal datas, and verification collection includes 8
NIR light modal data, test set also include 16 NIR light modal datas.Wherein calibration set is for establishing online migration models, and is tested
Card collection is then for selecting optimal the number of hidden nodes.Test set is for showing algorithm Generalization Capability.
1st, the selection of the number of hidden nodes L optimal to parameter
In order to choose optimal the number of hidden nodes for migration models, inventor has used the mode of 8 folding cross validations
Choose.The root-mean-square error picture that different the number of hidden nodes are obtained by cross validation is gone out.Spectrum after wherein migrating is equal
The calculation formula of square error is as follows:
Formula (1) is exactly the formula for calculating root-mean-square error, and wherein NumSample represents this batch NIR spectra to be migrated
Number of samples.As can be seen from Figure 3 with the increase of the number of hidden nodes purpose, RMSEP slowly rises after falling before, and works as hidden layer
When node reaches certain amount, the growth rate of RMSEP is substantially accelerated.It can be drawn from figure:With regard to the root mean square of spectrum migration
For error, it is the most suitable that hidden node is arranged to 19.
Therefore, the hidden node L of model is arranged to optimal number of nodes 19.Give algorithm 32 sp when initialmasterNIR spectra
Data sample and 32 spslaveNIR spectra data sample.Respectively as outputting and inputting for model, its essence namely will
spslaveTo spmasterMigration.The output weight matrix β of starting can be drawn by establishing initial migration models0, it is then assumed that subsequently
There is a NIR spectra data sample to arrive every time, the thought based on OSELM calculates the output weight matrix β after renewal(k).When having
Need to substitute into the migration models established during the slave spectrum migrated.NIR light modal data is so just realized based on OSELM
The purpose that algorithm is migrated to master spectral spaces.
Subgraph (a) above in Fig. 4 is to spslaveTestThe result obtained after migration subtracts corresponding spmasterTestKnot
Fruit, subgraph (b) below is spslaveTestDirectly subtract spmasterTestResult.(a) is it can be seen that pass through OSELM from the graph
The modeling of transfer learning is realized, most sp in subgraph (a) aboveslaveTestThere is preferable migration effect, after migration
spslaveTestWith spmasterTestDifference all have slight fluctuation near zero substantially.And the slave spectrum without migration with
The difference of master spectrum fluctuates between 0.04~0.06.
2nd, influence of the sample number for carrying out initially just entering during calibration migration to spectrum to arithmetic accuracy
The model is the transfer learning model established based on OSELM.There is the problem of initialization in OSELM algorithms just
Be using how many a samples as it is initial when the sample that is possessed generate basic model.Sample difference is right when very possible initial
The migration models of foundation can have an impact, this form 1 below will show influence of the initial sample number to model.
InitNum represents the NIR spectra number of samples possessed during initialization in table 1, and RMSEP, which remains unchanged, represents the equal of prediction
Square error.This time experiment hidden node is both configured to 15 and sample arrives one by one.Drawn from the data in table 1
Such conclusion:The number of NIR spectra data sample has to be more than or equal to hidden during the initialization of the migration models based on OSELM
Node layer number.Sample number when initially is more than or equal to the number of hidden nodes, then the error of prediction is maintained for constant.On in fact
The result in face also complies with the intension of OSELM algorithms, final profit during the NIR light spectrum data set line modeling that online streaming arrives
The data sample of all arrivals is used, OSELM You nothing different in fact from the essence of ELM, and OSELM is as line modeling
Algorithm is only the reduction of the unnecessary amount of computing repeatedly.But if the NIR sample numbers possessed when initial are less than hidden node
Number, then solving output matrix β(0)When may be related to the problem of matrix is violated in solution, the result of such migration process is equal
Square error will be bigger.
The relation of 1 initialization sample number of table and RMSEP
3rd, influence of the data block size that order arrives every time to arithmetic accuracy
All assume that what sample arrived one by one in experiment more than, look at the side that NIR light modal data arrives below
Whether formula can have an impact the precision of model.The default setting of experiment is that hidden node is 15, and sample number is 32 when initial.
Step represents the NIR spectra number of samples come every time in table 2.No matter it can be seen that NIR spectra data sample from the content in table
It is that one piece one piece next or that comes one by one does not all influence the performance of algorithm.The NIR samples of arrival are all finally sharp
Use, thus for test when, as long as they model when used equal number of NIR samples and the number of hidden nodes L, that
Root-mean-square error on prediction is exactly the same.
