CN104730140B - A kind of parameter optimization method in electronic nose detection honey - Google Patents
A kind of parameter optimization method in electronic nose detection honey Download PDFInfo
- Publication number
- CN104730140B CN104730140B CN201310323227.5A CN201310323227A CN104730140B CN 104730140 B CN104730140 B CN 104730140B CN 201310323227 A CN201310323227 A CN 201310323227A CN 104730140 B CN104730140 B CN 104730140B
- Authority
- CN
- China
- Prior art keywords
- sample
- honey
- head space
- parameter
- electronic nose
- 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.)
- Active
Links
Landscapes
- Sampling And Sample Adjustment (AREA)
- Investigating Or Analysing Biological Materials (AREA)
Abstract
A kind of parameter optimization method in electronic nose detection honey,Signal difference maximum turns to guiding between different honey samples,The optimal electronic nose testing conditions for distinguishing effect of selection,It is characterized in that detection parameters are divided into head space parameter and detection parameters,Wherein detection parameters can be divided into sample introduction parameter and signal acquisition parameter again,In view of the detection feature for instrument of detection parameters reaction,When instrument stabilizer,Its influence to testing result is smaller,Head space parameter then influences the generation of sample headspace gas,And headspace gas is then the direct detection object of electronic nose,Directly affect final testing result,The head space parameter of electronic nose mainly includes head space temperature and head space time,Simultaneously final testing result can be also influenceed in view of the difference of different sample sizes in ml headspace bottle,Therefore final choice sample size,Head space temperature and head space time are optimization object,Three factors are optimized using orthogonal experiment.
Description
Technical field
The application is related to the parameter optimization method in a kind of electronic nose detection honey.
Background technology
China's honey output occupied first place in the world, and yield is always maintained at the trend of rapid growth in recent years, by 2001
25.2 ten thousand tons of increase to 2009 40.2 ten thousand tons, account for Gross World Product and also brought up to by nearly 20% more than 30%.But due to economy
The driving of interests, honey market is adulterated serious at present, causes adulterated honey to occupy the 20%~30% of honey market, some areas
The bee product of adulterated fraud accounts for 50% or so, badly damaged consumer's interests, influence honey industry develop in a healthy way, strike outlet
Trade is earned foreign exchange.
Due to lacking the influence of detection means, cause adulterated strike difficulties, its basic reason is as follows：（1）Due to honeybee
The main matter of honey itself is relatively simple for structure, comprising water and carbohydrate content, to adulterated condition of providing convenience, meanwhile, depend merely on
Detect the number of this several content of material can not discriminate whether at all it is adulterated；（2）Because honey is by nectariferous plant species, honeybee
Group's gesture is strong and weak, sweet time phase length, the temperature and humidity of air, and many factors shadow such as the processing of honey, storage, crystallization
Ring, cause the content range of honey main matter to change greatly so that honey adulteration is simply, conveniently；（3）The detection of adulterations such as C4 take
With height, actually detected and law enforcement can not be used on a large scale.
Fragrance is one of important attribute that product quality embodies, and product fragrance, which characterizes, to be needed to protrude its objectivity, authenticity
With it is comprehensive.Gas-chromatography (GC), gas chromatography-mass spectrography (GC-MS) and gas-chromatography-smell distinguish the side such as (GC-O) at present
Method, monomer aroma substance limited in product can only be detected, and phenomena such as collaboration, modified tone between these fragrance be present, it is difficult to
Reflect the flavouring essence quality of sample on the whole.And intelligent Olfaction System（Electronic nose）Mankind's smelling feature, consolidated statement can be simulated
The Global Information of fragrance is levied, embodies the olfactory characteristic and overall quality of fragrance, at the same it is more objective than the smell of people, reliable.Mesh
It is preceding to be defined in food freshness, edible deterioration of oil differentiation, the detection of fruits and vegetables maturity, tea-leaf producing area variety ecotype, drinks brand
Etc. carried out correlative study.
Contain more than 300 kinds of aromatic substance in honey, therefore it is the important sample that the intelligent smell of research characterizes；It is simultaneously different
Nectar source, its flavor substance of different sources are different, and honey adulteration whether or quality can have been embodied on overall fragrance,
So that fragrance turns into one of important indicator of honey quality detection and adulterated discriminating；Absolutely prove and honey is characterized using intelligent smell
Quality has feasibility, and it is a kind of quick, economical, accurate and beneficial in real time should also to be provided for honey quality detection and adulterated discriminating
Detection method.Therefore selection honey has Practical significance as research object, to the more far-reaching valency of its industry healthy development
Value.
Product quality differentiation or adulterated discriminatory analysis are carried out using electronic nose, its essence is to utilize the whole of intelligent smell collection of illustrative plates
Body note gas information, finds the otherness of sample room, and its core is the profile information of otherness between searching representative sample, i.e. " difference
Change information ", also cry " the differentiation profile information of intelligent smell ".But the sensor array of electronic nose has cross-sensitivity,
That is every sensor has different degrees of response to each fragrance, therefore the aroma-producing substance collection of illustrative plates gathered by electronic nose has
Wide spectrum, it is overlapping the features such as, it is difficult to individually with the naked eye distinguish different samples from collection of illustrative plates, it is necessary to carry out " signal excavation ", particularly
" excavation of differentiation information between representative sample ", the otherness information of excavation is more, more contributes to efficiently to distinguish product
Feature and quality.But it is also very weak in terms of differentiation information excavating at present, and restrict the bottleneck of electronic nose development..
The content of the invention
A kind of parameter optimization method in electronic nose detection honey, signal difference maximum, which turns to, between different honey samples leads
To the optimal electronic nose testing conditions for distinguishing effect of selection, it is characterised in that detection parameters are divided into head space parameter and detection parameters,
Wherein detection parameters can be divided into sample introduction parameter and signal acquisition parameter again, it is contemplated that the detection for instrument of detection parameters reaction is special
Point, when instrument stabilizer, its influence to testing result is smaller, and head space parameter then influences the generation of sample headspace gas, and pushes up
Air body is then the direct detection object of electronic nose, that is, directly affects final testing result, the head space parameter of electronic nose is main
Final detection knot can be also influenceed including head space temperature and head space time, while in view of the difference of different sample sizes in ml headspace bottle
Fruit, therefore final choice sample size, head space temperature and head space time are optimization object, three factors are carried out using orthogonal experiment excellent
Change, when the varying level to each factor selects, it is contemplated that ml headspace bottle 10ml needs concussion to heat in head space room, is anti-
Only sample introduction needle touches fluid sample and influences instrument performance, and sample maximum may not exceed the 1/2 of ml headspace bottle, according to the close of honey
Spend for 1.4g/ml, three levels for determining headspace sampling amount are respectively 4g, 5g, 6g, head space temperature it is horizontally selected in, choosing
Three levels are respectively 40,50,60 DEG C, in head space selection of time, select the shorter head space time, three levels are respectively
120s, 180s, 240s, i.e. sample size 4g, 5g, 6g, head space temperature 40,50,60 DEG C, 180s, 240s, are ground head space time 120s
It is 5 kinds of different nectar source honey samples, respectively rape honey, acacia honey, chaste honey, honey of lychee flowers, linden to study carefully selected honey sample
Tree is sweet, every 6 parts of class nectar source sample, altogether 30 parts of samples.
