Event-based diaster occurrence detection method and system using deep learning

Abstract

The present invention relates to an event-based disaster occurrence detection method and system using deep learning. More specifically, the present invention relates to a method by which an information system collects social data distributed through a social network (SNS) and then analyzes the social data through deep learning to detect and provide the risk of occurrence of a natural disaster or a social disaster, and the information system using same. According to the present invention, morphemes of social data having an unstructured data structure are analyzed to extract a point of interest (POI), an object, and an event and prepare structural data, and then situation information is reconstructed through a combination of the structural data to find action events for each POI included in the social data, and the occurrence or sign of a disaster can be found through temporal and spatial correlation analysis of the action events. Therefore, a sign or occurrence of a disaster can be determined in the early stages for each POI of a country.

Description

Event-based disaster detection method and system using deep learning

[0001] The present invention relates to a method and system for event-based disaster detection using deep learning. More specifically, the present invention relates to a method and an information system utilizing such a method, wherein an information system collects social data distributed through social networks (SNS) and analyzes the social data using deep learning to detect and provide information on the risk of natural or social disasters. The present invention analyzes the morphemes of social data with an unstructured data structure to extract points of interest (POIs), objects, and events, transforming them into structured data. These are then combined to reconstruct contextual information. This method identifies behavioral events for each POI contained in the social data, and through temporal and spatial correlation analysis, detects the occurrence or signs of a disaster. Therefore, it enables early detection of signs or the presence of a disaster at each POI of a country. Furthermore, even when disaster risks are detected through social data, the authenticity of the data is verified through a fake news verification model, thereby preventing social chaos caused by fake news. In addition, it provides a sensor integration verification model, so that the authenticity of disaster risks detected from social data can be verified using sensor data collected at the relevant location, thereby improving the reliability of disaster occurrence detection results.

[0002] Disasters come in various forms, causing massive damage to life and property. Large-scale disasters can severely impact national finances and spark social unrest, posing significant risks to the nation. While natural disasters, such as earthquakes, floods, landslides, and inundation, are caused by natural phenomena, social disasters can also arise from human error or social or institutional deficiencies. These include fires, gas leaks, facility collapses, major traffic accidents, and crushing accidents in crowded areas.

[0003] Once these large-scale disasters occur, they cause widespread loss of life and property. In particular, the damage caused by large-scale disasters doesn't stop at the moment of occurrence. Most large-scale disasters, if left unchecked, not only spread over time, but also trigger secondary and tertiary disasters, ultimately leading to uncontrollable consequences and increasing damage. Therefore, while disaster prevention is crucial, rapid early detection is paramount.

[0004] To facilitate disaster prevention activities and early detection and suppression, the government is working to establish systems, correct practices, and develop technologies for these activities. The private sector is also actively participating in this technological development. In particular, research and development efforts are underway across various fields to apply ICT technologies to disaster prevention and other areas. As part of the development of ICT-based disaster response technologies, research is also underway on smart disaster safety systems that can be applied across all areas of disaster prevention, prediction, preparedness, response, and recovery.

[0005] Despite these studies and efforts, the problem is that the likelihood of large-scale disasters occurring in our society remains high, and they are occurring frequently. In particular, specific locations, such as shopping malls, sports stadiums, and other crowded multi-use facilities, as well as large-scale event venues for special events, as well as facilities associated with mass transportation like subway stations, railway stations, and airports, are highly susceptible to large-scale disasters for various reasons. However, due to these various factors, early detection of signs or events of disasters in such locations is extremely difficult.

[0006] In particular, megacities, such as special cities and metropolitan cities, are overpopulated, and even downtown streets are usually densely populated with people and vehicles. This creates numerous potential disaster-prone areas. Furthermore, once a disaster occurs, the dense population and vehicle density can significantly increase the scope and speed of damage. Fortunately, these potential disaster-prone areas are equipped with various state-of-the-art surveillance equipment, such as sensors and CCTV. (In this invention, these potential disaster-prone areas are referred to as "Points of Interest.")

[0007] However, despite the state-of-the-art surveillance equipment at these points of interest, institutional, systemic, and environmental deficiencies frequently hinder timely and accurate disaster detection. This is due to human factors, such as a shortage of surveillance personnel or a lack of expertise, as well as environmental issues, such as signal interruptions due to defective or malfunctioning surveillance sensors. Furthermore, the fundamental limitations of surveillance systems, which cannot accurately detect disasters based solely on fragmentary analysis of sensor signals, can be attributed to this. Therefore, with the current surveillance system, it is practically impossible to predict the signs of a disaster before it occurs, and even rapid early detection of a disaster is extremely challenging.

[0008] Meanwhile, social networks play a role in disseminating disaster information in various forms during natural or social disasters. Especially during large-scale disasters, social networks served as unofficial channels for spreading information, playing a more crucial role in crisis communication than traditional media or government disaster dissemination systems. For example, in Japan, when an earthquake struck, information related to the earthquake rapidly spread, collecting disaster information and utilizing it for disaster preparedness, monitoring, and response information.

[0009] The present invention, which was created to solve the above-described problems, aims to provide a method and system that can detect signs of a natural disaster or social disaster that may occur at a national point of interest in advance by analyzing social data distributed through a social network through deep learning, as well as detect the occurrence of a disaster quickly and early in the event of one.

[0010] The present invention also aims to provide a method and system for preventing the spread of a disaster by allowing an information system to analyze social data collected around points of interest in time and space using deep learning, thereby finding phenomena related to a disaster and reporting them as disaster information, thereby enabling the relevant authorities to be aware of the fact of a disaster or signs of a disaster occurring in the early stages or before a disaster occurs, and thereby enabling immediate response and action accordingly.

[0011] Another object of the present invention is to provide a method and system capable of preventing false detections and side effects caused by fake news by providing a method for identifying social data containing fake news in the process of detecting signs of disaster by analyzing social data through deep learning.

[0012] Another object of the present invention is to provide a method for increasing the reliability of detection by verifying disaster-related social data generated at a point of interest using data from various sensors collected at the point of interest in the process of detecting signs of a disaster by analyzing social data through deep learning, thereby improving the reliability of detection.

[0013] The technical problems to be solved by the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned can be clearly understood by a person having ordinary skill in the technical field to which the present invention belongs from the description below.

[0014] The present invention, in order to achieve the above-described object, provides a method for detecting the risk of occurrence of a natural disaster or a social disaster by collecting and analyzing event information distributed through a social network (SNS) using a deep learning program, the method comprising: a social data collection step of collecting social data distributed through one or more social networks; a tagging step of performing the following process using a deep learning algorithm; - a process of analyzing and extracting morphemes of words in each sentence of the social data; - a process of finding and classifying points of interest (Point of Interest), entities, and events among the extracted words; - a process of generating an action event, which is a short sentence expressing a phenomenon occurring in relation to the entity name, by associating the event with the entity name; - a process of generating situation information, which is a short sentence expressing the action event occurring at the point of interest, by associating the action event with the point of interest; - a process of tagging the situation information to the social data;

[0015] It is preferable to provide an event-based disaster occurrence detection method using deep learning, characterized by comprising: a classification step of classifying 'social data including a disaster situation' among the social data by performing event association analysis, behavior association analysis, and time association analysis on the situation information for each point of interest using an association analysis model; a verification step of verifying the authenticity of the 'social data including a disaster situation' by applying one or more verification models; a disaster event generation step of creating a disaster event using the 'social data including a disaster situation' verified in the verification step; and a disaster information posting step of creating and posting disaster information including the disaster event.

[0016] The present invention also includes a fake news verification model in addition to the above-described features, and in the verification step, the fake news verification model is characterized in that it verifies the authenticity by analyzing the context, event change and context pattern of the 'social data including a disaster situation' with a recurrent neural network deep learning model, or in addition, when the deep learning model analyzes the context, the behavior event (CT) included in the social data t ) is the contextual characteristic of the event (E n ) and performs sigmoid calculation as shown below. It is also desirable to use an event-based disaster occurrence detection method using deep learning.

[0017]

[0018] (b is the bias updated through learning)

[0019] The present invention also preferably uses a deep learning-based disaster occurrence detection method, characterized in that, in addition to the above-described features, the information system, when determining the risk of occurrence of the natural disaster or the social disaster, collects and analyzes sensor information measured by public sensors installed in various facilities to determine the risk, and the verification model includes a sensor integration verification model that verifies the on-site situation of the point of interest, and in the verification step, the sensor integration verification model collects the sensor information measured by the public sensors and verifies the authenticity of the 'social data including a disaster situation' through point of interest location verification, population analysis, density analysis, and event signal analysis for the sensor information.

[0020] The present invention also preferably comprises, in addition to the above-described features, a point of interest dictionary, a named entity dictionary and an event dictionary in the information system, wherein the point of interest dictionary, the named entity dictionary and the event dictionary each include national point of interest names, standard named entities and standard event names corresponding to abbreviations, synonyms, neologisms, vulgar words and trendy words, and in the tagging step, - when generating the action event by combining the event with the named entity, a standard named entity corresponding to a word classified as the named entity is searched for in the named entity dictionary, a standard named event corresponding to a word classified as the event is searched for in the event dictionary, and then the standard named event is combined with the standard named entity to generate the action event, and - when generating the situation information by combining the action event with the point of interest, a national named point of interest corresponding to a word classified as the point of interest is searched for in the named entity dictionary and combined with the action event to generate the situation information, wherein an event-based disaster occurrence detection method using deep learning is characterized in that:

[0021] The present invention also provides, in addition to the above-described features, the information system including a point-of-interest association dictionary, wherein the point-of-interest association dictionary includes entity names and events that are related to each point of interest, and in the classification step, - the event association analysis analyzes the degree of correlation between the points of interest, entity names, and events included in the situation information and the entity names and events classified by each point of interest in the point-of-interest association dictionary, and classifies social data having the degree of correlation above a certain level as first social data, - the situation association analysis analyzes the degree of behavior continuity indicating the degree to which the same behavior event occurs continuously for the first social data, and then classifies the first social data having the degree of behavior continuity above a certain level as second social data, - the time association analysis analyzes the degree of time correlation using a probabilistic time distance value that continues between the same behavior events for the second social data, and then classifies the second social data having the degree of time correlation above a certain level as third social data, - the information system classifies the third social data It is also desirable to use an event-based disaster detection method using deep learning, characterized by classifying data into the above-mentioned 'social data including disaster situations'.

[0022] The present invention also preferably uses a deep learning-based disaster occurrence detection method, characterized in that, in addition to the above-described features, the action continuity divides the action C included in the first social data into time units, and calculates and applies the action continuity CT, in which records of time units t that occurred at the point of interest for the action event E tagged in the first social data are continuously recorded, as follows.

[0023]

[0024] The present invention also calculates a temporal correlation H using a probabilistic temporal distance value of each behavior event in units of time flow t for the behavior events included in the second social data in addition to the above-described features, with a temporal weight, wherein the temporal distance value is a CT of one behavior event that occurred at the point of interest. t CT of previous action event (t-i) It is also desirable to use an event-based disaster occurrence detection method using deep learning, characterized in that the probabilistic distance is a time interval that continues to a certain extent, and the temporal correlation H is calculated and applied as follows.

[0025]

[0026] The present invention also provides an information system that detects and provides the risk of occurrence of a natural disaster or social disaster by collecting event information distributed through a social network (SNS) and analyzing the collected event information using a deep learning program. It is also preferable to use a disaster occurrence detection system based on an event using deep learning that can generate and post disaster information using any one of the above methods.

[0027] As discussed above, the present invention has a configuration that analyzes the morphemes of words in each sentence of social data using a deep learning algorithm, extracts points of interest, entity names, and events contained in each social data, reconstructs them, tags them to each social data, and then analyzes them. Therefore, despite the social data having an unstructured data structure, it is possible to systematically utilize them to make early judgments about the signs of a disaster or whether or not a disaster has occurred.

[0028] The present invention also utilizes an information system that classifies and learns social data using a deep learning model to classify "social data containing disaster situations" by point of interest. This allows for the early detection or even preemption of potential disasters occurring at points of interest. This allows for the fundamental prevention of disasters, and even if a disaster does occur, it facilitates early detection and suppression.

[0029] The present invention also includes a correlation analysis model that enables an information system to analyze correlations using a deep learning model and compose situational information about actions to perform correlation analysis according to the passage of time and other factors related to the occurrence of an event. Therefore, when a disaster occurs at a point of interest, the fact of the disaster is accurately and quickly reported using social data, thereby having the effect of minimizing damage caused by the disaster.

[0030] The present invention also has the effect of increasing the accuracy and sensitivity of disaster occurrence judgments because it can determine the truth of social data sentences having disaster meaning by analyzing the context of social data, the event change stage, and the pattern of the context, by having a configuration in which the information system uses a fake news verification model that verifies the accuracy of the event through deep learning.

[0031] The present invention also includes a sensor-integrated verification model that learns about disaster situations occurring at points of interest using image and signal processing, which is a deep learning model of the information system, so that the authenticity of the results of event-based detection can be verified using sensor information, thereby further increasing the accuracy and sensitivity of disaster occurrence judgment.

[0032] Figure 1 is a configuration diagram of an event-based disaster occurrence detection system using deep learning according to the present invention.

[0033] Figure 2 illustrates the concepts of points of interest, entity names, and events used in the present invention.

[0034] Figure 3 illustrates the entire flow of performing the event-based disaster occurrence detection method according to the present invention.

[0035] Figure 4 illustrates a verification step using a fake news verification model in the present invention.

[0036] Figure 5 illustrates another embodiment of an event-based disaster occurrence detection method according to the present invention.

[0037] Figure 6 illustrates a verification step using a sensor integration verification model in the present invention.

[0038] Figure 7 illustrates the detailed concept of point of interest and entity name classification association analysis in the present invention.

[0039] A method for detecting the risk of a natural disaster or social disaster by collecting and analyzing event information distributed through a social network (SNS) using a deep learning program, the method comprising: a social data collection step of collecting social data distributed through one or more social networks; a tagging step of performing the following steps using a deep learning algorithm; - a step of analyzing and extracting morphemes of words in each sentence of the social data; - a step of finding and classifying points of interest (Points of Interest), entities, and events among the extracted words; - a step of generating an action event, which is a short sentence expressing a phenomenon occurring in relation to the entity, by associating the event with the entity; - a step of generating situation information, which is a short sentence expressing the action event occurring at the point of interest, by associating the action event with the point of interest; - a step of tagging the situation information to the social data; a classification step of performing event association analysis, action association analysis, and time association analysis on the situation information for each point of interest using an association analysis model to classify 'social data including a disaster situation' among the social data; The most preferable method is an event-based disaster occurrence detection method using deep learning, characterized by comprising: a verification step of verifying the authenticity of the 'social data including a disaster situation' by applying one or more verification models; a disaster event generation step of creating a disaster event using the 'social data including a disaster situation' verified in the verification step; and a disaster information posting step of creating and posting disaster information including the disaster event.

[0040] Hereinafter, the present invention will be described in detail so that the aforementioned purposes and features become clear, and accordingly, those skilled in the art to which the present invention pertains will be able to easily practice the technical concepts of the present invention. In addition, in describing the present invention, if it is determined that a detailed description of a known technology related to the present invention, which is already well known in the art, may unnecessarily obscure the gist of the present invention, such a detailed description will be omitted.

[0041] In addition, the terms used in the present invention are selected from widely used general terms as much as possible. However, in certain cases, there are terms arbitrarily selected by the applicant, and in such cases, the meanings thereof are described in detail in the description of the relevant invention. Therefore, it is to be noted that the present invention should be understood based on the meaning of the terms, not simply the names of the terms. The terms used in the description of the embodiments are used only to describe specific embodiments and are not intended to limit the embodiments. The singular expression includes the plural expression unless the context clearly indicates otherwise.

[0042] The embodiments may be modified in various ways and may have various additional embodiments. Specific embodiments are illustrated in the drawings and described in detail herein. However, this is not intended to limit the embodiments to a specific form, but rather to encompass all modifications, equivalents, or alternatives that fall within the spirit and technical scope of the embodiments.

[0043] In the descriptions of various embodiments, expressions such as "first," "second," "first," or "second" may describe various components of the embodiments, but do not limit those components. For example, the expressions do not limit the order and / or importance of the components. The expressions may be used to distinguish one component from another.

[0044] The present invention will be described below with reference to the attached drawings. Social networks are a place where diverse users post stories about their various interests and those around them. The amount of data is growing exponentially and contains information describing various phenomena. These social networks post short messages, images, tagging keywords, and videos, and have the unique characteristic of allowing users to simultaneously view information about disaster situations and necessary solutions. Representative social network services include Twitter, Instagram, Facebook, and microblogging, and these services are recently being utilized as social surveillance systems. Furthermore, big data and artificial intelligence technologies are being applied to analyze various phenomena and provide data that explains the situation. Furthermore, various collection environments are provided to collect social data circulating on social networks, allowing for effective data collection and subsequent analysis of various phenomena.

[0045] Social data is increasingly being utilized in disaster response situations. In Japan, when an earthquake and tsunami damaged communication infrastructure, social networks were used to share near-real-time information about the disaster situation and damage, providing people in the disaster area with an opportunity to minimize damage. In this way, social networks, in place of communication networks, can rapidly disseminate disaster information, situational information, and response information to the affected area. This allows for rapid response to disasters occurring at specific locations, and modeling can be linked to various disasters that occur simultaneously.

[0046] The event-based disaster detection method and system using deep learning according to the present invention collects and analyzes event information distributed through social networks (SNS) to detect the risk of natural or social disasters. Specifically, various social data distributed on various social networks are collected, and when a situation occurs based on national and individual points of interest (POIs), the situation is designated as an event and contextual information is generated. The generated contextual information is analyzed for continuity (homeostasis) and temporal correlation to determine whether a disaster has occurred. If a disaster is determined to have occurred, a disaster event is generated. Furthermore, by posting disaster information including disaster events, a series of information systems can be constructed that support decision-making to prevent damage and spread for disaster safety management.

[0047] FIG. 1 illustrates a configuration diagram of an event-based disaster detection system using deep learning according to the present invention. The disaster detection system according to the present invention is an information system (200) that collects and analyzes social data from various social networks (100), and includes a deep learning program for analyzing the collected social data. In addition, it is preferable to include a social data collection means (210) that collects social data from the social networks (100), a preprocessing means (220) that preprocesses the collected social data before analyzing it, an analysis / classification means (230) that analyzes and extracts morphemes of words from the preprocessed social data and then classifies them to classify the 'social data including a disaster situation' through the analysis, a verification means (240) that verifies the classified social data, and a posting means (25) that posts the verified disaster information.

[0048] Meanwhile, the present invention uses the concepts of point of interest, entity name, and event, and FIG. 2 illustrates the concepts of the point of interest (310), entity name (320), and event (330) used in the present invention. As described above, the present invention collects and analyzes social data distributed through social networks. For example, in the content shown in FIG. 1, in the SNS message “There are a lot of people at City Hall Station, so I think I’ll be late. ㅠㅠ, 2022.10.16. 08:24,” the point of interest (310) will be ‘City Hall Station.’ In addition, the entity name (320) will be ‘people,’ and the event (330) will be ‘many,’ ‘late,’ etc.

[0049] In this way, the above points of interest (310) are major national facilities of interest, and as shown in Fig. 1, they include facilities or places where people are crowded in normal times or special cases, such as ‘City Hall Station’, ‘Seoul Station’, ‘Itaewon 00 Street’, and ‘World Cup Stadium’, and they are facilities or places that are expected to suffer great damage in the event of a disaster.

[0050] And, the above entity name (320) may be a person, facility, or place that may be included in the point of interest (310). For example, in FIG. 1, entity names (320) that may be included in City Hall Station may be subway, person, exit, screen door, stairs, escalator, etc. In addition, the above event (330) means an incident or phenomenon that may occur in the point of interest (310). For example, in FIG. 1, events (330) that may be included in City Hall Station may be non-stop, passing, late, many, crowded, flooding, traffic jam, etc.

[0051] Ultimately, by combining the point of interest (310), the entity name (320), and the event (330), it becomes possible to summarize and systematically express the phenomenon occurring around the point of interest (310). For example, by combining the event (330) with the entity name (320) in the social data described in FIG. 1, it will be possible to generate an action event, which is a short sentence expressing a phenomenon occurring in relation to the entity name (320). By combining the entity name (320) called 'person' and the event (330) called 'many,' it will be possible to generate an action event called 'many people.'

[0052] In addition, by combining the above-mentioned behavioral event with the point of interest (310), situational information such as “people are crowded at the city hall subway station” can be created, and by analyzing the collected situational information using deep learning, that is, by analyzing the frequency of occurrence of the situational information, the relationship with the point of interest (310), the continuity, persistence, and temporal correlation of the behavioral event using a deep learning model, it will be possible to detect the risk of disaster occurrence in advance for each point of interest (310).

[0053] Meanwhile, FIG. 3 illustrates a flowchart of the entire process of performing an event-based disaster occurrence detection method using deep learning according to the present invention. It is preferable that the information system (200) included in the present invention first performs a social data collection step (s100) of collecting social data distributed through one or more social networks. It is preferable that the social data collection step (s100) be collected in real time from the one or more social networks (100). In addition, it is preferable to perform a preprocessing step (s101) of removing unnecessary information from the collected social data so that only the necessary information for analyzing the social data collected from the social networks can be extracted.

[0054] To collect social data, big data lake technology should be applied to collect and store large amounts of data in a big data environment. Furthermore, it is also desirable to collect government-provided public IoT data and public statistics, centered on points of interest, and manage them within a big data environment. The collected data structure consists of structured public statistics data, semi-structured public IoT data, and unstructured social data. This includes establishing an integrated data structure model to integrate and manage these data.

[0055] Meanwhile, in order to classify the collected social data centered on the point of interest (310) and to track the behavioral event occurring centered on the point of interest (310), it is preferable to first analyze the morphemes of the words in each sentence of the social data to extract the words (s102), and then perform a process (s103) to find and classify the point of interest (310), the entity name (320), and the event (330) among the extracted words.

[0056] Then, it is preferable to perform a step of generating the action event using the entity name (320) and the event (330) extracted from the social data (s105), combining the action events generated above by point of interest (310) to generate the situation information, and then tagging the situation information to the social data (s106).

[0057] That is, after the information system (200) collects social data distributed through one or more social networks (100) (s100) and completes preprocessing (s101), it performs a process of analyzing and extracting morphemes of words in each sentence of the social data using a deep learning algorithm (s102), and then performs a process of finding and classifying the point of interest (310), the entity name (320), and the event (330) among the extracted words (s103), and then generates an action event, which is a short sentence expressing a phenomenon occurring in relation to the entity name, by combining the event (330) with the entity name (320) (s105), and then generates situation information, which is a short sentence expressing the action event occurring at the point of interest (310), by combining the action event with the point of interest (310), and then performs a process of tagging the situation information to the social data, which is the tagging step (s106).

[0058] Here, it is also desirable that the information system (200) further includes a point of interest dictionary (261), a named entity dictionary (262), and an event dictionary (263). Furthermore, it is desirable that each of the point of interest dictionary (261), the named entity dictionary (262), and the event dictionary (263) include national point of interest names, standard entity names, and standard event names corresponding to abbreviations, synonyms, neologisms, vulgarities, and trendy words.

[0059] In the case where the information system (200) includes the point of interest dictionary (261), the entity name dictionary (262), and the event dictionary (263), when generating the action event by combining the event (330) with the entity name (320), it is preferable to search for a standard entity name corresponding to a word classified as the entity name (320) in the entity name dictionary (262), search for a standard event name corresponding to a word classified as the event (330) in the event dictionary (263), and then combine the standard event name with the standard entity name to generate the action event. In addition, when generating the situation information by combining the action event with the point of interest (310), it is preferable to search for a national point of interest name corresponding to a word classified as the point of interest (310) in the point of interest dictionary (261), and combine it with the action event to generate the situation information (s104).

[0060] And then, it is preferable to perform a classification step (s110) to classify 'social data including disaster situations' among the social data by performing event association analysis, behavior association analysis, and time association analysis on the situation information for each point of interest (310) using the association analysis model. Details on the process (s115) of analyzing the social data using the association analysis model will be described later with reference to FIG. 7.

[0061] Furthermore, it is desirable to perform a verification step (s140) that verifies the authenticity of the "social data containing a disaster situation" by applying one or more verification models. While various verification models may be used, it is desirable to include a fake news verification model that filters out fake news as a basic feature. Fake news is a major issue on social networks, causing problems such as providing misinformation and causing information errors. Therefore, a process for verifying the "social data containing a disaster situation" by detecting false disaster precursors due to fake news is absolutely necessary.

[0062] Figure 4 illustrates the basic concept of a fake news verification model (s140) used in the present invention. The fake news verification model is preferably configured to verify the authenticity of the 'social data including a disaster situation' by analyzing the context of the 'social data including a disaster situation' (s140a), analyzing the event change (s140b), and analyzing the context pattern (s140c) using a recurrent neural network deep learning model (s140). When the deep learning model analyzes the context, the behavior event (CT) included in the social data t ) is the contextual characteristic of the event (E n ) and perform a sigmoid calculation as shown in Equation (1) below. This allows for learning a clear event pattern for fake news, thereby enabling the learning of criteria for what constitutes fake news. This can provide the effect of preemptively blocking fake news information related to behavioral events on social networks.

[0063]

[0064] (b is the bias updated through learning)

[0065] And, after performing a verification step (s140) of verifying the authenticity of the 'social data including a disaster situation' by applying one or more of the verification models, it is preferable that the information system (200) performs a disaster event creation step (s150) of creating a disaster event using the 'social data including a disaster situation' verified (s131) in the verification step (s140) and a disaster information posting step (s160) of creating and posting disaster information including the disaster event.

[0066] Meanwhile, when determining the risk of occurrence of the natural disaster or the social disaster, it is preferable that the information system (200) collects and analyzes sensor information measured by public sensors installed in various facilities to make a determination. This is because even if the authenticity of the 'social data including a disaster situation' is determined using a fake news verification model, there may be an erroneous determination. Therefore, it is preferable that the verification model include a sensor integration verification model (s145) that verifies the on-site situation of the point of interest. FIG. 5 illustrates the execution process of the present invention including the sensor integration verification model (s145), and FIG. 6 illustrates the process of executing the sensor integration verification model (s145).

[0067] As shown in FIGS. 5 and 6, when the sensor integration verification model (s145) is included, in the verification step (s130), it is preferable that the sensor integration verification model collects the sensor information measured by the public sensors and verifies the authenticity of the 'social data including disaster situation' through verification of the location of points of interest, analysis of the number of individuals, analysis of density, and analysis of event signals for the sensor information.

[0068] Figure 4 illustrates the process of securing public IoT sensor data for the above-mentioned points of interest and verifying events. While sensor data has limitations in terms of the information obtainable from points of interest, it assumes the collection of data open to the public. CCTV, in particular, is a public data source capable of observing various activities. Public CCTV information for points of interest is secured and collected by applying location information. CCTV data is collected using streaming technology, and YOLO, a deep learning model, is used to analyze the number of people distributed at points of interest and their density to confirm the verification process.

[0069] CCTV data is collected using streaming technology, and after first verifying the location of points of interest (s145a), an object count analysis (s145b) of the distribution of people in points of interest is performed using YOLO, a deep learning model. The commonly known YOLO model calculates the object frequency by counting the number of objects recognized as people. The object frequency calculates the size of the space by marking the size of the point of interest as a box unit. Then, a density analysis (s145c) is performed, which uses the object count analysis and the space size to calculate the density frequency of people or buildings in the event of a disaster in a space unit. Finally, event signal analysis (s145d) detects actions that may occur depending on the number of objects and the degree of density as sensor event signals. The sensor event signal is applied to YOLO, a deep learning model that has learned rescue signals or hand signals related to rescue among various actions as sensor event signals. The signal is processed through a series of steps, including the number of objects, density, and sensor event signals at the point of interest, to verify the disaster situation.

[0070] Meanwhile, FIG. 7 illustrates a procedure flowchart for explaining the behavior-related analysis model (s120) included in the present invention. Hereinafter, the behavior-related analysis model (s120) will be described with reference to FIG. 7. In order to perform the behavior-related analysis model (s120), it is preferable to include a point-of-interest association dictionary (264) in the information system (200). It is preferable that the point-of-interest association dictionary (264) be able to find and provide entity names (320) and events (330) that are related to each point of interest (310). That is, the point-of-interest association dictionary (264) contains entity names (320) and events (330) that can be included in each point of interest (310). For example, if the point of interest (310) is the City Hall subway station, the point of interest association dictionary (264) will include subway, stairs, etc. in the entity name (320), and the audience seats will not be included because they are not related at all. In addition, if the point of interest (310) is the World Cup stadium, the point of interest association dictionary (264) will include events (330) such as many and pushed because they are related, but the event (330) called non-stop will not be included because it is not related at all.

[0071] In this way, in the classification step (s110), the event association analysis (s121) included in the behavior association analysis model (s120) is preferably a process of analyzing the association between the point of interest (310), entity name (320) and event (330) included in the situation information and the entity name (320) and event (330) classified by the point of interest (310) in the point of interest association dictionary (264) and classifying social data having the association level above a certain level as first social data (s122). Here, the association (E) applies the following equation (2) to evaluate the association between the entity name and the event classified centered on the point of interest. Equation (2) is a process of evaluating the association between the time-based event keyword (E) and the entity name. t ) is evaluated as a highly reliable event keyword by using a weight that evaluates reliability. The weight (w) is the event tagging information (E) linked to the entity name. t ) between event keywords obtained from the flow of time units (t) through conditional probability for (t-i) ) calculates the strength of association.

[0072]

[0073] In addition, it is preferable that the above situational correlation analysis (s123) be a process of analyzing the behavior continuity, which indicates the degree to which the same behavioral event occurs continuously for the first social data, and then classifying the first social data with the behavior continuity above a certain level as the second social data (s124). As described above, the event correlation analysis (s121) evaluates the strength of the correlation between events, and extracts only the first social data by removing unnecessary information and leaving only highly reliable event information (s122). In the above situational correlation analysis (s123), in order to analyze the behavior continuity for the first social data derived in this process, the behavior continuity of the point of interest (310) and the correlation between the event are calculated by applying equation (3) to have an independent meaning. Equation (3) divides the action C of the point of interest (310) into time units, and calculates the constancy of the action event (E) in which the records of the time units that occurred at the point of interest (310) continue continuously for the action event (E). That is, the action of the point of interest (310) can be determined as an action event (E) according to the action by determining whether the conditional probability that the action occurred due to the action event (E) is highly or low. Here, it is preferable to calculate the action continuity CT by dividing the action C included in the first social data into time units, and applying the following Equation (3) to the action event (E) tagged in the first social data, in which the records of the time units t that occurred at the point of interest continue continuously.

[0074]

[0075] Finally, the second social data with high behavior continuity obtained from the above situational correlation analysis (s123) is analyzed for temporal correlation (H). The temporal correlation (H) calculates the time weight using the probabilistic distance value of each keyword in units of time flow t for the behavior-related event information that occurred at the point of interest (310). The temporal distance value is the behavior event (CT) that occurred at the point of interest. t ) occurred before the previous action event (CT) (t-i) ) calculates the probabilistic distance from the point of interest to some extent. For example, if the point of interest is Namsan Tower, when there are keywords for behavior events mentioned in social networks about Namsan Tower, the transportation tools include cable cars, stairs, and cars, and in the case of cars, when keywords such as “blocked,” “delayed,” “crowded,” and “fast” appear in the behavior events, the time distance at which each keyword occurred can be calculated to predict the reason why moving by car is slow based on the present. In addition, since it is possible to verify whether the reason for the slowness is due to an accident or whether there are many moving vehicles, it is possible to predict the cause of the behavior-related event.

[0076] That is, the temporal correlation analysis is preferably a process of analyzing the temporal correlation with the probabilistic temporal distance value that continues between the same behavior events for the second social data, and then classifying the second social data with the temporal correlation above a certain level as third social data. Here, the temporal correlation is calculated by using the probabilistic temporal distance value of each behavior event in units of time flow t for the behavior events included in the second social data, and calculating the temporal correlation H with a time weight, wherein the temporal distance value is one behavior event CT that occurred at the point of interest. t CT of previous action event (t-i)It is a probabilistic distance that leads to a certain time zone, and it is desirable to calculate the above temporal correlation H by applying the following equation (4).

[0077]

[0078] Then, it is preferable that the information system classify the third social data derived after performing the temporal correlation analysis as "social data containing disaster situations." In this way, the present invention collects and systematizes social network data with an unstructured data structure using a big data collection method, and then analyzes it using a deep learning model, an artificial intelligence learning model, thereby enabling the early detection of signs and facts of disaster events.

[0079] While the present invention has been described with numerous examples, it is not necessarily limited to these examples, and various modifications may be made without departing from the scope of the present invention. Therefore, the examples disclosed herein are intended to illustrate, rather than limit, the technical concept of the present invention, and the scope of the present invention is not limited by these examples. The scope of protection of the present invention should be interpreted according to the following claims, and all technical ideas within the scope equivalent thereto should be interpreted as being included within the scope of the present invention.

[0080] The present invention has a configuration in which the morphemes of words in each sentence of social data are analyzed using a deep learning algorithm to extract points of interest, entity names, and events contained in each social data, and these are reconstructed, tagged to each social data, and then analyzed. Therefore, despite the social data having an unstructured data structure, a system can be provided that can systematically utilize this to early determine signs of a disaster or whether or not it has occurred.

Claims

1. A method for detecting the risk of natural disasters or social disasters by collecting and analyzing event information distributed through social networks (SNS) using a deep learning program. A social data collection step that collects social data distributed through one or more social networks; Tagging step that performs the following process using a deep learning algorithm; - The process of analyzing and extracting the morphemes of words in each sentence of the above social data - The process of finding and classifying points of interest, entities, and events among the extracted words. - A process of creating an action event, which is a short sentence expressing a phenomenon occurring in relation to the entity name, by combining the event with the entity name. - A process of generating context information, which is a short sentence expressing the action event occurring at the point of interest, by combining the above action event with the point of interest. - The process of tagging the above situation information to the above social data A classification step of classifying social data that includes a disaster situation among the social data by performing event association analysis, behavior association analysis, and time association analysis on the situation information for each point of interest using an association analysis model; A verification step of verifying the authenticity of social data containing the above disaster situation by applying one or more verification models; A disaster event generation step for generating a disaster event using social data including the disaster situation verified in the above verification step; and A method for detecting disaster occurrence based on an event using deep learning, characterized in that it includes a disaster information posting step of generating and posting disaster information including the above disaster event.

2. In paragraph 1, The above verification model includes a fake news verification model. In the above verification step, the fake news verification model is characterized by analyzing the context, event changes and context patterns of social data including the disaster situation using a recurrent neural network deep learning model to verify the authenticity, and is an event-based disaster occurrence detection method using deep learning.

3. In paragraph 2, When the above deep learning model analyzes the above context, the above behavior event (CT) included in the above social data t ) is the contextual characteristic of the event (E n ) and performs sigmoid calculation as follows, using deep learning as an event-based disaster occurrence detection method. (b is the bias updated through learning) 4. In paragraph 1 The above information system, when determining the risk of occurrence of the above natural disaster or the above social disaster, collects and analyzes sensor information measured by public sensors installed in various facilities to make a determination. The above verification model includes a sensor integration verification model that verifies the on-site situation of the above point of interest. In the above verification step, the sensor integration verification model collects the sensor information measured by the public sensors and verifies the authenticity of the social data including the disaster situation through verification of the location of the point of interest, analysis of the number of objects, analysis of the density, and analysis of the event signal for the sensor information, which is an event-based disaster occurrence detection method using deep learning.

5. In paragraph 1 The above information system includes a point of interest dictionary, a named entity dictionary, and an event dictionary. The above point of interest dictionary, the above entity name dictionary and the above event dictionary each include national point of interest names, standard entity names and standard event names corresponding to abbreviations, synonyms, neologisms, vulgar words and trendy words. In the above tagging step, - When generating the action event by combining the above entity name with the above event, a standard entity name corresponding to a word classified as the entity name is found in the entity name dictionary, a standard event name corresponding to a word classified as the event is found in the event dictionary, and then the action event is generated by combining the standard event name with the standard entity name. - An event-based disaster occurrence detection method using deep learning, characterized in that when generating the situation information by combining the above-mentioned behavior event with the above-mentioned point of interest, the name of the national point of interest corresponding to the word classified as the point of interest is found in the above-mentioned point of interest dictionary and combined with the above-mentioned behavior event to generate the situation information.

6. In paragraph 1 The above information system includes a dictionary of related words of points of interest, The above point of interest association dictionary includes entity names and events related to each point of interest. In the above classification step, - The above event association analysis analyzes the association between the points of interest, entity names, and events included in the above situation information and the entity names and events classified by the points of interest in the point-of-interest association dictionary, and classifies social data with a certain level of association or higher as first social data. - The above situational correlation analysis analyzes the behavior continuity, which indicates the degree to which the same behavior event occurs continuously for the first social data, and then classifies the first social data with a certain level of behavior continuity or higher as the second social data. - The above temporal correlation analysis analyzes temporal correlation with the probabilistic temporal distance value between the same behavior events for the second social data, and then classifies the second social data with a temporal correlation above a certain level as third social data. - An event-based disaster occurrence detection method using deep learning, characterized in that the above information system classifies the third social data into social data including the disaster situation.

7. In paragraph 6 The above-mentioned action continuity is characterized by calculating and applying the action continuity CT, in which the records of the time unit t that occurred at the point of interest for the action event E tagged in the first social data are continuously continued, as follows: an event-based disaster occurrence detection method using deep learning.

8. In paragraph 6 The above temporal correlation is calculated by using the probabilistic temporal distance value of each behavior event in units of time flow t for the behavior events included in the second social data, and the temporal correlation H is calculated by using the temporal weight, and the temporal distance value is one behavior event C that occurred at the point of interest. t The previous action event C (t-i) A method for detecting disaster occurrence based on an event using deep learning, characterized in that the probabilistic distance is a time interval, and the temporal correlation H is calculated and applied as follows.

9. An information system that collects event information distributed through social networks (SNS) and analyzes the collected event information using a deep learning program to detect and provide the risk of natural or social disasters. An event-based disaster occurrence detection system using deep learning that can generate and post disaster information by any one of the methods in Articles 1 through 8.

Images