CN106770967B - The non-targeted interference gas recognition methods of electronic nose based on a kind of local expression model - Google Patents
The non-targeted interference gas recognition methods of electronic nose based on a kind of local expression model Download PDFInfo
- Publication number
- CN106770967B CN106770967B CN201710009931.1A CN201710009931A CN106770967B CN 106770967 B CN106770967 B CN 106770967B CN 201710009931 A CN201710009931 A CN 201710009931A CN 106770967 B CN106770967 B CN 106770967B
- Authority
- CN
- China
- Prior art keywords
- sample
- gas
- expression
- tested
- interference
- 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
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
- G01N33/0004—Gaseous mixtures, e.g. polluted air
Landscapes
- Chemical & Material Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Medicinal Chemistry (AREA)
- Food Science & Technology (AREA)
- Combustion & Propulsion (AREA)
- Physics & Mathematics (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Immunology (AREA)
- Pathology (AREA)
- Investigating Or Analyzing Materials By The Use Of Fluid Adsorption Or Reactions (AREA)
Abstract
The present invention provide the non-targeted interference gas recognition methods of electronic nose based on a kind of local expression model, it the following steps are included: step 1, in fixed object gas sample X s In find a kind of local expression model of sample to be tested y；Step 2 calculates expression coefficient vector α；Step 3 obtains Optimal error detection threshold valueT*；Step 4, the residual error for acquiring sample to be tested y and expression valueRESIfRES≤T*, then sample to be tested y is object gas sample；IfRES>T*, then sample to be tested y is abnormal interference gas sample.The solution have the advantages that: reduce operand and subtract, and improves the accuracy of gas detection.
Description
Technical field
The invention belongs to a kind of field of gas detection of electronic nose.
Background technique
Electronic nose, i.e. a kind of Artificial Olfactory of mimic biology nose, by metal oxide sensor, Signal Pretreatment system
The part such as system and pattern recognition unit forms.The current application of electronic nose is very extensive, such as to meat, dairy products, tealeaves, pollution
The detection and diagnosis of gas and disease.Since smell is the concentrated expression of Multiple components, the gas perception part of electronic nose is logical
Often by multiple there are different selective metal oxide sensor arrays to constitute, it can be complete using the cross-sensitivity of sensor
At the simulated implementation of Biologic Olfaction function.There are two large problems for electronic nose: first is that the value of sensor generates drift with the working time
It moves；Second is that sensor will receive the interference of abnormal gas, such as alcohol, perfume, oil smoke etc..Sensor drift problem is by domestic
The effort of outer scholar already has effective solution scheme at present, but does not have for the processing of abnormal gas interference problem
Obtain biggish progress.
Why there is abnormal interference gas in electronic nose, be attributed to the cross-sensitivity of metal oxide sensor.
When sensor generates response to object gas to be detected, while also to ambient enviroment existing interference signal at random
It is responded, and the sensor response that abnormal interference gas generates often is much higher than object gas.In this case,
Obtained sensor response differs greatly with expected response value, and electronic nose olfactory system is caused to can not work normally.
Abnormal interference gas has a following characteristics in actual environment: first, interference gas is all likely to occur at any time, electronic nose
Can not predict when which kind of interference gas can generate, therefore there is great stochastic uncertainty；Second, in real life,
Interference gas specimen types are a lot of, we can not obtain potential all interference gas samples, if known using traditional mode
Other method, such as SVM, the classification methods such as BP neural network and naive Bayesian then need the mould of known every kind of interference gas
Formula.And in practical application, the type of interference gas may have thousands of or even up to ten thousand kinds, obtain the mode letter of all interference gas
Breath is unpractical.Therefore, traditional mode identification method is used in the detection of electronic nose interference gas becomes not conforming to reality
Border.
The electronic nose that Chinese patent literature CN106124700A discloses a kind of band from expression on November 16th, 2016 is non-
Target jamming Gas Distinguishing Method, it includes step 1, the multi-class targets gas data for taking electronic nose to detect and a small amount of non-targeted dry
Disturb gas data；Step 2, according to object gas training set, solve expression coefficient matrix α=[α1,α2,…,αN]；Step 3 makes
With object gas training set and interference gas error threshold training set, calculate the two training sets i.e. mean error collection e1 of sum and
e2；Step 4 determines the search range [E for distinguishing object gas and non-targeted interference gas threshold value T according to E=[e1, e2]min,
Emax], the detection for obtaining the accuracy P1 and non-targeted interference gas training set of object gas training set for each T value is accurate
Spend P2；Step 5, P=P1+P2, the maximum corresponding T value of P value is selected desired threshold.
CN106124700A establishes a self model using object gas data, to realize object gas and interference gas
Differentiation under test gas is determined as object gas if under test gas is consistent with object gas model, conversely, then to be different
Normal interference gas, such interference gas test problems translate into oneclass classification problem, are and be not asking for object gas
Topic.Find out from the patent, the inherent attribute that object gas itself has, it is weaker or complete with abnormal interference gas correlation
It is uncorrelated, that is to say, that object gas can be carried out indicating certainly by the linear combination of known target gas sample collection, and table
It is smaller up to error or even level off to 0；Since abnormal interference gas and object gas correlation are unobvious, abnormal interference gas
It can not accurately be indicated by the linear combination of object gas sample, i.e. expression error is larger.It is dry that the patent solves magnanimity
The problem of disturbing gas data modeling, still, which uses target complete gas sample collection as a dictionary to be expressed
Object gas model (known target gas sample X of the present patent application the fixation for expressionsRegard dictionary as.It is deposited in dictionary
What is put is each sample.Expression sample for each sample to be tested is selected in dictionary), so being asked there are following
Topic: first is that operation is complicated；Second is that causing detection effect not good enough because of the participation of the weak related objective sample in part.
Summary of the invention
For the technical problems in the prior art, the technical problem to be solved by the invention is to provide one kind to be based on
The non-targeted interference gas recognition methods of the electronic nose of a kind of local expression model, it passes through selection and the maximally related mesh of sample to be tested
Mark gas sample as one group of base (i.e. with the most similar target sample X of sample to be tested yk) to improve the performance of expression model, it is real
Existing operand is reduced, and improves the accuracy of gas detection.
Insight of the invention is that having selected the k object gas sample nearest from sample to be tested Euclidean distance as one group
Effective base Xk, to express sample to be tested y.This k object gas sample be in object gas sample set X with sample to be tested y most
It is similar, that is to say, that compared with other object gas samples in addition to this k sample, this k object gas sample is with more
A possibility that big, realizes the optimum expression of sample to be tested y, and then efficiently avoids other weak related objective gas samples to table
The negative effect reached.
" a kind of local expression model " described in present patent application refers to not by known target gas sample XsIn own
Sample express y, but have chosen XsIn k object gas sample X maximally related from sample to be tested yk, i.e. fractional sample table
It reaches；So-called one kind, i.e., (electronic nose detects multiple gases to only a kind of sample, then these types of gas is for electricity in training set
All be object gas for sub- nose, therefore, object gas be attributed to same class gas, i.e., a kind of sample), realize other class gas
The detection of body.
In order to solve the above technical problems, the present invention the following steps are included:
Step 1, in fixed object gas sample XsIn find a kind of local expression model of sample to be tested y
The optimized-type of sample to be tested y one kind local expression model is as follows:
In formula,It is expression coefficient vector, XkIt is in object gas sample XsIn with it is to be measured
The maximally related k object gas sample of sample y；0 λ≤1 < and 0 μ≤1 < are regular coefficient, and R (α) is canonical standardization norm；
αp,αq∈ α, αpAnd αqIt is any two value expressed in coefficient vector α, shows respectively XkIn p-th of sample and q-th of sample
Expression coefficient, wpqFor p-th of neighbour's object gas sample xpWith q-th of neighbour's object gas sample xqSimilarity degree；
Step 2 calculates expression coefficient vector α
Express the solution formula of coefficient vector α are as follows:
In formula, L=D-M
R (α) has two kinds of regularization modes of a norm and two norms, MpqCorresponding value is arranged for pth row q in matrix M；D is
One diagonal matrix, DppIt is the value of p-th of position of diagonal line in matrix D；σ2It is the variance of Gaussian function, is a constant, this
Step is solved using the method for ADMM；
Step 3 obtains Optimal error detection threshold value T*
The expression error of object gas training sample are as follows:
The expression error of interference gas training sample can be expressed as:
αwiIndicate single target gas training sample wiThe expression coefficient vector of ∈ W, αhiIndicate single interference gas training
Sample hiThe expression coefficient vector of ∈ H,Indicate wiK neighbour's target sample,Indicate hiK neighbour's target sample,
N is object gas number of training, and n is interference gas number of training；
Optimal error detection threshold value T* are as follows:
Step 4 acquires sample to be tested y and expression value XkThe residual error RES of α
After obtaining expression coefficient vector α by step 2, the expression error of sample to be tested y passes through y and expression value Xkα's is residual
Difference is indicated, i.e.,
Judge that sample to be tested y belongs to object gas or interference gas by judging the size of residual error RES, if RES
≤ T*, then sample to be tested y is object gas sample；If RES > T*, sample to be tested y is abnormal interference gas sample
This.
Compared with the non-targeted interference gas recognition methods of existing electronic nose, the present invention has the advantage that:
1.XkIt is based on dictionary XsThe obtained object gas sample with superperformance is screened, there is excellent sample
Energy；
2. the present invention is not expressed with all samples in dictionary when identifying sample to be tested y, word is selected
The several object gas sample Xs nearest from y in allusion quotationk, the fortune of model will be substantially reduced in this way by being expressed by less data
Calculate complexity；
3. due to the object gas sample in dictionary of the invention there are multiclass, and sample to be tested y may with it is therein certain
A kind of or two classes are more related, completely uncorrelated to other classes, therefore, pass through k nearest object gas sample X of selectionkJust
It can avoid influencing brought by other incoherent classes well, so that y is by the expression of k neighbour's object gas sample
It is more accurate, substantially increase the accuracy of gas detection.
Detailed description of the invention
Detailed description of the invention of the invention is as follows:
Fig. 1 be L2 norm when target-interference gas discrimination with threshold value T change curve；
Fig. 2 be L1 norm when target-interference gas discrimination with threshold value T change curve；
Test response curve of the sensor to interference data set 1 when Fig. 3 is L2 norm；
Test response curve of the sensor to interference data set 2 when Fig. 4 is L2 norm；
Test response curve of the sensor to interference data set 1 when Fig. 5 is L1 norm；
Test response curve of the sensor to interference data set 2 when Fig. 6 is L1 norm.
Specific embodiment
Present invention will be further explained below with reference to the attached drawings and examples:
Symbol description used in present patent application:It is known object gas sample,It is
In XsIn with the maximally related k object gas sample of sample to be tested y, T* is Optimal Error detection threshold value,Indicate to be measured
Sample, D are characterized dimension, NsFor object gas total sample number, NkFor with the maximally related k object gas total sample number of y；Generation
Table L2 norm, | | | |1Represent L1 norm, ()TIndicate transposition operation, ()-1Indicate inversion operation.Full text is capitalized thick black
Body representing matrix, the extrabold of small letter indicate that vector, variable are indicated with italics.
Step of the invention is:
Step 1, in fixed object gas sample XsIn find a kind of local expression model of sample to be tested y
By sample to be tested y respectively with dictionary XsIn each sample calculate Euclidean distance, select Euclidean distance nearest k
Neighbour's object gas sample.
For the single sample y in sample to be tested collection Y, it is found first in dictionary XsK interior neighbour's object gas sample
This Xk=[x1,x2,...,xk], sample y and the relationship of this k neighbour's object gas sample can be expressed as
Y=x1α1+x2α2+...xkαk+e (1)
In order to enable this k neighbour's object gas sample to express sample to be tested y to greatest extent, that is to say, that expression misses
Poor e is the smaller the better, therefore show that the optimization of model is expressed as follows:
In formula (2),It is expression coefficient vector；
Due to k neighbour's object gas sample around sample to be tested y it is known that then further utilizing the priori of neighbor relationships
(priori knowledge refers to that k neighbour's object gas sample of sample to be tested y has obtained to knowledge, belongs to known object gas sample
This, then the distribution situation of these object gas samples and spatial relation are all known, then these can be utilized
Known spatial relation constrains α) as follows to α progress manifold canonical restrict:
In formula (3), αp,αq∈ α, αpAnd αqIt is any two value expressed in coefficient vector α, shows respectively XkMiddle pth
The expression coefficient of a sample and q-th of sample, wpqFor p-th of neighbour's object gas sample xpWith q-th of neighbour's object gas sample
This xqSimilarity degree, acquired by gaussian kernel function；
If xpWith xqIt is more similar, by the calculated w of kernel functionpqWill be larger, about by the optimization of solution (3) formula
Beam can make αp-αqValue it is smaller or close to 0, to reach expression factor alphapAnd αqThe purpose that is closer to of value.Letter
For it, as neighbour's object gas sample xpWith xqIt is more similar, it can to express factor alpha by the canonical constraint of (3) formulapWith
αqValue it is also more close, to reach the mesh for making the holding structural integrity of k neighbour's object gas around sample to be tested y
's.
In conjunction with (2) formula and (3) formula, the optimized-type for obtaining a kind of local expression model is as follows:
In formula (4), 0 λ≤1 < and 0 μ≤1 < are regular coefficient, and it is to expression factor alpha that R (α), which is canonical standardization norm,
Constraint, expression is
R (α)=| | α | |p (5)
In formula (5), | | | |pRepresent p norm.
If p=1, then by the α that (4) formula solves be it is sparse, if p=2, the α solved be it is smooth, both
Canonical constraint all will be so that α has more robustness.
Step 2 calculates expression coefficient vector α
When solving formula (4), if carrying out direct solution to manifold canonical bound term therein, it will give calculating band
It is next difficult, since the purpose of this is constrained expression coefficient vector α.Can by following solution procedure to formula (3) into
Row abbreviation:
In formula (10),
Here, MpqCorresponding value is arranged for pth row q in matrix M；D is a diagonal matrix, i.e., only has value on diagonal line,
Remaining position is 0, DppIt is the value of p-th of position of diagonal line in matrix D；σ2It is the variance of Gaussian function, is a constant.
For convenience of calculating, spy enables Laplacian Matrix L=D-M, brings the manifold regular terms derived into formula (4) and obtains:
λ is the coefficient of popular regular terms, since there are two types of regularization mode (L1 norm and L2 norm), L2 norms for R (α) tool
It is relatively easy to closed solution, therefore to its solution；And L1 norm is a non-convex problem, cannot pass through the side of immediate derivation
Method is solved, the method that this step uses ADMM, and the specific solution procedure of two kinds of regularization situations is as follows:
◆ as p=1, what R (α) was indicated is a norm, and formula (11) can be write
Formula (12) is a kind of sparse optimization problem, and the present invention solves it using ADMM method.Due to objective function
Only one variable, therefore a variable γ is quoted, so that form of the objective function with ADMM is as follows
S.t.-γ=0 α (13)
The Lagrange's equation that augmentation form can be constructed by formula (13), specifically shaped like following formula:
In formula (14), μ is regularization coefficient, and ρ is the coefficient of ADMM method, and z is Lagrange multiplier.
Since ADMM method needs successive iteration to find optimal solution, it is contemplated that there are three variables for formula (14), therefore when to one
When a variable is iterated, need to fix remaining two variables in the hope of optimal solution, substantially optimization process is as follows:
zm+1=zm+ρ(αm+1-γm+1) (15)
Update α:
Optimization formula when can must update α according to formula (14) is
Obtain the optimal solution of α, need to regard other with the incoherent item of α as known invariant, and to formula (16) into
Row derivation, i.e.,
Formula (17) is solved and can be obtained
-2Xk Ty+2Xk TXk+ 2 λ L α+z+ ρ (α-γ)=0 of α
Therefore, α in an iterative processm+1Expression formula is
αm+1=(2Xk TXk+2λL+ρI)-1(ργm+2Xk Ty-zm) (18)
In formula, I is unit matrix, i.e., is 1 on diagonal line, is elsewhere 0.
Update γ:
Firstly, since what γ took is a norm, need to be unfolded in solution procedure, it can according to original formula (14)
Formula, which must be optimized, is
The optimal solution of γ is obtained, the solution to the following formula can be converted into
Formula (20) is solved and can be obtained
Due to | γi| value and γiIt is positive and negative related, therefore carry out following discussion
Update z:
To sum up, when taking a norm, the specific solution procedure of formula (12) is as follows
Step 1: initialization α0, z0, and y is inputted, Xk, the value of λ, μ and ρ；
Step 2: α is updated according to formula (18)m+1；
Step 3: it is updated according to formula (22)Wherein i=1,2 ..., Nk；
Step 4: it is updated according to formula (23)Wherein i=1,2 ..., Nk；
Step 5: m=m+1；
Step 6: if not reaching the condition of convergence, repeating step 2, and three and step 4, otherwise execute step 7
Step 7: output αm+1；
◆ as p=2, what R (α) was indicated is two norms, and formula (11) can be write
This is a least squares problem, can be as follows in the hope of closed solution
α=(Xk TXk+λL+μI)-1Xk Ty (25)
In formula, I is a diagonal matrix.
Step 3 obtains Optimal error detection threshold value T*
Error-detection threshold T* is by a part of object gas training sample and a small amount of abnormal interference gas training sample
It is determined.DefinitionFor calculate Optimal error detection threshold value T* object gas training sample set,For
A small amount of abnormal interference gas training sample set of Optimal error detection threshold value T* is calculated, N is object gas number of training, and n is
Interference gas number of training (n < < N).αwiIndicate single target gas training sample wiThe expression coefficient vector of ∈ W, αhiIt indicates
Single interference gas training sample hiThe expression coefficient vector of ∈ H.Wherein α is the expression vector acquired for sample to be tested y；And
αwiIt is for object gas training sample wiThe expression vector acquired, αhiIt is for interference gas training sample hiThe expression acquired
Vector, their solution procedure be it is similar, then the expression error of object gas training sample are as follows:
Similarly, the expression error of interference gas training sample can be expressed as:
In formula (6)Indicate wiK neighbour's target sample, in formula (7)Indicate hiK neighbour's target sample
This.It readily appreciating that, the expression error of object gas training sample is smaller than the expression error of abnormal interference gas training sample, because
The search range of this error-detection threshold T should be between RESWMinimum value and RESHMaximum value between, and Optimal error examine
The accuracy of identification of object gas training sample W and interference gas training sample H should be comprehensively considered by surveying threshold value T*, i.e.,
Value range [the ESZ of error-detection threshold TW, ESZH]；Optimal error detection threshold value T* is determined according to the following steps:
Step (1), initialization T=ESZW, set increments of change delta；
Step (2), the value according to T, for the ESZ of formula (6)WThe ESZ of error amount collection and formula (7)ZError amount collection, obtains mesh
The accuracy in detection of the accuracy in detection Accuracy (W) of standard gas body training sample set W and abnormal interference gas training sample set H
Accuracy(H)；
Step (3): Accuracy=(Accuracy (W)+Accuracy (H))/2
Step (4): enabling T=T+delta, if T < ESZH, return step (2)；Otherwise, step (5) are executed；
Step (5): take the corresponding T value of maximum Accuracy as optimal threshold T*。
Step 4 acquires sample to be tested y and expression value XkThe residual error RES of α
After obtaining expression coefficient vector α by step 2, the expression error of sample to be tested y passes through y and expression value Xkα's is residual
Difference is indicated, i.e.,
Judge that sample to be tested y belongs to object gas or interference gas by judging the size of residual error RES, if RES
≤ T*, then sample to be tested y is object gas sample；If RES > T*, sample to be tested y is abnormal interference gas sample
This.
Embodiment:
Electric nasus system for the present embodiment is metal oxide semiconductor sensor array composition, and sensor has
TGS2602, TGS2620, TGS2201A and TGS2201B, for detecting six kinds of polluted gas common in life, including first
This six kinds of polluted gas are uniformly considered as object gas here by aldehyde, benzene, toluene, carbon monoxide, nitrogen dioxide, ammonia etc., are formed
Experiment sample collection be object gas sample set, in order to verify method of the invention, this experiment alcohol is as a kind of office of training
The interference gas sample of portion's expression model, then use additional alcohol as the interference sample of test model performance, experimental setup with
A kind of non-targeted interference gas recognition methods of electronic nose of band from expression disclosed in CN106124700A is consistent, and specific sample is retouched
It states as shown in table 1.
Table 1
Real-time testing data
(1) data set 1 is interfered
The data set is adopted in climatic chamber in the environment of electric nasus system to be placed in only non-targeted interference gas
Collection.The sampling number of each sensor is 2400.In the experiment collection process of sample, four-stage is divided to infuse in case respectively
Enter two kinds of non-targeted interference gas of perfume and floral water, the first two stage is perfume, latter two stage is floral water: being done by perfume
Sensor response signal region substantially 95~308 sampled points disturbed and 709~958 sampled points；The sensing interfered by floral water
Device response signal region substantially 1429~1765 sampled points and 2056~2265 sampled points；After injection interference acquisition every time is complete
Air pump can be used to carry out pumping cleaning to climatic chamber with environment in purifying box.
(2) data set 2 is interfered
In order to examine identification validity of the model when target and non-targeted gas exist simultaneously, in this experimental selection room
The formaldehyde object gas often occurred is as reference gas.Experimentation is divided into following three phases:
Stage 1: electric nasus system is placed in climatic chamber, is injected formaldehyde, is waited stable state to be achieved；Start to inject
Alcohol waits after stablizing, and stops pumping ten minutes later with pumping gas；
Stage 2: injection formaldehyde waits stable state to be achieved；Start to inject floral water interference smell, waits after stablizing, use
Pumping gas stops pumping ten minutes later；
Stage 3: injection formaldehyde waits stable state to be achieved；Start injection perfume and mix smell with orange, waits and stablizing
Afterwards, with pumping gas, after acquiring data, stop pumping.
The purpose of the experimental method is to study the injection under object gas environment and interfere smell and under interference environment
When injecting object gas, the application effect of AF panel model.According to above-mentioned experimentation, obtain the data set the length is
2400,3 response window regions of sensor PARA FORMALDEHYDE PRILLS(91,95) are 102~250 sampled points, 719~880 sampled points and 1380~1580
Sampled point；The window area that sensor is interfered by alcohol is 260~410 sampled points；Sensor interfered by floral water one
A window area is 881~1064 sampled points；Sensor is 1599 by a window area of the mixing interference of perfume and orange
~1899 sampled points.
Step according to the invention, the judgement to sample to be tested y:
1, it selects 1/5th of target sample in table 1 as dictionary Xs, then by partial target gas sample and interference
The sample of gas (for training) carries out optimal threshold T* (i.e. by the selected of step 3) of the invention；
2, the sample of sample to be tested y interference gas (for testing) in table 1, through step 1 of the invention, step 2 and
Step 3 respectively obtains α, T*, and presses step 4 of the invention, and RES is compared with T*, detects the attribute of y；
When L2 norm constraint, target-interference gas discrimination is with the change curve of threshold value T as shown in Figure 1, can from figure
To find out, with the increase of threshold value T value, object gas discrimination is gradually increased, and interference gas discrimination gradually decreases.This experiment
By calculating object recognition rate and disturbance ecology rate and maximum principle come selected threshold T*, T*=0.0887 is obtained
When L1 norm constraint, target-interference gas discrimination is with the change curve of threshold value T as shown in Fig. 2, can from figure
To find out, with the increase of threshold value T value, object gas discrimination is gradually increased, and interference gas discrimination gradually decreases.This experiment
By calculating object recognition rate and disturbance ecology rate and maximum principle come selected threshold T*, T*=0.1427 is obtained.
3, it for following each sample to be tested, repeats step 2 and is determined.
The comparison of recognition accuracy
Table 2 is the present invention and support vector machines (SVM), AF panel (PMIE) based on mode mispairing and
A kind of recognition accuracy comparison for the non-targeted interference gas recognition methods of electronic nose that band is expressed certainly disclosed in CN106124700A.
The accuracy rate of several recognition methods of table 2
The present invention is directed to target respectively and interference gas is tested, and since experiment condition is limited, target in this experiment
Sample size with interference be it is unbalanced, therefore the present embodiment used two kinds of metric forms more liberally verifying proposed
The validity of method.The specific representation of measurement one and measurement two is as follows
Measurement 1:
Measure (target sample recognition accuracy+interference specimen discerning accuracy rate)/2 2: recognition accuracy 2=
From expression formula as can be seen that accuracy rate is only expressed as identifying correct sample by measurement one, then divided by total sample
This, this is most commonly seen discrimination representation method, but is not in fact distinguish target and disturbance ecology degree.In reality
In there may be such situations, i.e. target sample detects very well, but sample is interfered almost not detected, and target sample
Number is far longer than interference sample, also can be quite high using the recognition accuracy of measurement one, but in fact such measurement effect
The detection case of sample will cannot be interfered to also illustrate that out simultaneously；Using measurement two can avoid well more than problem,
The recognition accuracy of target and interference sample is respectively calculated, finally integrates them, is thus balanced well
Relationship between target and interference gas.
As can be seen from Table 2: the test result of a kind of local expression model of the invention is better than support vector machines (SVM),
AF panel (PMIE) based on mode mispairing and from expression, a kind of local expression model of the invention, L1 norm is again
It is substantially better than L2 norm, this is because the sparsity of solution can be enhanced in L1 norm, is removed automatically different from sample to be tested mode
Sample is expressed, the expression sample that can preferably indicate under test gas is only left, so that sample to be tested obtains maximum
The expression of limit, accuracy of identification will also increase accordingly.
The constraint of expression model L1 norm and L2 norm will cause the performance difference of model, in actual use, select effect
The preferable norm of fruit, therefore the present invention has selected the expression model based on L1 norm.
A kind of band disclosed in the CN106124700A compared with the non-targeted interference gas recognition methods of electronic nose of expression, this
The recognition accuracy of a kind of local expression model L1 norm of invention is: the recognition accuracy of object gas sample by
The 91.25% of CN106124700A is increased to 100%, so, the present invention improves the accuracy of gas detection.
Sensor tests response curve
(1) when taking L2 norm, to interference data set 1 and interfere the identification situation of data set 2 as shown in Figure 3 and Figure 4；
(2) when taking L1 norm, to interference data set 1 and interfere the identification situation of data set 2 as shown in Figure 5 and Figure 6；
In Fig. 3~Fig. 6, for interference, the rectangular window that dotted line marks is we for the horizontal line part that sensor response is risen suddenly
Method invents the interference range identified, and it is just interference region that test response curve, which can see rectangular window,.In addition, test response
Curve can distinguish which partly belong to interference range, which partly belong to object gas area.
Claims (4)
1. the non-targeted interference gas recognition methods of electronic nose based on a kind of local expression model, characterized in that including following step
It is rapid:
Step 1, in fixed object gas sample XsIn find a kind of local expression model of sample to be tested y, sample to be tested y
A kind of local expression model optimization formula it is as follows:
In formula,It is expression coefficient vector, XkIt is in object gas sample XsIn with sample to be tested y
Maximally related k object gas sample；0 λ≤1 < and 0 μ≤1 < are regular coefficient, and R (α) is canonical standardization norm；αp,αq
∈ α, αpAnd αqIt is any two value expressed in coefficient vector α, shows respectively XkIn p-th of sample and q-th of sample table
Up to coefficient, wpqFor p-th of neighbour's object gas sample xpWith q-th of neighbour's object gas sample xqSimilarity degree；
Step 2 calculates expression coefficient vector α
Express the solution formula of coefficient vector α are as follows:
In formula, L=D-M
R (α) has two kinds of regularization modes of L1 norm and L2 norm, MpqCorresponding value is arranged for pth row q in matrix M；D is one
Diagonal matrix, DppIt is the value of p-th of position of diagonal line in matrix D；σ2It is the variance of Gaussian function, is a constant, this step
It is solved using the method for ADMM；
Step 3 obtains Optimal error detection threshold value T*
The expression error of object gas training sample are as follows:
The expression error of interference gas training sample can be expressed as:
αwiIndicate single target gas training sample wiThe expression coefficient vector of ∈ W, αhiIndicate single interference gas training sample
hiThe expression coefficient vector of ∈ H,Indicate wiK neighbour's target sample,Indicate hiK neighbour's target sample, N is
Object gas number of training, n are interference gas number of training；
Optimal error detection threshold value T* are as follows:
Step 4 acquires sample to be tested y and expression value XkThe residual error RES of α
After obtaining expression coefficient vector α by step 2, the expression error of sample to be tested y passes through y and expression value XkThe residual error of α come into
Row expression, i.e.,
Judge that sample to be tested y belongs to object gas or interference gas by judging the size of residual error RES, if RES≤T*,
So sample to be tested y is object gas；If RES > T*, sample to be tested y are abnormal interference gas.
2. the non-targeted interference gas recognition methods of the electronic nose according to claim 1 based on a kind of local expression model,
It is characterized in that in step 2, R (α) takes the specific solution procedure of L1 norm as follows:
Step 1: initialization α0, z0, and y is inputted, Xk, the value of λ, μ and ρ；μ is regularization coefficient, and ρ is the coefficient of ADMM method, λ
For popular regular coefficient,
Step 2: α is updated according to the following formulam+1；
αm+1=(2Xk TXk+2λL+ρI)-1(ργm+2Xk Ty-zm)
Wherein I is unit matrix；
Step 3: it updates according to the following formulaWherein i=1,2 ..., Nk；
Step 4: it updates according to the following formula
Step 5: m=m+1；
Step 6: if not reaching the condition of convergence, repeating step 2, and three and step 4, otherwise execute step 7；
Step 7: output αm+1。
3. the non-targeted interference gas recognition methods of the electronic nose according to claim 1 based on a kind of local expression model,
It is characterized in that in step 2, R (α) takes the solution of L2 norm are as follows:
α=(Xk TXk+λL+μI)-1Xk TY,
Wherein I is unit matrix.