Parameter optimization method according to claim 1, for the orthogonal experiment of similar Almost Sure Sample Stability evaluation index,
The lower 5 kinds of honey of every group of experiment takes 3 parts of samples, and Different categories of samples is believed by calculating 18 sensors of electronic nose under the experiment condition
Number standard deviation average, to weigh the stability C that electronic nose under the conditions of this detects to similar sample signal, computational methods such as following formula
It is shown
Wherein p is nectar source species, CkIt is electronic nose number of probes for the stability of kth class sample, m, nkFor k classes nectar source
The number of middle sample,For the response signal of i-th of sample j root sensor in the sample of k classes nectar source,For k classes nectar source sample
In the response signal average of jth root sensor.
Parameter optimization method according to claim 1, to the otherness of inhomogeneity sample according to inhomogeneity nectar source sample
This variance D of sample average under identical conditions is calculated, and computational methods are shown below
Wherein p be nectar source species,For the signal average of i-th sensor of k classes sample,For all samples
The signal average of i-th sensor, nkFor the number of sample in same nectar source.
Brief description of the drawings
Fig. 1 abnormity point elimination results：(a) mahalanobis distance differentiates result；(b) lever value differentiates result；
Feature extraction results of the Fig. 2 based on variance ratio
The feature point extraction result that Fig. 3 is differentiated based on individual event amount
Fig. 4 ant group algorithm flow charts
Feature extraction results of the Fig. 5 based on ant group algorithm
Feature point extraction results of the Fig. 6 based on core principle component analysis
Feature point extraction results of the Fig. 7 based on independent component analysis
Fig. 8 searches the SVMs parameter optimization result of element based on grid
SVMs parameter optimization results of the Fig. 9 based on genetic algorithm
SVMs parameter optimization results of the Figure 10 based on particle cluster algorithm
Embodiment
1 collects and prepares on sample
To make studied nectar source difference representative, according to China geographic area（Western part, south China, North China, East China,
Northeast）Division, select 5 kinds of different nectar sources to be respectively as research samples：1）Rape honey, pick up from the Chongqing Fu in west area
Ling Qu and Yongchuan District；2）Honey of lychee flowers, pick up from the Nanning of South China；3）Chaste honey, pick up from the Miyun Region of Beijing of North China
Etc. ground；4）Acacia honey, pick up from the Laiyang Shandong Province in East China；5）Mel, pick up from Jilin Dunhua and the Heilungkiang Harbin in northeast etc.
Ground.To ensure the authenticity and accuracy of experiment sample, the interference of market business honey processing technology is avoided, sample passes through Chinese agriculture
Honeybee research institute of the industry academy of sciences is directly bought by beekeeper.
It is respectively placed in after sample collection according to different nectar sources, different sources in different reagent bottles.To ensure that research is not examined
The interference of condition difference is surveyed, is stored in after sample collection under the conditions of -18 DEG C, treats uniformly to be tried after all samples collection
Test.Before experiment, after sample takes out at -18 DEG C, 5 kinds of nectar source samples respectively take 60g or so, are placed in 40 DEG C of constant water bath box, water
Bath heating 15min, melts honey sample, remaining sample, which continues to be placed at -18 DEG C, to be preserved.It is to ensure that sample melts during heating water bath
Change complete, nodeless mesh, concussion is needed per 3min during water-bath once.After the completion of sample water-bath, taking-up is placed in cooling more than 1h at room temperature,
Until sample temperature and room temperature（20℃）Unanimously.
2 electronic nose detection methods
Electronic nose is carried out using gas sensor array from the Adsorption of different volatile ingredients to testing sample honey
Detection.Honey volatile ingredient adsorbs with sensor characteristics（Including physical absorption and chemisorbed）Afterwards, semiconductor sensing is changed
Device top layer current strength.Changed by numeral, obtain the response curve of each sample, so as to be tested and analyzed to sample.This hair
It is bright to use the type electronic noses of Fox 4000（Alpha MOS, France）, the electronic nose is by 18 Metal Oxide Semiconductor Gas Sensings
Sensor（MOS）Formed with HS100 headspace autosamplers.
Instrument concrete operations flow is as follows：
1）The honey sample that room temperature is cooled to after water-bath is added as requested in the ml headspace bottle that volume is 10ml.It will install
The ml headspace bottle of sample is placed on pallet.HS100 automatic samplers accommodate up to 2 pallets, and each pallet can place 32 head spaces
Bottle.
2）Setting instrument testing conditions as requested, including head space sampling and electronic nose testing conditions.According to nectar source
Species and detection ordering, each ml headspace bottle on pallet is encoded.
3）Ml headspace bottle is placed into head space room according to the condition of setting and heated, and ml headspace bottle interval is shaken during heating, is protected
Demonstrate,prove headspace gas homogeneity.After head space sample preparation terminates, headspace gas is extracted, in Injection Detector, and by ml headspace bottle from head space room
Interior taking-up.Fox 4000 is that continuous type air-flow is injected, anti-with each sensor generation adsorption and de-adsorption after gas entrance detection gas
Should, and the response curve of each self-generating response.
Simple sample can obtain 18（18 sensors）*t（Detection time）Signal matrix.Conventional method is by each sensor
Maximum（It is small）Value is analyzed as the response of the sensor.
The 3 honey quality modeling methods based on electronic nose information
SVMs discrimination model is established using the electronic nose characteristic information extracted, the sample in different nectar sources is carried out
Classification.Traditional mode identification method is built upon the progressive theory on the basis of great amount of samples, but individual in production application, by
In the limitation of each side condition, substantial amounts of sample number is often difficult to preferably be ensured, under conditions of small sample, according to biography
The statistical basis of system, it is difficult to obtain comparatively ideal results of learning and extensive effect.But SVMs is applied to small sample bar
Modeling requirement under part, pattern-recognition judgement thus is carried out to different nectar source samples.
