CN102507677B - Drift rejection method of electronic nose based on multiple self-organizing neural networks - Google Patents

Abstract

The invention provides a drift rejection method of an electronic nose based on multiple self-organizing neural networks. The method uses a sample cache matrix to cache the neuron mean center of each self-organizing neural network and uses cyclic iterative for update to get the compensation incremental array of cyclic drift rejection compensation each time. Each neuron in the self-organizing neural network matched with gas adopted by the drift compensation training conducts drift compensation, meanwhile the compensation incremental array is utilized to conduct drift rejection compensation for gas-sensitive characteristic electrical signal values adopted by the drift compensation training in the neurons of other self-organizing neural networks, so that the neurons of each self-organizing neural network after drift compensation tend to approach actual electrical signal output array values of the matched gas detected by a gas sensor array, and therefore the purpose of rejecting drift is achieved. The drift compensation execution efficiency of the electronic nose is increased, and the electronic nose after drift rejection compensation still keeps good recognition performance.

Description

A kind of Electronic Nose drift rejection method based on multiple self-organizing neural networks

Technical field

The invention belongs to detection by electronic nose, training technique field, relate in particular to a kind of Electronic Nose drift rejection method based on multiple self-organizing neural networks.

Background technology

Electronic Nose is to utilize the response collection of illustrative plates of gas sensor array to identify electronic installation or the system of smell, mainly gas sensor array, Signal Pretreatment unit and pattern recognition unit three parts, consists of.The principle of work of Electronic Nose is: in face of certain smell is presented on the gas sensor that a kind of active material makes, gas sensor can convert the input of the chemistry of this gas to electric signal output, adopt a plurality of gas sensors to form gas sensor array, the plurality of gas sensor has just formed the electric signal output array of gas sensor array to this smell to a kind of response of smell; For realizing the qualitative or quantitative test to smell, the output of the electric signal of gas sensor must be carried out to suitable pre-service and (eliminate noise, feature extraction, signal amplification, normalization etc.), after, by pattern recognition unit, adopt suitable pattern recognition analysis method to carry out identifying processing to it; In theory, every kind of smell is for gas sensor array and Yan Douhui has its characteristic of correspondence array of electrical signals value, neuron using different odor characteristic of correspondence array of electrical signals value as Electronic Nose is stored in pattern recognition unit, carry out when gas detects the electric signal output array value of gas sensor array and neuron are contrasted to coupling identification, just can distinguish different gas, also can utilize gas sensor array to be measured the cross-sensitivity of multiple gases simultaneously, by suitable analytical approach, realize analysing mixed.Because detection by electronic nose has short, low cost and other advantages of time, in fields such as food inspection, medical diagnosis on disease, environment measurings, broad research and concern have been obtained at present.

The neuron of detection by electronic nose identification gas can be trained and set up acquisition by benchmark, utilize the multiple gases sample that Electronic Nose can be identified it to carry out the priori detection, obtain the feature array of electrical signals value of gas sensor array to this multiple gases, it is stored as the neuron of this multiple gases coupling, in order to the identification benchmark as this multiple gases.Yet, the electric signal output array value of Electronic Nose gas sensor array sensing same gas is not unalterable, and between the electric signal output array value of gas sensor array detection gas and the feature array of electrical signals value (being neuron) of this gas coupling, having occurred to drift about often affects a key factor of Electronic Nose recognition effect just.The reason that the detected value of gas sensor array produces drift mainly contains two: the one, due to the variation of Electronic Nose working environment, such as temperature, humidity etc., the electric signal output array value that makes gas sensor array the detect drift of fluctuating in the neuron value environs be complementary; Another reason is to change because the phenomenons such as gas sensor is aging make self physicochemical property, and then affect the size of its electric signal output valve, cause the electric signal output array value of gas sensor array and the neuron generation deviation be complementary, form drift.Wherein, the drift that the former causes belongs to a kind of transient state drift, does not affect in itself the accuracy of identification of Electronic Nose.And the drift caused by the latter generally is referred to as long term drift, it will follow use long-term existence the accumulation of gas sensor, if do not take braking measure will cause the accuracy of detection of Electronic Nose obviously to reduce, therefore the impact of inhibition or reduction gas sensor long term drift seems particularly important for the accuracy of detection that guarantees Electronic Nose and effect.

From the current study, the long term drift that mainly by neuronic value in the correction Electronic Nose, suppresses Electronic Nose.The method of existing inhibition Electronic Nose drift mainly contains two large classes: a kind of is that the electric signal output array value drift of Electronic Nose gas sensor array is considered as to a road independent signal, by principal component analysis (PCA), independent component analysis, the mathematical methods such as Orthogonal Decomposition are rejected it (referring to file " Bouwmans T, BafF E, Vachon B.Background modeling using mixture of Gaussians for foreground detection-a survey.Recent Patents on Computer Science, 2008.1 (3): 219-237 " and document " Piccardi M.Background subtraction techniques:a review.In:Proceeding of the IEEE International Conference on Systems, Man and Cybernetics.The Hague, Netherlands:IEEE 2004.3099-3104 " etc.) from the electric signal of sensor output, on this theoretical method, effect is better, but complete drift prior imformation need to be arranged as the foundation of rejecting, yet the regularity of its electric signal output array value drift of gas with various sensor is difficult to accurately summed up and grasp, therefore will set up complete drift prior imformation has quite high technical difficulty.Another kind method is without complete drift prior imformation, main individual layer self organizing neural network (the Self Organizing Maps that adopts, being called for short the SOM network) mode of carrying out the drift compensation training compensates the electric signal output of the gas sensor array that contains drift, be about to all neurons as a self organizing neural network, each neuron is the feature array of electrical signals value of a kind of gas with respect to gas sensor array, and be provided with a plurality of neurons for each gas, the plurality of neuron is got different value and then is formed between the neuron cog region to the gas that is complementary in certain interval, and then the gas that utilizes neuron to mate carries out the drift compensation training to Electronic Nose, at the drift compensation training period, if the central value generation deviation between the neuron cog region of the electric signal output array value that Electronic Nose gas sensor array sensing obtains and this gas coupling, show that drift has occurred the electric signal output array value that gas sensor array detects this gas, according to the size of drift value, each neuron between the neuron cog region of gas coupling is carried out to drift compensation, thereby reach the purpose that suppresses the detected value drift (referring to document " Kohonen T.The Self-organizing Maps[J] Proceedings of the IEEE, 1990, 78 (9): 1464-1480 "), but because the individual layer self organizing neural network is equivalent to all neurons of storage are all enumerated in a neural network plane, overlapping through occurring likely and between the neuron cog region of other gas between drift-compensated neuron cog region, the information loss of the partial nerve unit be overlapped in causing between the neuron cog region of other gas, this not only can the drift compensation effect, seriously even can cause neuron information confusion in whole self organizing neural network if overlapping between between the different neuron cog regions that cause because of drift compensation, have a strong impact on the recognition performance of Electronic Nose.

There is in recent years the scholar to propose to use multiple self-organizing neural networks (Multiple Self Organizing Maps, be called for short MSOM) the solution drifting problem, be about to enumerate at one independently on self organizing neural network for a kind of a plurality of neurons of gas, detect and construct multiple self-organizing neural networks for multiple gases, therefore between the neuron cog region of the corresponding a kind of gas of each self organizing neural network, thereby carry out drift compensation on the drift compensation training self organizing neural network that only meeting be complementary at gas for a kind of gas, the mode that its drift compensation to single self organizing neural network is processed is with identical (referring to document " Marzia Zuppa to the drift compensation processing mode between single neuron cog region in SOM network drift compensation method, Cosimo Distante, Pietro Siciliano, et al.Drift counteraction with multiple self-organising maps for an electronic nose[J] Sensors and Actuators B, 2004, 98:305-317 "), drift compensation just can not affect the neuron that is matched with other gas of other self organizing neural network like this, avoided occurring that neuron is because disturbing mutually the situation that information loss occurs, but this makes again the neuron on some self organizing neural networks that do not carry out for a long time the drift compensation training almost can not get any drift compensation at the drift compensation training period, cause its drift situation to continue to worsen, when again carrying out the gas detection, be not able to gas that drift-compensated self organizing neural network mates and detect recognition accuracy and just can significantly reduce, affected the recognition performance of Electronic Nose.Unless all kinds gas that adopts Electronic Nose to identify carries out the drift compensation training to Electronic Nose respectively, the self organizing neural network that guarantees every kind of gas coupling all can be able to drift compensation, just can overcome the generation of above-mentioned situation, but keep for a long time the drift compensation training of all kinds gas to be difficult to accomplish, and drift compensation training amount is large, troublesome poeration, causes drift inhibition and Electronic Nose recognition performance all to be difficult to be guaranteed.

Summary of the invention

For the above-mentioned problems in the prior art, the present invention proposes a kind of Electronic Nose drift rejection method based on multiple self-organizing neural networks in order to strengthen Electronic Nose for the whole drift compensation ability of long term drift, to improve the drift compensation execution efficiency of Electronic Nose, guarantee that Electronic Nose still keeps good recognition performance after drift suppresses compensation.

For achieving the above object, the present invention has adopted following technological means:

A kind of Electronic Nose drift rejection method based on multiple self-organizing neural networks, comprise the steps:

A) initialization step, it is specially:

A1) set up sample buffer memory matrix ${\tilde{X}}_{\mathrm{Rct}}\left(t\right)=\{X^1\left(t\right),X^2\left(t\right),...,X^K\left(t\right)\};$ Wherein, X ^k(t) array cache for comprising i element, k ∈ 1,2 ..., K}, K means the quantity of the self organizing neural network of Electronic Nose, and i means the number of gas sensor in the Electronic Nose gas sensor array, and t means constantly;

A2) initialization moment t=0;

A3) at current time, sample buffer memory matrix

In the value of each array cache be:

$X^k\left(t\right)=\frac{1}{M_k}\underset{m=1}{\overset{M_k}{\mathrm{\Σ}}}{W}_m^k\left(t\right),k=\mathrm{1,2},...,K;$

Wherein, $W_m^k\left(t\right)=[w_{m,1}^k\left(t\right),w_{m,2}^k\left(t\right),...,w_{m,i}^k\left(t\right)]$ Be illustrated in m neuron in k self organizing neural network of current time Electronic Nose,

Mean respectively described neuron In i feature value of electrical signals, M _kNeuronic quantity in k self organizing neural network of expression Electronic Nose;

B) drift suppresses compensation process; Be specially:

B1) t adds 1 certainly constantly;

B2) obtain the electric signal output array X of current time Electronic Nose gas sensor array _Ts(t):

X _ts(t)＝[x _ts，1(t)，x _ts，2(t)，…，x _ts，i(t)]；

Wherein, x _{Ts, 1}(t), x _{Ts, 2}(t) ..., x _{Ts, i}(t) be illustrated respectively in the electric signal output valve of i gas sensor in current time Electronic Nose gas sensor array;

B3) ask for coupling triumph self organizing neural network sequence number k _1stWith coupling time triumph self organizing neural network sequence number k _2nd:

Wherein,

Be illustrated in m neuron in k self organizing neural network of this previous moment (t-1) Electronic Nose; Symbol

Mean to get normalized value,

With

Mean to get respectively described electric signal output array X _Ts(t) normalized value and get described neuron

Normalized value;

Expression is got

With

Euclidean distance;

Expression is got

With

Euclidean distance all k ∈ 1,2 ..., K} and m ∈ 1,2 ..., M _kMinimum value in situation, Expression is got

With

Euclidean distance all k ∈ 1,2 ..., K} and m ∈ 1,2 ..., M _kOnly be greater than in situation

Inferior minimum value;

B4) if k _1st=k _2nd, according to the following formula each neuron in each self organizing neural network of Electronic Nose is carried out to drift compensation:

If k _1st≠ k _2nd, according to the following formula each neuron in each self organizing neural network of Electronic Nose is carried out to drift compensation:

Wherein,

Be illustrated in m neuron in k self organizing neural network of current time Electronic Nose; Δ W (t) means the compensation incremental array of current time, and

X ^K1st(t-1) be illustrated in this previous moment (t-1) sample buffer memory matrix In k _1stIndividual array cache,

Mean to get described array cache X ^K1st(t-1) normalized value; A is the compensating proportion coefficient, and span is 0<a≤0.5; B is the compensation increment coefficient, and span is 0<b≤0.5;

B5) according to the following formula to sample buffer memory matrix In the value of each array cache carry out the iteration renewal:

Wherein, X ^k(t) mean current time sample buffer memory matrix In k array cache; X ^k(t-1) be illustrated in this previous moment (t-1) sample buffer memory matrix

In k array cache,

Mean to get described array cache X ^k(t-1) normalized value;

C) circulation execution step B), until Electronic Nose stops drift inhibition work.

Further, described symbol

The concrete operation formula that means to get normalized value is as follows:

Wherein, F means any array that comprises i element, f ₁, f ₂..., f _iMean respectively i the element that described array F comprises.

Than prior art, the present invention has following beneficial effect:

1, in Electronic Nose drift rejection method of the present invention, adopting any one gas to carry out drift compensation training can make in the neuron of each self organizing neural network the feature value of electrical signals of this gas sensitization all drift about and suppressed to compensate, therefore as long as adopt, can make in gas sensor array all gas sensor all be able to the responsive several gases a kind of or few in number that detect to carry out the drift compensation training, can guarantee that in Electronic Nose, each neuron of each self organizing neural network all is able to drift compensation, strengthened the whole drift compensation ability of Electronic Nose for long term drift, improved the drift compensation execution efficiency of Electronic Nose.

2, in Electronic Nose drift rejection method of the present invention, by sample buffer memory matrix, to the neuron average center of each self organizing neural network, in addition buffer memory loop iteration upgrade, the compensation incremental array that suppresses compensation in order to ask for each cyclic process, in the self organizing neural network that the drift compensation training institute gas that adopts is mated, each neuron carries out the drift-compensated while, also utilize the compensation incremental array to carry out drift compensation to the neuron in other self organizing neural network, the loop iteration of sample buffer memory matrix upgrades and has guaranteed that the compensation incremental array always obtains suitable value, make the neuron of each self organizing neural network all be tending towards approaching the actual electrical signal output array value that gas sensor array detects its coupling gas after drift compensation, thereby reach the purpose that suppresses drift, guaranteed that Electronic Nose still keeps good recognition performance after drift suppresses compensation.

The accompanying drawing explanation

Fig. 1 is the FB(flow block) that the present invention is based on the Electronic Nose drift rejection method of multiple self-organizing neural networks;

The structural representation that Fig. 2 is the Electronic Nose test platform built voluntarily.

Embodiment

One, the limitation of existing MSOM drift compensation method.

According to the principle of work of MSOM drift compensation method, after supposing that Electronic Nose is trained by benchmark, store K self organizing neural network, be respectively used to the different gas of coupling identification K class; If adopt k self organizing neural network (k ∈ 1,2 ..., the continuous Δ t of gas sample that K}) mates constantly inputs to Electronic Nose and carries out the drift compensation training; T of this Δ constantly in, each neuron in k self organizing neural network can effectively obtain drift compensation, still the neuron in other self organizing neural network except k self organizing neural network almost can not obtain drift-compensated chance.If constantly the new gas sample of a non-k self organizing neural network coupling of Electronic Nose input is detected at (Δ t+1), due to through before this Δ t constantly after, drift has likely occurred to the electric signal output array value of this new gas sample in the Electronic Nose gas sensor array, and the self organizing neural network of this new gas sample coupling does not obtain any drift compensation, therefore can't guarantee Electronic Nose at (Δ t+1) constantly to the identification accuracy of this new gas sample.

Two, solution of the present invention.

For above-mentioned present situation and the deficiency of prior art, the present invention is directed to existing multiple self-organizing neural networks drift compensation method and be further improved, a kind of Electronic Nose drift rejection method based on multiple self-organizing neural networks is proposed.Electronic Nose drift rejection method of the present invention mainly comprises initialization and drift suppresses compensation two large steps, and carry out drift inhibition compensation process by circulation the compensation incremental array is carried out to the loop iteration renewal, to realize that each neuron in each self organizing neural network of Electronic Nose is carried out to the loop iteration drift compensation, make the neuron of each self organizing neural network level off to all the time gas sensor array and detect the actual electrical signal output array value of its coupling gas, thereby reach the purpose that suppresses drift.

The present invention is based on multiple self-organizing neural networks the Electronic Nose drift rejection method process flow diagram as shown in Figure 1, its concrete steps are as follows:

A) initialization step.

This step is mainly in order to set up sample buffer memory matrix, for the neuron average center to each self organizing neural network in addition buffer memory using the data source of upgrading as loop iteration, initialization sample buffer memory matrix and the value of t simultaneously constantly.It should be noted that, constantly t being not used in means concrete time value (as t second, t minute etc.), moment t in the inventive method is as distinguishing identifier use constantly, in order to indicate the different time points in the loop iteration process, realizes the differentiation of front and back data in the loop iteration process.

The particular content of initialization step is as following step a1)～a3) as described in:

A1) set up sample buffer memory matrix ${\tilde{X}}_{\mathrm{Rct}}\left(t\right)=\{X^1\left(t\right),X^2\left(t\right),...,X^K\left(t\right)\};$ Wherein, X ^k(t) array cache for comprising i element, k ∈ 1,2 ..., K}, K means the quantity of the self organizing neural network of Electronic Nose, and i means the number of gas sensor in the gas sensor array of Electronic Nose, and t means constantly.

A2) initialization moment t=0; T=0 means that current time is the original state moment constantly, to be different from follow-up loop iteration constantly.

A3) at current time, sample buffer memory matrix In the value of each array cache be:

$X^k\left(t\right)=\frac{1}{M_k}\underset{m=1}{\overset{M_k}{\mathrm{\&Sigma;}}}{W}_m^k\left(t\right),k=\mathrm{1,2},...,K;$ Formula (1);

In formula (1), $W_m^k\left(t\right)=[w_{m,1}^k\left(t\right),w_{m,2}^k\left(t\right),...,w_{m,i}^k\left(t\right)]$ Be illustrated in m neuron in k self organizing neural network of current time Electronic Nose,

Mean respectively described neuron

In i feature value of electrical signals, namely neuron

The gas be complementary corresponds respectively to i feature value of electrical signals of i gas sensor in the Electronic Nose gas sensor array, M _kNeuronic quantity in k self organizing neural network of expression Electronic Nose.Neuronic quantity M in each self organizing neural network of Electronic Nose _kNumber be as required and self-defined setting, the M that different self organizing neural networks are corresponding _kValue can be different, therefore can think M _kThe function of k, in any k self organizing neural network that its funtcional relationship is set by definition neuronic quantity and determine.

Current time is t=0, so sample buffer memory matrix now

In each array cache X ^k(t) value is original state value constantly; Array cache X ^k(t) original state value constantly is

The i.e. neuron average center of k self organizing neural network.Without relying on again the neuron average center ask for each self organizing neural network of Electronic Nose to determine the value of sample buffer memory matrix, but upgrade the neuron average center of determining that each self organizing neural network of Electronic Nose is new by loop iteration in follow-up loop iteration process.

B) drift suppresses compensation process.

Drift suppresses compensation process, to utilize Electronic Nose at the sample buffer memory matrix of the neuron of this each self organizing neural network of previous moment and this previous moment, the neuron of each self organizing neural network of current time Electronic Nose to be drifted about and suppresses compensation, and the sample buffer memory matrix to current time carries out the iteration renewal, to treat that an after this drift of loop iteration constantly suppresses compensation process and calls.

Drift suppresses the particular content of compensation process as following step b1)～b5) as described in:

B1) t adds 1 certainly constantly, for the sign as current time, to distinguish this previous moment.

B2) obtain the electric signal output array X of current time Electronic Nose gas sensor array _Ts(t):

X _Ts(t)=[x _{Ts, 1}(t), x _{Ts, 2}(t) ..., x _{Ts, i}(t)]; Formula (2);

In formula (2), x _{Ts, 1}(t), x _{Ts, 2}(t) ..., x _{Ts, i}(t) be illustrated respectively in the electric signal output valve of i gas sensor in current time Electronic Nose gas sensor array.

The electric signal output array X of Electronic Nose gas sensor array now _Ts(t), be equivalent to detect the gases used and actual electrical signal output array that obtains of drift compensation training at the gas sensor array of current time Electronic Nose.The training of this drift compensation can be by the drift compensation training experiment, to be undertaken separately, can be also utilizing Electronic Nose to carry out in carrying out the process of gas detection simultaneously; The latter is equivalent to the gas as detected object is used as drift compensation training gas sample simultaneously, drift about when Electronic Nose is carried out the gas detection and suppress operation, can carry out respectively gas by the multithreading task and detect operation and drift inhibition operating process in practical programs, not interfere with each other.This electric signal output array X _Ts(t) will drift about and suppress the reference data of compensation as current time, the neuron in the self organizing neural network that makes the drift compensation training institute gas that adopts mate by subsequent step levels off to this electric signal output array X after drift compensation _Ts(t).

B3) ask for coupling triumph self organizing neural network sequence number k _1stWith coupling time triumph self organizing neural network sequence number k _2nd:

Formula (3);

In formula (3),

Be illustrated in m neuron in k self organizing neural network of this previous moment (t-1) Electronic Nose; Symbol Mean to get normalized value,

With

Mean to get respectively described electric signal output array X _Ts(t) normalized value and get described neuron

Normalized value;

Expression is got

With

Euclidean distance;

Expression is got

With

Euclidean distance all k ∈ 1,2 ..., K} and m ∈ 1,2 ..., M _kMinimum value in situation,

Expression is got

With

Euclidean distance all k ∈ 1,2 ..., K} and m ∈ 1,2 ..., M _kOnly be greater than in situation

Inferior minimum value.

The coupling triumph self organizing neural network sequence number k so tried to achieve _1stWith coupling time triumph self organizing neural network sequence number k _2nd, be equivalent to: among each neuron of each self organizing neural network of Electronic Nose, will be after normalization the electric signal output array X of its Euclidean distance and current time Electronic Nose gas sensor array _Ts(t) self organizing neural network at nearest neuron place is regarded as coupling triumph self organizing neural network, using its self organizing neural network sequence number as k _1stWill be after normalization the electric signal output array X of its Euclidean distance and current time Electronic Nose gas sensor array _Ts(t) self organizing neural network at inferior nearest neuron place is regarded as coupling time triumph self organizing neural network, using its self organizing neural network sequence number as k _2nd.Ask for coupling triumph self organizing neural network sequence number k _1stWith coupling time triumph self organizing neural network sequence number k _2ndPurpose mainly contain two aspects: be on the one hand by coupling triumph self organizing neural network sequence number k _1stRegard as the self organizing neural network sequence number that the current time drift compensation training gas that adopts is complementary so that for drift compensation train the self organizing neural network that mates of the gas that adopts and unmatched self organizing neural network distinguished and carried out respectively different drift compensations; Be on the other hand by comparing k _1stAnd k _2ndValue know the big or small degree of Electronic Nose in current time drift, if k _1st=k _2ndShow that drift value that Electronic Nose detects the drift compensation training gas that adopts at current time is still in the identification range in self organizing neural network, if k _1st≠ k _2ndShow that the drift that Electronic Nose detects the drift compensation training gas that adopts at current time has reached the identification range edge of self organizing neural network, thereby the different big or small degree of drift for Electronic Nose of being convenient in subsequent step is taked different drift compensation strategies.

B4) if k _1st=k _2nd, according to formula (4), each neuron in each self organizing neural network of Electronic Nose is carried out to drift compensation:

Formula (4);

If k _1st≠ k _2nd, according to formula (5), each neuron in each self organizing neural network of Electronic Nose is carried out to drift compensation:

Formula (5);

In formula (4) and formula (5), Be illustrated in m neuron in k self organizing neural network of current time Electronic Nose; Δ W (t) means the compensation incremental array of current time, and

X ^K1st(t-1) be illustrated in this previous moment (t-1) sample buffer memory matrix

In k _1stThe value of individual array cache, i.e. sample buffer memory matrix

In k (get k=k _1st) value of array cache;

Mean to get described array cache X ^K1st(t-1) normalized value; A is the compensating proportion coefficient, and span is 0<a≤0.5; B is the compensation increment coefficient, and span is 0<b≤0.5.

The concrete value of compensating proportion coefficient a and compensation increment coefficient b, the sensitivity that the active material that need to adopt according to the gas sensor of Electronic Nose in actual conditions detects drift to gas is determined, the gas sensor of making for different active materials, the concrete value of its corresponding compensating proportion coefficient a and compensation increment coefficient b is not quite similar; In the methods of the invention, close numerical value interval (0 after opening before the span of compensating proportion coefficient a and compensation increment coefficient b all is confined to, 0.5] within, a ≠ b (certainly not getting rid of the situation that has compensating proportion coefficient a that indivedual active materials are corresponding and compensation increment coefficient b to have a=b) generally; The drift that active material detects gas is more responsive, and the value of its corresponding compensating proportion coefficient a and compensation increment coefficient b is larger.

In above-mentioned drift compensation process, can see, for K self organizing neural network of Electronic Nose:

1. at current time, with the gases used k be complementary of current time drift compensation training _1stEach neuron of individual self organizing neural network

(get k=k _1stIf, k _1st=k _2ndK _1stAnd k _2ndMean same self organizing neural network), the actual electrical signal output array X that directly utilizes the current time gas sensor array to obtain the gas that the adopts detection of drift compensation training institute _Ts(t) with the value of this neuron in this previous moment

(get k=k _1st) difference carry out the drift compensation of a times of ratio, make the neuron of current time

(get k=k _1st) level off to X _Ts(t).

2. for k _1st≠ k _2ndSituation, k _2ndIndividual self organizing neural network is exactly to train a self organizing neural network of the sensitive features correlativity maximum of the gas that adopts with the current time drift compensation, that is to say, the gas sensor array of Electronic Nose is to the sensitive features of the drift compensation training gas that adopts and to k _2ndThere is quite high correlativity in the sensitive features of the gas that individual self organizing neural network mates, if Electronic Nose trains the detection of the gas that adopts that drift has occurred to the current time drift compensation, means that Electronic Nose is to k _2ndThe gas of individual self organizing neural network coupling detects the drift situation that also correspondingly exists; Simultaneously, k _1st≠ k _2ndAlso just mean that Electronic Nose detects drift compensation at current time and trains the drift degree of the gas that adopts larger, if therefore Electronic Nose detects k at current time _2ndThe gas of individual self organizing neural network coupling also certainly exists larger drift value; Comprehensive this two aspects factor consideration, therefore need to be to current time k _2ndEach neuron of individual self organizing neural network carries out the drift compensation that amplitude is larger, so that its more drift is suppressed.Visible by the described drift compensation process of above-mentioned formula (5), at current time to k _2ndEach neuron of individual self organizing neural network

(get k=k _2nd) drift compensation be:

Get k=k _2nd

Formula (6);

The size that compensates incremental array Δ W (t) due to current time is the electric signal output array X by current time Electronic Nose gas sensor array _Ts(t) normalized value and coupling triumph self organizing neural network sequence number k _1stCorresponding this previous moment (t-1) sample buffer memory matrix

In k _1stThe difference of the normalized value of individual array cache determines, embodied the current time gas sensor array to detecting the drift degree of the drift compensation training gas that adopts, so in formula (6), adopted the compensation incremental array Δ W (t) of current time to carry out the drift compensation of b times of ratio; Simultaneously, consider again k _2ndThe neuron of individual self organizing neural network

(get k=k _2nd) train the sensitive features correlativity of the gas that adopts with the current time drift compensation, so adopted the electric signal output array X of current time Electronic Nose gas sensor array in formula (6) _Ts(t) to neuron (get k=k _2nd) drift compensation carried out the correction of a times of ratio.

3. for each neuron of the self organizing neural network outside above-mentioned two situations

(get k=1,2 ..., K and k ≠ k _1stOr k _2nd), at current time, in order to make neuron

(get k=1,2 ..., K and k ≠ k _1stOr k _2nd) in more responsive to the drift compensation training institute gas that adopts and exist the feature value of electrical signals that detects drift all to be able to more appropriate drift to suppress to compensate, so utilize the compensation incremental array Δ W (t) of current time to carry out the drift compensation of b times of ratio; So the principle of compensation is, owing to having:

Wherein, Δ w ₁(t), Δ w ₂(t) ..., Δ w ₂(t) be respectively i the element of compensation incremental array Δ W (t), and have:

X _ts(t)＝[x _ts，1(t)，x _ts，2(t)，…，x _ts，i(t)]，

${\mathrm{<figure-callout\; id="1"\; label="X\; k"\; filenames="HDA0000104591580000021.png"\; state="\{\{state\}\}">X</figure-callout>}}^k(t-1)=[x_1^{\mathrm{<figure-callout\; id="1"\; label="k"\; filenames="HDA0000104591580000021.png"\; state="\{\{state\}\}">k</figure-callout>}1\mathrm{st}}(t-1),x_2^{\mathrm{<figure-callout\; id="1"\; label="k"\; filenames="HDA0000104591580000021.png"\; state="\{\{state\}\}">k</figure-callout>}1\mathrm{st}}(t-1),...,x_i^{k1\mathrm{st}}(t-1)];$

Be respectively array cache X ^K1st(t-1) an i element, their value equals respectively this previous moment (t-1) k _1stI the feature value of electrical signals at the neuron average center of individual self organizing neural network; Therefore have:

Formula (7);

In formula (7),

Mean respectively x _{Ts, 1}(t), x _{Ts, 2}(t) ..., x _{Ts, i}(t) by the value after normalization,

Mean respectively By the value after normalization.If p gas sensor in gas sensor array (p ∈ 1,2 ..., i) for the current time drift compensation, train gases used susceptibility very low, cause p gas sensor corresponding

With

Value all approach 0, from formula (7), can see,

Value will approach 0; If q gas sensor in gas sensor array (q ∈ 1,2 ..., i) for the current time drift compensation, train gases used detection not drift about, have

Therefore Value also will approach 0.So, i the element Δ w of the compensation incremental array Δ W (t) of current time ₁(t), Δ w ₂(t) ..., Δ w ₂(t) among, only non-vanishing for those element values of the gases used comparatively sensitivity of current time drift compensation training and existence detection drift, therefore utilize the compensation incremental array Δ W (t) of current time to carry out the drift compensation of b times of ratio, can make the neuron of other self organizing neural network of current time

(get k=1,2 ..., K and k ≠ k _1stOr k _2nd) i feature value of electrical signals in more responsive and exist the feature value of electrical signals that detects drift all to be able to more appropriate drift inhibition to compensate for the drift compensation training institute gas that adopts, allow neuron

(get k=1,2 ..., K and k ≠ k _1stOr k _2nd) in actual electrical signal value output when obtaining drift-compensated feature value of electrical signals and being tending towards approaching gas sensor and detecting its coupling gas, thereby allow each neuron of the self organizing neural network outside above-mentioned two situations

(get k=1,2 ..., K and k ≠ k _1stOr k _2nd) also can obtain corresponding drift inhibition.

B5) according to the following formula to sample buffer memory matrix

In the value of each array cache carry out the iteration renewal:

Formula (8);

In formula (8), X ^k(t) mean current time sample buffer memory matrix In k array cache; X ^k(t-1) be illustrated in this previous moment (t-1) sample buffer memory matrix In k array cache,

Mean to get described array cache X ^k(t-1) normalized value.

Formula (8) is equivalent to sample buffer memory matrix

In each array cache all using the compensation incremental array Δ W (t) of current time and carry out the iteration renewal as iterative increment.So far, drift compensation and iteration renewal process are accomplished.

C) circulation execution step B), until Electronic Nose stops drift inhibition work.

Circulation is carried out the gap periods (cycle frequency in other words) of drift inhibition compensation process and can be determined according to actual system hardware situation and application demand.If the handling property of system hardware is very good, system clock frequency can be carried out to the cycle frequency that drift suppresses compensation process as circulation; The user also can be according to practical situations, the self-defined gap periods that circulation execution drift inhibition compensation process is set, the gap periods that circulation for example is set is 0.5 second, 10 seconds, 5 minutes, 2 hours ... etc., or the gas of carrying out certain number of times is set detects just circulation and carry out once drift and suppress compensation process; For, the aobvious Electronic Nose in confused situation of drifting about very good for system stability, even can interval a couple of days circulation carry out once drift and suppress compensation process.Until stop drift inhibition work, circulation stops.

The above-mentioned concrete steps that are the Electronic Nose drift rejection method that the present invention is based on multiple self-organizing neural networks.

What be worth paying special attention to is, the step b4 of the inventive method drift inhibition compensation process) in, through type (4) or formula (5) are to each neuronic drift compensation amount of each self organizing neural network of Electronic Nose, and step b5) in, through type (8) is to the iteration of each array cache of sample buffer memory matrix new increment more, and the two is numerically also unequal.The reason of processing like this is, if the drift of carving at a time suppresses in compensation process to occur a little specific factors, causes step b2) the electric signal output array X that obtains _Ts(t) accidental detection error appears, but due to step b4) in through type (4) or formula (5) to each neuronic drift compensation amount of each self organizing neural network of Electronic Nose by X _TsOr Δ W (t) has carried out a doubly or the dwindling of b times of ratio, so X (t) _Ts(t) the accidental detection error occurred can't fully, directly be embodied among each self organizing neural network neuron of Electronic Nose after this moment drift compensation; Simultaneously, step b5) in through type (8) to the iteration of each array cache of sample buffer memory matrix more new increment be Δ W (t), make X _Ts(t) the accidental detection error occurred is embodied directly in sample buffer memory matrix

Each array cache among, be conducive to like this in after this drift constantly suppresses compensation process, by an after this renewal that constantly compensates incremental array Δ W (t+1), by X _Ts(t) the accidental detection error occurred instant compensation is again returned, thereby has avoided X _Ts(t) the accidental detection error occurred is carried out long-term the existence or error accumulation in the process that suppresses compensation process of drifting about in circulation.The present invention strengthens the robust performance of the Electronic Nose drift rejection method based on multiple self-organizing neural networks just by this measure, simultaneously, exactly because also Electronic Nose drift rejection method of the present invention has good robust performance, just make the inventive method can be applied in the gas testing process of Electronic Nose, to as drift compensation training gas sample, use as the gas of detected object simultaneously, drift about when Electronic Nose is carried out the gas detection and suppress operation, unnecessary worry Electronic Nose occurs that because of accidentalia the detection error causes this error to be retained in drift and suppresses to affect the drift rejection in the middle of flow process in testing process.

On the other hand, at the inventive method step b4) to each self organizing neural network in each neuronic drift compensation and step b5) to sample buffer memory matrix

in the iteration of value of each array cache all got the normalized value of operational parameter in upgrading, this is because the gas sample that suppresses to adopt in compensation process that drifts about may concentration be had nothing in common with each other, gas sensor array also can there are differences for its detection electric signal output array value of gas sample of the same race of variable concentrations, and normalization can be eliminated the checkout discrepancy that gas sample concentration causes, and retain the responsive sign of gas sensor array to gas sample, therefore can avoid Electronic Nose to suppress in compensation process the difference of gas sample concentration is mistaken for it to detect drift and the wrong drift compensation that carries out in drift, also correspondingly strengthened the antijamming capability of Electronic Nose identification gas with various simultaneously.The concrete operation mode of getting normalized value has a lot, and different those skilled in the art, in different application scenarios, can adopt different normalization algorithms according to the needs of its technology custom or actual conditions.

Three, experiment effect checking.

Below in conjunction with drawings and Examples, technical scheme of the present invention is further described.

In order to verify the actual effect of the inventive method, adopt the Electronic Nose test platform of building voluntarily to be tested.The Electronic Nose test platform of building voluntarily as shown in Figure 2.The Electronic Nose that experiment adopts consists of gas sensor array 4 and computing machine 5; The gas sensor array that this experiment adopts is comprised of four metal-oxide gas transducers, be respectively GSBT11, TGS2620, TGS2602 and TGS2201, inner integrated two gas sensors of TGS2201 wherein, include altogether 5 gas sensors (in the inventive method, being equivalent to get i=5) therefore be equivalent to gas sensor array; Computing machine serves as Signal Pretreatment unit and two parts of pattern recognition unit of Electronic Nose, for storing each neuron of multiple self-organizing neural networks, and executing data collection and detection identification work, in the multiple self-organizing neural networks that this test Computer is stored, being divided into is that 3 self organizing neural networks are (in the inventive method, be equivalent to get K=3), be respectively used to mate pure air (as the gas sample of base-line data), carbon monoxide (CO) and formaldehyde (CH ₂0) three class gases, the data acquisition module of computing machine adopts 12 bits serial A/D chip TLC2543, acquisition time is spaced at every turn is about 1 second, and the data that collect are transferred in computing machine and are detected identifying processing by serial ports, detects recognition processing software and adopts MATLAB7.1.In experiment, adopt the Electronic Nose test platform of building voluntarily shown in Fig. 2, tested gas sample is formulated in Storage Time in Gas Collecting Bag 1, then adopt pump suction mode to utilize air intake pump 2 that gas is squeezed in test chamber 3; The gas sensor array 4 of Electronic Nose is positioned in test chamber 3, and gas is detected, and its signal output array is gathered and identified by computing machine 5; Finally, the gas that utilizes out air pump 7 to discharge in test chamber 3, and controlled by 6 pairs of air outputs of valve.

In this experiment, adopt the above-mentioned Electronic Nose test platform of building voluntarily, within the time of one month, utilize respectively existing MSOM drift compensation training method and the inventive method to carry out separately 50 gas sample tests, these 50 times the test for the test gas sample be respectively in order: for carbon monoxide (CO), test 10 times, for formaldehyde (CH ₂O) test 15 times, then test 25 times for carbon monoxide (CO); Each test, first utilize pure air to gather 50 of base-line datas, 100 of recycling test gas sample collection test datas, i.e. and each test gathers 150 of the electric signal output array data of gas sensor matrix altogether; Each electric signal output array includes 5 electric signal output valves, i.e. the electric signal output valve of 5 gas sensors in the Electronic Nose gas sensor array.Therefore, test altogether in this experiment and obtain 7500 of electric signal output array data, 1st～1500 is the electric signal output array data (comprising 500 base-line datas) for the CO gas sample, and 1501st～3750 is for CH ₂The electric signal output array data of O gas sample (comprising 750 base-line datas), 3751st～7500 is the electric signal output array data (comprising 1250 base-line datas) for the CO gas sample.In the test that adopts the inventive method, the compensating proportion coefficient a=0.3 got, compensation increment coefficient b=0.2, the normalization operational formula adopted is:

Wherein, symbol

Mean to get normalized value; F means any array that comprises i element, and i means the number of gas sensor in the Electronic Nose gas sensor array; f ₁, f ₂..., f _iMean respectively i the element that described array F comprises.For this test, comprise altogether 3 self organizing neural networks in the multiple self-organizing neural networks of Electronic Nose used, therefore get i=3.

Finally, utilize existing MSOM drift compensation training method and the inventive method respectively to carry out the detection recognition result that above-mentioned 50 gas samples test obtains respectively as shown in table 1:

Table 1

Adopt in 50 tests that existing MSOM drift compensation training method carries out, for 1st～3750 test datas, (10 test datas for the CO gas sample of front and 15 times are for CH ₂The test data of O gas sample) recognition effect is good, recognition accuracy is 100%, but since the 3751st test data (being last 25 test datas for the CO gas sample), Electronic Nose only can correctly be identified base-line data, for the test of CO gas sample whole identification errors, therefore recognition accuracy is only 33%, and the recognition accuracy be equivalent to the CO gas sample is 0%.Its reason is to be, for CH ₂During 1501st～3750 test datas that the O gas sample is tested, MSOM drift compensation training method is not carried out drift compensation to the neuron in the self organizing neural network that the CO gas sample is mated, yet the gas sensor matrix is for CO gas and CH ₂The sensitive features correlativity of O gas is larger, for CH ₂The O gas sample tested during the aging of gas sensor that cause also can affect relatively its detection to CO gas, the electric signal output array value that makes gas sensor array detect the CO gas sample is drifted about, and has caused last 25 identification errors for the test of CO gas sample.This also should demonstrate,prove the Limitation Analysis of the present invention to existing MSOM drift compensation training method.

And in 50 tests that employing the inventive method is carried out, because having for responsive correlativity neuron, the inventive method carries out drift-compensated characteristic, make Electronic Nose still can keep good recognition performance after drift suppresses compensation, so the discriminations that Electronic Nose is in the end tested for the CO gas sample for 25 times still can reach 100%.Visible by above-mentioned contrast, the Electronic Nose drift rejection method that the present invention is based on multiple self-organizing neural networks is compared to existing MSOM drift compensation training method, the drift compensation execution efficiency of Electronic Nose and recognition performance is had to add significantly and improve.

Finally explanation is, above embodiment is only unrestricted in order to technical scheme of the present invention to be described, although with reference to preferred embodiment, the present invention is had been described in detail, those of ordinary skill in the art is to be understood that, can modify or be equal to replacement technical scheme of the present invention, and not breaking away from aim and the scope of technical solution of the present invention, it all should be encompassed in the middle of claim scope of the present invention.

Claims (2)

1. the Electronic Nose drift rejection method based on multiple self-organizing neural networks, is characterized in that, comprises the steps:

A) initialization step, it is specially:

A1) set up sample buffer memory matrix ${\tilde{X}}_{\mathrm{Rct}}\left(t\right)=\{X^1\left(t\right),X^2\left(t\right),\&CenterDot;\&CenterDot;\&CenterDot;,X^K\left(t\right)\};$ Wherein, X ^k(t) array cache for comprising i element, k ∈ 1,2 ..., K}, K means the quantity of the self organizing neural network of Electronic Nose, and i means the number of gas sensor in the Electronic Nose gas sensor array, and t means constantly;

A2) initialization moment t=0;

A3) at current time, sample buffer memory matrix

In the value of each array cache be:

$X^k\left(t\right)=\frac{1}{M_k}\underset{m=1}{\overset{M_k}{\mathrm{\&Sigma;}}}{W}_m^k\left(t\right),k=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,K;$

Wherein,

Be illustrated in m neuron in k self organizing neural network of current time Electronic Nose,

Mean described neuron

In i feature value of electrical signals, M _kNeuronic quantity in k self organizing neural network of expression Electronic Nose;

B) drift suppresses compensation process; Be specially:

B1) t adds 1 certainly constantly;

B2) obtain the electric signal output array X of current time Electronic Nose gas sensor array _Ts(t):

X _ts(t)=[x _ts,1(t),x _ts,2(t),…,x _ts,i(t)]；

Wherein, x _{Ts, 1}(t), x _{Ts, 2}(t) ..., x _{Ts, i}(t) be illustrated in the electric signal output valve of i gas sensor in current time Electronic Nose gas sensor array;

B3) ask for coupling triumph self organizing neural network sequence number k _1stWith coupling time triumph self organizing neural network sequence number k _2nd:

Wherein,

Be illustrated in m neuron in k self organizing neural network of this previous moment (t-1) Electronic Nose; Symbol [[]] means to get normalized value, [[X _Ts(t)]], and

Mean to get respectively described electric signal output array X _Ts(t) normalized value and get described neuron

Normalized value;

Mean to get [[X _Ts(t)]] with

Euclidean distance;

Expression is got

With

Euclidean distance all k ∈ 1,2 ..., K} and m ∈ 1,2 ..., M _kMinimum value in situation,

Expression is got

With

Euclidean distance all k ∈ 1,2 ..., K} and m ∈ 1,2 ..., M _kOnly be greater than in situation

Inferior minimum value;

B4) if k _1st=k _2nd, according to the following formula each neuron in each self organizing neural network of Electronic Nose is carried out to drift compensation:

If k _1st≠ k _2nd, according to the following formula each neuron in each self organizing neural network of Electronic Nose is carried out to drift compensation:

Wherein,

Be illustrated in m neuron in k self organizing neural network of current time Electronic Nose; △ W (t) means the compensation incremental array of current time, and $\mathrm{\&Delta;W}\left(t\right)=\left[\right[X_{\mathrm{ts}}\left(t\right)\left]\right]-\left[\right[X^{k_{1\mathrm{st}}}(t-1)\left]\right],$

Be illustrated in this previous moment (t-1) sample buffer memory matrix

In k _1stIndividual array cache,

Mean to get described array cache

Normalized value; A is the compensating proportion coefficient, and span is 0<a≤0.5; B is the compensation increment coefficient, and span is 0<b≤0.5;

B5) according to the following formula to sample buffer memory matrix

In the value of each array cache carry out the iteration renewal:

$X^k\left(t\right)=\left\{\begin{array}{cc}\left[\right[X_{\mathrm{ts}}\left(t\right)\left]\right],& k=k_{1\mathrm{st}}\\ \left[\right[X^k(t-1)\left]\right]+\mathrm{\&Delta;W}\left(t\right),& k\&NotEqual;k_{1\mathrm{st}}\end{array}\right.;$

Wherein, X ^k(t) mean current time sample buffer memory matrix

In k array cache; X ^k(t-1) be illustrated in this previous moment (t-1) sample buffer memory matrix In k array cache, [[X ^k(t-1)]] mean to get described array cache X ^k(t-1) normalized value;

C) circulation execution step B), until Electronic Nose stops drift inhibition work.

2. the Electronic Nose drift rejection method based on multiple self-organizing neural networks according to claim 1, is characterized in that, it is as follows that described symbol [[]] means to get the concrete operation formula of normalized value:

$\left[\right[F\left]\right]=\frac{F}{\sqrt{{\left(f_1\right)}^2+{\left(f_1\right)}^2+\&CenterDot;\&CenterDot;\&CenterDot;+{\left(f_i\right)}^2}};$

Wherein, F means any array that comprises i element, f ₁, f ₂..., f _iMean i the element that described array F comprises.

Images