4. the non-targeted interference gas identification of the electronic nose according to claim 2 or 3 based on a kind of local expression model
Method, characterized in that in step 3, the solution procedure of Optimal error detection threshold value T* is:
Step (1), initialization error detection threshold value T=ESZW, set increments of change delta；
Step (2), the value according to T, for the ESZ for calculating acquisitionWError amount collection and ESZHError amount collection obtains object gas instruction
Practice the accuracy in detection Accuracy of the accuracy in detection Accuracy (W) and abnormal interference gas training sample set H of sample set W
(H)；
Step (3): Accuracy=(Accuracy (W)+Accuracy (H))/2
Step (4): enabling T=T+delta, if T < ESZH, return step (2)；Otherwise, step (5) are executed；
Step (5): take the corresponding T value of maximum Accuracy as optimal threshold T*。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710009931.1A CN106770967B (en) | 2017-01-06 | 2017-01-06 | The non-targeted interference gas recognition methods of electronic nose based on a kind of local expression model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710009931.1A CN106770967B (en) | 2017-01-06 | 2017-01-06 | The non-targeted interference gas recognition methods of electronic nose based on a kind of local expression model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106770967A CN106770967A (en) | 2017-05-31 |
| CN106770967B true CN106770967B (en) | 2019-02-12 |
Family
ID=58950790
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710009931.1A Active CN106770967B (en) | 2017-01-06 | 2017-01-06 | The non-targeted interference gas recognition methods of electronic nose based on a kind of local expression model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106770967B (en) |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7894072B1 (en) * | 2008-11-10 | 2011-02-22 | The United States Of America As Represented By The Secretary Of The Navy | Laser-based gas differential spectral analysis |
| CN102866179B (en) * | 2012-09-13 | 2014-06-18 | 重庆大学 | Online recognition and inhibition method based on non-target interference smell in electronic nose of artificial intelligent learning machine |
| CN103020940B (en) * | 2012-12-26 | 2015-07-15 | 武汉大学 | Local feature transformation based face super-resolution reconstruction method |
| CN103413144A (en) * | 2013-07-29 | 2013-11-27 | 西北工业大学 | Airport detection and recognition method based on local global feature joint decision |
| CN105678343B (en) * | 2015-08-25 | 2019-03-15 | 浙江工业大学 | Hydropower Unit noise abnormality diagnostic method based on adaptive weighted group of sparse expression |
| CN105678338B (en) * | 2016-01-13 | 2020-04-14 | 华南农业大学 | Target tracking method based on local feature learning |
| CN106124700B (en) * | 2016-06-20 | 2018-01-16 | 重庆大学 | A kind of electronic nose non-targeted interference Gas Distinguishing Method of band from expression |
-
2017
- 2017-01-06 CN CN201710009931.1A patent/CN106770967B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN106770967A (en) | 2017-05-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zio | A study of the bootstrap method for estimating the accuracy of artificial neural networks in predicting nuclear transient processes | |
| CN108664690A (en) | Long-life electron device reliability lifetime estimation method under more stress based on depth belief network | |
| CN105572572B (en) | Analog-circuit fault diagnosis method based on WKNN-LSSVM | |
| CN105447857A (en) | Feature extraction method of pulsed eddy-current infrared thermal image | |
| CN102866179A (en) | Online recognition and inhibition method based on non-target interference smell in electronic nose of artificial intelligent learning machine | |
| CN105044722B (en) | The full Bayesian Discriminating Features extracting method of synthetic aperture radar target | |
| Altomare et al. | Objective Bayesian search of Gaussian directed acyclic graphical models for ordered variables with non‐local priors | |
| CN110501122A (en) | A kind of adaptive leakage detection method of pressure vessel based on cluster | |
| CN113591215B (en) | Abnormal satellite component layout detection method based on uncertainty | |
| Dalla Libera et al. | A novel multiplicative polynomial kernel for volterra series identification | |
| Zhu et al. | Adaptive fault diagnosis of HVCBs based on P-SVDD and P-KFCM | |
| CN110889207B (en) | Deep learning-based intelligent assessment method for credibility of system combination model | |
| CN105203327B (en) | A kind of gas circuit measurement parameter selection method applied to engine air passage analysis | |
| Mills et al. | Aio-p: Expanding neural performance predictors beyond image classification | |
| CN106770967B (en) | The non-targeted interference gas recognition methods of electronic nose based on a kind of local expression model | |
| CN109657733B (en) | Variety discriminating method and system based on constituent structure feature | |
| CN108507607A (en) | A kind of method for detecting weak signals based on kernel function | |
| CN110531362A (en) | A kind of object detection method of high-resolution moving sonar Knowledge-based | |
| Ristic et al. | Achievable accuracy in parameter estimation of a Gaussian plume dispersion model | |
| Spaaks et al. | Resolving structural errors in a spatially distributed hydrologic model using ensemble Kalman filter state updates | |
| Kurz et al. | Investigating model-data inconsistency in data-informed turbulence closure terms | |
| CN110489602A (en) | Knowledge point ability value predictor method, system, equipment and medium | |
| Wang et al. | FCM algorithm and index CS for the signal sorting of radiant points | |
| Spagnolo et al. | Forensic metrology: uncertainty of measurements in forensic analysis | |
| CN106772306B (en) | A kind of detection method and server of object |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |