CN103575764A - Honey detection method based on support vector machine algorithm optimization - Google Patents

Abstract

The invention relates to a honey detection method based on support vector machine algorithm optimization. The honey detection method comprises the following steps: selecting 5 different honey sources including oilseed rape honey extracted from Fuling district and Yongchuan district of Chongqing of the western China, honey of lychee flowers extracted from Nanning of Guangxi of the south China, chaste honey extracted from Miyun of Beijing of the north China, acacia honey extracted from Laiyang of Shandong and basswood honey extracted from Dunhua of Jilin and Harbin of Heilongjiang of northeast as study samples; detecting the to-be-detected sample honey by utilizing adsorption difference of a gas sensor array with different volatile components; changing surface layer current intensity of a semiconductor sensor after the volatile components of the honey are adsorbed with sensor features, and obtaining response curves of each sample through digital conversion so as to detect and analyze the samples and establish a support vector machine discrimination model by utilizing extracted electronic nose feature information to classify the samples of different honey sources.

Description

A kind of honey detection method of optimizing based on algorithm of support vector machine

Technical field

The application relates to Electronic Nose sensor technology, is specifically related to a kind of honey detection method of optimizing based on algorithm of support vector machine.

Background technology

China's honey output occupies first place in the world, and output keeps the trend of rapid growth always in recent years, and 25.2 ten thousand tons of 40.2 ten thousand tons of being increased to 2009 by 2001, account for Gross World Product and also by nearly 20%, bring up to more than 30%.But the driving due to economic interests, honey market is adulterated serious at present, cause adulterated honey to occupy 20%～30% of honey market, the bee product of some regional adulterated fraud accounts for 50% left and right, badly damaged consumer's interests, affects honey industry and develops in a healthy way, hits the export trade and earn foreign exchange.

Owing to lacking the impact of detection means, cause adulterated strike difficulties, its basic reason is as follows: (1) is because the main matter of honey itself is relatively simple for structure, comprise water and carbohydrate content, give the adulterated condition of providing convenience, meanwhile, depend merely on detect these several content of material number can not differentiate at all whether adulterated; (2) due to honey, be subject to the temperature and humidity of nectariferous plant kind, hive gesture power, sweet time phase length, air, and the various factors such as processing of honey, storage, crystallization, cause the content range of honey main matter to change greatly, make honey adulterated simple, convenient; (3) the detection of adulterations expense such as C4 high, cannot for reality, detect and law enforcement on a large scale.

Fragrance is one of important property of product quality embodiment, and product fragrance characterizes need to give prominence to its objectivity, authenticity and comprehensive.Current gas chromatography (GC), gas chromatography-mass spectrography (GC-MS) and gas chromatography-smell methods such as distinguishing (GC-O), limited monomer aroma substance in can only testing product, and between these fragrance, there is the phenomenons such as collaborative, modified tone, be difficult to reflect on the whole the flavouring essence quality of sample.And Intelligent Olfaction System (Electronic Nose) can be smelt news feature by simulating human, the Global Information of comprehensive characterization fragrance, olfactory characteristic and the overall quality of embodiment fragrance, simultaneously more objective, reliable than people's sense of smell.At aspects such as food freshness, the rotten differentiation of edible oil, the detection of fruits and vegetables degree of ripeness, the identification of tea-leaf producing area kind, drinks brand define, carried out correlative study at present.

In honey, contain more than 300 kind of aromatic substance, so it is the important sample that the intelligent sense of smell of research characterizes; Simultaneously different nectar sources, its flavor substance of the different place of production are different, and whether adulterated honey is or quality can embody to some extent on whole fragrance, makes fragrance become honey quality and detect one of important indicator with adulterated discriminating; Absolutely prove that adopting intelligent sense of smell to characterize honey quality has feasibility, also for honey quality detects and adulterated discriminating provide a kind of fast, economical, accurately and be beneficial to the detection method of real-time application.Therefore select honey to there is Practical significance as research object, its industry healthy development is had more to far-reaching value.

Adopt Electronic Nose to carry out product quality differentiation or adulterated discriminatory analysis, its essence is to utilize the whole fragrance information of intelligent sense of smell collection of illustrative plates, find the otherness of sample room, its core is to find the figure spectrum information of otherness between representative sample, i.e. " differentiation information ", is also " the differentiation figure spectrum information of intelligent sense of smell ".But the sensor array of Electronic Nose has cross-sensitivity, be that every sensor has response in various degree to each fragrance, therefore the aroma-producing substance collection of illustrative plates gathering by Electronic Nose has wide spectrum, the feature such as overlapping, be difficult to the naked eye from collection of illustrative plates, distinguish different samples separately, need to carry out " signal excavation ", particularly " excavation of differentiation information between representative sample ", the otherness information of excavation is more, just more contributes to distinguish efficiently product feature and quality.But also very weak aspect differentiation information excavating at present, be also the bottleneck of restriction Electronic Nose development.

Summary of the invention

The honey detection method that algorithm of support vector machine is optimized, selects 5 kinds of different nectar sources as study sample, is respectively: 1) rape honey, picks up from Fuling Chongqing and the Yongchuan District in west area; 2) honey of lychee flowers, picks up from the Nanning of South China; 3) chaste honey, picks up from the ground such as Miyun Region of Beijing of North China; 4) acacia honey, picks up from the Laiyang Shandong Province in East China; 5) Mel, picks up from Jilin Dunhua and the Harbin, Heilungkiang in northeast; Utilize gas sensor array from the Adsorption of different volatile ingredients, testing sample honey to be detected; Wherein after honey volatile ingredient and sensor characteristics absorption, change semiconductor transducer top layer strength of current, pass through digital conversion, obtain the response curve of each sample, thereby sample is detected to the Electronic Nose characteristic information that analysis and utilization extracts and set up support vector machine discrimination model, the sample in different nectar sources is classified.

Described honey detection method, for guaranteeing the not interference of examined condition difference of research, after described sample collection, be stored under-18 ℃ of conditions, treat to unify to test after all samples collection, 5 kinds of nectar source samples are respectively got 60g left and right, be placed in 40 ℃ of constant water bath box, heating water bath 15min, melts honey sample, remaining sample continues to be placed at-18 ℃ to be preserved, during heating water bath, for guaranteeing that sample melts completely, without crystallization, during water-bath, need every 3min concussion once; After sample water-bath completes, take out and to be placed under room temperature more than cooling 1h, until sample temperature is consistent with room temperature.

Described honey detection method, described gas sensor array is to adopt Fox 4000 type Electronic Nose, this Electronic Nose is comprised of 18 Metal Oxide Semiconductor Gas Sensing sensors and HS100 head space automatic sampler; Simple sample can obtain the signal matrix of 18 sensor * t detection time.

Described honey detection method, the pattern-recognition step of described support vector machine discrimination model is:

Select suitable kernel function K;

Solve corresponding optimization method, obtain support vector;

Obtain optimal classification function f (x);

According to the value of classification function, determine the classification of differentiating.

Accompanying drawing explanation

Fig. 1 abnormity point elimination result: (a) mahalanobis distance is differentiated result; (b) lever value is differentiated result;

The feature extraction result of Fig. 2 based on variance ratio

The feature point extraction result that Fig. 3 differentiates based on individual event amount

Fig. 4 ant group algorithm process flow diagram

The feature extraction result of Fig. 5 based on ant group algorithm

The feature point extraction result of Fig. 6 based on core principle component analysis

The feature point extraction result of Fig. 7 based on independent component analysis

Fig. 8 searches plain support vector machine parameter optimization result based on grid

The support vector machine parameter optimization result of Fig. 9 based on genetic algorithm

The support vector machine parameter optimization result of Figure 10 based on particle cluster algorithm

Embodiment

1 about sample collection and preparation

For making studied nectar source difference representative, according to the division in China geographic area (western part, south China, North China, East China, northeast), select 5 kinds of different nectar sources as research sample, be respectively: 1) rape honey, picks up from Fuling Chongqing and the Yongchuan District in west area; 2) honey of lychee flowers, picks up from the Nanning of South China; 3) chaste honey, picks up from the ground such as Miyun Region of Beijing of North China; 4) acacia honey, picks up from the Laiyang Shandong Province in East China; 5) Mel, picks up from the ground such as the Jilin Dunhua in northeast and Harbin, Heilungkiang.For guaranteeing authenticity and the accuracy of experiment sample, avoid the interference of market business honey processing technology, sample is directly buied by beekeeper place by China Agriculture Industitute Bee Research Center.

After sample collection, according to different nectar sources, the different place of production, be placed in respectively different reagent bottles.For guaranteeing the not interference of examined condition difference of research, after sample collection, be stored under-18 ℃ of conditions, treat to unify to test after all samples collection.Before experiment, after sample takes out from-18 ℃, 5 kinds of nectar source samples are respectively got 60g left and right, are placed in 40 ℃ of constant water bath box, and heating water bath 15min melts honey sample, and remaining sample continues to be placed at-18 ℃ to be preserved.During heating water bath, for guaranteeing that sample melts completely, without crystallization, during water-bath, need every 3min concussion once.After sample water-bath completes, take out and to be placed under room temperature more than cooling 1h, until sample temperature is consistent with room temperature (20 ℃).

2 detection by electronic nose methods

Electronic Nose utilizes gas sensor array from the Adsorption of different volatile ingredients, testing sample honey to be detected.After honey volatile ingredient and sensor characteristics absorption (comprising physisorption and chemisorption), change semiconductor transducer top layer strength of current.By digital conversion, obtain the response curve of each sample, thereby sample is detected to analysis.The present invention adopts Fox 4000 type Electronic Nose (Alpha MOS, France), and this Electronic Nose is comprised of 18 Metal Oxide Semiconductor Gas Sensing sensors (MOS) and HS100 head space automatic sampler.

Instrument concrete operations flow process is as follows:

1) honey sample that is cooled to room temperature after water-bath being added as requested to volume is in the head space bottle of 10ml.The head space bottle that installs sample is placed on pallet.HS100 automatic sampler holds at most 2 pallets, and each pallet can be placed 32 head space bottles.

2) set as requested instrument testing conditions, comprise head space sampling and detection by electronic nose condition.According to nectar source kind and detection order, each head space bottle on pallet is encoded.

3) head space bottle is placed into according to the condition arranging that head space is indoor to be heated, and during heating, head space bottle intermittently shakes, and guarantees head space gas homogeneity.After head space sample preparation finishes, extract head space gas, in Injection Detector, and by head space bottle from the indoor taking-up of head space.Fox 4000 is the injection of continuous type air-flow, and gas enters and detects after gas and each sensor generation adsorption and desorption reaction enclosure, and the response curve of each self-generating response.

Simple sample can obtain 18(18 root sensor) * t(detection time) signal matrix.Classic method is analyzed the maximum of each sensor (little) value as the response of this sensor.

3 honey quality modeling methods based on Electronic Nose information

The Electronic Nose characteristic information that utilization extracts is set up support vector machine discrimination model, and the sample in different nectar sources is classified.Traditional mode identification method is the progressive theory being based upon on great amount of samples basis, but individual in production application, restriction due to each side condition, a large amount of sample numbers is often difficult to be guaranteed preferably, under the condition of small sample, according to traditional statistical basis, be difficult to obtain comparatively ideal results of learning and extensive effect.But support vector machine is applicable to the modeling demand under condition of small sample, thus different nectar sources sample is carried out to pattern-recognition judgement.

Support vector machine (Support Vector Machine, SVM) theory is Vapnik(1995) on traditional statistical learning basis, integrated structure principle of minimization risk, proposes for the feature of finite sample.The method can effectively reduce the random row of setting parameter in traditional mode model of cognition, has overcome the deficiency of empiric risk and expected risk generation bigger difference in model process of establishing, and concrete SVM is theoretical as follows.

In pattern-recognition, obtain an optimization function f (x, w), make it to unknown sample collection (x _i, y _i) (i=1,2 ..., n; y

for specimen number) while assessing, expected risk is minimum:

Wherein, F(x, y) be joint distribution probability, L (y, f (x, w)) predicts y and the loss that causes with f (x, w), be called loss function, for two quasi-mode identification problems, L can be defined as:

In conventional learning algorithms, employing be empirical risk minimization principle,

But in fact, minimizing of training error is difficult to the optimum efficiency that assurance is predicted, often easily occurs the phenomenon of over-fitting, meanwhile, through further studying and show, there is following relation in empiric risk and practical risk:

Be abbreviated as

The VC dimension that wherein h is function, η is confidence level, n is training sample number.

As can be seen from the above equation, for making the classification function practical risk of design minimum, not only to make empiric risk reduce as far as possible, also will increase training set number simultaneously or reduce function VC dimension, could reduce practical risk.This thought is structural risk minimization.

Based on above theory, to sample set (x _i, y _i) (i=1,2 ..., n; X is the proper vector of sample i, y

for specimen number) while differentiating, find discriminant function

, W and b are normalized after also equal proportion adjusting, make can meet for all samples

, now the classifying distance of two class samples is spaced apart .Therefore for obtaining better classification prediction effect, should make two class samples separate as far as possible, ask minimum value.Meet point, inner classification plan range is minimum, they have determined optimal classification function, these points are referred to as support vector (Support Vector, SV).

Under this condition, to the problem of optimal classification function, can be converted into optimization problem:

Optimization problem is converted into dual problem and can be expressed as:

α wherein _ifor the Lagrange for constraint condition (7) (Lagrange) factor, i=1,2 ... n, the slope that W is classification function, the intercept that b is classification function.

For linearly inseparable problem, it is theoretical that V.Vapanik introduces kernel function, at lower dimensional space, data are passed through in Nonlinear Mapping projection value higher dimensional space, can prove, if select suitable kernel function, the inseparable data of lower dimensional space neutral line can be converted into the data that higher dimensional space neutral line can divide.Introduce after kernel function, full scale equation can be converted into:

Wherein K is selected kernel function.

By solving kernel function, finally can determine corresponding classification function:

The pattern-recognition step of SVM integral body can be summarized as several steps:

(1) select suitable kernel function K;

(2) solve corresponding optimization method, obtain support vector;

(3) obtain optimal classification function f (x)

(4) according to the value of sgnf (x), determine the classification of differentiating;

Parameter optimization in 4 detection by electronic nose honey

4.1 determine parameter to be optimized and level

Detection by electronic nose parameter can be divided head space parameter and detected parameters.Wherein detected parameters can be divided into again sample introduction parameter and signals collecting parameter.What consider detected parameters reaction is the detection feature of instrument, and when instrument stabilizer, its impact on testing result is less.Head space parameter affects the generation of sample headspace gas, and head space gas is the direct-detection object of Electronic Nose, directly affects final testing result.Therefore, in the present invention, emphatically head space parameter is optimized.The head space parameter of Electronic Nose mainly comprises head space temperature and head space time, considers that the difference of different sample sizes in head space bottle also can affect final testing result simultaneously, and therefore final selection sample size, head space temperature and head space time are optimization object.For selecting optimum combination, utilize orthogonal experiment to be optimized three factors.When the varying level of each factor is selected, consider that head space bottle (10ml) needs concussion heating head space is indoor, for preventing that sample introduction needle from touching fluid sample and affecting instrument performance, sample maximum can not surpass 1/2 of head space bottle.According to the density of honey (being about 1.4g/ml), determine that three levels of headspace sampling amount are respectively 4g, 5g, 6g.In the level of head space temperature is selected, according to list of references, honey sample character under higher than 68 ℃ of conditions easily changes, and therefore selected three levels are respectively 40,50,60 ℃.In head space selection of time, consider rate request, the short-term stability of honey sample under hot environment that large sample amount detects, and the effumability feature of honey sample, select the shorter head space time, three levels are respectively 120s, 180s, 240s.The final optimal conditions of determining Three factors-levels, i.e. sample size 4g, 5g, 6g, 40,50,60 ℃ of head space temperature, head space time 120s, 180s, 240s, each factor and level are as shown in table 1.Studying selected honey sample is 5 kinds of different nectar sources honey samples, is respectively rape honey, acacia honey, chaste honey, honey of lychee flowers, Mel, and every class nectar source 6 duplicate samples amount to 30 duplicate samples.

Testing selected nominal price experiment table is L ₉(3) ⁴, the design of experiment table is as shown in table 2

all the other testing conditions are as shown in table 3

4.2 optimize definite evaluation index and method

The present invention turns to guiding with signal difference maximum between different honey samples, selects the best detection by electronic nose condition of distinguishing effect.By optimum detection condition, expect signal stabilization between similar honey sample, and differ greatly between inhomogeneity honey sample, thus the maximization of Electronic Nose signal difference between assurance honey sample.

(1) similar Almost Sure Sample Stability evaluation index

In orthogonal experiment, test lower 5 kinds of honey for every group and get 3 duplicate samples, by calculating the standard deviation average of 18 sensors of Electronic Nose to Different categories of samples signal under this experiment condition, weigh the stability that under this condition, Electronic Nose detects similar sample signal.Computing method are as shown in formula 15,16

Wherein p is that nectar source kind, m are Electronic Nose number of probes, n _kfor the number of sample in k class nectar source, for the response signal of i sample j root sensor in the sample of k class nectar source,

for the response signal average of k class nectar source sample at j root sensor.(2) otherness of inhomogeneity sample

The otherness of inhomogeneity sample calculates according to the variance of inhomogeneity nectar source sample sample average under identical conditions, and computing method as shown in Equation 17

（17）

Wherein

for the signal average of k class sample,

signal average for all samples

(3) overall assessment index

The optimum optimizing condition that research institute's expectation obtains is signal stabilization between similar sample, and between inhomogeneity sample, has larger difference.Therefore, final definite evaluation index as shown in Equation 18

4.3 based on the maximized Electronic Nose parameter optimization of signal difference between honey sample result

Using the evaluation index determined as the observed reading of orthogonal experiment, and Orthogonal experiment results is as shown in table 4

As can be seen from Table 4, three kinds of factors all have a certain impact to Electronic Nose differentiation effect.Wherein factor A, factor B, sample size and head space temperature are directly proportional to signaling zone calibration, and factor C and sample headspace time dead space calibration are inversely proportional to, and that blank group changes is relatively little.This is that in head space bottle, the volatile concentration of testing sample increases gradually due to the increase of sample size and head space temperature, can increase for the special component content of distinguishing, and distinguishes effect better.And along with the increase of head space time, under the higher environment of temperature, sample volatile ingredient character changes, such variable effect the differentiation of sample class, therefore distinguish effect and decline.Control group is stable shows no significant difference between similar sample repeatedly, and testing result is reliable.For further Orthogonal experiment results being analyzed, experimental result is carried out to variance analysis, variance analysis is as shown in table 5.

in the result of variance analysis, can find out, three factors all have appreciable impact to the discrimination of sample, wherein factor A and factor B conspicuousness large (P<0.01), the variance contribution ratio of three kinds of factors is respectively 65.34%, 22.16% and 9.66%.It is larger that this result shows that sample size and head space temperature affect sample area calibration, and the head space time is less on the impact of sample area calibration.

Based on the above results, selecting optimum optimization combination condition is A3B3C1, i.e. sample size 6g, and head space temperature 60 C, head space time 120s, the inhomogeneity sample obtaining under this condition is distinguished best results.

Exceptional sample point in 5 detection by electronic nose honey is rejected

5.1 exceptional sample point eliminating principles

Before the Electronic Nose signal to honey carries out formal analysis and honey quality modeling, be the stability that guarantees analysis result and model, need to reject the exceptional sample in the whole sample of honey, guarantee accuracy and the reliability of picked up signal and sample information.Exceptional sample in detection by electronic nose honey comprises the abnormal of abnormal (as sample number, the classification) of sample information and testing result.Exceptional sample usually easily affects the variation tendency of overall signal, destroys the stable of disaggregated model and thinks.Therefore the rejecting of exceptional sample is extremely necessary.

Cause the factor of Electronic Nose exceptional sample mainly comprise following some: (a) error of sample collection is mainly divided with the mistake of collecting sample, and error coded is main; (b) variation of sample shelf time character, because sample has certain interval by collecting detect to analyze, during this time every under, easily there is the variation of chemistry, physical property in some unsettled samples; (c) error of operation, comprises and claims the error of sample, the cleanliness factor of container etc.; (d) error that instrument detects, due to the variation of testing environment and sensor properties, even by the pre-treatment of signal, still has and is difficult to completely detection signal be proofreaied and correct, and still has part signal and population mean signal to have significant difference.

For obtaining larger sample size, the present invention is directed to three kinds of nectar source honey that existing market occupation rate is larger and analyze, rape honey, Mel and acacia honey, research is for analysis and the rejecting of the exceptional sample point in great amount of samples.Wherein 76 rape honeys, 56 Mels, 112 acacia honeys, amount to 244 samples.This is organized in the research of the same application of sample and 7,8.

5.2 exceptional sample point elimination methods

(1) mahalanobis distance is differentiated

Mahalanobis distance (Mahalanobis) is the evaluation index of vectorial intensity in hyperspace, is a kind of important method that multivariate data exceptional value detects.Mahalanobis distance is by calculating the mean vector of sample data and the departure degree between covariance matrix comparative sample signal, and circular is as follows:

Wherein sample average is vectorial, and S is sample covariance matrix.

for the mahalanobis distance average of sample,

for mahalanobis distance standard deviation, λ is the threshold value of accepting scope, x _tbe the proper vector of t sample, T is proper vector average.In this patent, set hard-threshold λ=3. for acceptable mahalanobis distance scope under this threshold condition.

(2) lever value is differentiated

The size of sample thick stick embodies the degree of dependence of model to this sample, and lever value is larger, relies on greatlyr, on model, affects larger.The sample that is usually located at character to be analyzed two ends has larger lever value.Excessive lever value has considerable influence to model, is unfavorable for the stable of model.By the analysis to sample lever value, reject the larger special sample of model impact, thus the stability of increase model.It is as follows that lever value is differentiated concrete grammar:

1, by PCA, calculate the score matrix T of sample to be tested;

2, calculate test matrix H: ;

3, the lever value hi of each sample: hi is i diagoned vector in test matrix H,

;

Similar with mahalanobis distance differentiation, lever value is differentiated by setting hard-threshold, and remove and there is the special sample point compared with big lever value, thus the stability of assurance later stage forecast model.

Exceptional sample point in 5.3 honey detection by electronic nose is rejected validity check

For being rejected to effect, sample verifies, select Bayes (Bayes) method of discrimination to evaluate the accuracy rate of discrimination model before and after abnormity point elimination, compare and other mode identification methods, Bayes method of discrimination is more simple, without relevant parameter is optimized.The basic thought that Bayes differentiates is that supposition has certain understanding before sampling to studied object (totally), and conventional prior probability distribution is described this understanding.Then the sample based on extracting is revised priori understanding again, obtain so-called posterior probability and distribute, and various statistical inference all distributes to carry out based on posterior probability.Bayesian Decision is different from classical statistical method, and its distinguishing feature is exactly in the situation that guaranteeing that risk of policy making is as far as possible little, applies all possible information as far as possible.

Fig. 1 is that the mahalanobis distance of analyzing samples (76 parts of rape honeys, 56 parts of Mels, 112 parts of acacia honeys) is differentiated (a) and lever value is differentiated result (b).Utilize mahalanobis distance to differentiate, reject altogether 36,53,76,79,85,99,117,240,244, amount to 9 abnormity point; And lever is differentiated and to be rejected altogether 36,79,216,240,244 totally 5 abnormity point.

The data that adopt Bayes to differentiate after two kinds of abnormity point methods are processed are carried out pattern-recognition prediction, predict the outcome as shown in table 6.By table 6, find, compare lever and differentiate, it is more that mahalanobis distance is differentiated the peculiar sample spot of rejecting, but accuracy rate there is no too large difference.This result shows the mahalanobis distance exceptional sample points of rejecting more, although compare and differ larger with the population distribution of sample, but to the result accuracy rate of differentiating not being affected greatly, therefore, in order to fully take into account the complete characteristic of sample, select lever to differentiate, reject 5 sample spot, i.e. the 36th, 76,216,240,244 5 exceptional samples, wherein rape honey, each 1 of Mel, 3 of acacia honeys.

6 honey characteristic perfumes are analyzed and honey fragrance simulated system is set up

Apply dynamic head space (Itex) and extract honey aroma-producing substance in conjunction with circulation beneficiation technologies, at chromatographic column end, carry out after the distribution of 1:1 fragrance content, application gas chromatography mass spectrometry (GC-MS) is measured its volatility with gas chromatography-measurement of olfaction (GC-Olfactometry, GC-O) technology simultaneously and is fragrant composition and sensory characteristic.Crystallized honey carries out heating water bath, is then cooled to rapidly room temperature, and in holding chamber, temperature constant state gathers aroma-producing substance.

Wherein, in GC-MS, utilize mass spectrum (library searching), Relative Retention Indices (RI) and smell three kinds of methods of news and determine the volatile ingredient of honey, and carrying out inner mark method ration.GC-O technology is the method that proportion detects and detected intensity combines, and by 5, preferably smells the GC-O evaluation group that the person of distinguishing forms, definite honey head perfume, front end fragrance, body note and active fragrance of the characteristic flavor on basis in four volatilization stages of bottom note of representing respectively.

According to the characteristic perfume contamination ratio in four volatilization stages, proportioning builds basic honey fragrance simulated system A.On the basis of system A, build the four group systems variant with it.The difference of every group of system and primary structure A is embodied in two aspects, different in certain volatilization stage or its characteristic perfume content, or its characteristic perfume component is different, and the aroma component of other three phases and content are all constant.

7 characterize the intelligent sense of smell collection of illustrative plates feature extraction of honey otherness

7.1 feature extracting methods based on variance ratio

Each signaling point of every sensor is calculated the middle variance of its sample and plants internal variance ratio, according to the size of variance ratio, signaling point is selected.The computing method of variance are with the evaluation index q in optimal conditions.Different with other embedded feature selecting, variance ratio is selected directly to pass through relatively under each information point, between kind, variance is recently selected information point with kind internal variance, need be by the method for other pattern discriminations, so the selection result of the method can be because selecting different mode recognition methods change.But Variance ratio method belongs to the method for exhaustion, larger to the operand of large sample.

Fig. 2 shows the variance ratio of a signaling point.From figure, can find, the information point that variance ratio differs greatly more concentrates on 900-1200 and 1800-2160, and the 8th to the 10th, 15-the 18th sensor.For same root sensor, variance differs greatly and a little concentrates on detection time is that 20s is in the signaling point of 35s.In this time period, be mainly the adsorption time of volatilization gas and sensor, and the signaling point in the detection later stage of each sensor is that in desorption time point, difference is less.Experimental selection variance ratio is greater than 1 signaling point as unique point, and its entrained information of signaling point that meets this condition can be reacted the difference of different sample rooms, selects altogether 798 unique points.With this understanding, utilize SVM discrimination model to integrate than predicting as the ratio of 2:1 in the checking of modeling collection, finally differentiating 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

All unique points are carried out to pattern-recognition one by one, relatively when each signaling point, differentiate the difference of accuracy rate during as single features.The method embeds mode identification method in feature selecting, by conjunction with method of discrimination, can obtain the ability of each signaling point to sample prediction.Different from front a kind of filtering method, the method relies on selected mode identification method, and selection result can change a lot with the change of method of discrimination.The discriminant criterion of selecting in this research is Bayes diagnostic method

From Fig. 3, find, different from variance selection result, in unidirectional amount selection, between different sensors, the difference of accuracy rate is less, and with differing greatly between the information point under different detection times in root sensor.But the variation tendency of time point Bayes differentiation accuracy rate is roughly consistent with variance ratio variation tendency in different sensors, detect initial stage signal differentiation accuracy rate higher, concentrate on the front 30s of each sensor in detection time, the effect that detects the later stage is poor.Selection differentiation accuracy rate is greater than 60% signaling point as unique point, and totally 598 unique points, utilize svm to verify.SVM predictablity rate is 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, need to calculate one by one each unique point.When unique point is more, calculated amount can be very large, this also fundamentally limit value its feature extraction for a large amount of signaling points.Ant group algorithm belongs to Heuristic Feature system of selection, utilizes the automatic Iterative of algorithm to evolve, and unique point is selected to carry out automatic optimal, until obtain optimal result.

Ant group algorithm (Ant Colony Optimization, ACO) is applied to travelling salesman's Path Selection at first, shortest path is optimized.In this experiment, ant group algorithm is applied to the selection of unique point.Algorithm simulation genetic algorithm is utilized binary coding to pass proper vector and is encoded each, and this information point is selected in 1 representative, and this information point is given up in 0 representative.Utilize Bayes differentiation accuracy rate and selected feature after each unique point is selected to count as fitness function, seek optimum vector combination.This algorithm main innovate point comprises: (a) unique point is selected number to add in fitness function, and set cost parameter, by parameter, regulate, can to feature, count as required and differentiate accuracy rate and accept or reject; (b) for to keep away the renewal anisotropy causing due to particular point, optimum collection is set, with optimal set, replace single optimum point to select; (c) pheromones renewal degree improves and is directly proportional to fitness function, and algorithm optimization is effective, and renewal amplitude increases; (d) for accelerating computing velocity, the vector that effect is poor is accelerated to evaporation rate, reduce pheromone concentration, reduce it later stage is calculated and disturbed.Algorithm flow as shown in Figure 4.

In ant group algorithm, each parameter is selected as follows: finally select ant group scale (m)=20; Pheromones volatilization concentration (rho)=0.003; Outstanding ant collection (n1)=3; Poor ant collection (n2)=3; Characteristic number punishment ratio (A)=400; The concentration of pheromones when figure below is a final generation.

The 9th, the signaling point of 15-18 root sensor as can be seen from Figure 5, the selected unique point of ant group algorithm focuses mostly near 1000 and 1500 to 2160.From result, can find out, inspire class algorithm owing to being automatic evolutional algorithm, although algorithm has certain randomness, to the overall distribution of feature letter signal is had to good selection.With this understanding, finally select feature to count 206, differentiation accuracy rate is 94.94%(75/79, and wherein rape honey 22/23, Mel 16/17, acacia honey 37/39).Ant group algorithm result is compared first two exhaust algorithm, differentiates accuracy rate and has a certain upgrade, and selected characteristic signal point still less, has more representativeness simultaneously.

7.4 feature extracting methods based on core principle component analysis

First three is planted extracting method and only signal itself is screened, and the selected unique point of these class methods has certain chemical sense, in conjunction with chemical results, can make an explanation preferably.But in the vector of giving up more after all, entrained some information of having selected vector not possess, are difficult to the accuracy rate that guarantees that the later stage is differentiated.Except direct extraction, utilize dimension reduction method, by matrixing, data message is compressed, effective information is carried out to enrichment, can significantly reduce characteristic signal quantity.In this research, selected two kinds of dimension reduction methods of core principle component analysis and independent component analysis to extract.

Core principle component analysis (Kernel Principal Component Analysis, KPCA) is introduced kernel function in principal component analysis (PCA) (Principal Component Analysis, PCA).KPCA utilizes kernel function, by data projection to higher dimensional space.Owing to more disperseing each other after data projection, therefore some inseparable signals in lower dimensional space can be distinguished, and extract and have more representational feature and extract.In this experiment, KPCA selects radial basis kernel function.

Under KPCA, the SVM predictablity rate of different number of principal components as shown in Figure 6.Fig. 6 demonstration is down to original signal under different dimensions through KPCA, the accuracy rate of forecast set.As can be seen from the figure, SVM differentiation accurately takes the lead in increasing with dimension, occurs afterwards of short duration plateau, 25, during to 30 dimension, differentiating accuracy rate increases and significantly increases with dimension, and accuracy rate is relatively stable afterwards, after 100 dimensions, the accuracy rate of sample increases and reduces with dimension.This Changing Pattern shows, in dimension lower-order section, each proper vector all carries the accurate classified information of sample, and between information, existence of redundant is less, increases dimension and contributes to improve and differentiate accuracy rate.After 50 dimensions, there is redundancy, covering in the information between individual characteristic quantity, and accuracy rate is differentiated in the now increase of dimension impact on promoting declines.When the later stage, the redundancy between feature has had influence on differentiation effect, therefore now differentiates accuracy rate and starts to decline.Finally select when dimension d=81, predictablity rate is the highest, is 93.67(74/79, and 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 (PCA) (comprising core principle component analysis) is all to maximize and to classify according to variance between data, i.e. the second moment of data, but ignored the independence of data in High Order Moment.Independent component (Independent components analysis, ICA) utilizes the High Order Moment between computational data to convert matrix, can further reduce the associated row between proper vector, strengthens signal compression effect.

The ICA method that this experiment adopts is fastica.Before algorithm carries out, data are carried out to albefaction processing.Under different independent components, svm differentiates accuracy rate as shown in Figure 7

The overall variation trend of ICA and KPCA type, relax but change relatively, and rate of accuracy reached required intrinsic dimensionality when stablizing is few compared with KPCA.Result shows, when independent component is 14, differentiation accuracy rate is 94.94%(75/79, and wherein rape honey 22/23, Mel 16/17, acacia honey 37/39).

The foundation of 8 support vector machine disaggregated model optimization methods

This research and utilization svm classifier device is as the disaggregated model of different nectar sources sample.Wherein, selected kernel function is that radial basis kernel function (RBF) is:

Wherein r is the form parameter of RBF function, x _iwith x _jfor two samples in sample set.

Compare and other kernel functions, select RBF to mainly contain following two kinds of reasons: 1) RBF can complete the linear nonlinear mapping of arriving, by mathematic(al) manipulation, can prove that linear kernel function is only a kind of special case of RBF; 2) than polynomial kernel function, RBF parameter is less, and model is relatively simple, and this has also guaranteed the stability of model.

Meanwhile, consider that being difficult to all training points of requirement meets constraint function (7), training points is introduced to slack variable ξ, constraint function (7) can be changed into

（22）

Slack variable ξ=(ξ ₁, ξ ₂, ξ ₃... ξ _n) ', embodied all training sets by mistake minute situation.Therefore introduce penalty c as the weighted value of interval between balanced class and wrong minute degree, majorized function (6) can be converted to

（23）

Take in the svm classifier device that RBF is kernel function, the classifying quality under different parameters (form parameter r and penalty parameter c) has larger difference, and therefore, this research and utilization Different Optimization method, is optimized the penalty coefficient in the middle r of RBF and punishment parameter.The data of selecting in this research are the data after ICA dimensionality reduction in 7.

Concrete Optimizing Flow is as follows:

1, divide in proportion training set and checking collection, ratio is 2:1.By training set according to five folding cross validation methods, be about to training set and be divided into 5 non-cross subsets, select in turn 4 subsets wherein to carry out parameter training, the parameter of selecting with a remaining sub-set pair is verified, calculates the classification accuracy of training set under different parameters.

2,, according to selected kernel function, set kernel functional parameter r and penalty parameter c.

3, with selected parameter according to five folding cross-validation methods in 1 to model training, and calculate the accuracy rate of different parameters drag.

4, whether judgment models accuracy rate is up to standard, otherwise change parameter value.

5, repeat 2,3,4 steps, until obtain best model differentiation rate, or reach stopping criterion for iteration.

In this patent, utilize different parameter searching methods, form parameter r and penalty parameter c in step 2 are optimized, finally obtain optimization model.

8.1 SVM parameters based on grid optimization are selected

Grid optimization utilizes the method for exhaustion, in the span of pre-estimating by certain step-length to searching for a little one by one in scope, determine final optimal parameter.Take 2 as the truth of a matter, 2 ^-4to 2 ¹⁰between r and c are carried out to exhaustive search.Work as c=5.2780, during r=0.1088, it is the highest that training set sample is differentiated accuracy rate, is 96.25% as shown in Figure 8.With this understanding, set up model, utilize forecast set to test.Final differentiation accuracy rate is 96.20% (76/79, wherein rape honey 23/23, Mel 16/17, acacia honey 37/39).

8.2 SVM parameters based on genetic algorithm are selected

Genetic algorithm (Genetic Algorithm, GA) is carried out heuristic search by using for reference biological evolution theory to data, and first this algorithm is proposed by the J.Holland professor of the U.S. the earliest for 1975.The fundamental operation process of genetic algorithm is as follows:

1) data initialization: maximum evolutionary generation is set, the random number of individuals generating and the colony forming thereof.In this research, select 20 of number of individuals, 100 generations of maximum iteration time.

2) individual evaluation: calculate each individual fitness in colony, the accuracy rate that in this research, fitness is sample classification.

3) Selecting operation: utilize and select operator to select at random each individuality in colony.In this research, utilize roulette method in conjunction with the accuracy rate of individual evaluation, individuality to be selected, thereby the higher individual information of fitness can be genetic to the next generation.

4) crossing operation: utilize crossover operator to the restructuring that superposes of the individuality in individuality, thereby produce new individuality, the characteristic information in integrated previous generation's individuality.

5) variation computing: utilize mutation operator to carry out random variation to individuality by probability, guaranteed the generation of new individuality.

Colony is through obtaining colony of future generation after the computing of selecting, intersect, make a variation.

6) stop judgement: if iterations reaches maximum algebraically or fitness, reach necessary requirement and stop iteration.

After optimizing, the accuracy rate of training set is 96.25%, c=3.2277, r=0.1354, as shown in Figure 9.With this understanding, differentiating accuracy rate is 97.4684% (77/79, wherein rape honey 23/23, Mel 16/17, acacia honey 38/39).

8.3 SVM parameters based on particle cluster algorithm are selected

Particle swarm optimization algorithm (Particle Swarm Optimization, PSO) similar with genetic algorithm, all by initialization random individual, but in PSO, each initial population was intersected and variation without the later stage, but carried out individuality more by calculating gap between current ideal adaptation degree functional value and colony's optimal-adaptive value.Compare GA, in PSO, on optimizing the guiding of direction, affected by optimum individual, and be not that all individualities carry out cross exchanged.Therefore, the speed of convergence of PSO has larger improvement compared with GA.In this experiment, select 200 generations of iteration, population number 20.Figure 10 is PSO optimum results

Optimum results is: the high-accuracy of training set is 91.25%, and c=32.3362, and r=0.0100.Under this condition, predictablity rate is 88.61%(71/79, and wherein rape honey 21/23, Mel 14/17, acacia honey 36/39).

Than genetic algorithm, particle cluster algorithm convergence is faster, just reaches optimum point about 6 generations.But its effect of optimization is poor, very large reason is because colony's optimal value representativeness is more unilateral, has been absorbed in local smallest point.From the result of trellis algorithm, can find out, under different parameters, SVM differentiation accuracy rate function is not single convex function.Therefore, with this understanding, although GA speed of convergence is slower, consider that the overall situation is individual, effect of optimization is better compared with PSO.

In three kinds of optimized algorithms, trellis algorithm needs certain priori conditions, as roughly determined r and the span of c and the step-length of search.When scope is large or step-length hour, search efficiency declines.Comparatively speaking, GA algorithm optimization effect is better, about 14 generations, can reach optimum solution, and more initial (75/79) of predictablity rate (77/79) after simultaneously optimizing improves.

Research is finally determined and is utilized genetic algorithm combination supporting vector machine mode identification method, and the Electronic Nose sensor signal of obtaining is carried out to discriminant classification, and after optimizing, finally differentiating accuracy rate is 97.46%.

Claims (4)

1. a honey detection method of optimizing based on algorithm of support vector machine, selects 5 kinds of different nectar sources as study sample, is respectively: 1) rape honey, picks up from Fuling Chongqing and the Yongchuan District in west area; 2) honey of lychee flowers, picks up from the Nanning of South China; 3) chaste honey, picks up from the ground such as Miyun Region of Beijing of North China; 4) acacia honey, picks up from the Laiyang Shandong Province in East China; 5) Mel, picks up from Jilin Dunhua and the Harbin, Heilungkiang in northeast; Utilize gas sensor array from the Adsorption of different volatile ingredients, testing sample honey to be detected; Wherein after honey volatile ingredient and sensor characteristics absorption, change semiconductor transducer top layer strength of current, pass through digital conversion, obtain the response curve of each sample, thereby sample is detected to the Electronic Nose characteristic information that analysis and utilization extracts and set up support vector machine discrimination model, the sample in different nectar sources is classified.

2. honey detection method according to claim 1, for guaranteeing the not interference of examined condition difference of research, after described sample collection, be stored under-18 ℃ of conditions, treat to unify to test after all samples collection, 5 kinds of nectar source samples are respectively got 60g left and right, be placed in 40 ℃ of constant water bath box, heating water bath 15min, melts honey sample, remaining sample continues to be placed at-18 ℃ to be preserved, during heating water bath, for guaranteeing that sample melts completely, without crystallization, during water-bath, need every 3min concussion once; After sample water-bath completes, take out and to be placed under room temperature more than cooling 1h, until sample temperature is consistent with room temperature.

3. honey detection method according to claim 1, described gas sensor array is to adopt Fox 4000 type Electronic Nose, this Electronic Nose is comprised of 18 Metal Oxide Semiconductor Gas Sensing sensors and HS100 head space automatic sampler; Simple sample can obtain the signal matrix of 18 sensor * t detection time.

4. honey detection method according to claim 1, the pattern-recognition step of described support vector machine discrimination model is:

A selects suitable kernel function K;

B solves corresponding optimization method, obtains support vector;

C obtains optimal classification function f (x);

D determines the classification of differentiating according to the value of classification function.

Images