CN103776797B - A kind of near infrared spectrum differentiates the method for flat interest Herb Gynostemmae Pentaphylli - Google Patents

Abstract

The present invention provide a kind of near infrared spectrum differentiate flat interest Herb Gynostemmae Pentaphylli method, comprise the steps: A, set up flat interest Herb Gynostemmae Pentaphylli near infrared spectrum differentiate model: A 1, select spectral region 4000 12500cm^‑1, scan flat interest Herb Gynostemmae Pentaphylli near infrared light spectrogram；A 2, to spectral region 4000 9500cm^‑1Data carry out pretreatment；A 3, extraction main constituent；A 4, set up artificial nerve network model: take artificial neural network algorithm, determine that according to inputoutput data feature the structure of neutral net, recycling training data train this neutral net；MATLAB software is used to set up the BP artificial nerve network model of input layer 10 hidden layer node 5 output layer node 2；B, the discriminating of unknown sample: unknown sample scans near infrared light spectrogram under the same conditions, choose main constituent number, according to the neural network model trained to judge the true and false of unknown sample, output node represents with binary code respectively, 10 representatives are flat interest Herb Gynostemmae Pentaphylli, and 01 representative is non-flat interest Herb Gynostemmae Pentaphylli.

Description

A kind of near infrared spectrum differentiates the method for flat interest Herb Gynostemmae Pentaphylli

Technical field

The present invention relates to a kind of method that near infrared spectrum differentiates flat interest Herb Gynostemmae Pentaphylli, particularly relate to one Kind near-infrared spectrum technique combines the method that artificial neural network algorithm differentiates flat interest Herb Gynostemmae Pentaphylli, Belong near infrared spectrum detection analysis field.

Background technology

Herb Gynostemmae Pentaphylli, also known as Herba Gynostemmatis and gynostemma pentaphyllum makino etc., has a blood pressure lowering, blood fat reducing, blood sugar lowering, Protect the heart to protect the liver, regulating lipid and reducing weight, have the appellation of " elixir grass ".From the quality inspection of 2004 countries Since Shaanxi flat interest Herb Gynostemmae Pentaphylli is implemented the protection of region, original producton location by general bureau, the personal value of flat interest Herb Gynostemmae Pentaphylli Multiplication, adulterates in market, fake and forged phenomenon happens occasionally, for effectively identifying that difference is produced The flat interest Herb Gynostemmae Pentaphylli on ground, the rights and interests of protection consumer, set up efficient Herb Gynostemmae Pentaphylli Production area recognition technology Imperative.

Near infrared spectrum has that response speed is fast, informative, pretreatment is few, do not pollute ring The advantages such as border, are used widely in a lot of fields, become and study the most popular spectrum at present One of analytical technology.Near infrared spectrum contains the bulk information of sample, therefore, by near-infrared Analytical technology combines with mode identification method, more effectively sample can be carried out grade and class The differentiation belonged to.Near-infrared mode identification technology is that the method for applied chemistry pattern recognition is from material Near-infrared data deduce the technology that material belongs to.All methods of Chemical Pattern Recognition can be used in The research of near-infrared pattern recognition.At present, based near infrared mode identification technology by extensively It is applied to the fields such as agricultural, medicine, food, oil, in true and false differentiation, grade separation, former The aspects such as place of production qualification played an important role.But the identification model that pattern recognition is set up is all For specific products, specificity is stronger.Applicant has used near infrared spectroscopy to combine horse Family name's distance algorithm, and qualification testing effectively authenticated Xiangshui County's rice；Utilize fi sher to differentiate to calculate Method successfully differentiates virgin oil and olea europaea fruit residual oil.The present invention is based on near-infrared spectrum technique Flat interest Herb Gynostemmae Pentaphylli is differentiated in conjunction with artificial neural network algorithm.At present, Chinese scholars Gynostemma pentaphyllum Makino Research be concentrated mainly in chemical composition and its pharmacological action and study.It mainly contains soap Glycosides [1], polysaccharide [2] and aminoacid [4], flavonoid [3], organic acid and trace element [4] etc. are many Plant chemical composition.These reports confirm that the composition of different sources Herb Gynostemmae Pentaphylli there are differences, because of These these methods have a some reference value for the Herb Gynostemmae Pentaphylli true and false differentiating different sources, but at present Have no the report discerned the false from the genuine utilizing its component difference to carry out flat interest Herb Gynostemmae Pentaphylli.

Summary of the invention

It is an object of the invention to provide one and it is an object of the invention to provide a kind of quick, accurate True combines the artificial neural network algorithm discriminating flat interest Herb Gynostemmae Pentaphylli true and false with near-infrared spectrum technique Method.

The technical scheme is that the method that this near infrared spectrum differentiates flat interest Herb Gynostemmae Pentaphylli, Comprise the steps:

A, set up flat interest Herb Gynostemmae Pentaphylli near infrared spectrum differentiate model

A-1, selection spectral region 4000-12500cm^-1, scan flat interest Herb Gynostemmae Pentaphylli near infrared spectrum Figure；

A-2, to spectral region 4000-12500cm^-1Data carry out pretreatment；

A-3, extraction main constituent；

A-4, set up artificial nerve network model: take artificial neural network algorithm, according to input Output data characteristics determines that the structure of neutral net, recycling training data train this neutral net, Obtain the discriminating model of flat interest Herb Gynostemmae Pentaphylli；

B, the discriminating of unknown sample

Unknown sample scans near infrared light spectrogram under the same conditions, chooses main constituent number, depends on According to the neural network model trained to judge the true and false of unknown sample, output node is respectively with two Carry system code represents, 10 representatives are flat interest Herb Gynostemmae Pentaphylli, and 01 representative is non-flat interest Herb Gynostemmae Pentaphylli.

Described near infrared spectrum differentiates the method for flat interest Herb Gynostemmae Pentaphylli, the scanning described in step A-1 Flat interest Herb Gynostemmae Pentaphylli near infrared light spectrogram includes: pulverized by the flat interest Herb Gynostemmae Pentaphylli sample drying of effective dose After be homogeneously disposed in quartz sample pool, use Fourier near infrared spectrometer to carry out absorption spectrum and sweep Retouch；Scan pattern is for rotating diffuse-reflectance, and resolution is 8cm^-1, each scan sample repeatedly, takes Averaged spectrum is the analysis spectrum that sample is final；

Described near infrared spectrum differentiates the method for flat interest Herb Gynostemmae Pentaphylli, the flat interest described in step A-2 Herb Gynostemmae Pentaphylli near infrared light spectral data pretreatment includes: Gynostemma pentaphyllum Makino sample spectrum carries out polynary dissipating Penetrate correction+appropriate normalized pretreatment, eliminate that sample is uneven by this process, light scattering and The impact of the interference factors such as noise of instrument, improves precision of prediction and the stability of model.

Described near infrared spectrum differentiates the method for flat interest Herb Gynostemmae Pentaphylli, the extraction described in step A-3 Main component is, by principal component analytical method, spectrogram information is carried out dimensionality reduction, takes front 10 masters Composition contribution rate of accumulative total is 99.99%, and limited amount input reduces the computation complexity of model, carries The precision of prediction of high model.

Described near infrared spectrum differentiates the method for flat interest Herb Gynostemmae Pentaphylli, the flat interest described in step A-4 Herb Gynostemmae Pentaphylli is set up artificial nerve network model and includes using MATLAB software to set up input layer The BP artificial nerve network model of 10-hidden layer node 5-output layer node 2:

A-4-1, the determination of input layer number: taking 10 principal component scores is parameter, really The input layer determining network is 10；

A-4-2, the determination of the number of hidden nodes: obtain with lower formula:

$L=\sqrt{0.43mn+0.12n^2+2.54m+0.77n+0.35}+0.51$

Wherein m, n are respectively input node and output node number, and hidden node number can be by Formula obtains an initial value, then utilizes progressively growth method correction, obtains empirical value 5；

A-4-3, the determination of output layer nodes: according to judging that Herb Gynostemmae Pentaphylli sample is belonging to flat interest and produces Still fall within two kinds of non-flat interest place of production result, determine that the output node of neutral net is 2.

