CN107843695A - Tobacco and the electronic nose instrument evaluation method of tobacco product aesthetic quality - Google Patents
Tobacco and the electronic nose instrument evaluation method of tobacco product aesthetic quality Download PDFInfo
- Publication number
- CN107843695A CN107843695A CN201711060781.3A CN201711060781A CN107843695A CN 107843695 A CN107843695 A CN 107843695A CN 201711060781 A CN201711060781 A CN 201711060781A CN 107843695 A CN107843695 A CN 107843695A
- Authority
- CN
- China
- Prior art keywords
- mrow
- seconds
- tobacco
- msubsup
- flue gas
- 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
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
- G01N33/0001—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00 by organoleptic means
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Chemical & Material Sciences (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Pathology (AREA)
- Data Mining & Analysis (AREA)
- Analytical Chemistry (AREA)
- Immunology (AREA)
- Medicinal Chemistry (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Biochemistry (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Food Science & Technology (AREA)
- Manufacturing Of Cigar And Cigarette Tobacco (AREA)
- Other Investigation Or Analysis Of Materials By Electrical Means (AREA)
Abstract
Tobacco of the present invention and one of the electronic nose instrument evaluation method, feature of tobacco product aesthetic quality are to simulate the process of smokeing panel test of professional, it is determined that detection and assessment process of the tobacco electronic nose instrument to tobacco and tobacco product sample;The two of feature are, by testing a large amount of standard specimens, establish gas sensor array response and brand mark and the tobacco big data of aesthetic quality's index evaluation value relation;The three of feature are that tobacco sensory quality assessment problem is converted into 1. identification and 2. aesthetic quality's index forecasting problem, solved with a kind of modular neural network cascade model;The four of feature are the single output neural networks with single hidden layer structures for proposing effective task analytic approach, the S types activation functions of amendment and optimization, realize the neutral net Fast Learning towards big data;The five of feature are to propose most of ballot decision rules and quantizing rule based on multiple neural network modules, realize the identification and the prediction of aesthetic quality's index of extensive tobacco and tobacco product.
Description
Technical field
The present invention-tobacco and the electronic nose instrument evaluation method of tobacco product aesthetic quality, it is related to computer, accurate survey
Amount, precision optical machinery, automatically control, analytical chemistry, tobacco product field, towards tobacco and tobacco product process of producing product quality
Control and market surpervision demand, mainly solve tobacco based on electronic nose instrument and tobacco product identification is pre- with aesthetic quality's index
Survey problem.
Background technology
" smokeing panel test " refers in certain circumstances, and the certain organs such as the oral cavity of employment, nasal cavity, throat are to tobacco and tobacco system
The quality of product and flavouring essence for tobacco and effect carry out naked eyes evaluation, i.e. cigarette interior quality-aesthetic quality's index evaluation.Existing rank
Section, tobacco determine by the sensory evaluating smoking of people completely with tobacco product quality good or not.Compared with the taste compounds such as wine, tea, food, cigarette
Grass and the unique distinction of tobacco product be, quality evaluation places one's entire reliance upon the sense organ of people, there is no practicable reason so far
Change Indexs measure and analysis method.
Existing professional standard《Tobacco and tobacco product sensory evaluation method》Yc/T138-1998 is by State Tobacco Monopoly Bureau
1998.03.12 day is issued, and implements 1998.05.01 days, is widely used in tobacco enterprise production and industry and market management.
The standard provides that member of smoking is evaluated tobacco and tobacco product sample using the overall method of smokeing panel test that circulates, and fills in each of table 1
Item aesthetic quality index score.Wherein, gloss and harmony digit synbol divide unit to be 0.5, and fragrance, miscellaneous gas, excitant and pleasant impression refer to
It is 1.0 to mark a point unit.
Table 1, cigarette (tobacco manufactured goods) aesthetic quality examine tables of original record (YC/T138-1998)
Note:It is full marks value in bracket.
In addition to color and luster this external quality index, in matter in this 5, fragrance, coordination, miscellaneous gas, excitant, pleasant impression in table 1
Figureofmerit all has direct or indirect relation with nasal receptor.In order that Analyses Methods for Sensory Evaluation Results is as objective as possible, just, accurate,
The personnel that smoke panel test have to pass through Specialized Theory training and technology is taken exercise, and keep correct psychological condition and good physical qualification;Comment
Standard requirement should be reached by inhaling environmental condition, calibrated and sought unity of talking with calibration sample before smokeing panel test.Smoke panel test group number of members should foot
It is more than enough, typically should be 7 people and more than, all members of group average values of giving a mark are some tobacco product sample aesthetic quality's index
Final score, and carry out significance test with bi-distribution and chi-square distribution.
At present, " smoke panel test " and be not only to determine tobacco and tobacco product quality uniquely practicable method, and be also true
Determine the basis of tobacco and tobacco product product formula structure, be exploitation new product, keep existing product style and steady quality
Necessary and decisive means.For the smoker of consumer one, smoking is a kind of physiological stimulation and fine enjoyment, in the absence of work
Problem;And for the personnel that smoke panel test, identification of smokeing panel test is then an extremely arduous, elaboration, and thought will during smokeing panel test
High concentration, concentrate, to be judged within short a few minutes.
According to existing professional standard YC/T138, cigarette interior quality subjective appreciation is typically using " overall to circulate method of smokeing panel test ":
Flue gas is sucked oral cavity by member of smoking, is swallowed by throat, is then discharged slowly from nasal cavity again, during entirely inhaling, gulp down, tell
Evaluated using sense organ of all smokeing panel test.Degree difference, physiology are caught due to sense organ sensitivity and to indices
The influence of the factors such as, environmental condition difference different with psychologic status, cause smoking result description and the difference judged between member of smoking
Different, this absolutely proves that sensory evaluating smoking's method has significant limitation.
Cigarette brand is numerous, and tobacco source is numerous, and tobacco aromaticss essence is numerous.Identification cigarette brand is people to tobacco electricity
The basic demand of sub- nose instrument.If on the basis of brand recognition, further requirement tobacco electronic nose realizes tobacco and tobacco
The quantitative prediction of aesthetic quality's indexs such as product fragrance, coordination, miscellaneous gas, excitant, pleasant impression, and it evaluates credit rating accordingly,
Which forms big data to analyze and process problem, and challenge is proposed to existing machine learning method.In order to by electronic nose instrument and
Method is used in tobacco and tobacco product aesthetic quality's metrics evaluation, it would be desirable to which invention is towards extensive tobacco and tobacco product
The machine learning method of Site Detection, identification and aesthetic quality's grade forecast.
Using the neutral net of standard sigmoid activation functions f (x)=1/ (1+exp (- x)) typically by data set transformation
To [0,1] scope, the actual acquiescence component average of this way is about 0.5.If input component is transformed to certain limit by us,
It can suitably amplify between tobacco brand at the interval of the sample input space, be advantageous to neutral net and accelerate pace of learning, improve
Learn precision and improve Generalization Ability.
To the multi-class problem of big data, the study of the multi output machine learning model of multi input one of monolithic devices and generalization
Can be often undesirable.For example, the Multiple input-output neutral net of monolithic devices is easily absorbed in local minimum in learning process
Point.Moreover, brand recognition is carried out simultaneously with tobacco product to extensive tobacco and multinomial organoleptic indicator's quantitative prediction is this
The classification and function that problem is related in machine learning field approaches (nonlinear regression) two research directions, it is necessary to invent new machine
Device learning model and algorithm, including task analytic approach, model topology optimization method, fast learning algorithm and decision-making technique.
The content of the invention
The present invention is in existing patent of invention《A kind of machine olfaction device and its olfactory analog method of testing》(referring to patent
Application number:02111046.8)、《A kind of machine olfaction odor distinguishing method based on modular combination neutral net》(referring to special
Sharp application number:03141537.7)、《A kind of small automatic machine smell instrument and smell analysis method》(referring to patent application
Number:200710036260.4)、《A kind of olfactory analog instrument and a variety of smell qualitative and quantitative analysis methods》(referring to patent application
Number:201010115026.2)、《Towards gas sensor selection, replacing and the bearing calibration of olfactory analog instrument》(referring to patent
Application number:201310419648.8) and《A kind of olfactory analog instrument and predetermined substance gas (smell) taste grade field assay method》
(referring to number of patent application:201310315482.5) on the basis of, a kind of electronic nose instrument analytical method is invented to solve tobacco
With tobacco product scene automatic detection, identification and aesthetic quality's quantification of targets forecasting problem.
To achieve these goals, the present invention-tobacco and the electronic nose instrument evaluation method of tobacco product aesthetic quality, its
In tobacco electronic nose instrument include gas sensor array module, flue gas automatic sample handling system, computer control and data point
Analysis system, automatic ignition device, realize tobacco and tobacco product identification and the prediction of aesthetic quality's index score.
Tobacco electronic nose instrument is 5 minutes to the flue gas sampling cycle of a tobacco and tobacco product sample, gas sensing
Device array recovers (210 seconds), pure air Accurate Calibration (40 seconds), balance (2 seconds), second mouthful of flue gas suction (2 after preliminary
Second), surrounding air rinse (46 seconds) totally 6 stages.
During flue gas sampling, under the control of the computer, flue gas automatic sample handling system is 1050 ml/mins with 17.5 milliliters/seconds
The flow of clock aspirates second mouthful of flue gas, passes through gas sensor array annular working chamber, skims over gas sensor sensitive membrane
Surface, continue 2 seconds, gather 35 milliliters of flue gas, therefore gas sensor array produces sensitive response.The quarter that self-balancing state starts
Rise, computer control and data analysis system start recording response data, successively record balance (2 seconds), second mouthful of flue gas suction
(2 seconds), surrounding air rinse the gas sensor array voltage responsive value in (first 36 seconds) this 3 stages, total duration 40 seconds.Flue gas
The data of other time in sampling period do not record.
Within the data record time of 40 seconds, single gas sensor to the voltage response curves stable state of second mouthful of flue gas most
Big value is extracted as characteristic component, therefore the array of 16 gas sensor compositions produces one 16 dimension voltage responsive vector.
Data record terminate after 10 seconds in, computer control with data analysis system according to this response vector to tobacco and tobacco system
Product sample carries out brand, the place of production, true and false identification and fragrance, coordination, miscellaneous gas, excitant, pleasant impression totally 5 aesthetic quality's index scores
Prediction.
Computer controls to be tried tobacco and tobacco product with data analysis system using modular neural network cascade model
Sample carries out 1. identification and 2. aesthetic quality's index score is predicted.1. the modular neural network cascade model first order is by n (n-1)/2
Individual single output nerve network forms side by side, forms n identification group of voting, is grown tobacco the identification with tobacco product, including product for n
Board, the place of production and true and false identification.2. the modular neural network cascade model second level is arranged side by side by the single output nerve network in n × 5
Composition, every 5 one group, grown tobacco and the fragrance of tobacco product, coordination, miscellaneous gas, excitant, pleasant impression this 5 aesthetic qualities for n
Index score is predicted.Each single output nerve network undergo the study stage to training set X and treat random sample this x identification and
Aesthetic quality's index score forecast period.
Identification and aesthetic quality index score of the tobacco electronic nose instrument to tobacco and tobacco product sample undetermined are predicted, are wrapped
Include following steps:
(1) start shooting:Instrument preheats 30 minutes, and surrounding air flows through the one or two successively with the flow of 6500 ml/mins
Two-way electromagnetic valve, gas sensor array annular working chamber, the 5th two-position two-way solenoid valve, the 4th two-position two-way solenoid valve, most
Outdoor is discharged to eventually.Gas sensor array constant temperature indoor temperature reaches constant 55 ± 0.1 DEG C from room temperature.
(2) the flue gas sampling cycle starts:Operating personnel click on " starting to detect " button of screen drop-down menu, and instrument enters
The flue gas sampling cycle of 5 minutes is lasted, computer automatically generates the text of one entitled " xxx " in specified folder, with
Record response data of the gas sensor array to flue gas.
(3) it is preliminary to recover:In the 0.00-210.00 seconds in flue gas sampling cycle, surrounding air is with the stream of 6500 ml/mins
Amount flows through the first two-position two-way solenoid valve, gas sensor array annular working chamber, the 5th two-position two-way solenoid valve, the 4th successively
Two-position two-way solenoid valve, it is ultimately drained into outdoor.In the case where flow is the surrounding air flushing action of 6500 ml/mins, adhesion
It is flushed away in the flue gas scent molecule of gas sensor sensitive membrane surface and inner-walls of duct, gas sensor array tentatively returns to
Normal condition, last 210 seconds.
(4) pure air Accurate Calibration:In the 210.00-250.00 seconds in flue gas sampling cycle, (a) pure air is accurately marked
Fixed and (b) cigarette insertion, automatic ignition, first flue gas are aspirated, the moon/spontaneous combustion the two links are carried out simultaneously, last 40 seconds.
(4a), pure air Accurate Calibration:Pure air with 17.5 milliliters/seconds be the flow of 1050 ml/mins successively
Flow through second throttle, the 6th two-position two-way solenoid valve, gas sensor array annular working chamber, the 3rd bi-bit bi-pass electromagnetism
Valve, interior is finally discharged to, lasts 40 seconds.Pure air makes gas sensor array Exact recovery to normal condition.
(4b.1) cigarette inserts:It is pure air Accurate Calibration state in the 210.00-225.00 seconds in flue gas sampling cycle
Initial 15 seconds in, screen display " cigarette insertion " printed words, operating personnel by tested cigarette sample filter end insertion cigarette clamp
Device, insertion depth are 9.0 ± 0.5 millimeters.
(4b.2) automatic ignition:It is pure air Accurate Calibration state in the 225.00-231.00 seconds in flue gas sampling cycle
15.00-21.00 seconds, ignition coil are powered;At the same time, ignition head, which is moved to the left 9 millimeters, makes ignition coil and tested sample
Contact and light sample, continue 6 seconds.In the 231.00-269.00 seconds in flue gas sampling cycle, magnet coil is powered, ignition coil
Power off and be returned to reference position, last 38 seconds, including latter 1 second (the 2nd second) of the suction of first flue gas, 18 seconds of the moon/spontaneous combustion,
2 seconds of balance, 2 seconds of second mouthful of flue gas suction, residual stub took out 15 seconds of operation.
(4b.3) first flue gas aspirates:It is that pure air is accurately marked in the 230.00-232.00 seconds in flue gas sampling cycle
Determine the state 20.00-22.00 seconds, under minipump swabbing action, flue gas is 1050 ml/mins with 17.5 milliliters/seconds
Flow, successively via the second two-position two-way solenoid valve, first throttle valve, flowmeter, be directly discharged to outdoor, continue 2 seconds.
(4b.4) the moon/spontaneous combustion:It is pure air Accurate Calibration state in the 232.00-250.00 seconds in flue gas sampling cycle
22.00-40.00 seconds, the second two-position two-way solenoid valve disconnect, and cigarette enters the moon/spontaneous combustion state, lasts 18 seconds.
(5) balance:In the 250.00-252.00 seconds in flue gas sampling cycle, all magnetic valves are in off-state, air-sensitive
Sensor array annular working intracavitary flows without gas, and cigarette lasts 2 seconds still in the moon/spontaneous combustion state.
(6) second mouthfuls of flue gas suctions:In the 252.00-254.00 seconds in flue gas sampling cycle, the first two-position two-way solenoid valve
Turned on the 5th two-position two-way solenoid valve, remaining four two-position two-way solenoid valve disconnects, and cigarette smoke is with 17.5 milliliters/seconds
The flow of 1050 ml/mins, pass sequentially through the first two-position two-way solenoid valve, gas sensor array annular working chamber, the 5th
Two-position two-way solenoid valve, first throttle valve, flowmeter, are finally discharged to outdoor, continue 2 seconds, gather 35 milliliters of flue gas.
(7) surrounding air rinses:In the 254.00-300.00 seconds in flue gas sampling cycle, room air is with 6500 ml/mins
The flow of clock flows through gas sensor array annular working chamber, is adhered to the cigarette of gas sensor sensitive membrane surface and inner-walls of duct
Gas scent molecule is tentatively washed away, and gas sensor array enters preliminary recovery state.Wherein,
(7.1) remain stub and take out operation:In the 254.00-269.00 seconds in flue gas sampling cycle, operating personnel were at 15 seconds
Residual stub is taken out in clock and is abandoned.During this period, the 3rd two-position two-way solenoid valve, the 4th two-position two-way solenoid valve, the five or two
Position two-way electromagnetic valve conducting, its excess-three two-position two-way solenoid valve disconnect, and indoor air is with the stream of 6500 ml/mins
Amount, pass sequentially through the 3rd two-position two-way solenoid valve, gas sensor array annular working chamber, the 5th two-position two-way solenoid valve, the
Four two-position two-way solenoid valves, are ultimately drained into outdoor, last 15 seconds.
(7.2) after remaining stub taking-up:In the 269.00-300.00 seconds in flue gas sampling cycle, the first bi-bit bi-pass electromagnetism
Valve, the 4th two-position two-way solenoid valve, the conducting of the 5th two-position two-way solenoid valve, its excess-three two-position two-way solenoid valve disconnect, ring
Border air flows through the first two-position two-way solenoid valve, gas sensor array annular working successively with the flow of 6500 ml/mins
Chamber, the 5th two-position two-way solenoid valve, the 4th two-position two-way solenoid valve, are ultimately drained into outdoor, last 31 seconds.The electricity in this stage
It is identical that magnet valve position and surrounding air flow condition tentatively recover state with gas sensor array.
(8) data record:From the 250.00th second flue gas sampling cycle, i.e., since the quarter poised state, computer
Voltage responsive caused by 16 gas sensors is stored in " xxx " text text by 16 passage, 16 high-accuracy data collection cards
In part, until the 290.00th second flue gas sampling cycle was surrounding air rinse stage the 36.00th second only, including second mouthful of flue gas is taken out
Inhale, this 3 processes after residual stub takes out operation, residual stub takes out, a length of 40 seconds during data record.
(9) feature extraction:Within a flue gas sampling cycle, computer is from 40 seconds " xxx " document data record of duration
In, each gas sensor voltage responsive stable state maximum is extracted as characteristic component, is substantially the sound to second mouthful of flue gas
Should, therefore tested tobacco product sample is converted into the measurement sample of one 16 dimension, and be stored in the tobacco of hard disc of computer with
In tobacco product sample data set file.
(10) identification and the prediction of aesthetic quality's index score:It is data in the 290.00-300.00 seconds in flue gas sampling cycle
In 10 seconds of record end ,-n ballot identification groups of the modular neural network cascade model first order are according to most of voting rules
Sample x brand, the place of production are determined with true and false, the modular neural network cascade model second level-ballot identification group is corresponding with winning
That score prediction group prediction x fragrance, coordination, miscellaneous gas, excitant, pleasant impression this 5 aesthetic quality's index score value, and lead to
Display is crossed to show.Repeat step (2)~(10), tobacco electronic nose instrument are realized to multiple tobaccos and tobacco product sample
Test, identification and the prediction of aesthetic quality's index score of flue gas.One complete tobacco is with the tobacco product sample testing cycle
300 seconds.
In data acquisition phase, tobacco electronic nose instrument test is judged by the group of smokeing panel test and provides quality index sense organ and obtains
The tobacco divided and tobacco product standard specimen.Single standard specimen measurement period is 5 minutes, extracts each gas sensor to the
As characteristic component, the voltage that one 16 dimension is obtained to p-th of standard specimen rings the voltage responsive stable state maximum of two mouthfuls of flue gases
Answer vector x 'p=(xp1' ..., xpi' ..., xp16’)T∈R16.Pass through the test to N number of standard specimen, tobacco electronic nose instrument
Obtain gas sensor array voltage responsive master sample collection X ' ∈ RN×16, and establish X ' and fragrance d1∈RN, coordinate d2∈RN、
Miscellaneous gas d3∈RN, excitant d4∈RN, pleasant impression d5∈RNThe one-to-one relationship of totally 5 quality index sensory scores.Each brand
A, B, C Three Estate respectively measure 10 standard specimens.If brand number is n, N=30n.
Gas sensor array response criteria sample set X ' all characteristic components through direct proportion preprocessing transformation to [0.0,
6.0] scope.If gas sensor i is x to the voltage responsive stable state maximum of p-th of standard specimenpi', the value after transformation of scale
For:
Here, max (X ') and min (X ') is respectively X ' maxima and minima, xpiIt is gas sensor i to p-th
Voltage responsive stable state maximum of the standard specimen after transformation of scale, voltage responsive vector x 'pTherefore it is changed into 16 dimension sample xp=
(xp1..., xpi..., xp16)T∈R16.Max (X ') and min (X ') is as master data deposit computer.Master sample collection X '
It is referred to as training set after direct proportion preprocessing transformation, is designated as X.
Identification and aesthetic quality's index score forecast period in sample x undetermined, gas sensor i voltage responsives stable state is most
Big value xpi' still use the max (X ') and min (X ') of master sample collection to carry out direct proportion conversion with formula (1).
It is the modular neural network cascade model first order-n (n-1)/2 single output nerve e-learning stage, right first
Training set X imposes one-to-one (one-against-one, OAO) decomposition, and X is broken down intoIndividual two classification (binary-
Class) training subset, then, this n (n-1)/2 subset are anti-using error by n (n-1)/2 single output nerve network respectively
Propagation algorithm is learnt one by one.All single output nerve network structures are single hidden layer, input number of nodes m=16, hidden section
Count as s1=8, output node number is 1.Target output is encoded using { 0.0,3.0 }, all hidden node and output node activation
Function is
For example, brand ωjAnd ωkThis two classification based trainings subset is Xjk={ Xj, Xk, by tobacco electronic nose instrument test
The two brand whole standard specimens and obtain, sample number Njk=Nj+Nk=60, by single output nerve networkStudy, learn
Habit step-length is ηjk=10/Njk=0.17.
To training sample xp=(xp1..., xpi..., xp16)T∈R16, single output nerve networkHidden node h reality
Export and be:
In formula,It is hidden node h to sample xpThe weighted sum of all input components:
Wherein,For threshold value, constant term xp0=+6.0.
To training sample xp, single output nerve networkReality output be:
In formula,For neutral netOutput node is to the weighted sums of all hidden node reality outputs, i.e.,:
Wherein,For threshold value, constant term
In modular neural network cascade model second level n × 5 single output nerve e-learning stage, training set X quilts
N training subset is resolved into, each training subset is made up of whole samples of a brand, and each score prediction group is by 5
Single output nerve network composition, be fitted respectively gas sensor array response through transformation of scale and corresponding brand fragrance, coordination,
Non-linear relation between this 5 aesthetic quality's index scores of miscellaneous gas, excitant, pleasant impression.Each single output nerve network structure
For single hidden layer, input number of nodes m=16, Hidden nodes s2=8, output node number is 1;All hidden nodes and output node
Activation functions are stillLearning Step is ηj=5/Nj=0.17, learning algorithm is still error back propagation algorithm.
For example, training subset XjOnly by from brand ωjWhole Nj=30 sample compositions, score prediction group Λj5
Individual single output nerve network is fitted X respectivelyjWith brand ωjFragrance, coordination, miscellaneous gas, excitant, pleasant impression this 5 aesthetic qualities refer to
Mark the non-linear relation between score.XjTarget output be brand ωjQuality index sensory scores arrived through transformation of scale
The scope of [0.15,2.85].
To from brand ωjSample xpIf r-th of quality index sensory scores isThen r-th of single output nerve
Target of the network after transformation of scale, which exports, is:
For brand ωjAll NjR-th of quality index sense of=30 standard specimens
Official's score vector.
Score prediction group ΛjThe hidden node h of r-th of single output nerve network reality output is:
In formula,It is hidden node h to sample xpThe weighted sum of all input components, i.e.,:
Wherein,For threshold value, constant term xp0=+6.0.
Score prediction group ΛjThe reality output of r-th of single output nerve network is
In formula,Weighted sum for r-th of single output nerve network output node to all hidden node reality outputs, i.e.,
Wherein,For threshold value, constant term
To the 1. modular neural network cascade model first order one that tobacco and tobacco product are identified per (n-1) individual list
Output nerve network forms a ballot identification group, represents a tobacco and tobacco product brand, highest number of votes obtained is n-1.Often
Individual single output nerve network and must only participate in wherein 2 ballot identification groups, therefore n (n-1)/2 single output nerve network divides
Not Zu Cheng n ballot identification group, and carry out decision-makings using most of ballot (majority vote) rules.
For example, single output nerve networkTwo jth, k ballot identification group Ω must be participated injAnd ΩkBallot.In jth
In group, ifReality output y(jk)> 1.5, then predict that sample x undetermined belongs to brand ωjPossibility assume to obtain 1 ticket.
In k groups, if y(jk)< 1.5, then predict that x belongs to brand ωkPossibility assume to obtain 1 ticket.
It is that x belongs to the product representated by that most ballot identification group of number of votes obtained to the sample x decision rules being identified
Board.If the two or more numbers of votes obtained launched of ballot identification groups are equal and are highest poll, decision-making:X is not belonging to existing
One brand.
To tobacco and tobacco product the modular neural network cascade model second level that 2. aesthetic quality's index score is predicted
One every 5 single output nerve networks form a score prediction group, be each responsible for the fragrance of the corresponding brand of prediction one, coordination,
This 5 aesthetic quality's index scores of miscellaneous gas, excitant, pleasant impression.The single output nerve network in n × 5 is divided into n score prediction group,
Corresponded with n ballot identification group.
In sample x 5 aesthetic quality's index score forecast periods, in the modular neural network cascade model first order
Identification group of voting ΩjOn the premise of number of votes obtained is most, the cascade model second level-represent brand ω is only neededjScore prediction group Λj
Prediction is participated in, other score prediction groups are not required to participate in.
If score prediction group ΛjThe reality output of r-th of single output nerve network is z(jr), then x belong to brand .j r-th
Aesthetic quality's index score predicted value is:
If on the basis of existing n kinds brand, increase identifies a kind of new brand, need only increase n single output nerve nets
Network simultaneously learns, and therefore, the modular neural network cascade model first order increases to n (n+1)/2 list from existing n (n-1)/2
Output nerve network.For example, to the brand ω newly increasedn+1, increasing single output nerve network that adduction learns newly is
Correspondingly, in order to newly increasing brand progress aesthetic quality's index score prediction, modular neural network level gang mould
The type second level newly increases 5 single output nerve networks and learnt, and increases to (n+1) × 5 from existing n × 5.False brand or
The existing same brand of another manufacturer production be seen as a kind of single brand be identified it is pre- with aesthetic quality's index score
Survey.
Brief description of the drawings
Fig. 1 is electronic nose instrument evaluation method-tobacco electronic nose instrument of the present invention-tobacco and tobacco product aesthetic quality
Fundamental diagram (second mouthful of flue gas aspiration phases).
Fig. 2 is the present invention-tobacco and the electronic nose instrument evaluation method of tobacco product aesthetic quality-in sectionThe sigmoid activation functions of amendmentWith the sigmoid activation functions of standardFirst-order partial derivative curve and its ratio change curve.(a)First-order partial derivativeCurve (solid line) withFirst-order partial derivativeCurve (dotted line);(b) first-order partial derivative
RatioCurve.
Fig. 3 is electronic nose instrument evaluation method -2 kinds of identification cigarette product of the present invention-tobacco and tobacco product aesthetic quality
Board { ωj, ωkSingle output nerve mixed-media network modules mixed-media
Fig. 4 is the present invention, and-tobacco and the electronic nose instrument evaluation method of tobacco product aesthetic quality-determines an examination undetermined
During the brand of sample, n (n-1)/2 single output nerve network is divided into the situation of n ballot identification group ballot.
Fig. 5 is the present invention-tobacco and the electronic nose instrument evaluation method of tobacco product aesthetic quality-by 5 single output god
The aesthetic quality's index score prediction group Λ formed through mixed-media network modules mixed-mediaj。
Fig. 6 is the present invention-tobacco and the electronic nose instrument evaluation method of tobacco product aesthetic quality-to a tobacco and cigarette
Straw-made articles sample carries out 1. identification and 2. the modular neural network cascade model decision-making of aesthetic quality's index score prediction simultaneously
Process schematic.
Embodiment
The present invention is described in further detail below in conjunction with the accompanying drawings.
Fig. 1 is electronic nose instrument evaluation method-tobacco electronic nose instrument of the present invention-tobacco and tobacco product aesthetic quality
Fundamental diagram, gas circuit and the position of magnetic valve now are second mouthful of flue gas suction operation state.Tobacco electronic nose instrument bag
Include gas sensor array module I, flue gas automatic sample handling system II, computer control and data analysis system III, automatic ignition
Device IV and tested cigarette sample V.
The main component units of gas sensor array module I include:Gas sensor array I-1, gas sensor array
Annular working chamber I-2, heat-insulation layer I-3 and resistance heating wire and fan, positioned at tobacco electronic nose instrument upper right quarter.Wherein, gas
Dependent sensor array is mainly made up of serial 16 gas sensors of TGS800 and TGS2000, available model include TGS800,
TGS813、TGS816、TGS821、TGS822、TGS823、TGS826、TGS830、TGS832、TGS2600、TGS2602、
TGS2603, TGS2610, TGS2611, TGS2612, TGS2620, and TGS3830.The effect of gas sensor array module I
It is the analog voltage signal that the cigarette smoke of complicated ingredient is converted into 0~10V.
Flue gas automatic sample handling system II component units include:First two-position two-way solenoid valve II-1, the second bi-bit bi-pass electricity
Magnet valve II-2, the 3rd two-position two-way solenoid valve II-3, the 4th two-position two-way solenoid valve II-4, the 5th two-position two-way solenoid valve II-
5, the 6th two-position two-way solenoid valve II-6, cigarette clamper (being commonly called as cigarette holder) II-7, minipump II-8, first throttle valve
II-9, flowmeter II-10, second throttle II-11, pure air II-12, ashtray II-13, and overflow flue gas discharge dress
Put II-14.
It is computer motherboard III-1 that computer, which is controlled with the main component units of data analysis system III, 16 circuit-switched datas gather
Block III-2, driving and control module III-3, multi-channel DC voltage module III-4, display III-5, and hard disk, network interface card,
Video card, mouse, keyboard etc., positioned at tobacco electronic nose instrument left part.Computer controls the main function with data analysis system III
It is the collection, analysis and processing of (1) gas sensor array response signal;(2) multiple two of flue gas automatic sample handling system II
Two-way electromagnetic valve and minipump, automatic ignition device IV and computer control and data analysis system III itself driving
With control.
Cigarette sample V automatic ignitions and flue gas automatic sample handling system II suction first smoke behavior, the two or two two
Three-way electromagnetic valve II-2 is turned on, the first two-position two-way solenoid valve II-1, the 4th two-position two-way solenoid valve II-4 and the 5th bi-bit bi-pass
Electromagnetic valve II -5 disconnects.In automatic ignition state, automatic ignition device IV ignition coil IV-1 is powered, and magnet coil IV-4 breaks
Electricity, valve element IV-3 levels in the presence of compression spring IV-6 are moved to the left 9 millimeters so that ignition coil IV-1 and tested cigarette
Sample V is contacted, 5 seconds duration.Then, gas sensor is not passed through completely in first flue gas aspiration phases, cigarette smoke
Array module I, but under minipump II-8 swabbing action, via the second two-position two-way solenoid valve II-2, first segment
Valve II-9, flowmeter II-10 are flowed, outdoor is directly discharged to the flow that 17.5 milliliters/seconds are 1050 ml/mins, continues 2
Second, equivalent to 35 milliliters of drawing smoke.
While cigarette sample V automatic ignitions and flue gas automatic sample handling system II suction first cigarettes, gas sensor
Array I-1 be in pure air demarcation state, two-position two-way solenoid valve II-3 and II-6 conducting, pure air with 1050 milliliters/
Minute flow flow successively through second throttle II-11, the 6th two-position two-way solenoid valve II-6, gas sensor array module I,
3rd two-position two-way solenoid valve II-3, is finally discharged in room atmosphere.This is also that pure air demarcated for the 2nd stage, continues 7
Second.
When aspirating second mouthful of flue gas, the first two-position two-way solenoid valve II-1 and the 5th two-position two-way solenoid valve II-5 are led
Logical, remaining four two-position two-way solenoid valve disconnects, cigarette smoke with the flow that 17.5 milliliters/seconds are 1050 ml/mins, according to
It is secondary to pass through the first two-position two-way solenoid valve II-1, gas sensor array I-1, the 5th two-position two-way solenoid valve II-5, first segment
Valve II-9, flowmeter II-10 are flowed, outdoor is finally discharged to, continues 2 seconds, equivalent to 35 milliliters of flue gas of collection.
In second mouthful of cigarette smoke flow process, gas sensor array I-1 productions Upi(max) raw sensitive response.To one
Individual specific tobacco and tobacco product sample p, single gas sensor i voltage response curves stable state maximum are extracted as
Characteristic component xpi', i.e. xpi'=Upi(max), therefore the array I-1 of 16 gas sensor compositions produces the voltage of one 16 dimension
Response vector xp'=(xp1', xp2' ..., xpi' ..., xp16’)T∈R16.This 16 dimension voltage responsive stable state maximum to
Amount, which is tobacco electronic nose instrument, tobacco and tobacco product to be identified and the foundation of aesthetic quality's metrics evaluation.
The one cigarette sample V flue gas sampling cycle is 300 seconds, and data record time span therein is 40 seconds.From flat
From the quarter that weighing apparatus state starts, i.e. the 250th second flue gas sampling cycle starts, computer control and data analysis system III record gas
Dependent sensor array transient voltage response data, a length of 40 seconds during record, including equilibrium stage 2 seconds, 2 seconds flue gas sampling stages,
36 seconds before surrounding air rinse stage, voltage responsives of the gas sensor array I-1 to cigarette smoke is that sampled data is stored in
In one text.In 40 second data record lengths, voltage responsives of the gas sensor i to cigarette sample p flue gas is steady
State maximum is as characteristic component xpi', thus obtain sound of the gas sensor array to p-th of cigarette sample, second mouthful of flue gas
Should, referred to as voltage responsive sample xp’∈R16.In 10 seconds after data record terminates, computer control and data analysis
System III is according to sample xp' provide sample p brand recognition result and fragrance, coordination, miscellaneous gas, excitant, the sense of this 5, pleasant impression
Official's quality index prediction result.
Why cigarette smoke sampling flow is set to 17.5 milliliters/seconds i.e. 1050 ml/mins and sampling duration is set to 2
SecondA large amount of statistical results point out that people takes a pull at flue gas volume as 35 milliliters, average duration 2 seconds.National standard《Conventional analysis is used
Smoking machine one defines and standard conditions》GB/T 16450-2004 provide that cigarette smoke standard aspiration capacity is 35 milliliters, flow
For 17.5 milliliters/seconds, the single port puff duration of standard is 2.00 ± 0.02 seconds.The cigarette smoke sampling flow of the present invention,
Sampling volume, sampling time are consistent with national standard GB/T 16450-2004.For the sake of unification, pure air flow is still defined as
1050 ml/mins.
Why useThe sigmoid activation functions of this amendmentBecause neural networks with single hidden layer uses
Error back propagation algorithm, on the premise of not vibrating, the error sum of squares between neutral net reality output and desired output
Function is bigger on the partial derivative of weights and threshold parameter (First-order Gradient), and neural network learning speed is faster.Fig. 2 gives
SectionThe sigmoid activation functions of amendmentWith standard Sigmoid activation functionsFirst-order partial derivative and its ratio curve.Fig. 2 (a) is that single order of the two activation functions in this section is inclined
DerivativeWithCurve, solid line are the activation functions of amendmentFirst-order partial derivativeIt is bent
Line, dotted line are standard Sigmoid activation functionsFirst-order partial derivativeCurve.Fig. 2 (b) is amendment
The sigmoid activation functions with standard first-order partial derivative ratio Curve.Fig. 2 shows, activation functionsFirst-order Gradient more than standard Sigmoid activation functionsIt is big.WhenWhen, the two
First-order Gradient ratio reaches p=732.63.Matched, input component transforms to [0,6] scope.The consideration of this way is,
Regardless of sample distribution state, each component averages of training dataset X are near 3.0.
Using standard Sigmoid activation functionsNeutral net typically by data set transformation to [0,1] model
Enclose, the actual acquiescence component average of this way is about 0.5.By comparison, inputting the advantages of component transforms to [0,6] scope is,
More original [0,1] is spaced between sample interval and class and is exaggerated 6 times, is advantageous to neutral net and adds on the premise of not vibrating
Fast pace of learning, improve study precision and improve Generalization Ability.
Fig. 3 is two kinds of cigarette brand { ω of identificationj, ωkSingle output neural networks with single hidden layer moduleStructural representation.
The neural network moduleThere are m=16 input node, s1=8 hidden nodes, 1 output node;All input component direct ratios
Example transforms to [0,6] scope, and target output is encoded using { 0.0,3.0 };All hidden node and output node activation functions areNeural network moduleLearnt using error back propagation algorithm, Studying factors ηjk=10/Njk=
0.17。
The present invention uses one-to-one (one-against-one, OAO) task analytic approach.Tobacco training set X is broken down intoIndividual two classification (binary-class) subproblem.This n (n-1)/2 subproblem is defeated by n (n-1)/2 list respectively
Go out neural network module to learn and solve one by one, thus the comprising modules neutral net cascade model first order.Table 2 gives ginseng
Add n (n-1)/2 single output of study
Table 2, for n tobacco of identification and tobacco product brand, n (n-1)/2 single output nerve network for participating in study arranges
Table
Neutral net list
For sample x undetermined brand, n (n-1)/2 single output god of the modular neural network cascade model first order
Most of voting rule decision-makings are used through network.Table 3 gives to grow tobacco and tobacco product to determine that sample x undetermined belongs to n
During which kind of brand, n (n-1)/2 single output nerve network packet list taking part in a vote.Table 3 is that master is diagonal symmetrical, one
Single output nerve mixed-media network modules mixed-media must participate in 2 ballot groups.For example, single output nerve mixed-media network modules mixed-mediaBoth taken part in a vote identification group
ΩjBallot, identification group of also taking part in a vote ΩkBallot.When Fig. 4 show in particular the brand for determining cigarette sample x undetermined, n
(n-1)/2 single output nerve network is divided into the situation that n ballot identification group is taken part in a vote, and each group has n-1 member.
Score prediction group shares n, is corresponded with ballot identification group, each score prediction group is by 5 single output nerves
Network forms.Fig. 5 is the quality index score prediction group Λ being made up of 5 single output nerve networksj, 5 members predict respectively
Belong to brand ωjCigarette sample x fragrance, coordination, miscellaneous gas, excitant, pleasant impression this 5 aesthetic quality's index scores.Each
Single output nerve network structure is,
Table 3, when determining that n is grown tobacco with tobacco product brand, n (n-1)/2 single output nerve network taken part in a vote divides
Groups List
M=16 input node, s2=8 hidden nodes, 1 output node;It is all input component direct proportions transform to [0,
6] scope, target output are encoded using { 0.0,3.0 };All hidden node and output node activation functions are
Target output is brand ωjScope of the quality index sensory scores through transformation of scale to [0.15,2.85].Learning algorithm is mistake
Poor back propagation algorithm, Learning Step ηj=0.17.
Fig. 6 is the brand and prediction aesthetic quality's index score that modular neural network cascade model determines tobacco sample x
Process schematic.When carrying out brand, the place of production, true and false identification to tobacco sample x, n (n-1)/2 single output nerve of the first order
Network full entry, is divided into n ballot identification group, every group of n-1 member, and a single output nerve network participates in two therein
Ballot identification group.N ballot identification group uses most of voting rule decision-makings, and who gets the most votes, which organizes, to win.When two or more
The draw in votes for identification group of voting and when being most, decision rule is:Sample x is not belonging to any brand of existing database.
When the prediction of quality index score is carried out to tobacco sample x, in 5n single output nerve networks of the second level, only thrown with winning
5 participation of score prediction group corresponding to ticket identification group, respectively prediction belong to the perfume (or spice) for the brand representated by ballot identification group of winning
This 5 quality index scores of gas, coordination, miscellaneous gas, excitant and pleasant impression.
Claims (7)
1. tobacco and the electronic nose instrument evaluation method of tobacco product aesthetic quality, it is characterized in that, tobacco electronic nose instrument includes
Gas sensor array module, flue gas automatic sample handling system, computer control and data analysis system, automatic ignition device, it is real
Existing tobacco and tobacco product identification and the prediction of aesthetic quality's index score;
Described tobacco electronic nose instrument is 5 minutes to the flue gas sampling cycle of a tobacco and tobacco product sample, and air-sensitive passes
Sensor array recovers (210 seconds), pure air Accurate Calibration (40 seconds), balance (2 seconds), second mouthful of flue gas suction (2 after preliminary
Second), surrounding air rinse (46 seconds) totally 6 stages;
During flue gas sampling, under the control of the computer, flue gas automatic sample handling system is 1050 ml/mins with 17.5 milliliters/seconds
Flow aspirates second mouthful of flue gas, passes through gas sensor array annular working chamber, skims over gas sensor sensitive membrane surface,
Continue 2 seconds, gather 35 milliliters of flue gas, therefore gas sensor array produces sensitive response;From the quarter that self-balancing state starts, meter
Calculation machine controls and data analysis system start recording response data;Successively record balance (2 seconds), second mouthful of flue gas suction (2 seconds),
Surrounding air rinses the gas sensor array voltage responsive value in (first 36 seconds) this 3 stages, total duration 40 seconds;Flue gas sampling week
The data of other time phase do not record;
Within the data record time of 40 seconds, voltage response curves stable state maximum of the single gas sensor to second mouthful of flue gas
Characteristic component is extracted as, therefore the array of 16 gas sensor compositions produces one 16 dimension voltage responsive vector;In data
In 10 seconds after record end, computer control is tried tobacco and tobacco product with data analysis system according to this response vector
Sample carries out brand, the place of production, true and false identification and fragrance, coordination, miscellaneous gas, excitant, pleasant impression, and totally 5 aesthetic quality's index scores are pre-
Survey.
Described computer control is with data analysis system using modular neural network cascade model to tobacco and tobacco product
Sample carries out 1. identification and 2. aesthetic quality's index score is predicted;1. the modular neural network cascade model first order is by n (n-
1)/2 a single output nerve network forms side by side, forms n ballot identification group, grows tobacco the identification with tobacco product, wrap for n
Include brand, the place of production and true and false identification;2. the modular neural network cascade model second level is by the single output nerve network in n × 5
Composition side by side, every 5 one group, grown tobacco and the fragrance of tobacco product, coordination, miscellaneous gas, excitant, pleasant impression this 5 sense organs for n
Quality index score is predicted;Each single output nerve network undergoes the study stage to training set X and treats this x of random sample knowledge
Not with aesthetic quality's index score forecast period.
Identification and aesthetic quality index score prediction of the tobacco electronic nose instrument to tobacco and tobacco product sample undetermined, including with
Lower step:
(1) start shooting:Instrument preheats 30 minutes, and surrounding air flows through the first bi-bit bi-pass successively with the flow of 6500 ml/mins
Magnetic valve, gas sensor array annular working chamber, the 5th two-position two-way solenoid valve, the 4th two-position two-way solenoid valve, final row
Go out to outdoor;Gas sensor array constant temperature indoor temperature reaches constant 55 ± 0.1 DEG C from room temperature;
(2) the flue gas sampling cycle starts:Operating personnel click on " starting to detect " button of screen drop-down menu, and instrument, which enters, to last
The flue gas sampling cycle of 5 minutes, computer automatically generate the text of one entitled " xxx " in specified folder, with record
Response data of the gas sensor array to flue gas;
(3) it is preliminary to recover:In the 0.00-210.00 seconds in flue gas sampling cycle, surrounding air with the flow of 6500 ml/mins according to
It is secondary flow through the first two-position two-way solenoid valve, gas sensor array annular working chamber, the 5th two-position two-way solenoid valve, the four or two
Two-way electromagnetic valve, it is ultimately drained into outdoor;In the case where flow is the surrounding air flushing action of 6500 ml/mins, gas is adhered to
The flue gas scent molecule of dependent sensor sensitive membrane surface and inner-walls of duct is tentatively washed away, and gas sensor array tentatively returns to
Normal condition, last 210 seconds;
(4) pure air Accurate Calibration:In the 210.00-250.00 seconds in flue gas sampling cycle, (a) pure air Accurate Calibration and
(b) cigarette insertion, automatic ignition, the suction of first flue gas, the moon/spontaneous combustion the two links are carried out simultaneously, last 40 seconds:
(4a), pure air Accurate Calibration:Pure air is flowed successively through with the flow that 17.5 milliliters/seconds are 1050 ml/mins
Second throttle, the 6th two-position two-way solenoid valve, gas sensor array annular working chamber, the 3rd two-position two-way solenoid valve, most
After be discharged to interior, last 40 seconds;Pure air makes gas sensor array Exact recovery to normal condition;
(4b.1) cigarette inserts:Be in the 210.00-225.00 seconds in flue gas sampling cycle pure air Accurate Calibration state most
In first 15 seconds, tested cigarette sample filter end is inserted cigarette clamper by screen display " cigarette insertion " printed words, operating personnel, is inserted
Enter depth for 9.0 ± 0.5 millimeters;
(4b.2) automatic ignition:It is pure air Accurate Calibration state in the 225.00-231.00 seconds in flue gas sampling cycle
15.00-21.00 seconds, ignition coil are powered;At the same time, ignition head, which is moved to the left 9 millimeters, makes ignition coil be connect with tested sample
Touch and light sample, continue 6 seconds;In the 231.00-269.00 seconds in flue gas sampling cycle, magnet coil is powered, and ignition coil breaks
Electricity is simultaneously returned to reference position, lasts 38 seconds, including latter 1 second (the 2nd second) of the suction of first flue gas, 18 seconds of the moon/spontaneous combustion, it is flat
2 seconds of weighing apparatus, 2 seconds of second mouthful of flue gas suction, residual stub took out 15 seconds of operation;
(4b.3) first flue gas aspirates:It is pure air Accurate Calibration shape in the 230.00-232.00 seconds in flue gas sampling cycle
The state 20.00-22.00 seconds, under minipump swabbing action, flue gas is with stream that 17.5 milliliters/seconds are 1050 ml/mins
Amount, successively via the second two-position two-way solenoid valve, first throttle valve, flowmeter, is directly discharged to outdoor, continues 2 seconds;
(4b.4) the moon/spontaneous combustion:It is pure air Accurate Calibration state in the 232.00-250.00 seconds in flue gas sampling cycle
22.00-40.00 seconds, the second two-position two-way solenoid valve disconnect, and cigarette enters the moon/spontaneous combustion state, lasts 18 seconds;
(5) balance:In the 250.00-252.00 seconds in flue gas sampling cycle, all magnetic valves are in off-state, gas sensing
Device array annular working intracavitary flows without gas, and cigarette lasts 2 seconds still in the moon/spontaneous combustion state;
(6) second mouthfuls of flue gas suctions:In the 252.00-254.00 seconds in flue gas sampling cycle, the first two-position two-way solenoid valve and
Five two-position two-way solenoid valves turn on, and remaining four two-position two-way solenoid valve disconnects, and cigarette smoke is 1050 with 17.5 milliliters/seconds
The flow of ml/min, pass sequentially through the first two-position two-way solenoid valve, gas sensor array annular working chamber, the five or two
Two-way electromagnetic valve, first throttle valve, flowmeter, are finally discharged to outdoor, continue 2 seconds, gather 35 milliliters of flue gas;
(7) surrounding air rinses:In the 254.00-300.00 seconds in flue gas sampling cycle, room air is with 6500 ml/mins
Flow flows through gas sensor array annular working chamber, is adhered to the flue gas gas of gas sensor sensitive membrane surface and inner-walls of duct
Taste molecule is tentatively washed away, and gas sensor array enters preliminary recovery state;Wherein,
(7.1) remain stub and take out operation:In the 254.00-269.00 seconds in flue gas sampling cycle, operating personnel are in 15 seconds
Take out residual stub and abandon;During this period, the 3rd two-position two-way solenoid valve, the 4th two-position two-way solenoid valve, the five or two two
Three-way electromagnetic valve turns on, and its excess-three two-position two-way solenoid valve disconnects, indoor air with the flow of 6500 ml/mins,
Pass sequentially through the 3rd two-position two-way solenoid valve, gas sensor array annular working chamber, the 5th two-position two-way solenoid valve, the four or two
Position two-way electromagnetic valve, is ultimately drained into outdoor, lasts 15 seconds;
(7.2) after remaining stub taking-up:In the 269.00-300.00 seconds in flue gas sampling cycle, the first two-position two-way solenoid valve,
Four two-position two-way solenoid valves, the conducting of the 5th two-position two-way solenoid valve, its excess-three two-position two-way solenoid valve disconnect, surrounding air
First two-position two-way solenoid valve, gas sensor array annular working chamber, the 5th are flowed through with the flow of 6500 ml/mins successively
Two-position two-way solenoid valve, the 4th two-position two-way solenoid valve, are ultimately drained into outdoor, last 31 seconds;The magnetic valve position in this stage
Put that tentatively to recover state with surrounding air flow condition with gas sensor array identical;
(8) data record;From the 250.00th second flue gas sampling cycle, i.e., since the quarter poised state, computer passed through
Voltage responsive caused by 16 gas sensors is stored in " xxx " text by 16 16, passage high-accuracy data collection cards
In, stop within the 36.00th second until the 290.00th second flue gas sampling cycle was surrounding air rinse stage, including second mouthful of flue gas is taken out
Inhale, this 3 processes after residual stub takes out operation, residual stub takes out, a length of 40 seconds during data record;
(9) feature extraction;Within a flue gas sampling cycle, computer carries in 40 seconds " xxx " document data record of duration
Each gas sensor voltage responsive stable state maximum is taken as characteristic component, is substantially the response to second mouthful of flue gas, one
Therefore individual tested tobacco product sample is converted into the measurement sample of one 16 dimension, and be stored in tobacco and the tobacco system of hard disc of computer
In product sample data set file;
(10) identification and the prediction of aesthetic quality's index score;It is data record in the 290.00-300.00 seconds in flue gas sampling cycle
In 10 seconds terminated ,-n ballot identification groups of the modular neural network cascade model first order determine according to most of voting rules
Sample x brand, the place of production with true and false, the modular neural network cascade model second level one it is corresponding with ballot identification group of winning that
Individual score prediction group prediction x fragrance, coordination, miscellaneous gas, excitant, pleasant impression this 5 aesthetic quality's index score value, and by aobvious
Show that device is shown;
Repeat step (2)~(10), tobacco electronic nose instrument realize test, the knowledge to multiple tobaccos and tobacco product sample flue gas
Do not predicted with aesthetic quality's index score;One complete tobacco is 300 seconds with the tobacco product sample testing cycle.
2. tobacco according to claim 1 and the electronic nose instrument evaluation method of tobacco product aesthetic quality, it is characterized in that,
In data acquisition phase, tobacco electronic nose instrument test is judged by the group of smokeing panel test and provides the tobacco of quality index sensory scores
With tobacco product standard specimen;Single standard specimen measurement period is 5 minutes, extracts each gas sensor to second mouthful of flue gas
Voltage responsive stable state maximum as characteristic component, p-th of standard specimen is obtained the voltage responsive vector x of one 16 dimension 'p
=(xp1' ..., xpi' ..., xp16’)T∈R16;By the test to N number of standard specimen, tobacco electronic nose instrument obtains air-sensitive
Sensor array voltage responsive master sample collection X ' ∈ RN×16, and establish X ' and fragrance d1∈RN, coordinate d2∈RN, miscellaneous gas d3∈
RN, excitant d4∈RN, pleasant impression d5∈RNThe one-to-one relationship of totally 5 quality index sensory scores;Each brand A, B, C tri-
Grade respectively measures 10 standard specimens;If brand number is n, N=30n.
3. tobacco according to claim 1 and the electronic nose instrument evaluation method of tobacco product aesthetic quality, it is characterized in that,
Gas sensor array response criteria sample set X ' all characteristic components are through direct proportion preprocessing transformation to [0.0,6.0] model
Enclose;If gas sensor i is x to the voltage responsive stable state maximum of p-th of standard specimenpi', the value after transformation of scale is:
<mrow>
<msub>
<mi>x</mi>
<mrow>
<mi>p</mi>
<mi>i</mi>
</mrow>
</msub>
<mo>=</mo>
<mn>6</mn>
<mo>&times;</mo>
<mfrac>
<mrow>
<msup>
<msub>
<mi>x</mi>
<mrow>
<mi>p</mi>
<mi>i</mi>
</mrow>
</msub>
<mo>&prime;</mo>
</msup>
<mo>-</mo>
<mi>min</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>X</mi>
<mo>&prime;</mo>
</msup>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<mi>m</mi>
<mi>a</mi>
<mi>x</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>X</mi>
<mo>&prime;</mo>
</msup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mi>min</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>X</mi>
<mo>&prime;</mo>
</msup>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
Here, max (X ') and min (X ') is respectively X ' maxima and minima, xpiIt is gas sensor i to p-th of standard
Voltage responsive stable state maximum of the sample after transformation of scale, voltage responsive vector x 'pTherefore it is changed into 16 dimension sample xp=
(xp1, ..., xpi..., xp16)T∈R16;Max (X ') and min (X ') is as master data deposit computer;Master sample collection
X ' is referred to as training set after direct proportion preprocessing transformation, is designated as X;
Identification and aesthetic quality's index score forecast period in sample x undetermined, gas sensor i voltage responsive stable state maximums
xpi' still use the max (X ') and min (X ') of master sample collection to carry out direct proportion conversion with formula (1).
4. tobacco according to claim 1 and the electronic nose instrument evaluation method of tobacco product aesthetic quality, it is characterized in that,
In the modular neural network cascade model first order-n (n-1)/2 single output nerve e-learning stage, training set X is applied first
Decomposed with one-to-one (one-against-one, OAO), X is broken down intoIndividual two classification (binary-class) instruction
Practice subset;Then, this n (n-1)/2 subset uses error back propagation algorithm one by n (n-1)/2 single output nerve network respectively
One is learnt;All single output nerve network structures are single hidden layer, input number of nodes m=16, Hidden nodes s1
=8, output node number is 1;Target output is encoded using { 0.0,3.0 }, and all hidden node and output node activation functions are
For example, brand ωjAnd ωkThis two classification based trainings subset is Xjk={ Xj, Xk, by tobacco electronic nose instrument test this two
Individual brand whole standard specimen and obtain, sample number Njk=Nj+Nk=60, by single output nerve networkStudy, study step
A length of ηjk=10/Njk=0.17;
To training sample xp=(xp1, ..., xpi..., xp16)T∈R16, single output nerve networkHidden node h reality output
For:
In formula,It is hidden node h to sample xpThe weighted sum of all input components:
Wherein,For threshold value, constant term xp0=+6.0;
To training sample xp, single output nerve networkReality output be:
<mrow>
<msubsup>
<mi>y</mi>
<mi>p</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mi>&phi;</mi>
<mi>p</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>j</mi>
<mi>k</mi>
</mrow>
<mo>)</mo>
</mrow>
</msubsup>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<mn>3</mn>
<mrow>
<mn>1</mn>
<mo>+</mo>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<msubsup>
<mi>&phi;</mi>
<mi>p</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>j</mi>
<mi>k</mi>
</mrow>
<mo>)</mo>
</mrow>
</msubsup>
<mo>/</mo>
<mn>3</mn>
</mrow>
</msup>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>4</mn>
<mo>)</mo>
</mrow>
</mrow>
In formula,For neutral netOutput node is to the weighted sums of all hidden node reality outputs, i.e.,:
<mrow>
<msubsup>
<mi>&phi;</mi>
<mi>p</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>j</mi>
<mi>k</mi>
</mrow>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<msup>
<mi>&theta;</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>+</mo>
<msubsup>
<mi>w</mi>
<mn>1</mn>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<msubsup>
<mi>&xi;</mi>
<mrow>
<mi>p</mi>
<mn>1</mn>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>+</mo>
<mo>...</mo>
<mo>+</mo>
<msubsup>
<mi>w</mi>
<mn>1</mn>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<msubsup>
<mi>&xi;</mi>
<mrow>
<mi>p</mi>
<mi>h</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>+</mo>
<mo>...</mo>
<mo>+</mo>
<msubsup>
<mi>w</mi>
<mn>8</mn>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<msubsup>
<mi>&xi;</mi>
<mrow>
<mi>p</mi>
<mn>8</mn>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>h</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mn>8</mn>
</munderover>
<msubsup>
<mi>w</mi>
<mi>h</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<msubsup>
<mi>&xi;</mi>
<mrow>
<mi>p</mi>
<mi>h</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>5</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein,For threshold value, constant term
5. tobacco according to claim 1 and the electronic nose instrument evaluation method of tobacco product aesthetic quality, it is characterized in that,
In modular neural network cascade model second level n × 5 single output nerve e-learning stage, training set X is broken down into n
Training subset;Each training subset is made up of whole samples of a brand, and each score prediction group is by 5 single output nerves
Network forms, and is fitted gas sensor array response and corresponding brand fragrance, coordination, miscellaneous gas, stimulation through transformation of scale respectively
Non-linear relation between property, pleasant impression this 5 aesthetic quality's index scores;Each single output nerve network structure is single hidden layer
, input number of nodes m=16, Hidden nodes s2=8, output node number is 1;All hidden node and output node activation functions
Still it isLearning Step is ηj=5/Nj=0.17, learning algorithm is still error back propagation algorithm;
For example, training subset XjOnly by from brand ωjWhole Nj=30 sample compositions, score prediction group Λj5 lists it is defeated
Go out neutral net and be fitted X respectivelyjWith brand ωjFragrance, coordination, miscellaneous gas, excitant, pleasant impression this 5 aesthetic quality's index scores
Between non-linear relation;XjTarget output be brand ωjQuality index sensory scores through transformation of scale to [0.15,
2.85] scope;
To from brand ωjSample xpIf r-th of quality index sensory scores isThen r-th of single output nerve network
Target after transformation of scale, which exports, is:
<mrow>
<msubsup>
<mi>d</mi>
<mi>p</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>j</mi>
<mi>r</mi>
</mrow>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<mn>0.15</mn>
<mo>+</mo>
<mn>2.70</mn>
<mo>&times;</mo>
<mfrac>
<mrow>
<msubsup>
<mi>d</mi>
<mi>p</mi>
<mrow>
<mrow>
<mo>(</mo>
<mrow>
<mi>j</mi>
<mi>r</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>&prime;</mo>
</mrow>
</msubsup>
<mo>-</mo>
<mi>min</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>d</mi>
<mrow>
<mrow>
<mo>(</mo>
<mrow>
<mi>j</mi>
<mi>r</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>&prime;</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<mi>max</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>d</mi>
<mrow>
<mrow>
<mo>(</mo>
<mrow>
<mi>j</mi>
<mi>r</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>&prime;</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mi>min</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>d</mi>
<mrow>
<mrow>
<mo>(</mo>
<mrow>
<mi>j</mi>
<mi>r</mi>
</mrow>
<mo>)</mo>
</mrow>
<mo>&prime;</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>6</mn>
<mo>)</mo>
</mrow>
</mrow>
For brand ωjAll NjR-th of quality index sense organ of=30 standard specimens obtains
Divide vector.
Score prediction group ΛjThe hidden node h of r-th of single output nerve network reality output is
In formula,It is hidden node h to sample xpThe weighted sum of all input components, i.e.,:
Wherein,For threshold value, constant term xp0=+6.0;
Score prediction group ΛjThe reality output of r-th of single output nerve network is
<mrow>
<msubsup>
<mi>z</mi>
<mi>p</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mi>&phi;</mi>
<mi>p</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>j</mi>
<mi>r</mi>
</mrow>
<mo>)</mo>
</mrow>
</msubsup>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<mn>3</mn>
<mrow>
<mn>1</mn>
<mo>+</mo>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<msubsup>
<mi>&phi;</mi>
<mi>p</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>j</mi>
<mi>r</mi>
</mrow>
<mo>)</mo>
</mrow>
</msubsup>
<mo>/</mo>
<mn>3</mn>
</mrow>
</msup>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>9</mn>
<mo>)</mo>
</mrow>
</mrow>
In formula,Weighted sum for r-th of single output nerve network output node to all hidden node reality outputs, i.e.,
<mrow>
<msubsup>
<mi>&phi;</mi>
<mi>p</mi>
<mrow>
<mo>(</mo>
<mrow>
<mi>j</mi>
<mi>r</mi>
</mrow>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<msup>
<mi>&theta;</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>+</mo>
<msubsup>
<mi>w</mi>
<mn>1</mn>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
</mrow>
</msubsup>
<msubsup>
<mi>&zeta;</mi>
<mrow>
<mi>p</mi>
<mn>1</mn>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>+</mo>
<mo>...</mo>
<mo>+</mo>
<msubsup>
<mi>w</mi>
<mn>1</mn>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
</mrow>
</msubsup>
<msubsup>
<mi>&zeta;</mi>
<mrow>
<mi>p</mi>
<mi>h</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>+</mo>
<mo>...</mo>
<mo>+</mo>
<msubsup>
<mi>w</mi>
<mn>5</mn>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
</mrow>
</msubsup>
<msubsup>
<mi>&zeta;</mi>
<mrow>
<mi>p</mi>
<mn>5</mn>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>h</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mn>5</mn>
</munderover>
<msubsup>
<mi>w</mi>
<mi>h</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
</mrow>
</msubsup>
<msubsup>
<mi>&zeta;</mi>
<mrow>
<mi>p</mi>
<mi>h</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>10</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein,For threshold value, constant term
6. tobacco according to claim 1 and the electronic nose instrument evaluation method of tobacco product aesthetic quality, it is characterized in that,
To the 1. modular neural network cascade model first order one that tobacco and tobacco product are identified per (n-1) individual single output nerve
Network forms a ballot identification group, represents a tobacco and tobacco product brand, highest number of votes obtained is n-1;Each single output
Neutral net and must only participate in wherein 2 ballot identification groups, therefore n (n-1)/2 single output nerve network separately constitutes n
Ballot identification group, and using most of ballot (majority vote) rule progress decision-makings;
For example, single output nerve networkTwo jth, k ballot identification group Ω must be participated injAnd ΩkBallot;In jth group,
IfReality output y(jk)> 1.5, then predict that sample x undetermined belongs to brand ωjPossibility assume to obtain 1 ticket;In kth group
In, if y(jk)< 1.5, then predict that x belongs to brand ωkPossibility assume to obtain 1 ticket;
It is that x belongs to the brand representated by that most ballot identification group of number of votes obtained to the sample x decision rules being identified;
If the two or more numbers of votes obtained launched of ballot identification groups are equal and are highest poll, decision-making:X is not belonging to existing any product
Board;
To tobacco and tobacco product the modular neural network cascade model second level one every 5 that 2. aesthetic quality's index score is predicted
Individual single output nerve network forms a score prediction group, be each responsible for the fragrance of the corresponding brand of prediction one, coordination, miscellaneous gas,
This 5 aesthetic quality's index scores of excitant, pleasant impression;The single output nerve network in n × 5 is divided into n score prediction group, with n
Identification group of voting corresponds;
Obtained stage by stage in forecast sample x 5 aesthetic quality's indexs, in the ballot of the modular neural network cascade model first order
Identification group ΩjOn the premise of number of votes obtained is most, the cascade model second level one is only needed to represent brand ωjScore prediction group ΛjParticipate in
Prediction, other score prediction groups are not required to participate in;
If score prediction group ΛjThe reality output of r-th of single output nerve network is z(jr), then x belong to r-th of sense organ matter of brand j
Figureofmerit score predicted value is:
<mrow>
<msup>
<mi>z</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
<mo>&prime;</mo>
</mrow>
</msup>
<mo>=</mo>
<mrow>
<mo>(</mo>
<msup>
<mi>z</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>-</mo>
<mn>0.15</mn>
<mo>)</mo>
</mrow>
<mo>*</mo>
<mfrac>
<mrow>
<mi>m</mi>
<mi>a</mi>
<mi>x</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>d</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
<mo>&prime;</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mi>m</mi>
<mi>i</mi>
<mi>n</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>d</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
<mo>&prime;</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
</mrow>
<mn>2.70</mn>
</mfrac>
<mo>+</mo>
<mi>m</mi>
<mi>i</mi>
<mi>n</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>d</mi>
<mrow>
<mo>(</mo>
<mi>j</mi>
<mi>r</mi>
<mo>)</mo>
<mo>&prime;</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>11</mn>
<mo>)</mo>
</mrow>
</mrow>
7. tobacco according to claim 1 and the electronic nose instrument evaluation method of tobacco product aesthetic quality, it is characterized in that,
If on the basis of existing n kinds brand, increase identifies a kind of new brand, need only increase n single output nerve networks and learn,
Therefore, the modular neural network cascade model first order increases to n (n+1)/2 from existing n (n-1)/2 single output nerve network
It is individual;For example, to the brand ω newly increasedn+1, increasing single output nerve mixed-media network modules mixed-media that adduction learns newly is
Correspondingly, in order to carry out the prediction of aesthetic quality's index score to newly increasing brand, modular neural network cascade model the
Two level newly increases 5 single output nerve networks and learnt, and increases to (n+1) × 5 from existing n × 5;False brand or another
The existing same brand of manufacturer production is seen as a kind of single brand and is identified and the prediction of aesthetic quality's index score.
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711060781.3A CN107843695B (en) | 2017-10-31 | 2017-10-31 | The electronic nose instrument evaluation method of tobacco and tobacco product aesthetic quality |
US16/651,722 US10895560B2 (en) | 2017-10-31 | 2018-03-18 | Electronic nose instrument for sensory quality evaluation of tobacco and tobacco product |
PCT/CN2018/079389 WO2019085369A1 (en) | 2017-10-31 | 2018-03-18 | Electronic nose instrument and sensory quality evaluation method for tobacco and tobacco product |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711060781.3A CN107843695B (en) | 2017-10-31 | 2017-10-31 | The electronic nose instrument evaluation method of tobacco and tobacco product aesthetic quality |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107843695A true CN107843695A (en) | 2018-03-27 |
CN107843695B CN107843695B (en) | 2018-08-28 |
Family
ID=61682198
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711060781.3A Active CN107843695B (en) | 2017-10-31 | 2017-10-31 | The electronic nose instrument evaluation method of tobacco and tobacco product aesthetic quality |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107843695B (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108896706A (en) * | 2018-05-17 | 2018-11-27 | 华东理工大学 | The foul gas multiple spot centralization electronic nose instrument on-line analysis of big data driving |
CN109241983A (en) * | 2018-09-13 | 2019-01-18 | 电子科技大学 | A kind of cigarette image-recognizing method of image procossing in conjunction with neural network |
CN109709267A (en) * | 2018-12-14 | 2019-05-03 | 武汉东昌仓贮技术有限公司 | A kind of tobacco-alcoholizing quality evaluation method and system based on electronic nose |
WO2019218395A1 (en) * | 2018-05-17 | 2019-11-21 | 华东理工大学 | Online monitoring and analysis method for malodorous gas multi-point centralized electronic nose instrument |
CN111443161A (en) * | 2020-01-23 | 2020-07-24 | 华东理工大学 | Electronic nose gas-sensitive-chromatographic information fusion and flavor substance on-site detection and analysis method |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20020000115A1 (en) * | 2000-07-03 | 2002-01-03 | Shimadzu Corporation | Odor measuring apparatus |
CN101101299A (en) * | 2007-06-25 | 2008-01-09 | 华东理工大学 | Parallel-tandem mode identification method and its uses in machine scent |
CN101382557A (en) * | 2008-09-04 | 2009-03-11 | 广东中烟工业有限责任公司 | Cigarette quality analytical method |
CN101419209A (en) * | 2008-12-04 | 2009-04-29 | 哈尔滨工程大学 | Cigarette sensing appraise and flume index immune neural net prediction method |
CN102353798A (en) * | 2011-07-01 | 2012-02-15 | 华东理工大学 | Biological fermentation process-oriented olfactory analog instrument and online analysis method |
CN103472094A (en) * | 2013-07-24 | 2013-12-25 | 华东理工大学 | Olfactory analog instrument and on-site analysis method for odor grade of specific substance |
CN104820077A (en) * | 2015-05-29 | 2015-08-05 | 广东中烟工业有限责任公司 | Method and device for recognizing cigarette |
CN105758897A (en) * | 2016-03-29 | 2016-07-13 | 福建农林大学 | White peony fragrance grade discriminating method based on electronic nose detection information |
CN105954451A (en) * | 2016-06-06 | 2016-09-21 | 广东中烟工业有限责任公司 | Rapid cigarette type distinguishing method based on electronic nose full chromatographic data |
-
2017
- 2017-10-31 CN CN201711060781.3A patent/CN107843695B/en active Active
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20020000115A1 (en) * | 2000-07-03 | 2002-01-03 | Shimadzu Corporation | Odor measuring apparatus |
CN101101299A (en) * | 2007-06-25 | 2008-01-09 | 华东理工大学 | Parallel-tandem mode identification method and its uses in machine scent |
CN101382557A (en) * | 2008-09-04 | 2009-03-11 | 广东中烟工业有限责任公司 | Cigarette quality analytical method |
CN101419209A (en) * | 2008-12-04 | 2009-04-29 | 哈尔滨工程大学 | Cigarette sensing appraise and flume index immune neural net prediction method |
CN102353798A (en) * | 2011-07-01 | 2012-02-15 | 华东理工大学 | Biological fermentation process-oriented olfactory analog instrument and online analysis method |
CN103472094A (en) * | 2013-07-24 | 2013-12-25 | 华东理工大学 | Olfactory analog instrument and on-site analysis method for odor grade of specific substance |
CN104820077A (en) * | 2015-05-29 | 2015-08-05 | 广东中烟工业有限责任公司 | Method and device for recognizing cigarette |
CN105758897A (en) * | 2016-03-29 | 2016-07-13 | 福建农林大学 | White peony fragrance grade discriminating method based on electronic nose detection information |
CN105954451A (en) * | 2016-06-06 | 2016-09-21 | 广东中烟工业有限责任公司 | Rapid cigarette type distinguishing method based on electronic nose full chromatographic data |
Non-Patent Citations (6)
Title |
---|
DAQI GAO等: "Quantitative analysis of multiple kinds of volatile organic compounds using hierarchical models with an electronic nose", 《SENSORS AND ACTUATORS B: CHEMICAL》 * |
DEHAN LUO等: "Application of ANN with extracted parameters from an electronic nose in cigarette brand identification", 《SENSORS AND ACTUATORS B》 * |
GAO DAQI等: "Simultaneous estimation of odor classes and concentrations using an electronic nose with function approximation model ensembles", 《SENSORS AND ACTUATORS B》 * |
殷勇等: "用基于神经网络的电子鼻评定卷烟香气质量", 《仪器仪表学报》 * |
焦禹顺等: "电子鼻检测和定量表达分析相结合对烟草香气突变体进行鉴定", 《中国烟草学报》 * |
田书霞等: "卷烟主流烟气粒相物的电子鼻分析", 《中国烟草学报》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108896706A (en) * | 2018-05-17 | 2018-11-27 | 华东理工大学 | The foul gas multiple spot centralization electronic nose instrument on-line analysis of big data driving |
CN108896706B (en) * | 2018-05-17 | 2019-04-16 | 华东理工大学 | The foul gas multiple spot centralization electronic nose instrument on-line analysis of big data driving |
WO2019218395A1 (en) * | 2018-05-17 | 2019-11-21 | 华东理工大学 | Online monitoring and analysis method for malodorous gas multi-point centralized electronic nose instrument |
US10948467B2 (en) | 2018-05-17 | 2021-03-16 | East China University Of Science And Technology | Online centralized monitoring and analysis method for multi-point malodorous gases using electronic nose instrument |
CN109241983A (en) * | 2018-09-13 | 2019-01-18 | 电子科技大学 | A kind of cigarette image-recognizing method of image procossing in conjunction with neural network |
CN109709267A (en) * | 2018-12-14 | 2019-05-03 | 武汉东昌仓贮技术有限公司 | A kind of tobacco-alcoholizing quality evaluation method and system based on electronic nose |
CN109709267B (en) * | 2018-12-14 | 2021-06-29 | 武汉东昌仓贮技术有限公司 | Tobacco mellowing quality evaluation method and system based on electronic nose |
CN111443161A (en) * | 2020-01-23 | 2020-07-24 | 华东理工大学 | Electronic nose gas-sensitive-chromatographic information fusion and flavor substance on-site detection and analysis method |
CN111443161B (en) * | 2020-01-23 | 2021-01-05 | 华东理工大学 | Electronic nose gas-sensitive-chromatographic information fusion and flavor substance on-site detection and analysis method |
WO2021147275A1 (en) * | 2020-01-23 | 2021-07-29 | 高大启 | Gas sensitivity and chromatography multi-information fusion and flavor substance on-site test and analysis method for electronic nose instrument |
Also Published As
Publication number | Publication date |
---|---|
CN107843695B (en) | 2018-08-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10895560B2 (en) | Electronic nose instrument for sensory quality evaluation of tobacco and tobacco product | |
CN107843695A (en) | Tobacco and the electronic nose instrument evaluation method of tobacco product aesthetic quality | |
Hazar | An Analysis of the Relationship between Digital Game Playing Motivation and Digital Game Addiction among Children. | |
Luo et al. | Application of ANN with extracted parameters from an electronic nose in cigarette brand identification | |
CN101692053A (en) | Multi-sensing information fusion based instrumental intelligent evaluation method for quality of famous tea | |
Lepš | Vegetation dynamics in early old field succession: a quantitative approach | |
CN107807199B (en) | One grows tobacco and tobacco product electronic nose instrument detection method | |
CN101419209A (en) | Cigarette sensing appraise and flume index immune neural net prediction method | |
CN104569062B (en) | Electronic nose air chamber based on molecular sieve filtration | |
CN105092789B (en) | Tobacco style changing features trend analysis in a kind of sheet nicotinyl alcohol process | |
CN103562718A (en) | Disease detection in plants | |
CN103234918A (en) | Method for discriminating liquor odor by liquid array taste sense bionic sensor | |
CN109920551A (en) | Autism children social action performance characteristic analysis system based on machine learning | |
CN102692488A (en) | Jinhua ham grading and identifying method based on electronic nose technology | |
CN104713849A (en) | Method for quickly predicting tobacco mildew based on near-infrared spectrum analysis technique | |
CN110307871A (en) | A method of quickly detection The Quality of Sliced Herbal Medicine is merged using electronics sense organ | |
CN107491447A (en) | Establish inquiry rewriting discrimination model, method for distinguishing and corresponding intrument are sentenced in inquiry rewriting | |
Garamszegi et al. | The relationship between syllable repertoire similarity and pairing success in a passerine bird species with complex song | |
Mammides et al. | Measurement of species associations in mixed‐species bird flocks across environmental and human disturbance gradients | |
CN104680314A (en) | Method for carrying out sensory evaluation on style characteristic of tobacco leaf | |
CN103399050A (en) | Method for rapidly evaluating ginseng-adulterated American ginseng based on mouth feel information | |
CN108414471A (en) | A method of sense organ characterization information is differentiated based near infrared spectrum and sensory evaluating smoking's mutual information | |
Karimi | C-test and Vocabulary Knowledge | |
Brown | Transfer performance in children's oddity learning as a function of dimensional preference, shift paradigm and overtraining | |
Tudu et al. | Normalization techniques for gas sensor array as applied to classification for black tea |
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 |