KR102811004B1 - Apparatus for generating target gas detection model based on CNN learning and method thereof - Google Patents

Abstract

The present invention relates to a device and method for generating a target gas detection model based on CNN learning. According to the present invention, a mixed gas detection method is provided, including the steps of collecting time series data of different time lengths observed from a multi-channel structure sensor array for a plurality of gas samples in which at least one gas is mixed, the steps of searching for a shortest time length among the time lengths of the collected plurality of time series data and generating a plurality of candidate length values having the shortest time length as a maximum value, the steps of applying a time window of L×M size (L: the candidate length value, M: the number of channels) corresponding to the candidate length values to the time series data and sliding it by a set time unit to process a plurality of matrix data of L×M size from the time series data, the steps of training a CNN-based classification model by using a plurality of matrix data processed from the corresponding time series data for each gas sample as input values and a labeling value regarding the presence or absence of a target gas in the gas sample as an output value, and the steps of determining an optimal candidate length value from among the plurality of candidate length values from a result of training the classification model for each of the plurality of candidate length values.
According to the present invention, the performance of a classification model for target gas detection can be improved by securing a sufficient amount of learning data from a small amount of time series data.

Description

{Apparatus for generating target gas detection model based on CNN learning and method thereof}

The present invention relates to a device and method for generating a target gas detection model based on CNN learning, and more specifically, to a mixed gas detection device and method capable of improving the performance of a classification model by securing a sufficient amount of sample data from a small amount of time series data.

In order to predict the type of gas present in a mixed gas sample based on a learning model, a large amount of time series data of various types is required.

However, it is impossible to secure a large amount of time series data of various types in a short period of time, and since a small amount of time series data is not enough to secure learning data, there is a disadvantage in that if a learning model is built using the existing universal data composition method, it will show poor performance.

Figure 1 is a diagram explaining a learning data composition method using conventional PCA.

This figure 1 corresponds to a method of concatenating rows of time series data to form a vector representing the time series data. Each sample data illustrated in Figure 1 is assumed to be sensor data observed from a gas sample using a multi-channel sensor array.

Each row of sample data represents a time slot, and each column represents each sensor channel of the sensor array. As shown in Fig. 1, when the number of channels of the sensor array is 4, each sample data obtains 4 sensing values for each time slot. At this time, it can be seen that the time length of each sample data is different. This is because the collection time length of the sample data cannot always be the same depending on the observation conditions, field environment, and setting information.

The configuration method 1 illustrated in Fig. 1 is a method of taking a time interval of the same size for multiple sample data having different time lengths, concatenating the data of the time intervals into a single line of multidimensional vectors, and then reconstructing the data by reducing the dimensionality using the PCA technique (principal component analysis). The reconstructed data is used as input data (training data) for learning the classification model.

For example, after taking the time interval of 2 columns from sample data 1, it is spread out into one line, an 8-dimensional vector x1 is processed, and then the vector dimension is reduced to 4 dimensions through the PCA technique. In this way, vectors x1, x2, and x3 with the same number of dimensions are obtained corresponding to sample data 1, 2, and 3, respectively.

As a result, according to the configuration method 1, sample data 1, 2, and 3 of different time lengths are finally converted into vectors x1, x2, and x3 of the same dimension, and the data converted in this way are used as input data for building a learning model and trained.

This configuration method 1 has the advantage that sensor value characteristics that vary over time are reflected in model learning because it obtains learning data by taking time blocks, but it requires an additional dimensionality reduction technique due to high-dimensional data transformation through CONCAT, and since learning data is created only as many as the number of collected sample data, it has the limitation that only a very small amount of learning data can be secured.

Figure 2 is a diagram schematically explaining a learning data composition method (method 2) using a conventional differential method.

This Fig. 2 considers each row as an independent sample through the difference of time series data, and unlike Fig. 1, multiple learning data are obtained per sample data, so the learning data can be greatly increased.

In the case of this configuration method 2, although a large amount of data can be secured, there is a disadvantage in that it is difficult to view each row as an independent input data even after difference is made because each row within the same sensor data actually has a very high correlation.

Therefore, a new type of time series data classification methodology is required that can guarantee excellent performance with only a small amount of time series data while overcoming the shortcomings of the methods of Figs. 1 and 2.

The technology that serves as the background for the present invention is disclosed in Korean Patent No. 10-1852074 (announced on April 25, 2018).

The purpose of the present invention is to provide a device and method for generating a target gas detection model based on CNN learning, which can secure a sufficient amount of learning data from a small amount of time series data and improve the performance of a classification model for target gas detection.

The present invention provides a method for generating a target gas detection model using a target gas detection model generating device based on CNN learning, the method comprising: a step of collecting time series data of different time lengths observed from a multi-channel structure sensor array for a plurality of gas samples in which at least one gas is mixed; a step of searching for a shortest time length among the time lengths of the collected plurality of time series data and generating a plurality of candidate length values having the shortest time length as a maximum value; a step of processing a plurality of matrix data of L×M size from the time series data by applying a time window of L×M size (L: the candidate length value, M: the number of channels) corresponding to the candidate length value to the time series data and sliding it by a set time unit; a step of training a CNN-based classification model by using a plurality of matrix data processed from the time series data of each gas sample as input values and a labeling value regarding the presence or absence of a target gas in the gas sample as an output value; and a step of determining an optimal candidate length value from among the plurality of candidate length values from a result of training the classification model for each of the plurality of candidate length values.

Additionally, the sensor array may be implemented including a plurality of sensors selected from a metal oxide (MOX) sensor, an electrochemical sensor, a photoionization detector (PID), a temperature sensor, and a humidity sensor.

In addition, the step of generating the plurality of candidate length values can generate a list of candidate length values (L = 1,2,…,l) from 1 to the shortest time length (l).

In addition, the target gas may include at least one of toluene, dimethyl sulfide (DMS), and butyl acetate, and the classification model may include at least one of a first classification model learned to classify the presence or absence of toluene in a gas sample, a second classification model learned to classify the presence or absence of dimethyl sulfide, and a third classification model learned to classify the presence or absence of butyl acetate.

In addition, the step of processing multiple matrix data from the time series data can process the multiple matrix data by applying a time window of the size of L×M to the time series data and sliding it by one time slot.

In addition, the mixed gas detection method may further include a step of acquiring time series data for an arbitrary gas sample from the sensor array, a step of processing a plurality of matrix data having a size of L'×M corresponding to the optimal candidate value (L') for the acquired time series data, and a step of sequentially inputting the plurality of processed matrix data into the classification model to provide a classification for each hour of the presence or absence of a target gas to be detected.

And, the present invention provides a mixed gas detection device utilizing a small number of sensor data, comprising: a data collection unit for collecting time series data of different time lengths observed from a multi-channel structure sensor array for a plurality of gas samples in which at least one gas is mixed; a candidate value generation unit for searching for a shortest time length among the time lengths of the plurality of collected time series data and generating a plurality of candidate length values having the shortest time length as a maximum value; a data processing unit for processing a plurality of matrix data of L×M size from the time series data by applying a time window of L×M size (L: the candidate length value, M: the number of channels) corresponding to the candidate length value to the time series data and sliding it by a set time unit; a learning unit for learning a CNN-based classification model using a plurality of matrix data processed from the time series data of each gas sample as input values and a labeling value regarding the presence or absence of a target gas in the gas sample as an output value; and a control unit for determining an optimal candidate length value from among the plurality of candidate length values from a result of learning the classification model for each of the plurality of candidate length values.

In addition, the mixed gas detection device may further include a data acquisition unit that acquires time series data for an arbitrary gas sample from the sensor array, and a classification unit that sequentially inputs a plurality of matrix data of the size of L'×M, processed by the data processing unit in response to the optimal candidate value (L') for the acquired time series data, into the classification model to provide a classification for the presence or absence of a target gas to be detected every hour.

According to the present invention, a sufficient amount of learning data can be secured from a small amount of time series data to improve the performance of a classification model for target gas detection, and since input data can be configured in a matrix form rather than a vector form, there is an advantage in that CNN learning is possible.

In addition, the present invention applies time windows of multiple candidate lengths to time series data of different time lengths observed from a sensor array, and individually conducts model learning from a large amount of learning data obtained by sliding, thereby determining the optimal candidate length with the highest classification performance.

In addition, the input data can be reproduced by sliding the time window of the determined optimal candidate length applied to the time series data of the gas sample to be inspected, and this can be applied to the learned classification model to accurately detect the presence of the target gas.

Figure 1 is a diagram explaining a learning data composition method using conventional PCA.
Figure 2 is a diagram explaining a learning data composition method using a conventional differential method.
FIG. 3 is a drawing showing the configuration of a target gas detection model generation device according to an embodiment of the present invention.
Figure 4 is a drawing explaining a model creation method using Figure 3.
FIG. 5 is a drawing explaining a data processing method according to an embodiment of the present invention.
FIG. 6 is a diagram illustrating a concept of detecting a target gas using a classification model generated according to an embodiment of the present invention.
FIG. 7 is a diagram illustrating time series sensor data collected according to an embodiment of the present invention.
Figure 8 is a diagram expressing labeling values for each sample data in an embodiment of the present invention.
FIGS. 9 and 10 are diagrams summarizing the conventional data configuration methods 1 and 2 illustrated in FIGS. 1 and 2.
Figure 11 is a drawing explaining a data configuration method according to an embodiment of the present invention.
Figure 12 is a diagram showing the number of samples derived for each experiment.
Figures 13 to 15 are diagrams showing the classification performance of classification models 1, 2, and 3 for each experiment.

Hereinafter, with reference to the attached drawings, embodiments of the present invention will be described in detail so that those with ordinary skill in the art can easily implement the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In addition, in order to clearly describe the present invention in the drawings, parts that are not related to the description are omitted, and similar parts are assigned similar drawing reference numerals throughout the specification.

Throughout the specification, when a part is said to be "connected" to another part, this includes not only the case where it is "directly connected" but also the case where it is "electrically connected" with another element in between. Also, when a part is said to "include" a component, this does not mean that it excludes other components, but rather that it may include other components, unless otherwise stated.

The present invention relates to a technique for generating a target gas detection model based on CNN learning, and proposes a technique for generating a CNN-based classification model that detects the presence or absence of a target gas by applying a time window of a set length to time series data collected by observing a mixed gas (hereinafter, “gas sample”) in which at least one gas is mixed and manufactured through a multi-channel sensor array and processing a plurality of sample data by sliding the data, and using the data as input data.

FIG. 3 is a drawing showing the configuration of a target gas detection model generation device according to an embodiment of the present invention, and FIG. 4 is a drawing explaining a model generation method using FIG. 3.

Referring to FIGS. 3 and 4, a target gas detection model generation device (100) according to an embodiment of the present invention includes a data collection unit (110), a candidate value generation unit (120), a data processing unit (130), a learning unit (140), a control unit (150), and may further include a data acquisition unit (160), a classification unit (170), and an output unit (180).

Here, the operation of each part (110, 120, 130, 140, 160, 170, 180) and the data flow between each part can be controlled by the control part (150).

First, the data collection unit (110) collects raw data for learning the classification model. Specifically, the data collection unit (110) collects time series data of different time lengths observed from a sensor array of a multi-channel structure for multiple gas samples in which at least one gas is mixed (S410).

Here, the data collection unit (110) can also collect information (labeling value corresponding to the correct answer) on at least one gas actually existing in the gas sample. In addition, among the collected gas samples, the target gas to be detected may not exist at all, or at least one target gas may exist.

The data collection unit (110) can obtain time series data from each of various types of gas samples mixed in different combinations, thereby improving the detection performance of target gases from an unknown mixed gas sample to be inspected.

In an embodiment of the present invention, the target gas may include at least one of toluene, dimethyl sulfide (DMS), and butyl acetate. Of course, the target gas is not necessarily limited to the examples described above, and the device and technique according to an embodiment of the present invention may be utilized to detect a wider variety of target gases.

Additionally, the sensor array can be implemented including a plurality of sensors selected from a metal-oxide (MOX) sensor, an electrochemical sensor, a photoionization detector (PID), a temperature sensor, and a humidity sensor.

Table 1 below shows an example of a sensor array configuration.

Information on array sensors Sensor number Target gases Sensor type 1 VOCs, air contaminants MOX 2 Amine, sulfurous gas MOX 3 VOCs with ionization potential < 10.6 eV PID 4 Formaldehyde Electrochemical 5 H ₂ S Electrochemical 6 S0 ₂ Electrochemical 7 N0 ₂ Electrochemical

Of course, in addition to the types of sensors listed, various types of sensors that are used to detect various hazardous substances, odors, accidents, etc. can be additionally arranged. In addition, all sensor values can be converted (normalized) into values of 0-4000 corresponding to 0-5V through an ADC (analog-to-digital converter).

Figure 5 illustrates three sample data (time series data) observed from three gas samples for which the correct answer is known. Each sample data includes sensing values for the time axis and the channel axis.

Each row of sample data represents a time slot (a unit of measurement time), and each column represents each sensor channel of the sensor array. If the number of channels of the sensor array is 4, each sample data obtains 4 sensing values for each time slot. Sample data 1 represents time series data observed from a 4-channel sensor array during three time slots.

Additionally, sample data 1 is time series data obtained over three time intervals, sample data 2 is time series data obtained over four time intervals, and sample data 3 represents time series data obtained over two time intervals.

In this way, the time length of each sample data is different, which takes into account the situation in which the data collection length varies depending on the field environment, observation conditions, and setup information, as explained in the background technology above.

For convenience of explanation, this Figure 5 illustrates sample data of a very short time length, but in reality, observed sample data may consist of tens to hundreds or more time slots.

Next, the candidate value generation unit (120) searches for the shortest time length among the time lengths of the collected multiple time series data and generates multiple candidate length values with the shortest time length (l) as the maximum value (S330). At this time, the candidate value generation unit (120) can generate a candidate length value list (L = 1, 2,…, l) including ‘1’ to the shortest time length (l).

In Fig. 5, the time lengths of sample data 1, 2, and 3 are 3, 4, and 2, respectively, and the sample data with the shortest time length is 3. At this time, since the time length of sample data 3 is '2', the candidate length values are generated from 1 to 2, and are defined as L∈{1,2}.

Of course, as another example, if the time lengths of sample data 1, 2, and 3 were 100, 25, and 40, respectively, multiple candidate length values are generated from 1 to 25, and in this case, L∈{1,2,…,25} is defined.

Thereafter, the data processing unit (130) applies a time window of size L×M corresponding to the candidate length value (L: the candidate length value, M: number of channels) to the time series data and slides it by a set time unit, thereby processing multiple matrix data of size L×M from the time series data (S440).

In the case of Fig. 5, a 2×4 sized window is applied to each sample data and the data is slid by each time slot to reproduce multiple matrix data from each sample data.

For sample data 1, a 2×4 sized window was applied and slid by 1 time slot, thereby processing two 2×4 sized matrix data (X1, X2). In the same way, for sample data 2 and 3, three 2×4 sized matrix data (X3, X4, X5) and one 2×4 sized matrix data (X6) were processed, respectively.

In this way, in the data configuration method proposed in the present invention, each matrix created by cutting time series data into specific time window units is considered as one sample.

Here, since the collected time length of each sample data is different, the number of matrix data to be reproduced is also different. Fig. 5 shows the case where a 2×4 sized window is applied when L=2 among the candidate length values (L∈{1,2}). Similarly, when L=1 is applied, 3, 4, and 2 matrix data can be processed from sample data 1, 2, and 3, respectively.

If the window size is too small, data from various time periods cannot be entered into a single window, but the number of matrix data to be reproduced increases. If the window size is too large, data from various time periods can be entered into a single window, but the number of matrix data to be reproduced decreases.

However, if a large number of sample data are reproduced from limited sample data resources and applied to model learning through these, learning performance and classification accuracy can be improved compared to when a model is learned with a small number of sample data.

Therefore, in an embodiment of the present invention, in order to determine the optimal window size, the learning performance of the classification model is identified for each case where each candidate value is used, and the candidate length value that derives the highest classification performance therefrom is selected as the optimal candidate length value.

Specifically, the learning unit (140) trains a CNN-based classification model by using a plurality of matrix data processed from the time series data for each gas sample as input values and a label value regarding the presence or absence of a target gas in the gas sample as an output value.

In the case of a CNN model, it is a model capable of learning input data of a matrix structure, and an embodiment of the present invention inputs matrix data into a CNN-based classification model and trains the classification model to follow the corresponding label value corresponding to the input matrix data.

That is, when the input data is matrices X1 and X2, the model is trained to follow the correct labeling value corresponding to sample data 1, when the input data is matrices X3, X4 and X5, the model is trained to follow the correct labeling value corresponding to sample data 2, and when the input data is matrix X6, the model is trained to follow the correct labeling value corresponding to sample data 3.

Then, the control unit (150) determines the optimal candidate length (L') from the results of learning the classification model for each of the multiple candidate length values (S450).

Specifically, the control unit (150) determines the optimal candidate length value among the plurality of candidate length values from the result of learning the classification model through the matrix data generated for each of the plurality of candidate length values (L=1,2). For example, if the classification performance of the classification model for each candidate length value is analyzed and the case where L=2 is applied is the best, L'=2 is determined.

Afterwards, by applying a window corresponding to L=2 to the time series data observed from an actual random gas sample, multiple matrix data are obtained, and by inputting this into the learned classification model above, the presence or absence of the target gas in the gas sample can be accurately detected.

Specifically, after step S450, the data acquisition unit (160) acquires time series data for an arbitrary gas sample from the sensor array (S460). Here, the data acquisition unit (160) can receive sensor data of each channel in real time while connected to the sensor array through a network.

Next, the data processing unit (130) processes multiple matrix data of size L'×M (2×4) by applying a sliding window of size L'×M (2×4) corresponding to the optimal candidate value (L'=2) to the time series data acquired in step S460 (S470).

Next, the classification unit (170) can sequentially input the processed multiple matrix data into the classification model to provide real-time classification of the presence or absence of the target gas to be detected every hour (S480). For example, the classification model can output 0 if the target gas (e.g., toluene gas) is detected, and 1 if it is not detected.

Thereafter, the output unit (180) can provide data output from the classification model in real time to a display or a network-connected administrator terminal, user terminal, etc.

FIG. 6 is a diagram illustrating a concept of detecting a target gas using a classification model generated according to an embodiment of the present invention. As shown in FIG. 6, a classification model can be generated for each type of target gas.

For example, the classification model may include at least one of a first classification model (classification model 1) trained to classify the presence or absence of toluene (T) in a gas sample, a second classification model (classification model 2) trained to classify the presence or absence of dimethyl sulfide (DMS), and a third classification model (classification model 3) trained to classify the presence or absence of butyl acetate (BA).

Accordingly, by inputting the matrix data processed from the time series data of the random gas sample to be detected into classification model 1, the presence or absence of toluene in the sample can be confirmed, and by inputting the same matrix data into classification models 2 and 3, the presence or absence of dimethyl sulfide and butyl acetate in the sample can be confirmed, respectively.

According to the above-described method, the present invention has the advantage of being able to secure a large amount of input data from sample data of limited length and of being able to configure input data in the form of a matrix rather than a vector, thereby enabling CNN learning.

Below, the results of learning a classification model using a data configuration technique according to an embodiment of the present invention and the existing data configuration techniques (configuration 1, configuration 2) shown in FIGS. 1 and 2 and comparing the classification performance thereof are described.

Fig. 7 is a diagram illustrating time series sensor data collected according to an embodiment of the present invention. Fig. 7 shows sensor data acquired over time from a sensor array having a total of 11 sensor channels for an arbitrary sample. The time series data serves as input data in the problem of determining whether a specific gas is included in a mixed sample.

For the experiment, sensor data for a total of 100 samples were collected, and Fig. 7 illustrates one sample data among them. Each of the 100 sample data has between 40 and 100 row values (time slots). Here, Fig. 7 illustrates only three columns, which simply illustrates data for three time slots out of the total time slots.

Figure 8 is a diagram expressing labeling values for each sample data in an embodiment of the present invention.

The left figure of Fig. 8 is a numerical representation of the concentration ratio of the target gas actually present in each sample, with 0 representing no presence at all. The right figure is the result of converting each value shown in the left figure into a binary value of '1' or '0' corresponding to the presence or absence of the target gas.

The labeling value serves as output data in the problem of determining whether a specific target gas is included in a mixed sample. Since the embodiment of the present invention is for determining whether a target gas is included in a sample, each value is converted to a binary value. Here, the labeling value for toluene for each sample can be used for learning classification model 1, and the labeling values for DMS and BS can be used for learning classification models 2 and 3, respectively.

FIGS. 9 and 10 are diagrams summarizing the conventional data configuration methods 1 and 2 illustrated in FIGS. 1 and 2.

Figure 9 shows the methodology of various cases belonging to the data composition method 1 (First experiment) of Figure 1, and there are a total of 8 cases. In this case, it can be seen that classification models based on logistic regression and random forest structures were utilized, respectively.

Figure 10 shows the methodology of various cases belonging to the data composition method 2 (Second experiment) of Figure 2, and there are a total of four cases. Here, logistic regression and classification models based on random forest structure are utilized respectively.

Figure 11 is a drawing explaining a data configuration method according to an embodiment of the present invention.

As shown in Fig. 11, the data composition method (Third experiment) proposed in the present invention utilizes a classification model of a CNN structure, and there are a total of two cases. In the case of the present invention, baseline CNN and weighted CNN were utilized as classification models. Weighted CNN is a CNN technique that learns by applying different weights to each class.

Figure 12 is a diagram showing the number of samples derived for each experiment. From the results of these 12, it can be seen that different amounts of sample data are obtained from the same amount of sample data. In other words, it can be seen that the proposed data composition method (Third experiment) can process a much larger number of sample data than the existing method 1 (First experiment) and can secure a large amount of sample data almost similar to the existing method 2 (Second experiment).

Figures 13 to 15 are diagrams showing the classification performance of classification models 1, 2, and 3 for each experiment. Figure 13 shows the classification performance confirmed by applying the sample data acquired for each experiment to classification model 1 (toluene detection model), and Figures 14 and 15 show the classification performance confirmed by applying the data to classification model 2 (DMS detection model) and classification model 3 (BS detection model), respectively.

Based on the methodologies mentioned above, a total of 14 models were trained for each gas sample, and the performance of each model was measured using AUC (Area Under the Curve) and F1 score as evaluation metrics. In each table of Figs. 13 to 15, the first 8 columns out of the total 14 columns represent the performance of 8 methodologies of data composition method 1 (First experiment), the next 4 columns represent the performance of 4 methodologies of data composition method 2 (Second experiment), and the last 2 columns represent the performance of 2 methodologies of the data composition method of the present invention (Third experiment).

For toluene, the classification model based on CNN showed the best performance in terms of both AUC and F1 score (Fig. 13). For DMS, the logistic regression in the second experiment showed the highest performance in terms of AUC, but the classification model of the present invention based on CNN also showed significant performance results (Fig. 14). For butyl acetate, the classification model of the present invention based on CNN showed the highest performance in terms of F1 score (Fig. 15).

According to the present invention as described above, a sufficient amount of learning data can be secured from a small amount of time series data to improve the performance of a classification model for target gas detection, and since input data can be configured in a matrix form rather than a vector form, there is an advantage in that CNN learning is possible.

Although the present invention has been described with reference to the embodiments shown in the drawings, these are merely exemplary, and those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Accordingly, the true technical protection scope of the present invention should be determined by the technical idea of the appended claims.

100: Target gas detection model generation device
110: Data collection section 120: Candidate value generation section
130: Data processing department 140: Learning department
150: Control unit 160: Data acquisition unit
170: Classification section 180: Output section

Claims (12)

A method for generating a target gas detection model using a target gas detection model generating device based on CNN learning,
A step of collecting time series data of different time lengths observed from a sensor array of a multi-channel structure for a plurality of gas samples in which at least one gas is mixed;
A step of searching for the shortest time length among the time lengths of a plurality of collected time series data and generating a plurality of candidate length values having the shortest time length as the maximum value, wherein a list of candidate length values (L = 1,2,…,l) from 1 to the shortest time length (l) is generated;
A step of processing a plurality of matrix data of L×M size from the time series data by applying a time window of L×M size (L: candidate length value, M: number of channels) corresponding to the candidate length value to the time series data and sliding it by a set time unit;
A step of training a CNN-based classification model by using a plurality of matrix data processed from the corresponding time series data for each gas sample as input values and a label value regarding the presence or absence of a target gas in the gas sample as an output value; and
A method for generating a target gas detection model, comprising a step of determining an optimal candidate length value among the plurality of candidate length values from a result of learning the classification model for each of the plurality of candidate length values. In claim 1,
The above sensor array,
A method for generating a target gas detection model implemented by including a plurality of sensors selected from a metal oxide (MOX) sensor, an electrochemical sensor, a photoionization detector (PID), a temperature sensor, and a humidity sensor. delete In claim 1,
The above target gas is,
Containing at least one of toluene, dimethyl sulfide (DMS), and butyl acetate;
The above classification model is,
A method for generating a target gas detection model, comprising at least one of a first classification model trained to classify the presence or absence of toluene in a gas sample, a second classification model trained to classify the presence or absence of dimethyl sulfide, and a third classification model trained to classify the presence or absence of butyl acetate. In claim 1,
The step of processing multiple matrix data from the above time series data is:
A method for generating a target gas detection model by processing the plurality of matrix data while applying the time window of the above L×M size to the above time series data and sliding it by one time slot unit. In claim 1,
A step of acquiring time series data for any gas sample from the above sensor array;
A step of processing a plurality of matrix data of sizes L'×M corresponding to the optimal candidate length value (L') for the acquired time series data; and
A method for generating a target gas detection model further comprising a step of sequentially inputting a plurality of processed matrix data into the classification model to provide a classification result every hour for the presence or absence of a target gas to be detected. In a device for generating a target gas detection model based on CNN learning,
A data acquisition unit for collecting time series data of different time lengths observed from a sensor array of a multi-channel structure for a plurality of gas samples in which at least one gas is mixed;
A candidate value generation unit that searches for the shortest time length among the time lengths of multiple collected time series data and generates multiple candidate length values with the shortest time length as the maximum value, and generates a candidate length value list (L = 1,2,…,l) from 1 to the shortest time length (l);
A data processing unit that processes multiple matrix data of L×M size from the time series data by applying a time window of L×M size (L: the candidate length value, M: number of channels) corresponding to the candidate length value to the time series data and sliding it by a set time unit;
A learning unit that learns a CNN-based classification model by using a plurality of matrix data processed from the corresponding time series data for each gas sample as input values and a label value regarding the presence or absence of a target gas in the gas sample as an output value; and
A target gas detection model generation device including a control unit that determines an optimal candidate length value among the plurality of candidate length values from the results of learning the classification model for each of the plurality of candidate length values. In claim 7,
The above sensor array,
A target gas detection model generation device implemented including a plurality of sensors selected from a metal oxide (MOX) sensor, an electrochemical sensor, a photoionization detector (PID), a temperature sensor, and a humidity sensor. delete In claim 7,
The above target gas is,
Containing at least one of toluene, dimethyl sulfide (DMS), and butyl acetate;
The above classification model is,
A target gas detection model generating device comprising at least one of a first classification model trained to classify the presence or absence of toluene in a gas sample, a second classification model trained to classify the presence or absence of dimethyl sulfide, and a third classification model trained to classify the presence or absence of butyl acetate. In claim 7,
The above data processing unit,
A target gas detection model generation device that processes the plurality of matrix data by applying the time window of the above L×M size to the above time series data and sliding it by one time slot unit. In claim 7,
A data acquisition unit for acquiring time series data for any gas sample from the above sensor array; and
A target gas detection model generation device further comprising a classification unit for sequentially inputting a plurality of matrix data of size L'×M, processed by the data processing unit in response to the optimal candidate length value (L') of the acquired time series data, into the classification model to provide a classification for the presence or absence of a target gas to be detected every hour.