Described near infrared spectrum differentiates the method for flat interest Herb Gynostemmae Pentaphylli, the main constituent of described sample by Following method determines:

If x₁, x₂..., x_nFor taking from the sample of overall x, wherein x_i=(x_i1, x_i2..., x_ip) ' (i=1, 2 ... n)；

Note sample observations matrix is:

$X=\left[\begin{array}{cccc}{x}_{11}& x_{12}& \mathrm{...}& x_{1p}\\ x_{21}& x_{22}& \mathrm{...}& x_{2p}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ x_{n1}& x_{n2}& \mathrm{...}& x_{np}\end{array}\right]$

The corresponding sample of every a line of x, the corresponding variable of every string；

Note sample covariance matrix and sample correlation coefficient matrix are respectively as follows:

$S=\frac{1}{n-1}\underset{i=1}{\overset{n}{\Σ}}(x_i-\stackrel{\‾}{x}){(x_i-\stackrel{\‾}{x})}^{\′}=\left(s_{ij}\right)$

$\hat{R}=\left(r_{ij}\right),r_{ij}=\frac{s_{ij}}{\sqrt{s_{ii}}\sqrt{s_{jj}}}$

Wherein,For sample mean；

Using S as the estimation of Σ,As the estimation of R, from S orSet out and can try to achieve the main one-tenth of sample Point.

Described near infrared spectrum differentiates the method for flat interest Herb Gynostemmae Pentaphylli, it is characterised in that: described sample This main constituent is by from sample correlation coefficient matrixSet out and solve:

IfForP eigenvalue,For corresponding positive presentate Position characteristic vector, then p main constituent of sample is

${\hat{y}}^{*}={\hat{t}}_i{{}^{*}}^{\′}{x}^{*},i=1,2,...,p$

By sample x_iObservation after standardizationSubstitute into jth main constituent, i.e. can get sample x_i? J principal component scores

${\hat{y}}_{ij}^{*}={{\hat{t}}_j^{*}}^{\′}{x}_i^{*},(i=1,2,...,n;j=1,2,...,p).$

The present invention selects spectral region 4000-9500cm^-1, according to original spectrum analysis, this district Between the characteristic peak containing the main near infrared absorption of Herb Gynostemmae Pentaphylli sample, Gynostemma pentaphyllum Makino sample spectrum Carry out the normalized pretreatment of multiplicative scatter correction+vector, eliminate that sample is uneven, light scattering and The impact of the interference factors such as noise of instrument, improves precision of prediction and the stability of neural network model； The present invention extracts the principal component scores of front 10 main constituents and inputs as new variables, limited amount defeated Enter to reduce the computation complexity of model, improve the precision of prediction of model；Can quickly realize mirror The true and false of other flat interest Herb Gynostemmae Pentaphylli, model training collection and forecast set differentiate that accuracy is 100%.Accordingly The discrimination model set up is significant for the genuine/counterfeit discriminating realizing flat interest Herb Gynostemmae Pentaphylli.

Accompanying drawing explanation

Fig. 1 neural network structure figure

Each letter representation in figure:

x_jThe input of expression input layer jth node, j=1 ..., M；

w_ijRepresent that hidden layer i-th node is to the weights between input layer jth node；

θ_iRepresent the threshold value of hidden layer i-th node；

Φ_xRepresent the excitation function of hidden layer；

w_kjRepresent output layer kth node to the weights between hidden layer i-th node, I=1 ..., q；

a_kThe threshold value of expression output layer kth node, k=1 ..., L；

Ψ_xRepresent the excitation function of output layer；

O_kRepresent the output of output layer kth node.

Fig. 2 flat interest Herb Gynostemmae Pentaphylli near infrared light spectrogram

The accumulation contribution rate of front ten principal component scores of Fig. 3

Fig. 4 is the programme diagram that neural computing realizes

Detailed description of the invention

The present invention is further detailed explanation with detailed description of the invention below in conjunction with the accompanying drawings.

1, instrument and reagent: use the MPA near infrared spectrometer of Brooker company of Germany and overflow Reflection accessory gathers sample spectra, and instrument carries analysis software and MATLAB software.

1-1, noise of instrument

Configuration regulated power supply, it is substantially stabilized that start is preheated to instrument, it is ensured that suitably test environment Temperature 15-25 DEG C；

1-2, wavelength accuracy and repeatability

Accurate with low pressure mercury lamp and methylene blue solution tuning wavelength that mass fraction is 0.005% Property, prevent drift.

2, sample preparation and spectral scan: Herb Gynostemmae Pentaphylli sample is the most at source bought.

Table 1 Herb Gynostemmae Pentaphylli sample message table

3, the near infrared spectrum setting up geographical sign Herb Gynostemmae Pentaphylli differentiates model

3-1, scanning Herb Gynostemmae Pentaphylli near infrared light spectrogram

Flat interest Herb Gynostemmae Pentaphylli sample is dried 4 hours at 60 DEG C, pulverizes, and crosses 60 mesh sieves, takes about 50g sample is homogeneously disposed in sample cell, uses the MPA near infrared spectrum of Brooker company of Germany Instrument and diffuse-reflectance adnexa gather the rotation diffuse-reflectance spectrum of sample, as shown in Figure 2.Light source for instrument For 20W tungsten sodium lamp, spectral region is 4000-12500cm^-1.The spectra collection software used is cloth The OPUS 6.5 of Lu Ke company, scanning times 64, resolution is 8cm^-1, with the built-in back of the body of instrument Scape is as reference.Each scan sample 3 times, the averaged spectrum taking 3 times is sample spectrum.

3-2, spectroscopic data is carried out pretreatment

At wave band 4000-9500cm^-1On carry out the foundation of discrimination model, with MATLAB software pair Sample spectra processes, and uses Ricoh at multiplicative scatter correction+vector normalization preprocess method Compose and use principal component analytical method Gynostemma pentaphyllum Makino sample spectra to carry out dimensionality reduction.

3-3, extraction main constituent；

Described near infrared spectrum differentiates the method for flat interest Herb Gynostemmae Pentaphylli, the main constituent of described sample by Following method determines:

If x₁, x₂..., x_nFor taking from the sample of overall x, wherein x_i=(x_i1, x_i2..., x_ip) ' (i=1, 2 ... n)；

Note sample observations matrix is:

$X=\left[\begin{array}{cccc}{x}_{11}& x_{12}& \mathrm{...}& x_{1p}\\ x_{21}& x_{22}& \mathrm{...}& x_{2p}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ x_{n1}& x_{n2}& \mathrm{...}& x_{np}\end{array}\right]$

The corresponding sample of every a line of x, the corresponding variable of every string；

Note sample covariance matrix and sample correlation coefficient matrix are respectively as follows:

$S=\frac{1}{n-1}\underset{i=1}{\overset{n}{\Σ}}(x_i-\stackrel{\‾}{x}){(x_i-\stackrel{\‾}{x})}^{\′}=\left(s_{ij}\right)$

$\hat{R}=\left(r_{ij}\right),r_{ij}=\frac{s_{ij}}{\sqrt{s_{ii}}\sqrt{s_{jj}}}$

Wherein,For sample mean；

Using S as the estimation of Σ,As the estimation of R, from S orSet out and can try to achieve the main one-tenth of sample Point.

Described near infrared spectrum differentiates the method for flat interest Herb Gynostemmae Pentaphylli, it is characterised in that: described sample This main constituent is by from sample correlation coefficient matrixSet out and solve:

IfForP eigenvalue,For corresponding positive presentate Position characteristic vector, then p main constituent of sample is

${\hat{y}}^{*}={\hat{t}}_i{{}^{*}}^{\′}{x}^{*},i=1,2,...,p$

By sample x_iObservation after standardizationSubstitute into jth main constituent, i.e. can get sample x_i? J principal component scores

${\hat{y}}_{ij}^{*}={{\hat{t}}_j^{*}}^{\′}{x}_i^{*},(i=1,2,...,n;j=1,2,...,p).$

MATLAB software is used to take front 10 principal component scores input layer as network, In Fig. 3, front ten main constituent accumulation contribution rates are to 99.99%, as shown in table 2.Can represent Most effective informations of Herb Gynostemmae Pentaphylli, so using the score of front ten main constituents as input Node layer parameter.

2 front ten main constituent accumulation contribution rates of table

The output point of network then uses the output in the binary code representation place of production, as shown in table 3 eight The sample in the place of production is only divided into two big classes, and one is original producton location flat interest Herb Gynostemmae Pentaphylli, and it uses binary code 10 represent；Another kind of is non-original producton location Herb Gynostemmae Pentaphylli, represents with binary code 01.

Table 3 output node place of production code

3-4, the foundation of artificial nerve network model

3-4-1, the determination of input layer number: taking 10 principal component scores is parameter, really The input layer determining network is 10, and sample is divided into again training set and forecast set, and training set is to use Train neutral net, set up the artificial nerve network model of the present invention.Forecast set is used for verifying The accuracy of inventive network model, if it find that the rate of accuracy reached of this method is less than requiring, The most later go training pattern, namely those parameters of re-optimization, until setting up neural accurately Network model.

The present embodiment gathers 405 samples altogether, and the flat interest place of production gathers 90 samples, other places of production 405 samples are randomly divided into training set and forecast set two parts by 45 samples of each collection, its Middle training set sample number is 270, it was predicted that collection sample number is 135.

3-4-2, the determination of the number of hidden nodes: obtain with lower formula:

$L=\sqrt{0.43mn+0.12n^2+2.54m+0.77n+0.35}+0.51$

Wherein m, n are respectively input node and output node number, and hidden node number can be by Formula obtains an initial value, then utilizes progressively growth method correction, obtains empirical value 5；

3-4-3, the determination of output layer nodes: according to judging that Herb Gynostemmae Pentaphylli sample is belonging to flat interest and produces Still fall within two kinds of non-flat interest place of production result, determine that the output node of neutral net is 2.

According to calculation procedure shown in Fig. 4, determine the transmission function of network input layer and hidden layer Tansig, output layer transmission function is traingdx function, and training objective is set to 1x10^-6, net The learning rate of network is 0.05, and the training iterations set is as 1000 times, and the number of hidden nodes is 5.MATLAB software is used to set up input layer 10-hidden layer node 5-output layer The BP artificial nerve network model of node 2.

4, unknown sample differentiates

4-1, sample treatment: take about 50g unknown sample and dry 4 hours at 60 DEG C, pulverize, Cross 60 mesh sieves, be homogeneously disposed in sample cell, gather the rotation diffuse-reflectance spectrum of sample.Spectrum model Enclose for 4000-12500cm^-1, scanning times 64, resolution is 8cm^-1, with the built-in back of the body of instrument Scape is as reference.Each scan sample 3 times, the averaged spectrum taking 3 times is sample spectrum.

4-2, neural network recognization model

Choose unknown sample spectral region 4000-9500cm^-1Interior near infrared light spectrogram, uses polynary Scatter correction+vector normalization preprocess method processes, and uses MATLAB software to extract first 10 The score input discrimination model of main constituent, if gained model output result 10 is judged as flat interest Herb Gynostemmae Pentaphylli, if being output as 01, is judged as non-flat interest Herb Gynostemmae Pentaphylli.Result shows, 270 training The recognition correct rate of sample is 100%, and the recognition correct rate of 135 prediction samples is 100%, tool The result of body is as shown in table 4.

Table 4 artificial neural network judged result to flat interest Herb Gynostemmae Pentaphylli identification feasibility

Foregoing description is only used as near infrared spectrum of the present invention and differentiates that the method for flat interest Herb Gynostemmae Pentaphylli can be implemented Technical scheme propose, not as the single restrictive condition to its technical scheme itself.

List of references:

[1] this length of bamboo pine, the saponin component of the composition Study Herb Gynostemmae Pentaphylli of cucurbitaceous plant. external Medical science Chinese medicine fascicle volume 7 in April, 1985 the 5th phase

[2] Wang Zhaojing, sieve mountain peak brightness. alkali proposes the research of Herb Gynostemmae Pentaphylli water soluble polysaccharide. food research with open Send out volume 27 in May, 2006 the 5th phase

[3] Wang Qinghao, Zhang Xionglu. refining of macroporous adsorbent resin Gynostemma pentaphyllum Makino flavone compound Technical study [J] forest chemical engineering communication volume 39 in June, 2005 the 6th phase

[4] Deng Shilin, Zhou Xiaojuan. aminoacid, vitamin and multiple chemical element in Herb Gynostemmae Pentaphylli Analyze [J]. Hunan Medical University's journal volume 19 in June, 1994 the 6th phase

Claims (2)

1. the method that near infrared spectrum differentiates flat interest Herb Gynostemmae Pentaphylli, its feature comprises the steps:

A, set up flat interest Herb Gynostemmae Pentaphylli near infrared spectrum differentiate model

A-1, selection spectral region 4000-12500cm^-1, scanning flat interest Herb Gynostemmae Pentaphylli near infrared light spectrogram: It is homogeneously disposed in quartz sample pool after the flat interest Herb Gynostemmae Pentaphylli sample drying of effective dose is pulverized, uses Fu Vertical leaf near infrared spectrometer carries out absorption spectrum scanning；Scan pattern is for rotating diffuse-reflectance, resolution For 8cm^-1, repeatedly, be averaged spectrum is the final analysis spectrum of sample to each scan sample；

A-2, chosen spectrum scope 4000-9500cm^-1Data carry out pretreatment: Gynostemma pentaphyllum Makino sample This spectrum carries out multiplicative scatter correction+appropriate normalized pretreatment, eliminates sample by this process Uneven, light scattering and the impact of noise of instrument interference factor, improve the precision of prediction of model and steady Qualitative；

A-3, extraction main constituent: spectrogram information is carried out dimensionality reduction by principal component analytical method, Taking front 10 main constituent contribution rate of accumulative total is 99.99%, and limited amount input reduces the calculating of model Complexity, improves the precision of prediction of model；

The main constituent of described sample determines by the following method:

If x₁, x₂..., x_nFor taking from the sample of overall x, wherein x_i=(x_i1, x_i2..., x_ip) ', i=1, 2 ... n；

Note sample observations matrix is:

$X=\left[\begin{array}{cccc}{x}_{11}& x_{12}& ...& x_{1p}\\ x_{21}& x_{22}& ...& x_{2p}\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ \\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ x_{n1}& x_{n2}& ...& x_{np}\end{array}\right]$

The corresponding sample of every a line of x, the corresponding variable of every string；

Note sample covariance matrix and sample correlation coefficient matrix are respectively as follows:

$S=\frac{1}{n-1}\underset{i=1}{\overset{n}{\Σ}}(x_i-\stackrel{\‾}{x}){(x_i-\stackrel{\‾}{x})}^{\′}=\left(s_{ij}\right)$

$\hat{R}=\left(r_{ij}\right),r_{ij}=\frac{s_{ij}}{\sqrt{s_{ii}}\sqrt{s_{jj}}}$

Wherein,For sample mean；

Using S as the estimation of ∑,As the estimation of R, from S orSet out and can try to achieve the main one-tenth of sample Point；

A-4, set up artificial nerve network model: take artificial neural network algorithm, according to input Output data characteristics determines that the structure of neutral net, recycling training data train this neutral net, Obtain the discriminating model of flat interest Herb Gynostemmae Pentaphylli；

B, the discriminating of unknown sample

Unknown sample scans near infrared light spectrogram under the same conditions, chooses main constituent number, foundation The neural network model trained is to judge the true and false of unknown sample, and output node uses binary system respectively Numeral represents, 10 representatives are flat interest Herb Gynostemmae Pentaphylli, and 01 representative is non-flat interest Herb Gynostemmae Pentaphylli；

The main constituent of sample is by from sample correlation coefficient matrixSet out and solve:

IfForP eigenvalue,For corresponding positive presentate Position characteristic vector, then p main constituent of sample is

${\hat{y}}^{*}={{\hat{t}}_i}^{{*}^{\′}}{x}^{*},i=1,2,...,p$

By sample x_iObservation after standardizationSubstitute into jth main constituent, i.e. can get sample x_iJth Principal component scores

${\hat{y}}_{ij}^{*}={\hat{t}}_j^{{*}^{\′}}{x}_i^{*},i=1,2,...,n;j=1,2,...,p.$

Near infrared spectrum the most according to claim 1 differentiates the method for flat interest Herb Gynostemmae Pentaphylli, its It is characterised by: the artificial nerve network model of setting up described in step A-4 includes using MATLAB soft Part sets up the BP artificial neural network of input layer 10-hidden layer node 5-output layer node 2 Model:

A-4-1, the determination of input layer number: taking 10 principal component scores is parameter, determines The input layer of network is 10；

A-4-2, the determination of the number of hidden nodes: obtain with lower formula:

$L=\sqrt{0.43mn+0.12n^2+2.54m+0.77n+0.35}+0.51$

Wherein m, n are respectively input node and output node number, and hidden node number is by formula Obtain an initial value, then utilize progressively growth method correction, obtain empirical value 5；

A-4-3, the determination of output layer nodes: according to judging that Herb Gynostemmae Pentaphylli sample is belonging to flat interest and produces Still fall within two kinds of non-flat interest place of production result, determine that the output node of neutral net is 2.