Background technology
The mankind obtain food, not only meet psychological need, also meet psychological needs and spiritual enjoyment.Food with good or peculiar flavour, except direct increase appetite, promote digestive juice secretion, the function strengthening digesting and assimilating, also can make people obtain happiness on sense organ, the glamour place of Here it is flavour of food products.Flavour of food products refers to before and after the food entrance the sensory stimulations such as eye to human body, nose, tongue, thereby causes the comprehensive impression of people to its general characteristic, is in particular in the many aspects such as color, shape.For a long time, the check of flavour of food products and judge rely on the artificial sense check usually, but the artificial sense check has some limitations itself: 1) the artificial sense check needs the process of a long-term training and experience accumulation, be generally to rely on well-trained expert to complete, ordinary consumer does not possess this ability; 2) the artificial sense check easily is subject to the interference of external factor, even well-trained expert, their sensory sensitivity also can be subject to the interference of the factors such as sex, experience, the state of mind, health, motivation attitude and territorial environment, has influence on accuracy and the consistance of evaluation result.Therefore, invent a kind of intelligentized sensory analysis method, with indirect labor's organoleptic examination, proofread and correct artificial sense check deviation, improve objectivity and the consistance of sensory analysis result, to instructing food production, guaranteeing that food quality and increase food added value etc. have extremely important meaning.
In recent years, along with the development of computing machine, microelectronics and material science, the Novel bionic sensor technologies such as vision, sense of smell and the sense of taste are come out one after another, and they have obtained application in the food quality intelligent Evaluation.Food product packets is containing multiple organoleptic quality indexs such as color, shapes, the quantity of information that the single-sensor technology is obtained is comprehensive not, finally have influence on the accuracy of testing result, in recent years, multi-sensor information fusion technology, also having obtained Preliminary Applications aspect the food quality intelligent Evaluation, is compared with single detection means, the application of many information fusion in the Intelligent Food organoleptic examination, can make to detect the information acquire more comprehensively, result is more objective, and similar with the human perception process.
From literature searches such as paper and patents, can find out, at present in the food quality intelligent Evaluation, biomimetic sensor adopts business-like instrument mostly, only play the effect of analytical instrument, detecting mechanism and artificial sense checks mechanism also to differ greatly, truly do not possess mankind's similar " perceptional function ", and from the paper periodical of having delivered and the patent of having authorized, can know and have the people that multi-information merging technology is applied to food, the Non-Destructive Testing aspect of the qualities such as agricultural product, as " the quality of famous tea intelligent equipment based on multi-sensor data fusion is evaluated method " (ZL200910232916.9) and " a kind of food sensory quality assessment method and system " (ZL201110142990.9), but on multiple sensor information fusion, can find it is substantially all that several sensor technologies are simply connected, to the physical significance span, larger heat transfer agent is carried out the arbitrary decision stack, this is not fusion truly in fact, also with the mankind's perception behavior, differ greatly.The human perception behavior is a very complicated process, the topological structure formed by sensory neuron, nervous centralis and cerebral cortex sensory area.Current research shows that the sensory neuron end is connected with a kind of star-shaped glial cell, causes the space connection of this topological structure very complicated, and the perception information that different sense organs produce exists mutually and intersects in transmitting procedure; That is to say, in sensory analysis, a plurality of organ actings in conjunction of the mankind, influence each other.
Given this, the invention provides a kind of sensory method merged across the many information interaction inductions of perception that realizes simulating the people.
Summary of the invention
With single detection means, compare, the application of many information fusion in the Intelligent Food organoleptic examination, can make the information that detect to obtain more comprehensively, result is more objective, and the perception of more pressing close to the mankind.How to realize utilizing the technology of many information fusion to carry out the food quality intelligent Evaluation and become patent of the present invention one guardian technique place.
The technical scheme that this patent provides is: in the bionical evaluation of a kind of Intelligent Food, across perception information, respond to alternately fusion method, utilize nose, eye, the tongue perceptual organ of dissimilar biomimetic sensor from sense of smell, vision, sense of taste aspect simulation people, extract the characteristic variable of food, obtain a matrix, the heat transfer agent that comprises different physical significances and different dimension, magnitude, by principal component analysis (PCA), all characteristic variables of extracting are recombinated, obtained several mutually orthogonal major component PC
i, realize that the mutual induction of dissimilar heat transfer agent is merged; Utilize front n major component PC1, PC2, PC3 ... PCn is returned sense of smell, vision and the sense of taste score of artificial sense check, obtains respectively the dummy variable of corresponding sense of smell, vision and the sense of taste; The finally input using these dummy variables as decision system, build non-linear decision system by Genetic Neural Network Method, realized that sense of smell, vision, the sense of taste merge across the induction of detecting sensor information interaction.
The concrete steps of the inventive method are:
(1) sense of smell information acquisition, select olfactory sensor equipment simulating people's nose to gather the smell information of food, obtains reflecting p characteristic variable a of food smell information according to different olfactory sensor arrays
1, a
2, a
3... .a
p;
(2) visual information collection, the eyes of selecting vision sensor to simulate the people gather the exterior quality information of food, and the outward appearance quality information is carried out to color characteristic and analysis of texture, obtain reflecting q characteristic variable b of appearance information
1, b
2, b
3... b
q; (3) collection of sense of taste information, select taste sensor equipment simulating people's tongue to gather the flavour information of food, obtains reflecting t characteristic variable c of food flavour
1, c
2, c
3... c
t;
(4) characteristic variable that (1)-(3) obtain is combined into to m is capable, the matrix of (p+q+t) row;
(5) adopt the method for principal component analysis (PCA) to complete purification, dimensionality reduction and the screening process to above-mentioned characteristic variable, principal component analysis (PCA) is recombinated to former all characteristic variables in the higher-dimension Virtual Space, obtains several mutually orthogonal major components;
(6) extract front n major component PC1, PC2, PC3 ... PCn, set up the score dummy variable L1, L2, L3 of sense of smell, vision, three types of the sense of taste respectively to front n major component PC1, PC2, PC3 by multiple linear regression ... the equation of linear regression of PCn;
(7) score dummy variable L1, L2, the L3 of the sense of smell of all samples, vision, three types of the sense of taste and the artificial sense of answering are in contrast estimated to obtain graduation as the input layer substitution multi-sensor data fusion model based on BP-ANN, sample is divided into calibration set and two groups of forecast set at random according to the ratio of 3:2;
(8) according to BP-ANN Fusion Model Output rusults, check the height of forecast set discrimination, verify the stability of institute's established model.
Adopt in this patent and set up dummy variable: (1) sense of smell information acquisition, select olfactory sensor equipment simulating people's nose to gather the smell information of food, obtain reflecting the characteristic variable a of food smell information according to different olfactory sensor arrays
1, a
2, a
3... .a
p.(2) visual information collection, select vision sensor technical modelling people's eyes to gather the exterior quality information of food, and the outward appearance quality information is carried out to color characteristic and analysis of texture, obtains reflecting the characteristic variable b of appearance information
1, b
2, b
3... b
q.(3) collection of sense of taste information, select taste sensor equipment simulating people's tongue to gather the flavour information of food, obtains reflecting the characteristic variable c of food flavour
1, c
2, c
3... c
t.The heat transfer agent of three kinds of different physical significances obtaining and different dimension, magnitude is combined into to m is capable, the huge matrix of (p+q+t) row:
In the face of like this huge data, usually according to unified approach, be mapped to a higher dimensional space, and take out several virtual vectors and complete the fusion process across perception information.This patent adopts the method for principal component analysis (PCA) (PCA) to complete purification, dimensionality reduction and the screening process to characteristic variable, and principal component analysis (PCA) is recombinated to former all characteristic variables in the higher-dimension Virtual Space, obtains several mutually orthogonal major components.Patent of the present invention is selected front n major component PC1, PC2, PC3 with the height of variance contribution ratio ... PCn, take that it can represent that the quantity of information of food quality more than 90 percent is as best.Obviously, after processing through principal component analysis (PCA), can obtain:
PC1=k
11a
1+k
12a
2+k
13a
3……+k
1pa
p+k
1(p+1)b
1+k
1(p+2)b
2+……k
1(p+q)b
q+k
1(p+q+1)c
1+k
1(p+q+2)c
2+……k
1(p+q+t)c
t,
PC2=k
21a
1+k
22a
2+k
23a
3……+k
2pa
p+k
2(p+1)b
1+k
2(p+2)b
2+……k
2(p+q)b
q+k
2(p+q+1)c
1+k
2(p+q+2)c
2+……k
2(p+q+t)c
t,
PC3=k
31a
1+k
32a
2+k
33a
3……+k
3pa
p+k
3(p+1)b
1+k
3(p+2)b
2+……k
3(p+q)b
q+k
3(p+q+1)c
1+k
3(p+q+2)c
2+……k
3(p+q+t)c
t,
PC4=k
41a
1+k
42a
2+k
43a
3……+k
4pa
p+k
4(p+1)b
1+k
4(p+2)b
2+……k
4(p+q)b
q+k
4(p+q+1)c
1+k
4(p+q+2)c
2+……k
4(p+q+t)c
t,
………………………………………………………………………………………………………………………
………………………………………………………………………………………
PCn=k
n1a
1+k
n2a
2+k
n3a
3……+k
npa
p+k
n(p+1)b
1+k
n(p+2)b
2+……k
n(p+q)b
q+k
n(p+q+1)c
1+k
n(p+q+2)c
2+……k
n(p+q+t)c
t
Each major component forms by all sensing characteristic variables in theory, has embodied interactivity.Then extract front n major component PC1, PC2, PC3 ... PCn, set up the score dummy variable L1, L2, L3 of sense of smell, vision, three types of the sense of taste respectively to front n major component PC1, PC2, PC3 by multiple linear regression (MLR) ... the equation of linear regression of PCn:
Virtual sense of smell variables L 1=a+b
11pC1+b
12pC2+b
13pC3+b
14pC4+ ... + b
1npCn;
Virtual vision variables L 2=a+b
21pC1+b
22pC2+b
23pC3+b
24pC4+ ... + b
2npCn;
Virtual sense of taste variables L 3=a+b
31pC1+b
32pC2+b
33pC3+b
34pC4+ ... + b
3npCn
The sense of smell obtained by structure, vision, three score dummy variables of the sense of taste, just truly realized that the numeral of mutual induction between human olfactory, vision, the sense of taste is bionical.
Because characteristic variable and the corresponding relation between sensory quality of food of detecting sensor is not simple linear relationship, usually the non-linear means such as genetic neural network, fuzzy support vector machine are incorporated in Fusion Model foundation, the Fusion Model obtained is close with the human perception behavior as much as possible.This patent adopts error back propagation neural network (BP-ANN) means to be dissolved in model foundation and goes.Using three dummy variables and people's true sensory evaluation score (generally by the specialized review expert, providing evaluation score) as input layer, be updated in the BP-ANN model and go, obtain the comprehensive distinguishing result of food quality based on across the detecting sensor information fusion technology, so as the organs such as the eye of human simulation, nose, tongue to food quality as comprehensive sensory evaluation.
The present invention utilizes the modern data method for digging to be purified and dimensionality reduction raw data, from the mutual incoherent three class sensor information data of magnanimity, screens characteristic variable; Design intelligent learning algorithm is coupled itself without practical significance characteristic variable and corresponding artificial sense result, give its corresponding perceptional function, with realization character variable perception meaning, resolves.This patent provides the heat transfer agent of different physical significances and different dimension, magnitude is mapped to a higher dimensional space according to unified approach, and takes out the virtual vector of several scores and complete across perception information and merge.In view of characteristic variable and the corresponding relation between sensory quality of food of detecting sensor is not often simple linear relationship, research is planned the non-linear means such as genetic neural network, fuzzy support vector machine and is incorporated in Fusion Model foundation, and the Fusion Model obtained is close with the human perception behavior as much as possible.
Embodiment
In the present embodiment, the intellectuality that this patent is intended adopting many biomimetic sensors integration technology to be applied to tealeaves millet paste quality detects, simulation people's eye, nose, the large sense organ of tongue three.Flow process is as Fig. 1, and the sensor by three types gathers the sensing characteristic information of different physical significances and different dimension, magnitude.Multi-sensor information gatherer process: (1) sense of smell information acquisition, select the Electronic Nose instrument and equipment PEN3 Electronic Nose of company (German AIRSENSE) simulation people's nose to gather the smell information of millet paste, obtain reflecting the information of millet paste smell according to different olfactory sensor arrays, extract 3 eigenwerts such as each sensor signal maximal value, minimum value and average, electric nasus system has 10 mos sensors, can extract so altogether 30 odor characteristics variablees; (2) visual information collection, select colour examining colour-difference-metre (the full-automatic colour examining colour-difference-metre of DC-P3 type) to gather the color and luster information of millet paste, millet paste color and luster information is carried out to signature analysis, obtain reflecting the aberration brightness L of soup look information, red green degree a, Huang Lan degree b, △ L, △ a, △ b and △ E value, and calculate its derivative value, comprise form and aspect b/a, tone chroma Cab, color saturation Sab and hue angle Hab, totally 11 color and luster characteristic variables; (3) collection of sense of taste information, select the electronic tongue instrument equipment ASTREE II electronic tongues of company (French Alpha MOS) simulation people's tongue to gather the flavour information of millet paste, obtain reflecting 7 working electrode signals of millet paste flavour information, extract the sensor stationary value, a sample obtains 7 sense of taste characteristic variables.
Utilize principal component analysis (PCA) (PCA) to be purified and dimensionality reduction the raw data of resulting three kinds of sensor informations, screen characteristic variable from mass data.PCA is mapped to many information of different physical significances and different dimension, magnitude in a higher dimensional space and goes according to unified approach, realize the mutual restructuring between multiple mutual incoherent characteristic variable information, obtain several mutually orthogonal major components, set up vision, sense of smell, three score dummy variables of the sense of taste respectively to the equation of linear regression of PC1, PC2, PC3, PC4, front 6 major components of PC5, PC6 by multiple linear regression (MLR) again, merge from the mutual induction of truly accomplishing many heat transfer agents.Design intelligent learning algorithm, adopt some nonlinear pattern recognition algorithms that itself is coupled without practical significance characteristic variable and corresponding artificial sense result, gives its corresponding perceptional function, with realization character variable perception meaning, resolves.Artificial sense evaluation score generally is comprised of professional evaluation expert group, comprehensively according to millet paste soup look, fragrance, flavour, provide the evaluation score according to artificial sense test stone (as table 1), then divide according to millet paste is divided into to 1 grade, 2 grades, 3 grades three different quality grades (as shown in table 2) according to integrated level.
Table 1 artificial sense test stone
PTS=0.4*X1+0.3*X2+0.3*X3.
Table 2 integrated level is divided foundation
Integrated level |
The PTS scope |
1 grade |
90-99 |
2 grades |
80-89 |
3 grades |
70-79 |
This patent adopts error back propagation neural network (BP-ANN) means to be dissolved in model foundation and goes.Using three score dummy variables and people's true perception, estimate to such an extent that classify as input layer, be updated in the BP-ANN model and go, the Fusion Model obtained is close with the human perception behavior as much as possible, so as the organs such as the eye of human simulation, nose, tongue to the millet paste quality as comprehensive sensory evaluation.Realization is made reasonable, standard compliant evaluation result across bionical sensing technology to millet paste.
Many information fusion of multisensor intelligent Evaluation result: select 45 sample (Three Estates, 15 samples of each grade) set up model as calibration set substitution BP-ANN, choose 30 sample (Three Estates, 10 samples of each grade) verify the stability of institute's established model as forecast set, resulting predict the outcome into, the goodness of fit of evaluating the grade result with artificial sense is 90%.