Cross validation based intelligent disaster classification and alerting system

Abstract

The present invention relates to a cross-verification-based intelligent disaster classification and warning system, and more specifically, to a cross-verification-based intelligent disaster classification and warning system that overcomes the limitations of existing methods limited to the simple transmission of disaster messages by fusing external disaster messages in the form of unstructured text and structured sensor data within the user environment in a multimodal manner and cross-verifying them, and can classify disaster grades by quantifying the impact of the disaster at the user's actual location.

Description

Technology Field

[0001] The present invention relates to a cross-verification-based intelligent disaster classification and warning system, and more specifically, to a cross-verification-based intelligent disaster classification and warning system that overcomes the limitations of existing methods limited to the simple transmission of disaster messages by fusing external disaster messages in the form of unstructured text and structured sensor data within the user environment in a multimodal manner and cross-verifying them, and can classify disaster grades by quantifying the impact of the disaster at the user's actual location. Background Technology

[0003] With the recent frequent occurrence of various types of disasters, such as earthquakes, typhoons, fine dust, infectious diseases, and chemical leaks, the importance of disaster warning systems for rapidly conveying disaster situations to the public is becoming increasingly prominent. Accordingly, governments around the world operate emergency disaster text message systems to send disaster information via text messages to an unspecified number of people in affected areas when a disaster occurs. In Korea as well, emergency disaster text messages via the Disaster Safety Communication Network are widely utilized to provide rapid alerts to the public regarding various disaster situations, including earthquakes, heavy rainfall, heatwaves, fine dust, and infectious diseases.

[0004] However, these existing disaster alert systems operated by sending identical messages in bulk across wide areas, which had limitations in that they could not reflect the actual impact of the disaster in the individual user's location. For example, when an earthquake occurred in a specific region and a disaster alert was sent out across the entire area, the same level of warning was delivered to users located at a considerable distance from the epicenter, causing unnecessary fear. Conversely, in the case of disasters with significant localized variations, such as fine dust or hazardous gas leaks, it was difficult to accurately assess the level of risk in the user's actual living space using only wide-area alerts.

[0005] Meanwhile, to address these issues, a technology has been developed that installs sensors within the user environment and detects disaster situations based on their measurements. This involves installing various sensors, such as fire detection sensors, earthquake detection sensors, and fine dust meters, inside buildings to monitor situations like fires, vibrations, and air quality deterioration in real time.

[0006] However, since these sensor-based disaster detection methods rely on a single sensor or a single type of data, there is a high probability of false alarms caused by sensor malfunctions or measurement errors. Furthermore, they have limitations in that they cannot respond to types of disasters that are impossible to detect with individual sensors, such as the spread of infectious diseases or the indirect impact of distant disasters. Additionally, because judgments are made based solely on local data from the specific space where the sensor is installed, there is also the problem of being unable to grasp the overall context of a disaster unfolding outside that area.

[0007] Therefore, there is a need to develop an intelligent disaster classification and warning system that can quantitatively calculate the risk level of a disaster and precisely classify it into multiple grades by fusing external disaster text messages and internal sensor data in a multimodal manner and performing AI-based cross-verification. Prior art literature

[0009] Korean Registered Patent No. 10-2249901 (May 3, 2021) The problem to be solved

[0010] The present invention has been devised to solve the above problem, and aims to provide a cross-verification-based intelligent disaster classification and warning system capable of classifying disaster grades by quantifying the impact of a disaster at the user's actual location, thereby overcoming the limitations of existing methods that are limited to the simple transmission of disaster messages by fusing external disaster messages in the form of unstructured text with structured sensor data within the user environment in a multimodal manner and cross-verifying them.

[0011] Furthermore, the present invention aims to provide a cross-verification-based intelligent disaster classification and warning system capable of improving the accuracy of disaster grade classification in various disaster situations by dynamically calculating weights by comprehensively considering multiple environmental variables, such as the physical distance from the disaster location, the real-time reliability index of sensors, and temporal urgency factors, in addition to setting basic weights according to the type of disaster.

[0012] Furthermore, the present invention aims to provide a cross-verification-based intelligent disaster classification and warning system that can function as a practical disaster defense system beyond a simple notification system, as it can actively control disaster prevention devices within the user environment according to the disaster grade.

[0013] Furthermore, the present invention aims to provide a cross-validation-based intelligent disaster classification and warning system capable of implementing an evolutionary disaster classification system that becomes optimized for the characteristics of the user's residential environment over time by collecting the user's manual intervention history as feedback data and continuously correcting the learning weight matrix and bias values. means of solving the problem

[0015] The present invention for achieving the above objective is characterized by comprising: a receiving unit that receives a disaster message in the form of unstructured text from an external source; a collecting unit that collects structured sensor data from a plurality of sensors installed within a user environment; and an AI analysis unit that receives the disaster message from the receiving unit, analyzes its content, receives the sensor data from the collecting unit, analyzes the correlation and consistency with the content of the disaster message to perform cross-verification, and determines a disaster grade based on the cross-verification results.

[0016] In addition, the AI ​​analysis unit of the present invention is characterized by including a text analysis module that extracts the type of disaster by analyzing the disaster text through a natural language processing algorithm, a sensor analysis module that analyzes the sensor data through a time series analysis algorithm, and a level determination module that verifies whether there is a match between the type of disaster extracted from the text analysis module and the sensor data analyzed from the sensor analysis module, and determines the level of the disaster grade based on the verification result.

[0017] In addition, the text analysis module of the present invention is characterized by numerically converting the disaster character to calculate a text embedding vector, the sensor analysis module is characterized by numerically converting the sensor data to calculate a sensor embedding vector, and the level determination module is characterized by applying a reliability weight of the disaster character to the text embedding vector and applying an accuracy weight of the sensor data to the sensor embedding vector, respectively, and fusing them to calculate a disaster index for determining a disaster grade level.

[0018] In addition, the level determination module of the present invention is characterized in that the application ratio of the reliability weight of the disaster character or the accuracy weight of the sensor data is set differently depending on the type of disaster extracted.

[0019] In addition, the disaster index of the present invention is,

[0020]

[0021] (Softmax: A function that converts calculated scores into probabilities between 0 and 1, D: Disaster Index, α: Confidence weight of disaster messages, β: Accuracy weight of sensor data, h text : Text embedding vector, h sensor : Sensor embedding vector, : is characterized by a vector combination, W: an empirical learning weight matrix acquired by the AI ​​model through training, and b: a bias value (bias, the system's sensitivity index to a specific disaster).

[0022] In addition, the level determination module of the present invention is characterized by determining the disaster grade by discretizing it to include two or more of the interest stage, caution stage, warning stage, or severe stage by comparing the calculated disaster index with a plurality of preset threshold values.

[0023] In addition, the level determination module of the present invention is characterized by activating a disaster prevention device within the user environment when the disaster grade corresponds to a disaster grade greater than or equal to a set threshold.

[0024] In addition, the level determination module of the present invention is characterized by collecting the intervention history as feedback data and correcting the learning weight matrix (W) or the bias value (b) when an intervention event occurs in which a user manually changes the level for the determined disaster level.

[0025] In addition, the level determination module of the present invention is characterized by correcting the reliability weight of the disaster message according to the physical distance between the disaster occurrence location within the disaster message and the user environment location, or correcting the accuracy weight of the sensor data using the real-time reliability index of the sensor within the user environment as a variable.

[0026] In addition, the text analysis module of the present invention extracts a temporal urgency factor from the disaster text, and the sensor analysis module is characterized by variably controlling the collection cycle of the sensor data in real time in response to the extracted temporal urgency factor. Effects of the invention

[0028] The present invention overcomes the limitations of existing methods that are limited to the simple transmission of disaster messages by fusing external disaster messages in the form of unstructured text with structured sensor data within the user environment in a multimodal manner and cross-verifying them, and has the effect of classifying disaster grades by quantifying the impact of the disaster at the user's actual location.

[0029] In addition, the present invention has the effect of improving the accuracy of disaster grade classification in various disaster situations by dynamically calculating weights by comprehensively considering multiple environmental variables, such as the physical distance from the disaster location, the real-time reliability index of the sensor, and the temporal urgency factor, in addition to setting basic weights according to the type of disaster.

[0030] In addition, since the present invention can actively control disaster prevention devices within the user environment according to the disaster level, it has the effect of functioning as a practical disaster defense system beyond a simple notification system.

[0031] In addition, the present invention has the effect of implementing an evolutionary disaster classification system that becomes optimized for the characteristics of the user's living environment as time passes, by collecting the user's manual intervention history as feedback data and continuously correcting the learning weight matrix and bias values. Brief explanation of the drawing

[0033] FIG. 1 is a block diagram of a cross-verification-based intelligent disaster classification and warning system according to an embodiment of the present invention. FIG. 2 is a flowchart of the operation of the AI ​​analysis unit of a cross-verification-based intelligent disaster classification and warning system according to an embodiment of the present invention. FIG. 3 is a flowchart of the operation of the AI ​​analysis unit of a cross-verification-based intelligent disaster classification and warning system according to an additional embodiment of the present invention. Specific details for implementing the invention

[0034] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. The drawings presented below are provided as examples to ensure that the concept of the present invention is sufficiently conveyed to those skilled in the art. Accordingly, the present invention is not limited to the drawings presented below and may be embodied in other forms. Furthermore, throughout the specification, the same reference numerals indicate the same components.

[0035] Unless otherwise defined, technical and scientific terms used herein have the meanings commonly understood by those skilled in the art to which this invention pertains, and descriptions of known functions and configurations that could unnecessarily obscure the essence of the invention are omitted in the following description and accompanying drawings.

[0036] This specification clarifies the scope of the present invention and explains the principles of the present invention and discloses embodiments so that those skilled in the art can practice the present invention. The disclosed embodiments may be implemented in various forms and are not limited to the embodiments described herein.

[0037] Terms such as "comprising," "may compose," or "having," which may be used in various embodiments of the present invention, are intended to indicate the existence of the features described in the specification (e.g., functions, numbers, steps, actions, components, parts, or combinations thereof), and should be understood as not excluding in advance the existence or addition of one or more other features.

[0038] As used in this specification, singular expressions include plural expressions unless the context clearly indicates otherwise.

[0039] When it is stated that a component is "connected or coupled" to another component, it should be understood that the component may be directly connected or coupled to the other component, but that a new component may also exist between the component and the other component. On the other hand, when it is stated that a component is "directly connected" or "directly coupled" to another component, it should be understood that no new component exists between the component and the other component.

[0040] The terms "first," "second," etc., as used in this specification may be used to describe various components, but the components should not be limited by these terms. The terms are used solely for the purpose of distinguishing one component from another.

[0042] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

[0043] The present invention comprises a receiving unit (100) that receives a disaster message in the form of unstructured text from an external source, a collecting unit (200) that collects structured sensor data from a plurality of sensors (10) installed within a user environment, and an AI analysis unit (300) that receives a disaster message from the receiving unit (100), analyzes the content thereof, receives sensor data from the collecting unit (200), analyzes the correlation and consistency with the content of the disaster message to perform cross-verification, and determines a disaster grade based on the cross-verification results.

[0045] The receiving unit (100) plays the role of receiving disaster messages in the form of unstructured text sent by the government or public institutions from the outside. Here, disaster messages refer to warning information in the form of text sent to an unspecified number of people through the government disaster safety communication network or emergency disaster message system regarding various disaster situations such as earthquakes, typhoons, fine dust, infectious diseases, or chemical leaks. Due to their nature, these disaster messages are unstructured text that does not have a standardized structure, and contain information such as the type of disaster, location of occurrence, expected damage range, and response guidelines in the form of natural language. The receiving unit (100) receives these disaster messages through communication means such as API or SMS and provides the received disaster messages to the AI ​​analysis unit (300).

[0047] The collection unit (200) is responsible for collecting structured sensor data from multiple sensors (10) installed within the user environment. Here, the multiple sensors (10) include various sensors that can be used to detect disaster situations, such as seismometers, fine dust measuring instruments, flood detection sensors, or fire detection sensors. These sensors are installed within the space where the user resides or lives and measure physical values ​​such as temperature, humidity, vibration, and fine dust concentration in real time. The sensor data collected by the collection unit (200) is structured numerical data in the form of a time series that is continuously measured over time, and its data format is different from the disaster message in the form of unstructured text received by the receiving unit (100). The collection unit (200) provides the collected sensor data to the AI ​​analysis unit (300).

[0049] The AI ​​analysis unit (300) analyzes the content of the disaster text message provided by the receiving unit (100), analyzes the correlation and consistency with the sensor data provided by the collection unit (200) to perform cross-verification, and determines the disaster grade based on the cross-verification results. That is, the AI ​​analysis unit (300) performs a precise classification of disaster situations that reflect regional characteristics that cannot be determined by the disaster text message alone by comparing in real-time with internal sensor (10) data whether the content of the disaster text message sent on a wide-area basis is valid at the actual user's location.

[0050] To this end, the AI ​​analysis unit (300) is configured to include a text analysis module (310) that extracts the type of disaster by analyzing disaster text through a natural language processing algorithm, a sensor analysis module (320) that analyzes sensor data through a time series analysis algorithm, and a level determination module (330) that verifies whether there is a match between the type of disaster extracted from the text analysis module (310) and the sensor data analyzed from the sensor analysis module (320), and determines the level of the disaster grade based on the verification result.

[0051] The text analysis module (310) performs the role of extracting the type of disaster by analyzing the disaster text provided by the receiver (100) through a natural language processing algorithm. Here, the natural language processing algorithm may include a Transformer-based language model such as BERT or RoBERTa, and analyzes the keywords and context of the disaster text, which is unstructured text, to identify the type of disaster, such as whether it is an earthquake, fine dust, infectious disease, or fire, and extracts the expected risk level. This text analysis module (310) converts the disaster text numerically to produce a text embedding vector. That is, it converts the content of the disaster text into a vector, which is a numerical form that artificial intelligence can calculate. When a disaster sentence is received as input and passed through the encoder of a natural language processing model, a text embedding vector containing the semantic characteristics of the disaster text is produced.

[0052] Meanwhile, the text analysis module (310) can extract a temporal urgency factor from the disaster text. Here, the temporal urgency factor refers to urgency information related to time included in the context of the disaster text. For example, if the disaster text includes the expression "evacuate immediately," it is extracted as having a very high temporal urgency, and if it includes the expression "caution until tomorrow morning," it is extracted as having a relatively low temporal urgency. This temporal urgency factor is extracted separately from the type of disaster and can be used to adjust the sensor data collection cycle of the sensor analysis module (320) to be described later and to calculate weights for the level determination module (330).

[0053] The sensor analysis module (320) performs the role of analyzing sensor data provided by the collection unit (200) through a time series analysis algorithm. Here, the time series analysis algorithm may include a hybrid deep learning model such as CNN-LSTM and analyzes the patterns of multiple sensor (10) data over time. Since CNN is suitable for extracting local features of sensor data and LSTM is suitable for identifying long-term dependencies according to temporal order, combining the two allows for more precise analysis of the time series patterns of sensor data. This sensor analysis module (320) numerically converts the sensor data to produce a sensor embedding vector. When multiple sensor (10) data over time is input into the CNN-LSTM model, the time series data of each sensor is converted into a single condensed sensor embedding vector.

[0054] Meanwhile, the sensor analysis module (320) can variably control the collection cycle of sensor data in real time in response to the temporal urgency factor extracted by the text analysis module (310). For example, if the text analysis module (310) recognizes a high-urgency keyword such as "immediate evacuation" from a disaster text message, the sensor analysis module (320) switches the sensor data scan cycle to a high speed in seconds to collect and analyze more dense sensor data. Conversely, if the temporal urgency is relatively low, the normal collection cycle is maintained. Through this, the real-time nature of the sensor data is maximized in urgent disaster situations, thereby enabling more accurate cross-verification.

[0055] The level determination module (330) verifies whether there is a match between the type of disaster extracted from the text analysis module (310) and the sensor data analyzed from the sensor analysis module (320), and determines the level of the disaster grade based on the verification result. To this end, the level determination module (330) calculates a disaster index for determining the disaster grade level by applying a reliability weight of the disaster character to the text embedding vector and applying an accuracy weight of the sensor data to the sensor embedding vector, respectively, and fusing them.

[0056] The disaster index is calculated by the following formula.

[0057]

[0058] Here, Softmax is a function that converts the calculated scores into probabilities between 0 and 1, D is the disaster index, α is the confidence weight of the disaster character, β is the accuracy weight of the sensor data, and h text is a text embedding vector, and h sensor is a sensor embedding vector, and is a vector combination operator, W is an empirical learning weight matrix acquired by the AI ​​model through training, and b is a bias value, which is the system's sensitivity index to a specific disaster.

[0059] To explain the above formula in more detail, first, the text embedding vector is multiplied by the reliability weight α of the disaster message, and the sensor embedding vector is multiplied by the accuracy weight β of the sensor data. These two vectors, each with its respective weights applied, are combined into a single combined vector using a vector combination operator. Vector combination is a process that combines text analysis results and sensor numerical results—which are different forms of data—into a single computational space, serving to enable the artificial intelligence to consider both data simultaneously. The final score is calculated by multiplying the combined vector by the learning weight matrix W and adding the bias value b, and then converting this into a probability value between 0 and 1 using the Softmax function to derive the disaster index D.

[0060] Here, the learning weight matrix W represents the empirical judgment criteria acquired by the AI ​​model during training through tens of thousands of past disaster cases; it corresponds to the strength of neural connections in the network that link combined vectors to disaster grades. In other words, the learning weight matrix W acts as a filter to determine which disaster grade a combination of specific disaster text content and specific sensor measurements corresponds to. The bias value b is a fundamental tendency acting externally to the data, controlling how sensitively the system reacts to a specific disaster. For instance, in the case of an earthquake with a high risk of casualties, the bias value can be increased to adjust the model's baseline to classify the event as more conservatively dangerous.

[0061] Meanwhile, the level determination module (330) is characterized by having different application ratios for the reliability weight of the disaster text or the accuracy weight of the sensor data depending on the type of disaster extracted. This is because the information value of the disaster text and the sensor data varies depending on the type of disaster. For example, in the case of a disaster such as yellow dust or fine dust, the sensor data of a fine dust measuring device installed inside the user environment reflects the actual air quality condition of the location more accurately than the disaster text sent on a wide-area basis, so it is reasonable to increase the weight of the accuracy weight β of the sensor data. Conversely, in the case of a disaster such as an infectious disease, it is impossible to directly detect whether the infectious disease is spreading with individual sensors, so it is reasonable to overwhelmingly increase the weight of the reliability weight α of the disaster text.

[0062] However, simply setting the application ratio of such weights as a fixed value according to the type of disaster is artificial and has limitations in that it does not sufficiently reflect various actual situations. Therefore, in addition to setting the basic weight ratio according to the type of disaster, the level determination module (330) can correct the reliability weight of the disaster text according to the physical distance between the location of the disaster occurrence within the disaster text and the location of the user environment, or correct the accuracy weight of the sensor data using the real-time reliability index of the sensor (10) in the user environment as a variable.

[0063] More specifically, weight correction based on the physical distance between the location of the disaster and the location of the user environment is performed as follows. Disaster messages typically contain location information of the area where the disaster occurred, and a text analysis module (310) extracts this and calculates the physical distance by comparing it with the location information of the user environment. The closer the distance between the location of the disaster and the user environment, the higher the reliability weight α of the disaster message is corrected, and the further the distance, the lower the reliability weight α is corrected. For example, if an earthquake disaster message is received but the epicenter of the earthquake is in an area very far from the user environment, the reliability weight of the disaster message can be corrected downward to prevent unnecessary spread of fear.

[0064] In addition, weight correction based on the real-time reliability index of the sensor (10) in the user environment is performed using the sensor reliability index, which is calculated based on the sensor's recent malfunction history, the age of the device, and whether it is operating normally, as a variable. The higher the sensor's reliability index, the higher the accuracy weight β of the sensor data is maintained, and the lower the sensor's reliability index, the lower the accuracy weight β is corrected downward. Through this, even if the sensor's condition is poor or malfunction is suspected, the accuracy of the disaster index can be maintained by appropriately correcting the weight.

[0065] In this way, the level determination module (330) dynamically calculates reliability weight α and accuracy weight β in real time by comprehensively considering multiple environmental variables, such as the type of disaster, the physical distance between the disaster location and the user environment, the real-time reliability index of the sensor, and the temporal urgency factor extracted by the text analysis module (310), so that the actual impact of the disaster situation can be reflected more precisely compared to when fixed weights are used.

[0067] Meanwhile, the dynamic calculation of such reliability weight α and accuracy weight β is configured so that the AI ​​analysis unit (300) can determine the disaster grade even in exceptional situations where data is not normally provided from either the receiving unit (100) or the collecting unit (200).

[0068] For example, there may be cases where the receiving unit (100) is unable to receive external disaster text messages due to a failure of the disaster text message sending system, a communication network disconnection, an error in the API server, etc. In this case, the AI ​​analysis unit (300) detects the absence of data from the receiving unit (100), sets the reliability weight α of the disaster text message to 0 or a value close to 0, and corrects the accuracy weight β of the sensor data upward. Then, the disaster index can be calculated based only on the sensor data provided by the collection unit (200). However, in this case, since cross-verification of the information of the external disaster text message is not performed, the level determination module (330) may also notify the user that the reliability of the disaster grade is relatively low.

[0069] Conversely, if some or all of the multiple sensors (10) installed within the user environment are malfunctioning, the power to the sensors is cut off, or a failure occurs in the transmission path of the sensor data, there may be cases where the collection unit (200) cannot collect sensor data normally. In this case, the AI ​​analysis unit (300) detects the absence of data from the collection unit (200), sets the accuracy weight β of the sensor data to 0 or a value close to 0, and corrects the reliability weight α of the disaster text upward. Then, the disaster index can be calculated based only on the disaster text provided by the receiver (100). However, since cross-verification with the sensor data is not performed in this case as well, the level determination module (330) may also notify the user that the reliability of the disaster grade is relatively low.

[0070] In addition, if only some of the multiple sensors (10) in the user environment are malfunctioning or if data collection is impossible for only some of them, the AI ​​analysis unit (300) can calculate a sensor embedding vector by selectively utilizing only the data from the sensors (10) that are operating normally. At this time, it is preferable that the level determination module (330) be configured to prevent judgments based on incomplete sensor data from being over- or under-evaluated by proportionally lowering the accuracy weight β of the sensor data compared to when all sensors (10) are operating normally.

[0072] Based on the disaster index calculated in this way, the level determination module (330) determines a disaster grade that includes two or more of the interest level, caution level, warning level, or severe level by comparing the calculated disaster index with a plurality of preset threshold values. That is, the level determination module (330) classifies the disaster index, which is a continuous probability value between 0 and 1 derived through multimodal analysis, into a discrete disaster grade through comparison with a plurality of preset threshold values ​​and outputs it.

[0073] To give a more specific example, if the disaster index D is 0 or higher and less than 0.25, it is determined to be the "Attention" stage, which means that signs of a disaster are detected but there is no disruption to daily life. If the disaster index D is 0.25 or higher and less than 0.50, it is determined to be the "Caution" stage, which means that there is a high probability of a disaster occurring and careful observation is required. If the disaster index D is 0.50 or higher and less than 0.75, it is determined to be the "Alert" stage, which means that the disaster has become visible and damage has begun to occur. If the disaster index D is 0.75 or higher and less than or equal to 1.00, it is determined to be the "Severe" stage, which means that large-scale damage has occurred and immediate evacuation and response are required. However, the specific numerical values ​​of the above thresholds are merely examples, and it goes without saying that they may be set differently depending on the type of disaster or policy standards.

[0074] A specific example of the operation of such cross-verification is as follows. When a disaster message of "Earthquake occurred, please evacuate immediately" is received from the outside as a wide-area disaster message, the text analysis module (310) analyzes it to identify the type of disaster as an earthquake and identifies the temporal urgency as high. At the same time, the collection unit (200) collects sensor data from an earthquake sensor, vibration sensor, or acceleration sensor within the user environment, and the sensor analysis module (320) performs time-series analysis on the sensor data to calculate a sensor embedding vector. At this time, if the measurement values ​​of the internal earthquake sensor, vibration sensor, or acceleration sensor are within the normal range and no physical vibration is detected, the level determination module (330) detects a discrepancy between the high risk level conveyed by the disaster message and the low risk level indicated by the sensor data during the cross-verification process. Furthermore, the level determination module (330) calculates the physical distance between the disaster location and the user environment to lower the reliability weight of the disaster message and calculates a disaster index reflecting this. As a result, the disaster level may be downgraded from severe to caution, and customized disaster response information, such as "Earthquake in nearby area, current location safe but caution required," may be displayed to the user.

[0075] Meanwhile, the level determination module (330) is characterized by activating disaster prevention devices within the user environment when the disaster level corresponds to a disaster level above a set threshold. For example, if a disaster level of the warning stage or higher is calculated due to a hazardous gas leak or a severe fine dust situation, the level determination module (330) may output a command to activate disaster prevention devices, such as activating an air purifier within the user environment, switching the operation mode of a ventilation fan, or controlling a window opening / closing device. Furthermore, it is possible to control multiple disaster prevention devices in combination depending on the type and level of the disaster. For example, if a severe stage is calculated due to a hazardous gas leak, beyond simply closing the windows, active device control customized to the type of disaster can be performed, such as changing the air flow path between multiple purification filters in conjunction with an indoor air purification device or controlling the flow load to concentrate on a specific filter. Through this, the present invention can function as a practical disaster defense system beyond a simple disaster notification system.

[0076] Additionally, the level determination module (330) is characterized by collecting the intervention history as feedback data and correcting the learning weight matrix W or the bias value b when an intervention event occurs in which a user manually changes the level of a determined disaster level. This corresponds to a user intervention-based reinforcement learning mechanism, and when an intervention event occurs, such as a user manually adjusting the level upward or manually operating a disaster prevention device regarding the disaster level determined by the AI ​​analysis unit (300), the level determination module (330) collects the user's intervention history as feedback data. For example, if the AI ​​analysis unit (300) determined the level to the caution stage but the user felt anxious and manually adjusted it upward to the alert mode, this user intervention event serves as feedback data indicating that the judgment of the AI ​​analysis unit (300) in that situation was lower than the user's perceived risk. The level determination module (330) accumulates this feedback data and continuously corrects the neuron connection strength of the learning weight matrix W or the sensitivity baseline of the bias value b to match the user environment. Through this, the system of the present invention can function as an evolutionary disaster classification system that becomes increasingly optimized for the characteristics of the user's living environment as time passes.

[0078] The present invention, composed of the above-described configuration, overcomes the limitations of existing methods that are limited to the simple transmission of disaster messages by fusing external disaster messages in the form of unstructured text with structured sensor data within the user environment in a multimodal manner and cross-verifying them, and has the effect of classifying disaster grades by quantifying the impact of the disaster at the user's actual location.

[0079] In addition, the present invention has the effect of improving the accuracy of disaster grade classification in various disaster situations by dynamically calculating weights by comprehensively considering multiple environmental variables, such as the physical distance from the disaster location, the real-time reliability index of the sensor, and the temporal urgency factor, in addition to setting basic weights according to the type of disaster.

[0080] In addition, since the present invention can actively control disaster prevention devices within the user environment according to the disaster level, it has the effect of functioning as a practical disaster defense system beyond a simple notification system.

[0081] In addition, the present invention has the effect of implementing an evolutionary disaster classification system that becomes optimized for the characteristics of the user's living environment as time passes, by collecting the user's manual intervention history as feedback data and continuously correcting the learning weight matrix and bias values.

[0083] The foregoing description of the present invention is for illustrative purposes only, and those skilled in the art will understand that modifications can be easily made to other specific forms while maintaining the technical spirit or essential features of the present invention. Therefore, it should be understood that the embodiments described above are illustrative in all respects and are not intended to exclude variations other than those described.

[0084] For example, the present invention has been described with reference to the embodiments illustrated in the drawings, but this is merely illustrative and those skilled in the art will understand that various modifications to the embodiments are possible therefrom.

[0085] Accordingly, the true technical scope of protection of the present invention is defined by the technical concept of the claims, and all modifications or variations derived from the meaning and scope of the claims and equivalent concepts thereof should be interpreted as being included within the scope of the present invention. Explanation of the symbols

[0087] 10: Sensors in the user environment 100 : Receiver 200 : Collector 300: AI Analysis Unit 310: Text Analysis Module 320: Sensor Analysis Module 330: Level Determination Module

Claims

Claim 1 A receiving unit (100) that receives disaster text messages in the form of unstructured text from an external source; a collecting unit (200) that collects structured sensor data from a plurality of sensors (10) installed within a user environment; and an AI analysis unit (300) that receives the disaster text messages from the receiving unit (100), analyzes the content thereof, receives the sensor data from the collecting unit (200), analyzes the correlation and consistency with the content of the disaster text messages to perform cross-verification, and determines a disaster grade based on the cross-verification results; wherein the AI ​​analysis unit (300) includes a text analysis module (310) that extracts the type of disaster by analyzing the disaster text messages through a natural language processing algorithm, a sensor analysis module (320) that analyzes the sensor data through a time series analysis algorithm, and a level determination module (330) that verifies whether there is a match between the type of disaster extracted from the text analysis module (310) and the sensor data analyzed from the sensor analysis module (320), and determines the level of the disaster grade based on the verification results; wherein the text analysis module (310) numerically The sensor analysis module (320) converts the sensor data to produce a text embedding vector, and the sensor analysis module (320) numerically converts the sensor data to produce a sensor embedding vector. The level determination module (330) applies the reliability weight of the disaster character to the text embedding vector and applies the accuracy weight of the sensor data to the sensor embedding vector, respectively, and fuses them to produce a disaster index for determining the disaster grade level. The level determination module (330) sets the application ratio of the reliability weight of the disaster character or the accuracy weight of the sensor data differently depending on the type of disaster extracted. The disaster index is (Softmax: A function that converts calculated scores into probabilities between 0 and 1, D: Disaster Index, α: Confidence weight of disaster characters, β: Accuracy weight of sensor data, h text : Text embedding vector,h sensor : Sensor embedding vector, A cross-validation-based intelligent disaster classification and warning system characterized by: : vector combination, W: empirical learning weight matrix acquired by the AI ​​model through training, b: bias value, sensitivity index of the system to a specific disaster,). Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 delete Claim 6 A cross-verification-based intelligent disaster classification and warning system, wherein, in claim 1, the level determination module (330) determines the disaster grade by discretizing and calculating the calculated disaster index with a plurality of preset threshold values, and including two or more of the interest level, caution level, warning level, or severe level. Claim 7 In claim 6, the level determination module (330) is characterized by activating a disaster prevention device within the user environment when the disaster level corresponds to a disaster level greater than or equal to a set threshold value, in a cross-verification-based intelligent disaster classification and warning system. Claim 8 In claim 1, the level determination module (330) is characterized by collecting the intervention history as feedback data and correcting the learning weight matrix (W) or the bias value (b) when an intervention event occurs in which a user manually changes the level of the determined disaster level. This describes an intelligent disaster classification and warning system based on cross-validation. Claim 9 A cross-verification-based intelligent disaster classification and warning system, wherein, in claim 1, the level determination module (330) corrects the reliability weight of the disaster text according to the physical distance between the disaster occurrence location within the disaster text and the user environment location, or corrects the accuracy weight of the sensor data using the real-time reliability index of the sensor (10) within the user environment as a variable. Claim 10 A cross-verification-based intelligent disaster classification and warning system, wherein, in claim 1, the text analysis module (310) extracts a temporal urgency factor from the disaster text, and the sensor analysis module (320) variably controls the collection cycle of the sensor data in real time in response to the extracted temporal urgency factor.

Images