Influence of the 2 stream data sample size of table to RMSEP
4th, based on influence of the number of samples for precision of prediction in OSELM migration models
What is represented in Fig. 5 is the NIR light modal data after being migrated by OSELM as NIR spectra data sample order arrives
With spmasterTestRoot-mean-square error size.It can be seen that with on-line study carry out accumulation get up NIR data samples it is more next
More, the root-mean-square error general trend after being migrated to test set is less and less.In figure it can be found that about the 55th, 56
There is the situation for predicting error increase in a sample when arrival, this is probably more abnormal data sample occurred,
Situation is caused to be deteriorated.But with regard to predict general trend for, the time longer accumulation of on-line study on corn spectroscopic data
Knowledge also more and then transfer learning accuracy have also just been lifted.From the data in figure it is concluded that:Arrive
The models of the more foundation of sample number also more accurate more stable generalization ability is also stronger.This conclusion is equivalent to having responded
The necessity of line modeling, if data be not all of disposably arriving every time but one piece one piece or arrival one by one and
And there are the needs of prediction in the meantime, then just must be based on existing sample data and establish model, when there is sample next time
Only need more new model can be without from the beginning construction model again again, where this is the intension of OSELM during arrival.
In experiment above inquired on the spectrum and the root-mean-square error of principal spectrum after migration, but the migration
The ultimate purpose of model is made in SPECslaveOn the sp that collectsslaveIt can utilize in SPECmasterUpper foundation it is online more
New model, i.e., finally wish that SPEC can be passed throughmasterThe model prediction of upper foundation goes out on spslavePhysics and chemistry sex character.Therefore exist
Effect that will be to be migrated in terms of the height of physics and chemistry sex character in experiment below.Due to spslaveTo spmasterThe process of migration is
Carry out online, therefore for spmasterPrediction model between corresponding physics and chemistry sex character y is also required to dynamic and updates.Following reality
Test the sp that will be collected first on slave spectrometersslaveOnline moves to spmasterSpace where data, then will move
Spectrum after shifting, which substitutes into, is based on spmasterOn-line prediction model between y.Need to pass through sp in corn data setslaveData
Carry out on water content, the prediction of protein content, oil content, content of starch.
5th, the performances of migration models is treated in the height of physical and chemical index
Cross validation is rolled over by k first and chooses principal spectrum to the optimal the number of hidden nodes of physics and chemistry sex character prediction model.Below
Experiment in the spectroscopic data after migration substituted into most newly-established physics and chemistry sex character prediction model.From table 3 it is observed that L tables
Show migration models the number of hidden nodes, RMSEP represents the root-mean-square error on the prediction of physics and chemistry sex character.If it can be seen that hidden layer section
Points are set too small, then and the root-mean-square error of prediction result can be bigger, and if the number of hidden nodes is more than some point
When, prediction error can also start to increase.Extremely calculated in table it is apparent that when the number of hidden nodes is more than initial sample number
Root-mean-square error can drastically become larger.
Influence of 3 the number of hidden nodes of table to the RMSEP of physical and chemical index
Result is shown in order to apparent, the relation of RMSEP and the number of hidden nodes L can be used line chart
Draw.Fig. 6 more can intuitively find out that its trend is come.
Abscissa represents water in Fig. 7testValue and ordinate represents corresponding predicted value pre_water.Triangle represents
Directly by spslvaeTestBring the predicted value that the prediction model of physics and chemistry sex character obtains into, five-pointed star is represented spslaveTestMigrated
Bring the predicted value of the prediction model of physics and chemistry sex character into afterwards.Plus sige then represents spmasterTestBring the predicted value of model into.Can in Fig. 7
With find out the spectroscopic data after migration predict the physics and chemistry attribute and the actual value that come be very close to, if SPECslaveOn
The spectrum collected just directly uses SPEC without migrationmasterThe physics and chemistry sex character prediction model of upper foundation, then the error of prediction
It is very big.As can be seen from Figure 7 very big success is achieved on corn data set based on the migration models of OSELM.
As shown in figure 8, triangle, star and plus sige in subgraph are the same meanings with upper graph expression.Wherein subgraph
(a), (b), (c), the abscissa of (d) are followed successively by protein, starch, oil content and water content, and the ordinate of every width subgraph is successively
For the predicted value of corresponding abscissa.First width subgraph is described based on the prediction mould that corn physics and chemistry sex character is protein foundation
Type.It is the prediction model that starch (starch) is established that second width subgraph, which describes to be based on physics and chemistry sex character in corn data set,.The
Three width subgraphs describe the prediction model established based on oil content in corn.4th width figure is exactly figure 7 above.Result in Fig. 8
It can clearly find out the effect of spectroscopic data migration.In (a), (b), (c), (d) four width subgraph, between triangle and star
There is obvious decompose to be spaced.
6th, the operational efficiency of OSELM
The performance of 9 the first row of sectional drawing is that traditional ELM runs the time once needed.Second row represents that OSELM is migrated online
11 required times.The third line is represented if modeling the time of 11 needs with ELM algorithms.As can be seen from Figure 9
The time of this experiment ELM migration algorithm of operation of corn data set is about 0.02 second, and OSELM migration algorithms model
About run 0.04 second for 11 times.It is not difficult to draw based on the migration models of OSELM than being based only on ELM by simply calculating
Migration models block it is very much.It is because OSELM reduces the amount of computing repeatedly soon that why OSELM models repeatedly than traditional ELM.
