CN1828575A - Correlativity analysis method of physical, chemical data and sensing index of formulation product - Google Patents

Abstract

With the SVM gradient analysis, building a technique process to determine former correlation; then, finding out the correlated parameters to sense index, and creating the detection sample. This invention can provide quantitative analysis data for quality evaluation and level partition.

Description

The method of formulation product physicochemical data and sense index related analysis

Technical field

The present invention relates to a kind of method flow of data analysis, realize physicochemical data and sense index related analysis, grade estimation and grade classification specifically to instruct formulation product.

Background technology

In existing formulation product manufacturing, form at prescription and composition thereof that product adopted, need carry out aesthetic quality's evaluation in technological processes such as quality of production management and raw material classifications.For example, normally evaluated, to indicate its different brackets to the consumer with indexs such as odor type style, pungency, strength for cigarette products.

For the commercial production formulation product, evaluation process in the past mainly is to rely on to judge the expert, the mode of tasting by the scene, relies on individual's sensory experience to come divided rank, quality.Though the manufacturing enterprise of formulation product has accumulated the expert assessment and evaluation data of some through long-term production management, be personal behavior because implementation quality evaluates, thereby there are many human factors inevitably in these assessment data itself.As the expert in the grade estimation process, the interference that can be subjected to he or she's mood, health, individual sense organ hobby and experience factors such as degree of fatigue, objectively existing perceived error, the grade classification that finally is reflected in formulation product is inaccurate, be difficult to carry out the further raising and the optimization of production technology.And, organize the expert to carry out quality evaluation and also need higher expense and plenty of time.In addition, by analyzing a large amount of expert assessment and evaluation data as can be seen, exist inevitable close association between the aesthetic quality of formulation product and its formula material composition and the ratio thereof, the raw material of control formula for a product and the quality that ratio can directly be reflected in the aesthetic quality thereof are come up, for example, include the number of chemical component in the tobacco leaf, these chemical compositions react to each other and the sense organ that stimulates the people in smoking process.Know that at present the ratio regular meeting of nicotine composition directly has influence on characteristics such as stimulus to the sense organ and strength, and compositions such as potassium, chlorine grade with its flammability, ash sizable correlativity are arranged.

At the suitability for industrialized production scene, the single-chip microcomputer and the information processing technology have obtained widespread use at present.How to utilize the assessment of data of existing accumulation and sample standard, under the prerequisite that breaks away from interference from human factor, utilize mathematical statistics and logic analysis to point out the physicochemical data of formulation product and the correlativity between the organoleptic indicator, to realize instructing the grade estimation and the grade classification of formulation product, be formulation product manufacturing enterprise technical task anxious to be solved always.

Summary of the invention

The method of formulation product physicochemical data of the present invention and sense index related analysis its objective is the ladder analytic approach based on Support Vector Machine, realizes the measuring and calculating of degree of correlation between the physicochemical data measured value of production scene and the organoleptic indicator.The method of described formulation product physicochemical data and sense index related analysis is to set up a kind of technological process that can determine above-mentioned correlativity in the production run of formulation product.Can in each physicochemical data, find out the parameter relevant, and set up the detection sample to describe out this class correlativity respectively by this method with the organoleptic indicator, thereby the physicochemical data Direct Mark at production scene actual measurement goes out both indexs of correlation, thereby provides quantitative analytical data for the grade estimation and the grade classification of formulation product.

Existing formulation product includes consumer's various product used in everyday, as cigarette, food, spices, food additives etc.As in the raw tobacco material of producing cigarette, including the number of chemical component, these chemical compositions interact in smoking process, the common sense organ that stimulates the smoking people, comprise the sense of taste, sense of smell, sense of touch, analyze from the statistics angle, various cause-effect relationshiies be must form between these chemical compositions and formation people's the organoleptic indicator, simple line style and complicated Nonlinear Mapping relation promptly formed.

By accumulate the subjective appreciation data that the expert makes in large quantities when formulation product is assessed, can tentatively disclose the physicochemical data of these formulation products and organoleptic indicator's mapping relations.But in existing production process of formulated product, the small sample that often exists (the empirical data accumulation is few), higher-dimension, non-linear etc. data characteristics, thereby depend on the expert evaluation data and be difficult to reach due realization assessment accuracy.

Correlation analysis of the present invention, during correlativity between the physicochemical data of analyzing actual measurement and organoleptic indicator's type the ladder analytic approach is combined with Support Vector Machine, utilize the ladder sample from supporting vector machine model, to carry out Knowledge Extraction, thereby disclose physicochemical data and organoleptic indicator's correlativity about correlativity.

Support vector machine of the present invention is a kind of statistical theory that can realize self-teaching on small sample is promptly taked the basis of short run data, and its starting point is to be based on the hypothesis of limited sample.As publishing house of Tsing-Hua University disclosed content in " Statistical Learning Theory " book of publishing on February 1st, 2004 relevant for Support Vector Machine.

For example suppose that H1, H2 are respectively the solid dot and the hollow dots two class sample sets at bidimensional interface, H is a sorting track, and H1, H2 be respectively the nearest sample of distance classification line H and be parallel to the straight line of sorting track, and the distance between then above-mentioned straight line is called class interval (margin).If try to achieve so-called optimal classification line, require sorting track not only two classes correctly can be separated (the training error rate is 0) exactly, and make the class interval maximum.

During described support vector machine can be applied to preferably return and estimate.Consider with F (X)=W.X+b fitting data { x _i, y _i, i=1,2, Λ, n; x _i∈ R ^d, y _i∈ R ^dProblem.And suppose all training datas be mapped to high-dimensional feature space after the useable linear function error freely with precision ε match, promptly

$\left\{\begin{array}{c}{y}_i-w\·x_i-b\≤\mathrm{\ϵ},\\ w\·x_i+b-y_i\≤\mathrm{\ϵ},\end{array}\right.$ i＝1，…，n

Similar to maximization class interval in the optimal classification face, the method for control function collection complicacy is to make regression function the most smooth here, and it is equivalent to and minimizes 2/ ‖ W ‖ ²Consider the situation that allows error of fitting, introduce relaxation factor ξ _i ⁺〉=0 and ξ _i ^-〉=0, then above expression formula becomes,

$\left\{\begin{array}{c}{y}_i-w\·x_i-b\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i^{+},\\ w\·x_i+b-y_i\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i^{-},\end{array}\right.$ i＝1，…，n

Then objective function is converted into,

$\mathrm{\φ}(w,\mathrm{\ξ})=\frac{1}{2}{\left|\right|w\left|\right|}^2+C(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i^{+}+{\mathrm{\ξ}}_i^{-})$

Wherein C is a regularization parameter, is used to control the popularization performance of learning model.Utilization and the same optimization method of pattern-recognition obtain the dual problem of the problems referred to above, promptly in constraint,

$\underset{i=1}{\overset{n}{\mathrm{\Σ}}}({\mathrm{\α}}_i^{+}-{\mathrm{\α}}_i^{-})=0$

$0{\≤\mathrm{\α}}_i^{+},$ ${\mathrm{\α}}_i^{-}\≤C,$ i＝1，…，n

Down, to Lagrange multiplier α _i ⁺, α _i ^-The maximization objective function,

$Q({\mathrm{\α}}^{+},{\mathrm{\α}}^{-})=-\mathrm{\ϵ}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}({\mathrm{\α}}_i^{+}+{\mathrm{\α}}_i^{-})+\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i({\mathrm{\α}}_i^{+}-{\mathrm{\α}}_i^{-})-$

$\frac{1}{2}\underset{i,j=1}{\overset{n}{\mathrm{\Σ}}}({\mathrm{\α}}_i^{+}-{\mathrm{\α}}_i^{-})({\mathrm{\α}}_j^{+}-{\mathrm{\α}}_j^{-})K(x_i,x_j)$

The non-linear regression function that obtains is,

$F\left(x\right)=(w^{+}\·x)+b^{+}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}({\mathrm{\α}}_i^{{+}^{+}}-{\mathrm{\α}}_i^{{-}^{+}})K(x_i,x)+b^{+}$

Above-mentioned modeling process about Support Vector Machine can be realized the study of small sample data.

To achieve the object of the present invention, promptly the physical and chemical test data by input production scene reality obtain and the organoleptic indicator between the correlation analysis result, carry out feature selecting according to degree of relevancy, it is right to improve predictor, the present invention combines the ladder analytic approach with Support Vector Machine, utilize the ladder sample to carrying out Knowledge Extraction, to disclose the correlativity between input and the output by supporting vector machine model.

Expert evaluation data sample according to existing accumulation is constructed the ladder sample, particularly:

Suppose some physical and chemical index x of formulation product _iAnd exist correlativity between the organoleptic indicator Y, for obtaining the analysis conclusion of degree of correlation, this index x in the sample set that can fetch data _iMaximal value and minimum value, rule of thumb choose suitably little Δ x _i, set up and from the minimum value to the maximal value, differ Δ x one by one _iThe staged value.

Simultaneously, the input variable of other physical and chemical index is got some definite value Ci respectively, and all physical and chemical indexs (comprise other definite values Ci and x _iStep values composition data sample set one ladder sample.

According to the statistics viewpoint, suppose the sample Normal Distribution, definite value Ci should select the sample data average of other physical and chemical index input variables usually.Good ladder sample carries out Knowledge Extraction to the support vector machine model of training to utilize structure, discloses the correlativity of the two.Here to the process of model Knowledge Extraction, be exactly in fact that the ladder sample is sent into the process that the support vector machine model of training is tested.For multidimensional input sample, remove the physical and chemical index that will investigate, other index is definite value, so the Trendline (staircase chart) of test result (equally also being step values) can directly reflect the correlationship of the two.

The formed Trendline of above-mentioned test result maps out physical and chemical index x intuitively _iAnd the degree of correlation between the organoleptic indicator Y, defining this degree of correlation is the precipitous index ρ of above-mentioned Trendline, just at input variable (physical and chemical index) x _iWith any 1 x on output variable (organoleptic indicator) the y Trendline ₀Slope, it is defined as follows:

$\stackrel{\‾}{\mathrm{\ρ}}=E\left(\frac{\mathrm{dy}}{\mathrm{dx}}{|}_{x=x_0}\right)=\∫[\frac{\mathrm{dy}}{\mathrm{dx}}{|}_{x=x_0}*f\left(x_0\right)]d{x}_0---\left(1\right)$

The order of magnitude of this precipitous index ρ directly maps out a certain input variable (physical and chemical index) x _iWith the influence of output variable (organoleptic indicator) y have much, i.e. the power of the positive and negative correlativity between it.

At above-mentioned all physical and chemical index x _iMake the precipitous index ρ of above-mentioned trend staircase chart under the separate situation respectively, adopt the simple regression method can comprehensively derive all relevant physical and chemical indexs of assessment and this organoleptic indicator's the quantitative c of the degree of correlation, the physical and chemical index x that a certain organoleptic indicator y is relevant with all ₁, x ₂..., x _mExpression formula as follows:

Y＝F(x ₁，x ₂，...，x _m)

＝f(x ₁)+f(x ₂)+...+f(x _m)+c (2)

Order

y＝f(x ₁)+f(x ₂)+...+f(x _m) (3)

Send into and derive corresponding y value in the following formula (4) accumulating data that the formed training sample of assessment of data concentrates, then the quantitative c of the degree of correlation tries to achieve by following formula:

c＝∑Y _t/n-∑y _t/n (4)

Wherein, m is the item number of physical and chemical index, and n is that training sample is concentrated number of samples.

With the quantitative c substitution of the degree of correlation of (4) formula gained (2) formula, promptly get a certain organoleptic indicator Y physical and chemical index x relevant with all ₁, x ₂..., x _mExpression formula.

If above-mentioned each physical and chemical index is separate, then can directly sum up each relational expression; If each physical and chemical index is not independent each other, then should at first carry out orthogonal transformation to its sample, carry out analytical work again.

The method of formulation product physicochemical data of the present invention and sense index related analysis is based on the ladder analytic approach of Support Vector Machine, and its method flow is:

Every physical and chemical index of actual detected formulation product also organizes industry specialists that finished product is evaluated, with resulting assessment of data accumulate, record, and form with set of data samples;

Reject the wrong or special sample that above-mentioned data sample is concentrated according to expert's industry experience, in the hope of map out correlativity and the program size thereof between physicochemical data and the organoleptic indicator intuitively as far as possible;

Physicochemical data sample set after will putting in order according to indexs such as the place of production, grade, styles is divided into training sample set and checking sample set;

Use the ladder analytic approach and construct several ladder samples, be used for the support vector machine model is carried out Knowledge Extraction and correlation analysis;

Transfer the knowledge model of Support Vector Machine and to its each parameter initialization, training sample is sent in the model trained, and utilize the standard of the degree of conformity of checking sample prediction as preference pattern;

If degree of conformity is higher, then illustrate model basic studies to whole truly rules of sample inside; Otherwise, illustrate that the parameter selection of model is also improper, then readjust model and reach standard up to degree of conformity;

The ladder sample sent into test and obtain exporting the result in the model, the result (is that organoleptic indicator Y and this input data (are a certain physical and chemical index x according to output _i) obtain the precipitous index ρ of Trendline, promptly the staircase chart that mapped out of expression formula (1) carries out preliminary qualitative analysis correlationship;

Precipitous index ρ according to staircase chart reflects weeds out less index item, only keeps certain bigger physical and chemical index x _iFeature selecting as mapping organoleptic indicator Y;

By each physical and chemical index x _iObtain the quantitative c of the degree of correlation with the expression formula of organoleptic indicator Y, i.e. all relevant physical and chemical index x of being mapped out of expression formula (2) at a certain organoleptic indicator Y _iThe relationship between expression formula.

The flow process of described method finishes.

Precipitous index ρ and the quantitative c of the degree of correlation according to the determined formulation product of said method flow process can survey physical and chemical index x according to each in the formulation product actual production process _iInput value, directly map out the degree of correlation of organoleptic indicator Y intuitively, thereby carry out the whether qualified evaluation of grade classification, quality for the raw material of producing formulation product or the finished product of formulation product by above-mentioned expression formula (1) and expression formula (2).

In sum, the method for formulation product physicochemical data of the present invention and sense index related analysis, its advantage and beneficial effect are:

Utilize the assessment of data and the sample sample of existing accumulation, can be implemented in the simple expert's behavior of disengaging and directly map out each relevant organoleptic indicator by the field measurement data, realized analyzing the direct conversion of hardware device from human brain, thus the transition of having realized machine learning and analyzing evaluation.

Use said method and can rely on short run data sample and minimodel can set up the analysis process that instructs formulation product production and grade, simple possible, accuracy rate height, be easy to promote the use of in existing formulation product manufacturing enterprise.

Be no longer dependent on and organize the expert to carry out on-site evaluation repeatedly, can save a large amount of funds and time.

Along with using the correlation analysis findings data that said method obtains, the continuous training sample set and the checking sample set that use of extending method, thereby progressively improve the accuracy of analysis process, thereby having the upgrade mechanism of continuous study and self-perfection, dependability can constantly promote, this method has stronger adaptive faculty.

Description of drawings

Fig. 1 is the method flow diagram of described formulation product physicochemical data and sense index related analysis;

Fig. 2 determines each physical and chemical index x _iProcess flow diagram with analysis of the organoleptic indicator Y degree of correlation and the quantitative c of the degree of correlation;

Fig. 3 is the described process flow diagram of setting up training sample set and checking sample set;

Fig. 4 is that using said method is realized the process flow diagram that tobacco formulation is optimized;

Fig. 5 is a determined ladder trend map among Fig. 4.

