Apparatus for chemical material accident risk assessment using artificial intelligence and method thereof

Abstract

According to one embodiment of the present invention, a chemical accident risk assessment device using artificial intelligence comprises at least one processor, wherein the at least one processor trains a chemical accident occurrence prediction model and a chemical accident risk assessment model based on preset data, inputs data related to the current status of a business entity to assess chemical accident risk into the trained chemical accident occurrence prediction model to predict whether a chemical accident will occur, and if the trained chemical accident occurrence prediction model predicts that a chemical accident will occur at the business entity to assess chemical accident risk, determines input information based on the predicted chemical accident, and inputs the determined input information into the trained chemical accident risk assessment model to assess the chemical accident risk for the business entity to assess chemical accident risk.

Description

Technology Field

[0001] The following embodiments relate to a chemical accident risk assessment device using artificial intelligence and a chemical accident risk assessment method using artificial intelligence. Background Technology

[0002] With the rapid development of industry, humanity has enjoyed growth and prosperity, and chemical substances have become indispensable to our daily lives and industrial activities, while the volume of chemical substances in circulation has also increased annually. However, on the other hand, the mismanagement and handling of the environment caused by the growth and prosperity of the chemical industry and related technology sectors have reached a point where they are affecting the development of humanity.

[0003] Chemical accidents can have harmful effects not only on the workplace but also on the surrounding area, nearby residents, and ecosystems; since damage can manifest chronically over a long period, it is difficult to predict the post-accident impact. To mitigate damage, while it is important to predict the movement of chemical substances immediately after an accident and implement appropriate responses, it is equally important to assess the risks and prepare for such accidents.

[0004] Machine learning is a field of artificial intelligence that has evolved from research in pattern recognition and computer learning theory, referring to the development of algorithms and technologies that enable computers to learn. The core of machine learning lies in representation and generalization. Representation refers to the evaluation of data, while generalization refers to the processing of data that is not yet known. It is also a field of computational learning theory.

[0005] Deep learning is defined as a set of machine learning algorithms that attempt a high level of abstraction through a combination of various non-linear transformation techniques, and in a broad sense, it can be described as a field of machine learning that teaches computers human thinking methods.

[0006] A Deep Neural Network (DNN) is an Artificial Neural Network (ANN) composed of multiple hidden layers between an input layer and an output layer. Like general artificial neural networks, Deep Neural Networks can model complex non-linear relationships. Prior art literature

[0007] Republic of Korea Published Patent Application No. 10-2024-0086657 (Published June 18, 2024) The problem to be solved

[0008] According to one embodiment, if a chemical accident is predicted to occur in a business for evaluating the risk of a chemical accident, a chemical accident risk assessment device using artificial intelligence and a chemical accident risk assessment method using artificial intelligence can be provided to assess the risk of the predicted chemical accident.

[0009] In addition, according to another embodiment, a chemical accident risk assessment device using artificial intelligence and a chemical accident risk assessment method using artificial intelligence can be provided, which can train a model for predicting whether a chemical accident will occur and a model for assessing the risk of a chemical accident based on preset data.

[0010] In addition, according to another embodiment, a chemical accident risk assessment device and method capable of generating input information based on a predicted chemical accident can be provided. means of solving the problem

[0011] According to one embodiment of the present invention, a chemical accident risk assessment device using artificial intelligence comprises at least one processor, wherein the at least one processor trains a chemical accident occurrence prediction model and a chemical accident risk assessment model based on preset data, inputs data related to the current status of a business entity to assess chemical accident risk into the trained chemical accident occurrence prediction model to predict whether a chemical accident will occur, and if the trained chemical accident occurrence prediction model predicts that a chemical accident will occur at the business entity to assess chemical accident risk, determines input information based on the predicted chemical accident, and inputs the determined input information into the trained chemical accident risk assessment model to assess the chemical accident risk for the business entity to assess chemical accident risk.

[0012] In addition, the above at least one processor can process data related to the status of a business to generate first training data, train a chemical accident occurrence prediction model using the generated first training data, process data related to a chemical accident to generate second training data, and train a chemical accident risk assessment model using the generated second training data.

[0013] In addition, the above at least one processor can determine the chemical accident risk to any one of the preset steps.

[0014] According to another embodiment of the present invention, the method comprises the steps of: training a chemical accident occurrence prediction model and a chemical accident risk assessment model based on preset data; inputting data related to the current status of a business entity for assessing chemical accident risk into the trained chemical accident occurrence prediction model to predict whether a chemical accident will occur; determining input information based on the predicted chemical accident when the trained chemical accident occurrence prediction model predicts that a chemical accident will occur at the business entity for assessing chemical accident risk; and inputting the determined input information into the trained chemical accident risk assessment model to assess the chemical accident risk for the business entity for assessing chemical accident risk.

[0015] Additionally, the step of training the chemical accident occurrence prediction model and the chemical accident risk assessment model may include: a step of generating first training data by processing data related to the current status of the business; a step of training the chemical accident occurrence prediction model using the generated first training data; a step of generating second training data by processing data related to chemical accidents; and a step of training the chemical accident risk assessment model using the generated second training data.

[0016] In addition, the step of evaluating the chemical accident risk for a business entity to evaluate the chemical accident risk may include the step of determining the chemical accident risk as one of the preset steps. Effects of the invention

[0017] According to one embodiment, if a chemical accident is predicted to occur at a business for evaluating the risk of chemical accidents, there is an effect of being able to evaluate the risk of the predicted chemical accident.

[0018] In addition, it has the effect of enabling the training of a chemical accident occurrence prediction model and a chemical accident risk assessment model based on pre-set data.

[0019] In addition, it has the effect of generating input information based on predicted chemical accidents. Brief explanation of the drawing

[0020] FIG. 1 is a diagram showing the configuration of a chemical accident risk assessment device using artificial intelligence according to one embodiment. Figure 2 is a diagram showing the configuration of a chemical accident risk assessment model according to one embodiment. FIG. 3 is a flowchart illustrating a chemical accident risk assessment method using artificial intelligence according to one embodiment. Specific details for implementing the invention

[0021] Specific structural or functional descriptions regarding embodiments according to the concept of the present invention disclosed herein are provided merely for the purpose of explaining embodiments according to the concept of the present invention, and embodiments according to the concept of the present invention may be implemented in various forms and are not limited to the embodiments described herein.

[0022] Embodiments according to the concept of the present invention may be subject to various modifications and may take various forms; therefore, embodiments are illustrated in the drawings and described in detail in this specification. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosed forms, and includes all modifications, equivalents, or substitutions that fall within the spirit and scope of the present invention.

[0023] Terms such as "first" or "second" may be used to describe various components, but said components should not be limited by said terms. For the sole purpose of distinguishing one component from another, for example, without departing from the scope of rights according to the concept of the present invention, the first component may be named the second component, and similarly, the second component may be named the first component.

[0024] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. Conversely, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between. Other expressions describing the relationship between components, such as "between" and "exactly between," or "adjacent to" and "directly adjacent to," should be interpreted in the same way.

[0025] The terms used in this specification are used merely to describe specific embodiments and are not intended to limit the invention. The singular expression includes the plural expression unless the context clearly indicates otherwise.

[0026] In this specification, terms such as “comprising” or “having” are intended to specify the existence of the described features, numbers, steps, actions, components, parts, or combinations thereof, and should not be understood as precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0027] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which this invention pertains.

[0028] Terms such as those defined in commonly used dictionaries should be interpreted as having meanings consistent with their meanings in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this specification.

[0029] In the following description, identical identifiers denote identical configurations, and unnecessary redundant explanations and descriptions of known technologies will be omitted.

[0030] Hereinafter, the present invention will be described in detail by explaining preferred embodiments of the present invention with reference to the attached drawings.

[0031] FIG. 1 is a diagram showing the configuration of a chemical accident risk assessment device using artificial intelligence according to one embodiment.

[0032] Referring to FIG. 1, a chemical accident risk assessment device using artificial intelligence according to one embodiment (hereinafter referred to as the "chemical accident risk assessment device") (100) includes a processor (110), a model for predicting whether a chemical accident will occur (120), a model for assessing the risk of a chemical accident (130), an input / output interface module (140), and a memory (150).

[0033] The processor (110), chemical accident occurrence prediction model (120), chemical accident risk assessment model (130), input / output interface module (140), and memory (150) constituting the chemical accident risk assessment device (100) are interconnected and can transmit data to each other.

[0034] The processor (110) can execute programs or instructions stored in memory (150). At this time, an operation program (e.g., OS) for operating the chemical accident risk assessment device (100) may be stored in memory (150).

[0035] The processor (110) can execute a program to manage information about the chemical accident risk assessment device (100).

[0036] The processor (110) can execute a program to manage the operation of the chemical accident risk assessment device (100).

[0037] The processor (110) can execute a program to manage the operation of the input / output interface module (140).

[0038] i) Generation of the first training data

[0039] The processor (110) can obtain data related to the employment and industrial accident insurance enrollment status from the public data portal through the input / output interface module (140).

[0040] The processor (110) can obtain data related to the status of businesses by industry, factory registration status, status of hazardous chemical handling businesses and hazardous chemical accident status in a preset region (e.g., Gyeonggi-do, etc.) through the input / output interface module (140).

[0041] The processor (110) can convert each acquired data into a preset file format (e.g., Comma Separated Values, CSV, hereinafter referred to as "CSV").

[0042] The processor (110) can store each data converted into a preset file format (e.g., Comma Separated Values, CSV, etc.) in memory (150).

[0043] The processor (110) can store each data converted into a CSV file in memory (150).

[0044] The processor (110) can extract primary features (hereinafter referred to as "features") for each data from each data converted into a preset file format (e.g., Comma Separated Values, CSV, etc.). Here, the primary features may be features related to a chemical accident, such as a road name address, a lot number address, a CAS (Chemical Abstract Service Register Number, hereinafter referred to as "CAS number") number, or a physical characteristic, but the primary features It is not limited to this. In addition, the primary feature can be changed through a pre-set method (e.g., user input).

[0045] The processor (110) can delete empty data from which primary features were not extracted in each data.

[0046] The processor (110) can save each of the data, from which empty data with no primary feature extracted has been deleted, as a CSV file.

[0047] The processor (110) can store the column type for each CSV file in memory (150) in a preset file format (e.g., JavaScript Object Notation, JSON, hereinafter referred to as "JSON").

[0048] The processor (110) can extract secondary features from data in which column types for each CSV file are stored in a pre-set file format (e.g., JavaScript Object Notation, JSON, hereinafter referred to as "JSON"). Here, the secondary features may be features that allow for model training (e.g., chemical accident prediction model, chemical accident risk assessment model, etc.) by excluding features that are less related to chemical accidents from features related to chemical accidents (e.g., primary features), such as road name address, lot number address, CAS number, and characteristics, but the secondary features It is not limited to this. In addition, secondary features can be changed through a pre-set method (e.g., user input).

[0049] In order to prevent the process from being killed due to insufficient memory while the task (process) is in progress during data merging, the processor (110) can merge data related to the employment and industrial accident insurance enrollment status obtained from the public data portal with data related to the status of businesses by industry, factory registration status, status of businesses handling hazardous chemicals, and status of hazardous chemical accidents obtained for a pre-set region (e.g., Gyeonggi-do, etc.), by encoding the "address" (e.g., address label encoding) and merging with the corresponding encoded value.

[0050] The processor (110) can store the encoded "address" in an encoder (not shown).

[0051] The processor (110) can merge data based on address, year of investigation, and industry code.

[0052] The processor (110) can extract only the first two digits (medium classification) of the five-digit industry code from the merged data to use as input features for a model (e.g., chemical accident prediction model, chemical accident risk assessment model, etc.).

[0053] The processor (110) can store the merged data as a CSV file in memory (150).

[0054] The processor (110) can store the column types of the CSV file in memory (150) in a preset file format (e.g., JavaScript Object Notation, JSON, hereinafter referred to as "JSON").

[0055] The processor (110) can obtain a CSV file (data) of merged data in which the column type is saved in a preset file format (e.g., JavaScript Object Notation, JSON, hereinafter referred to as "JSON").

[0056] The processor (110) can extract the city / county from the CSV file of the merged data by performing decoding (e.g., address label decoding) on ​​the CSV file of the merged data.

[0057] The processor (110) can process missing values ​​related to city / county extracted from a CSV file of merged data using a preset method. Here, the preset method may be any one of a method of replacing with unused column values, a method of replacing with fixed values, or a method of replacing with random values, but the preset method is not limited to these.

[0058] The processor (110) can set target data based on whether there is a feature to be used as a target.

[0059] The processor (110) can determine that if there is a value of at least one of the features used as a target, it is 1, otherwise it is 0. Here, the features used as a target are features related to whether a chemical accident occurred, and the features used in the current model may include the year and date of the accident, the details of the accident, and the cause of the accident, but the features used as a target are not limited to these.

[0060] The processor (110) can convert the feature (feature data) used as the determined target into label data that can be used in a model (e.g., a chemical accident prediction model, a chemical accident risk assessment model, etc.).

[0061] The processor (110) can extract the finally required features from the generated target data.

[0062] The processor (110) can perform binary encoding or ordinal encoding for each extracted feature and then store it in an encoder (not shown).

[0063] The processor (110) can save to a scaler after performing a preset scaling for each feature. Here, the preset scaling may be any one of no scaling, MinMax, or Quantile Transformation (Uniform PDF), but the preset scaling is not limited to these.

[0064] The processor (110) can store the processed data in memory (150) as a CSV file.

[0065] The processor (110) can store the column types of the CSV file in memory (150) in a preset file format (e.g., JavaScript Object Notation, JSON, hereinafter referred to as "JSON").

[0066] The processor (110) can create a dataloader based on the saved CSV file.

[0067] The processor (110) can create a data loader based on a CSV file that has been saved after preprocessing, such as encoding and scaling, on the extracted features.

[0068] The processor (110) can generate a data loader based on the final preprocessed data.

[0069] The processor (110) can generate first training data based on the saved CSV file.

[0070] ii) Training a model to predict whether a chemical accident will occur

[0071] The processor (110) can train a chemical accident occurrence prediction model (120) using first training data including data related to whether hazardous chemicals are handled and the scale of the generated hazardous chemicals.

[0072] The processor (110) can train a chemical accident occurrence prediction model (120) such that when accident data (e.g., abnormal data, etc.) is input, the difference between the input and the output becomes larger.

[0073] iii) Generation of second training data

[0074] The processor (110) can obtain chemical accident case data through the input / output interface module (140). Here, the chemical accident case data may be chemical accident case study data from the Korea Occupational Safety and Health Agency, a collection of major accident cases, hazardous chemical accident status data from Gyeonggi-do, chemical accident occurrence status data from Suwon / Gunsan-si, etc., but the chemical accident case data is not limited to these.

[0075] The processor (110) can organize and merge the acquired chemical accident case data.

[0076] The processor (110) can organize and merge the acquired chemical accident case data through feature unification, data categorization, numerical unit unification, and input of chemical accident substance properties. Here, feature unification of the chemical accident case data may involve unifying features of the chemical accident, such as the accident area, accident time zone, accident location, accident cause, accident type, and chemical substance CAS number, to match the input shape of the learning model, but the feature unification of the chemical accident case data is not limited to this.

[0077] The processor (110) can obtain data for data verification through the input / output interface module (140). Here, the data for data verification may be the employment industry code of the Ministry of Employment and Labor, the CAS number (Chemical Abstract Service Register Number, hereinafter referred to as "CAS number") of the Korea Research Institute of Bioscience and Biotechnology, and the operational status of the National Research Safety Information System, but the data for data verification is not limited to these.

[0078] The processor (110) can convert the merged data into a CSV file and store it in memory (150).

[0079] The processor (110) can convert the acquired data for verification into a CSV file and store it in memory (150).

[0080] The processor (110) can extract primary features from the merged data in a CSV file (data). Here, the primary features are region, accident season, accident time zone, accident location, accident cause, accident type, employment industry code_medium category, fatalities, injuries, environmental damage, CAS number, properties, color, odor, pH, melting point_Celsius, initial boiling point_Celsius, flash point_Celsius, flammability range_upper limit_percentage, flammability range_lower limit_percentage, vapor pressure_atm_Celsius, solubility_g_ml, solubility_reference temperature_Celsius, vapor density_g_cm³, specific gravity_g_cm³, Log_Kow, autoignition temperature_Celsius, decomposition temperature_Celsius, viscosity_cP, viscosity reference_reference temperature_Celsius, molecular weight, flammability, irritant gas, corrosive gas, toxic gas, corrosive fume, toxic fume, toxic gas upon combustion, amount of accident substance_Kg, number of employees, number of accidents, condition_to_avoid_heat, Condition to Avoid_Spark, Condition to Avoid_Flame, Condition to Avoid_Moisture, Condition to Avoid_Contamination, Condition to Avoid_Other, Substance to Avoid_Water, Substance to Avoid_Oxidizer, Substance to Avoid_Acid, Substance to Avoid_Halode, Substance to Avoid_Metal, Substance to Avoid_Separation Group, Substance to Avoid_Flammable Substance, Substance to Avoid_Reducing Substance, Substance to Avoid_Fuel, Substance to Avoid_Other, etc., but the primary feature is not limited to these. Additionally, the primary feature can be changed through a pre-set method (e.g., user input, etc.).

[0081] The processor (110) can perform data processing on the merged data in a CSV file, such as deleting empty data, categorization, numerical processing, and unit standardization.

[0082] The processor (110) can compare the acquired verification data and perform data processing on the CSV file with the merged data.

[0083] The processor (110) can generate a CSV file for each chemical accident case data for which data processing has been performed, and store the generated CSV file in memory (150).

[0084] The processor (110) can store column types for each CSV file in memory (150) in a preset file format (e.g., JavaScript Object Notation, JSON, hereinafter referred to as "JSON").

[0085] The processor (110) can separate the training / validation / test data set index according to a preset selection method for each chemical accident case data saved as a CSV file.

[0086] The processor (110) can perform primary processing of missing values ​​using a preset selection method. Here, the preset selection method may be any one of simple random selection, random selection while maintaining the ratio of category values, or random selection applying a distribution similar to the feature distribution (e.g., logistic distribution, normal distribution, Pareto distribution, triangular distribution, chi-square distribution, etc.), but the preset selection method is not limited to these.

[0087] The processor (110) uses a weighted average of features to be used on the target data ( Can generate ).

[0088] The processor (110) can calculate a weighted average by assigning a weight of 5 to the deceased and a weight of 1 to the injured using the following [Equation 1].

[0089]

[0090] The processor (110) is a weighted average ( Random values ​​(using a random Poisson distribution with ) applied ) can be produced.

[0091] The processor (110) + Calculated using a random Poisson distribution The risk index can be calculated.

[0092] The processor (110) can categorize the risk level into "1" to "5" by applying a preset interval separation point to the calculated risk index. Here, the preset interval separation point is variable.

[0093] The processor (110) can determine that the calculated risk index is less than 10 as the risk level (risk level) 1.

[0094] The processor (110) can determine that the calculated risk index is 10 or more and less than 15 as the risk level (risk level) 2.

[0095] The processor (110) can determine that the calculated risk index is 15 or higher and less than 20 as the risk level (risk level) 3.

[0096] The processor (110) can determine that the calculated risk index is 20 or higher and less than 30 as the risk level (risk level) 4.

[0097] The processor (110) can determine the risk level (risk level) as 5 when the calculated risk index is 40 or higher.

[0098] The processor (110) can generate target data based on the determined risk level (risk level).

[0099] After categorizing the risk level, the processor (110) can generate target data using a preset encoding (e.g., one-hot encoding).

[0100] The processor (110) can multiply the target data to generate new target data.

[0101] The processor (110) can multiply the input data to generate new input data.

[0102] The processor (110) can amplify the learning / validation data.

[0103] The processor (110) can augment training and validation data of imbalanced labels using a preset method. Here, the preset method may be the Synthetic Minority Oversampling Technique (SMOTE) method, but the preset method is not limited to this.

[0104] The processor (110) can further augment data using a random substitution augmentation method on target data in which imbalanced data has been amplified.

[0105] The processor (110) can restore and extract original data in an amount equal to the size of the target data to be augmented.

[0106] The processor (110) can remove some column values ​​for each data from the training / validation data extracted with replacement. Here, the columns to be excluded can be determined by applying a random number of values ​​less than or equal to a certain value.

[0107] The processor (110) can fill in values ​​by reflecting missing value processing for columns from which values ​​have been excluded.

[0108] The processor (110) can perform secondary processing of missing values ​​in the input data of the training / validation dataset using a preset selection method. Here, the preset selection method may be any one of simple random selection, random selection while maintaining the ratio of categorical values, or random selection applying a distribution similar to the feature distribution (e.g., logistic distribution, normal distribution, Pareto distribution, triangular distribution, chi-square distribution, etc.), but the preset selection method is not limited to these.

[0109] The processor (110) can store each feature in an encoder (not shown) after encoding using a preset method. Here, the preset method may be binary encoding, ordinal encoding, or CatBoost, but is not limited to these.

[0110] The processor (110) can perform CatBoost encoding using the following [Equation 2].

[0111]

[0112] Here, TargetSum_i is the sum of target values ​​corresponding to category i, prior is a constant value using the average of the target values ​​(AllTargetSum / AlldataNum), and FeatureCount_i is the number of data points corresponding to category i.

[0113] When the processor (110) performs encoding using [Equation 2], it is encoded into the average value of the Target value corresponding to the i category variable.

[0114] The processor (110) can encode the CAS number using MATFeaturizer.

[0115] The processor (110) can create a dataloader based on a CSV file that has been saved after preprocessing, such as encoding and scaling, on the extracted features.

[0116] The processor (110) can generate a dataloader based on the final preprocessed data.

[0117] The processor (110) can generate second training data based on the encoded target data.

[0118] iv) Learning the chemical accident risk assessment model

[0119] The processor (110) can train a chemical accident risk assessment model (130) using the generated second training data. Here, the second training data may include information on the types of chemical accidents predicted to occur at a business to assess chemical accident risk, information on the causes of chemical accidents, information on environmental damage that may occur due to chemical accidents, information on toxic substances that may occur due to chemical accidents, information on chemical substances that cause chemical accidents, basic business information (employment industry code_medium classification, number of employees, etc.), accident time and location, damage status, etc., but the information that the second training data may include is not limited to this.

[0120] The processor (110) can generate input information using the generated second training data. Here, the input information may include information on the types of chemical accidents predicted to occur at a business to assess the risk of chemical accidents, information on the causes of chemical accidents, information on environmental damage that may occur due to chemical accidents, information on toxic substances that may occur due to chemical accidents, information on chemical substances that cause chemical accidents, basic business information (employment industry code_medium classification, number of employees, etc.), information related to chemical accidents such as the time and place of the accident, and the status of damage, or information on chemical substances related to chemical accidents, but the information that the input information may include is not limited thereto.

[0121] The processor (110) can train a chemical accident risk assessment model (130) to determine whether the level of chemical accident risk corresponds to any one of the levels from “Level 1” to “Level 5”.

[0122] v) Prediction of whether a chemical accident will occur

[0123] The processor (110) can obtain data related to the status of a business entity for evaluating chemical accident risk through an input / output interface module (140). Here, the data related to the status of a business entity for evaluating chemical accident risk may include data related to the employment and industrial accident insurance coverage status of the business entity for evaluating accident risk, data related to the status of businesses by industry in the region where the business entity for evaluating accident risk exists, data related to the factory registration status in the region where the business entity for evaluating accident risk exists, data related to the status of hazardous chemical handling businesses in the region where the business entity for evaluating accident risk exists, and data related to the status of hazardous chemical accidents in the region where the business entity for evaluating accident risk exists, but the data that may be included in the data related to the status of a business entity for evaluating chemical accident risk is not limited to these.

[0124] The processor (110) can input data related to the current status of the business to assess the risk of chemical accidents into a chemical accident occurrence prediction model (120) that has completed learning.

[0125] The processor (110) can obtain the prediction result regarding whether a chemical accident has occurred, which is output by the chemical accident occurrence prediction model (120) that has completed learning.

[0126] vi) Chemical accident risk assessment

[0127] The processor (110) can determine input information based on the predicted chemical accident when the trained chemical accident occurrence prediction model (120) predicts that a chemical accident will occur at a business to evaluate the risk of a chemical accident. Here, the input information may include information on the type of chemical accident predicted to occur at the business to evaluate the risk of a chemical accident, information on the cause of the chemical accident, information on environmental damage that may occur due to the chemical accident, information on toxic substances that may occur due to the chemical accident, information on chemical substances that cause the chemical accident, basic business information (employment industry code_medium classification, number of employees, etc.), information related to the chemical accident such as the time and place of the accident, and the status of damage, but the information that the input information may include is not limited to this.

[0128] The processor (110) can evaluate the chemical accident risk for a business entity by inputting the input information into a chemical accident risk assessment model (130) that has completed learning.

[0129] The processor (110) can determine the risk of a chemical accident assessed by the learned chemical accident risk assessment model (130) to be one of the preset stages (e.g., “Stage 1” to “Stage 5”, etc.).

[0130] The processor (110) can evaluate the chemical accident risk for a business entity to evaluate the chemical accident risk by determining the chemical accident risk evaluated by the learned chemical accident risk assessment model (130) to one of the preset stages (e.g., “Stage 1” to “Stage 5”).

[0131] The processor (110) can output the chemical accident risk assessment results for a business entity to assess the chemical accident risk determined through the input / output interface module (140).

[0132] The chemical accident occurrence prediction model (120) may be an RSR autoencoder that incorporates a Robust Subspace Recovery (hereinafter "RSR") layer into the Latent (Bottleneck) portion of the autoencoder. Here, the RSR layer can perform the role of mapping embedding vectors to low-dimensional vectors to provide emphasis on singular values.

[0133] A chemical accident prediction model (120) can be trained to produce an output of the same shape as the input.

[0134] A chemical accident prediction model (120) can be trained only with normal data (e.g., data unrelated to accident occurrence).

[0135] The chemical accident occurrence prediction model (120) can output a prediction result regarding whether a chemical accident will occur.

[0136] The chemical accident risk assessment model (130) may include a deep neural network (DNN) and a chemical feature extraction network (e.g., MAT networks, etc.).

[0137] A chemical accident risk assessment model (130) can combine the output of a deep neural network (DNN) and the output of a chemical feature extraction network (e.g., MAT networks, etc.).

[0138] A chemical accident risk assessment model (130) can be trained to determine whether the level of chemical accident risk corresponds to one of the levels “1” to “5” based on the output of a combined deep neural network (DNN) and the output of a chemical substance feature extraction network (e.g., MAT networks, etc.).

[0139] The chemical accident risk assessment model (130) can output the determined level of chemical accident risk.

[0140] The input / output interface module (140) can be communicated to an external device (e.g., a public data portal server, etc.) through a network.

[0141] The input / output interface module (140) can transmit data to an external device (e.g., a data storage server, etc.) through a network.

[0142] The input / output interface module (140) can receive data transmitted by an external device (e.g., public data portal server, USB, HDD, SSD, etc.) through a network.

[0143] The input / output interface module (140) can output data generated by the processor (110).

[0144] The input / output interface module (140) can output training data generated by the processor (110).

[0145] The input / output interface module (140) can display the training data generated by the processor (110).

[0146] The input / output interface module (140) can display the prediction result regarding whether a chemical accident will occur, which is output by the chemical accident occurrence prediction model.

[0147] The input / output interface module (140) can display the level of chemical accident risk output by the chemical accident risk assessment model.

[0148] The input / output interface module (140) can obtain user input.

[0149] The input / output interface module (140) can be provided as an integral part of the chemical accident risk assessment device (100).

[0150] The input / output interface module (140) can be provided separately from the chemical accident risk assessment device (100).

[0151] The input / output interface module (140) may be a separate device that is communicationally connected to the chemical accident risk assessment device (100).

[0152] The input / output interface module (140) may include a port (e.g., a USB port) for connecting to an external device.

[0153] The input / output interface module (140) may include a monitor, touchscreen, mouse, electronic pen, microphone, keyboard, speaker, earphones, headphones, or touchpad.

[0154] The memory (150) can store data received by the input / output interface module (140).

[0155] The memory (150) can store user input obtained by the input / output interface module (140).

[0156] The memory (150) can store data transmitted by the input / output interface module (140).

[0157] The memory (150) can store the CSV file generated by the processor (110).

[0158] The memory (150) can store the training data generated by the processor (110).

[0159] The memory (150) can store features determined by the processor (110).

[0160] The memory (150) can store features extracted by the processor (110).

[0161] Memory (150) can store column types as JSON for each CSV file.

[0162] The memory (150) can store the learning results of the chemical accident prediction model.

[0163] The memory (150) can store the prediction results of the chemical accident occurrence prediction model.

[0164] The memory (150) can store target data generated by the processor (110).

[0165] The memory (150) can store the encoder and scaler of the features generated by the processor (110).

[0166] The memory (150) can store data augmented by the processor (110).

[0167] The memory (150) can store the learning results of the chemical accident risk assessment model.

[0168] The memory (150) can store the level of chemical accident risk determined by the chemical accident risk assessment model.

[0169] It is obvious to those skilled in the art to which this invention pertains that the term 'device or module' used herein refers to a logical constituent unit and is not necessarily a physically distinct component.

[0170] Figure 2 is a diagram showing the configuration of a chemical accident risk assessment model according to one embodiment.

[0171] Referring to FIG. 2, the chemical accident risk assessment model (200) can receive chemical accident-related information (211) and chemical accident-related chemical substance information (212) as input information (210) and output a chemical accident risk level (270). Here, the chemical accident risk level (270) may be any one of "Level 1" to "Level 5", but the chemical accident risk level (250) is not limited to this.

[0172] The chemical accident risk assessment model (200) may include a deep neural network (DNN) (220), a MATFeaturizer (230), a chemical substance feature extraction network (240), and a classification model (Multi-class Classifier) ​​(260).

[0173] The deep neural network (DNN) (220) included in the chemical accident risk assessment model (200) can be constructed by stacking multiple dense blocks (221).

[0174] Each Dense Block (222) constituting the Deep Neural Network (DNN) (220) generates a mapping between input and output, and consists of a Dense Layer (222a) in which each connection has a weight and the node can have a bias, a Layer Normalization (222b) in which the mean / variance of feature values ​​within each data of each Layer is calculated and normalized based on this, and a SELU (222c) in which self-normalization is performed so that the mean and variance of the input values ​​do not become extremely large or small.

[0175] A deep neural network (DNN) (220) can prevent overfitting with Dropout (223), and in this case, normalization and activation processing may not be performed.

[0176] A deep neural network (DNN) (220) can receive input features other than the CAS number (e.g., chemical accident-related information (211), etc.) and generate a deep neural network output (DNN Output) (224). Here, the deep neural network output (DNN Output) (224) may be a vector, but the deep neural network output (DNN Output) (224) is not limited to this.

[0177] The MATFeaturizer (230) included in the chemical accident risk assessment model (200) can obtain an embedding vector (numerical vector) by encoding the input CAS number (e.g., chemical accident-related chemical substance information (212), etc.).

[0178] The chemical substance feature extraction network (240) included in the chemical accident risk assessment model (200) may be a Molecule Attention Transformer (MAT) network, but the chemical substance feature extraction network (240) is not limited to this.

[0179] The chemical accident risk assessment model (200) can input the embedding vector (numerical vector) obtained from the MATFeaturizer (230) into the chemical substance feature extraction network (240).

[0180] The chemical accident risk assessment model (200) can combine (250) the deep neural network output (224) generated by the deep neural network (DNN) (220) and the numerical vector generated by the chemical feature extraction network (240).

[0181] The chemical accident risk assessment model (200) can assess the chemical accident risk by combining (250) the deep neural network output (DNN Output) (224) generated by the deep neural network (DNN) (220) and the numerical vector generated by the chemical substance feature extraction network (240) to reflect chemical substance information related to the chemical accident in the chemical accident risk assessment.

[0182] The chemical accident risk assessment model (200) can input a vector formed by concatenating (250) the deep neural network output (DNN Output) (224) and the numerical vector generated by the chemical substance feature extraction network (240) into a classification model (Multi-class Classifier) ​​(260).

[0183] The chemical accident risk assessment model (200) can improve the performance of the classification model (Multi-class Classifier) ​​(260) in determining the chemical accident risk level by inputting a vector formed by combining (250) the deep neural network output (DNN Output) (224) and the numerical vector generated by the chemical substance feature extraction network (240) into the classification model (Multi-class Classifier) ​​(260).

[0184] The chemical accident risk assessment model (200) can determine which level of chemical accident risk corresponds to "Level 1" to "Level 5" by using a classification model (Multi-class Classifier) ​​(260) that distinguishes multiple classes.

[0185] The chemical accident risk assessment model (200) can output a determined chemical accident risk level (270).

[0186] FIG. 3 is a flowchart illustrating a chemical accident risk assessment method using artificial intelligence according to one embodiment.

[0187] Referring to FIG. 3, the method for evaluating chemical accident risk using artificial intelligence includes the following steps: a step (S300) in which a chemical accident risk evaluation device trains a chemical accident prediction model and a chemical accident risk evaluation model based on pre-set data; a step (S310) in which the chemical accident risk evaluation device inputs data related to the current status of a business to evaluate chemical accident risk into the trained chemical accident prediction model to predict whether a chemical accident will occur; a step (S320) in which, if the chemical accident risk evaluation device predicts that a chemical accident will occur at a business to evaluate chemical accident risk through the trained chemical accident prediction model, the chemical accident risk evaluation device determines input data based on the predicted chemical accident; a step (S330) in which the chemical accident risk evaluation device inputs the determined input data into the trained chemical accident risk evaluation model to evaluate the chemical accident risk for a business to evaluate chemical accident risk; and a step (S340) in which the chemical accident risk evaluation device outputs the evaluation result.

[0188] In step S300, the chemical accident risk assessment device can obtain data related to the status of employment and industrial accident insurance enrollment from the public data portal.

[0189] In step S300, the chemical accident risk assessment device can acquire data related to the status of businesses by industry, factory registration status, status of hazardous chemical handling operators, and hazardous chemical accident status in a pre-set region (e.g., Gyeonggi-do, etc.).

[0190] In step S300, the chemical accident risk assessment device can convert each acquired data into a preset file format (e.g., Comma Separated Values, CSV, hereinafter referred to as "CSV").

[0191] In step S300, the chemical accident risk assessment device can extract primary features (hereinafter referred to as "features") for each data from each data converted into a preset file format (e.g., Comma Separated Values, CSV, etc.). Here, the primary features may be features related to chemical accidents, such as road name addresses, lot number addresses, CAS (Chemical Abstract Service Register Number, hereinafter referred to as "CAS number") numbers, and properties, but the primary features are not limited to these. Additionally, the primary features may be modified through a preset method (e.g., user input, etc.).

[0192] In step S300, the chemical accident risk assessment device can delete empty data from which primary features were not extracted in each data.

[0193] In step S300, the chemical accident risk assessment device can save each of the data, from which empty data with primary features not extracted has been deleted, as a CSV file.

[0194] In step S300, the chemical accident risk assessment device can store the column type for each CSV file in memory (150) in a preset file format (e.g., JavaScript Object Notation, JSON, hereinafter referred to as "JSON").

[0195] In step S300, the chemical accident risk assessment device can extract secondary features from data in which the column type for each CSV file is saved in a pre-set file format (e.g., JavaScript Object Notation, JSON, hereinafter referred to as "JSON"). Here, the secondary features may be features that allow for the training of a model (e.g., chemical accident prediction model, chemical accident risk assessment model, etc.) by excluding features that are less related to chemical accidents from features related to chemical accidents (e.g., primary features), such as road name address, lot number address, CAS number, and characteristics, but the secondary features are not limited to this. Additionally, the secondary features may be modified through a pre-set method (e.g., user input, etc.).

[0196] In step S300, to prevent the chemical accident risk assessment device from being killed due to insufficient memory during the operation (process) when merging data related to the employment and industrial accident insurance enrollment status obtained from the public data portal with data related to the status of businesses by industry, factory registration status, status of hazardous chemical handling businesses, and hazardous chemical accident status obtained for a pre-set region (e.g., Gyeonggi-do, etc.), the device may encode the "address" (e.g., address label encoding) and merge using the corresponding encoded value.

[0197] In step S300, the chemical accident risk assessment device can store the encoded "address" in the encoder.

[0198] In step S300, the chemical accident risk assessment device can merge data based on address, investigation year, and industry code.

[0199] In step S300, the chemical accident risk assessment device can extract only the first two digits (medium classification) of the five-digit industry code from the merged data to use as input features for a model (e.g., chemical accident prediction model, chemical accident risk assessment model, etc.).

[0200] In step S300, the chemical accident risk assessment device can obtain a CSV file (data) of merged data in which the column types are saved in a preset file format (e.g., JavaScript Object Notation, JSON, hereinafter referred to as "JSON").

[0201] In step S300, the chemical accident risk assessment device can extract the city / county from the CSV file of the merged data by performing decoding (e.g., address label decoding) on ​​the CSV file of the merged data.

[0202] In step S300, the chemical accident risk assessment device can process missing values ​​related to cities and counties extracted from the CSV file of merged data using a preset method. Here, the preset method may be any one of replacing with unused column values, replacing with fixed values, or replacing with random values, but the preset method is not limited to these.

[0203] In step S300, the chemical accident risk assessment device can set target data based on whether there is a feature to be used as a target.

[0204] In step S300, the chemical accident risk assessment device may determine that if at least one of the features used as a target has a value, it is 1, otherwise it is 0. Here, the features used as targets are features related to whether a chemical accident has occurred, and the features used in the current model may include the year and date of the accident, the details of the accident, and the cause of the accident, but the features used as targets are not limited to these.

[0205] In step S300, the chemical accident risk assessment device can convert the feature (feature data) used as the determined target into label data that can be used in a model (e.g., a chemical accident prediction model, a chemical accident risk assessment model, etc.).

[0206] In step S300, the chemical accident risk assessment device can extract the finally necessary features from the generated target data.

[0207] In step S300, the chemical accident risk assessment device may save to the scaler after performing preset scaling for each feature. Here, the preset scaling may be any one of no scaling, MinMax, or Quantile Transformation (Uniform PDF), but the preset scaling is not limited to these.

[0208] In step S300, the chemical accident risk assessment device can generate a data loader based on the final preprocessed data.

[0209] In step S300, the chemical accident risk assessment device can generate first training data based on a CSV file.

[0210] In step S300, the chemical accident risk assessment device can train a chemical accident occurrence prediction model using first training data that includes data related to whether the generated hazardous chemical substance is handled and the scale of the damage.

[0211] In step S300, the chemical accident risk assessment device can train a chemical accident occurrence prediction model such that the difference between the Input and Output increases when accident data (e.g., abnormal data, etc.) is input.

[0212] In step S300, the chemical accident risk assessment device can acquire chemical accident case data. Here, the chemical accident case data may include chemical accident case study data from the Korea Occupational Safety and Health Agency, major accident casebooks, hazardous chemical accident status data from Gyeonggi-do, and chemical accident occurrence status data from Suwon / Gunsan-si, but the chemical accident case data is not limited to these.

[0213] In step S300, the chemical accident risk assessment device can organize and merge the acquired chemical accident case data through feature unification, data categorization, numerical unit unification, and inputting the physical properties of the chemical accident substances. Here, feature unification of the chemical accident case data may involve unifying features related to the chemical accident, such as the accident region, accident time zone, accident location, accident cause, accident type, and chemical substance CAS number, to match the input shape of the learning model, but the feature unification of the chemical accident case data is not limited to this.

[0214] In step S300, the chemical accident risk assessment device may acquire data for data verification. Here, the data for data verification may include the employment industry code of the Ministry of Employment and Labor and the CAS number (Chemical Abstract Service Register Number, hereinafter referred to as "CAS number") of the Korea Research Institute of Bioscience and Biotechnology, which may be the operational status of the National Research Safety Information System, but the data for data verification is not limited to these.

[0215] In step S300, the chemical accident risk assessment device can extract primary features from the merged data in a CSV file (data). Here, the primary features are region, accident season, accident time zone, accident location, accident cause, accident type, employment industry code_medium category, fatalities, injuries, environmental damage, CAS number, properties, color, odor, pH, melting point_Celsius, initial boiling point_Celsius, flash point_Celsius, flammability range_upper limit_percentage, flammability range_lower limit_percentage, vapor pressure_atm_Celsius, solubility_g_ml, solubility_reference temperature_Celsius, vapor density_g_cm³, specific gravity_g_cm³, Log_Kow, autoignition temperature_Celsius, decomposition temperature_Celsius, viscosity_cP, viscosity reference_reference temperature_Celsius, molecular weight, flammability, irritant gas, corrosive gas, toxic gas, corrosive fume, toxic fume, toxic gas upon combustion, amount of accident substance_Kg, number of employees, number of accidents, condition_to_avoid_heat, Condition to Avoid_Spark, Condition to Avoid_Flame, Condition to Avoid_Moisture, Condition to Avoid_Contamination, Condition to Avoid_Other, Substance to Avoid_Water, Substance to Avoid_Oxidizer, Substance to Avoid_Acid, Substance to Avoid_Halode, Substance to Avoid_Metal, Substance to Avoid_Separation Group, Substance to Avoid_Flammable Substance, Substance to Avoid_Reducing Substance, Substance to Avoid_Fuel, Substance to Avoid_Other, etc., but the primary feature is not limited to these. Additionally, the primary feature can be changed through a pre-set method (e.g., user input, etc.).

[0216] In step S300, the chemical accident risk assessment device can perform data processing on the merged data in a CSV file, such as deleting empty data, categorization, numerical processing, and unit standardization.

[0217] In step S300, the chemical accident risk assessment device compares the acquired verification data and can perform data processing on the CSV file with the merged data.

[0218] In step S300, the chemical accident risk assessment device can separate the training / validation / test data set index for each chemical accident case data stored as a CSV file according to a preset selection method.

[0219] In step S300, the chemical accident risk assessment device can perform primary processing of missing values ​​using a preset selection method. Here, the preset selection method may be any one of simple random selection, random selection while maintaining the proportion of categorical values, or random selection applying a distribution similar to the feature distribution (e.g., logistic distribution, normal distribution, Pareto distribution, triangular distribution, chi-square distribution, etc.), but the preset selection method is not limited to these.

[0220] In step S300, the chemical accident risk assessment device can calculate a risk index.

[0221] In step S300, the chemical accident risk assessment device can categorize risk levels into "1" to "5" by applying preset interval separation points to the calculated risk index. Here, the preset interval separation points are variable.

[0222] In step S300, the chemical accident risk assessment device can categorize the risk level and then generate target data using a preset encoding (e.g., one-hot encoding).

[0223] In step S300, the chemical accident risk assessment device can amplify target data to generate new target data.

[0224] In step S300, the chemical accident risk assessment device can multiply input data to generate new input data.

[0225] In step S300, the chemical accident risk assessment device can amplify learning / validation data.

[0226] In step S300, the chemical accident risk assessment device can augment learning and validation data of imbalanced labels using a preset method. Here, the preset method may be the Synthetic Minority Oversampling Technique (SMOTE) method, but is not limited to this.

[0227] In step S300, the chemical accident risk assessment device can further augment data using a random substitution augmentation method for target data in which imbalanced data has been amplified.

[0228] In step S300, the chemical accident risk assessment device can restore and extract original data up to the size of the target data to be augmented.

[0229] In step S300, the chemical accident risk assessment device can remove some column values ​​for each data point from the training / validation data extracted with replacement. Here, the columns to be excluded can be determined by applying a random number of values ​​below a certain threshold.

[0230] In step S300, the chemical accident risk assessment device can fill in values ​​by reflecting missing value processing for columns where values ​​have been excluded.

[0231] In step S300, the chemical accident risk assessment device can perform secondary processing of missing values ​​in the input data of the training / validation dataset using a preset selection method. Here, the preset selection method may be any one of simple random selection, random selection while maintaining the proportion of categorical values, or random selection applying a distribution similar to the feature distribution (e.g., logistic distribution, normal distribution, Pareto distribution, triangular distribution, chi-square distribution, etc.), but the preset selection method is not limited to these.

[0232] In step S300, the chemical accident risk assessment device may encode each feature using a preset method and store it in an encoder (not shown). Here, the preset method may be binary encoding, ordinal encoding, or CatBoost, but is not limited to these.

[0233] In step S300, the chemical accident risk assessment device can encode the CAS number using MATFeaturizer.

[0234] In step S300, the chemical accident risk assessment device can generate a dataloader based on the final preprocessed data.

[0235] In step S300, the chemical accident risk assessment device can generate second training data based on encoded target data.

[0236] In step S300, the chemical accident risk assessment device can train the chemical accident risk assessment model (130) using the second training data generated.

[0237] In step S300, the chemical accident risk assessment device can train a chemical accident risk assessment model to determine whether the level of chemical accident risk corresponds to any one of "Level 1" to "Level 5".

[0238] In step S310, the chemical accident risk assessment device may acquire data related to the current status of the business entity for assessing chemical accident risk. Here, the data related to the current status of the business entity for assessing chemical accident risk may include data related to the employment and industrial accident insurance coverage status of the business entity for assessing accident risk, data related to the status of business entities by industry in the region where the business entity for assessing accident risk is located, data related to the factory registration status in the region where the business entity for assessing accident risk is located, data related to the status of hazardous chemical substance handlers in the region where the business entity for assessing accident risk is located, and data related to the status of hazardous chemical substance accidents in the region where the business entity for assessing accident risk is located, but the data that may be included in the data related to the current status of the business entity for assessing chemical accident risk is not limited to these.

[0239] In step S310, the chemical accident risk assessment device can input data related to the current status of the business to assess the chemical accident risk into a trained chemical accident occurrence prediction model.

[0240] In step S310, the chemical accident risk assessment device can obtain the prediction result regarding whether a chemical accident will occur, which is output by the chemical accident occurrence prediction model that has completed training.

[0241] In step S320, the chemical accident risk assessment device can determine input information based on the predicted chemical accident when the chemical accident occurrence prediction model (120), which has completed learning, predicts that a chemical accident will occur at the business to assess the chemical accident risk. Here, the input information may include information regarding the type of chemical accident predicted to occur at the business to assess the chemical accident risk, information regarding the cause of the chemical accident, information regarding environmental damage that may occur due to the chemical accident, information regarding toxic substances that may occur due to the chemical accident, information regarding chemical substances that cause the chemical accident, basic business information (employment industry code_medium classification, number of employees, etc.), the time and place of the accident, damage status, etc., or chemical accident-related information or chemical substance information related to the chemical accident, but the information that the input information may include is not limited to these.

[0242] In step S330, the chemical accident risk assessment device can assess the chemical accident risk for a business entity by inputting the input information into a chemical accident risk assessment model that has completed learning.

[0243] In step S330, the chemical accident risk assessment device can determine the risk of a chemical accident assessed by a trained chemical accident risk assessment model to be one of the preset stages (e.g., "Stage 1" to "Stage 5", etc.).

[0244] In step S340, the chemical accident risk assessment device can output the chemical accident risk assessment results for a business entity to assess the chemical accident risk.

[0245] In the foregoing, although all components constituting the embodiments of the present invention have been described as being combined into one or operating together, the present invention is not necessarily limited to such embodiments. That is, within the scope of the purpose of the present invention, all components may be selectively combined into at least one to operate.

[0246] In addition, while all such components may each be implemented as a single independent piece of hardware, some or all of the components may be optionally combined to be implemented as a computer program having a program module that performs some or all of the combined functions on one or more pieces of hardware. The codes and code segments constituting the computer program will be readily deducible by those skilled in the art of the present invention.

[0247] An embodiment of the present invention can be implemented by storing such a computer program on a computer-readable storage medium and reading and executing it by a computer. The storage medium for the computer program may include a magnetic recording medium, an optical recording medium, etc.

[0248] Furthermore, terms such as "include," "constitute," or "have" as described above, unless specifically stated otherwise, imply that the relevant component may be inherent; therefore, they should be interpreted as allowing for the inclusion of additional components rather than excluding them.

[0249] All terms, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which this invention pertains, unless otherwise defined. Commonly used terms, such as those defined in advance, should be interpreted as consistent with their meaning in the context of the relevant technology and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this invention.

[0250] The above description is merely an illustrative explanation of the technical concept of the present invention, and those skilled in the art to which the present invention pertains will be able to make various modifications and variations within the scope of the essential characteristics of the present invention.

[0251] Accordingly, the embodiments disclosed in this invention are intended to illustrate, not limit, the technical concept of the invention, and the scope of the technical concept of the invention is not limited by these embodiments. The scope of protection of this invention shall be interpreted by the claims below, and all technical concepts within an equivalent scope shall be interpreted as being included within the scope of rights of this invention. Explanation of the symbols

[0252] 100... Chemical accident risk assessment device

Claims

Claim 1 A chemical accident risk assessment device using artificial intelligence comprises at least one processor, wherein the at least one processor trains a chemical accident occurrence prediction model and a chemical accident risk assessment model based on preset data, inputs data related to the current status of a business entity to assess chemical accident risk into the trained chemical accident occurrence prediction model to predict whether a chemical accident will occur, and if the trained chemical accident occurrence prediction model predicts that a chemical accident will occur at the business entity to assess chemical accident risk, determines input information based on the predicted chemical accident, and inputs the determined input information into the trained chemical accident risk assessment model to assess the chemical accident risk for the business entity to assess chemical accident risk, wherein the trained chemical accident risk assessment model comprises a Deep Neural Network (DNN), a MATFeaturizer, a chemical substance feature extraction network, and a classification model (Multi-class Classifier), wherein the Deep Neural Network (DNN) receives chemical accident-related information as input and generates a Deep Neural Network Output (DNN Output) which is a vector, and The MATFeaturizer obtains an embedding vector by encoding chemical substance information related to a chemical accident, the chemical substance feature extraction network generates a numeric vector based on the embedding vector obtained by the MATFeaturizer, and the chemical accident risk assessment model generates a vector by concatenating the deep neural network output (DNN Output) and the numeric vector generated by the chemical substance feature extraction network.A chemical accident risk assessment device using artificial intelligence that determines the chemical accident risk level by inputting the vector generated by concatenating the deep neural network output (DNN Output) and the numerical vector generated by the chemical substance feature extraction network into the classification model (Multi-class Classifier) ​​to improve the performance of determining the chemical accident risk level by reflecting chemical substance information related to the chemical accident in the chemical accident risk assessment. Claim 2 A chemical accident risk assessment device using artificial intelligence according to claim 1, wherein at least one processor processes data related to the status of a business entity to generate first training data, trains a chemical accident occurrence prediction model using the generated first training data, processes data related to a chemical accident to generate second training data, and trains a chemical accident risk assessment model using the generated second training data. Claim 3 In claim 1, the at least one processor is a chemical accident risk assessment device using artificial intelligence that determines the chemical accident risk to any one of preset steps. Claim 4 A step of training a chemical accident occurrence prediction model and a chemical accident risk assessment model based on pre-set data; a step of inputting data related to the current status of a business entity for assessing chemical accident risk into the trained chemical accident occurrence prediction model to predict whether a chemical accident will occur; and a step of determining input information based on the predicted chemical accident if the trained chemical accident occurrence prediction model predicts that a chemical accident will occur at the business entity for assessing chemical accident risk.The method includes the step of evaluating the chemical accident risk for a business entity by inputting the determined input information into the chemical accident risk assessment model that has completed training, wherein the chemical accident risk assessment model that has completed training includes a Deep Neural Network (DNN), a MATFeaturizer, a chemical substance feature extraction network, and a classification model (Multi-class Classifier); wherein the Deep Neural Network (DNN) receives chemical accident-related information as input and generates a Deep Neural Network Output (DNN Output) which is a vector; the MATFeaturizer encodes chemical substance information related to the chemical accident to obtain an embedding vector; the chemical substance feature extraction network generates a numeric vector based on the embedding vector obtained by the MATFeaturizer; the chemical accident risk assessment model generates a vector by concatenating the Deep Neural Network Output (DNN Output) and the numeric vector generated by the chemical substance feature extraction network; and the chemical accident risk assessment model reflects chemical substance information related to the chemical accident in the chemical accident risk assessment to evaluate the chemical accident A method for assessing chemical accident risk using artificial intelligence, wherein the method inputs the vector generated by concatenating the deep neural network output (DNN Output) and the numerical vector generated by the chemical substance feature extraction network into the classification model (Multi-class Classifier) ​​to determine the chemical accident risk level in order to improve the performance of determining the risk level. Claim 5 In claim 4, the step of training the chemical accident occurrence prediction model and the chemical accident risk assessment model comprises: a step of generating first training data by processing data related to the current status of a business entity; a step of training the chemical accident occurrence prediction model using the generated first training data; a step of generating second training data by processing data related to chemical accidents; and a step of training the chemical accident risk assessment model using the generated second training data. Claim 6 In paragraph 4, the step of evaluating the chemical accident risk for a business entity to evaluate the chemical accident risk comprises a step of determining the chemical accident risk as one of the preset steps, a method for evaluating chemical accident risk using artificial intelligence.

Images