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

Abstract

The invention provides an electronic nose drift suppression method based on multiple self-organizing neural networks, which uses a sample cache matrix to cache the neuron mean center of each self-organizing neural network and update iteratively in order to obtain the drift of each cycle The compensation incremental array for suppression compensation, while performing drift compensation for each neuron in the self-organizing neural network with gas matching used in drift compensation training, also uses the compensation incremental array to compensate for the neurons in other self-organizing neural networks. The gas-sensitive characteristic electrical signal values used in drift compensation training are all subjected to drift suppression compensation, so that the neurons of the neural network of each organization tend to approach the actual electrical signal output array value of the gas sensor array to detect its matching gas after drift compensation, so that The purpose of suppressing drift is achieved, the execution efficiency of drift compensation of the electronic nose is improved, and the electronic nose is guaranteed to maintain good recognition performance after drift suppression compensation.

Description

A drift suppression method for electronic nose based on multiple self-organizing neural networks

technical field

The invention belongs to the technical field of electronic nose detection and training, in particular to an electronic nose drift suppression method based on multiple self-organizing neural networks.

Background technique

The electronic nose is an electronic device or system that uses the response spectrum of a gas sensor array to identify odors. It mainly consists of three parts: a gas sensor array, a signal preprocessing unit, and a pattern recognition unit. The working principle of the electronic nose is: when a certain odor is presented in front of a gas sensor made of an active material, the gas sensor can convert the chemical input of the gas into an electrical signal output, and multiple gas sensors are used to form a gas sensor array. The response of multiple gas sensors to a smell constitutes the electrical signal output array of the gas sensor array to the smell; in order to realize the qualitative or quantitative analysis of the smell, the electrical signal output of the gas sensor must be properly preprocessed (eliminated Noise, feature extraction, signal amplification, normalization, etc.), the pattern recognition unit adopts an appropriate pattern recognition analysis method to identify and process it; in theory, each odor will have its corresponding characteristics for the gas sensor array Electrical signal array value, the characteristic electrical signal array value corresponding to different odors is stored in the pattern recognition unit as the neuron of the electronic nose, and the electrical signal output array value of the gas sensor array is compared and matched with the neuron when performing gas detection. Different gases can be distinguished, and at the same time, the gas sensor array can be used to measure the cross-sensitivity of multiple gases, and the mixed gas analysis can be realized through appropriate analysis methods. Due to the short time and low cost of electronic nose detection, it has been widely researched and paid attention to in the fields of food detection, disease diagnosis, and environmental detection.

The neuron for detecting and identifying gas by the electronic nose can be established and obtained through benchmark training, that is, the electronic nose is used to perform prior detection of various gas samples that can be identified, and the characteristic electrical signal array value of the gas sensor array for the various gases is obtained. It is stored as the neurons matching the multiple gases, and used as the identification reference for the multiple gases. However, the electrical signal output array value of the electronic nose gas sensor array sensing the same gas is not invariable. The occurrence of drift is often an important factor affecting the recognition effect of the electronic nose. There are two main reasons for the drift of the detection value of the gas sensor array: one is due to changes in the working environment of the electronic nose, such as temperature, humidity, etc., so that the electrical signal output array value detected by the gas sensor array is within the matching neuron value Another reason is that the physical and chemical properties of the gas sensor have changed due to aging and other phenomena, which in turn affects the size of its electrical signal output value, resulting in the electrical signal output array value of the gas sensor array matching the nerve sensor array value. The element deviates and forms a drift. Among them, the drift caused by the former belongs to a kind of transient drift, which does not affect the recognition accuracy of the electronic nose in essence. The drift caused by the latter is generally called long-term drift, which will exist and accumulate for a long time with the use of the gas sensor. If no suppression measures are taken, the detection accuracy of the electronic nose will be significantly reduced, so suppress or reduce the long-term drift of the gas sensor. The influence is particularly important to ensure the detection accuracy and effect of the electronic nose.

From the current research, the long-term drift of the electronic nose is suppressed mainly by correcting the values of the neurons in the electronic nose. There are two main types of existing methods for suppressing electronic nose drift: one is to treat the drift of the electrical signal output array value of the electronic nose gas sensor array as an independent signal, and through principal component analysis, independent component analysis, orthogonal decomposition, etc. Mathematical methods remove it from the electrical signal output by the sensor (see the document "Bouwmans T, Baf F 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 the literature "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.), this method works better in theory , but it needs to have complete drift prior information as the basis for elimination. However, it is difficult to accurately summarize and grasp the regularity of the electrical signal output array value drift of different gas sensors. Therefore, it is very important to establish complete drift prior information. technical difficulty. Another method does not require complete drift prior information, and mainly uses a single-layer self-organizing neural network (SelfOrganizing Maps, referred to as SOM network) for drift compensation training to compensate the electrical signal output of the gas sensor array with drift, that is, all Neurons are used as a self-organizing neural network, each neuron is the characteristic electrical signal array value of a gas relative to the gas sensor array, and multiple neurons are set for each gas, and the multiple neurons are set at a certain Different values are taken within the range of values to form the neuron recognition interval for the matching gas, and then the electronic nose is again used to perform drift compensation training on the gas matched by the neurons. During the drift compensation training, if the electronic nose gas sensor array senses The measured electrical signal output array value deviates from the central value of the neuron identification interval matched by the gas, indicating that the gas sensor array detects that the electrical signal output array value of the gas has drifted. Each neuron in the neuron identification interval performs drift compensation, so as to achieve the purpose of suppressing the drift of the detection value (see the literature "Kohonen T. The Self-organizing Maps [J] Proceedings of the IEEE, 1990, 78 (9): 1464-1480 ”), but because the single-layer self-organizing neural network is equivalent to listing all the stored neurons in a neural network plane, the neuron recognition interval after drift compensation may overlap with the neuron recognition interval of other gases, As a result, the information of the overlapped part of neurons in the neuron recognition intervals of other gases is lost, which will not only affect the effect of drift compensation, but if the overlap between different neuron recognition intervals caused by drift compensation is serious, it may even cause the entire self-organizing nerve The neuron information in the network is chaotic, which seriously affects the recognition performance of the electronic nose.

In recent years, some scholars have proposed to use multiple self-organizing neural network (Multiple Self Organizing Maps, referred to as MSOM) to solve the drift problem, that is, multiple neurons for one gas will be listed on an independent self-organizing neural network to detect multiple gases. Construct multiple self-organizing neural networks, so each self-organizing neural network corresponds to the neuron recognition interval of a gas, so that the drift compensation training for a gas will only perform drift compensation on a self-organizing neural network that matches the gas , the drift compensation processing method for a single self-organizing neural network is the same as the drift compensation processing method for a single neuron identification interval in the SOM network drift compensation method (see the literature "MarziaZuppa, Cosimo Distante, Pietro Siciliano, et al.Drift counteraction with multiple self-organizing maps for an electronic nose[J]Sensors and Actuators B, 2004, 98: 305-317"), so that the drift compensation will not affect the neurons of other self-organizing neural networks matching other gases, avoiding There is a situation where neurons lose information due to mutual interference; but this makes some neurons on the self-organizing neural network that have not been trained for drift compensation for a long time hardly receive any drift compensation during drift compensation training, resulting in their drift. As it continues to deteriorate, when gas detection is performed again, the accuracy of gas detection and recognition matched by the self-organizing neural network without drift compensation will be greatly reduced, which affects the recognition performance of the electronic nose. Unless all kinds of gases identified by the electronic nose are used to perform drift compensation training on the electronic nose separately, ensuring that the self-organizing neural network matched with each gas can be compensated for drift, in order to overcome the occurrence of the above situation, but keep the drift of all kinds of gases for a long time Compensation training is difficult to achieve, and drift compensation training has a large workload and troublesome operation, which makes it difficult to guarantee the drift suppression effect and electronic nose recognition performance.

Contents of the invention

Aiming at the above-mentioned problems existing in the prior art, the present invention proposes an electronic nose drift suppression method based on multiple self-organizing neural networks in order to enhance the overall drift compensation capability of the electronic nose for long-term drift, so as to improve the drift compensation execution efficiency of the electronic nose , to ensure that the electronic nose still maintains good recognition performance after drift suppression compensation.

To achieve the above object, the present invention adopts the following technical means:

A method for suppressing electronic nose drift based on multiple self-organizing neural networks, comprising the steps of:

A) initialization step, which is specifically:

a1) Establish a sample cache matrix ${\tilde{x}}_{\mathrm{Rct}}\left(t\right)=\{x^1\left(t\right),x^2\left(t\right),...,x^K\left(t\right)\};$ Among them, X ^k (t) is a cache array containing i elements, k ∈ {1, 2, ..., K}, K represents the number of self-organizing neural networks of the electronic nose, and i represents the gas sensor in the gas sensor array of the electronic nose The number of , t represents the time;

a2) initialization time t=0;

中各个缓存阵列的取值为：a3) At the current moment, the sample cache matrix

The value of each cache array in is:

${\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; ;</span>$

则分别表示所述神经元中的i个特征电信号值，M_k表示电子鼻的第k个自组织神经网络中神经元的数量；in, $W_m^k\left(t\right)=[w_{m,1}^k\left(t\right),w_{m,2}^k\left(t\right),...,w_{m,i}^k\left(t\right)]$ Indicates the mth neuron in the kth self-organizing neural network of the electronic nose at the current moment,

respectively represent the neuron The i characteristic electrical signal values in , M _k represents the number of neurons in the kth self-organizing neural network of the electronic nose;

B) Drift suppression compensation step; specifically:

b1) time t is incremented by 1;

b2) Obtain the electrical signal output array X _ts (t) of the electronic nose gas sensor array at the current moment:

_Xts (t) = [ _{xts, 1} (t), _{xts, 2} (t), ..., xts _{, i} (t)];

Wherein, x _{ts, 1} (t), x _{ts, 2} (t), ..., x _{ts, i} (t) respectively represent the electrical signal output values of i gas sensors in the electronic nose gas sensor array at the current moment;

b3) Calculate the self-organizing neural network serial number k _{1st of} the winning match and the self-organizing neural network serial number k _2nd of the winning match:

表示在此前一时刻(t-1)电子鼻的第k个自组织神经网络中第m个神经元；符号

表示取归一化值，

和

分别表示取所述电信号输出阵列X_ts(t)的归一化值和取所述神经元

的归一化值；

表示取

与

的欧氏距离；

表示取

与

的欧氏距离在所有k∈{1，2，…，K}和m∈{1，2，…，M_k}情况中的最小值，表示取

与

的欧氏距离在所有k∈{1，2，…，K}和m∈{1，2，…，M_k}情况中仅大于

的次最小值；in,

Indicates the mth neuron in the kth self-organizing neural network of the electronic nose at the previous moment (t-1); symbol

Indicates the normalized value,

and

Respectively represent to take the normalized value of the electrical signal output array X _ts (t) and take the neuron

normalized value of

Indicates to take

and

Euclidean distance;

Indicates to take

and

The minimum value of the Euclidean distance for all k ∈ {1, 2, ..., K} and m ∈ {1, 2, ..., M _k } cases, Indicates to take

and

The Euclidean distance of is only _greater than

the second minimum value of

b4) If k _1st =k _2nd , perform drift compensation for each neuron in each self-organizing neural network of the electronic nose according to the following formula:

If k _1st ≠k _2nd , then perform drift compensation for each neuron in each self-organizing neural network of the electronic nose according to the following formula:

表示在当前时刻电子鼻的第k个自组织神经网络中第m个神经元；ΔW(t)表示当前时刻的补偿增量阵列，且

X^k1st(t-1)表示在此前一时刻(t-1)样本缓存矩阵中第k_1st个缓存阵列，

表示取所述缓存阵列X^k1st(t-1)的归一化值；a为补偿比例系数，取值范围为0＜a≤0.5；b为补偿增量系数，取值范围为0＜b≤0.5；in,

Represents the mth neuron in the kth self-organizing neural network of the electronic nose at the current moment; ΔW(t) represents the compensation increment array at the current moment, and

X ^k1st (t-1) represents the sample buffer matrix at the previous moment (t-1) In the k _1st cache array,

Indicates that the normalized value of the cache array X ^k1st (t-1) is taken; a is the compensation proportional coefficient, the value range is 0<a≤0.5; b is the compensation increment coefficient, the value range is 0<b≤ 0.5;

b5) According to the following formula, the sample cache matrix The value of each cache array in is updated iteratively:

中的第k个缓存阵列，

表示取所述缓存阵列X^k(t-1)的归一化值；Among them, X ^k (t) represents the sample buffer matrix at the current moment The k-th cache array in ; X ^k (t-1) represents the sample cache matrix at the previous moment (t-1)

The kth cache array in ,

Represents taking the normalized value of the cache array X ^k (t-1);

C) Step B) is executed cyclically until the electronic nose terminates the drift suppression work.

表示取归一化值的具体运算公式如下：Further, the symbol

The specific calculation formula for taking the normalized value is as follows:

Wherein, F represents any array containing i elements, and f ₁ , f ₂ , . . . , f _i respectively represent i elements contained in the array F.

Compared with the prior art, the present invention has the following beneficial effects:

1. In the electronic nose drift suppression method of the present invention, using any kind of gas for drift compensation training can make the characteristic electrical signal values sensitive to the gas in the neurons of each self-organizing neural network be able to drift suppression and compensation, so as long as the gas is used One or a small number of gases that can make all the gas sensors in the gas sensor array can be sensitively detected can be trained for drift compensation, which can ensure that each neuron of each self-organizing neural network in the electronic nose can be drift-compensated and enhanced. The overall drift compensation capability of the electronic nose for long-term drift is improved, and the drift compensation execution efficiency of the electronic nose is improved.

2. In the electronic nose drift suppression method of the present invention, the neuron mean center of each self-organizing neural network is cached and cyclically iteratively updated through the sample cache matrix to obtain the compensation increment array for each cycle drift suppression compensation , while performing drift compensation for each neuron in the self-organizing neural network with gas matching used in drift compensation training, the compensation increment array is also used to compensate for the drift of neurons in other self-organizing neural networks, and the cycle of the sample buffer matrix The iterative update ensures that the compensation incremental array always obtains an appropriate value, so that the neurons of the neural network of each organization tend to approach the actual electrical signal output array value of the gas sensor array to detect its matching gas after drift compensation, so as to achieve the suppression of drift The purpose is to ensure that the electronic nose still maintains good recognition performance after drift suppression compensation.

Description of drawings

Fig. 1 is the flowchart block diagram of the electronic nose drift suppression method based on multiple self-organizing neural networks of the present invention;

Figure 2 is a schematic diagram of the structure of the self-built electronic nose test platform.

Detailed ways

1. Limitations of existing MSOM drift compensation methods.

According to the working principle of the MSOM drift compensation method, it is assumed that after the electronic nose passes the benchmark training, K self-organizing neural networks are stored, which are used to match and identify K different gases; if the k-th self-organizing neural network (k∈{ 1, 2,..., K}) matched gas samples are continuously input to the electronic nose for drift compensation training at Δt time; during this Δt time, each neuron in the kth self-organizing neural network can effectively obtain drift compensation , but the neurons in the self-organizing neural network other than the k-th self-organizing neural network can hardly obtain the chance of drift compensation. If a new gas sample that is not matched by the k-th self-organizing neural network is input to the electronic nose at (Δt+1) time, because after the previous Δt time, the electronic nose gas sensor array will detect the new gas sample. The value of the signal output array may have drifted, and the self-organizing neural network matching the new gas sample has not received any drift compensation, so the electronic nose cannot guarantee the accuracy of the identification of the new gas sample at (Δt+1) time .

Two, the solution of the present invention.

In view of the above-mentioned current situation and shortcomings of the prior art, the present invention further improves the existing multiple self-organizing neural network drift compensation method, and proposes a method for suppressing electronic nose drift based on multiple self-organizing neural networks. The electronic nose drift suppression method of the present invention mainly includes two steps of initialization and drift suppression compensation, and the compensation increment array is cyclically and iteratively updated by performing the drift suppression compensation step in order to realize the self-organizing neural networks of the electronic nose. The unit performs cyclic iteration drift compensation, so that the neurons of each self-organizing neural network always approach the gas sensor array to detect the actual electrical signal output array value of its matching gas, so as to achieve the purpose of suppressing drift.

The flow chart of the electronic nose drift suppression method based on multiple self-organizing neural networks of the present invention is shown in Figure 1, and its specific steps are as follows:

A) Initialization step.

This step is mainly to establish a sample cache matrix, which is used to cache the neuron mean center of each self-organizing neural network as a data source for loop iteration update, and at the same time initialize the sample cache matrix and the value of time t. It should be noted that the time t is not used to represent a specific time value (such as t seconds, t minutes, etc.), and the time t in the method of the present invention is used as a time distinguishing mark to mark different time points in the loop iteration process , to realize the distinction between before and after data in the loop iteration process.

The specific content of the initialization step is as described in the following steps a1) to a3):

a1) Establish a sample cache matrix ${\tilde{x}}_{\mathrm{Rct}}\left(t\right)=\{x^1\left(t\right),x^2\left(t\right),...,x^K\left(t\right)\};$ Among them, X ^k (t) is a cache array containing i elements, k ∈ {1, 2, ..., K}, K represents the number of self-organizing neural networks of the electronic nose, i represents the gas in the gas sensor array of the electronic nose The number of sensors, t represents the time.

a2) Initialization time t=0; time t=0 indicates that the current time is the initial state time, so as to distinguish it from the subsequent loop iteration time.

a3) At the current moment, the sample cache matrix The value of each cache array in is:

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

则分别表示所述神经元

中的i个特征电信号值，也就是神经元

相匹配的气体分别对应于电子鼻气体传感器阵列中i个气体传感器的i个特征电信号值，M_k表示电子鼻的第k个自组织神经网络中神经元的数量。电子鼻各自组织神经网络中神经元的数量M_k的多少是根据需要而自定义设定的，不同自组织神经网络对应的M_k的值可以是不同的，因此可以认为M_k是k的函数，其函数关系由定义设定的任意第k个自组织神经网络中神经元的数量的而确定。In formula (1), $W_m^k\left(t\right)=[w_{m,1}^k\left(t\right),w_{m,2}^k\left(t\right),...,w_{m,i}^k\left(t\right)]$ Indicates the mth neuron in the kth self-organizing neural network of the electronic nose at the current moment,

respectively represent the neuron

The i characteristic electrical signal values in , that is, the neuron

The matched gases correspond to i characteristic electrical signal values of i gas sensors in the electronic nose gas sensor array, and M _k represents the number of neurons in the kth self-organizing neural network of the electronic nose. The number of neurons M _k in each organization neural network of the electronic nose is customized according to the needs, and the value of M _k corresponding to different self-organizing neural networks can be different, so it can be considered that M _k is a function of k , and its functional relationship is determined by defining the number of neurons in any kth self-organizing neural network.

中各个缓存阵列X^k(t)的取值即为初始状态时刻取值；缓存阵列X^k(t)的初始状态时刻取值为

即第k个自组织神经网络的神经元均值中心。在后续的循环迭代过程中则无需再依靠求取电子鼻各自组织神经网络的神经元均值中心确定样本缓存矩阵的取值，而是通过循环迭代更新而确定电子鼻各自组织神经网络新的神经元均值中心。The current moment is t=0, so the sample buffer matrix at this time

The value of each cache array X ^k (t) is the value of the initial state moment; the value of the initial state moment of the cache array X ^k (t) is

That is, the neuron mean center of the kth self-organizing neural network. In the subsequent cyclic iteration process, it is no longer necessary to determine the value of the sample cache matrix by calculating the mean center of the neurons of the respective tissue neural networks of the electronic nose, but to determine the new neurons of the respective tissue neural networks of the electronic nose through cyclic iterative updates. mean center.

B) Drift suppression compensation step.

The drift suppression compensation step is to use the neurons of the electronic nose to organize the neural network at the previous moment and the sample buffer matrix at the previous moment to carry out drift suppression and compensation for the neurons of the electronic nose's respective neural network at the current moment, and to compensate the neurons of the neural network at the current moment. The sample cache matrix is iteratively updated, and is called by the drift suppression and compensation step of the loop iteration at a later moment.

The specific content of the drift suppression compensation step is as described in the following steps b1) to b5):

b1) The time t is incremented by 1, which is used as an identifier of the current time to distinguish the previous time.

b2) Obtain the electrical signal output array X _ts (t) of the electronic nose gas sensor array at the current moment:

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) respectively represent the electrical signal output of i gas sensors in the electronic nose gas sensor array at the current moment value.

At this time, the electrical signal output array X _ts (t) of the gas sensor array of the electronic nose is equivalent to the actual electrical signal output array obtained by detecting the gas used for drift compensation training by the gas sensor array of the electronic nose at the current moment. The drift compensation training can be carried out through the drift compensation training experiment alone, or it can be carried out simultaneously in the process of using the electronic nose for gas detection; the latter is equivalent to using the gas as the detection object as a drift compensation training gas sample at the same time, The drift suppression operation is performed while the electronic nose is performing gas detection. In the actual program, the gas detection operation and the drift suppression operation process can be executed separately through multi-threaded tasks without interfering with each other. The electrical signal output array X _ts (t) will be used as a reference for drift suppression compensation at the current moment, and through subsequent steps, the neurons in the self-organizing neural network with gas matching used in drift compensation training will approach this value after drift compensation. The electrical signal output array X _ts (t).

b3) Calculate the self-organizing neural network serial number k _{1st of} the winning match and the self-organizing neural network serial number k _2nd of the winning match:

Formula (3);

表示在此前一时刻(t-1)电子鼻的第k个自组织神经网络中第m个神经元；符号表示取归一化值，

和

分别表示取所述电信号输出阵列X_ts(t)的归一化值和取所述神经元

的归一化值；

表示取

与

的欧氏距离；

表示取

与

的欧氏距离在所有k∈{1，2，…，K}和m∈{1，2，…，M_k}情况中的最小值，

表示取

与

的欧氏距离在所有k∈{1，2，…，K}和m∈{1，2，…，M_k}情况中仅大于

的次最小值。In formula (3),

Indicates the mth neuron in the kth self-organizing neural network of the electronic nose at the previous moment (t-1); symbol Indicates the normalized value,

and

Respectively represent to take the normalized value of the electrical signal output array X _ts (t) and take the neuron

normalized value of

Indicates to take

and

Euclidean distance;

Indicates to take

and

The minimum value of the Euclidean distance for all k ∈ {1, 2, ..., K} and m ∈ {1, 2, ..., M _k } cases,

Indicates to take

and

The Euclidean distance of is only _greater than

the second minimum value of .

The matching winning self-organizing neural network serial number k _1st and the matching winning self-organizing neural network serial number k _2nd obtained in this way are equivalent to: in each neuron of each self-organizing neural network of the electronic nose, after normalization The self-organizing neural network whose Euclidean distance is the closest to the electrical signal output array X _ts (t) of the electronic nose gas sensor array at the current moment is identified as the matching winning self-organizing neural network, and its self-organizing neural network number is taken as k _1st ; after normalization, the self-organizing neural network where the nearest neuron is located in its Euclidean distance and the electrical signal output array X _ts (t) times of the electronic nose gas sensor array at the current moment is identified as the matching times winning self-organizing neural network , and take its self-organizing neural network number as k _2nd . The purpose of obtaining the matching winning self-organizing neural network number k _1st and the matching winning self-organizing neural network number k _2nd mainly has two aspects: one is to identify the matching winning self-organizing neural network number k _1st as the drift compensation training center at the current moment The serial number of the self-organizing neural network matching the gas is used to distinguish the self-organizing neural network and the non-matching self-organizing neural network used in the drift compensation training to perform different drift compensation respectively; on the other hand, by comparing The values of k _1st and k _2nd can be used to know the degree of drift of the electronic nose at the current moment. If k _1st = k _2nd , it indicates that the drift of the gas used in the drift compensation training of the electronic nose at the current moment is still within the recognition range of the self-organizing neural network , if k _1st ≠ k _2nd , it indicates that the electronic nose detects the drift of the gas used in drift compensation training at the current moment and has reached the edge of the recognition range of the self-organizing neural network, so that it is convenient to take different measures for different drifts of the electronic nose in the subsequent steps. drift compensation strategy.

b4) If k _1st =k _2nd , perform drift compensation for each neuron in each self-organizing neural network of the electronic nose according to formula (4):

Formula (4);

If k _1st ≠k _2nd , then perform drift compensation for each neuron in each self-organizing neural network of the electronic nose according to formula (5):

;Formula (5);

X^k1st(t-1)表示在此前一时刻(t-1)样本缓存矩阵

中第k_1st个缓存阵列的值，即样本缓存矩阵

中第k个(取k＝k_1st)缓存阵列的值；

表示取所述缓存阵列X^k1st(t-1)的归一化值；a为补偿比例系数，取值范围为0＜a≤0.5；b为补偿增量系数，取值范围为0＜b≤0.5。In formula (4) and formula (5), Represents the mth neuron in the kth self-organizing neural network of the electronic nose at the current moment; ΔW(t) represents the compensation increment array at the current moment, and

X ^k1st (t-1) represents the sample buffer matrix at the previous moment (t-1)

The value of the k _1st cache array in , the sample cache matrix

The value of the kth (take k=k _1st ) cache array among them;

Indicates that the normalized value of the cache array X ^k1st (t-1) is taken; a is the compensation proportional coefficient, the value range is 0<a≤0.5; b is the compensation increment coefficient, the value range is 0<b≤ 0.5.

The specific values of the compensation proportional coefficient a and the compensation increment coefficient b need to be determined according to the sensitivity of the active material used in the gas sensor of the electronic nose to the gas detection drift in the actual situation. For gas sensors made of different active materials , the specific values of the corresponding compensation proportional coefficient a and compensation increment coefficient b are different; Within the interval (0, 0.5], usually a≠b (of course, it does not rule out the case where a=b exists in the compensation proportional coefficient a and compensation increment coefficient b corresponding to individual active materials); the drift of the active material to gas detection The more sensitive it is, the larger the values of the corresponding compensation proportional coefficient a and compensation increment coefficient b are.

In the above drift compensation process, it can be seen that for the K self-organizing neural networks of the electronic nose:

(取k＝k_1st，若k_1st＝k_2nd则k_1st和k_2nd表示同一个自组织神经网络)，直接利用当前时刻气体传感器阵列对漂移补偿训练所采用气体检测得到的实际电信号输出阵列X_ts(t)与该神经元在此前一时刻的值

(取k＝k_1st)之差进行a倍比例的漂移补偿，使得当前时刻的神经元

(取k＝k_1st)趋近于X_ts(t)。①At the current moment, each neuron of the k _1st self-organizing neural network that matches the gas used for drift compensation training at the current moment

(take k=k _1st , if k _1st =k _2nd then k _1st and k _2nd represent the same self-organizing neural network), directly use the gas sensor array at the current moment to output the actual electrical signal output array obtained by the gas detection used in the drift compensation training X _ts (t) and the value of the neuron at the previous moment

(take k=k _1st ) difference to perform a drift compensation in proportion to a times, so that the neuron at the current moment

(taking k=k _1st ) approaches X _ts (t).

(取k＝k_2nd)的漂移补偿为：②For the case of k _1st ≠ k _2nd , the k _2nd self-organizing neural network is the self-organizing neural network most correlated with the sensitive characteristics of the gas used for drift compensation training at the current moment, that is to say, the gas sensor of the electronic nose The array has a high correlation between the sensitive characteristics of the gas used in the drift compensation training and the sensitive characteristics of the gas matched by the k _2nd self-organizing neural network. If the detection of the gas used in the current drift compensation training by the electronic nose occurs If there is no drift, it means that the electronic nose also drifts correspondingly to the gas detection matched by the k _2nd self-organizing neural network; at the same time, k _1st ≠ k _2nd means that the electronic nose detects drift compensation training at the current moment. The degree of gas drift is relatively large, so if the electronic nose detects the gas matched by the k _2nd self-organizing neural network at the current moment, there must be a large amount of drift; considering these two factors, it is necessary to analyze the current k _2nd Each neuron of a self-organizing neural network performs large-scale drift compensation to suppress its large-scale drift. Through the drift compensation process described in the above formula (5), it can be seen that at the current moment, each neuron of the k _2nd self-organizing neural network

(taking k=k _2nd ) the drift compensation is:

取k＝k_2nd；

Take k=k _2nd ;

Formula (6);

中第k_1st个缓存阵列的归一化值之差而决定的，体现了当前时刻气体传感器阵列的对检测漂移补偿训练所采用气体的漂移程度，因此式(6)中采用了当前时刻的补偿增量阵列ΔW(t)进行b倍比例的漂移补偿；同时，又考虑到第k_2nd个自组织神经网络的神经元

(取k＝k_2nd)与当前时刻漂移补偿训练所采用气体的敏感特征相关性，因此式(6)中采用了当前时刻电子鼻气体传感器阵列的电信号输出阵列X_ts(t)对神经元(取k＝k_2nd)的漂移补偿进行了a倍比例的修正。Since the size of the compensation increment array ΔW(t) at the current moment is determined by the normalized value of the electrical signal output array X _ts (t) of the electronic nose gas sensor array at the current moment and the matching winning self-organizing neural network serial number k _1st corresponding to the previous One-time (t-1) sample buffer matrix

It is determined by the difference of the normalized value of the k _1st cache array in the current moment, which reflects the drift degree of the gas sensor array used in the detection drift compensation training at the current moment. Therefore, the compensation at the current moment is used in formula (6) Incremental array ΔW(t) performs b-fold drift compensation; at the same time, taking into account the neuron of the k _2nd self-organizing neural network

(taking k=k _2nd ) is correlated with the sensitive characteristics of the gas used in drift compensation training at the current moment, so the electrical signal output array X _ts (t) of the electronic nose gas sensor array at the current moment is used in the formula (6) to affect the neuron (taking k=k _2nd ) the drift compensation is corrected by a times ratio.

(取k＝1，2，…，K且k≠k_1st或k_2nd)，在当前时刻，为了使得神经元

(取k＝1，2，…，K且k≠k_1st或k_2nd)中与对漂移补偿训练所采用气体比较敏感并且存在检测漂移的特征电信号值都得以较适量的漂移抑制补偿，因此利用当前时刻的补偿增量阵列ΔW(t)进行b倍比例的漂移补偿；如此补偿的原理在于，由于有：③ For each neuron of the self-organizing neural network other than the above two cases

(take k=1, 2, ..., K and k≠k _1st or k _2nd ), at the current moment, in order to make the neuron

(take k=1, 2,..., K and k≠k _1st or k _2nd ) and the characteristic electrical signal values that are more sensitive to the gas used in the drift compensation training and have detection drift can be compensated by a relatively appropriate amount of drift suppression, so Utilize the compensation incremental array ΔW(t) at the current moment to perform b-fold drift compensation; the principle of such compensation is that, due to:

Among them, Δw ₁ (t), Δw ₂ (t), ..., Δw ₂ (t) are the i elements of the compensation incremental array ΔW(t), and there are:

X _ts (t) = [x _{ts, 1} (t), x _{ts, 2} (t), ..., x _{ts, i} (t)],

${\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="1"\; label="k"\; filenames="HDA0000104591580000021.png"\; state="\{\{state\}\}">k</figure-callout></span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; st</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="1"\; label="k"\; filenames="HDA0000104591580000021.png"\; state="\{\{state\}\}">k</figure-callout></span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; st</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="1"\; label="k"\; filenames="HDA0000104591580000021.png"\; state="\{\{state\}\}">k</figure-callout></span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; st</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; st</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; ;</span>$

Be respectively the i elements of the buffer array X ^k1st (t-1), and their values are respectively equal to the i characteristic electrical signal values of the neuron mean value center of the k _1st self-organizing neural network at the previous moment (t-1); So there are:

Formula (7);

分别表示x_ts，1(t)，x_ts，2(t)，…，x_ts，i(t)被归一化过后的值，

分别表示被归一化过后的值。如果气体传感器阵列中第p个气体传感器(p∈1，2，…，i)对于当前时刻漂移补偿训练所用气体的敏感度非常低，导致第p个气体传感器对应的

和

的值都逼近0，则从式(7)可以看到，

的值必将逼近0；如果气体传感器阵列中第q个气体传感器(q∈1，2，…，i)对于当前时刻漂移补偿训练所用气体的检测并未发生漂移，则有

因此的值也必将逼近0。所以，当前时刻的补偿增量阵列ΔW(t)的i个元素Δw₁(t)，Δw₂(t)，…，Δw₂(t)之中，只有对于当前时刻漂移补偿训练所用气体较为敏感并且存在检测漂移的那些元素值不为零，因此利用当前时刻的补偿增量阵列ΔW(t)进行b倍比例的漂移补偿，能够使得当前时刻其它自组织神经网络的神经元

(取k＝1，2，…，K且k≠k_1st或k_2nd)的i个特征电信号值中对于漂移补偿训练所采用气体比较敏感并且存在检测漂移的特征电信号值都得以较适量的漂移抑制补偿，让神经元

(取k＝1，2，…，K且k≠k_1st或k_2nd)中得到漂移补偿的特征电信号值趋于接近气体传感器检测其匹配气体时的实际电信号输出值，从而让上述两种情况之外的自组织神经网络的各个神经元

(取k＝1，2，…，K且k≠k_1st或k_2nd)也能取得相应的漂移抑制效果。In formula (7),

Represent x _{ts, 1} (t), x _{ts, 2} (t), ..., x _{ts, i} (t) are normalized values,

Respectively Normalized value. If the pth gas sensor (p ∈ 1, 2, ..., i) in the gas sensor array has a very low sensitivity to the gas used for drift compensation training at the current moment, resulting in the pth gas sensor corresponding to

and

The values are all close to 0, then it can be seen from formula (7),

The value of must be close to 0; if the qth gas sensor (q∈1, 2, ..., i) in the gas sensor array does not drift in the detection of the gas used for drift compensation training at the current moment, then there is

therefore The value of will also be close to 0. Therefore, among the i elements Δw ₁ (t), Δw ₂ (t), ..., Δw ₂ (t) of the compensation increment array ΔW(t) at the current moment, only the gas used for drift compensation training at the current moment is more sensitive And the value of those elements that detect drift is not zero, so using the compensation increment array ΔW(t) at the current moment to perform b-fold drift compensation can make the neurons of other self-organizing neural networks at the current moment

(take k=1, 2, ..., K and k≠k _1st or k _2nd ) among the i characteristic electric signal values, the characteristic electric signal values that are more sensitive to the gas used in the drift compensation training and have detection drift can be compared. compensation for drift inhibition, allowing neurons

(take k=1, 2, ..., K and k≠k _1st or k _2nd ) the characteristic electrical signal value obtained by drift compensation tends to be close to the actual electrical signal output value when the gas sensor detects its matching gas, so that the above two Each neuron of the self-organizing neural network other than this case

(k=1, 2, . . . , K and k≠k _1st or k _2nd ) can also achieve a corresponding drift suppression effect.

中各个缓存阵列的取值进行迭代更新：b5) According to the following formula, the sample cache matrix

The value of each cache array in is updated iteratively:

式(8)；

Formula (8);

表示取所述缓存阵列X^k(t-1)的归一化值。In formula (8), X ^k (t) represents the sample buffer matrix at the current moment The k-th cache array in ; X ^k (t-1) represents the sample cache matrix at the previous moment (t-1) The kth cache array in ,

Indicates that the normalized value of the cache array X ^k (t-1) is taken.

中的各个缓存阵列都以当前时刻的补偿增量阵列ΔW(t)作为迭代增量进行迭代更新。至此，一次漂移补偿和迭代更新过程即得以完成。Equation (8) is equivalent to the sample cache matrix

Each cache array in is iteratively updated with the compensation increment array ΔW(t) at the current moment as the iteration increment. So far, a process of drift compensation and iterative update is completed.

C) Step B) is executed cyclically until the electronic nose terminates the drift suppression work.

The interval period (or cycle frequency) for cyclically executing the drift suppression and compensation step may be determined according to actual system hardware conditions and application requirements. If the processing performance of the system hardware is very good, the system clock frequency can be used as the cycle frequency for cyclic execution of the drift suppression compensation step; the user can also customize the interval period for the cyclic execution of the drift suppression compensation step according to the actual application situation, such as setting the cycle The interval period is 0.5 seconds, 10 seconds, 5 minutes, 2 hours, etc., or a certain number of gas detections is set to execute a drift suppression and compensation step in a cycle; for electronic devices with very good system stability and insignificant drift For noses, it is even possible to cycle through the drift suppression compensation steps several days apart. Until the drift suppression work is terminated, the cycle stops.

The above are the specific steps of the electronic nose drift suppression method based on multiple self-organizing neural networks in the present invention.

的各个缓存阵列之中，这样有利于在此后一时刻的漂移抑制补偿过程中，通过对此后一时刻补偿增量阵列ΔW(t+1)的更新，将X_ts(t)出现的偶然检测误差再即时的补偿回来，从而避免了X_ts(t)出现的偶然检测误差在循环执行漂移抑制补偿步骤的过程中长期存在或误差累积。本发明正是通过这一措施来增强基于多重自组织神经网络的电子鼻漂移抑制方法的鲁棒性能，同时，也正是因为本发明的电子鼻漂移抑制方法具有良好的鲁棒性能，才使得本发明方法能够应用到电子鼻的气体检测过程中，将作为检测对象的气体同时作为漂移补偿训练气体样本使用，在电子鼻进行气体检测的同时即进行漂移抑制操作，不必要担心电子鼻在检测过程中因偶然因素出现检测误差而导致该误差遗留在漂移抑制流程当中影响漂移抑制性能。It is worth noting that in the step b4) of the drift suppression compensation step of the method of the present invention, the drift compensation amount of each neuron in each self-organizing neural network of the electronic nose is calculated by formula (4) or formula (5), and in step b5) by Equation (8) is an iterative update increment for each cache array of the sample cache matrix, and the two are not equal in value. The reason for this treatment is that if some special factors appear in the drift suppression and compensation process at a certain moment, the electric signal output array X _ts (t) obtained in step b2) has occasional detection errors, but due to the step b4) through the formula (4) or formula (5) has reduced X _ts (t) or ΔW(t) by a or b times for the drift compensation of each neuron in each self-organizing neural network of the electronic nose, so X _ts ( t) The occasional detection error that appears will not be completely and directly reflected in each self-organizing neural network neuron of the electronic nose after drift compensation at this moment; at the same time, in step b5) through the formula (8) for each sample buffer matrix The iterative update increment of the cache array is ΔW(t), that is, the occasional detection error in X _ts (t) is directly reflected in the sample cache matrix

Among the various buffer arrays, it is beneficial to update the compensation increment array ΔW(t+1) at the next moment during the drift suppression compensation process at the next moment, and the accidental detection error that occurs in X _ts (t) Then it is compensated back immediately, thereby avoiding the long-term existence or error accumulation of the occasional detection error in X _ts (t) in the process of cyclic execution of the drift suppression compensation step. The present invention just enhances the robust performance of the electronic nose drift suppression method based on multiple self-organizing neural networks through this measure, and at the same time, it is precisely because the electronic nose drift suppression method of the present invention has good robust performance that it makes The method of the present invention can be applied to the gas detection process of the electronic nose, and the gas as the detection object is used as a drift compensation training gas sample at the same time, and the drift suppression operation is performed while the electronic nose is performing gas detection, and there is no need to worry about the electronic nose being detected. The detection error occurs due to accidental factors in the process, and the error is left in the drift suppression process and affects the drift suppression performance.

中各个缓存阵列的取值的迭代更新中都取了运算参数的归一化值，这是因为进行漂移抑制补偿过程中所采用的气体样本可能浓度各有不同，气体传感器阵列对于不同浓度的同种气体样本其检测电信号输出阵列值也会存在差异，而归一化能够消除气体样本浓度引起的检测差异，并且保留气体传感器阵列对气体样本的敏感体征，因此可以避免电子鼻在漂移抑制补偿过程中将气体样本浓度的不同将之误判为检测漂移而错误的进行漂移补偿，同时也相应地增强了电子鼻识别不同气体的抗干扰能力。取归一化值的具体运算方式有很多，不同的本领域技术人员在不同的应用场合，可以根据其技术习惯或实际情况的需要而采用不同的归一化算法。On the other hand, in step b4) of the method of the present invention, in the drift compensation of each neuron in each self-organizing neural network and in step b5) for the sample cache matrix

In the iterative update of the value of each buffer array in the method, the normalized value of the operation parameter is taken. This is because the gas samples used in the process of drift suppression and compensation may have different concentrations. There will also be differences in the detection electrical signal output array value of different gas samples, and normalization can eliminate the detection difference caused by the gas sample concentration, and retain the sensitive signs of the gas sensor array to the gas sample, so it can avoid the drift suppression compensation of the electronic nose. In the process, the difference in the concentration of the gas sample is misjudged as the detection drift and the drift compensation is wrongly performed. At the same time, the anti-interference ability of the electronic nose to identify different gases is correspondingly enhanced. There are many specific calculation methods for obtaining the normalized value, and different persons skilled in the art may use different normalization algorithms according to their technical habits or actual needs in different application occasions.

3. Experimental effect verification.

The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

In order to verify the actual effect of the method of the present invention, a self-built electronic nose test platform is used for experiments. The self-built electronic nose test platform is shown in Figure 2. The electronic nose used in the experiment is composed of a gas sensor array 4 and a computer 5; the gas sensor array used in this experiment is composed of four metal oxide gas sensors, namely GSBT11, TGS2620, TGS2602 and TGS2201, of which TGS2201 integrates Two gas sensors, so it is equivalent to that the gas sensor array contains 5 gas sensors (corresponding to the method of the present invention, which is equivalent to getting i=5); the computer acts as two parts of the signal preprocessing unit and the pattern recognition unit of the electronic nose , used to store each neuron of multiple self-organizing neural networks, and perform data acquisition and detection and identification work. In this test, the multiple self-organizing neural networks stored by the computer are divided into 3 self-organizing neural networks (corresponding to the present invention) In the method, it is equivalent to taking K=3), which are respectively used to match clean air (gas sample as baseline data), carbon monoxide (CO) and formaldehyde (CH ₂ 0) three types of gases, and the data acquisition module of the computer adopts 12-bit string The AD chip TLC2543 is used, and the time interval of each acquisition is set to about 1 second. The collected data is transmitted to the computer through the serial port for detection and identification processing. The detection and identification processing software adopts MATLAB7.1. In the experiment, the self-built electronic nose test platform shown in Figure 2 was used to prepare the gas samples to be tested in the gas collection bag 1, and then pump the gas into the test chamber 3 with the intake pump 2; the electronic nose The gas sensor array 4 is placed in the test chamber 3 to detect the gas, and its signal output array is collected and identified by the computer 5; finally, the gas in the test chamber 3 is discharged by the gas outlet pump 7, and the gas output is adjusted by the valve 6. control.

In this experiment, using the self-built electronic nose test platform mentioned above, the existing MSOM drift compensation training method and the method of the present invention were used to carry out 50 gas sample tests respectively within one month. The 50 tests were aimed at The test gas samples were tested in sequence: 10 times for carbon monoxide (CO), 15 times for formaldehyde (CH ₂ O), and 25 times for carbon monoxide (CO). For each test, clean air was used to collect baseline data 50 times. 1, and then use the test gas sample to collect 100 test data, that is, each test collects 150 electrical signal output array data of the gas sensor matrix; each electrical signal output array contains 5 electrical signal output values, that is, the electronic nose gas The electrical signal output values of the 5 gas sensors in the sensor array. Therefore, a total of 7,500 electrical signal output array data were obtained in this experiment, the 1st to 1500th are electrical signal output array data for CO gas samples (including 500 baseline data), and the 1501st to 3750th are for CH ₂ O The electrical signal output array data of the gas sample (including 750 baseline data), and the 3751st to 7500th are the electrical signal output array data (including 1250 baseline data) of the CO gas sample. In the test that adopts the inventive method, the compensation proportional coefficient a=0.3 of taking, the compensation increment coefficient b=0.2, the normalized computing formula that adopts is:

表示取归一化值；F表示包含i个元素的任意阵列，i表示电子鼻气体传感器阵列中气体传感器的个数；f₁，f₂，…，f_i分别表示所述阵列F包含的i个元素。针对本次试验而言，所用电子鼻的多重自组织神经网络中共包含3个自组织神经网络，因此取i＝3。Among them, the symbol

Indicates the normalized value; F represents any array containing i elements, and i represents the number of gas sensors in the electronic nose gas sensor array; f ₁ , f ₂ ,..., f _i represent the i elements. For this experiment, the multiple self-organizing neural network of the electronic nose contains 3 self-organizing neural networks, so i=3 is taken.

Finally, the detection and recognition results obtained by using the existing MSOM drift compensation training method and the method of the present invention to perform the above-mentioned 50 gas sample tests respectively are shown in Table 1:

Table 1

Among the 50 tests carried out using the existing MSOM drift compensation training method, the identification of the first to 3750 test data (the previous 10 test data for CO gas samples and 15 test data for CH ₂ O gas samples) The effect is good, and the recognition accuracy rate is 100%. However, starting from the 3751st test data (that is, the last 25 test data for CO gas samples), the electronic nose can only correctly identify the baseline data, while the test for CO gas samples does not All the identifications are wrong, so the identification accuracy rate is only 33%, which is equivalent to 0% identification accuracy rate for CO gas samples. The reason is that during the 1501st to 3750th test data period of the CH ₂ O gas sample test, the MSOM drift compensation training method did not perform drift compensation for the neurons in the self-organizing neural network that matched the CO gas sample , however, the gas sensor matrix is highly correlated with the sensitive characteristics of CO gas and CH ₂ O gas, and the aging of the gas sensor caused during the test for CH ₂ O gas samples will also correlately affect its detection of CO gas , so that the electrical signal output array value of the gas sensor array detecting the CO gas sample drifted, resulting in the last 25 identification errors for the CO gas sample test. This also proves the limitation analysis of the present invention to the existing MSOM drift compensation training method.

In the 50 tests carried out by the method of the present invention, since the method of the present invention has the characteristics of drift compensation for sensitive correlation neurons, the electronic nose can still maintain good recognition performance after drift suppression compensation, so the electronic nose is at the end. The recognition rate of 25 CO gas sample tests can still reach 100%. It can be seen from the above comparison that, compared with the existing MSOM drift compensation training method, the electronic nose drift suppression method based on the multiple self-organizing neural network of the present invention has significantly increased the implementation efficiency and recognition performance of the drift compensation of the electronic nose. improve.

Finally, it is noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be carried out Modifications or equivalent replacements without departing from the spirit and scope of the technical solution of the present invention shall be covered by the claims of the present invention.

Claims (2)

1. An electronic nose drift suppression method based on a multiple self-organizing neural network is characterized by comprising the following steps:

A) an initialization step, which specifically comprises:

a1) establishing a sample cache 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) is a cache array containing i elements, K belongs to {1,2, …, K }, K represents the number of the self-organizing neural networks of the electronic nose, i represents the number of the gas sensors in the gas sensor array of the electronic nose, and t represents the time;

a2) initialization time t = 0;

a3) at the current time, the sample buffer matrix

The values of each cache array in the cache memory are as follows:

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

wherein,

represents the mth neuron in the kth self-organizing neural network of the electronic nose at the current moment,

then the neuron is represented

I characteristic electrical signal values of M_kRepresenting the number of neurons in the kth self-organizing neural network of the electronic nose;

B) a drift suppression compensation step; the method specifically comprises the following steps:

b1) adding 1 at the time t;

b2) electric signal output array X for acquiring electronic nose gas sensor array at current moment_ts(t)：

X_ts(t)=[x_ts,1(t),x_ts,2(t),…,x_ts,i(t)]；

Wherein x is_ts,1(t),x_ts,2(t),…,x_ts,i(t) represents the electrical signal output values of i gas sensors in the electronic nose gas sensor array at the current moment;

b3) calculating the matching winning self-organizing neural network sequence number k_1stAnd matching the number k of the winning self-organizing neural network_2nd：

Wherein,

represents the mth neuron in the kth self-organizing neural network of the electronic nose at the previous time (t-1); symbol [ 2 ]]]Expressed as normalized value, [ [ X ]_ts(t)]]And are and

respectively representing the output arrays X for taking the electric signals_ts(t) normalizing value and taking said neuron

A normalized value of (d);

is represented by [ [ X ]_ts(t)]]And

the Euclidean distance of;

express get

And

is in all K e {1,2, …, K } and M e {1,2, …, M ∈_kThe minimum value of the cases (c) is,

express get

And

is in all K e {1,2, …, K } and M e {1,2, …, M ∈_kIn the case of only more than

The secondary minimum of (d);

b4) if k is_1st=k_2ndAnd performing drift compensation on each neuron in each self-organizing neural network of the electronic nose according to the following formula:

if k is_1st≠k_2ndAnd then carrying out drift compensation on each neuron in each self-organizing neural network of the electronic nose according to the following formula:

wherein,

representing the mth neuron in the kth self-organizing neural network of the electronic nose at the current moment; Δ W (t) represents the compensation increment array at the current time, and $\mathrm{\&Delta;W}\left(t\right)=\left[\right[X_{\mathrm{ts}}\left(t\right)\left]\right]-\left[\right[X^{k_{1\mathrm{st}}}(t-1)\left]\right],$

indicating the sample buffer matrix at the previous time (t-1)

Middle (k) th_1stA plurality of cache memory arrays, each of which is provided with a cache memory array,

representing fetching of the cache array

A normalized value of (d); a is a compensation proportionality coefficient with a value range of 0<a is less than or equal to 0.5; b is a compensation increment coefficient with a value range of 0<b≤0.5；

b5) Buffering the samples with a matrix according to

And carrying out iterative update on the values of each cache array:

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

wherein, X^k(t) sample buffer matrix at current time

The kth cache array of (1); x^k(t-1) represents the sample buffer matrix at the previous time instant (t-1)The kth cache array of [ [ X ]^k(t-1)]]Representing fetching of the cache array X^kA normalized value of (t-1);

C) and B) circularly executing the step B) until the electronic nose stops the drift suppression work.

2. The method for suppressing drift of an electronic nose based on multiple self-organizing neural networks as claimed in claim 1, wherein the symbol [ ] ] represents a specific operation formula taking a normalized value as follows:

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

where F represents an arbitrary array containing i elements, and F₁,f₂,…,f_iRepresenting the i elements that the array F contains.

Images