Embodiment

Embodiment 1, and as shown in Figure 1, the method flow of described formulation product physicochemical data and sense index related analysis is:

Every physical and chemical index of actual detected formulation product also organizes industry specialists that finished product is evaluated, with resulting assessment of data accumulate, record, and form with set of data samples;

Reject the wrong or special sample that above-mentioned data sample is concentrated according to expert's industry experience, in the hope of map out correlativity and the program size thereof between physicochemical data and the organoleptic indicator intuitively as far as possible;

Physicochemical data sample set after will putting in order according to indexs such as the place of production, grade, styles is divided into training sample set and checking sample set;

Use the ladder analytic approach and construct several ladder samples, be used for the support vector machine model is carried out Knowledge Extraction and correlation analysis;

Transfer the knowledge model of Support Vector Machine and to its each parameter initialization, training sample is sent in the model trained, and utilize the standard of the degree of conformity of checking sample prediction as preference pattern;

If degree of conformity is higher, then illustrate model basic studies to whole truly rules of sample inside; Otherwise, illustrate that the parameter selection of model is also improper, then readjust model and reach standard up to degree of conformity;

The ladder sample sent into test and obtain exporting the result in the model, the result (is that organoleptic indicator Y and this input data (are a certain physical and chemical index x according to output _i) obtain the precipitous index ρ of Trendline, promptly the staircase chart that mapped out of expression formula (1) carries out preliminary qualitative analysis correlationship;

Precipitous index ρ according to staircase chart reflects weeds out less index item, only keeps certain bigger physical and chemical index x _iFeature selecting as mapping organoleptic indicator Y;

By each physical and chemical index x _iObtain the quantitative c of the degree of correlation with the expression formula of organoleptic indicator Y, i.e. all relevant physical and chemical index x of being mapped out of expression formula (2) at a certain organoleptic indicator Y _iThe relationship between expression formula.

As shown in Figure 2, determine described each physical and chemical index x _iWith the flow process of organoleptic indicator Y degree of correlation analysis and the quantitative c of the degree of correlation be:

According to each physical and chemical index x that has made _iAnd the ladder trend map of correlativity between the organoleptic indicator Y returns estimation to the predicted value of ladder sample;

At each physical and chemical index x _iUnder the separate prerequisite, directly each function expression (2) is summed up; If each physical and chemical index x _iNot independent each other, then should at first carry out orthogonal transformation to its sample, carry out analytical work again;

Send into and derive the quantitative c of the degree of correlation in the following formula (4) accumulating data that the formed training sample of assessment of data concentrates;

The final a certain organoleptic indicator Y physical and chemical index x relevant that obtain with all ₁, x ₂..., x _mExpression formula

Y＝F(x ₁，x ₂，...，x _m)

＝f(x ₁)+f(x ₂)+...+f(x _m)+c (2)

On the basis of Fig. 1 and Fig. 2, to shown in Figure 5, in the production of cigarettes process, analyze tobacco leaf physical and chemical index in the single-tobacco-typed cigarette and the correlativity between aesthetic quality's index-pungency as Fig. 3, its operating process is:

The first step is carried out data aggregation, the data of each single-tobacco-typed cigarette that the typing detection obtains or the base attribute of tobacco product, physics and chemistry, sensory evaluating smoking's quality, fume indication.Form following table 1 and table 2.

Numbering	Total reducing sugar	Total nicotine	Reducing sugar	Protein	Total nitrogen	Chlorine	Potassium	Shi Muke	Sugar alkali ratio	Potassium chlorine ratio
											1	19.1	7.3	20.5	13.7	1.9	0.3	1.8	1.0	12.0	4.1
2	21.7	1.4	21.3	12.1	1.7	0.4	2.5	3.2	4.2	1.9
											3	19.1	8.5	25.6	11.8	2.0	0.9	2.6	2.4	10.4	3.0
...	...	...	...	...	...	...	...	...	...	...

Table 1

Numbering	Odor type	Fragrance matter	Perfume quantity	Assorted gas	Pungency	Concentration	Strength	Pleasant impression	Flammability	Ash content
											1	3	3	3	3	3	2	3	3	2	2
2	4	2	2	2	2	2	2	3	1	2
											3	6	2	2	2	2	2	1	3	2	1
...	...	...	...	...	...	...	...	...	...	...

Table 2

Second step, carry out data preparation, according to expertise the wrong sample with unusual is filtered out; Then each attribute orthogonalization process is formed data sample to be analyzed.

The 3rd step, construct each sample set, according to indexs such as the place of production, grade, styles the reduced data sample is divided into training sample set and checking sample set, the ratio of sample size is 5: 1, uses staircase method structure ladder sample simultaneously.

The 4th step, initiation parameter, the model parameter of initialization Support Vector Machine comprises the concrete form of regularization parameter, kernel function etc.

In the 5th step, training sample set is sent in the model of setting and trained.

The 6th step, the checking sample set is sent in the support vector machine model that trains, obtain test result.

The 7th step, judge whether the support vector machine model is suitable, and according to its degree of conformity of degree of conformity formula to calculating of enterprise formulation.Promptly with the ratio of the sum of errors permissible error of real number value and the desired value of output as the basis of calculation of assessing accuracy, as shown in the formula:

If degree of conformity is higher, illustrate that then model learnt the inherent laws of sample; Otherwise, revise model parameter and repeat the 5th, 6 liang of step operation, reach more than 70% up to degree of conformity.

In the 8th step, test ladder sample: will construct good ladder sample and send into and test in the model, and obtain stepped output result, this output result has promptly disclosed sample and has exported the Changing Pattern of importing with specific sample.

Numbering	Total reducing sugar	Total nicotine	Reducing sugar	Protein	Total nitrogen	Chlorine	Potassium	Execute wooden brother	Sugar alkali ratio	Potassium chlorine ratio		Pungency
													1	C1	C2	C3	C4	1.34	C6	C5	C7	C8	C9	2.2671
2	C1	C2	C3	C4	1.44	C6	C5	C7	C8	C9		2.2704
													3	C1	C2	C3	C4	1.54	C6	C5	C7	C8	C9	2.2738
4	C1	C2	C3	C4	1.64	C6	C5	C7	C8	C9		2.2774
													5	C1	C2	C3	C4	1.74	C6	C5	C7	C8	C9	2.281
...	...	...	...	...		...	...	...	...	...		...

The 9th goes on foot, and makes the ladder trend map of correlativity between physical and chemical index-total nitrogen and the organoleptic indicator-pungency, and input physical and chemical index data are horizontal ordinate, is ordinate with output organoleptic indicator Y, as shown in Figure 5.

As can be seen from Figure 5, total nitrogen becomes significant positive correlation with the pungency of tobacco leaf, and promptly along with the increasing of total nitrogen content in the tobacco leaf, the tobacco leaf pungency is increasing.

The tenth step calculated the precipitous index ρ that reflects according to the ladder trend map, as following table 3:

Nicotine	Reducing sugar	Total reducing sugar	Sugar subtracts ratio	Protein	Total nitrogen	Shi Muke	Potassium chlorine ratio	Potassium	Chlorine
										0.689	-0.468	-0.261	0.243	0.196	0.169	-0.137	-0.041	-0.0025	-0.0022

By table 3 as can be seen, nicotine, sugared alkali ratio, protein and total nitrogen become comparatively significant positive correlation with pungency, and reducing sugar, total reducing sugar and Shi Muke and pungency are negative correlation.

According to the size of ladder index absolute value, we can reject potassium chlorine ratio, potassium and three attribute items of chlorine, reach the purpose of feature selecting.When analyzing the pungency of tobacco leaf from now on, can not need to detect this three physical and chemical indexs.

According to the funtcional relationship that finally obtains all physical and chemical indexs and organoleptic indicator,

Pungency=-0.261* total reducing sugar+0.689* total nicotine-0.468* reducing sugar+0.196* protein+0.169* total nitrogen-0.0022* chloro-0.0025* potassium-0.137* executes wood gram+0.243* sugar alkali ratio-0.041* potassium chlorine and compares+1.96.

Wherein, the quantitative c=1.96 of the degree of correlation.

As mentioned above, promptly be the major programme of the method for described formulation product physicochemical data and sense index related analysis.For the suitable modification of this method, comprise that the combination of ladder analytic approach and other intelligent modeling all should be protection scope of the present invention.

Claims (3)

1, the method for a kind of formulation product physicochemical data and sense index related analysis, it is characterized in that: its method flow is, every physical and chemical index of actual detected formulation product also organizes industry specialists that finished product is evaluated, with resulting assessment of data accumulate, record, and form with set of data samples;

Reject the wrong or special sample that above-mentioned data sample is concentrated according to expert's industry experience, in the hope of map out correlativity and the program size thereof between physicochemical data and the organoleptic indicator intuitively as far as possible;

Physicochemical data sample set after will putting in order according to indexs such as the place of production, grade, styles is divided into training sample set and checking sample set;

Use the ladder analytic approach and construct several ladder samples, be used for the support vector machine model is carried out Knowledge Extraction and correlation analysis;

Transfer the knowledge model of Support Vector Machine and to its each parameter initialization, training sample is sent in the model trained, and utilize the standard of the degree of conformity of checking sample prediction as preference pattern;

If degree of conformity is higher, then illustrate model basic studies to whole truly rules of sample inside; Otherwise, illustrate that the parameter selection of model is also improper, then readjust model and reach standard up to degree of conformity;

The ladder sample sent into test and obtain exporting the result in the model, the result (is that organoleptic indicator Y and this input data (are a certain physical and chemical index x according to output _i) obtain the precipitous index ρ of Trendline, promptly the staircase chart that mapped out of expression formula (1) carries out preliminary qualitative analysis correlationship;

Precipitous index ρ according to staircase chart reflects weeds out less index item, only keeps certain bigger physical and chemical index x _iFeature selecting as mapping organoleptic indicator Y;

By each physical and chemical index x _iObtain the quantitative c of the degree of correlation with the expression formula of organoleptic indicator Y, i.e. all relevant physical and chemical index x of being mapped out of expression formula (2) at a certain organoleptic indicator Y _iThe relationship between expression formula.

2, the method for formulation product physicochemical data according to claim 1 and sense index related analysis is characterized in that: described organoleptic indicator Y and a certain physical and chemical index x _iThe precipitous index ρ of Trendline full in following formula,

$\stackrel{\‾}{\mathrm{\ρ}}=E\left(\frac{\mathrm{dy}}{\mathrm{dx}}{|}_{x=x_0}\right)=\∫[\frac{\mathrm{dy}}{\mathrm{dx}}{|}_{x=x_0}*f\left(x_0\right)]d{x}_0---\left(1\right)$

This precipitous index ρ promptly is input variable (physical and chemical index) x _iWith any 1 x on output variable (organoleptic indicator) the y Trendline ₀Slope.

3, the method for formulation product physicochemical data according to claim 2 and sense index related analysis is characterized in that: described definite each physical and chemical index x _iWith the flow process of organoleptic indicator Y degree of correlation analysis and the quantitative c of the degree of correlation be,

According to each physical and chemical index x that has made _iAnd the ladder trend map of correlativity between the organoleptic indicator Y returns estimation to the predicted value of ladder sample;

At each physical and chemical index x _iUnder the separate prerequisite, directly each function expression (2) is summed up; If each physical and chemical index x _iNot independent each other, then should at first carry out orthogonal transformation to its sample, carry out analytical work again;

To accumulate the formed training sample of assessment of data and concentrate, then the quantitative c of the degree of correlation satisfies following expression formula,

c＝∑Y _t/n-∑y _t/n (4)

Wherein, m is the item number of physical and chemical index, and n is that training sample is concentrated number of samples;

The final a certain organoleptic indicator Y physical and chemical index x relevant that obtain with all ₁, x ₂..., x _mCorrelativity, satisfy following expression formula,

Y＝F(x ₁，x ₂，...，x _m) (2)

＝f(x ₁)+f(x ₂)+...+f(x _m)+c

Wherein, c promptly is that the degree of correlation is quantitative.

Images