KR101656253B1 - Apparatus and method for constructing combine feature vector for gas classification - Google Patents

Abstract

The present invention relates to an apparatus for constructing a combination feature vector for gas classification, and a method thereof. The apparatus creates a combination characteristic by extracting characteristics of the entire area by using samples of the entire area in a total measurement section, extracting an area characteristic by using an area sample of a stabilization section, an exposure section and a purge section, and extracting a characteristic of an area and the entire area having enough information for discernment. According to the present invention, the apparatus for constructing a combination feature vector for gas classification including a sensor array constituted by a plurality of channels and a plurality of sections comprises: a gas measurement part which measures reaction by a gas flowing into a chamber to generate gas data; a determination characteristic extraction part which extracts a determination characteristic based on the gas data generated in the gas measurement part; and a combination characteristic vector generation part which generates a combination characteristic vector based on the gas data generated in the gas measurement part and the determination characteristic extracted in the determination characteristic extraction part.

Description

[0001] APPARATUS AND METHOD FOR CONSTRUCTION COMBINE FEATURE VECTOR FOR GAS CLASSIFICATION [0002]

The present invention relates to an apparatus and method for generating a joint feature vector for gas classification, and more particularly to an apparatus and method for generating a joint feature vector for gas classification that generates a joint feature vector, ≪ / RTI >

ELECTRONIC NOSE system is an electronic system that classifies a specific gas type using sensors. Since the olfactory function of the human body is easily fatigued, the kind of odor that can be taken without successively smelling other odor is limited, and the odor can be continuously monitored. Also, the electronic nose The system was developed.

The earliest electronic nose system uses a calorimetric sensor to display measurements of the gas in the form of an array of colors.

Conventional electronic nose systems include precision instruments such as a gas chromatography system and a machine intelligence coupled mass spectrometer, and require complicated analysis processes at various stages There is a restriction in the use environment.

Accordingly, the conventional electronic nose system has a limited application field such as quality control of the food and beverage field.

As electrochemical sensors and hardware and software technologies have developed recently, electronic nose systems have become intelligent and portable as compared to conventional electronic nose systems, , Environmental protection, and the cosmetics industry.

Generally, an electronic nose system consists of a sensor array and a computing system composed of several channels.

Each channel of the sensor has different characteristics and reacts differently depending on the kind of gas. At this time, the response of the sensor is converted into numerical data in vector form by the electronic interface.

The computing system extracts features that are useful for gas classification from numerical data transformed by the electronic interface. The computing system classifies the type of gas using a classifier.

Therefore, the sensitivity of the sensor array, the feature extraction, and the performance of the classifier are very important factors for determining the gas type analysis accuracy (i.e., performance) of the electronic nose system.

Of the various types of sensor arrays used in electronic nose systems, conductivity sensors for sensing materials include conductive polymer composites, intrinsically conducting polymers, metal oxides, and the like. It is mainly used.

Conductive polymer composites have the disadvantages of sensor drift, sensor life limitation, sensitivity to temperature and humidity, and the use of polymer materials compared to other materials, low cost, stable operation at room temperature, low power consumption And it is widely used as a gas sensor.

The measurement of the sensor consists of three stages: stabilization, exposure, and purge. Each channel of the sensor shows different responses depending on the type of gas. Lt; / RTI >

In the computing system, various kinds of pattern recognition techniques are used to discriminate the type of gas data inputted through the sensor array. The process of classifying the gas is divided into steps of extracting features useful for classification from the gas and preliminarily classifying the classifier using the extracted features.

The data samples acquired through the sensors in the electronic nose system are high-dimensional data composed of values measured for each channel and sampling point. Therefore, various kinds of dimension reduction methods are used to effectively reduce the amount of data processing while increasing the accuracy of gas classification.

PCA (Principle Component Analysis), which is a typical resource reduction method, projects data samples to a low-dimensional feature space that maximizes dispersion of data samples and performs classification. Other varieties of PCA, such as CC-PCA and CCCPCA, are used to classify gases.

Among resource reduction methods, Linear Discriminant Analysis (LDA) is a way to focus on finding the optimal linear discriminant function. LDA is generally better than PCA for classification problems. The LDA uses the class information of training data to narrow the distribution within the class and construct a feature space that broadens the distance between the classes. In addition, CLDA, which developed the LDA suitable for high dimensional data with large correlation of input variables, is used to classify the gas.

The conventional methods described above express the entire measurement values from the stabilization step to the purge step with respect to the gas sample as a single vector and construct a feature space through a statistical analysis process. However, it is effective to use the local information together with the global information because the measurement value of the sensor includes different information for each measurement interval.

Korean Patent Laid-Open No. 10-2003-0093682 (entitled "Small Electronic Nose System and Method for Detecting and Detecting Environmental Hazardous Gasses and Foreign Objects Included in Automobile Gasoline)

SUMMARY OF THE INVENTION The present invention has been proposed in view of the above circumstances, and it is an object of the present invention to extract global features by using global samples of the entire measurement interval, extract local features using local samples of the stabilization interval, It is an object of the present invention to provide an apparatus and method for generating a combined feature vector for classifying a gas in which a large number of global features and regional features are extracted to generate combined features.

According to an aspect of the present invention, there is provided an apparatus for generating a joint feature vector for gas classification according to an embodiment of the present invention. The apparatus includes a sensor array including a plurality of channels and a plurality of sections, A gas measurement unit for measuring and generating gas data; A discrimination feature extracting unit for extracting discrimination characteristics based on the gas data generated by the gas measuring unit; And a combined feature vector generation unit for generating a combined feature vector based on the gas data generated by the gas measurement unit and the discrimination feature extracted by the discrimination feature extraction unit.

The gas measuring unit is composed of a micromechanical sensor array composed of 16 channels, and the 16 channels are divided into three sections in which the cross electrode, the microheater and the polymer complex are respectively disposed, and the 16 channels are composed of different polymer complexes .

The gas measuring unit divides the entire measurement period into the stabilization period, the exposure period and the purge period, and measures the reaction by the gas at the set time interval set differently for each period.

The discrimination feature extraction unit performs an LDA to calculate a projection matrix composed of a plurality of projection vectors based on a dispersion matrix and an intra-class dispersion matrix satisfying the objective function, and performs a PCA to reduce the dimension of the plurality of learning data, Defines a projection matrix, defines a projection matrix of PCA and LDA defined based on a projection matrix and an integrated projection matrix, and generates a feature vector composed of a plurality of determination features based on the eigenvalues of the projection matrixes of PCA and LDA .

The discrimination feature extracting unit extracts

Using _{(W PCA = argmax W | W} T S T S | _{a, WLDA = argmax W {| W} T W PCA T S B W PCA W | / | | W T W PCA T S W W PCA W}) Defines the projection matrix of PCA and LDA.

The joint feature generation unit classifies the gas data measured by the gas measurement unit into a measurement value of the stabilization period, a measurement value of the exposure interval, and a measurement value of the fuzzy interval, calculates global characteristics based on the measured values of the entire interval, The regional characteristics of each section are calculated based on the classified measurements.

The joint feature generation unit constructs a basic feature set composed of a plurality of elements based on the calculated local features and global features, sets a weight vector of the regional feature and the global feature included in the basic feature set using the ReliefF method, And construct a combined feature vector based on the vector.

The joint feature generation unit calculates

(A _j is the weight of the jth feature, Ykj is the jth basic feature of the randomly selected kth sample, Hrj is the random number of samples selected from the samples having the same class for the randomly selected kth sample, s j-th basic characteristics of the sample, Mrj is the j-th basic feature of the in the sample belonging to the different classes for a selected randomly k th sample distance nearest s samples, N is the number of samples, P (C _l) is a class C _l And P (CYk) is the prior probability of the class to which Yk belongs) to calculate the weight vector of the local features and the global features included in the basic feature set.

According to another aspect of the present invention, there is provided a method for generating a combined feature vector for gas classification, the method comprising: generating a combined feature vector using a combined feature vector generation apparatus for gas classification, Measuring the reaction by the gas sensor to generate gas data; Extracting discrimination characteristics based on the generated gas data; And generating a combined feature vector based on the generated gas data and the extracted discrimination feature.

In the step of generating the gas data, the microchip sensor array is composed of 16 channels. The 16 channels are divided into three sections where the cross electrodes, the microheaters and the polymer complexes are disposed, respectively. Measure the response to gas entering the chamber through the complex.

The step of generating the gas data may include measuring the response of the sensor to the gas at a set time interval during the set time in the stabilization period; Measuring the response of the sensor to the gas at the set time interval during the set time in the exposure section; And measuring the response of the sensor to the gas at a set time interval during the set time in the purge interval, wherein the set time is set differently for each section.

The step of extracting the discriminating feature may comprise the steps of: defining a intra-class variance matrix and a inter-class variance matrix; Calculating a projection matrix composed of projection vectors based on a dispersion matrix in the class and a dispersion matrix between classes that satisfy the objective function by performing the LDA;

Performing a PCA to reduce a dimension of a plurality of learning data to define an overall projection matrix; Defining a projection matrix of PCA and LDA based on a projection matrix and an integrated projection matrix; And extracting the discriminant feature based on the eigenvalues of the projection matrixes of the PCA and the LDA.

In the step of defining the projection matrixes of PCA and LDA,

Using _{(W PCA = argmax W | W} T S T S | _{a, WLDA = argmax W {| W} T W PCA T S B W PCA W | / | | W T W PCA T S W W PCA W}) Defines the projection matrix of PCA and LDA.

The step of generating the combined feature vector comprises: classifying the measured gas data into a measured value of the stabilization interval, a measured value of the exposure interval, and a measured value of the fuzzy interval; Calculating global features and regional features based on measured values classified by measurement intervals; Constructing a basic feature set based on the calculated global feature and local feature; Setting a weight vector of a local feature and a global feature included in a basic feature set using a ReliefF method; And constructing a combined feature vector based on the weight vector.

In the step of setting the weight vector,

(A _j is the weight of the jth feature, Ykj is the jth basic feature of the randomly selected kth sample, Hrj is the random number of samples selected from the samples having the same class for the randomly selected kth sample, s j-th basic characteristics of the sample, Mrj is the j-th basic feature of the in the sample belonging to the different classes for a selected randomly k th sample distance nearest s samples, N is the number of samples, P (C _l) is a class C _l And P (CYk) is the prior probability of the class to which Yk belongs) to calculate the weight vector of the local features and the global features included in the basic feature set.

According to the present invention, an apparatus and method for generating a combined feature vector for classifying gases is characterized in that only the features with high discrimination power are selected from among the global features and the regional features of each section to remove the unnecessary information for each section, It can be used as an input of a classifier, so that it is possible to classify a robust gas even when noise occurs in the process of measuring gas data.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 and FIG. 2 illustrate an apparatus for generating a combined feature vector for gas classification according to an embodiment of the present invention. FIG.
3 is a view for explaining the gas measurement unit of FIG. 2;
4 is a flowchart illustrating a method for generating a combined feature vector for gas classification according to an embodiment of the present invention.
FIG. 5 is a flow chart for explaining the gas measurement step of FIG. 4;
FIG. 6 is a flowchart for explaining the discrimination feature extraction step of FIG. 4;
FIG. 7 is a flowchart for explaining the joint feature vector generation step of FIG. 4;
FIG. 8 to FIG. 16 are diagrams for explaining a result of gas classification experiment based on combined feature vectors generated through an apparatus and method for generating a combined feature vector for gas classification according to an embodiment of the present invention; FIG.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings in order to facilitate a person skilled in the art to easily carry out the technical idea of the present invention. . In the drawings, the same reference numerals are used to designate the same or similar components throughout the drawings. In the following description of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a combined feature vector generation apparatus for gas classification according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings. FIG. 1 and FIG. 2 are views for explaining a combined feature vector generation apparatus for gas classification according to an embodiment of the present invention. FIG. 3 is a view for explaining the gas measuring unit of FIG. 2. FIG.

A coupled feature vector generation apparatus for gas classification (hereinafter referred to as a coupled feature vector generation apparatus) generates a reference factor input to an analyzer (classifier) for gas classification. That is, the combined feature vector generation apparatus generates a combined feature vector as a reference factor in order to utilize all the advantages of the global feature and the regional feature of the measured value as a whole in the measurement period for gas classification.

As shown in FIG. 1, the combined feature vector generating device extracts global characteristic (s) from measured values (i.e., global samples) of the entire measurement interval using a principle component analysis (PCA) and a linear discriminant analysis And extracts the local features from the measured values of each measurement period (i.e., the local samples).

At this time, the extracted global features and local features include information that is not useful to the gas classification, but rather interferes with the classification of the gas due to the influence of various kinds of sides or noise.

Accordingly, the combined feature vector generating apparatus selects only the features suitable for the gas classification among the extracted features (i.e., the global features and the regional features) and uses them in the final recognition step.

For this purpose, the joint feature vector generation device measures class discrimination ability for each feature using the ReliefF method, which is one of representative feature selection methods. The combined feature vector generation device combines the best global features and local features for gas classification based on the measured class discrimination capabilities.

2, the combined feature vector generation apparatus 100 includes a gas measurement unit 120, a discrimination feature extraction unit 140, and a combined feature vector generation unit 160.

The gas measuring unit 120 is disposed inside the chamber 200 to measure the gas flowing into the chamber 200. To this end, the gas measuring unit 120 is constituted by a sensor array composed of a plurality of channels and a plurality of sections. That is, the gas measuring unit 120 is composed of a micromachined sensor array composed of 16 channels. In this case, each channel is divided into three sections, and a carbon-black (CB) polymer composites having an interdigitated electrode, a micro heater, and a machined membrane ) Sensor is disposed. Here, as shown in FIG. 3, the carbon black polymer composite sensor may be composed of a polymer composite of different materials for each channel.

The gas measuring unit 120 measures the reaction by the chemical bonding between the polymer complex disposed in each channel and the gas to be introduced. At this time, the gas measuring unit 120 measures the reaction by the chemical bonding of the gas and the polymer complex at a set time interval (for example, about 0.1 second interval) for a set time (for example, about 200 seconds).

The gas measuring unit 120 divides the entire measurement period into a stabilization period, an exposure period, and a purge period. The gas measuring unit 120 sets the set time for each section, and measures the reaction for each set time during the set time. That is, the gas measuring unit 120 sets a set time of about 30 seconds in a stabilization period in which a resistance signal is stabilized after the sensor array is placed in the chamber 200, measures the response at intervals of about 0.1 second do. The gas measuring unit 120 measures a reaction time at an interval of about 0.1 second by setting a set time of about 60 seconds in an exposure period, which is a period during which the target gas is introduced into the chamber 200 after stabilization of the sensor array. The gas measuring unit 120 sets a set time of about 110 seconds in a purging period which is an interval for discharging the target gas in the chamber 200 to the outside, and measures the reaction at intervals of about 0.1 second.

The discrimination feature extracting unit 140 extracts discrimination characteristics based on the gas data measured by the gas measuring unit 120. That is, the discrimination feature extracting unit 140 extracts discrimination characteristics suitable for gas classification from the gas data using LDA (Linear Discriminant Analysis), which is a type of dimension reduction method.

(k=1, …, N)가 주어졌을 때, 판별 특징 추출부(140)는 LDA를 이용하여 동일 클래스에 속한 학습 데이터 샘플들을 모으면서 동시에 클래스의 평균들 사이의 거리를 멀어지게 하는 특징 공간을 구성한다.N n-dimensional learning data samples with C classes

(k = 1, ..., N), the discrimination feature extracting unit 140 collects the learning data samples belonging to the same class using the LDA, and at the same time, .

To this end, the discrimination feature extracting unit 140 defines the within-class scatter matrix S _W as shown in Equation (1) and calculates a between-class scatter matrix S _B by the following mathematical expression Is defined as Equation (2).

Here, X _k denotes the m-th sample belongs to the class C _i and, N is the total number of sample, C is the number of classes, μ is the average (mean) of the entire sample, μ _i is the sample belongs to the class C _i .

The discrimination feature extractor 140 uses the LDA to calculate projection vectors W _l ^L , l = 1, ..., n 'from S _W and S _B satisfying the objective function defined in Equation (3) A projection matrix W consisting of _LDA = [W ₁ ^L , ... , W _n ^L ].

Here, W _l ^L should satisfy ^L = λ _{l l} W _B S S _W W _l ^L which is determined by calculating a specific vector (eigen-vector) of S _w S _b ^-1.

)와 같이 고차원 데이터인 경우, n이 S_W의 rank (N-C)보다 커진다. 그에 따라, S_W가 항상 단일화(singular)되는 SSS(Small Sample Size) 문제가 발생한다.At this time, the learning data is the gas data (in this embodiment,

), N is larger than the rank (NC) of S _W. As a result, a Small Sample Size (SSS) problem occurs in which S _W is always singular.

Accordingly, the discriminating feature extracting unit 140 performs PCA before the LDA is performed to reduce the dimension of the learning data to m? NC. The discrimination feature extracting unit 140 defines a total scatter matrix S _{T according} to the following equation (4).

Accordingly, the projection matrix of the PCA and the LDA is defined by the following equation (5).

Here, W = argmax _{PCA W} | a W _^T S ^T S | a, WLDA _W = argmax {| ^T W ^T W _{PCA PCA} S _B W W | / | | W ^{T T} S _W W W _{PCA PCA} W}.

If n < C-1 is selected in order of increasing eigenvalues among the projection vectors W _l ^PL constituting the projection matrix of the PCA and the LDA, the gas data sample X _k is divided into n 'number of discriminant features Dimensional feature vector is expressed as Equation (6) below.

The combined feature vector generator 160 separates the measurement values corresponding to the stabilization period, the exposure period, and the purging period from the global learning data sample Xk including all of the measured values measured in the entire measurement period. That is, the combined feature vector generation unit 160 outputs the measured values (X _k ^ST ∈ R ^{3000 × 1} ) measured in the stabilization period, the measured value (X _k ^SX ∈ R ^{6000 × 1} ) measured in the exposure period, And the measured values measured in the purge interval (X _k ^PU ∈ R ^{11000 × 1} ).

The combined feature vector generation unit 160 calculates the global feature and the regional feature based on the measured values classified by intervals. At this time, the joint feature vector generation unit 160 calculates the global feature and the regional feature using the measured values classified by the interval and the above-described Equation (6). In this case, the global feature (y _k ^tot ), the local feature (y _k ^st ) of the stabilization section, the local feature (y _k ^ex ) of the exposure section and the regional feature (y _k ^pu ) do.

The combined feature vector generation unit 160 constructs a basic feature set having 4 (C-1) elements by using the calculated global feature and local feature. At this time, the basic feature set is constructed as shown in Equation (8).

The combined feature vector generator 160 selects the representative feature through the ReliefF method to measure how useful each basic feature is to classify the pattern. Here, the ReliefF method basically distinguishes between the useful features and the non-useful features depending on how well the values of the respective features can be distinguished from each other.

To this end, the joint feature vector generator 160 defines a weight vector A, which is a value for each feature. Here, the weight vector A is defined by the following equation (9).

We define s samples with the closest distance in the feature space among the samples having the same class for the randomly selected kth sample (Xk) as "HIT" (H _r , r = 1, ..., s) , And defines 's Miss' (M _r , r = 1, ..., s) as the s closest samples among the samples belonging to other classes.

In this case, it means that for a j-th feature Y _j X _k and H _r is, if having a different value, characterized in Y _j are separated from each other a sample X _k, and H _r belonging to the same class. In this case, since the class discrimination becomes difficult, the joint feature vector generator 160 reduces the value of the _jth element A _j of the weight vector.

On the other hand, if the features Y _{j k} X _r and M having a different value, it means that the feature Y _j difficult to remove a sample X _k with M _r belonging to different classes. Accordingly, the joint feature vector generator 160 increases the value of the j-th element A _j of the weight vector.

When the initial value of A _j is defined as 0, the joint feature vector generator 160 calculates the weight A _j of the feature Y _j by the sample X _k through the following process.

1) Select any sample X _k .

2) Find the s H _r of from the main feature space.

3) Find s s M _r (C _l ) in class C _Yk of X _k and other classes C _l (l = 1, ..., C-1)

4) The weight Aj is calculated through the following equation (10).

Where A _j is the weight of the jth feature, Ykj is the jth fundamental feature of the randomly selected kth sample, Hrj is the randomly selected sample of the same class for the randomly selected sample, s J is the j-th basic characteristic of the samples, Mrj is the j-th basic characteristic of the s closest samples among the samples belonging to different classes for randomly selected kth samples, N is the number of samples, P ( _Cl ) _The prior probability of _l , P (CYk) is the prior probability of the class to which Yk belongs. The final coupling used for the combined feature vector generating unit 160 is the above process N samples _{X k (k = 1,2, ...} , N) after the gas classification, the value of A _j has large features repeated for And constructs the feature vector (y ^CF ). Here, the final configured joint feature vector y ^CF is input to the classifier 300 for classifying the gas, and is used as a reference factor for classifying the gas.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A method of generating a joint feature vector for gas classification according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings. 4 is a flowchart illustrating a method of generating a joint feature vector for gas classification according to an embodiment of the present invention. FIG. 5 is a flow chart for explaining the gas measuring step of FIG. 4, FIG. 6 is a flowchart for explaining the discriminating feature extracting step of FIG. 4, and FIG. 7 is a flowchart for explaining the combining characteristic vector generating step of FIG. .

A combined feature vector generating apparatus 100 (hereinafter referred to as a combined feature vector generating apparatus 100) for gas classification measures the gas flowing into the chamber 200 (S100). That is, the coupled feature vector generation apparatus 100 measures the reaction by the chemical coupling between the polymer complex constituting the sensor array composed of a plurality of channels and the gas flowing into the chamber 200. At this time, the coupled feature vector generation apparatus 100 measures the response due to the chemical bonding of the gas and the polymer complex at a set time interval (about 0.1 second) for a set time (about 200 seconds).

At this time, as shown in FIG. 5, the coupled feature vector generation apparatus 100 measures the gas for the set time in the stabilization period (S120). That is, the coupled feature vector generation apparatus 100 sets a set time of about 30 seconds in the stabilization period, which is a period in which the resistance signal is stabilized after the sensor array is placed in the chamber 200, .

The combined feature vector generation apparatus 100 measures the gas for a set time in the exposure section (S140). That is, the combined feature vector generation apparatus 100 sets a set time of about 60 seconds in the exposure period, which is a period in which the target gas is introduced into the chamber 200 after stabilization of the sensor array, The reaction is measured.

The coupled feature vector generation apparatus 100 measures the gas during the set time in the purge interval (S160). That is, the combined feature vector generation apparatus 100 sets a set time of about 110 seconds in a purge section, which is a section for discharging a target gas in the chamber 200 to the outside, and measures the response at a set time interval of about 0.1 second do.

The combined feature vector generation apparatus 100 extracts the discrimination characteristic based on the gas measurement result (S200). That is, the coupled feature vector generation apparatus 100 extracts the discrimination characteristic based on the measurement value (gas data) which is the gas measurement result in step S100. At this time, the coupled feature vector generation apparatus 100 extracts discrimination characteristics suitable for gas classification from gas data using LDA (Linear Discriminant Analysis) and PCA (Principle Component Analysis), which are one of the dimensional reduction methods. This will be described with reference to FIG. 6 attached hereto.

(k=1, …, N)가 주어졌을 때, 판별 특징 추출부(140)는 LDA를 이용하여 동일 클래스에 속한 학습 데이터 샘플들을 모으면서 동시에 클래스의 평균들 사이의 거리를 멀어지게 하는 특징 공간을 구성하기 위해서, 결합 특징 벡터 생성 장치(100)는 클래스 내 분산 행렬 S_W 및 클래스 간 분산 행렬 S_B를 정의한다.The coupled feature vector generation apparatus 100 defines a distributed matrix within a class and a distributed matrix between classes (S210). That is, N pieces of n-dimensional learning data samples having C classes

(k = 1, ..., N), the discrimination feature extracting unit 140 collects the learning data samples belonging to the same class using the LDA, and at the same time, The combined feature vector generation apparatus 100 defines an in-class variance matrix S _W and an inter-class variance matrix S _B.

The combined feature vector generation apparatus 100 calculates a projection matrix using the LDA (S230). That is, the combined feature vector generation apparatus 100 is constructed from a projection vector W _l ^L , l = 1, ..., n 'from S _W and S _B satisfying an objective function (see Equation 3) The projection matrix W _LDA = [W ₁ ^L , ... , W _n ^L ].

)와 같이 고차원 데이터인 경우, n이 S_W의 rank (N-C)보다 커져 S_W가 항상 단일화(singular)되는 SSS(Small Sample Size) 문제점을 방지하기 위해서, 결합 특징 벡터 생성 장치(100)는 LDA의 수행전에 PCA를 수행하여 학습 데이터의 차원을 m≤N-C로 줄인 종합 투영 행렬 S_T를 정의한다.The combined feature vector generation apparatus 100 calculates a projection matrix using the PCA (S250). That is, when the learning data is the gas data (in this embodiment,

), The combined feature vector generation apparatus 100 may use the LDA (Small Sample Size) in order to prevent the SSS (Small Sample Size) problem where n is greater than the rank (NC) of S _{W and} the S _w is always singular. The PCA is performed to define a general projection matrix S _T that reduces the dimension of the learning data to m? NC.

The combined feature vector generation apparatus 100 defines the LDA and PCA projection matrices based on the previously calculated projection matrices (S270). That is, the combined feature vector generation apparatus 100 defines the PCA and _LDA projection matrices using the projection matrix (W _LDA ) calculated using the _LDA and the integrated projection matrix (S _T , W _PCA ) calculated using the PCA.

The combined feature vector generation apparatus 100 extracts discrimination features based on the eigenvalues of the projection vectors constituting the LDA and PCA projection matrices (S290). That is, when selecting n? C-1 in the order of larger eigenvalues among the projection vectors W _l ^PL constituting the projection matrix of the PCA and the LDA, the combined feature vector generating apparatus 100 generates the gas data sample X _k extracts n 'dimensional feature vectors composed of n' discriminant features.

The combined feature vector generation apparatus 100 generates a combined feature vector based on the discrimination characteristic and the gas measurement result (S300). This will be described with reference to FIG. 7 attached hereto.

The combined feature vector generation apparatus 100 classifies the measured values, which are the gas measurement results, into the respective measurement intervals (S310). That is, the combined feature vector generation apparatus 100 classifies the measured values measured in step S100 into a stabilization period, an exposure period, and a purge period. At this time, the joint feature vector generation apparatus 100 calculates the measured values (X _k ^ST ∈ R ^{3000 × 1} ) measured in the stabilization period, the measured value (X _k ^SX ∈ R ^{6000 × 1} ) measured in the exposure period, And the measured values measured in the purge interval (X _k ^PU ∈ R ^{11000 × 1} ).

The combined feature vector generation apparatus 100 calculates a global feature and a local feature based on the measurement values classified by measurement interval (S330). That is, the combined feature vector generation apparatus 100 calculates the global feature and the regional feature using the measurement values classified according to the interval and the discrimination feature extracted in step S200 in step S310. In this case, the combined feature vector generation apparatus 100 may include a global feature (y _k ^tot ) using the measured values of the entire measurement interval, a local feature (y _k ^st ) of the stabilization interval using the measured values of the stabilization interval, (Y _k ^ex ) and the local feature (y _k ^pu ) of the fuzzy section using the measured values of the fuzzy interval.

The combined feature vector generation apparatus 100 forms a basic feature set based on the calculated global feature and local feature (S350). That is, the combined feature vector generation apparatus 100 constructs a basic feature set using the global feature, the local feature of the stabilization period, the local feature of the exposure period, and the regional feature of the fuzzy interval calculated in step S330. At this time, the combined feature vector generation apparatus 100 constitutes a basic feature set having 4 (C-1) elements (see Equation (7)).

The joint feature vector generation apparatus 100 defines a weight vector for each feature (S370). That is, the joint feature vector generation apparatus 100 defines a weight vector A = {A ₁ , A ₂ , ..., A _{4 (c-1)} } ^T meaning a value for a global feature and a local feature .

We define s samples with the closest distance in the feature space among the samples having the same class for the randomly selected kth sample (Xk) as "HIT" (H _r , r = 1, ..., s) , And defines 's Miss' (M _r , r = 1, ..., s) as the s closest samples among the samples belonging to other classes.

In this case, it means that for a j-th feature Y _j X _k and H _r is, if having a different value, characterized in Y _j are separated from each other a sample X _k, and H _r belonging to the same class. In this case, since the class discrimination becomes difficult, the combined feature vector generation apparatus 100 decreases the value of the _jth element A _j of the weight vector.

On the other hand, if the features Y _{j k} X _r and M having a different value, it means that the feature Y _j difficult to remove a sample X _k with M _r belonging to different classes. Therefore, the combined feature vector generation apparatus 100 increases the value of the j-th element A _j of the weight vector.

When defining the initial value of A _j to 0, the combined feature vector generator 100 selects a random sample X _k for weights A _j of the feature Y _j by the sample X _k, find s of H _r , We find s s M _r (C _l ) in the class C _Ym ^Tot of X _k and other classes C _l (l = 1, ..., C-1) and then calculate the weight Aj through Equation (10).

The combined feature vector generation apparatus 100 constructs a final combined feature vector based on the defined weight vector (S390). Combined feature vector generator 100 includes a final binding characteristics after repeated for the step S370 of the N samples _{X k (k = 1,2, ...} , N), has a great feature values of A _j to be used for gas classification (Y ^CF ). Here, the final configured joint feature vector y ^CF is input to the classifier 300 for classifying the gas, and is used as a reference factor for classifying the gas.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a result of a gas classification experiment based on a combined feature vector generated through an apparatus and method for generating a joint feature vector for gas classification according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings. FIGS. 8 to 16 are diagrams for explaining results of gas classification experiments based on combined feature vectors generated through an apparatus and method for generating a joint feature vector for gas classification according to an embodiment of the present invention.

Classification for VOC measurement results for 8 kinds of gases to confirm that the gas can be effectively classified in the electronic nose system using the combined feature vector generated by the feature vector generation device and method for gas classification I tried. Here, the eight gases are acetone, benzene, cyclo-hexane, and ethanol. heptane, methanol, propanol, and toluene.

Twenty samples were collected for each type of gas and a total of 160 samples were used for the experiment. Each sample consisted of measurements for 2000 time points measured at 200 Hz with a sampling rate of 10 Hz per channel. The measurements of 16 channels were stored as a 2000x16 matrix, (Lexicographic ordering operator).

As shown in FIG. 8, in order to check how effective the combined feature vector generated through the combined feature vector generation apparatus and method for classifying gas is when noise occurs in the sensing data, an original data sample, Data samples with Gaussian noise with standard deviations of 2, 3, and 4 were analyzed for each data sample.

The classification performance for a total of 160 data samples was evaluated using an 8-fold cross validation strategy. First, 160 samples are mixed using a random number generator based on a seed value, and then divided into eight folds so that the same number of samples are included. One sample is divided into test data (I.e., a total of 16 samples), and the remaining 140 samples were used as training data.

In the same manner, eight experiments are performed so that all the data samples of each fold are used as test data once to obtain an average value of the classification rates.

At this time, in order to increase the statistical reliability of the 8-fold cross validation method, the classification performance was evaluated by the average value of the results obtained by repeating this process 8 times using different random seed numbers.

 All data samples were normalized to the mean and standard deviation of the training data.

When the discriminant feature is extracted using PCA and LDA, the degree of reduction to the dimension (m) using the PCA has a great influence on the classification performance of the discriminant feature.

Based on the performance evaluation of various m values, we set m (= 105) to account for 99% of the total eigenvalue of S _T. NN (nearest neighbor) rule was used as the classifier, and l _Z norm was used as the distance between two samples.

First, we compare the classification rates of local features (y _k ^st , y _k ^ex , y _k ^pu ) extracted from three kinds of local samples (X _k ^st , X _k ^ex , X _k ^pu ). As shown in Figs. 9 to 12, for the original data with no noise, each local characteristic showed a good classification rate.

In FIGS. 10 to 12, the local characteristic of the exposure section intensively entering the sensor in the gas was comparatively better than the characteristic of the other sections, but it was confirmed that the recognition rate gradually decreased as the noise characteristic became worse .

In FIGS. 13 to 16, the classification performance of the proposed combination feature (y ^CF ), PCA and LDA global features (y _k ^tot ), CC-PCA features (y _PCA ^CC ), and CC-CPCA features (y _CPCA ^CC ) Respectively. y _PCA ^CC and y _CPCA ^C , PCA was applied once more after CC-PCA and CC-CPCA were performed.

As for the noise-free data, each method showed good performance. However, as the degree of noise increases, the recognition performance decreases sharply in other methods.

On the other hand, when the combined feature vector generated through the coupled feature vector generation apparatus and method for gas classification is used, even if Gaussian noise having a standard deviation of 5 occurs, the classification rate of 90% or more is maintained have.

However, in the coupled feature vector generation apparatus and method for gas classification, the global feature and various kinds of local features are used together to discriminate various types In this paper, we propose a new classification method for the classifier. In this paper, we propose a novel classification method for the classification of a classifier.

As a result, the results of Figs. 9 to 16 can confirm that the combining feature configured using the feature selection operates more robustly to noise than the other methods.

In the electronic nose system, measurement of gas data using sensors can be divided into several stages according to the process. The sensor response at each stage has different information, and various features for classification can be extracted from it.

In the feature vector generation system and method for gas classification, three types of regional features are extracted from the sections classified as stabilization, exposure, and fuzzy, and they are used for classifying together the global features extracted from the entire region and the combining features. Based on the discriminant analysis of individual global features and regional features, we used discriminant power of each feature to select only those features with a large amount of discriminatory information and construct the combined features. The proposed method showed good classification performance for eight types of gas data measured by a sensor array composed of 16 channels. In particular, when noises occur in the sensing process, only the global characteristic, Showed better classification performance than other feature methods.

As described above, an apparatus and method for generating a combined feature vector for classifying a gas classify only the features having high discriminative power among the global features and the regional features of each interval, thereby eliminating unnecessary information for each section, It can be used as an input of a classifier, so that it is possible to classify a robust gas even when noise occurs in the process of measuring gas data.

While the present invention has been described in connection with what is presently considered to be practical exemplary embodiments, it is to be understood that the invention is not limited to the disclosed embodiments, but many variations and modifications may be made without departing from the scope of the present invention. It will be understood that the invention may be practiced.

100: Combined feature vector generation device
120: gas measuring unit 140: discrimination feature extracting unit
160: Coupled feature vector generation unit 200:
300: classifier

Claims (15)

A gas measurement unit comprising a sensor array composed of a plurality of channels and a plurality of intervals, the gas measurement unit measuring a response of the gas introduced into the chamber to generate gas data;
A discrimination feature extracting unit for extracting discrimination characteristics based on the gas data generated by the gas measuring unit; And
And a combined feature vector generation unit for generating a combined feature vector based on the gas data generated by the gas measurement unit and the discrimination feature extracted by the discrimination feature extraction unit,
The discriminating feature extracting unit extracts,
LDA is performed to calculate a projection matrix composed of projection vectors based on a dispersion matrix and an interclass dispersion matrix satisfying an objective function, and a PCA is performed to reduce the dimension of a plurality of learning data to define an overall projection matrix And defines a projection matrix of PCA and LDA defined on the basis of the projection matrix and the integrated projection matrix and generates a feature vector composed of a plurality of discrimination features based on the eigenvalues of the projection matrixes of the PCA and the LDA And a combined feature vector generating unit for generating a combined feature vector for gas classification. The method according to claim 1,
Wherein the gas measuring unit comprises:
And sixteen channels are divided into three sections in which the cross electrodes, the micro heating elements and the polymer complexes are respectively disposed, and each of the sixteen channels is composed of different polymer complexes Characterized in that the combined feature vector generation device for gas classification. The method of claim 2,
Wherein the gas measuring unit comprises:
Wherein the total measurement period is divided into a stabilization period, an exposure period, and a purge period, and the reaction by gas is measured at set time intervals during different set times for each period. delete The method according to claim 1,
The discriminating feature extracting unit extracts,
Equation

Using _{(W PCA = argmax W | W} T S T S | _{a, WLDA = argmax W {| W} T W PCA T S B W PCA W | / | | W T W PCA T S W W PCA W}) And defining a projection matrix of the PCA and the LDA. The method according to claim 1,
Wherein the combining feature generation unit comprises:
The gas data measured by the gas measuring unit is classified into a measured value of the stabilization period, a measured value of the exposure period, and a measured value of the fuzzy period, and the global characteristic is calculated based on the measured value of the entire interval, And the local feature of each section is calculated on the basis of the value of the local feature. The method of claim 6,
Wherein the combining feature generation unit comprises:
Constructing a basic feature set composed of a plurality of elements based on the calculated local features and global features, setting a weight vector of the regional features and the global features included in the basic feature set using the ReliefF method, And a combined feature vector generating unit for generating a combined feature vector for the gas classification. The method of claim 7,
Wherein the combining feature generation unit comprises:
Equation

(A _j is the weight of the jth feature, Ykj is the jth basic feature of the randomly selected kth sample, Hrj is the random number of samples selected from the samples having the same class for the randomly selected kth sample, s j-th basic characteristics of the sample, Mrj is the j-th basic feature of the in the sample belonging to the different classes for a selected randomly k th sample distance nearest s samples, N is the number of samples, P (C _l) is a class C _l P (CYk) is the prior probability of the class to which Yk belongs)
And a weight vector of the local feature and the global feature included in the basic feature set is calculated using the weighted feature vector. A combined feature vector generation method using a combined feature vector generation apparatus for gas classification,
Generating gas data by measuring a reaction by a gas flowing into the chamber;
Extracting discrimination characteristics based on the generated gas data; And
Generating the combined feature vector based on the generated gas data and the extracted discrimination feature,
The step of extracting the discriminating characteristic comprises:
Defining a dispersion matrix in the class and a dispersion matrix between classes;
Calculating a projection matrix composed of projection vectors based on a dispersion matrix in the class and a dispersion matrix between classes that satisfy the objective function by performing the LDA;
Performing a PCA to reduce a dimension of a plurality of learning data to define an overall projection matrix;
Defining a projection matrix of the PCA and the LDA based on the projection matrix and the integrated projection matrix; And
And extracting a discriminant feature based on the eigenvalues of the projection matrixes of the PCA and the LDA. The method of claim 9,
In the step of generating the gas data,
Wherein the sixteen channels are divided into three sections in which the cross electrodes, the micro heating elements and the polymer complexes are disposed, respectively, and the sixteen channels are divided into three sections, Wherein the response to the gas flowing into the gas collection chamber is measured. The method of claim 9,
In the step of generating the gas data,
Measuring a response of the sensor to the gas at a set time interval during the set time in the stabilization period;
Measuring the response of the sensor to the gas at the set time interval during the set time in the exposure section; And
Measuring the response of the sensor to the gas at set time intervals during the set time in the purge interval,
Wherein the set time is set differently for each interval. delete The method of claim 9,
In the step of defining the projection matrixes of the PCA and the LDA,
Equation

Using _{(W PCA = argmax W | W} T S T S | _{a, WLDA = argmax W {| W} T W PCA T S B W PCA W | / | | W T W PCA T S W W PCA W}) And defining a projection matrix of the PCA and the LDA. The method of claim 9,
Wherein generating the combined feature vector comprises:
Classifying the measured gas data into a measured value of the stabilization period, a measured value of the exposure period, and a measured value of the purge period;
Calculating global features and regional features based on measured values classified by measurement intervals;
Constructing a basic feature set based on the calculated global feature and local feature;
Setting a weight vector of a local feature and a global feature included in the basic feature set using a ReliefF method; And
And constructing a combined feature vector based on the weight vector. 15. The method of claim 14,
In the step of setting the weight vector,
Equation

(A _j is the weight of the jth feature, Ykj is the jth basic feature of the randomly selected kth sample, Hrj is the random number of samples selected from the samples having the same class for the randomly selected kth sample, s j-th basic characteristics of the sample, Mrj is the j-th basic feature of the in the sample belonging to the different classes for a selected randomly k th sample distance nearest s samples, N is the number of samples, P (C _l) is a class C _l P (CYk) is the prior probability of the class to which Yk belongs)
And calculating a weight vector of a local feature and a global feature included in the basic feature set.

Images