SVMs（Support Vector Machine, SVM）Theory is Vapnik（1995）In traditional statistics
On learning foundation, integrated structure principle of minimization risk, proposed for the characteristics of finite sample.This method can effectively subtract
The random row of parameter setting in few traditional mode identification model, overcome empiric risk during model is established and sent out with expected risk
The deficiency of raw bigger difference, specific SVM are theoretical as follows.
In pattern-recognition, an optimization function f (x, w) is obtained, makes it to unknown sample collection (xi,yi)（i=1,
2…,n;yFor specimen number）When being assessed, expected risk R (W) is minimum：
Wherein, F（x,y）For Joint Distribution probability, L (y, f (x, w)) be with f (x, w) y is predicted and caused by damage
Lose, referred to as loss function, identify problem for two quasi-modes, L can be defined as：
In conventional learning algorithmses, using empiric risk Remp(W) minimization principle, i.e.,
But in fact, the minimum of training error is difficult to ensure that the optimum efficiency of prediction, tend to over-fitting occur
Phenomenon, meanwhile, by further investigations have shown that, experience Remp(W) following relation be present with practical risk R (W)：
It is abbreviated as
Wherein h is that the VC of function is tieed up, and η is confidence level, and n is training sample number.
As can be seen from the above equation, to make the classification function practical risk of design minimum, empiric risk is not only made to subtract as far as possible
It is small, while also to increase training set number or reduce function VC dimensions, practical risk could be reduced.This thought is structure risk
Minimization principle.
It is theoretical based on more than, to sample set (xi,yi)（i=1,2…,n;X be sample i characteristic vector, y
For specimen number）When being differentiated, discriminant function is found, W and b are normalized and equal proportion is adjusted
Afterwards, make to meet for all samples, now the classifying distance of two class samples at intervals of.Therefore it is to obtain
Prediction effect of preferably classifying is obtained, two class samples should be made as separated as possible, that is, askedMinimum value.MeetPoint,
In classify that plan range is minimum, and they determine optimal classification function, these points are referred to as supporting vector（Support
Vector, SV）.
Under this condition, optimization problem can be converted into the problem of optimal classification function：
Optimization problem is converted into dual problem and is then represented by：
Wherein αiFor for constraints（7）Lagrange（Lagrange）The factor, i=1,2 ... n, W are classification function
Slope, b be classification function intercept.
For linearly inseparable problem, V.Vapanik introduces kernel function theory, i.e., passes through data in lower dimensional space non-thread
Property mapping projections value higher dimensional space in, can prove, if the appropriate kernel function of selection, will linearly can not in lower dimensional space
The data divided are converted into the data of linear separability in higher dimensional space.After introducing kernel function, full scale equation can be converted into：
Wherein K is selected kernel function.
By solving kernel function, corresponding classification function may finally be determined：
Pattern recognition step overall SVM can be summarized as several steps：
（1）Select appropriate kernel function K；
（2）Optimization method corresponding to solution, obtain supporting vector；
（3）Obtain optimal classification function f (x)
（4）The classification for determining to differentiate according to sgnf (x) value；
Parameter optimization in 4 electronic noses detection honey
4.1 determine parameter and level to be optimized
Electronic nose detection parameters can divide head space parameter and detection parameters.Wherein detection parameters can be divided into again sample introduction parameter and
Signal acquisition parameter.In view of the detection feature for instrument of detection parameters reaction, when instrument stabilizer, it is to testing result
Influence smaller.Head space parameter then influences the generation of sample headspace gas, and headspace gas is then the direct detection object of electronic nose,
Directly affect final testing result.Therefore, head space parameter is optimized emphatically in the present invention.The head space parameter of electronic nose
It is main to include head space temperature and head space time, while can also influence final inspection in view of the difference of different sample sizes in ml headspace bottle
Result is surveyed, therefore final choice sample size, head space temperature and head space time are optimization object.To select optimum combination, using just
Experiment is handed over to optimize three factors.When the varying level to each factor selects, it is contemplated that ml headspace bottle（10ml）Pushing up
Concussion is needed to heat in empty room, to prevent sample introduction needle touches fluid sample from influenceing instrument performance, sample maximum may not exceed
The 1/2 of ml headspace bottle.According to the density of honey（About 1.4g/ml）, three levels determining headspace sampling amount be respectively 4g, 5g,
6g.Head space temperature it is horizontally selected in, according to bibliography, honey sample property under the conditions of higher than 68 DEG C easily becomes
Change, therefore selected three levels are respectively 40,50,60 DEG C.In head space selection of time, the speed of large sample size detection is considered
It is required that, the short-term stability of honey sample in high temperature environments, and the effumability feature of honey sample, select shorter
Head space time, three levels are respectively 120s, 180s, 240s.The final optimal conditions for determining Three factors-levels, i.e. sample size
4g, 5g, 6g, head space temperature 40,50,60 DEG C, head space time 120s, 180s, 240s, each factor and it is horizontal as shown in table 1.Research
Selected honey sample is 5 kinds of different nectar source honey samples, respectively rape honey, acacia honey, chaste honey, honey of lychee flowers, lime tree
Honey, per 6 parts of class nectar source sample, 30 parts of samples altogether.
The selected nominal price experiment table of experiment is L9(3)4, the design of experiment table is as shown in table 2
Remaining testing conditions is as shown in table 3
The evaluation index and method that 4.2 optimizations determine
Present invention signal difference maximum between different honey samples turns to guiding, selects the optimal electronic nose inspection for distinguishing effect
Survey condition.By optimizing detection condition, it is expected signal stabilization between similar honey sample, and inhomogeneity honey inter-sample difference compared with
Greatly, so as to ensureing the maximization of e-nose signal difference between honey sample.
（1）Similar Almost Sure Sample Stability evaluation index
In orthogonal experiment, the lower 5 kinds of honey of every group of experiment takes 3 parts of samples, by calculating electronic nose 18 under the experiment condition
Root sensor to the standard deviation average of Different categories of samples signal, come weigh electronic nose under the conditions of this similar sample signal is detected it is steady
It is qualitative.Computational methods are as shown in formula 15,16
Wherein p is nectar source species, CkIt is electronic nose number of probes for the stability of kth class sample, m, nkFor k classes nectar source
The number of middle sample,For the response signal of i-th of sample j root sensor in the sample of k classes nectar source,For k classes nectar source sample
In the response signal average of jth root sensor.（2）The otherness of inhomogeneity sample
According to inhomogeneity nectar source sample, the variance of sample average under identical conditions calculates the otherness of inhomogeneity sample
Arrive, computational methods are as shown in Equation 17
（17）
WhereinFor the signal average of k class samples,For the signal average of all samples
（3）Overall assessment index
The desired optimum optimizing condition obtained the signal stabilization for similar sample between of research, and between inhomogeneity sample with compared with
Big difference.Therefore, the evaluation index q finally determined is as shown in Equation 18
4.3 are based on the maximized electronic nose parameter optimization result of signal difference between honey sample
It is as shown in table 4 as the observation of orthogonal experiment, Orthogonal experiment results using the evaluation index of determination
As can be seen from Table 4, three kinds of factors are distinguished effect to electronic nose and had a certain impact.Wherein factor A, factor
B, i.e. sample size and head space temperature are directly proportional to signaling zone indexing, and factor C and the dead space indexing of sample headspace time are inversely proportional,
And blank control group change is relatively small.This is due to the increase of sample size and head space temperature, and testing sample waves in ml headspace bottle
Hair constituent concentration gradually increases, and is available for the specific component content increase distinguished, it is preferable to distinguish effect.And with the increasing of head space time
Add, under the higher environment of temperature, sample volatile ingredient property changes, and such changes the area that have impact on sample class
Point, therefore distinguish effect and decline.Control group stably show similar sample it is multiple between no significant difference, testing result is reliable.To enter
One step is analyzed Orthogonal experiment results, carries out variance analysis to experimental result, variance analysis is as shown in table 5.
Three factors have a significant impact to the discrimination of sample it can be seen from the result of variance analysis, wherein factor A and factor
B conspicuousnesses are larger（P<0.01）, the variance contribution ratio of three kinds of factors is respectively 65.34%, 22.16% and 9.66%.This result shows
Sample size and head space temperature have a great influence to sample discrimination, and the head space time is smaller on the influence of sample discrimination.
Based on the above results, it is A3B3C1 to select optimum optimization combination condition, i.e. sample size 6g, head space temperature 60 C, top
120s between space-time, the inhomogeneity sample obtained under the conditions of this distinguish best results.
Exceptional sample point in 5 electronic noses detection honey is rejected
5.1 exceptional sample point eliminating principle
Before formally analysis and honey quality modeling are carried out to the e-nose signal of honey, to ensure analysis result and model
Stability, it is necessary to rejected to the exceptional sample in honey entirety sample, ensure to obtain the accurate of signal and sample information
Property and reliability.Exceptional sample in electronic nose detection honey includes the exception of sample information（Such as sample number, classification）And inspection
Survey the exception of result.Exceptional sample usually easily influences the variation tendency of overall signal, and the stabilization for destroying disaggregated model is thought.Therefore
Rejecting to exceptional sample is extremely necessary.
Cause the factor of electronic nose exceptional sample mainly include it is following some：（a）The error of sample collection is mainly with collection
The mistake of sample point, based on error coded；（b）The change of sample storage phase property, because sample is had by collecting detection and analysis
Certain interval, here under, chemistry, the change of physical property easily occur for some unstable samples；（c）The mistake of operation
Miss, including cleanliness factor of the error of title sample, container etc.；（d）The error of instrument detection, due to detection environment and sensor properties
Change, even if by the pre-treatment of signal, still has and is difficult to completely be corrected detection signal, still has part signal to be put down with overall
Equal signal has significant difference.
To obtain larger sample size, the present invention is divided for three kinds of larger nectar source honey of existing market occupation rate
Analysis, rape honey, Mel and acacia honey, studies the analysis and rejecting for the exceptional sample point in great amount of samples.Wherein 76
Rape honey, 56 Mels, 112 acacia honeys, altogether 244 samples.The equally application of this group of sample with 7,8 research.
5.2 exceptional sample point elimination methods
（1）Mahalanobis distance differentiates
Mahalanobis distance（Mahalanobis）It is the evaluation index of vectorial intensity in hyperspace, is that multivariate data is different
A kind of important method of constant value detection.Mahalanobis distance is by calculating the mean vector and covariance matrix comparative sample of sample data
Departure degree between signal, circular are as follows：
Wherein sample mean vector, S are sample covariance matrix.For the mahalanobis distance average of sample,For geneva away from
From standard deviation, λ is the threshold value for receiving scope, xtFor the characteristic vector of t-th of sample, T is characterized vectorial average.Set in the present invention
Determine hard -threshold λ=3.For acceptable mahalanobis distance scope under the threshold condition.
（2）Lever value differentiates
The size of sample thick stick embodies model to the degree of dependence of the sample, and lever value is bigger, and dependence is bigger, and model is influenceed
It is bigger.Being usually located at the sample at property both ends to be analyzed has larger lever value.Excessive lever value has larger shadow to model
Ring, be unfavorable for the stabilization of model.By the analysis to sample lever value, the special sample being had a great influence to model is rejected, so as to
Increase the stability of model.Lever value differentiates that specific method is as follows：
1st, the score matrix T of sample to be tested is calculated by PCA；
2nd, test matrix H is calculated：；
3rd, the lever value hi of each sample：Hi is i-th of diagoned vector in test matrix H,；
Similar with mahalanobis distance differentiation, lever value is differentiated by setting hard -threshold, is removed with special compared with big lever value
Sample point, so as to ensure the stability of later stage forecast model.
Exceptional sample point in the detection of 5.3 honey electronic noses rejects validity check
Verified to reject effect to sample, select Bayes（Bayes）Method of discrimination before and after abnormity point elimination to sentencing
The accuracy rate of other model is evaluated, compared to other mode identification methods, Bayes method of discrimination is more simple, without to phase
Parameter is answered to optimize.The basic thought that Bayes differentiates assumes that the object to being studied（It is overall）It is existing certain before sampling
Understanding, conventional prior probability distribution describes this understanding.The sample for being then based on extracting is corrected to priori understanding again, is obtained
To so-called Posterior probability distribution, and various statistical inferences are all based on Posterior probability distribution to carry out.Bayesian Decision is different from warp
The statistical method of allusion quotation, its distinguishing feature are exactly that application is all as far as possible in the case where guarantee risk of policy making is as small as possible
Possible information.
Fig. 1 is analysis sample（76 parts of rape honeys, 56 parts of Mels, 112 parts of acacia honeys）Mahalanobis distance differentiate (a) and thick stick
Bar value differentiates result (b).Differentiated using mahalanobis distance, altogether rejecting 36,53,76,79,85,99,117,240,244,9 altogether
Abnormity point；And lever differentiates totally 5 abnormity points of rejecting 36,79,216,240,244 altogether.
Differentiated using Bayes and pattern-recognition prediction is carried out to the data after two kinds of abnormal point methods processing, prediction result is such as
Shown in table 6.Found by table 6, differentiated compared to lever, mahalanobis distance differentiates that the peculiar sample spot rejected is more, but accuracy rate is simultaneously
Without too big difference.This result shows the exceptional sample point that mahalanobis distance is rejected more, although being differed compared with the overall distribution of sample
It is larger, but to not affected greatly to the result accuracy rate of differentiation, therefore, in order to fully take into account all special of sample
Property, select lever to differentiate, reject 5 sample spots, i.e., the the 36th, 76,216,240,244 5 exceptional sample, wherein rape honey, linden
Tree honey each 1, acacia honey 3.
6
Honey characteristic perfume is analyzed and honey fragrance simulated system is established
Using Dynamic headspace（Itex）Honey aroma-producing substance is extracted with reference to circulation collection technology, 1 is carried out in chromatographic column end:
After the distribution of 1 fragrance content, using gas chromatography mass spectrometry（GC-MS）With gas-chromatography-measurement of olfaction（GC-Olfactometry, GC-O）
It is in fragrant composition and sensory characteristic that technology determines its volatility simultaneously.Crystallized honey carries out heating water bath, is then rapidly cooled to room
Temperature, and keep constant indoor temperature state acquisition aroma-producing substance.
In wherein GC-MS, mass spectrum is utilized（Library searching）, Relative Retention Indices（RI）Honey is determined with three kinds of methods of smelling
Volatile ingredient, and carry out inner mark method ration.GC-O technologies are the methods being combined using frequency detecting and detection intensity, by
5 GC-O evaluations groups for preferably smelling the person's of distinguishing composition, it is determined that representing honey head perfume, front end fragrance, body note and bottom note four respectively
The characteristic flavor on basis activity fragrance of volatilization period.
According to the characteristic perfume species and content ratio of four volatilization periods, proportioning builds basic honey fragrance simulated system
A.On the basis of system A, the structure four group systems variant with it.Every group of system and primary structure A difference are embodied in
Two aspects, otherwise that is, in some volatilization period, its characteristic perfume content is different, otherwise its characteristic perfume component is different, and other
The aroma component and content of three phases are all constant.
7 characterize the intelligent smell TuPu method extraction of honey otherness
7.1 feature extracting methods based on variance ratio
Variance and kind internal variance ratio among its sample are calculated to each signaling point of every sensor, according to the big of variance ratio
It is small that signaling point is selected.The computational methods of variance are the same as the evaluation index q in optimal conditions.With other embedded feature selectings
Difference, variance ratio are selected directly by the way that under more each information point, inter-species variance is selected information point with kind internal variance ratio,
The method by other pattern discriminations is not required to, therefore the selection result of this method will not be sent out because selecting different mode recognition methods
It is raw to change.It is larger to the operand of large sample but Variance ratio method belongs to the method for exhaustion.
Fig. 2 shows the variance ratio of a signaling point.It can be found that the larger letter of the bigger i.e. intermediate diversity of variance from figure
Breath point concentrates on 900-1200 and 1800-2160, and the 8th to the 10th, the sensors of 15- the 18th.It is poor for same root sensor, variance
Different larger point is concentrated in the signaling point that detection time is 20s to 35s.It is mainly volatilization gas and sensor in time period
Adsorption time, and the signaling point in the detection later stage of each sensor is that difference is smaller in desorption time point.Experimental selection variance ratio
Signaling point more than 1 meets that information of the signaling point of this condition entrained by it can react different sample rooms as characteristic point
Difference, 798 characteristic points are selected altogether.On this condition, collected using SVM discrimination models by modeling collection checking than being 2:1 ratio is entered
Row prediction, it is final to differentiate that result is 89.8734% (71/79, wherein rape honey 21/23, Mel 14/17, acacia honey 36/
39)。
7.2 feature extracting methods based on individual event diagnostic method
Pattern-recognition is carried out one by one to all characteristic points, compares and differentiates accuracy rate when each signaling point is as single features
Difference.Mode identification method is embedded in feature selecting by this method, by that with reference to method of discrimination, can obtain each signaling point pair
The ability of sample prediction.Different from former filtering method, this method is relatively relied on selected mode identification method, selection knot
Fruit can change a lot with the change of method of discrimination.The discriminant criterion selected in this research is Bayes diagnostic methods
Found from Fig. 3, it is different from variance selection result, during unidirectional amount selects between different sensors accuracy rate difference compared with
It is small, and with differing greatly between the information point under different detection times in root sensor.But time point Bayes in different sensors
Differentiate that variation tendency and the variance ratio variation tendency of accuracy rate are substantially uniform, that is, detect signal at initial stage and differentiate that accuracy rate is higher, collection
In in the preceding 30s detection times of each sensor, and the effect for detecting the later stage is then poor.Selection differentiates that accuracy rate is more than 60%
Signaling point is as characteristic point, and totally 598 characteristic point, is verified using svm.SVM predictablity rates be 84.8101% (67/79,
Wherein rape honey 20/23, Mel 13/17, acacia honey 34/39).
7.3 feature extracting methods based on ant group algorithm
First two algorithm is exhaustive system of selection, it is necessary to be calculated one by one each characteristic point.When characteristic point compared with
When more, amount of calculation can be very big, this also fundamentally its feature extraction for a large amount of signaling points of limit value.Ant group algorithm belongs to
Heuristic Feature system of selection, is evolved using the automatic Iterative of algorithm, automatic optimal is carried out to feature point selection, until obtaining most
Excellent result.
Ant group algorithm (Ant Colony Optimization, ACO) is initially applied to the Path Selection of travelling salesman,
Shortest path is optimized.Ant group algorithm is applied to the selection of characteristic point in this experiment.Algorithm simulation genetic algorithm, profit
Each biography characteristic vector is encoded with binary coding, 1 representative selects the information point, and the information point is given up in 0 representative.Using each
Bayes after feature point selection differentiates that accuracy rate and selected feature points are fitness function, seeks optimal vector
Combination.The algorithm main innovation point includes：（a）Feature point selection number is added in fitness function, and sets cost parameter, is passed through
Parameter regulation, feature can be counted as needed and differentiate that accuracy rate is accepted or rejected；（b）For keep away due to caused by particular point more
New direction mistake, optimal collection is set, selected with optimal set instead of single optimum point；（c）Pheromone update degree is with fitting
Function is answered to improve directly proportional, algorithm optimization effect is good, then updates amplitude increase；（d）It is poor to effect to accelerate calculating speed
Vector accelerates evaporation rate, reduces pheromone concentration, reduces it and calculates interference to the later stage.Algorithm flow is as shown in Figure 4.
In ant group algorithm, each parameter selection is as follows：Final choice ant colony scale（m）=20；Pheromones volatile concentrations（rho）=
0.003；Outstanding ant collection（n1）=3；Poor ant collection（n2）=3;Characteristic punishes ratio（A）=400；When figure below is a final generation
The concentration of pheromones.
From figure 5 it can be seen that the characteristic point selected by ant group algorithm focus mostly in 1000 nearby and 1500 to 2160, i.e.,
9th, the signaling point of 15-18 root sensors.As can be seen from the results, class algorithm is inspired due to being automatic evolutional algorithm,
Although algorithm has certain randomness, to believing that feature the overall distribution of signal has preferable selection.On this condition, most
Selection feature points 206 eventually, it is 94.94% to differentiate accuracy rate（75/79, wherein rape honey 22/23, Mel 16/17, acacia honey
37/39）.Ant group algorithm result compares first two exhaust algorithm, differentiates that accuracy rate has a certain upgrade, while selected feature is believed
Number point is less, more representative.
7.4 feature extracting methods based on core principle component analysis
First three extracting method is only screened in itself to signal, and the characteristic point selected by such method has certain change
Meaning is learned, can preferably be explained with reference to chemical results.But in the vector given up more after all it is entrained some selected
The information that vector does not possess, it is difficult to ensure the accuracy rate of later stage differentiation.In addition to directly extracting, using dimension reduction method, pass through matrix
Conversion, data message is compressed, effective information is enriched with, can substantially reduce characteristic signal quantity.In this research
Two kinds of dimension reduction methods of core principle component analysis and independent component analysis are have selected to be extracted.
Core principle component analysis（Kernel Principal Component Analysis, KPCA）Kernel function is introduced and led
Constituent analysis（Principal Component Analysis, PCA）In.KPCA utilizes kernel function, by data projection to higher-dimension
In space.Due to more scattered each other after data projection, therefore can inseparable signal enters in lower dimensional space by some
Row is distinguished, and is extracted more representational feature and extracted.KPCA selects Radial basis kernel function in this experiment.
Under KPCA, the SVM predictablity rates of different number of principal components are as shown in Figure 6.Fig. 6, which is shown, drops primary signal through KPCA
Under to different dimensions, the accuracy rate of forecast set.Accurately take the lead in increasing and increasing with dimension it can be seen that SVM differentiates, it
After there is of short duration plateau, 25 to 30 tie up when, differentiate accuracy rate with dimension increase and dramatically increase, accuracy rate is relatively steady afterwards
Fixed, after 100 dimensions, the accuracy rate of sample increases and reduced with dimension.The changing rule shows, in the dimension relatively low stage, each feature
Vector carries sample Accurate classification information, and existence of redundant is less between information, and it is accurate that increase dimension is favorably improved differentiation
Rate.After 50 tie up, there is redundancy, covering in the information between individual characteristic quantity, and the now increase of dimension differentiates the shadow of accuracy rate to lifting
Ring and decline.When the later stage, the redundancy between feature has had influence on differentiation effect, therefore now differentiates that accuracy rate is begun to decline.Most
Selection is as dimension d=81 eventually, predictablity rate highest, is 93.67（74/79, wherein rape honey 22/23, Mel 15/17,
Acacia honey 37/39）.
7.5 feature extracting methods based on independent component analysis
Principal component analysis（Including core principle component analysis）Be classified according to maximum variance between data, i.e. data
Second moment, but have ignored independence of the data in High Order Moment.Independent element（Independent components
analysis, ICA）Line translation then is entered to matrix using the High Order Moment calculated between data, can further be reduced between characteristic vector
Associated row, strengthen Signal Compression effect.
The ICA methods that this experiment uses is fastica.Whitening processing is carried out to data before algorithm is carried out.It is different independent
Svm differentiates that accuracy rate is as shown in Figure 7 under composition
ICA overall variation trend and KPCA types, but change it is relative relax, and rate of accuracy reached is to spy required when stablizing
It is few compared with KPCA to levy dimension.As a result show, when independent element is 14, it is 94.94% to differentiate accuracy rate（75/79, wherein rape honey
22/23, Mel 16/17, acacia honey 37/39）.
The foundation of 8 support vector cassification model optimization methods
Disaggregated model of this research and utilization SVM classifier as different nectar source samples.Wherein, selected kernel function is footpath
To base kernel function（RBF）For：
Wherein r be RBF functions form parameter, xiWith xjFor two samples in sample set.
Mainly there are following two reasons compared to other kernel functions, selection RBF：1）RBF can complete linearly to arrive non-linear
Mapping, can prove that linear kernel function is only a kind of RBF special case by mathematic(al) manipulation；2）Compared to Polynomial kernel function,
RBF parameters are less, and model is relatively simple, and this guarantees the stability of model.
Simultaneously, it is contemplated that be difficult to all training points of requirement and meet constraint function（7）, slack variable ξ is introduced to training points, then
Constraint function（7）It can be changed to
（22）
Then slack variable ξ=(ξ1, ξ2, ξ3... ξn) ', all training sets are embodied by mistake point situation.Therefore punishment letter is introduced
Number c divides the weighted value of degree as interval between balanced class and mistake, then majorized function（6）It can be converted to
（23）
Using RBF as in the SVM classifier of kernel function, different parameters（Form parameter r and penalty parameter c）Under classification effect
Fruit has larger difference, therefore, this research and utilization Different Optimization method, the penalty coefficient in RBF middle r and punishment parameter is entered
Optimization is gone.The data selected in this research are the data in 7 after ICA dimensionality reductions.
Specific Optimizing Flow is as follows：
1st, divide training set in proportion with checking to collect, ratio 2:1.By training set according to five folding cross validation methods, i.e.,
Training set is divided into 5 non-cross subsets, selects 4 subsets therein to carry out parameter training in turn, with remaining one
Subset is verified to the parameter of selection, calculates the classification accuracy of training set under different parameters.
2nd, according to selected kernel function, kernel functional parameter r and penalty parameter c are set.
3rd, model is trained according to five folding cross-validation methods in 1 with selected parameter, and calculates mould under different parameters
The accuracy rate of type.
4th, whether judgment models accuracy rate is up to standard, otherwise changes parameter value.
5th, 2,3,4 steps are repeated, until obtaining best model differentiation rate, or reach stopping criterion for iteration.
Different parameter searching methods is utilized in the present invention, form parameter r in step 2 is optimized with penalty parameter c,
It is final to obtain optimal models.
The 8.1 SVM parameters selections based on grid optimization
Grid optimization utilizes the method for exhaustion, in the span pre-estimated by certain step-length in scope institute a little
Searched for one by one, determine final optimal parameter.It is the truth of a matter with 2,2-4To 210Between exhaustive search is carried out to r and c.When c=
When 5.2780, r=0.1088, training set sample differentiates accuracy rate highest, is 96.25% as shown in Figure 8.On this condition, establish
Model, tested using forecast set.It is final to differentiate that accuracy rate is 96.20% (76/79, wherein rape honey 23/23, Mel
16/17, acacia honey 37/39).
The 8.2 SVM parameters selections based on genetic algorithm
Genetic algorithm（Genetic Algorithm, GA）Heuristic search is carried out to data by using for reference biological evolution theory
Rope, the algorithm are taught 1975 by the J.Holland in the U.S. propose first earliest.The basic operation process of genetic algorithm is as follows：
1) data initialization：Maximum evolutionary generation is set, the number of individuals generated at random and its colony formed.This research
Middle selection number of individuals 20, the generation of maximum iteration 100.
2) individual evaluation：Each individual fitness in colony is calculated, fitness is the accurate of sample classification in this research
Rate.
3) Selecting operation:Each individual in colony is randomly choosed using selection opertor.Wheel is utilized in this research
The accuracy rate of disk gambling method combination individual evaluation is selected individual, so as to which the higher individual information of fitness can be genetic to
It is of future generation.
4) crossing operation：Restructuring is overlapped to the individual in individual using crossover operator, so as to produce new individual, collection
Characteristic information into previous generation's individual.
5) mutation operator：Random variation is carried out by probability to individual using mutation operator, ensure that the generation of new individual.
Colony obtains colony of future generation after selection, intersection, mutation operator.
6) terminate and judge:Reach if iterations if maximum algebraically or fitness reach necessary requirement and stop iteration.
The accuracy rate of training set is 96.25%, c=3.2277 after optimization, r=0.1354, as shown in Figure 9.On this condition,
Differentiation accuracy rate is 97.4684% (77/79, wherein rape honey 23/23, Mel 16/17, acacia honey 38/39).
The 8.3 SVM parameters selections based on particle cluster algorithm
Particle swarm optimization algorithm（Particle Swarm Optimization, PSO）It is similar with genetic algorithm, all it is logical
Initialization random individual is crossed, but each initial population was intersected and made a variation without the later stage in PSO, but fitted by calculating current individual
Gap carries out individual more between response functional value and colony's adaptive optimal control value.Compared to GA, to optimizing the guiding in direction only by most in PSO
It is excellent individual influence, and and it is not all individual carry out cross exchanged.Therefore, PSO convergence rate has larger improvement compared with GA.This experiment
The middle selection generation of iteration 200, population number 20.Figure 10 is PSO optimum results
Optimum results are：Training set highest accuracy rate is 91.25%, c=32.3362, r=0.0100.Under the conditions of this, in advance
It is 88.61% to survey accuracy rate（71/79, wherein rape honey 21/23, Mel 14/17, acacia honey 36/39）.
Compared to genetic algorithm, particle cluster algorithm is restrained faster, is just optimal a little in or so 6 generations.But its effect of optimization
Poor, very big reason is due to that colony's optimal value representativeness is more unilateral, has been absorbed in local minimum points.From the result of trellis algorithm
As can be seen that it is single convex function that SVM, which differentiates that accuracy rate function is not, under different parameters.Therefore, on this condition, although GA
Convergence rate is slower, but preferable compared with PSO in view of global individual, effect of optimization.
In three kinds of optimized algorithms, trellis algorithm needs certain priori conditions, such as substantially determine r and c span with
And the step-length of search.When scope is larger or step-length is small, search efficiency declines.Comparatively speaking, GA algorithm optimizations are better,
In or so 14 generations, can be optimal solution, while the predictablity rate after optimization（77/79）More initial（75/79）There is certain carry
It is high.
Research finally determines the use of genetic algorithm combination supporting vector machine mode identification method, and the electronic nose of acquisition is sensed
Device signal carries out discriminant classification, final to differentiate that accuracy rate is 97.46% after optimization.
Claims (2)
1. the parameter optimization method in a kind of electronic nose detection honey, signal difference maximum, which turns to, between different honey samples leads
To the optimal electronic nose testing conditions for distinguishing effect of selection, it is characterised in that testing conditions are divided into head space parameter and detection parameters,
Wherein detection parameters can be divided into sample introduction parameter and signal acquisition parameter again, it is contemplated that the detection for instrument of detection parameters reaction is special
Point, when instrument stabilizer, its influence to testing result is smaller, and head space parameter then influences the generation of sample headspace gas, and pushes up
Air body is then the direct detection object of electronic nose, that is, directly affects final testing result, the head space parameter of electronic nose is main
Final detection knot can be also influenceed including head space temperature and head space time, while in view of the difference of different sample sizes in ml headspace bottle
Fruit, therefore final choice sample size, head space temperature and head space time are optimization object, three factors are carried out using orthogonal experiment excellent
Change, when the varying level to each factor selects, it is contemplated that ml headspace bottle 10ml needs concussion to heat in head space room, is anti-
Only sample introduction needle touches fluid sample and influences instrument performance, and sample maximum may not exceed the 1/2 of ml headspace bottle, according to the close of honey
Spend for 1.4g/ml, three levels for determining headspace sampling amount are respectively 4g, 5g, 6g, head space temperature it is horizontally selected in, choosing
Three levels are respectively 40,50,60 DEG C, in head space selection of time, select the shorter head space time, three levels are respectively
120s, 180s, 240s, i.e. sample size 4g, 5g, 6g, head space temperature 40,50,60 DEG C, 180s, 240s, are ground head space time 120s
It is 5 kinds of different nectar source honey samples, respectively rape honey, acacia honey, chaste honey, honey of lychee flowers, linden to study carefully selected honey sample
Tree is sweet, every 6 parts of class nectar source sample, altogether 30 parts of samples；
For the orthogonal experiment of similar Almost Sure Sample Stability evaluation index, the lower 5 kinds of honey of every group of experiment takes 3 parts of samples, passes through meter
The standard deviation average of 18 sensors of electronic nose under the experiment condition to Different categories of samples signal is calculated, to weigh electronic nose under the conditions of this
To the stability C of similar sample signal detection, computational methods are shown below
Wherein p is nectar source species, CkIt is electronic nose number of probes for the stability of kth class sample, m, nkFor sample in k classes nectar source
This number,For i-th of sample in the sample of k classes nectar source jth root sensor response signal,Exist for k classes nectar source sample
The response signal average of jth root sensor；
To the otherness of inhomogeneity sample, according to inhomogeneity nectar source sample, the variance D of sample average is calculated under identical conditions
Arrive, computational methods are shown below：
Wherein p be nectar source species,For the signal average of i-th sensor of k classes sample,For i-th sensor of all samples
Signal average, nkFor the number of sample in same nectar source；
(3) the desired optimum optimizing condition obtained the signal stabilization for similar sample between of research, and between inhomogeneity sample with compared with
Big difference, the evaluation index q finally determined are shown below：
。
2. parameter optimization method according to claim 1, it is characterised in that according to China geographic area western part, south China, China
North, the division in East China, northeast, 5 kinds of different nectar sources of selection are respectively as study sample：1) rape honey, west area is picked up from
Fuling Chongqing and Yongchuan District；2) honey of lychee flowers, the Nanning of South China is picked up from；3) chaste honey, the north of North China is picked up from
Capital Miyun；4) acacia honey, the Laiyang Shandong Province in East China is picked up from；5) Mel, Jilin Dunhua and the Heilungkiang Harbin in northeast are picked up from；
It is stored in after the sample collection under the conditions of -18 DEG C, treats uniformly to be tested after all samples collection, 5 kinds of nectar sources
Sample respectively takes 60g, is respectively placed in 40 DEG C of constant water bath box, heating water bath 15min, melts honey sample, remaining sample after
Continuous be placed at -18 DEG C preserves, and per 3min, concussion is once during heating water bath；After the completion of sample water-bath, taking-up is placed in be cooled down at room temperature
More than 1h, until sample temperature is consistent with room temperature, then load honey sample in 10mL ml headspace bottle；
The Adsorption of different volatile ingredients is detected to testing sample honey using gas sensor array, the electricity
Sub- nose detection parameters are electronic nose sample size 1.000mL, extract μ L/s of speed 1000,65 DEG C of sample introduction needle temperature, acquisition time
120s, collection delay 600s.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310323227.5A CN104730140B (en) | 2013-07-30 | 2013-07-30 | A kind of parameter optimization method in electronic nose detection honey |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310323227.5A CN104730140B (en) | 2013-07-30 | 2013-07-30 | A kind of parameter optimization method in electronic nose detection honey |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104730140A CN104730140A (en) | 2015-06-24 |
| CN104730140B true CN104730140B (en) | 2018-02-02 |
Family
ID=53454238
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310323227.5A Active CN104730140B (en) | 2013-07-30 | 2013-07-30 | A kind of parameter optimization method in electronic nose detection honey |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104730140B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112036619B (en) * | 2020-08-17 | 2023-07-11 | 中国标准化研究院 | Method for judging whether roast duck exceeds goods shelf end point by combining electronic nose with Bayes algorithm |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5788833A (en) * | 1995-03-27 | 1998-08-04 | California Institute Of Technology | Sensors for detecting analytes in fluids |
| WO2002003047A2 (en) * | 2000-07-03 | 2002-01-10 | Michigan State University | Method and apparatus for the detection of volatile products in a sample |
| CN101382557B (en) * | 2008-09-04 | 2012-05-09 | 广东中烟工业有限责任公司 | Cigarette quality analytical method |
| CN102749362B (en) * | 2012-07-19 | 2014-06-11 | 浙江大学 | Method of optimizing peanut freshness detection conditions of electronic nose of gas sensor |
| CN102879436A (en) * | 2012-10-22 | 2013-01-16 | 浙江工商大学 | Method of using electronic nose for detecting freshness of river crucian carp |
-
2013
- 2013-07-30 CN CN201310323227.5A patent/CN104730140B/en active Active
Non-Patent Citations (2)
| Title |
|---|
| Electronic nose and neural network use for the classification of honey;Simona BENEDETTI et al.;《Apidologie》;20040831;第35卷;第1-5页 * |
| 电子鼻测定酸奶气味过程中测定参数的研究;郭奇慧 等;《乳业科学与技术》;20080831(第4期);第170-172页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104730140A (en) | 2015-06-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103487465B (en) | A kind of intelligence of the characterization of variation of honey otherness based on variance ratio sense of smell TuPu method extracting method | |
| CN103499646B (en) | Honey characteristic fragrance analysis and honey fragrance system simulation method | |
| CN103575764B (en) | A kind of honey detection method optimized based on algorithm of support vector machine | |
| CN103487558B (en) | A kind of method detecting exceptional sample during the pattern recognition analysis applying intelligent sensory signal to carry out tea leaf quality | |
| CN103487463B (en) | A kind of honey detection method of the support vector machine classifier Selecting parameter based on grid optimization | |
| Jiang et al. | A novel framework for analyzing MOS E-nose data based on voting theory: Application to evaluate the internal quality of Chinese pecans | |
| CN103487537A (en) | Detection method for producing areas of Xihulongjing tea based on genetic algorithm optimization | |
| CN103499609B (en) | A kind of method that honey fragrance intelligence sense of smell dynamic response feature and differentiation information dynamic characterization are studied | |
| CN104749219B (en) | A kind of honey detection method of the support vector machine classifier parameter selection based on particle cluster algorithm | |
| Shi et al. | Optimization of electronic nose sensor array by genetic algorithms in Xihu-Longjing Tea quality analysis | |
| CN106022528A (en) | Short-term power prediction method of photovoltaic power plant based on density peak hierarchical clustering | |
| Kang et al. | Determination of the quality of tea from different picking periods: An adaptive pooling attention mechanism coupled with an electronic nose | |
| CN109273096A (en) | A kind of risk management grading evaluation method based on machine learning | |
| CN103499608B (en) | A kind of intelligence of the characterization of variation of honey otherness based on ant group algorithm sense of smell TuPu method extracting method | |
| CN103499606B (en) | A kind of sign Mel diversity intelligence olfactory sensation TuPu method extracting method based on core principle component analysis | |
| Canizo et al. | Data mining approach based on chemical composition of grape skin for quality evaluation and traceability prediction of grapes | |
| CN106295686A (en) | The tea aroma sorting technique of parameter optimization support vector machine | |
| CN103499663B (en) | A kind of system of selection based on sensor in the Longjing tea Quality Detection Grade Model of genetic algorithm | |
| CN103499607B (en) | Exceptional sample point elimination method in a kind of detection by electronic nose honey | |
| Su et al. | Wine characterisation according to geographical origin using analysis of mineral elements and rainfall correlation of oxygen isotope values | |
| CN103488868B (en) | A kind of method of the intelligent smell discrimination model for setting up honey quality difference | |
| CN104730140B (en) | A kind of parameter optimization method in electronic nose detection honey | |
| CN103487464B (en) | A kind of sign Mel diversity intelligence olfactory sensation TuPu method extracting method based on individual event diagnostic method | |
| CN103487466B (en) | The Mel detection method that a kind of support vector machine classifier parameter based on genetic algorithm selects | |
| CN103499610B (en) | A kind of intelligence of the characterization of variation of honey otherness based on independent component analysis sense of smell TuPu method extracting method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |