KR20240043388A - Apparatus for classification model building to classify fire accelerators and method thereof - Google Patents

Abstract

The present invention relates to an apparatus and method for building a classification model for classifying fire accelerants.
A classification model building device for classifying fire accelerants according to the present invention includes a database that collects GC-MS data obtained through suspected arson cases, a data preprocessor that preprocesses the collected GC-MS data, and the preprocessed GC-MS data. A model building unit that generates a data set using MS data, determines hyperparameters for building multiple classification models using the generated data sets, and extracts fire substances through the constructed multiple classification models. Includes a verification unit that verifies performance.

Description

Apparatus for classification model building to classify fire accelerators and method thereof}

The present invention relates to an apparatus and method for building a classification model for classifying fire accelerators. More specifically, the present invention relates to building a classification model based on machine learning to classify accelerators used in fires using GC-MS data. It relates to a device and method for doing so.

In modern society, financial losses associated with fire damage are an important consideration with a variety of implications. Therefore, it is important to determine whether the fire occurred naturally or was caused by arson. Several fire accelerants are used in most common arson incidents, such as hydrocarbon fuels such as gasoline, kerosene, diesel, and organic solvents or candles.

Conventionally, several marker components of fire accelerants are separated and identified through analysis using Gas Chromatography - Mass Spectrometer (GC-MS), Solid Phase Micro-Extraction (SPME), or solvent extraction. did.

Because gas chromatography mass spectrometry (GC-MS) provides continuous and natural information about samples, many attempts have been made to develop machine learning models using gas chromatography mass spectrometry (GC-MS) data.

Republic of Korea Patent Publication No. 10-2375679 (announced on March 18, 2022)

The technical problem to be achieved by the present invention is to provide an apparatus and method for building a classification model based on machine learning to classify accelerants used in fires using GC-MS data.

In order to achieve this technical task, a database that collects GC-MS data obtained through suspected cases of arson according to an embodiment of the present invention, a data preprocessor that preprocesses the collected GC-MS data, and the preprocessed GC-MS data A model building unit that generates a data set using and determines hyperparameters for building a plurality of classification models using the generated data set, and the performance of extracting fire substances through the constructed plurality of classification models. Includes a verification unit that verifies.

The database stores the GC-MS data by classifying it into a plurality of groups according to the substance that caused the fire, and the substance may include at least one of fire accelerants, gasoline, kerosene, diesel, candles, and organic solvents.

The plurality of classification models may include a random forest (RF) model, a support vector machine (SVM) model, and a convolutional neural network (CNN) model.

When the classification model to be learned is the random forest (RF) model and the support vector machine (SVM) model, the preprocessor analyzes the GC-MS data to extract a plurality of peaks corresponding to the chemical substance, and extracts Cosine similarity can be calculated by comparing the peak and the reference peak.

The preprocessor may extract material information included in the GC-MS data if the calculated similarity is greater than a reference value.

When the classification model to be learned is the convolutional neural network (CNN) model, the preprocessor can preprocess the GC-MS data using the following equation.

here, represents the standardized value of each element, represents each element of the matrix.

The model building unit may determine hyperparameters of the random forest (RF) model using the number of estimators and the maximum depth of the model.

The model building unit may determine hyperparameters using a radial basis function (RBF) as the kernel of a support vector machine (SVM) model.

The model building unit places a dropout layer behind the convolutional layer, and the activation function of all middle layers is a ReLU (Rectified Linear Unit) function, and the sigmoid function is applied to the output layer to create a convolutional neural network (CNN) model. Hyperparameters can be determined.

In addition, in the classification model building method using the classification model building device according to an embodiment of the present invention, collecting GC-MS data obtained through suspected arson cases to build a database, the collected GC-MS data A preprocessing step, generating a data set using the preprocessed GC-MS data, determining hyperparameters for building a plurality of classification models using the generated data set, and the constructed plurality of classification models. It includes the step of verifying the performance of extracting fire substances through.

In this way, according to the present invention, it is possible to classify fire accelerants from fire residues suspected of actual arson using artificial intelligence, which can help in effective and accurate investigation of fire accidents.

1 is a configuration diagram illustrating a classification model building device according to an embodiment of the present invention.
Figure 2 is a flowchart of a classification model building method using a classification model building device according to an embodiment of the present invention.
FIG. 3 is an example diagram illustrating conditions for setting optimal hyperparameters for a random forest (RF) model and a support vector machine (SVM) model in step S230 shown in FIG. 2.
Figure 4 is an example diagram for explaining the convolutional neural network (CNN) model built in step S230 shown in Figure 2.
Figure 5 is an example diagram to explain average GC-MS data for each category.
Figure 6 is an example diagram showing feature points extracted from GC-MS data.
Figure 7 is an exemplary diagram showing the performance evaluation results of a classification model according to an embodiment of the present invention.
Figure 8 is an exemplary diagram showing the ROC curve and PR curve of a classification model according to an embodiment of the present invention.

Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the attached drawings. In this process, the thickness of lines or sizes of components shown in the drawing may be exaggerated for clarity and convenience of explanation.

In addition, the terms described below are terms defined in consideration of functions in the present invention, and may vary depending on the measurement subject, operator's intention, or custom. Therefore, definitions of these terms should be made based on the content throughout this specification.

Hereinafter, a classification model building device for classifying fire accelerants according to an embodiment of the present invention will be described in more detail using FIG. 1.

1 is a configuration diagram illustrating a classification model building device according to an embodiment of the present invention.

As shown in FIG. 1, the classification model building apparatus 100 according to an embodiment of the present invention includes a database 110, a data preprocessing unit 120, a model building unit 130, and a verification unit 140. .

First, the database 110 stores raw data obtained through cases of suspected arson. To elaborate, the database 110 collects GC-MS data obtained through gas chromatography mass spectrometry (GC-MS) and classifies the collected GC-MS data according to the type of fire catalyst.

The data preprocessing unit 120 performs preprocessing on GC-MS data to use it as learning data. As an example, the data preprocessor 120 extracts a two-dimensional matrix from the GC-MS data to use the GC-MS data as training data for a convolutional neural network (CNN) model, and the extracted 2 Standardize the matrix of dimensions.

The model building unit 130 builds a plurality of classification models based on machine learning. To explain this again, the model building unit 130 obtains a classification model based on a random forest (RF) model, a support vector machine (SVM) model, and a convolutional neural network (CNN) model. .

Then, the model building unit 130 classifies the learning data that has been preprocessed for a plurality of models into training data and validation data, and determines hyperparameters corresponding to each model using the classified training data and validation data.

Finally, the verification unit 140 performs verification of the classification model built with the determined hyperparameters.

Hereinafter, a method of building a classification model for classifying fire accelerants according to an embodiment of the present invention will be described in more detail using FIGS. 2 to 8.

Figure 2 is a flowchart of a classification model building method using a classification model building device according to an embodiment of the present invention.

As shown in FIG. 2, the classification model building device 100 according to an embodiment of the present invention collects GC-MS data obtained through suspected arson cases and builds a database (S210).

To elaborate, the classification model building device 100 receives GC-MS data from the server of a research institute that analyzes suspected cases of arson, such as the National Institute of Forensic Science.

At this time, the GC-MS data is the result of analysis according to the standard analysis conditions listed in Table 1 below.

The above standard analysis conditions are not essential, and various analysis conditions may be modified arbitrarily for more appropriate analysis.

The received GC-MS data is classified according to fire accelerant, gasoline, kerosene, diesel, candle, and organic solvent and stored in the database 110.

Next, the data preprocessing unit 120 performs preprocessing on the GC-MS data stored in the database 110 to use it as learning data (S220).

The GC-MS data stored in the database 110 is obtained from actual fire cases, resulting in an imbalance between categories.

Therefore, the data preprocessing unit 120 preprocesses the GC-MS data according to the classification model.

To explain this in more detail, the classification model building device 100 according to an embodiment of the present invention creates a classification model using a random forest (RF) model, a support vector machine (SVM) model, and a convolutional neural network (CNN) model. I want to build it.

Therefore, the data preprocessing unit 120 preprocesses GC-MS data to improve the performance of the random forest (RF) model, support vector machine (SVM) model, and convolutional neural network (CNN) model.

Below, we will describe in more detail how to preprocess GC-MS data to learn each classification model.

In the case of a random forest (RF) model and a support vector machine (SVM) model, the data preprocessor 120 preprocesses the GC-MS data into one-dimensional data. To explain this again, the data preprocessor 120 sets the peaks represented by substances frequently found in fire residues as the reference peak. Then, the data preprocessing unit 120 analyzes the GC-MS data and detects peaks corresponding to each chemical substance. Then, the data pre-processing unit 120 compares the plurality of detected peaks and the reference peak to determine whether similar peaks appear. Then, when it is determined that a similar peak has occurred, the data preprocessor 120 calculates the cosine similarity between the similar peak and the reference peak. And if the calculated similarity exceeds the reference value, the data preprocessor 120 extracts material information corresponding to the corresponding peak.

Additionally, when the classification model is a convolutional neural network (CNN) model, the data preprocessor 120 preprocesses the two-dimensional matrix included in the GC-MS data into a relatively weak signal using Equation 1 below.

here, represents the standardized value of each element, represents each element of the matrix.

When step S220 is completed, the model building unit 130 determines hyperparameters to apply to each classification model using the preprocessed GC-MS data (S230).

To explain this in more detail, the model building unit 130 classifies GC-MS data corresponding to approximately 20% of the total GC-MS data for which preprocessing has been completed as a test set, and approximately 25% of the remaining GC-MS data is classified into a test set. GC-MS data corresponding to % are classified as the validation set, and the remaining data are classified as the learning set.

Next, the model building unit 130 performs cross-validation using training data and validation data to determine hyperparameters to apply to the classification model.

Figure 3 is an example diagram for explaining conditions for setting optimal hyperparameters for the random forest (RF) model and support vector machine (SVM) model in step S230 shown in Figure 2, and Figure 4 is shown in Figure 2. This is an example diagram to explain the convolutional neural network (CNN) model built in step S230.

When the classification model is a random forest (RF) model, the model construction unit 130 determines hyperparameters using the number of estimators and the maximum depth of the model. Then, as shown in (a) of Figure 3, the random forest (RF) model has a maximum depth of 30 and It shows the highest accuracy when composed of an estimator.

Additionally, when the classification model is a support vector machine (SVM) model, the model building unit 130 uses a radial basis function (RBF) as the kernel of the support vector machine (SVM) model. The model building unit 130 optimizes hyperparameters by performing grid search using a five-layer cross-validation method. Then, as shown in (b) of Figure 3, the ranges of the kernel variables of the optimal hyperparameters are all inside The C value, defined as the reciprocal of the normalization parameter, is 10, and the kernel variable (γ) is optimized at 0.01.

Lastly, when the classification model is a convolutional neural network (CNN) model, the model building unit 130 builds a convolutional neural network (CNN) model composed of a neural network structure such as AlphaNet or GoogLeNet, but uses a 1x1 convolutional layer to optimize performance. Add .

Additionally, as shown in Figure 4, the dropout layer of the convolutional neural network (CNN) model is placed behind the convolutional layer. The activation function of all middle layers is a ReLU (Rectified Linear Unit) function, but the sigmoid function is applied to the output layer.

When step S230 is completed, the verification unit 140 performs verification of the plurality of constructed classification models (S240).

Figure 5 is an example diagram for explaining average GC-MS data for each category, and Figure 6 is an example diagram showing feature points extracted from GC-MS data.

As shown in Figure 5, when the GC-MS data includes fire residues for gasoline or organic solvents, it shows a short retention time, and when it includes diesel or paraffin, it shows a long retention time.

Also, as shown in Figure 6, the GC-MS spectra of gasoline and solvent show similar patterns. This is because both gasoline and solvents contain alkyl ring benzene. The presence of alkylated benzene is generally considered strong evidence for the presence of gasoline, but most alkylated benzenes also produce strong signals in other petroleum solvents.

In addition, fire accelerators with higher alkanes, such as kerosene, diesel, and candles, commonly contain isoparaffin as a distinguishable component, so when the fire catalyst is a candle, the paraffin wax component is used as a classification marker.

Therefore, the apparatus 100 for building a classification model for classifying fire accelerants according to an embodiment of the present invention acquires important material information from GC-MS data through data preprocessing in step S220. The classification model building device 100 collects signal intensity corresponding to the obtained material information.

Then, the classification model building device 100 trains a plurality of classification models using the collected signal intensities, and causes the learned classification model to output the corresponding substance from the input GC-MS data.

Figure 7 is an exemplary diagram showing the performance evaluation results of a classification model according to an embodiment of the present invention.

As shown in FIG. 7, the verification unit 140 evaluates the performance of a random forest (RF) model, a support vector machine (SVM) model, and a convolutional neural network (CNN) model using a test set. Looking at the results, the weighted average F1-scores, defined as the harmonic average of the positive identification rate among the actual positive cases (precision) and the true positive rate among the positive identification cases (recall) of each model, were evaluated as 0.88, 0.88, and 0.92.

Figure 8 is an exemplary diagram showing the ROC curve and PR curve of a classification model according to an embodiment of the present invention.

The area under the receiver operating characteristic (ROC) curve (AU-ROC) is a widely used indicator to determine the performance of classification tools. The closer the value of the area under the ROC curve (AU-ROC) is to 0.5, the more the model is judged to be unable to completely distinguish between positive and negative numbers. The closer the value of the area under the ROC curve (AU-ROC) is to 1, the more positive the model is. It is judged that it is possible to completely distinguish between and negative numbers.

As shown in Figure 8, the area under the ROC curve (AU-ROC) of the random forest (RF) model, support vector machine (SVM) model, and convolutional neural network (CNN) model according to an embodiment of the present invention is 0.98, Since it is 0.98 and 0.99, each classification model has excellent reliability.

Area under the ROC curve (AU-ROC) is a powerful metric, but it is not sufficient for evaluating the performance of models with imbalanced data sets. Therefore, in the present invention, the area under the PR curve (AU-PRC) of each model is also evaluated. Similar to AU-ROC, the area under the PR curve (AU-PRC) can be evaluated as being closer to 1, indicating better model performance. Since the AU-PRC values of the random forest (RF) model, support vector machine (SVM) model, and convolutional neural network (CNN) model are 0.94, 0.93, and 0.96, respectively, the reliability of the corresponding classification model can be judged to be excellent.

In this way, according to the present invention, fire accelerants can be classified from fire residues suspected of actual arson using artificial intelligence, which can help in effective and accurate investigation of fire accidents.

The present invention has been described with reference to the embodiments shown in the drawings, but these are merely illustrative, and those skilled in the art will understand that various modifications and other equivalent embodiments are possible therefrom. will be. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the patent claims below.

100: Classification model building device
110: database
120: data preprocessing unit
130: model construction unit
140: verification unit

Claims (18)

A database that collects GC-MS data obtained from suspected arson cases;
A data preprocessing unit that preprocesses the collected GC-MS data,
A model building unit that generates a data set using the preprocessed GC-MS data and determines hyperparameters for building a plurality of classification models using the generated data set, and
A classification model building device including a verification unit that verifies the performance of extracting fire substances through the constructed plurality of classification models. According to paragraph 1,
The database is,
The GC-MS data is classified and stored into multiple groups according to the substance that caused the fire,
The substance is,
A classification model building device comprising at least one of fire accelerants, gasoline, kerosene, diesel, candles, and organic solvents. According to paragraph 1,
The plurality of classification models are,
A classification model building device including a random forest (RF) model, a support vector machine (SVM) model, and a convolutional neural network (CNN) model. According to paragraph 3,
If the classification model you want to learn is the random forest (RF) model and support vector machine (SVM) model,
The preprocessor,
A classification model building device that analyzes the GC-MS data to extract a plurality of peaks corresponding to chemical substances, and calculates cosine similarity by comparing the extracted peaks with the reference peak. According to clause 4,
The preprocessor,
A classification model building device that extracts material information included in the GC-MS data when the calculated similarity is greater than the reference value. According to paragraph 3,
If the classification model you want to learn is the convolutional neural network (CNN) model,
The preprocessor,
Classification model building device that preprocesses GC-MS data using the following equation:

here, represents the standardized value of each element, represents each element of the matrix. According to paragraph 1,
The model building unit,
A classification model building device that determines the hyperparameters of a random forest (RF) model using the number of estimators and the maximum depth of the model. According to paragraph 1,
The model building unit,
A classification model building device that determines hyperparameters by using the Radial Basis Function (RBF) as the kernel of a support vector machine (SVM) model. According to paragraph 1,
The model building unit,
A dropout layer is placed after the convolutional layer, the activation function of all intermediate layers is a ReLU (Rectified Linear Unit) function, and the sigmoid function is applied to the output layer to determine the hyperparameters of the convolutional neural network (CNN) model. Classification model building device. In the method of building a classification model using a classification model building device,
Building a database by collecting GC-MS data obtained through suspected arson cases,
Preprocessing the collected GC-MS data,
Creating a data set using the preprocessed GC-MS data and determining hyperparameters for building a plurality of classification models using the generated data set, and
A classification model construction method including the step of verifying the performance of extracting fire substances through the constructed plurality of classification models. According to clause 10,
The step of building the database is,
The GC-MS data is classified and stored into multiple groups according to the substance that caused the fire,
The substance is,
A method for building a classification model that includes at least one of the following: fire accelerant, gasoline, kerosene, diesel, candle, and organic solvent. According to clause 10,
The plurality of classification models are,
Methods for building classification models, including random forest (RF) models, support vector machine (SVM) models, and convolutional neural network (CNN) models. According to clause 12,
If the classification model you want to learn is the random forest (RF) model and support vector machine (SVM) model,
The preprocessing step is,
A classification model construction method that extracts a plurality of peaks corresponding to chemical substances by analyzing the GC-MS data and calculates cosine similarity by comparing the extracted peaks with the reference peak. According to clause 13,
The preprocessing step is,
A classification model construction method for extracting material information included in the GC-MS data when the calculated similarity is greater than the reference value. According to clause 12,
If the classification model you want to learn is the convolutional neural network (CNN) model,
The preprocessing step is,
How to build a classification model that preprocesses GC-MS data using the following equation:

here, represents the standardized value of each element, represents each element of the matrix. According to clause 10,
The step of determining the hyperparameters is,
A method of building a classification model that determines the hyperparameters of a random forest (RF) model using the number of estimators and the maximum depth of the model. According to clause 10,
The step of determining the hyperparameters is,
A method of building a classification model that determines hyperparameters by using the Radial Basis Function (RBF) as the kernel of a support vector machine (SVM) model. According to clause 10,
The step of determining the hyperparameters is,
A dropout layer is placed after the convolutional layer, the activation function of all intermediate layers is a ReLU (Rectified Linear Unit) function, and the sigmoid function is applied to the output layer to determine the hyperparameters of the convolutional neural network (CNN) model. How to build a classification model.