OSELM remains knowledge during last modeling, and if new data arrive, it only needs to model new knowledge.
7th, the summary based on the experiment of corn data set:
The NIR spectra Data Migration model based on OSELM is can be seen that on corn data set from experimental result above
Show more satisfactory performance.Spectrum after migration substitutes into the prediction model knot on physics and chemistry sex character with the spectrum without migration
Fruit has obvious difference.The advantages of algorithm is opposite to be had faster with traditional on-line learning algorithm, and adaptability is stronger.In corn
No matter data are that each arrival or two, three or more, the algorithm model one by one can solve on data set.And
Size of the algorithm before bringing into operation next time without the specified data volume brought.The algorithm makes salve NIR spectras without weight
Rebuild mould and corresponding water content, fat content, protein content, starch are directly predicted using the model on master instruments
Content is possibly realized.
1.3.2 the experiment based on tablet
Data in tablet data set are collected on two different spectrometers.Facilitate this to state below
In spectrometer 1 is denoted as SPEC1, spectrometer 2 is denoted as SPEC2。SPEC1The NIR light modal data collected is denoted as sp1, SPEC2Collection
Obtained NIR light modal data is denoted as sp2.Having carried 9 different tables when data are introduced above in data set is respectively:Calibration
Collect (calibrate_1, calibrate_2, calibrate_Y), verification collection (validate_1, validate_2,
Validate_Y), test set (test_1, test_2, test_Y).Label 1,2 therein represents that NIR light modal data comes from respectively
Spectrometer SPEC1And SPEC2.Calibrarte set are usually to be used for establishing model and calibrate in machine learning algorithm
Set is then the hyper parameter for selection algorithm model.Test set can generally be used for the performance of testing model, carry it into foundation
Good model is obtained a result.To model generalization ability, accuracy and whether can occur the one of over-fitting by the result
Sample judgment mode.But this data set calibrate set there was only 155 samples and test up to 460 samples.So
Situation and do not meet the setting of test set in general machine learning algorithm, therefore experimental data is drawn again in this experiment
Point.This experiment it is total temporarily should not validate set, the hyper parameter in algorithm can be showed by the result of test of many times
It changes the change for causing algorithm performance.Assume that the data for testing are 40 NIR spectra data samples in this experiment.
1st, optimal the number of hidden nodes is chosen based on spectrum migration
For different data sets, the selection of optimal the number of hidden nodes is necessarily different.Tablet data set possesses phase
Compared with the more NIR spectra number of samples of corn data set.Therefore seek β establishing preliminary examination migration models(0)When, initial sample number
It is arranged to 100.In this context, it is desirable to probe into the root-mean-square error based on spectrum migration and the selection of the number of hidden nodes L.
Abscissa hiddenNum in lower Figure 10 represents the corresponding ordinate of hidden node number in migration algorithm
RMSEP represents the root-mean-square error of spectrum and corresponding principal spectrum after migration.It can be seen that RMSEP is with hidden node in Figure 10
The increase of number shows the unstable state fluctuated up and down, but is remained for general trend and downward trend is presented, when
When hiddenNum is about 59, the value of RMSEP is minimum.As hiddenNum continues to increase, RMSEP starts to show on slowly
The trend risen.When the number of hidden nodes exceedes it is initial when assign sample number when can see RMSEP in figure and can sharply increase.
As seen from Figure 11 by OSELM algorithms migration after NIR light modal data substantially concentrate on 0 nearby fluctuation and substantially
Scope is [- 0.02,0.02].But the migration effect for having a few sample is not fine.This is probably because original data sheet
The difference of body is with regard to very little.If examine figure to can be found that in SPEC1NIR data of upper collection and in SPEC2Upper collection
Difference of the NIR data differences between [- 0.1,0.1] and as can be seen from the figure between most of wave band, spectrum is basic
Level off to 0.
2nd, influence of the selection of activation primitive to arithmetic accuracy
The sometimes selection of activation primitive has algorithm performance direct influence, such as tanh functions map end value
To between [- 1,1], and between end value is then mapped to [0,1] by sigmoid functions, so the result difference drawn also just than
It is larger.Therefore influence of the activation primitive to arithmetic accuracy is probed into this part Preparatory work of experiment.Used here as common sigmoid,
Tanh, tribas, hardlim function.Following experiment is just started with from the optimal activation primitive of selection.In table 4 RMSEP represent on
The root-mean-square error size of first physics and chemistry sex character predicted value in tablet data set, N represent to participate in the number of samples of modeling, this reality
Middle 20 NIR samples of initial input are tested, it is 15 to set hidden node L.From experimental result it can be found that no matter which kind of activation chosen
Function, with model learning to sample number increase RMSEP all be present downward trend.Data in contrast form can be known
The performance that sigmoid functions are shown on tablet data set is better than tanh, tribas, hardlim function.
Influence of 4 activation primitive of table to arithmetic accuracy
3rd, the relation of input sample number and hidden node is initialized
Experimental Research is when migration models migrate tablet NIR light modal data below, the NIR light modal data that when initialization inputs
The setting of sample number and the number of hidden nodes L, which can produce the performance of algorithm what kind of, to be influenced.Transverse axis represents the number of hidden nodes in table 5
L, Sn represent the NIR sample numbers that input during initialization, the digital representation RMSEP among form.Place under table 5 is hollow represents this
It is very big to locate error.From the numeral change in form will become apparent from once the number of hidden nodes set sample number when exceeding initial that
RMSEP will increased dramatically, and error can reach thousands of sometimes.Therefore confirm the number of hidden nodes no more than just again
The sample number during beginning.Numeral in form can also find out the NIR spectra number of samples for when hidden node is constant, initializing input
Change substantially on RMSEP not influence (but sample number needs to be more than or equal to the number of hidden nodes).
The relation of input sample number and hidden node is initialized in 5 tablet data set of table
4th, influence of the data block size to arrive every time to arithmetic accuracy
Following form is used for representing whether the data block size that streaming arrives has an impact the precision of algorithm.In this reality
In testing, hidden node is arranged to 90, and the NIR sample numbers of initialization are 100.
It can be seen that from the data in table 6 with OSELM algorithm process stream datas, the size that each data arrive is not
It can influence the precision of algorithm.OSELM algorithms are requiring no knowledge about the data block size that arrives next time.And OSELM algorithms have
The advantages of other most of algorithms do not have is exactly that can handle the data sample to arrive one by one or handle big
The data block of small change.
The relation of the fast sample number size of 6 stream data of watch and RMSEP
5th, the performance based on OSELM migration models is treated from the height of physical and chemical index
Using tablet NIR data sets, first the spectroscopic data on-line proving measured from instrument SPEC2 is migrated to main instrument
The spectral space of SPEC2, then substitutes into the prediction model on the first active ingredient in tablet using the spectroscopic data after migration
Obtain predicted value.From can be by Figure 12 on the first active ingredient effect to going out from the Forecast of Spectra collected on instrument
Arrive.It is pre- to represent that test data substitution tablet the first active ingredient prediction model of main instrument obtains by SP1_predict in Figure 12
Measured value.TLSP2_predict is represented using the predicted value based on the first active ingredient prediction model after OSELM migration models.
SP2_predict represents that SP2 is directly substituted into without migration and obtains predicted value on the first active ingredient prediction model in tablet.Figure
In it can be seen that have between triangle and plus sige and star it is obvious decompose line, therefore for the calibration migration effect from spectrum still
Compare significant.
That drawn in Figure 13 is SP1_test, TLSP2_test, SP2 active on the second active ingredient in tablet and the 3rd
The predicted value of component.Although the first width subgraph five-pointed star and plus sige do not revert to straight line, this can only illustrate that parameter is not chosen
Or effect of the data set on physics and chemistry sex character prediction model is bad, but still it can be seen that the prediction of plus sige and five-pointed star
It is worth essentially the same, and triangle clearly has obvious segmentation between five-pointed star and plus sige.The performance that second width subgraph is shown
It is similar with the first width subgraph, just no longer discuss herein.
1.3.3 contrast test
Whether it is better than other algorithms, it is necessary to can know by comparing in models established above.Used in contrast test
The PDS algorithms and CCA algorithms introduced in background technology are arrived.It is window size wherein to have hyper parameter in PDS algorithms, different
Window size can produce the effect of algorithm certain influence.Showed below in this table by choosing different window value sizes
Performances of the PDS under different windows.By PDS, CCA, TLOSELM (migration algorithm based on OSELM proposed in the present invention)
Salve spectroscopic datas are migrated, then the model after migration is substituted into the prediction mould on physics and chemistry sex character established online
Type (hyper parameter of the prediction model is selected by cross validation), can predict the corresponding physicochemical characteristics of spectrum after migration
Value.Only water content and protein both physics and chemistry sex character progress root-mean-square error size are contrasted in corn data set.
Then carry out contrasting the root-mean-square error of the predicted value of three kinds of active ingredients in tablet data set.
Table 7 and table 8 represent that distribution represents that three kinds of algorithms predict the contrast of error on corn content and protein content.
Table 9, the distribution of table 10,11 represent that PDS, CCA, TLOSELM are distributed to the pre- of first, second and third active ingredient on tablet data set
Monitoring error.The alphabetical N in all tables represents to be used for the sample number for building migration models above, and RMSEP is represented on physics and chemistry sex character
Root-mean-square error.It can be drawn in corn data set from the data in table 7 and Fig. 9, when M5 spectrum are migrated to MP5 spectrum
Be better than PDS algorithms based on CCA algorithm migration models, but MP6 to MP5 spectrum migrate when PDS effect it is better.Just
For whole corn NIR light spectrum data set, TLOSELM algorithms (the moving based on online order limit learning machine i.e. in the present invention
Move algorithm) performance capabilities is optimal in these three algorithms.Three kinds of algorithms are all as the increase of modeling sample number, TLOSELM are calculated
The RMSEP of method model and CCA algorithm models constantly reduces, but PDS algorithms show unstable one side, just start with
The increase of sample, RMSEP are begun to decline, and as sample number continues to increase, RMSEP starts to show the trend of growth again.Table 9,
10th, 11 can draw, TLOSELM algorithms remain best performance in three kinds of algorithms in tablet data set.Except for first
The prediction CCA of active ingredient is better than PDS, and PDS is then better than CCA algorithms in the prediction on second and third active ingredient.Figure
14 be prediction effect of three kinds of algorithms to corn content, and Figure 15 is prediction of three kinds of algorithms to the 3rd active ingredient of tablet data
As a result.
7 corn data set of table predicts error under algorithms of different to water
Prediction error of the 8 corn data set of table under algorithms of different to protein
The prediction error of 9 tablet data set the first active ingredient of table
The prediction error of 10 tablet data set second active ingredient of table
The prediction error of 11 tablet data set the third active ingredient of table
Draw a conclusion from experiment above:On corn and tablet data set, TLOSELM algorithms of the invention compared to
PDS algorithms and CCA algorithms are possessing better performance from spectrum into principal spectrum transition process.Work as use in corn data set
MP6 be used as from spectrum MP5 as principal spectrum when, although it is optimal to remain TLOSELM, but it is found that CCA algorithms are not
Have more excellent than PDS algorithm too many.And tablet integrate or M5 be used as from spectrum MP5 as principal spectrum when, CCA and TLOSELM compared to
PDS has very big advantage, and the TLOSELM algorithm performances of the present invention are optimal.
Claims (10)
1. a kind of sample composition content assaying method based on online order limit learning machine, it is characterised in that gather sample
Spectroscopic data sample, and be modeled using online order limit learning machine algorithm;Utilize established model to sample into
Point content is measured.
2. the sample composition content assaying method according to claim 1 based on online order limit learning machine, its feature
It is, specifically includes following steps:
S1, according to initial principal spectrum SPmaster(0)And corresponding sample composition content y0With hidden node number L, hidden layer is calculated
To the initial weight matrix α of output layer(0), wherein, SPmaster(0)And y0Include M0A sample;
S2, when there is new principal spectrum SPmaster(k+1)And corresponding sample composition content yk+1During arrival, according to online order limit
Learning machine algorithm calculates hidden layer to the weight matrix α of output layer(k+1);Wherein, the data SP of+1 arrival of kthmaster(k+1)With
yk+1Include Mk+1A sample;k≥0;
S3, if also there are new principal spectrum SPmaster(k+1)With corresponding sample composition content yk+1, then k=k+1 is made, and go to
S2, otherwise goes to S4;
S4, the spectroscopic data sp for obtaining sample is calculated according to the following formulamasterPreCorresponding component content predicted value pre_y:
Pre_y=Hpreα(new)
Wherein,
α(new)For current newest hidden layer to the weight matrix of output layer, spmasterPreInclude N number of sample;W and b are respectively random
The orthogonal input weight matrix of generation and biasing;G(w,spmasterPre, b) and it is activation primitive.
3. the sample composition content assaying method according to claim 1 based on online order limit learning machine, its feature
It is, further includes:Utilize online order limit learning machine algorithm, the spectrum of the sample gathered to key light spectrometer and from spectrometer
Data sample is modeled, and realizing will migrate to the spectroscopic data space of key light spectrometer from the spectroscopic data of spectrometer;Then it is sharp
The content prediction model established with key light spectrometer carries out the measure of sample composition content.
4. the sample composition content assaying method according to claim 3 based on online order limit learning machine, its feature
It is, the online order limit learning machine algorithm of the utilization, the spectrum of the sample gathered to key light spectrometer and from spectrometer
Data sample is modeled, and realizing will migrate to the spectroscopic data space of key light spectrometer from the spectroscopic data of spectrometer including following
Step:
S01, according to initial principal spectrum SPmaster0With from spectrum spslave0And the number of hidden nodes L, the power of generation hidden layer to output layer
Weight matrix β(0), wherein SPmaster0In the number of samples that includes be M0;
S02, new M is included when havingk+1The SP of a samplemaster(k+1)And spslave(k+1)During arrival, according to online order limit
Habit machine algorithm calculates hidden layer to the weight matrix β of output layer(k+1);Wherein, k >=0;
S03, if the SP of also new samplemaster(k+1)And spslave(k+1)Arrive, then make k=k+1, and go to S02, otherwise
Go to S04;
S04, according to the following formula to the test data sp comprising N number of sampleslaveTestMigrated:
spslaveTomaster=H'testβnew
Wherein, spslaveTomasterRepresent the spectroscopic data after migration;βnewRepresent newest hidden layer to the weight square of output interlayer
Battle array;W and b is respectively random raw
Into orthogonal input weight matrix and biasing;G(w,spslaveTest, b) and it is activation primitive.
5. the sample composition content assaying method based on online order limit learning machine according to claim 2 or 4, it is special
Sign is that the hidden node number L is less than or equal to initial number of samples M0。
6. the sample composition content assaying method according to claim 5 based on online order limit learning machine, its feature
It is, in step S1,
α(0)=(H0 TH0)-1H0 Ty0;
Wherein,
7. the sample composition content assaying method according to claim 5 based on online order limit learning machine, its feature
It is, in step S2,
<mrow>
<msup>
<mi>&alpha;</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</msup>
<mo>=</mo>
<msup>
<mi>&alpha;</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>+</mo>
<msub>
<mi>q</mi>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<msubsup>
<mi>H</mi>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
<mi>T</mi>
</msubsup>
<mrow>
<mo>(</mo>
<msub>
<mi>y</mi>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>H</mi>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<msup>
<mi>&alpha;</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
</mrow>
Wherein,
<mrow>
<msub>
<mi>H</mi>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>=</mo>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<mrow>
<mi>G</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>sp</mi>
<mrow>
<mi>m</mi>
<mi>a</mi>
<mi>s</mi>
<mi>t</mi>
<mi>e</mi>
<mi>r</mi>
<mo>_</mo>
<mrow>
<mo>(</mo>
<msubsup>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mi>k</mi>
</msubsup>
<msub>
<mi>M</mi>
<mi>j</mi>
</msub>
<mo>+</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mi>w</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<msub>
<mi>b</mi>
<mn>1</mn>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<mrow>
<mi>G</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>sp</mi>
<mrow>
<mi>m</mi>
<mi>a</mi>
<mi>s</mi>
<mi>t</mi>
<mi>e</mi>
<mi>r</mi>
<mo>_</mo>
<mrow>
<mo>(</mo>
<msubsup>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mi>k</mi>
</msubsup>
<msub>
<mi>M</mi>
<mi>j</mi>
</msub>
<mo>+</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mi>w</mi>
<mi>L</mi>
</msub>
<mo>,</mo>
<msub>
<mi>b</mi>
<mi>L</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mi>G</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>sp</mi>
<mrow>
<mi>m</mi>
<mi>a</mi>
<mi>s</mi>
<mi>t</mi>
<mi>e</mi>
<mi>r</mi>
<mo>_</mo>
<mrow>
<mo>(</mo>
<msubsup>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msubsup>
<msub>
<mi>M</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mi>w</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<msub>
<mi>b</mi>
<mn>1</mn>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<mrow>
<mi>G</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>sp</mi>
<mrow>
<mi>m</mi>
<mi>a</mi>
<mi>s</mi>
<mi>t</mi>
<mi>e</mi>
<mi>r</mi>
<mo>_</mo>
<mrow>
<mo>(</mo>
<msubsup>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msubsup>
<msub>
<mi>M</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mi>w</mi>
<mi>L</mi>
</msub>
<mo>,</mo>
<msub>
<mi>b</mi>
<mi>L</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>.</mo>
</mrow>
8. the sample composition content assaying method according to claim 5 based on online order limit learning machine, its feature
It is, in step S01,
<mrow>
<msup>
<mi>&beta;</mi>
<mrow>
<mo>(</mo>
<mn>0</mn>
<mo>)</mo>
</mrow>
</msup>
<mo>=</mo>
<msup>
<mrow>
<mo>(</mo>
<msup>
<mi>H</mi>
<mo>&prime;</mo>
</msup>
<msup>
<msub>
<mrow></mrow>
<mn>0</mn>
</msub>
<mi>T</mi>
</msup>
<msub>
<msup>
<mi>H</mi>
<mo>&prime;</mo>
</msup>
<mn>0</mn>
</msub>
<mo>)</mo>
</mrow>
<mrow>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msup>
<mo>-</mo>
<msup>
<mi>H</mi>
<mo>&prime;</mo>
</msup>
<msup>
<msub>
<mrow></mrow>
<mn>0</mn>
</msub>
<mi>T</mi>
</msup>
<msub>
<mi>sp</mi>
<mrow>
<mi>m</mi>
<mi>a</mi>
<mi>s</mi>
<mi>t</mi>
<mi>e</mi>
<mi>r</mi>
<mn>0</mn>
</mrow>
</msub>
<mo>;</mo>
</mrow>
Wherein,
9. the sample composition content assaying method according to claim 5 based on online order limit learning machine, its feature
It is, in step S02,
<mrow>
<msup>
<mi>&beta;</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</msup>
<mo>=</mo>
<msup>
<mi>&beta;</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>+</mo>
<msub>
<mi>P</mi>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<msubsup>
<msup>
<mi>H</mi>
<mo>&prime;</mo>
</msup>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
<mi>T</mi>
</msubsup>
<mrow>
<mo>(</mo>
<msub>
<mi>sp</mi>
<mrow>
<mi>m</mi>
<mi>a</mi>
<mi>s</mi>
<mi>t</mi>
<mi>e</mi>
<mi>r</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
</msub>
<mo>-</mo>
<msub>
<msup>
<mi>H</mi>
<mo>&prime;</mo>
</msup>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<msup>
<mi>&beta;</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
</mrow>
Wherein,
<mrow>
<msub>
<msup>
<mi>H</mi>
<mo>&prime;</mo>
</msup>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>=</mo>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<mrow>
<mi>G</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>sp</mi>
<mrow>
<mi>s</mi>
<mi>l</mi>
<mi>a</mi>
<mi>v</mi>
<mi>e</mi>
<mo>_</mo>
<mrow>
<mo>(</mo>
<msubsup>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mi>k</mi>
</msubsup>
<msub>
<mi>M</mi>
<mi>j</mi>
</msub>
<mo>+</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mi>w</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<msub>
<mi>b</mi>
<mn>1</mn>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<mrow>
<mi>G</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>sp</mi>
<mrow>
<mi>s</mi>
<mi>l</mi>
<mi>a</mi>
<mi>v</mi>
<mi>e</mi>
<mo>_</mo>
<mrow>
<mo>(</mo>
<msubsup>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mi>k</mi>
</msubsup>
<msub>
<mi>M</mi>
<mi>j</mi>
</msub>
<mo>+</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mi>w</mi>
<mi>L</mi>
</msub>
<mo>,</mo>
<msub>
<mi>b</mi>
<mi>L</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mi>G</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>sp</mi>
<mrow>
<mi>s</mi>
<mi>l</mi>
<mi>a</mi>
<mi>v</mi>
<mi>e</mi>
<mo>_</mo>
<mrow>
<mo>(</mo>
<msubsup>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msubsup>
<msub>
<mi>M</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mi>w</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<msub>
<mi>b</mi>
<mn>1</mn>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<mrow>
<mi>G</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>sp</mi>
<mrow>
<mi>s</mi>
<mi>l</mi>
<mi>a</mi>
<mi>v</mi>
<mi>e</mi>
<mo>_</mo>
<mrow>
<mo>(</mo>
<msubsup>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mrow>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msubsup>
<msub>
<mi>M</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mi>w</mi>
<mi>L</mi>
</msub>
<mo>,</mo>
<msub>
<mi>b</mi>
<mi>L</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>.</mo>
</mrow>
10. the sample composition content assaying method based on online order limit learning machine according to claim 2 or 4, its
It is characterized in that, the method that cross validation is rolled over by k determines optimal the number of hidden nodes L;The activation primitive uses sigmoid letters
Number.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711068234.XA CN107918718B (en) | 2017-11-03 | 2017-11-03 | Sample component content determination method based on online sequential extreme learning machine |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711068234.XA CN107918718B (en) | 2017-11-03 | 2017-11-03 | Sample component content determination method based on online sequential extreme learning machine |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107918718A true CN107918718A (en) | 2018-04-17 |
CN107918718B CN107918718B (en) | 2020-05-22 |
Family
ID=61896071
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711068234.XA Active CN107918718B (en) | 2017-11-03 | 2017-11-03 | Sample component content determination method based on online sequential extreme learning machine |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107918718B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10515715B1 (en) | 2019-06-25 | 2019-12-24 | Colgate-Palmolive Company | Systems and methods for evaluating compositions |
CN112414966A (en) * | 2019-08-21 | 2021-02-26 | 东北大学秦皇岛分校 | Near infrared spectrum multi-target calibration migration method based on affine change |
CN112834546A (en) * | 2020-12-01 | 2021-05-25 | 上海纽迈电子科技有限公司 | Method for testing water content and oil content in plant grains and application thereof |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105095652A (en) * | 2015-07-10 | 2015-11-25 | 东北大学 | Method for testing component in sample based on stacking extreme learning machine |
CN106596450A (en) * | 2017-01-06 | 2017-04-26 | 东北大学秦皇岛分校 | Incremental method for analysis of material component content based on infrared spectroscopy |
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 |
CN106803124A (en) * | 2017-01-21 | 2017-06-06 | 中国海洋大学 | Field migration extreme learning machine method based on manifold canonical and norm canonical |
US20170257452A1 (en) * | 2016-03-02 | 2017-09-07 | Huawei Technologies Canada Co., Ltd. | Systems and methods for data caching in a communications network |
-
2017
- 2017-11-03 CN CN201711068234.XA patent/CN107918718B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105095652A (en) * | 2015-07-10 | 2015-11-25 | 东北大学 | Method for testing component in sample based on stacking extreme learning machine |
US20170257452A1 (en) * | 2016-03-02 | 2017-09-07 | Huawei Technologies Canada Co., Ltd. | Systems and methods for data caching in a communications network |
CN106596450A (en) * | 2017-01-06 | 2017-04-26 | 东北大学秦皇岛分校 | Incremental method for analysis of material component content based on infrared spectroscopy |
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 |
CN106803124A (en) * | 2017-01-21 | 2017-06-06 | 中国海洋大学 | Field migration extreme learning machine method based on manifold canonical and norm canonical |
Non-Patent Citations (2)
Title |
---|
ADITIE GARG等: "Fault Classification, Location in a Series Compensated Power Transmission Network using Online Sequential Extreme Learning Machine", 《IEEE》 * |
刘月: "基于NIR光谱的半监督在线序列ELM回归算法研究", 《中国优秀硕士学位论文全文数据库 基础科学辑》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10515715B1 (en) | 2019-06-25 | 2019-12-24 | Colgate-Palmolive Company | Systems and methods for evaluating compositions |
US10839942B1 (en) | 2019-06-25 | 2020-11-17 | Colgate-Palmolive Company | Systems and methods for preparing a product |
US10839941B1 (en) | 2019-06-25 | 2020-11-17 | Colgate-Palmolive Company | Systems and methods for evaluating compositions |
US10861588B1 (en) | 2019-06-25 | 2020-12-08 | Colgate-Palmolive Company | Systems and methods for preparing compositions |
US11315663B2 (en) | 2019-06-25 | 2022-04-26 | Colgate-Palmolive Company | Systems and methods for producing personal care products |
US11342049B2 (en) | 2019-06-25 | 2022-05-24 | Colgate-Palmolive Company | Systems and methods for preparing a product |
US11728012B2 (en) | 2019-06-25 | 2023-08-15 | Colgate-Palmolive Company | Systems and methods for preparing a product |
CN112414966A (en) * | 2019-08-21 | 2021-02-26 | 东北大学秦皇岛分校 | Near infrared spectrum multi-target calibration migration method based on affine change |
CN112414966B (en) * | 2019-08-21 | 2022-06-10 | 东北大学秦皇岛分校 | Near infrared spectrum multi-target calibration migration method based on affine change |
CN112834546A (en) * | 2020-12-01 | 2021-05-25 | 上海纽迈电子科技有限公司 | Method for testing water content and oil content in plant grains and application thereof |
Also Published As
Publication number | Publication date |
---|---|
CN107918718B (en) | 2020-05-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104166731B (en) | A kind of overlapping community discovery system and method for social networks | |
CN106815643B (en) | Infrared spectroscopy Model Transfer method based on random forest transfer learning | |
CN103854305B (en) | A kind of Model Transfer method based on multi-scale Modeling | |
CN107918718A (en) | Sample composition content assaying method based on online order limit learning machine | |
CN108152239A (en) | The sample composition content assaying method of feature based migration | |
CN106596450B (en) | Incremental method based on infrared spectrum analysis material component content | |
US20230349818A1 (en) | Cross-validation based calibration of a spectroscopic model | |
CN106680238B (en) | Method based on infrared spectrum analysis material component content | |
CN107563596A (en) | A kind of evaluation index equilibrium state analysis method based on Bayes's causal network | |
JP7063389B2 (en) | Processing equipment, processing methods, and programs | |
WO2018096683A1 (en) | Factor analysis method, factor analysis device, and factor analysis program | |
CN111859249B (en) | Ocean numerical forecasting method based on analytical four-dimensional set variation | |
CN108681697A (en) | Feature selection approach and device | |
Karaman et al. | Sparse multi-block PLSR for biomarker discovery when integrating data from LC–MS and NMR metabolomics | |
CN111144017A (en) | FF-RVM-based multi-period intermittent process soft measurement modeling method | |
CN112287601B (en) | Method, medium and application for constructing tobacco leaf quality prediction model by using R language | |
CN102663495A (en) | Neural net data generation method for nonlinear device modeling | |
Jiang et al. | Qualitative and quantitative analysis in solid-state fermentation of protein feed by FT-NIR spectroscopy integrated with multivariate data analysis | |
CN110070004B (en) | Near-earth hyperspectral data expansion method applied to deep learning | |
CN105956605A (en) | Three-dimensional structure similarity clustering method based on parallel k-means clustering | |
CN107367467A (en) | A kind of content of material quantitative analysis method | |
Eberhardt et al. | Modeling technology and technological change in manufacturing: how do countries differ? | |
CN110231653B (en) | Method and device for constructing bidirectional constraint initial model | |
CN104008493A (en) | Data acquisition method and device | |
Neumeyer | Testing independence in nonparametric regression |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |