Data processing system

Abstract

The present application relates to a nationwide distributed disaster response social infrastructure, which collects and integrates various data such as personnel movement, traffic infrastructure, shelters, material inventory, disaster occurrence, and the like possessed by municipalities and related organizations nationwide in real time, constructs a digital twin that reproduces the real world with high precision on a virtual space, uses generative AI as a must item, accurately predicts and simulates the number of evacuees and the amount of required materials in each shelter, the operation status of traffic infrastructure, and the optimal material distribution plan at the time of disaster occurrence, visualizes and shares information, and evolves the model through federated learning, thereby achieving rapid and accurate support for disaster victims without relying on manual operation or human judgment in the past.

Description

Technical Field

[0001] The technology disclosed herein relates to a system. Background Technology

[0002] Patent Document 1 discloses a personalized chatbot control method executed by at least one processor, the method comprising: receiving user speech; adding the user speech to a prompt containing instructions related to a chatbot role; encoding the prompt; and inputting the encoded prompt into a language model to generate chatbot speech in response to the user speech.

[0003] Patent document 1: Japanese Patent Application Publication No. 2022-180282.

[0004] This invention addresses a complex challenge faced by autonomous regions and related institutions nationwide: the inability to obtain real-time and comprehensive information on population movement, transportation infrastructure, shelter capacity, and material inventory and demand during disasters. Information is often transmitted manually via telephone or fax, leading to delays in information sharing between institutions, mismatches between material needs and inventory levels, difficulty in developing optimal material distribution plans, disparities in response capabilities among autonomous regions, and challenges in flexibly responding to unprecedented disasters or new scenarios. Initial disaster response actions are also delayed, hindering rapid and accurate support for victims. The key to solving these problems lies in combining distributed digital twins with generative AI and federated learning to achieve real-time nationwide information collection, integration, analysis, prediction, visualization, and optimization. This approach eliminates reliance on manual operations or human judgment, enabling the automatic and optimal delivery of necessary materials and support to disaster-stricken areas immediately after a disaster. This significantly improves the disaster response capabilities of autonomous regions and related institutions, and builds a social infrastructure to protect the lives and living environment of disaster victims. Summary of the Invention

[0005] As a means to solve the problem, the following structure is adopted: A data collection and integration department is set up to collect and integrate various data in real time from autonomous regions and related institutions across the country, including data on mobile personnel, transportation infrastructure operation status, shelter capacity, material inventory, and disaster occurrence information; a digital twin generation and updating department is set up to use the data obtained from the data collection and integration department to accurately reproduce the status of real-world cities and regions in virtual space, on a national and regional basis; a disaster response prediction and simulation department is set up to reflect disaster events in real time in the model constructed by the digital twin generation and updating department, using generative AI. Based on past disaster data and real-time data, the system accurately predicts and simulates the shift in the number of evacuees, the required quantity of supplies, the operational status of transportation infrastructure, and the material distribution plan for each shelter. It automatically generates the optimal material distribution plan. The disaster response prediction and simulation unit operates with generative AI as a mandatory component, using federated learning to continuously learn from the dispersed data across the country, enabling it to flexibly respond to unknown disaster patterns and regionally specific challenges. Furthermore, by setting up an information sharing and visualization unit, information such as the material needs, inventory status, distribution plans, and transportation progress of each shelter can be shared and visualized in real time among autonomous regions, the nation, and relevant agencies, thereby achieving rapid and accurate support for disaster victims. Attached Figure Description

[0006] Figure 1 This is a conceptual diagram illustrating an example of the configuration of a data processing system according to the first embodiment.

[0007] Figure 2 This is a conceptual diagram illustrating an example of the main functions of the data processing apparatus and smart device according to the first embodiment.

[0008] Figure 3 This is a conceptual diagram illustrating an example of the data processing system configuration in the second embodiment.

[0009] Figure 4 This is a conceptual diagram illustrating an example of the main functions of the data processing device and smart glasses according to the second embodiment.

[0010] Figure 5 This is a conceptual diagram illustrating an example of the data processing system configuration in the third embodiment.

[0011] Figure 6 This is a conceptual diagram illustrating an example of the main functions of the data processing device and head-mounted terminal according to the third embodiment.

[0012] Figure 7 This is a conceptual diagram illustrating an example of the data processing system configuration in the fourth embodiment.

[0013] Figure 8This is a conceptual diagram illustrating an example of the functions of the main parts of the data processing device and robot according to the fourth embodiment.

[0014] Figure 9 It represents an emotion graph that maps multiple emotions.

[0015] Figure 10 It represents an emotion graph that maps multiple emotions.

[0016] Explanation of reference numerals in the attached figures Data processing systems 10, 210, 310, and 410 12 Data processing device 14 Smart devices 214 Smart Glasses 314 Head-mounted terminal 414 Robot. Detailed Implementation

[0017] Hereinafter, an example of an implementation of the system involved in this disclosure will be described with reference to the accompanying drawings.

[0018] First, let's explain the terms used in the following description.

[0019] In the following embodiments, the processor (hereinafter referred to as "processor") can be a single computing device or a combination of multiple computing devices. Furthermore, a processor can be a single computing device or a combination of multiple computing devices. Examples of computing devices include CPU (Central Processing Unit), GPU (Graphics Processing Unit), GPGPU (General-Purpose computing on Graphics Processing Units), APU (Accelerated Processing Unit), etc.

[0020] In the following implementation, the labeled RAM (Random Access Memory) is a memory that temporarily stores information and is used by the processor as working memory.

[0021] In the following embodiments, the labeled memory is one or more non-volatile storage devices used to store various programs and various parameters, etc. Examples of non-volatile storage devices include flash memory (SSD (Solid State Drive)), disk (e.g., hard disk) or magnetic tape, etc.

[0022] In the following implementation, the labeled Communication I / F (Interface) is an interface that includes a communication processor and an antenna, etc. The Communication I / F is responsible for communication between multiple computers. Examples of communication standards applicable to the Communication I / F include wireless communication standards such as 5G (5th Generation Mobile Communication System), Wi-Fi (registered trademark), or Bluetooth (registered trademark).

[0023] In the following implementation, "A and / or B" is synonymous with "at least one of A and B". That is, "A and / or B" means that it can be only A, only B, or a combination of A and B. Furthermore, in this specification, when "and / or" connects more than three items, the same approach as "A and / or B" applies.

[0024] [First Implementation] Figure 1 An example of the configuration of the data processing system 10 according to the first embodiment is shown.

[0025] like Figure 1 As shown, the data processing system 10 includes a data processing device 12 and an intelligent device 14. An example of the data processing device 12 is a server.

[0026] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0027] The smart device 14 includes a computer 36, a receiver 38, an output device 40, a camera 42, and a communication I / F 44. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the receiver 38, output device 40, camera 42, and communication I / F 44 are also connected to the bus 52.

[0028] The receiving device 38 includes a touchscreen 38A and a microphone 38B, etc., for receiving user input. The touchscreen 38A receives user input generated by contact with an indicator (e.g., a pen or finger). The microphone 38B receives user input generated by sound by detecting the user's voice. The control unit 46A in the processor 46 sends data representing user input received through the touchscreen 38A and microphone 38B to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the data representing the user input.

[0029] The output device 40 includes a display 40A and a speaker 40B, and presents data to the user 20 by outputting data in a form perceptible to the user 20 (e.g., sound and / or text). The display 40A displays visual information such as text and images according to instructions from the processor 46. The speaker 40B outputs sound according to instructions from the processor 46. The camera 42 is a small digital camera equipped with an optical system such as a lens, aperture, and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor.

[0030] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for sending and receiving various information between processor 46 and processor 28 via network 54.

[0031] Figure 2 An example of the main functions of the data processing device 12 and the smart device 14 is shown.

[0032] like Figure 2 As shown, in the data processing apparatus 12, specific processing is performed by the processor 28. A specific processing program 56 is stored in the memory 32. The specific processing program 56 is an example of a "program" as understood in this disclosure. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 based on the specific processing program 56 executed on the RAM 30.

[0033] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290. The specific processing unit 290 can use the emotion-specific model 59 to infer the user's emotions and perform specific processing based on the user's emotions. The emotion inference function using the emotion-specific model 59 (emotion-specific function) includes various inferences and predictions related to the user's emotions, but is not limited to this example. Furthermore, emotion inference and prediction also include, for example, emotion analysis (analysis).

[0034] In the smart device 14, the processor 46 performs the receive / output processing. The memory 50 stores the receive / output program 60. The receive / output program 60 is used in conjunction with the data processing system 10 and the specific processing program 56. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on the RAM 48. The receive / output processing is implemented by the processor 46 acting as a control unit 46A based on the receive / output program 60 executed on the RAM 48. Furthermore, the smart device 14 may also have the same data generation model and emotion-specific model as the data generation model 58 and the emotion-specific model 59, and use these models to perform the same processing as the specific processing unit 290. The receive / output processing is implemented by the processor 46 acting as a control unit 46A based on the receive / output program 60 executed on the RAM 48.

[0035] In addition, other devices besides the data processing device 12 may also have a data generation model 58.

[0036] For example, a server device (e.g., a generation server) may have a data generation model 58.

[0037] In this case, the data processing device 12 obtains the processing results (prediction results, etc.) using the data generation model 58 by communicating with a server device that has the data generation model 58. Furthermore, the data processing device 12 can be a server device or a user-owned terminal device (e.g., a mobile phone, robot, home appliance, etc.). Next, an example of the processing performed by the data processing system 10 of the first embodiment will be described.

[0038] (Example 1) The specific processing flow in Example 1 will be described below. Each part of the system described below is implemented by the data processing device 12 and the smart device 14. Furthermore, the data processing device 12 is referred to as the "server" and the smart device 14 is referred to as the "terminal".

[0039] (Implementation methods for carrying out the present invention) As an embodiment of this invention, a system is constructed that collects and integrates various information held by autonomous regions and related institutions, infrastructure operators, logistics operators, and shelter operators across the country in real time. This system constructs a nationwide, distributed digital twin that accurately reproduces the real world in virtual space. During disasters, generative AI is used as an essential component to accurately predict and simulate the shift in the number of evacuees, the required quantity of supplies, the operational status of transportation infrastructure, and material distribution plans for each shelter, automatically generating the optimal material distribution plan. The following details the various components and methods of utilizing generative AI, as well as the role division between servers and terminals, through numerous specific examples. Specifically, this system operates in a distributed computing environment composed of multiple computing nodes. This system employs a composite artificial intelligence architecture combining large-scale language models and graph neural networks. This system unfolds heterogeneous mixed data asynchronously sent from various institutions as multidimensional tensors into memory. This system uses natural language processing to convert unstructured text data into fixed-length embedding vectors. This system constructs a spatiotemporal graph structure integrating spatial and temporal information. This system performs graph convolution operations. The system also establishes dependencies between computing nodes. This system uses a self-attention mechanism to extract important features. It outputs the shift in the number of refugees in each shelter as a time-series numerical array. It outputs the required quantity of supplies as a probability distribution for each item. It outputs the operational status of transportation infrastructure as edge weights. These outputs are then fed into a reinforcement learning agent. The system performs reward maximization calculations based on Markov decision processes. It automatically generates the optimal material distribution plan as a node transition sequence. The generated plan is then transmitted to a subsequent visualization module. This system completely eliminates existing planning processes that rely on human intuition or experience. It solves nonlinear optimization problems in high-dimensional spaces in milliseconds. It dramatically improves processing speed while suppressing computational resource consumption. It prevents fatal delays in the time-constrained use case of disaster response. This system is applied to logistics optimization during wide-area disasters. It simultaneously evaluates multiple alternative delivery routes. It performs robust optimization considering the probability of road blockage. It uses a linear combination of cross-entropy error and mean squared error as the loss function. It dynamically updates the network weights using gradient descent. This system thus continuously improves prediction accuracy.

[0040] First, the Data Collection and Integration Department collects in real time mobile personnel data (e.g., location information of portable terminals, passenger and boarding history of transportation IC cards, and GPS data of vehicles) held by local governments, transportation operators, infrastructure managers, shelter operators, logistics operators, medical institutions, meteorological observation agencies, and telecommunications operators nationwide; transportation infrastructure operation data (e.g., road closure information, railway or bus operation status, and bridge or tunnel damage status); shelter capacity data (e.g., real-time number of people, age groups, and information on those requiring care from the shelter access management system); material inventory data (e.g., inventory levels, consumption rates, and expiration dates of reserves in various local governments); disaster occurrence information data (e.g., rapid reports from seismographs or weather radar, and disaster reports from SNS or notification applications); medical institution operation status (e.g., number of beds, drug inventory, and number of patients that can be transported); and communication infrastructure operation status (e.g., whether base stations are operational and areas with communication failures). This data is aggregated to the server through various means, including API integration, IoT sensors, data integration with existing business systems, manual input from on-site personnel via terminals, and automatic collection through image or voice recognition. For example, on-site personnel at shelters can input the number of evacuees or required supplies on a tablet terminal, take photos, and upload them; the situation on-site is then instantly reflected on the server. Furthermore, image data of disaster-stricken areas captured by drones or surveillance cameras is automatically collected, and AI-powered image analysis extracts collapsed buildings or flooded areas. Specifically, this data collection and integration department uses a distributed message queue system to receive millions of streaming data messages per second. The received data is temporarily stored in an in-memory database. The department represents mobile personnel data as a 3D vector containing latitude, longitude, and timestamps. It represents transportation infrastructure operational status data as a graph-structured edge attribute tensor. It represents shelter capacity data as an array of integer scalar values. It represents material inventory data as a typical data structure consisting of pairs of item IDs and quantities. This data collection and integration department acquires disaster occurrence information data as natural language text. This department acquires the operational status of medical institutions as multi-dimensional numerical vectors. This department acquires the operational status of communication infrastructure as an array of binary flags. This department maps these heterogeneous data into a common spatiotemporal coordinate system. This department supplements missing values ​​using linear interpolation algorithms. This department eliminates outliers using statistical anomaly detection algorithms. This department normalizes the preprocessed data. This department converts the normalized data into a 512-dimensional feature vector. This department applies convolutional neural networks to image data.This data collection and integration unit receives input image tensors, for example, as a 3-channel RGB image with a resolution of 224x224. It outputs a continuous value score between 0 and 1 representing the probability of building collapse. It also outputs a segmentation mask representing the extent of flooding. Furthermore, it outputs an array of coordinates representing the bounding box of the fire's location. These outputs are then transmitted to the subsequent digital twin generation and update unit. This unit completely eliminates data input delays caused by human intervention. It utilizes a dedicated tensor processing processor to improve the inference speed of machine learning models. This unit significantly enhances the comprehensiveness and real-time performance of data collection. It is applicable to use cases involving situational awareness during wide-area disasters or infrastructure monitoring. It uses a Kalman filter to automatically resolve inconsistencies between multiple data sources. Finally, it mathematically guarantees the reliability of the data.

[0041] Next, the digital twin generation and updating unit uses a variety of data obtained from the data collection and integration unit on the server side to accurately reproduce the state of cities and regions in the real world in virtual space, organized by country and region. For example, it integrates map data obtained from geographic information systems, 3D models of buildings, road networks, terrain data of rivers or mountains, location information of shelters, population distribution data, and the location and operational status of important infrastructure (power plants, substations, water supply and drainage facilities, communication base stations, etc.) to faithfully reproduce the state of the real world. In the event of a disaster, it simulates the impact range of earthquakes, floods, mudslides, tsunamis, fires, power outages, large-scale communication failures, etc., and instantly reflects the disaster scenario, inaccessible sections, isolated areas, the capacity of shelters, the operational status of medical institutions, and the operational status of logistics networks onto the digital twin. For example, in the event of an earthquake with intensity 7, the system instantly calculates and displays predictions of building damage around the epicenter, road blockages, bridge damage, railway shutdowns, the capacity of shelters, the occurrence of isolated villages, and the number of patients that medical institutions can receive, and displays these predictions in virtual space. Furthermore, in the event of typhoons or heavy rains, it also simulates river flooding or predicts the extent of inundation. Specifically, this digital twin generation and update unit constructs a voxel grid data structure in memory to manage three-dimensional spatial information. This digital twin generation and update unit assigns collected feature vectors to each voxel. This digital twin generation and update unit executes a physics engine that mimics the physical laws of the real world. This digital twin generation and update unit uses the finite element method to calculate the propagation of seismic waves. This digital twin generation and update unit simulates the resonance between the natural frequencies of buildings and seismic waves. This digital twin generation and update unit uses a computational fluid dynamics model based on shallow water equations to calculate river flooding. This digital twin generation and update unit inputs these physical simulation results into a graph neural network. This digital twin generation and update unit uses a 1024-dimensional tensor containing the geographic coordinates and attribute information of each node as input data. The unit also inputs an adjacency matrix representing edge connectivity. It dynamically calculates the influence degree between nodes using a graph attention network. As output, it generates numerical vectors representing the probability of functional malfunction for each infrastructure element. It outputs a binary mask tensor representing impassable sections of roads. It outputs a heatmap image representing the probability of isolated areas. These outputs are then provided to the subsequent disaster response prediction and simulation unit. This unit eliminates disaster assumptions based on human imagination. It rigorously calculates complex interactions in high-dimensional space. Compared to existing methods, this unit significantly improves the spatial resolution of the simulation. It offloads the computational load to a distributed processing cluster.This digital twin generation and update unit enables real-time visualization of disaster conditions. It can be applied to urban planning and disaster prevention training. This unit automatically generates scenarios of complex disasters involving multiple cascading events. It uses the Monte Carlo method to quantify uncertainty. It calculates the confidence interval for the prediction.

[0042] The disaster response prediction and simulation department uses generative AI as a mandatory component on the server side. Based on past disaster data and current real-time data, it accurately predicts and simulates the shift in the number of evacuees, the required supplies, the operational status of transportation infrastructure, and the material distribution plan for each shelter. For example, generative AI that has learned from past earthquake evacuation patterns or material consumption patterns, the timing of shelter openings, and the material consumption tendencies of each evacuee's attributes (medicine or nursing supplies for the elderly, formula or diapers for infants, and menstrual products for women, etc.) automatically calculates the number of evacuees or the required quantities of food, water, blankets, medicine, hygiene products, cold-weather gear, generators, fuel, communication equipment, etc., for each shelter in the coming hours to days during this disaster. Furthermore, based on the damage to roads or railways, it also simulates the feasibility or time required for material transportation routes. For example, when major roads are blocked, it automatically assesses the necessity of exploring alternative routes or using helicopters or drones for airlift. Furthermore, generative AI uses federated learning to continuously learn from distributed data across the country, enabling it to flexibly address unknown disaster patterns and regionally specific issues. For example, specific disaster patterns that have occurred in a region in the past, changes in the demand for supplies needed only in specific regions, seasonal factors (heating appliances in winter, cooling equipment or heatstroke prevention supplies in summer), and increased demand for masks or disinfectants during infectious disease outbreaks are all reflected in the model through federated learning. Specifically, this disaster response prediction and simulation unit uses a large language model based on an autoregressive Transformer architecture as its core. This disaster response prediction and simulation unit receives time-series tensors representing past disaster histories as input data. This disaster response prediction and simulation unit receives current real-time data as embedding vectors. This disaster response prediction and simulation unit encodes categorical variables representing shelter attribute information using one-hot encoding and inputs them. This disaster response prediction and simulation unit sets the dimension of the input data to, for example, 768 dimensions. This disaster response prediction and simulation unit uses a multi-head attention mechanism to extract long-term dependencies within the time-series data. This disaster response prediction and simulation unit uses a feedforward neural network to perform nonlinear transformations. The Disaster Response Prediction Simulation Unit outputs a floating-point array representing the shift in evacuation numbers over the next 72 hours. It outputs the required material quantity as a demand prediction score for each item. It outputs the estimated recovery time of transportation infrastructure as a probability distribution. These outputs are then transmitted to the subsequent automatic material distribution plan generation unit. The unit executes a federated learning algorithm. It aggregates the gradient information calculated individually from each region's computing nodes to the central server. It updates the weights of the global model using a federated averaging method. The unit does not transmit raw data over the network. This approach improves the model's generalization performance while protecting privacy.This disaster response prediction simulation completely eliminates demand forecasting based on human experience rules. It discovers hidden correlations in a high-dimensional feature space. It significantly reduces prediction error from the perspective of mean absolute error. It can be applied to use cases such as wide-area evacuation planning or optimal allocation of stockpiled supplies. It incorporates seasonal variations or weather factors as external characteristics into the model. It uses Poisson regression loss as the loss function. It optimizes the prediction accuracy for count data.

[0043] In the automatic generation of supply distribution plans, generative AI automatically generates the optimal supply distribution plan based on predicted supply demand, inventory at each distribution point, road conditions, and the operational status of transportation vehicles such as trucks, helicopters, and drones. For example, if a shelter in City A is short of 5,000 liters of water, 200 blankets, and 100 sets of medicine, the required supplies are gathered from the nearest supply distribution point or the reserve warehouses of other autonomous regions, accessible roads are selected, and trucks are dispatched. If major roads are blocked, alternative routes are explored, and air transport via helicopters or drones is combined as needed. Simultaneously, deliveries to multiple shelters are optimized, automatically calculating delivery routes and sequences, the number of vehicles required, estimated arrival times, loading plans, driver allocation, and fuel replenishment plans. Push-based support is also supported; without waiting for requests from autonomous regions, generative AI automatically arranges essential supplies according to priority. For example, immediately after an earthquake, based on the number of people in shelters and the rate of supply consumption, the system automatically arranges the delivery of necessities such as food, water, blankets, and medicine, instantly formulating a plan for delivery along the optimal route. Furthermore, in the event of trucks encountering congestion or traffic restrictions during delivery, the route is recalculated in real time, and relevant personnel are notified. Specifically, this automatic material distribution plan generation unit executes a deep reinforcement learning algorithm. This unit constructs an environment model based on a Markov decision process. This unit defines the state space as the vector of material shortages for each shelter and the tensor of the current position coordinates of each transport vehicle. This unit defines the action space as the combination of the next destination node for each vehicle and the load mass. This unit sets the reward function as a linear combination of the negative penalty for material arrival delay time and the negative penalty for transportation cost. This unit trains the policy network using a proximal policy optimization algorithm. This unit receives a 256-dimensional state vector representing the current state as input. This unit generates a Softmax distribution representing the probability of choosing each action as output. This unit outputs a baseline prediction of the scalar value representing the value of the current state. This unit outputs a sequence of node IDs representing the specific delivery route. This automated material distribution plan generation unit transmits the output results to the subsequent information sharing and visualization unit. This unit completely eliminates intuitive vehicle allocation planning performed by human dispatchers. This unit solves the combinatorial explosion-prone traveling salesman problem at high speed using a hybrid approach of heuristic search and neural networks. This unit reduces computation time from exponential to polynomial levels. This unit is applicable to emergency transportation during disasters or last-mile delivery use cases. This unit immediately recalculates when road closure information is updated.This automatic material distribution plan generation unit implements dynamic routing. It optimizes and combines multimodal transport plans using multiple modes of transport. It also employs a constraint satisfaction problem solver. This automatic material distribution plan generation unit guarantees solutions that rigorously satisfy hard constraints such as loading weight limits or driver working time limits.

[0044] The Information Sharing and Visualization Department distributes information generated on the server side, such as the material needs, inventory status, delivery plans, and transportation progress of each shelter, to the dashboard or map display screen on the terminal side in real time. For example, the material shortage status of shelters nationwide, the current location and estimated arrival time of delivery trucks, the shift in the number of evacuees in each shelter, material consumption patterns, the operational status of transportation infrastructure, the number of patients that medical institutions can receive, and the status of isolated villages are visualized on the map. Relevant personnel can instantly grasp the situation from terminals (PCs, tablets, smartphones, etc.) and make decisions such as supplementary arrangements or route changes as needed. For example, in the event of a sharp increase in the number of evacuees in a shelter, when the on-site personnel input information from the terminal, the generative AI on the server side immediately predicts the new material needs, automatically updates the delivery plan, and notifies the terminals of relevant personnel. In addition, the dashboard can also display the material consumption rate or remaining inventory of each shelter graphically, visualizing future shortage predictions or replenishment opportunities. Specifically, this Information Sharing and Visualization Department uses the WebSocket protocol to establish a two-way real-time communication session between the server and the terminal. This information sharing visualization department continuously sends JSON-formatted data streams to the front-end rendering engine. It uses a map drawing library to plot vector tile data on the screen. It configures icons representing the status of each shelter based on geographic coordinates. It overlays material shortage status as color-coded heatmaps. It dynamically animates the current location of delivery trucks based on time-series data from GPS coordinates. It displays the output from generative AI as a natural language summary text on the side panel of the screen. It receives screen operation events from the user as input. It parses the operation events and calculates the bounding box coordinates of the display area. It selectively retrieves data from the server that is only contained within the calculated coordinate range. As output, it generates an optimized graphical user interface on the user's terminal screen. As output, it generates a pop-up window displaying detailed information when a specific shelter is clicked. As output, it generates a time-series line graph representing the rate of material consumption. This information sharing visualization department provides these outputs directly to the user's visual perception process. This information sharing visualization department completely eliminates the need for manual statistical analysis of massive amounts of raw data. It dramatically improves the visual recognizability and speed of understanding of information. It implements data compression algorithms to minimize network bandwidth consumption. This information sharing visualization department is applicable to use cases such as large-screen displays at disaster response headquarters or mobile terminal displays for on-site personnel.This information sharing visualization unit dynamically adjusts the number of polygons drawn based on the terminal's screen resolution or processing power. This information sharing visualization unit performs level-of-detail control. This information sharing visualization unit always maintains the responsiveness of the user interface in milliseconds.

[0045] Furthermore, disaster response performance data and new event data are continuously learned on the server side through federated learning, evolving the simulation accuracy and predictive model of generative AI. For example, the AI ​​learns the actual amount of supplies needed or delivery time required during this disaster, the movement of evacuees, the reception status of medical institutions, and the recovery speed of communication infrastructure, enabling more accurate predictions and plans for future disasters. The characteristics of each region and past lessons are also reflected in the model. For instance, it learns the increased demand for heating appliances or cold-weather gear in snowy areas during winter, the likelihood of delivery delays due to traffic congestion in urban areas, the potential for isolated villages in mountainous regions, and the increased demand for shelters or hygiene products during infectious disease outbreaks, automatically generating optimal plans that take these factors into account for the next disaster. Specifically, this federated learning evolution unit operates on a distributed machine learning framework. It distributes initialized neural network weights to the computing nodes in each region. The unit performs learning using gradient descent on each computing node using locally collected disaster response performance data. This federated learning evolution unit uses numerical vectors representing actual material consumption or scalar values ​​representing delivery time as input data. It calculates the mean squared error between predicted and actual values ​​as the loss function. The unit uses backpropagation to calculate the weight gradients of each layer. It encrypts the calculated gradient tensors. It sends the encrypted gradient tensors to a central aggregation server. Secure multi-party computation is performed on the aggregation server. The unit performs an additive average without decrypting the gradient tensors from each node. As output, it generates an updated global model weight parameter tensor. As output, it generates a graph representing the learning progress of each node, showing the loss values. As output, it generates an evaluation metric score representing the model's generalization performance. The unit applies these outputs to the next inference loop. This unit completely eliminates existing centralized learning methods that centralize raw data in one place. It simultaneously achieves data privacy protection and saves communication bandwidth. This federated learning evolutionary unit incorporates personalized federated learning algorithms that consider region-specific data distribution biases. It dynamically blends and optimizes local models tailored to regional characteristics with a general global model. This unit improves the accuracy of zero-shot inference for unknown disaster patterns. It is applicable to use cases such as building nationwide disaster prevention networks or wide-area data sharing in healthcare. It detects poisoning attacks from malicious nodes. It automatically eliminates anomalous gradient vectors through a robust aggregation algorithm. This unit mathematically guarantees the model's security and integrity.

[0046] Furthermore, generative AI plays a crucial role as the core of the entire system in various aspects, including predicting the number of people seeking refuge or the demand for supplies, optimizing distribution plans, automatically summarizing information or providing decision support, model evolution through federated learning, automating information collection during disasters (extracting information from SNS or notification applications, and assessing the disaster situation through image recognition), providing suggestions and prompts to on-site personnel (e.g., "Due to the large number of elderly people in this shelter, please prioritize replenishing medicines and nursing supplies"), automatically generating or customizing disaster response guidelines, and supporting communication among relevant personnel (through chatbot consultation or decision support). Specifically, the generative AI constituting the core of the system is installed as a massive Transformer model with hundreds of billions of parameters. This generative AI has the ability to process multimodal input data. This generative AI receives text prompts representing the shelter's condition as input. This generative AI receives image tensors representing the on-site disaster situation as input. This generative AI receives time-series numerical arrays from sensors as input. This generative AI projects these heterogeneous data into a common latent space vector. This generative AI uses a cross-attention mechanism to fuse the feature quantities of text and images. This generative AI generates structured JSON data representing the predicted material needs of each shelter as output. It also generates natural language text representing specific action instructions for on-site personnel. Furthermore, it generates Markdown-formatted documents representing a summary report of the disaster situation. These outputs are asynchronously distributed to various sub-modules within the system. This generative AI completely replaces the information interpretation or judgment process performed by human operators. It performs complex reasoning tasks in parallel while maintaining high-dimensional contextual information. It minimizes processing latency. This generative AI is applied to the dynamic generation of disaster response guidelines or through chatbot conversational interfaces. It integrates retrieval-enhanced generation techniques. It extracts relevant past disaster cases from external vector databases through similarity retrieval. It appends the extracted cases as contextual clues. It probabilistically suppresses the generation of unfounded outputs, often referred to as illusions. It guarantees the reliability of the outputs from an information retrieval perspective. It uses human feedback from reinforcement learning to align the model's outputs with human intent. This generative AI ensures ethical and safe decision support.

[0047] Thus, the system of this invention, by employing a structure where the server side primarily handles data collection, integration, analysis, simulation, plan generation, and learning, while the terminal side handles data input from the field, status monitoring via dashboards, decision support, notification reception, uploading of on-site photos or voice data, alarm notifications to on-site personnel, and on-site material receipt confirmation or delivery completion reports, enables real-time and optimized nationwide disaster response, fundamentally solving the challenges faced by autonomous regions and related institutions. Generative AI, as the core of the entire system, does not rely on existing manual operations or human judgment, becoming a unique social infrastructure for achieving rapid and accurate support to disaster victims. Specifically, this system adopts a layered distributed processing architecture that links the cloud computing environment with edge devices. On the server side, the system performs large-scale matrix operations or deep learning model inference tasks on a cluster of graphics processing units. On the terminal side, the system performs edge computing using a lightweight inference model. The system receives asynchronous data streams from on-site terminals as input. The system receives high-frequency sampled data from various sensors as input. The system outputs a system log representing the overall system operation status. This system outputs metrics representing the processing latency of each submodule. It also outputs a multi-dimensional data structure representing the overall optimized disaster response plan. These outputs are transmitted to the system administrator's monitoring dashboard. The system completely eliminates human-based communication methods such as manual data transcription or telephone calls. It ensures data consistency and timeliness throughout the system. The system employs a microservice architecture. Each functional module is deployed as an independent container. The system dynamically and automatically scales the number of container replicas based on load changes. The system maintains system availability above 99.99% during periods of concentrated bursts of access. This system is applicable to use cases involving nationwide, wide-area disaster response or large-scale event crowd control. The system does not use generative AI merely as a simple text generation tool, but rather as a control orchestrator for the entire system. The system uses AI to autonomously route data flows between modules. The system uses AI to dynamically optimize hardware resource allocation. This system significantly enhances the autonomy and self-healing capabilities of the computer system itself. This system represents a fundamental improvement in information processing technology.

[0048] (System Composition) The system involved in this implementation includes a data collection and integration department, a digital twin generation and update department, a disaster response prediction and simulation department, a material distribution plan automatic generation department, an information sharing and visualization department, a federated learning and evolution department, an on-site input support department, an alarm notification department, and a generative AI prompt management department.

[0049] The Data Collection and Integration Department collects real-time data from various entities across the country, including autonomous regions, transportation operators, infrastructure managers, shelter operators, logistics operators, medical institutions, meteorological observation agencies, and telecommunications operators. This data includes mobile personnel data, transportation infrastructure operation data, shelter occupancy data, material inventory data, disaster occurrence information data, medical institution operation status data, telecommunications infrastructure operation status data, disaster notification data from SNS or notification applications, and image data obtained through drones or surveillance cameras. This data is aggregated to the server through API linkage, IoT sensors, data linkage with existing business systems, manual input from on-site personnel at terminals, and automatic collection through image recognition or voice recognition. After encryption and anonymization, the data is integrated. For example, examples include: on-site personnel at shelters inputting the number of evacuees or required supplies on a tablet terminal, taking photos and uploading them, thus instantly reflecting the situation on the server; automatically obtaining information on prohibited passage zones or congestion from road management systems, and the current location of trucks; obtaining real-time disaster information such as earthquakes or typhoons from meteorological observation systems; and extracting collapsed buildings or flooded areas from aerial images of disaster-stricken areas taken by drones through AI image analysis. Specifically, this data collection and integration department is equipped with a distributed stream processing engine. This data collection and integration department serializes and receives payloads sent from various data sources in protocol buffer format. This data collection and integration department receives an N-row, 3-column floating-point tensor containing mobile personnel data as input data. This data collection and integration department receives 4K resolution video streams from drones as input data. This data collection and integration department receives variable-length string arrays containing text submissions from social networking sites as input data. This data collection and integration department uses advanced encryption standards algorithms to perform end-to-end encryption on this data. This data collection and integration department uses hash functions to irreversibly anonymize personal information. This data collection and integration unit applies a single-pass real-time object detection neural network to image data. It generates coordinate tensors of bounding boxes surrounding collapsed buildings within the image as output. It also generates pairs of category labels and confidence scores for each bounding box as output. Furthermore, it generates frequency vectors of disaster keywords extracted from SNS text as output. These outputs are then loaded into a subsequent data warehouse. This unit completely eliminates image verification by human visual inspection or manual data cleaning. Initial filtering is performed at the edge device level. This unit significantly reduces network bandwidth consumption on the central server. This unit is applicable to smart city surveillance systems or wide-area infrastructure maintenance use cases. It uses a dynamic time warping algorithm to synchronize time-series data with different sampling rates along the time axis.This data collection and integration department improves the accuracy of correlation analysis between heterogeneous data. This department implements the Isolation Forest algorithm as an anomaly detection model. This department blocks illegal data caused by sensor malfunctions or noise in real time.

[0050] The Digital Twin Generation and Update Department uses diverse data obtained from the Data Collection and Integration Department to accurately reproduce the state of real-world cities and regions in virtual space, on a national and regional basis. For example, it integrates map data from Geographic Information Systems, 3D models of buildings, road networks, terrain data of rivers or mountains, location information of shelters, population distribution data, and the location and operational status of critical infrastructure to faithfully reproduce the state of the real world. In the event of disasters, it simulates the impact range of earthquakes, floods, mudslides, tsunamis, fires, power outages, and large-scale communication failures, reflecting disaster scenarios, impassable sections, isolated areas, the capacity of shelters, the operational status of medical institutions, and the operational status of logistics networks in real time onto the digital twin. Examples include: in the event of an earthquake of intensity 7, real-time calculations of damage predictions for buildings around the epicenter, road blockages, bridge damage, railway shutdowns, shelter capacity, the occurrence of isolated villages, and the number of patients that medical institutions can receive, and displaying these predictions in virtual space; simulating river flooding or predicting inundation areas during typhoons or heavy rains; visualizing the expected recovery of power infrastructure during power outages; and reflecting the operational status of communication infrastructure on a map during communication failures. Specifically, this digital twin generation and update unit uses an octree algorithm as the spatial index structure to perform hierarchical segmentation and management of the three-dimensional virtual space. This digital twin generation and update unit receives point cloud tensors containing the three-dimensional coordinates and attribute information of each infrastructure as input data. This digital twin generation and update unit receives grid data of wind speed and precipitation from a meteorological simulation model as input data. This digital twin generation and update unit receives acceleration waveform data from a seismograph as input data. This digital twin generation and update unit integrates these input data and performs multiphysics simulations. This digital twin generation and update unit generates a continuous score from 0 to 100 representing the structural damage degree of each building as output. It also generates a weight matrix representing the passability probability of each edge on the road network as output. Furthermore, it generates a two-dimensional floating-point matrix representing the spatial distribution of flooding depth as output. These outputs are then forwarded to the subsequent disaster response prediction simulation unit and information sharing visualization unit. This unit completely eliminates existing disaster simulation methods based on static two-dimensional maps. It maximizes the parallel computing capabilities of the graphics processing unit. It performs complex physical calculations involving the interaction of millions of objects in real time. This unit is applicable to use cases such as building virtual testing environments for autonomous vehicles or environmental assessments for urban development. Finally, it uses the Markov chain Monte Carlo method to perform sensitivity analysis on the initial values ​​of the simulation.This digital twin generation and update unit probabilistically generates the worst-case and best-case scenarios for disaster prediction. It supports decision-making under conditions of uncertainty based on mathematical principles. This unit updates the time-series difference data only in memory. This unit minimizes the consumption of computational resources.

[0051] The Disaster Response Prediction and Simulation Department uses generative AI as a mandatory component. Based on past disaster data and current real-time data, it accurately predicts and simulates the shift in the number of evacuees, the required supplies, the operational status of transportation infrastructure, and the material distribution plan for each shelter. For example, generative AI that has learned from past earthquake evacuation patterns or material consumption patterns, the timing of shelter opening, and the material consumption tendencies of each evacuee can automatically calculate the number of evacuees or the required quantities of food, water, blankets, medicine, hygiene products, cold-weather gear, generators, fuel, and communication equipment for each shelter in the coming hours to days during this disaster. Examples include simulating the feasibility or time required for material transport routes based on the damage to roads or railways; automatically determining the necessity of exploring alternative routes or airlifting by helicopters or drones when major roads are blocked; and predicting the need for dense shelter evacuation or increased demand for hygiene products during infectious disease outbreaks. Specific examples of prompts that generative AI can read include: "Based on the latest shelter entry and exit data, material inventory data, and road traffic conditions, predict the required quantity of supplies and the optimal delivery route for each shelter in the next 24 hours," "Based on earthquake disaster data from the past 10 years and the current population distribution and transportation infrastructure operation status of the affected areas, simulate the shift in the number of evacuees and changes in material demand in the next 72 hours," and "Given the large number of elderly people and infants in shelters in City A, predict the required quantities of medicines, formula, diapers, and nursing supplies, and develop the optimal delivery plan." Specifically, this disaster response prediction simulation unit employs a hybrid architecture integrating a Long Short-Term Memory network (a type of recurrent neural network) and a large-scale language model. The unit receives multivariate time-series tensors representing the shift in the number of evacuees during past disasters as input data. It also receives context vectors representing current weather conditions or infrastructure damage as input data. Finally, it receives prompts written in natural language as input data. The disaster response prediction and simulation department converts the prompt text into an array of integer indices using a word segmenter. It generates hourly evacuation number predictions for each shelter over the next 72 hours as a floating-point matrix as output. It also generates character-based data consisting of pairs of required material item IDs and quantities as output. Finally, it generates a natural language text string explaining the basis of the predictions as output. These outputs are then forwarded to the subsequent material distribution plan automatic generation department. The department completely excludes heuristic demand forecasting based on human experts. It models the complex interaction between past patterns and the current situation in a high-dimensional nonlinear space.This disaster response prediction simulation significantly reduces the mean squared error of predictions compared to existing statistical models. It is applicable to use cases such as pandemic forecasting for infectious diseases or supply chain demand forecasting. This simulation analyzes attention weights. It determines which input feature contributes most to the prediction results. It ensures the interpretability of deep learning models that are easily black-boxed. It uses Dropout to calculate the uncertainty during inference in a Bayesian inference manner. It dynamically outputs the confidence intervals of the predictions.

[0052] The automated material distribution plan generation unit automatically generates optimal material distribution plans based on generative AI-predicted material demand, inventory at each distribution point, road conditions, and the operational status of transportation vehicles such as trucks, helicopters, and drones. For example, in a scenario where a shelter in City A is short of 5000 liters of water, 200 blankets, and 100 sets of medicine, the unit gathers the required materials from the nearest distribution point or other municipal reserve warehouses, selects passable roads, and arranges trucks. It also explores alternative routes when major roads are blocked and combines air transport via helicopters or drones as needed. Furthermore, it optimizes deliveries to multiple shelters, automatically calculating delivery routes and sequences, the number of vehicles required, estimated arrival times, loading plans, driver allocation, and refueling plans. In cases where trucks encounter congestion or road closures, the unit recalculates routes in real time and notifies relevant personnel. Specifically, this automated material distribution plan generation unit executes a route optimization engine that combines graph neural networks and deep reinforcement learning. The unit receives a matrix of node feature values ​​representing the material demand of each shelter as input data. This automated material distribution plan generation unit receives marginal feature tensors representing road accessibility probabilities and travel times as input data. It also receives state vectors representing the loading capacity and current location of each transport vehicle as input data. This unit uses the Actor-Critic algorithm to learn the optimal strategy. It generates a sequence of integer values ​​representing the IDs of the nodes each vehicle should visit next as output. It generates a vector representing the continuous values ​​of the amount of material unloaded at each node as output. It generates a scalar value representing the overall plan's evaluation score as output. This unit forwards these outputs to the subsequent information sharing and visualization unit and alarm notification unit. This unit completely excludes routing based on rules of thumb by human vehicle allocation personnel. It recalculates the plan in milliseconds for dynamically changing traffic conditions or new material demands. This unit addresses the trade-off between transportation costs and delivery delays from a Pareto optimal perspective. This automated material distribution plan generation unit is applied to use cases such as dynamic delivery route optimization for express delivery or vehicle allocation systems for on-demand public transportation. It integrates a constraint programming solver into the output layer of a neural network. This unit guarantees executable solutions that absolutely do not violate hard physical and legal constraints such as maximum vehicle capacity or driver continuous driving time limits. It incorporates multi-agent reinforcement learning. This unit achieves autonomous, decentralized control where multiple vehicles coordinate with each other to maximize overall delivery efficiency.

[0053] The Information Sharing and Visualization Department distributes information generated on the server side, such as the material needs, inventory status, delivery plans, and transportation progress of each shelter, to the dashboard or map display screen on the terminal side in real time. Examples include visualizing the material shortage status of shelters nationwide, the current location and estimated arrival time of delivery trucks, the shift in the number of refugees in each shelter, material consumption patterns, the operational status of transportation infrastructure, the number of patients that medical institutions can receive, and the status of isolated villages on a map. Relevant personnel can instantly grasp the situation from the terminal and make decisions such as supplementary arrangements or route changes as needed. The dashboard can also chart the material consumption rate or inventory balance of each shelter, visualizing future shortage predictions or replenishment opportunities. When on-site personnel input information from the terminal, the server-side generative AI instantly predicts new material needs, automatically updates the delivery plan, and notifies relevant personnel's terminals. Specifically, this Information Sharing and Visualization Department constructs an asynchronous data stream processing pipeline based on a reactive programming paradigm. This department receives JSON-formatted update event streams pushed and distributed from the server side as input data. This information sharing visualization unit receives normalized scalar values ​​between 0 and 1, representing the sufficiency rate of supplies at each shelter, as input data. It also receives a time-series array of GPS coordinates for delivery vehicles as input data. This unit stores this data in the client-side memory state management storage. It generates a difference tree of a virtual document object model as output. It generates an array of vertex coordinates for a polygon layer on a map as output. It generates control point parameters for a Bézier curve representing supply consumption trends as output. This unit renders these outputs on a physical display using a graphics rendering API. This unit completely eliminates the need for manual data collection from multiple systems and aggregation using spreadsheet software. It instantly links data changes to visual elements on the screen. It keeps the user interface rendering latency within a standard 60 frames per second. This unit is applicable to use cases such as real-time monitoring dashboards in financial markets or IoT device monitoring systems in factories. It considers the constraints of client-side computing resources. It skips the rendering of objects that exist outside the screen through culling. This information sharing visualization department uses clustering algorithms to visually aggregate a large number of data points. It prevents browser memory leaks, ensuring stable operation over extended periods. Furthermore, it implements layout optimization algorithms that minimize user eye movement.

[0054] The Federated Learning Evolution Department continuously learns from disaster response performance data and new event data on the server side through federated learning, thereby evolving the simulation accuracy and predictive models of generative AI. For example, it can learn from the actual quantity of supplies or delivery time required during a disaster, the movement of evacuees, the reception status of medical institutions, and the recovery speed of communication infrastructure, enabling more accurate predictions and planning for future disasters. It can also learn from the characteristics of each region, such as increased demand for heating appliances or cold-weather gear in snowy winter areas, delivery delays due to traffic congestion in urban areas, the potential for isolated villages in mountainous regions, and increased demand for shelters or hygiene products during infectious disease outbreaks, automatically generating optimal plans that take these factors into account for the next disaster. Furthermore, it can integrate scattered data from across the country through federated learning to flexibly address unknown disaster patterns or region-specific issues, and automatically generate or customize disaster response guidelines. Specifically, this Federated Learning Evolution Department employs a distributed parameter server architecture with a secure multi-party computation protocol installed. This federated learning evolution unit receives the weight gradient tensors of the neural network computed at each edge node as input data. This federated learning evolution unit receives scalar values ​​representing the size of the learning data for each node as input data. This federated learning evolution unit receives the loss values ​​for validation data against the local model as input data. This federated learning evolution unit performs computations while homomorphically encrypting these input data. This federated learning evolution unit generates a aggregated global model weight parameter tensor as output. This federated learning evolution unit generates a vector representing the Shapley values ​​of each node as output. This federated learning evolution unit generates a list of hyperparameter settings to be used in the next learning round as output. This federated learning evolution unit broadcasts these outputs to all participating nodes on the network. This federated learning evolution unit completely eliminates the existing centralized data lake approach of aggregating highly confidential personal information or medical data to a central server. This federated learning evolution unit physically prevents data from leaking from edge devices to the outside world. This federated learning evolution unit simultaneously addresses the opposing challenges of privacy protection and performance improvement of machine learning models. This federated learning evolution unit is applied to use cases such as collaborative learning of medical image diagnostic models among multiple hospitals or sharing of improper detection models among financial institutions. This federated learning evolution unit overcomes communication bandwidth constraints. This federated learning evolution unit applies sparsity and quantization algorithms to the gradient tensor. This federated learning evolution unit compresses the amount of communication data to less than 1% of the original size. This federated learning evolution unit applies a federated proximal algorithm to edge node groups with non-independent and identically distributed data characteristics. This federated learning evolution unit suppresses overfitting of local models and dramatically improves the convergence speed of the global model.

[0055] The Field Input Support Unit provides an interface for field personnel to input and upload information such as the status of shelters, inventory of supplies, disaster conditions, photos, or voice data from terminals, and sends the input data to the server. Examples include inputting the number or attributes of shelters, the types and quantities of required supplies, photos or videos of the scene, voice memos, detailed descriptions of the disaster conditions, etc., from tablets or smartphones; confirming on-site material receipt or reporting delivery completion from terminals; field personnel quickly registering information using voice input or image recognition; and confirming on-site alarm notifications or instructions for receipt on terminals. Specifically, this Field Input Support Unit operates as a native application closely integrated with the hardware sensors of the mobile terminal. It receives PCM format audio streams representing the user's voice as input data. It also receives RGB image tensors from the terminal's camera as input data. Furthermore, it receives trajectory data of gesture coordinates from the touchscreen as input data. The Field Input Support Unit executes a lightweight speech recognition neural network at the edge. The on-site input support unit transcribes speech into text and generates structured JSON objects as output. It also generates feature vectors extracted from images as output. Furthermore, it generates transaction records with timestamps indicating input completion. These outputs are sent to a central server via an asynchronous communication queue. The unit completely eliminates the need for handwritten input on paper forms or tedious keyboard-based text input. Even in offline environments, it temporarily stores data in a local database. It automatically performs synchronization processing the moment network connectivity is restored. It prevents data loss in harsh environments such as disaster-stricken areas with unstable communication infrastructure. The unit is used in construction site progress management systems or for recording on-site care. It performs real-time input completion and error checking on input text data using natural language processing. It ensures the quality of input data at the edge. Finally, it uses optical character recognition (OCR) technology to automatically read barcodes or product names printed on cardboard boxes of relief supplies. The on-site input support department has reduced the input load for inventory management to the limit.

[0056] The Alert Notification Department sends real-time notifications to relevant personnel's terminals regarding important information or anomaly detections from generative AI or the system as a whole, changes in delivery plans, a sharp increase in the number of people in shelters, predicted shortages of supplies, road closures, medical facility access restrictions, and communication failures. Examples include sending alerts for additional supplies when shelters experience a sharp increase in refugees, route change instructions when delivery trucks encounter road closures, delivery instructions to nearby medical facilities when medical facility access restrictions occur, and alerts prompting the use of alternative communication methods in the event of a communication failure. Specifically, this Alert Notification Department uses an event-driven messaging architecture based on a publish-subscribe model. It receives JSON payloads of anomaly detection events issued by various modules within the system as input data. It also receives device tokens and current location coordinates from each user's terminal as input data. Furthermore, it receives profile data representing the user's role or permissions as input data. Finally, it executes a rule engine and a machine learning-based importance determination model. This alert notification unit generates a formatted message payload as output for sending to the push notification service. It also generates a probability score between 0 and 1, representing the urgency of the alert. Furthermore, it generates a log record representing the notification's distribution history. These outputs are distributed to the target terminals with a delay of milliseconds. This alert notification unit completely eliminates the inefficient monitoring task of constantly visually inspecting screens to find anomalies. It prevents excessive notifications, known as alert storms. It aggregates multiple similar alert events into a single notification based on temporal and spatial proximity. It significantly reduces the cognitive load on users. This alert notification unit is applicable to use cases such as anomaly detection systems in factories or unauthorized access detection in financial systems. It tracks whether the user has enabled the notification confirmation signal. If the notification is not enabled within a certain time, it autonomously performs a fallback process to SMS or automated voice calls as alternative communication methods. This alert notification unit completely prevents the omission of critical information at the system level.

[0057] The Generative AI Prompt Management Department manages the prompts required when utilizing generative AI within the overall system and provides the optimal prompts to the generative AI based on the situation. For example, it manages prompts for forecasting material demand or developing distribution plans for each shelter, prompts for simulated instructions combining past disaster data with the current situation, prompts for automated responses to inquiries from on-site personnel, and prompts for automatically generated instructions in disaster response guidelines, and provides these prompts to the generative AI as needed. Specific examples of prompts could include: "Based on the latest shelter entry / exit data, material inventory data, and road traffic data, predict the required quantity of supplies and the optimal delivery route for each shelter over the next 24 hours," "Given the large number of elderly people and infants in shelters in City A, predict the required quantities of medicine, formula, diapers, and nursing supplies, and develop the optimal delivery plan," "Based on earthquake disaster data from the past 10 years and the current population distribution and transportation infrastructure operation status of the affected areas, simulate the shift in the number of evacuees and changes in material demand over the next 72 hours," and "Analyze disaster report photos from the scene, extract collapsed buildings or flooded areas, and summarize the disaster situation," etc. Specifically, this generative AI prompt management unit possesses a dynamic template engine that automates the prompting process. This generative AI prompt management unit receives a dictionary of context variables representing the current system state as input data. This generative AI prompt management unit receives a query string in natural language representing the user's request intent as input data. This generative AI prompt management unit receives an evaluation metric representing the success rate of past prompts as input data. This generative AI hint management unit uses a vector database to retrieve similar past best hint templates. It generates optimized hint strings as output for input into a large language model. It also generates a list of parameters embedded in the hints as output. Finally, it generates integer values ​​representing the number of tokens in the generated hints as output. These outputs are then sent to the generative AI's inference API. This unit completely eliminates the human intervention required for manually adjusting hints through trial and error. It automatically summarizes and truncates input data to avoid exceeding the context window limit. It maximizes output quality while minimizing the inference cost of large language models. This unit is applicable to use cases such as automated response systems for customer support or automated code generation tools. It implements security filters to detect hint injection attacks. It neutralizes malicious input strings. It blocks the risk of unexpected actions caused by the AI ​​model at system boundaries. This generative AI prompt management system uses reinforcement learning to continuously optimize the composition of prompts.

[0058] In this way, the system involved in this implementation, through the coordinated operation of its various components, enables real-time and optimized nationwide disaster response. Utilizing generative AI as its core, it eliminates reliance on traditional manual operations or human judgment, becoming a unique social infrastructure for providing rapid and accurate support to disaster victims. Specifically, this system forms the aforementioned sub-modules as a loosely coupled, autonomous, decentralized information processing infrastructure. As input data, this system receives terabytes of streaming data per second from a nationwide sensor network. As input data, this system receives asynchronous transaction data from tens of thousands of on-site personnel. As input data, this system receives periodic batch data from open data APIs of external organizations. This system processes this massive amount of data in real time. As output data, this system generates a multi-dimensional decision tree representing a national-level disaster response strategy. As output data, this system generates an optimized matrix representing resource allocation to various autonomous regions. As output data, this system generates a health check report representing the self-diagnostic results of the entire system. This system automatically links these output results to the core systems of the government or autonomous regions. This system is capable of processing information at a scale and speed exceeding the limits of human cognition. This system integrates previously isolated disaster prevention systems into a single intelligent platform. It enables data-driven decision-making to be implemented in society. Its applications include operating systems for smart nations and global climate change response platforms. This system does not merely position generative AI as a dialogue agent, but rather as a cognitive engine that controls the entire system. It highly integrates hardware computing power with software algorithms. This system embodies a paradigm shift in computer technology itself. In today's highly uncertain society, this system provides the most robust and resilient technological solutions for protecting life and property. Through continuous machine learning, its performance improves indefinitely over time.

[0059] (Implementation steps) Step 1: Data Collection and Integration. Real-time collection of various data from local governments, transportation operators, infrastructure managers, shelter operators, logistics operators, medical institutions, meteorological observation agencies, and telecommunications operators nationwide, including mobile personnel data, transportation infrastructure operation status data, shelter occupancy data, material inventory data, disaster occurrence information data, medical institution operation status, telecommunications infrastructure operation status, disaster notification data from SNS or reporting applications, and image data from drones or surveillance cameras. This data is aggregated to the server through API integration, IoT sensors, data integration with existing business systems, manual input from on-site personnel, and automatic collection via image or voice recognition. After encryption and anonymization, the data is integrated. Examples include: on-site personnel at shelters inputting the number of evacuees or required supplies on a tablet terminal and uploading photos, thus instantly reflecting the situation on-site to the server; automatically obtaining information on prohibited traffic zones or congestion from road management systems, and the current location of trucks; obtaining real-time disaster information such as earthquakes or typhoons from meteorological observation systems; and extracting collapsed buildings or flooded areas from aerial images of disaster-stricken areas taken by drones through AI image analysis. Specifically, this data collection and integration department performs the processing of mapping heterogeneous, mixed multimodal data to a common high-dimensional vector space. This system receives data streams asynchronously sent from various data sources. This processing department automatically applies normalization and missing value imputation preprocessing to the received raw data. This system uses an autoencoder-type neural network to perform data cleaning. Input data to this neural network includes, for example, one-dimensional time-series arrays of temperature or humidity obtained from various sensors, and RGB image tensors with a resolution of 1920×1080 obtained from drones. This neural network removes noise from these input data. The output data from this neural network consists of a noise-removed 512-dimensional feature vector and a probability score representing outliers. If the probability score exceeds a predetermined threshold, this processing unit discards the data as an outlier. The system forwards the feature vector deemed normal to the subsequent digital twin generation module. This processing unit uses a natural language processing model to parse text data from social networking sites (SNS). The input data to this model is word embedding vectors representing user-submitted text. The output data from this model consists of category labels representing the type of disaster and floating-point numbers from 0 to 1 representing the urgency level. This system dynamically adjusts the data processing priority based on this urgency level. Through these processes, this system completely eliminates manual data verification by humans. This system reduces data processing latency to milliseconds. This system significantly improves prediction accuracy by guaranteeing the quality of input data in subsequent AI inference. This system prevents system downtime even during massive traffic surges in widespread disasters.

[0060] Step 2: Digital Twin Generation and Updates. Using diverse data obtained from the Data Collection and Integration Department, the state of real-world cities and regions is reproduced with high precision in virtual space, organized by country and region. This involves integrating map data from Geographic Information Systems (GIS), 3D models of buildings, road networks, topographic data of rivers or mountains, location information of shelters, population distribution data, and the location and operational status of critical infrastructure to faithfully reproduce the real-world state. In the event of disasters, the impact range of earthquakes, floods, mudslides, tsunamis, fires, power outages, and large-scale communication failures is simulated, and the disaster scenario, inaccessible sections, isolated areas, the capacity of shelters, the operational status of medical institutions, and the operational status of logistics networks are reflected in the digital twin in real time. Examples include real-time calculations of damage predictions for buildings around the epicenter, road blockages, bridge damage, railway shutdowns, shelter capacity, the occurrence of isolated villages, and the number of patients that medical institutions can receive in the event of an earthquake of intensity 7, and displaying these results in a virtual space; simulations of river flooding or prediction of inundation areas during typhoons or heavy rains; visualization of expected power infrastructure recovery during power outages; and the mapping of communication infrastructure operation status during communication failures. Specifically, this digital twin generation and update unit uses a dynamic graph neural network to mathematically model the physical and logical structure of the city. The system constructs a large-scale directed graph in memory, with intersections or shelters as nodes and roads or communication networks as edges. The processing unit binds real-time changing attribute data to each node and edge. The input data to this neural network consists of feature vectors output by the data collection and integration unit, and spatial tensors representing the 3D coordinates of buildings. This neural network executes a message passing algorithm between adjacent nodes. The output data from this neural network consists of a score array from 0 to 1 representing the probability of each node being affected by disaster, and a weight matrix representing the probability of each edge being impassable. This processing unit uses this weight matrix to dynamically update the cost function of the shortest path exploration algorithm. The updated graph structure is then input into the subsequent simulation engine. This simulation engine uses a hydrodynamic model to calculate the inundation extent of a tsunami or river flood. The input data to this engine is a two-dimensional grid tensor representing the elevation data of the terrain, and a time-series vector representing precipitation. The output data from this engine is a floating-point matrix representing the inundation depth of each grid. This system forwards this matrix to the rendering pipeline, rewriting the texture data in virtual space. Through these processes, this system performs unconventional high-dimensional spatial operations, unlike previous static database retrieval methods. This system calculates the complex interactions of the entire city in real time. This system delivers a significant reduction in the time required to assess the situation during a disaster, supporting rapid decision-making.

[0061] Step 3: Disaster Response Prediction and Simulation using Generative AI. Generative AI is used as a mandatory component, using past disaster data and current real-time data to accurately predict and simulate the shift in evacuation numbers, required supplies, operational status of transportation infrastructure, and material distribution plans for each shelter. For example, generative AI that has learned from past earthquake evacuation patterns or material consumption patterns, the timing of shelter openings, and the material consumption tendencies of each evacuee's attributes can automatically calculate the number of evacuees or the required quantities of food, water, blankets, medicine, hygiene products, cold-weather gear, generators, fuel, and communication equipment for each shelter in the coming hours and days during this disaster. Based on the damage to roads or railways, the feasibility or time required for material transport routes is simulated. When major roads are blocked, the necessity of exploring alternative routes or using helicopters or drones for airlift is automatically assessed. During infectious disease outbreaks, the increased need for dense shelter evacuation or hygiene products is predicted. Specific examples of prompts that generative AI can read include: "Based on the latest shelter entry and exit data, material inventory data, and road traffic conditions, predict the required quantity of supplies and the optimal delivery route for each shelter in the next 24 hours," "Based on earthquake disaster data from the past 10 years and the current population distribution and transportation infrastructure operation status of disaster-stricken areas, simulate the shift in the number of evacuees and changes in material demand in the next 72 hours," and "Given the large number of elderly people and infants in shelters in City A, predict the required quantities of medicines, milk powder, diapers, and nursing supplies, and formulate the optimal delivery plan." Specifically, this disaster response prediction simulation department performs integrated inference using a large-scale language model and a time series prediction model based on the Transformer architecture. The system uses a self-attention mechanism to extract long-term dependencies within the input data. The input data to this AI model consists of a multi-dimensional time series tensor representing the shift in the number of evacuees during past disasters, and an environmental variable vector representing current weather conditions. Furthermore, the system tokenizes prompts described in natural language and inputs them as embedding vectors into the AI ​​model. The AI ​​model integrates these input data and performs nonlinear transformations in the latent space. The output data from this AI model is a numerical matrix representing the hourly material requirements for each shelter over the next 72 hours. Additionally, the AI ​​model outputs a probability distribution tensor representing the confidence level of the prediction. If this confidence level falls below a predetermined threshold, the processing unit initiates a feedback loop requesting additional data collection. The system forwards the output numerical matrix to the subsequent material distribution planning module. This system uses a Long Short-Term Memory (LSTM) network to predict the recovery time of transportation infrastructure. The input data to this network is a one-hot encoded array of categorical variables representing the degree of damage to bridges or roads. The output data from this network is an array of integer values ​​representing the estimated time until recovery of each link.Through these processes, this system completely replaces previous predictive methods that relied on human intuition or experience. This system mathematically unravels the complex relationships between multiple variables. This system effectively prevents surpluses or shortages of resources and optimally allocates limited resources.

[0062] Step 4: Automatic Generation of Supply Distribution Plans. Based on generative AI-predicted supply needs, inventory at each distribution point, road conditions, and the operational status of transportation vehicles such as trucks, helicopters, and drones, the optimal supply distribution plan is automatically generated. For example, if a shelter in City A is short of 5,000 liters of water, 200 blankets, and 100 sets of medicine, the required supplies are gathered from the nearest supply point or the reserve warehouses of other autonomous regions, accessible roads are selected, and trucks are dispatched. If major roads are blocked, alternative routes are explored, and air transport via helicopters or drones is combined as needed. Simultaneously, deliveries to multiple shelters are optimized, automatically calculating delivery routes and sequences, the number of vehicles required, estimated arrival times, loading plans, driver allocation, and refueling plans. If trucks encounter congestion or are prohibited from passing during delivery, the route is recalculated in real time, and relevant personnel are notified. Specific examples of prompts for generative AI to read include "Automatically generate the best delivery route and delivery method based on the latest supply demand forecasts and road conditions for each shelter" and "Develop a simultaneous delivery plan to multiple shelters at the lowest cost, and calculate the required number of vehicles and estimated arrival time." Specifically, this supply delivery plan generation unit uses deep reinforcement learning algorithms to solve combinatorial optimization problems. The system constructs a multi-agent reinforcement learning environment where delivery vehicles or drones are defined as autonomous agents. The input data to this AI model consists of request vectors representing the supply demand of each shelter, state tensors representing the current location and load capacity of each vehicle, and adjacency matrices representing the travel costs of the road network. This AI model executes a policy gradient method based on Markov decision processes. The output data from this AI model is an array of indices of the nodes each vehicle should visit next, and a floating-point matrix representing the unloading amount at each node. Based on this output data, the processing unit generates a sequence of GPS coordinates representing the specific driving routes of each vehicle. The system transmits the generated route information to the onboard terminals of each vehicle via a wireless communication network. When this system receives new traffic closure information while a vehicle is in motion, it immediately updates the reward function and triggers a relearning process. The output data from this relearning process is a dynamically corrected sequence of nodes for the detour route. The system pushes this detour route notification to the onboard terminal in real time. Through these processes, this system derives suboptimal solutions within a practical computational time for NP-hard problems such as the Traveling Salesman Problem or the Delivery Planning Problem. This system minimizes delivery delays while suppressing computational resource consumption. This system delivers the technical effect of ensuring reliable and efficient material transportation even under chaotic traffic conditions during disasters.

[0063] Step 5: Information Sharing and Visualization. Information generated on the server side, including material needs, inventory status, delivery plans, and transportation progress for each shelter, is distributed in real-time to the dashboard or map display on the terminal side. The shortage status of materials in shelters nationwide, the current location and estimated arrival time of delivery trucks, the shift in the number of refugees in each shelter, material consumption patterns, the operational status of transportation infrastructure, the number of patients that medical institutions can accept, and the status of isolated villages are visualized on the map. Relevant personnel can instantly grasp the situation from the terminal and make decisions such as supplementary arrangements or route changes as needed. The dashboard displays the material consumption rate or remaining inventory of each shelter in charts, visualizing future shortage predictions or replenishment opportunities. When on-site personnel input information from the terminal, the server-side generative AI instantly predicts new material needs, automatically updates the delivery plan, and notifies the terminals of relevant personnel. Specifically, this information sharing and visualization department executes an algorithm to optimize the data stream processing to the front-end terminals. This system uses the WebSocket protocol to establish a bidirectional, low-latency communication channel between the server and the client. This processing unit uses a spatial indexing structure to dynamically extract data contained only within the user's display area. Input data to this unit is a differential data stream in JSON format representing the state updates of the entire system. The unit calculates the geometric information required for depiction from this differential data. Output data from this unit consists of an array of vertex coordinates for depicting polygons on the map, and rendering attribute vectors specifying the color or transparency of each object. This system forwards this output data directly to the graphics processing unit of the client terminal. The terminal uses hardware acceleration to perform high-speed re-depicting of the screen. Upon receiving new text input or numerical data from the field terminal, the system immediately activates a natural language processing model. Output data from this model is a score representing the urgency of the input content. If the score exceeds a predetermined threshold, the system sets a flag variable to true to make the corresponding icon on the screen blink. Through these processes, the system minimizes network bandwidth consumption. Even when displaying tens of thousands of dynamic objects simultaneously, the system completely prevents frame rate degradation. This system effectively reduces the cognitive load on users caused by information overload, enabling intuitive situational awareness.

[0064] Step 6: Model Evolution through Federated Learning. On the server side, federated learning continuously learns from disaster response performance data and new event data, evolving the simulation accuracy and predictive model of the generative AI. For example, the AI ​​learns the actual amount of supplies or delivery time required during a disaster, the movement of evacuees, the reception status of medical institutions, and the recovery speed of communication infrastructure, enabling more accurate predictions and plans for future disasters. It learns the characteristics of each region, such as increased demand for heating appliances or cold-weather gear in snowy areas during winter, delivery delays due to traffic congestion in urban areas, the potential for isolated villages in mountainous regions, and increased demand for shelters or hygiene products during infectious disease outbreaks, automatically generating optimal plans that take these factors into account for the next disaster. Specific examples of prompts for the generative AI to read include "Integrate past disaster response performance data and new event data to update the material demand prediction model for the next disaster" and "Generate a disaster response simulation model that reflects the characteristics of each region or past lessons." Specifically, this federated learning evolution department executes a federated averaging algorithm in a distributed computing environment. This system enables local servers in each autonomous region or network point to function as edge nodes. Each edge node uses raw data collected from its own network point to relearn a local neural network model. The input data for this learning process is an error tensor representing the difference between predicted material demand and actual consumption. The edge node updates its weight parameters using stochastic gradient descent. The output data from the edge node is a gradient matrix representing the difference in the updated weight parameters. The system does not send the raw data externally; instead, it encrypts and sends this gradient matrix to the central server. The central server performs a weighted average of the gradient matrices received from most edge nodes to update the global model. The output data from the central server is the weight tensor of the optimized new global model. The system broadcasts this weight tensor to all edge nodes, updating the inference engine of each node through hot-swapping. To perform domain adaptation for each region, the system uses a meta-learning algorithm. The input data to this algorithm is a vector of environmental variables representing climate conditions or population dynamics for each region. The output data from this algorithm is an array of micro-call parameters inherent to each region. Through these processes, the system fully protects the privacy of personal or confidential data. This system significantly reduces the load on communication networks. It also provides a technique to continuously improve the model's generalization performance to unknown disaster patterns and suppress overfitting.

[0065] (specific use case) For example, imagine a large-scale earthquake occurring in a vast municipality encompassing coastal areas, resulting in massive damage spanning multiple municipalities. In this scenario, the data collection and integration department would collect various information in real time, including: evacuation numbers or attribute information from the evacuation management system of each municipality; material inventory from the reserve warehouse management system; information on restricted passage sections or bridge damage from the road management system; railway or bus operation status from transportation operators; hospital bed availability from medical institutions; truck operation status from logistics operators; aftershock or tsunami warnings from meteorological observation agencies; communication infrastructure operation status from telecommunications operators; reports on the status of evacuations or material shortages entered by on-site personnel via terminals; and image data of the disaster area obtained from drones or surveillance cameras. This information would then be integrated on the server side. Based on this data, the digital twin generation and updating unit recreates geographic information system map data, 3D building models, population distribution, location information of shelters, medical institutions, and logistics networks in virtual space, as well as the operational status of critical infrastructure. It instantly reflects the impact of earthquakes on building collapses, road blockages, isolated villages, the capacity of shelters, the number of patients that medical institutions can receive, and the operational status of logistics networks. Specifically, this system activates a unique hyperparallel data processing pipeline during wide-area disasters. The data collection and integration unit uses a message queue system to buffer data packets arriving asynchronously from millions of endpoints. The processing unit uses a convolutional neural network to analyze aerial images from drones in real time. The input data to this network is a multispectral image tensor with a resolution of 4096×2160. This network performs convolutional operations for edge detection and feature extraction. The output data from this network is an array of bounding box coordinates for collapsed buildings or blocked roads, and a probability score from 0 to 1 representing the severity of the disaster. If the probability score exceeds a predetermined threshold, the system assigns a disaster status to the coordinates. This digital twin generation and update unit receives an integrated state vector containing the disaster marker. This processing unit uses a spatial hashing algorithm to quickly retrieve and update 3D objects within the virtual space. The input data to this processing unit is the GPS coordinate vector assigned to the disaster marker. The output data from this processing unit is a vertex displacement matrix used to deform the polygon mesh of the corresponding object in the virtual space. This system applies this matrix to the rendering engine to generate physically based damage representations in real time. Through these processes, this system can process massive amounts of disaster information spanning a wide area without delay. This system optimizes computer memory usage and completely prevents system crashes. This system leverages the technology to synchronize real-world physical changes to virtual space with millisecond precision.

[0066] The Disaster Response Prediction and Simulation Department utilizes generative AI as a mandatory tool. It learns from past earthquake disaster data, the shift in evacuation numbers at each shelter, material consumption patterns, road and railway damage, the operational history of medical institutions, and the delivery performance of logistics networks. In this disaster, it automatically calculates the evacuation number shift at each shelter over the next few hours to days, as well as the required quantities of food, water, blankets, medicine, hygiene products, cold-weather gear, generators, fuel, and communication equipment. For example, AI predicts a higher proportion of elderly and infants in coastal shelters, leading to a greater demand for medicine, formula, diapers, and nursing supplies. It also automatically identifies the potential for isolated villages in mountainous areas, necessitating airlifts by helicopter or drone. When major roads are blocked, AI explores alternative routes, optimizes simultaneous deliveries to multiple shelters, and automatically calculates delivery routes and sequences, the number of vehicles required, estimated arrival times, loading plans, driver allocation, and fuel replenishment plans. Furthermore, during infectious disease outbreaks, AI predicts increased evacuation at shelters and increased demand for hygiene products, prioritizing the necessary supplies. Specifically, to address complex dependencies in a wide-area space, this disaster response prediction simulation unit employs a graph attention network. The system applies a self-attention mechanism to the graph structure with shelters and logistics points as nodes, dynamically calculating the influence between nodes. The input data to this AI model consists of attribute vectors representing the population dynamics of each shelter and an adjacency matrix tensor representing the probability of disaster on surrounding roads. The AI ​​model extracts hidden layer features from these input data. The output data from this AI model consists of time-series tensors of demand forecasts for each material category and probability scores representing the isolated risk of each node. If the isolated risk score exceeds a predetermined threshold, the processing unit automatically initiates a branching process to activate the air transport arrangement module. This system uses a genetic algorithm to solve the route planning problem for multiple vehicles. The input data to this algorithm consists of the demand forecast tensor and the initial state vectors of each vehicle. The algorithm repeatedly applies crossover and mutation operators to evolve the solution population. The output data from this algorithm consists of a node sequence representing the optimized delivery route and a floating-point array of the estimated arrival times of each vehicle. This system writes the output data into the subsequent vehicle allocation management database. Through these processes, the system avoids getting trapped in local optimization and derives the globally optimal solution for the entire wide area. This system achieves efficient allocation of computing resources. This system provides the technical effect of preventing cascading shortages of supplies and traffic congestion before they occur.

[0067] The Information Sharing and Visualization Department distributes real-time data generated on the server side, including material needs, inventory status, distribution plans, transportation progress, evacuation numbers, the status of isolated villages, the number of patients that medical institutions can receive, and the operational status of transportation infrastructure, to dashboards or map displays on the terminal side. Relevant personnel can instantly grasp the situation from their terminals and make decisions such as additional arrangements or route changes as needed. When on-site personnel input an increase in the number of people in a shelter or a shortage of materials, the server-side generative AI predicts new material needs, automatically updates the distribution plan, and notifies the relevant personnel's terminals. The dashboard displays the rate of material consumption or remaining inventory for each shelter in charts, and future shortage predictions or replenishment opportunities are also visualized. Specifically, this Information Sharing and Visualization Department executes dynamic data binding and differential update algorithms to minimize the rendering load on client terminals. The system maintains a state management tree on the server side, extracting only nodes that have changed. The data input to this processing unit consists of natural language vectors of text sent from on-site terminals and arrays of numerical data. This processing unit inputs this input data into the natural language processing model in real time. The output data from this model consists of a score representing the urgency of the input content and a category label. Based on this score, the system reorders the priority queue of payloads sent to clients. The output data from this processing unit is a serialized JSON-formatted differential data packet. The system uses a publish-subscribe model to multicast this differential data packet to the corresponding terminals. The rendering engine on the terminal side parses the received differential data packet. The output data from this engine is an array of control point coordinates for the Bézier curve used to depict the graph on the screen. When the urgency score exceeds a predetermined threshold, the system highlights the corresponding area on the screen and triggers a forced push notification. Through these processes, the system eliminates unnecessary full-screen re-rendering, significantly reducing terminal battery consumption and CPU usage. Even in disaster environments with limited communication bandwidth, this system effectively ensures the timeliness and consistency of information.

[0068] The Federal Learning and Evolution Department enables AI to learn the actual needs of a disaster, such as the quantity of supplies or delivery time required, the movement of evacuees, the reception status of medical facilities, and the recovery speed of communication infrastructure, allowing for more accurate predictions and planning for future disasters. For example, it learns the characteristics of each region, such as increased demand for heating appliances or cold-weather gear in snowy areas during winter, the likelihood of delivery delays due to traffic congestion in urban areas, the potential for isolated villages in mountainous regions, and the increased demand for shelters or hygiene products during infectious disease outbreaks. This allows for the automatic generation of optimal plans that take these factors into account for future disasters. Specifically, this Federal Learning and Evolution Department executes a domain-adaptive algorithm to correct for data distribution biases in each region. The system runs independent learning processes in the background on local servers in each region. The data input to this learning process is an error tensor representing the difference between the estimated delivery time output by the predictive model and the actual delivery time obtained from GPS logs. This process uses backpropagation to fine-tune the weights of the local model. The output data from this process is a weight update matrix that has learned the region-specific biases. This weight update matrix is ​​then sent to the central server. This central server performs weighted averaging based on regional similarity on the weight update matrices from each region. The input data to this process consists of environmental feature vectors representing population density or elevation variance for each region. The output data from this process comprises the parameter tensor of the base model shared across all regions, and the parameter arrays of the adapter modules specific to each region. The system dynamically combines these two parameters during inference. Through these processes, the system completely prevents overfitting to data specific to a particular region. Even in sparsely populated areas where data collection is difficult, the system can perform zero-shot learning of the learning outcomes from urban areas. This system delivers a continuous and self-regulating improvement in the overall prediction accuracy.

[0069] Specific examples of prompts that can be read by generative AI include: "Based on the latest data on shelter entry and exit, material inventory, road conditions, and medical facility operations, predict the required quantity of supplies and the optimal delivery route for each shelter over the next 24 hours," "Based on earthquake disaster data from the past 10 years and the current population distribution, transportation infrastructure operation, and logistics network operation in disaster-stricken areas, simulate the shift in the number of evacuees and changes in material demand over the next 72 hours," "In coastal shelters with a high number of elderly people and infants, predict the required quantities of medicines, formula, diapers, and nursing supplies, and develop the optimal delivery plan," "In the case of isolated villages, develop the optimal material delivery plan, including air transport via helicopters or drones," and "Considering the increased demand for hygiene products in shelters during infectious disease outbreaks, create a priority list of supplies and a delivery plan," etc. Specifically, this generative AI prompt management department possesses an algorithm for automatically executing dynamic prompt engineering. The system uses user input or changes in system state as triggers to select the optimal prompt template from the database. The data input to this processing unit consists of a multi-dimensional vector representing the current system state and an integer ID representing the user request category. This processing unit performs string replacement processing on the variable placeholders within the selected template, embedding real-time numerical or textual data. The output data from this processing unit is a labeled array of context-optimized hints. This labeled array is then passed to the input layer of a large language model. This model performs autoregressive inference based on the input labeled array. The output data from this model consists of natural language text representing prediction results or plans, and structured JSON-formatted data objects. This processing unit performs schema validation on the output JSON data. Upon detecting missing required fields or inconsistent data types, this processing unit performs a loop that automatically generates hints for error correction and requests a re-inference. Through these processes, this system physically suppresses the illusions inherent in large language models. This system fully guarantees the consistency of the inference results' format. This system leverages the technical effectiveness of subsequent system modules in safely and reliably parsing the AI's output.

[0070] Thus, the system of this invention collects and integrates various data in real time immediately after a disaster, accurately reproducing the real world on a digital twin. It utilizes generative AI as a necessary component, accurately predicting and simulating the shift in evacuation numbers or material needs in each shelter, the operational status of transportation infrastructure, and optimal material distribution plans. It visualizes and shares information, and through federated learning, the model evolves, thus becoming a unique social infrastructure for providing rapid and accurate support to disaster victims, independent of previous manual operations or human judgment. Specifically, this system achieves a fundamental improvement in computer technology itself through the coordinated actions of the aforementioned modules. This system adopts a microservice architecture in a distributed computing environment. The system asynchronously processes the inference of each AI model, performing load balancing of tasks via a message queue system. The data input to this orchestration module includes the network traffic logs of the entire system and resource tensors representing the CPU and memory usage of each container. This module predicts future load peaks based on this input data. The output data from this module includes control commands instructing the horizontal scaling of specific inference containers and resource allocation matrices. This system dynamically configures virtual machines based on these control commands. Through these processes, this system completely eliminates single points of failure, dramatically improving the fault tolerance of the entire system. Even under sudden surges in access during disasters, this system ensures scalability and provides a complete guarantee of real-time processing. This system transcends simple business automation, achieving the ultimate in hardware resource optimization and computational efficiency.

[0071] (Application Example 1) The specific processing flow corresponding to Use Case 1 will be described below. The various parts of the system described below are implemented by the data processing device 12 and the intelligent device 14. Furthermore, the data processing device 12 is referred to as the "server" and the intelligent device 14 is referred to as the "terminal".

[0072] (Implementation methods for carrying out the present invention) As an embodiment of the present invention, a detailed and comprehensive description of the structure and operation of an AI system for optimizing urban operations using a real-time digital twin of the entire city is provided. Specifically, this system employs a hyperparallel distributed processing architecture for handling massive data streams at the city scale. This architecture has a layered structure of an edge computing layer and a cloud computing layer. The system executes lightweight data filtering algorithms on each edge device. The data input to this edge layer is a raw data stream generated in milliseconds from millions of IoT sensors within the city. This edge layer applies noise removal and downsampling processing to this raw data. The output data from this edge layer is a normalized and compressed feature vector. The system only sends this feature vector to the cloud layer. The cloud layer routes the received vector to the respective specialized AI modules. Through these processes, the system drastically reduces the consumption of communication bandwidth to the cloud. This system minimizes data processing latency, achieving the technical effect of real-time monitoring of the entire city's status.

[0073] This system comprises a city data collection and integration department, a city digital twin generation and updating department, a city operation optimization department, an information sharing and visualization department, a federated learning and evolution department, a field input support department, an alarm notification department, and a generative AI prompt management department. Specifically, the various components of this system communicate with each other via a high-speed internal bus and shared memory space. The system features a task scheduler based on a directed acyclic graph (DAG) to manage data dependencies between modules. The data input to this scheduler consists of event signals indicating the completion of each module's processing, and pointers to memory addresses storing the output data. The scheduler executes a topology sorting algorithm to determine the next executable task. The output data from this scheduler includes the start signal for the next module to execute and resource allocation parameters for the thread pool. Based on this output data, the system triggers tensor operations on the GPU cluster in parallel. Through these processes, the system completely eliminates bottlenecks caused by standby time between modules. This system maximizes hardware utilization, resulting in a technical effect that maximizes the throughput of the entire system.

[0074] The Urban Data Collection and Integration Department collects and integrates various data from all sensors and systems within the city in real time. For example, it obtains vehicle counts, average speeds, and traffic volume categorized by vehicle type from traffic volume sensors installed on major roads or intersections; it obtains operational status or delay information for buses, railways, subways, and trams from public transportation operation management systems; it obtains signal switching timing or fault information from traffic light control systems; it obtains parking space availability or utilization rates from parking management systems; it obtains electricity, gas, and tap water consumption, and the operational status of solar power generation or batteries from energy management systems for households, buildings, and factories; it obtains temperature, humidity, precipitation, wind speed, and sunshine duration from meteorological observation systems; and it obtains air quality (PM2.5, NOx, CO2), noise, vibration, and the location of garbage collection trucks from environmental sensors. This data collection system comprehensively gathers all information about urban activities, including: mobile records from residents' smartphones or transportation IC cards; resident service usage records; administrative procedure records; event information; disaster information (earthquakes, typhoons, rainstorms, tsunamis, fires, power outages, communication failures); infrastructure operation status (water supply and drainage, electricity, communications, roads, bridges, tunnels, railways, airports, ports); disaster, congestion, and infrastructure anomaly reports from SNS or reporting applications; the operation status of IoT home appliances; the operation status of medical institutions (number of beds, emergency transport reception status, drug inventory); attendance information or school safety information from educational institutions; and the operation status of logistics outlets or delivery vehicle location information. This data is aggregated on a server through various means, such as API linkage, IoT sensors, data linkage with existing business systems, manual input from on-site personnel, automatic collection through image or voice recognition, and automatic analysis of images and video data from drones or surveillance cameras. After encryption and anonymization, the data is integrated. For example, during events, data from pedestrian flow sensors or social media submissions can be combined to understand congestion conditions, and real-time location data of garbage trucks and fullness sensor information of garbage bins can be integrated to automatically generate optimal collection routes. For example, during hot summer days, electricity consumption data from various buildings or households can be aggregated and used for peak power prediction or sending energy-saving alerts. For example, during heavy rain, real-time data from river level sensors or road flooding sensors, shelter occupancy status, and road flooding information can be collected and used for disaster response. Specifically, the city's data collection and integration department performs processing that embeds multimodal data streams into a common high-dimensional vector space. This processing department runs in parallel a one-dimensional convolutional neural network for time-series data and a visual Transformer for image data. Input data to this processing department includes, for example, 1920×1080 resolution RGB image tensors from intersection cameras and floating-point arrays of temperature, humidity, and PM2.5 concentration from environmental sensors.This processing unit applies a separate feature extraction pipeline to these input data. The output data from this processing unit is a 512-dimensional feature vector extracted from various data sources. The system integrates these feature vectors at a consolidation layer to generate a single integrated state tensor representing the current state of the city. This integrated state tensor is then passed to the subsequent digital twin generation module. For data containing personal information, the system performs irreversible anonymization processing using locality-sensitive hashing at the hardware level. Through these processes, the system mathematically extracts the semantic relationships between heterogeneous data. This system standardizes the input dimensions of subsequent prediction models, resulting in a significant improvement in computational efficiency.

[0075] The city digital twin generation and updating department uses diverse data obtained from the city data collection and integration department to reproduce buildings, roads, transportation, energy infrastructure, environment, resident activities, infrastructure operation status, logistics outlets, medical institutions, educational institutions, event venues, etc., with high precision in virtual space, taking the entire city and its regions as units, and updating the city's status in real time. For example, it integrates map data obtained from geographic information systems, 3D models of buildings, road networks, topographic data of rivers or mountains, location information of shelters or medical institutions, logistics outlets, population distribution data, and the operation status of important infrastructure to faithfully reproduce the urban structure and activities in the real world. In the event of disasters, it simulates the impact range of earthquakes, floods, mudslides, tsunamis, fires, power outages, large-scale communication failures, etc., and reflects the disaster scenario, inaccessible sections, isolated areas, the capacity of shelters, the number of patients that medical institutions can receive, and the operation status of logistics outlets in real time on the digital twin. For example, during an earthquake, the risk of building collapse, road blockages, bridge damage, and railway shutdowns can be recreated in virtual space, visualizing which areas of the city will be isolated. For example, during an event, changes in pedestrian or traffic flow around the venue and public transportation congestion can be recreated in real time, simulating the effectiveness of congestion mitigation measures. For example, during winter snowfall, the operation of snowplows or snow depth data can be reflected, visualizing traffic obstacles or logistics delays. Specifically, the city's digital twin generation and updating department uses a dynamic graph neural network to model the city's physical and logical structure. This network constructs a large-scale graph with intersections or buildings as nodes and roads or communication networks as edges. The data input to this network is the integrated state tensor and multidimensional vectors representing the static attributes of each node and edge. This network executes a message-passing algorithm incorporating a time decay function to calculate changes from past states. The output data from this network is a tensor representing the congestion score of each node, the traffic flow prediction value of each edge, and the disaster probability distribution during a disaster. Based on these output tensors, this processing unit updates the dynamic rendering parameters of 3D objects in virtual space. The system then feeds the updated parameters to the graphics pipeline, performing recalculation of textures or geometry. Through these processes, the system mathematically represents the complex interactions of the city. This system reduces the computational order of the simulation, resulting in a level of real-time performance.

[0076] The Urban Operations Optimization Department utilizes generative AI as a mandatory tool, analyzing data from the city's digital twin to optimize urban operations and automate disaster responses. For example, in the transportation sector, generative AI analyzes traffic volume data, public transportation operation status, traffic light status, parking space availability, resident movement history, and event information to predict future congestion risks. It then simulates, proposes, and executes real-time optimizations for signal control, including increasing public transportation frequency, proposing detour routes, and automatically issuing traffic control notices. For instance, if congestion is predicted at major intersections during morning and evening rush hours, the system automatically adjusts signal switching timings, proposes increasing bus or train services, and notifies residents of optimal travel routes. Similarly, if congestion is predicted in surrounding areas during an event, the system automatically notifies residents of temporary public transportation services or recommends park-and-ride options. In the energy sector, generative AI analyzes data on electricity, gas, and water consumption in households or buildings, the operational status of solar power generation or batteries, and meteorological data to predict increased consumption due to future temperature rises or events. To optimize the balance of electricity supply and demand during peak periods, it automatically sends energy-saving alerts or instructions to utilize decentralized power generation. For example, in areas where electricity supply and demand are tight on hot days, residents or operators are notified of decentralized power generation or energy-saving alerts.

[0077] For example, during power outages, generative AI can automatically prioritize power supply to critical infrastructure or automatically schedule generators. In the environmental field, generative AI analyzes data from air quality sensors, noise sensors, temperature and humidity sensors, garbage collection status, and CO2 absorption by green spaces to predict increases in PM2.5 or noise, and automatically proposes and implements measures such as traffic control, greening initiatives, and optimization of garbage collection routes. For example, in industrial areas where there are signs of deteriorating air quality, it can propose traffic control or adjustments to factory operations. For example, AI can optimize garbage truck routes to minimize fuel consumption and CO2 emissions. In disaster response, generative AI analyzes information on disasters such as earthquakes, rainstorms, and typhoons, the capacity of shelters, the damage to infrastructure, residents' location information, and data from social media reports to automatically formulate and execute disaster expansion predictions or evacuation guidance, infrastructure restoration priorities, and material distribution plans. For example, immediately after an earthquake, based on building damage data, shelter capacity, and road blockage information, it can formulate plans for isolating high-risk areas and prioritizing evacuation guidance and material distribution. For example, during heavy rain, areas with a high risk of river flooding can be identified, and residents can be automatically notified of evacuation instructions.

[0078] In resident services, based on residents' attributes or activity history, health data, event participation history, administrative procedure history, etc., generative AI automatically proposes and notifies individuals of the best administrative services or health advice, provides event guidance, and individual evacuation instructions during disasters. For example, it automatically sends individual evacuation instructions or health management advice to the elderly or those in need of care. It also automatically guides families with young children to the availability of childcare facilities or vaccination schedules. Specifically, the city operations optimization department implements a hybrid inference engine that combines deep reinforcement learning algorithms and large-scale language models. This engine constructs a multi-agent reinforcement learning environment where various city departments act as self-regulating agents. The data input to this engine is the state vectors of each area extracted from the digital twin, and tensors representing past policy implementation history. This engine uses the Actor-Critic method to simultaneously optimize the action value function and policy function of each agent. The output data from this engine consists of numerical data representing the optimal green light extension seconds for traffic lights, a matrix representing the output adjustment of distributed power supplies, and an array of natural language tags constituting the alert text for residents. This processing unit directly sends the output control parameters to the API endpoints of various infrastructure control systems. It also transmits alarm texts to the push notification distribution server. Through these processes, the system autonomously resolves the trade-offs between multiple opposing optimization objectives, such as congestion mitigation and energy conservation. This system achieves multivariate simultaneous optimization that is impossible for humans, maximizing the operational efficiency of the entire city.

[0079] The Information Sharing and Visualization Department uses dashboards or notifications to visualize and share the city's status, forecasts, optimization proposals, and disaster response plans in real time with city operators, infrastructure operators, residents, and other relevant stakeholders. For example, it visualizes the city's traffic conditions, energy supply and demand, environmental indicators, disaster damage, shelter capacity, and infrastructure recovery progress on maps or charts, enabling stakeholders to stay informed. For instance, it notifies residents via smartphone applications of optimal travel routes, energy-saving alerts, and disaster evacuation instructions; and it displays city operators' overall risks and optimization proposals on dashboards. It also provides real-time notifications to logistics operators of optimized delivery routes and traffic congestion information. Specifically, this Information Sharing and Visualization Department performs adaptive data filtering based on user permissions and current location. Using a spatial indexing structure, it quickly extracts data only from areas of user interest. The input data consists of GPS coordinate vectors representing the user's current location and a tensor of alert information for the entire system. The department applies a geometric cross-validation algorithm to this input data. The data output from this processing unit consists of structured JSON data containing only alerts relevant to the user, and an array of polygon vertices for map rendering. This system pushes and publishes the extracted data to the client terminal. Based on the received data, the terminal performs hardware-accelerated rendering processing. Through these processes, the system significantly reduces the memory consumption of the client terminal. This system also optimizes network bandwidth and prevents frame rate degradation during rendering.

[0080] The Federated Learning Evolution Unit continuously learns from performance data of city operations and new event data through federated learning, enabling the generative AI's predictive and optimization models to evolve. For example, the AI ​​learns from past disaster response performance, the effectiveness of traffic congestion mitigation measures, the performance of energy supply and demand adjustments, the effectiveness of environmental policies, and the utilization of resident services, allowing for more accurate predictions and optimizations in future city operations or disaster responses. For instance, it learns the characteristics of different regions, such as increased demand for heating appliances or cold-weather gear in snowy areas during winter, the likelihood of delivery delays due to traffic congestion in urban areas, the potential for isolated villages in mountainous regions, and the increased demand for shelters or hygiene products during infectious disease outbreaks. This allows for the automatic generation of optimal plans that take these characteristics into account during the next disaster. Specifically, this Federated Learning Evolution Unit performs gradient calculations for local models on edge servers in each city region and applies them to a central server using a federated averaging algorithm. This algorithm learns multiple patterns across the entire city while protecting data privacy. The data input to this edge server is a time-series tensor representing the actual traffic volume or energy consumption observed in each region. This edge server uses this data to perform error backpropagation for the local models. The data output from this edge server is the weight update difference tensor of the local model. This system only sends this difference tensor to the central server. The data output from this central server is the new weight parameter matrix of the global model after aggregation and averaging. This system redistributes the updated global model to each edge server and updates the inference engine through hot-swapping. Through these processes, this system eliminates the need for network transmission of raw data, significantly reducing communication load. This system leverages the technical advantage of building models with high generalization performance that eliminate region-specific biases.

[0081] The On-Site Input Support Department provides interfaces for on-site personnel to input and upload information such as the status of shelters, inventory of supplies, disaster conditions, photos, or voice data from terminals, and sends the input data to the server. For example, input can be made from tablets or smartphones, including the number of people or attributes in shelters, the types and quantities of required supplies, photos or videos of the scene, voice memos, and detailed descriptions of the disaster situation. For example, on-site personnel can confirm receipt of supplies or report delivery completion from the terminal. For example, on-site personnel can utilize voice input or image recognition to quickly register information. Specifically, this On-Site Input Support Department executes real-time data preprocessing and compression algorithms using edge AI. To maximize the use of the limited computing resources of mobile terminals, this processing department runs a lightweight convolutional neural network on the terminal. The data input to this network is high-resolution on-site image tensors obtained from the terminal's camera and voice waveform data from the microphone. This network performs object detection within the images and text-to-speech processing locally. The data output from this network is an array of bounding box coordinates of the disaster-stricken areas extracted from the images, and a natural language vector of the text converted from the speech. This system only sends these lightweight feature data to the server side. The server then maps the received data to a digital twin in real time. Through these processes, the system ensures reliable data transmission even in unstable communication environments during disasters. This system achieves the technical effect of distributing the processing load from the server side to the edge, thereby improving the overall system response speed.

[0082] The Alarm Notification Unit sends real-time notifications to relevant personnel's terminals regarding important information or anomaly detections from the generative AI or the entire system, changes in delivery plans, a sharp increase in the number of people in shelters, predicted shortages of supplies, road closures, restrictions on medical facility access, and communication failures. For example, in the event of a sharp increase in shelter residents, an alarm is sent requesting additional supplies. For example, if a delivery truck encounters a road closure, an alarm is sent instructing on a route change. For example, if there are restrictions on medical facility access, instructions are sent to transport the goods to a nearby facility. For example, in the event of a communication failure, an alarm is sent prompting the use of alternative communication methods. Specifically, this Alarm Notification Unit continuously monitors the anomaly score or predicted probability distribution output from the generative AI. If the anomaly score exceeds a pre-set dynamic threshold, the Alarm Notification Unit immediately generates a hardware interrupt. Using this interruption as a trigger, the Alarm Notification Unit performs low-latency data transmission using a push notification protocol to the relevant personnel's portable terminals. When generating notification content, the Alarm Notification Unit dynamically generates natural language text based on the situation using a large language model. This large-scale language model takes as input a time-series array of sensor data that serves as an anomaly detection trigger, categorical variables representing anomaly types, and geographic coordinate vectors of the target region. As output, this large-scale language model generates structured data containing a numerical score representing urgency, a list of recommended response actions, and notification text for relevant personnel. Based on the urgency score contained in the structured data, the alarm notification unit performs QoS control, prioritizing bandwidth allocation for the communication network. Through this process, the system achieves the technical effect of reliably delivering critical alarm information even in communication congestion environments such as during disasters. The alarm notification unit monitors the arrival confirmation response of sent alarms. If no confirmation response is received within a certain time, the alarm notification unit automatically performs retransmission processing using an alternative communication path or a different communication protocol. This configuration is not merely a simple information display, but rather an optimization of communication data transmission closely linked to the communication layer. This technology significantly improves the reliability and immediacy of information delivery in use cases such as emergency response systems during wide-area disasters or anomaly monitoring systems in large-scale factories.

[0083] The Generative AI Prompt Management Department manages the prompts needed when utilizing generative AI throughout the system and provides the optimal prompts to the generative AI based on the situation. For example, it manages prompts for forecasting material demand or developing distribution plans for each shelter, prompts combining past disaster data and current conditions for simulation instructions, prompts for automated responses to inquiries from on-site personnel, and prompts for automatically generated instructions from disaster response manuals, and provides them to the generative AI as needed. Examples include: "Based on the latest shelter entry / exit data, material inventory data, and road traffic data, predict the required quantity of materials and the optimal delivery route for each shelter over the next 24 hours," "Given the large number of elderly and infants in shelters in City A, predict the required quantities of medicines, formula, diapers, and nursing supplies, and develop the optimal distribution plan," "Based on earthquake disaster data from the past 10 years and the current population distribution and transportation infrastructure operation status of the affected area, simulate the shift in the number of evacuees and changes in material demand over the next 72 hours," and "Analyze disaster report photos from the scene, extract collapsed buildings or flooded areas, and summarize the disaster situation," etc. Specifically, this generative AI prompt management unit performs dynamic prompt construction processing based on the situational context. This generative AI prompt management unit maintains multiple prompt templates as vector representations in a database. This generative AI prompt management unit generates a multi-dimensional state vector representing the current system state. This multi-dimensional state vector includes elements numerically representing the type of disaster, elapsed time, disaster scale, and types of available data sources. This generative AI prompt management unit calculates the cosine similarity between the multi-dimensional state vector and the vector representation of the prompt template. This generative AI prompt management unit selects the template with the highest similarity. This generative AI prompt management unit automatically embeds real-time sensor data or database retrieval results into placeholders for the selected template. During the embedding process, to ensure that the number of tags in the input data does not exceed the context window limit of the large language model, this generative AI prompt management unit applies an importance-based text summarization algorithm. This generative AI prompt management unit inputs the constructed prompts into the inference engine of the large language model. Based on the input prompts, this inference engine outputs a text stream containing prediction results or plans. This generative AI prompt management unit performs syntactic parsing on the output text stream and converts it into structured data that can be processed by subsequent system modules. Through this processing, unlike existing methods that rely on manual human input prompts, this system achieves fully automated prompt optimization based on mechanical state judgments. This approach effectively improves the inference accuracy of large language models while preventing the waste of computational resources caused by invalid prompts. This technology can be applied to autonomous control systems in smart cities or dynamic optimization systems for complex supply chains.

[0084] In this way, the embodiments of the present invention collect and integrate various data within the city in real time, generate high-precision digital twins for the entire city and each region, use generative AI as a necessary component to optimize city operations or automatically respond to disasters, visualize and share information, and evolve the model through federated learning, thereby becoming a unique social infrastructure for achieving continuous optimization and resilience improvement of the entire city.

[0085] (System Composition) The system involved in this implementation includes a city data collection and integration department, a city digital twin generation and updating department, a city operation optimization department, an information sharing and visualization department, a federated learning and evolution department, a field input support department, an alarm notification department, and a generative AI prompt management department.

[0086] The Urban Data Collection and Integration Department has the capability to collect various data from all sensors and systems within the city in real time and integrate them on a server. For example, it can obtain data such as the number of vehicles, average speed, and traffic volume categorized by vehicle type from traffic volume sensors installed on major roads or intersections; the operational status or delay information of buses, railways, subways, and trams from public transportation operation management systems; signal switching timing or fault information from traffic light control systems; parking space availability or utilization rates from parking management systems; electricity, gas, and tap water consumption, or the operational status of solar power generation and batteries from energy management systems for households, buildings, and factories; temperature, humidity, precipitation, wind speed, and sunshine duration from meteorological observation systems; air quality (PM2.5, NOx, CO2), noise, and vibration from environmental sensors; the location or collection status of garbage trucks; mobile data from residents' smartphones or transportation IC cards; and resident service usage history or administrative procedure history, event information, and disaster information. The city's data collection and integration department collects comprehensive data on urban activities, including: earthquake, typhoon, rainstorm, tsunami, fire, power outage, and communication failure information; infrastructure operation status (water supply and drainage, electricity, communications, roads, bridges, tunnels, railways, airports, and ports); disaster, congestion, and infrastructure anomaly reports from SNS or notification applications; IoT appliance operation status; medical institution operation status (number of beds, emergency transport reception status, and drug inventory); attendance information or school safety information from educational institutions; and logistics network operation status or delivery vehicle location information. This data is aggregated on a server through various means, such as API linkage, IoT sensors, data linkage with existing business systems, manual input from on-site personnel, automatic collection via image or voice recognition, and automatic parsing of images and video data from drones or surveillance cameras. After encryption and anonymization, the data is integrated. Specifically, the department executes a preprocessing pipeline that maps heterogeneous data streams to a unified multidimensional tensor space. A distributed message queue system is used to temporarily buffer data packets asynchronously sent from various sensors. The city's data collection and integration department applies timestamp synchronization and spatial coordinate normalization to the buffered data. For data streams containing missing values, the department uses Kalman filters or Long Short-Term Memory (LSTM) interpolation algorithms to generate continuous time-series data. For image or video data, the department applies a convolutional neural network object detection model. This object detection model takes high-resolution image tensors as input and outputs structured data including bounding box coordinates, category labels, and confidence scores for detected objects.The city's data collection and integration department applies a Transformer-based natural language processing model to text data described in natural language, extracting feature vectors representing sentiment polarity or event categories. This multi-format data is then stored in a graph database with a common spatiotemporal index. Nodes in this graph database represent physical entities, and edges represent interactions or physical distances between entities. Through this processing, the system effectively transforms massive amounts of unstructured urban data into a structured format that can be efficiently computed by subsequent artificial intelligence models. This approach is not merely data accumulation, but rather the semantic integration of data in a high-dimensional space. This technology significantly reduces data processing latency in use cases such as integrated monitoring platforms for smart cities or predictive maintenance systems for wide-area infrastructure.

[0087] The City Digital Twin Generation and Update Department utilizes diverse data obtained from the City Data Collection and Integration Department to accurately reproduce buildings, roads, transportation, energy infrastructure, environment, resident activities, infrastructure operation status, logistics networks, medical institutions, educational institutions, event venues, etc., in virtual space, taking the entire city and its regions as units, and updates the city's status in real time. For example, it integrates map data obtained from geographic information systems, 3D models of buildings, road networks, topographic data of rivers or mountains, location information of shelters or medical institutions, logistics networks, population distribution data, and the operational status of critical infrastructure to faithfully reproduce the city's structure and activities in the real world. During disasters, it simulates the impact range of earthquakes, floods, mudslides, tsunamis, fires, power outages, large-scale communication failures, etc., instantly reflecting disaster scenarios, inaccessible sections, isolated areas, the capacity of shelters, the number of patients that medical institutions can receive, and the operational status of logistics networks onto the digital twin. For example, during an earthquake, the virtual space can recreate the risk of building collapse, road blockages, bridge damage, and railway shutdowns, visualizing which areas of the city will be isolated. For example, during an event, the virtual space can recreate real-time changes in pedestrian or traffic flow around the venue, public transportation congestion, and simulate the effectiveness of congestion mitigation measures. For example, during winter snowfall, the virtual space can reflect the operation of snowplows or snow depth data, visualizing traffic obstacles or logistical delays. For example, during a power outage, the virtual space can reflect the expected restoration of power infrastructure or the priority supply to critical facilities. Specifically, the city's digital twin generation and update department employs a hybrid simulation architecture that tightly integrates the computation of the physics engine with the inference of the deep learning model. The city's digital twin generation and update department converts the city's static structural data into a voxel mesh representation and assigns physical attributes to each voxel in three-dimensional space. The city's digital twin generation and update department maps dynamic sensor data onto the voxel mesh in real time, constructing a four-dimensional spatiotemporal tensor. To predict the spread of disasters during an event, the city's digital twin generation and update department executes a graph convolutional network. This graph convolutional network takes as input an adjacency matrix representing the infrastructure network topology and a feature matrix representing the current operating state of each node. As output, it outputs a probability distribution of each node's potential for malfunction at a specific future time. Based on this probability distribution, the city's digital twin generation and update unit performs parallel simulations of multiple scenarios using the Monte Carlo method. The unit then summarizes the simulation results to determine the expected or worst-case scenario of the disaster. Based on the determined scenario, the unit dynamically changes the rendering parameters of objects in the virtual space. Through this process, the system achieves the technical effect of simulating complex interactions in the real world at high speed and with high precision on a computer. This configuration mathematically elucidates the mechanism of cascading failures that are difficult to predict through human intuition or purely rule-based methods.This technology provides quantitative predictive data as a basis for decision-making in use cases such as national-level disaster prevention simulators or dynamic routing systems for large-scale transportation networks.

[0088] The Urban Operations Optimization Department utilizes generative AI as a mandatory tool, analyzing data from the city's digital twin to optimize urban operations and enable automated disaster response. For example, in the transportation sector, generative AI analyzes traffic volume data, public transportation operation status, traffic light status, parking space availability, resident movement history, and event information to predict future congestion risks. It then simulates, proposes, and executes real-time optimizations for signal control, including increasing public transportation frequency, proposing detour routes, and automatically issuing traffic control notices. For instance, if congestion is predicted at major intersections during rush hour, it automatically adjusts signal switching timings, proposes increasing bus or train services, and notifies residents of optimal travel routes. Similarly, if congestion is predicted in surrounding areas during an event, it automatically provides recommendations for temporary public transportation services or park-and-ride options. In the energy sector, generative AI analyzes data on electricity, gas, and water consumption in households or buildings, the operational status of solar power generation or batteries, and meteorological data to predict increased consumption due to rising temperatures or upcoming events. To optimize peak-hour electricity supply and demand balance, it automatically sends energy-saving alerts or instructions to utilize decentralized power generation. For example, in areas experiencing power shortages during hot weather, it notifies residents or operators of decentralized power generation or energy-saving alerts. It also automatically prioritizes power supply to critical infrastructure or automatically schedules generators during power outages. In the environmental sector, generative AI analyzes data from air quality sensors, noise sensors, temperature / humidity sensors, garbage collection status, and CO2 absorption in green spaces to predict increases in PM2.5 or noise levels. It automatically proposes and implements measures such as traffic control, greening initiatives, and optimization of garbage collection routes. For example, in industrial areas showing signs of deteriorating air quality, it proposes traffic control or adjustments to factory operations. It also optimizes garbage truck routes to minimize fuel consumption and CO2 emissions. In disaster response, generative AI analyzes information such as earthquakes, rainstorms, and typhoons, the capacity of shelters, the damage to infrastructure, residents' location information, and data from social media platforms to automatically formulate and execute disaster expansion predictions, evacuation guidance, infrastructure restoration priorities, and material distribution plans. For example, immediately after an earthquake, based on building damage data, shelter capacity, and road blockage information, it can formulate plans for isolating high-risk areas and prioritizing evacuation guidance and material distribution. Similarly, during rainstorms, it can identify areas with high risk of river flooding and automatically notify residents of evacuation instructions. In resident services, generative AI, based on residents' attributes, activity history, health data, event participation history, and administrative procedures history, automatically proposes and notifies personalized best administrative services or health advice, event guidance, and individual evacuation instructions during disasters. For example, it can automatically send individual evacuation instructions or health management suggestions to the elderly or those requiring care.For example, it can automatically guide families with young children to information on childcare facility availability or vaccination schedules. Specific examples of prompts that generative AI can read include: "Based on current traffic volume data, public transportation operation status, and event information, predict the risk of congestion at each major intersection in the next 30 minutes and propose optimal traffic mitigation measures such as signal control or increasing bus frequency," "Based on current energy consumption data, weather forecasts, and event information, identify areas with tight supply-demand balance in the next 2 hours and formulate the optimal plan for sending out decentralized power generation or energy-saving alerts," and "Based on immediate post-earthquake building damage data, evacuation shelter capacity, and road blockage information, formulate evacuation guidance and material distribution plans for areas with high isolation risk." Specifically, the City Operations Optimization Department functions as an autonomous agent system integrating deep reinforcement learning algorithms and large-scale language models. The City Operations Optimization Department obtains multi-dimensional state vectors extracted from the city's digital twin as environmental observations. The City Operations Optimization Department inputs these observations into the policy network of the deep reinforcement learning model. This policy network takes as input a continuous vector containing vehicle congestion lengths at each intersection, power grid frequency deviations, and shelter capacity. As output, it generates an action vector containing the green light extension time in seconds, battery discharge command values, and the destination node ID of the delivery vehicle. The city operations optimization department inputs this action vector into a large language model, which performs a natural language-based inference process to verify the appropriateness of the action. The large language model retrieves a database of similar past cases and assesses the potential secondary impacts of the proposed action. The city operations optimization department sends the action vector as an actual control command to the control devices of each infrastructure only if the evaluation score from the large language model exceeds a predetermined threshold. The city operations optimization department uses a reward function to evaluate environmental changes after control execution and updates the weight parameters of the policy network using a proximal policy optimization algorithm. Through this process, the system achieves the technical effect of globally optimizing a complex and intertwined group of urban subsystems. This structure breaks through the limitations of individual optimization based on the rules of thumb of human operators, enabling continuous optimal solution exploration in high-dimensional space via computers. This technology can dramatically improve the overall energy efficiency and safety of systems in use cases such as next-generation smart grid control systems or swarm control systems for fully autonomous vehicles.

[0089] The Information Sharing and Visualization Department has the capability to visualize and share the city's status, forecasts, optimization proposals, and disaster response plans in real time with city operators, infrastructure operators, residents, and other relevant stakeholders through dashboards or notifications. For example, it can visualize the entire city's traffic conditions, energy supply and demand, environmental indicators, disaster damage or shelter capacity, and infrastructure recovery progress on maps or through charts, enabling stakeholders to stay informed. For instance, it can notify residents of optimal travel routes or energy-saving alerts and disaster evacuation instructions via smartphone applications, and provide city operators with a dashboard overview of city-wide risks or optimization proposals. It can also provide real-time notifications to logistics operators of delivery route optimization results or traffic congestion information. Furthermore, it can visualize emergency transport reception status or drug inventory status for medical institutions, supporting regional medical collaboration. Specifically, this Information Sharing and Visualization Department possesses a rendering engine that dynamically reduces the dimensionality of massive, multi-dimensional data, adapting to the limitations of human cognition to generate optimal visual representations. The department calculates the priority of information to be displayed based on the user's role or the current context. To emphasize high-priority information, this information sharing visualization department applies heatmap generation algorithms or uses visualization techniques based on fluid simulations of particle systems. Through hardware acceleration, it rapidly renders node and edge information retrieved from a backend graph database on a browser. Using a large language model, it automatically generates natural language summary text explaining the trends or outliers of complex data charts. This large language model takes as input statistical features of time-series data, alert indicators from anomaly detection models, and profile information representing the user's level of expertise. As output, it generates explanatory text adjusted for technical difficulty or a list of actions to be taken. The department integrates the generated text and visual elements to construct an interactive dashboard interface. Through this process, the system effectively reduces cognitive load caused by information overload, supporting rapid and accurate decision-making. This structure is not merely a data listing, but a fusion of computer-generated semantic interpretation and visual optimization of data. This technology maximizes information delivery efficiency in use cases such as large-screen display systems for disaster response headquarters or personalized disaster prevention portals for citizens.

[0090] The Federated Learning Evolution Department has the capability to continuously learn from performance data of city operations and new event data through federated learning, enabling the generative AI's predictive and optimization models to evolve. For example, the AI ​​learns from past disaster response performance, the effectiveness of traffic congestion mitigation measures, the performance of energy supply and demand adjustments, the effectiveness of environmental policies, and the utilization of resident services, allowing for more accurate predictions and optimizations in future city operations or disaster responses. For instance, it learns the characteristics of each region, such as increased demand for heating appliances or cold-weather gear in snowy winter areas, the likelihood of delivery delays due to traffic congestion in urban areas, the potential for isolated villages in mountainous regions, and increased demand for shelters or hygiene products during infectious disease outbreaks, automatically generating optimal plans that take these factors into account for the next disaster. For example, it learns from the effectiveness of congestion mitigation measures during past events, automatically proposing optimal crowd control strategies for the next event. Specifically, this Federated Learning Evolution Department implements a distributed machine learning framework that balances privacy protection and improved model generalization performance. The department distributes the initial weight parameters of the global model to edge nodes configured on local servers of each autonomous region or infrastructure operator. Each edge node uses highly confidential performance data stored in its local memory to perform gradient calculations for its local model. This federated learning evolution unit collects gradient vectors calculated from each edge node only via encrypted communication channels. The federated learning evolution unit applies a secure aggregation algorithm to the collected gradient vectors, adding noise to mathematically make individual data recovery mathematically impossible. The federated learning evolution unit uses the aggregated gradient vectors to update the weights of the global model on the central server using a federated averaging method. The federated learning evolution unit uses a validation dataset to evaluate the performance of the updated global model, and if improved accuracy is confirmed, it is deployed as a new model throughout the system. Through this process, the system leverages the technical advantage of integrating knowledge from a wide range of diverse data sources rather than aggregating raw data in one place. This architecture overcomes data governance constraints while enabling the construction of robust AI models that eliminate region-specific biases. This technology brings a breakthrough in overcoming data sharing barriers in use cases such as the collaborative development of supply chain optimization models among multiple enterprises or the construction of disease prediction models in healthcare network institutions.

[0091] The Field Input Support Unit provides an interface for field personnel to input and upload information such as shelter status, material inventory, disaster situation, photos, or voice data from terminals, and sends the input data to the server. For example, users can input the number of people or attributes in the shelter, the type and quantity of required materials, photos or videos of the scene, voice memos, and detailed descriptions of the disaster situation from tablets or smartphones. For example, users can confirm the receipt of materials or report delivery completion from the terminal. For example, field personnel can quickly register information using voice input or image recognition. For example, users can confirm the receipt of alarm notifications or instructions on the terminal. Specifically, this Field Input Support Unit is installed using edge computing technology that ensures reliable data collection even in harsh communication environments or when personnel are extremely fatigued. This Field Input Support Unit keeps a lightweight inference model residing on the local memory of the portable terminal. The Field Input Support Unit immediately performs semantic segmentation using the lightweight inference model on the image data of the scene taken by the personnel. This lightweight inference model is received as input as image tensors adjusted to low resolution. This lightweight inference model, as output, generates a probability map representing whether each pixel in an image belongs to a specific category. The Field Input Support Unit (NPS) processes voice input data into text using a speech recognition engine running on the edge, and further performs named entity extraction using a natural language processing model. The NPS combines the extracted structured data and image parsing results to generate a very small payload. In cases of unstable network connectivity, the NPS queues the payload in local non-volatile memory and asynchronously sends it to the server in batches the moment communication is restored. Through this processing, the system effectively prevents bandwidth strain caused by sending large amounts of raw data, distributing the processing load from the server side to the edge. This structure is not merely a provision of input forms, but rather achieves high-level preprocessing and autonomous communication control on the terminal side. This technology significantly improves the timeliness and accuracy of data collection in use cases such as information collection systems for first responders at disaster sites or inspection systems in offline environments like underground facilities.

[0092] The alarm notification unit has the function of providing real-time notifications to relevant personnel's terminals regarding important information or anomaly detections from generative AI or the system as a whole, changes in delivery plans, a sharp increase in the number of people in shelters, predicted shortages of supplies, road closures, restrictions on medical facility access, and communication failures. For example, it sends an alarm for additional supplies arrangements when a shelter experiences a sharp increase in refugees. For example, it sends an alarm for route change instructions when a delivery truck encounters a road closure. For example, it notifies personnel of delivery instructions to nearby medical facilities when medical facility access restrictions occur. For example, it sends an alarm prompting the use of alternative communication methods in the event of a communication failure. Specifically, this alarm notification unit constantly monitors the anomaly scores or predicted probability distributions output from the generative AI. When the anomaly score exceeds a pre-set dynamic threshold, the alarm notification unit immediately triggers a hardware interruption. Taking advantage of this interruption, the alarm notification unit performs low-latency data transmission using a push notification protocol to the portable terminals of relevant personnel. When generating notification content, the alarm notification unit dynamically generates natural language text based on the situation using a large language model. This large-scale language model takes as input a time-series array of sensor data that serves as anomaly detection triggers, categorical variables representing anomaly types, and geographic coordinate vectors of the target area. As output, this large-scale language model generates structured data containing a numerical score representing urgency, a list of recommended response actions, and notification text for relevant personnel. Based on the urgency score contained in the structured data, the alarm notification unit performs QoS control, prioritizing the allocation of communication network bandwidth. Through this process, the system achieves the technical effect of reliably delivering extremely important alarm information even in communication congestion environments such as during disasters. The alarm notification unit monitors the arrival confirmation response of sent alarms. If no confirmation response is received within a certain time, the alarm notification unit automatically performs retransmission processing using alternative communication paths or different communication protocols. This structure is not merely a simple information display, but rather an optimization of data transmission closely linked to the communication layer. This technology significantly improves the reliability and immediacy of information delivery in use cases such as emergency response systems during wide-area disasters or anomaly monitoring systems in large-scale factories.

[0093] The Generative AI Prompt Management Department manages the prompts required when utilizing generative AI within the overall system and provides the optimal prompts to the generative AI based on the situation. For example, it manages prompts for forecasting material demand or developing distribution plans for each shelter, prompts combining past disaster data and current conditions with simulated instructions, prompts for automated responses to inquiries from on-site personnel, and prompts for automatically generated instructions for disaster response guidelines, providing them to the generative AI as needed. Specific prompts include: "Based on current traffic volume data, public transportation operation status, and event information, predict the risk of congestion at each major intersection in the next 30 minutes and propose the best traffic mitigation measures, such as signal control or increasing bus frequency"; "Based on current energy consumption data, weather forecasts, and event information, identify areas with tight supply-demand balance in the next 2 hours and develop the best plan for sending out decentralized power generation or energy-saving alarms"; "Based on immediate post-earthquake building damage data, shelter occupancy status, and road blockage information, develop evacuation guidance and material distribution plans for areas with high isolation risk." Specifically, this generative AI prompt management unit performs dynamic prompt construction processing based on the situational context. This generative AI prompt management unit maintains multiple prompt templates as vector representations in the database. This generative AI prompt management unit generates a multi-dimensional state vector representing the current system state. This multi-dimensional state vector includes elements numerically representing the type of disaster, elapsed time, disaster scale, and types of available data sources. This generative AI prompt management unit calculates the cosine similarity between the multi-dimensional state vector and the vector representation of the prompt template. This generative AI prompt management unit selects the template with the highest similarity. This generative AI prompt management unit automatically embeds real-time sensor data or database retrieval results into placeholders for the selected template. During the embedding process, to ensure that the number of tags in the input data does not exceed the context window limit of the large language model, this generative AI prompt management unit applies an importance-based text summarization algorithm. This generative AI prompt management unit inputs the constructed prompts into the inference engine of the large language model. Based on the input prompts, this inference engine outputs a text stream containing prediction results or plans. This generative AI prompt management unit performs syntactic parsing on the output text stream, converting it into structured data that can be processed by subsequent system modules. Through this processing, unlike existing methods that rely on manual human input prompts, this system achieves fully automated prompt optimization based on mechanical state judgments. This structure effectively improves the inference accuracy of large language models while preventing the waste of computational resources caused by invalid prompts. This technology can be applied to autonomous control systems in smart cities or dynamic optimization systems for complex supply chains.

[0094] In this way, the system involved in this implementation collects and integrates various data within the city in real time, generates high-precision digital twins for the entire city and each region, uses generative AI as a necessary component to optimize city operations and automatically respond to disasters, visualizes and shares information, and evolves the model through federated learning, thus becoming a unique social infrastructure for achieving continuous optimization and resilience improvement of the entire city.

[0095] (Implementation steps) Step 1: Real-time collection and integration of urban data Data is obtained from traffic volume sensors installed on major urban roads or intersections, including vehicle count, average speed, and traffic volume categorized by vehicle type; from public transportation management systems, including bus, railway, subway, and tram operation status or delay information; from traffic light control systems, including signal switching timing or fault information; from parking management systems, including parking space availability or utilization rate; from energy management systems, including electricity, gas, and water consumption in households, buildings, and factories, or the operation status of solar power generation and batteries; from meteorological observation systems, including temperature, humidity, precipitation, wind speed, and sunshine duration; from environmental sensors, including air quality (PM2.5, NOx, CO2), noise, and vibration; and from garbage collection trucks, including location and collection status. The system obtains mobile data from residents' smartphones or transportation IC cards, including resident service usage records or administrative procedure records, event information, disaster information, infrastructure operation status, disaster, congestion, and infrastructure anomaly reports from SNS or notification applications, IoT appliance operation status, medical institution operation status, educational institution attendance or school safety information, logistics network operation status, and delivery vehicle location information. Through API integration, IoT sensors, existing business system integration, on-site personnel terminal input, image or voice recognition, and image analysis from drones or surveillance cameras, all data related to urban activities is aggregated and integrated into a server. This data undergoes encryption and anonymization before consolidation. Specifically, in this step, the system uses an in-memory data grid to rapidly ingest massive data streams asynchronously sent from a distributed array of edge devices. The system parses the header information of each received data packet, extracting the data type, source ID, and timestamp. Based on the extracted metadata, the system dynamically routes the data to the appropriate processing pipeline. This system applies outlier detection algorithms to time-series numerical data, filtering out spike noise caused by sensor malfunctions or communication errors in real time. For image or video data, the system executes convolutional neural networks on a parallel computing cluster to extract objects such as vehicles, pedestrians, and obstacles as bounding boxes. The system then transforms the coordinates of the extracted objects to a common geospatial coordinate system. This preprocessed heterogeneous mixed data is registered as nodes and edges in a graph database, and semantic relationships between data are used as links. Through this processing, the system achieves the technical effect of continuously generating high-quality and structured datasets that can be immediately computed by subsequent artificial intelligence models. This structure is not merely a simple accumulation of data, but rather an automated semantic integration and quality assurance of data. In use cases such as wide-area control systems for autonomous vehicles or real-time supply and demand adjustment systems for smart grids, this technology minimizes data processing latency.

[0096] Step 2: Generation and updating of the city's digital twin Using diverse data obtained from the city's data collection and integration department, the system recreates buildings, roads, transportation, energy infrastructure, environment, resident activities, infrastructure operation status, logistics networks, medical institutions, educational institutions, event venues, etc., in virtual space with high precision, taking the entire city and its regions as units, and updates the city's status in real time. For example, it integrates map data obtained from geographic information systems, 3D models of buildings, road networks, topographic data of rivers or mountains, location information of shelters or medical institutions, logistics networks, population distribution data, and the operation status of important infrastructure to faithfully reproduce the urban structure and activities of the real world. In the event of disasters, the system simulates the impact range of earthquakes, floods, mudslides, tsunamis, fires, power outages, large-scale communication failures, etc., and reflects the disaster scenario, inaccessible sections, isolated areas, the capacity of shelters, the number of patients that medical institutions can receive, and the operation status of logistics networks in real time on the digital twin. Specifically, in this step, the system dynamically reconstructs a voxel grid in three-dimensional space based on information from the integrated graph database. This system assigns multidimensional vectors representing physical, thermodynamic, and electromagnetic properties to each voxel. Whenever real-time sensor data is updated, the corresponding voxel vectors for spatial coordinates are overwritten and updated. Upon receiving a disaster trigger signal, the system initiates a structural analytical simulator using the finite element method. This simulator takes seismic wave acceleration data or water depth data as boundary conditions to calculate the stress distribution of buildings or the probability of infrastructure failure. For objects whose calculated failure probability exceeds a predetermined threshold, the system changes their state to "functional stop," logically severing the edges connecting them to the corresponding nodes in the graph database. This topology change is propagated throughout the network to determine the impact range of cascading failures. Through this process, the system achieves the technical effect of synchronizing real-world physical changes to virtual space with millisecond-level latency. This structure is not limited to the display of static 3D models but rigorously simulates dynamic state transitions based on physical laws on a computer. This technology provides an extremely realistic predictive environment in use cases such as the construction of digital twins for large-scale factories or city-level disaster prevention training simulators.

[0097] Step 3: Urban operation optimization and disaster response simulation through generative AI Generative AI is used as an essential component to analyze data from urban digital twins, enabling the optimization of urban operations and automated disaster response. For example, in the transportation sector, it analyzes traffic volume data, public transportation operation status, traffic light status, parking space availability, resident movement history, and event information to predict future congestion risks, and simulate, propose, and execute real-time optimizations for signal control, including increasing public transportation frequency, proposing detour routes, and automatically issuing traffic control orders. In the energy sector, it analyzes consumption and meteorological data to automatically send energy-saving alerts and instructions for utilizing decentralized power generation to optimize peak-hour electricity supply and demand balance. In the environmental sector, it analyzes air quality and noise data to automatically propose and execute measures such as traffic control, greening initiatives, and optimized waste collection routes. In disaster response, it automatically formulates and executes disaster expansion predictions and evacuation guidance, prioritizes infrastructure restoration, and plans for material distribution. Specific examples of prompts that generative AI can read include: "Based on current traffic volume data, public transportation operation status, and event information, predict the risk of congestion at each major intersection within the next 30 minutes and propose the best traffic mitigation measures, such as signal control or increasing bus frequency"; "Based on current energy consumption data, weather forecasts, and event information, identify areas where supply and demand balance becomes strained within the next 2 hours and formulate the best plan for sending out decentralized power generation or energy-saving alerts"; "Based on post-earthquake building damage data, shelter capacity, and road blockage information, formulate plans for prioritizing evacuation guidance and material distribution to areas with high isolation risk." Specifically, in this step, the system inputs the current state vector of the digital twin into a hybrid architecture of a large-scale language model and a deep reinforcement learning model. The system first uses a deep reinforcement learning model to rapidly explore a candidate set of optimal control parameters from a vast action space. This deep reinforcement learning model, as input, receives the congestion matrix of the traffic network or the load vector of the power grid. As output, this deep reinforcement learning model generates a continuous vector representing the extension of the green light for traffic lights or the discharge rate of a battery. This system converts the generated candidate set of control parameters into natural language prompts and inputs them into a large-scale language model. This large-scale language model compares the input prompts with a database of similar past cases, logically inferring the social impact and secondary risks of each candidate. As output, this large-scale language model generates structured data containing evaluation scores for each candidate, explanations of risk factors, and the ultimately recommended action plan. The system automatically selects the action plan with the highest evaluation score and sends control commands to the corresponding infrastructure control devices. Through this process, the system achieves advanced decision-making that integrates mathematical optimization and common-sense reasoning. This approach lends interpretability to the judgment process of AI, which is easily rendered opaque, making safe and reliable automatic control possible.This technology dramatically improves operational efficiency in use cases such as integrated control centers in smart cities or dynamic routing systems in wide-area logistics networks.

[0098] Step 4: Information Sharing and Visualization Through dashboards or notifications, the system provides real-time visualization and sharing of city status, forecasts, optimization proposals, and disaster response plans with city operators, infrastructure operators, residents, and other relevant stakeholders. For example, it visualizes the city's traffic conditions, energy supply and demand, environmental indicators, disaster damage or shelter capacity, and infrastructure recovery progress on maps or charts, enabling stakeholders to stay informed. For residents, it notifies them of optimal travel routes, energy-saving alerts, and disaster evacuation instructions via smartphone applications. For city operators, it lists city-wide risks and optimization proposals on dashboards. Specifically, in this step, the system dynamically converts the format of the generated optimization plans or forecast data based on the user's terminal characteristics and network bandwidth. For city operators' high-resolution displays, the system performs 3D particle rendering to visually represent microscopic changes in traffic or energy flow. For residents' smartphones, the system uses coordinate-based geofencing technology to filter and send relevant information only within a radius of a few kilometers of the user's current location. To minimize data transmission, the system distributes map data in vector tile format and performs rendering processing on the terminal side. This system uses a large language model to automatically generate natural language summary text that explains the characteristics of complex data charts, and provides visual cues along with the summary text. Through this process, the system effectively optimizes data cues for the information recipient's context and reduces cognitive load. This architecture is not merely a simple screen display, but rather a distributed rendering and adaptive data distribution system that fully utilizes the computing resources of the client terminal. This technology improves information comprehension and responsiveness in use cases such as personalized disaster prevention applications for citizens or augmented reality navigation systems for field workers.

[0099] Step 5: Model Evolution via Federated Learning By continuously learning from past city operation performance data and new event data through federated learning, the predictive and optimization models of generative AI evolve. For example, AI learns from past disaster response performance, the effectiveness of traffic congestion mitigation measures, the performance of energy supply and demand adjustments, the effectiveness of environmental policies, and the utilization of resident services, enabling more accurate predictions and optimizations for future city operations or disaster responses. It learns the characteristics of each region, such as increased demand for heating appliances or cold-weather gear in snowy areas during winter, delivery delays caused by traffic congestion in urban areas, the potential for isolated villages in mountainous regions, and increased demand for shelters or hygiene products during infectious disease outbreaks, automatically generating optimal plans that take these characteristics into account for the next disaster. Specific examples of prompts for the generative AI to read include: "Integrate past disaster response performance data with new event data to update the material demand prediction model for the next disaster," and "Generate a disaster response simulation model that reflects the characteristics of each region or past lessons." Specifically, in this step, the system uses a cluster of edge nodes distributed across local servers in various regions to execute the federated learning protocol. This system broadcasts the latest global model weight matrix from the central server to each edge node. Each edge node uses existing environmental data or resident behavior records collected within its jurisdiction to relearn its local model. During this relearning process, each edge node calculates the error between predicted and actual results as a loss function and uses stochastic gradient descent to calculate the updated weight vector. The central server receives the updated vectors from each edge node and integrates them using a secure aggregation algorithm. The system uses the integrated updated vectors to update the parameters of the global model, improving the model's generalization performance. The system evaluates the updated model using a test dataset and automatically deploys it to the production environment only if the accuracy exceeds a pre-defined benchmark. Through this process, the system achieves the technical effect of continuously evolving the intelligence of the entire system without centralizing highly confidential personal information or detailed infrastructure data in one place. This constitutes a mathematical solution to the trade-off between protecting data privacy and improving the performance of machine learning models. This technology enables secure and efficient model application in use cases such as wide-area collaborative disaster prevention systems across multiple autonomous systems or decentralized supply chain optimization for global enterprises.

[0100] (specific use case) For example, consider a scenario where a large-scale sporting event is held in the city center. In this case, the city's data collection and integration department obtains real-time data on vehicle numbers or average speeds, and traffic volume categorized by vehicle type, from traffic volume sensors located on major roads or intersections. Simultaneously, it collects data from public transportation management systems (bus or rail operation status or delays), traffic light control systems (signal switching timing or fault information), parking management systems (vacancy status or utilization rates), pedestrian flow sensors around the event venue (inflow and outflow data), and congestion and traffic disruption reports from social networking sites or notification applications. Furthermore, it integrates data from meteorological observation systems (temperature, humidity, precipitation, wind speed), environmental sensors (air quality or noise data), garbage truck locations or garbage bin fullness sensors, residents' mobile history from smartphones or transportation IC cards, and delivery vehicle location information from logistics outlets. Based on this data, the city's digital twin generation and updating department recreates a high-precision virtual spatial model of the entire city and the area surrounding the event venue, providing real-time visualization of traffic, pedestrian flow, environment, and infrastructure operations. Specifically, to address the unique context of the event, this system employs adaptive sensor control that dynamically increases the sampling rate of data collection. For a sensor group within a specific radius centered on the event venue, the system multicasts control commands to set the data transmission frequency to several times the normal rate. To handle the surge in data traffic, the system initiates stream processing engines on edge computing nodes, distributing primary processing such as moving average calculations or noise removal. To estimate crowd density from camera images, the system applies a convolutional neural network specialized for crowd counting. This neural network receives image tensors of the crowd as input and generates a density map as output, representing the probability that each pixel represents a human head. By spatially integrating the density map, the system calculates the accurate number of people remaining within a specific area. The system combines the calculated number data with the historical data from the transportation IC card ticket gates on a time axis to generate a vector field of crowd flow. This high-resolution data is mapped onto the voxel space of a digital twin, constructing a state capable of simulating the interaction between microscopic crowd movement and macroscopic traffic flow. Through this process, the system achieves the technical effect of capturing nonlinear changes in urban conditions caused by sudden events with zero latency and accuracy. This architecture does not rely on static data collection scheduling, but rather autonomously optimizes the observation network in response to changing conditions. In use cases such as security support systems for large-scale festivals or dynamic traffic control systems around stadiums, this technology enables the prevention and smooth guidance of crowd incidents.

[0101] The Urban Operations Optimization Department utilizes generative AI as a mandatory tool, analyzing traffic volume data, public transportation operation status, traffic light status, parking lot availability, event information, pedestrian flow sensor data, and SNS notification data to predict future congestion or mixed risks. For example, when predicting congestion or pedestrian concentration at major intersections or around stations before or after an event, generative AI simulates, proposes, and executes real-time optimizations for signal control, including temporary increases in bus or train services, recommendations for park-and-ride services, notifications of optimal routes to residents or attendees, and automatic issuance of traffic control measures. Furthermore, based on sensor information regarding the fullness of trash cans or the location of garbage trucks around the event venue, generative AI automatically generates optimal garbage collection routes to maximize post-event cleanup efficiency. Moreover, based on temperature or air quality data, in cases of increased risk of heatstroke or air pollution, generative AI automatically notifies event organizers or residents of potential precautions or countermeasures. Specifically, the City Operations Optimization Department collects high-resolution pedestrian flow vector fields and traffic flow data, inputting them into a spatiotemporal prediction model combining graph neural networks and Transformers. This spatiotemporal prediction model receives a multidimensional time-series tensor containing the number of people crossing and vehicles using intersections and pedestrian crossings over the past few hours, as well as weather conditions. As output, it outputs the predicted shift in congestion scores for each node over the next few hours. The City Operations Optimization Department identifies nodes whose predicted congestion scores exceed a safe threshold and calculates control actions to restrict inflow to those nodes. Using reinforcement learning algorithms, the department dynamically calculates the optimal values ​​for traffic light cycle length, green light ratio, and phase difference. The calculated control parameters are then input into a large-scale language model, which automatically generates detour guidance text for those arriving at the event. This large-scale language model receives the current congestion status, a list of recommended detour route coordinates, and the event's schedule information as input. As output, it generates a guidance message using simple language to avoid confusing those arriving at the event. The city operations optimization department pushes the generated messages to digital signage or a dedicated smartphone application. Through this process, the system proactively disperses traffic before localized congestion escalates into city-wide traffic paralysis. This approach is not a reactive measure, but rather a proactive crowd control system based on highly accurate future predictions. This technology significantly improves participant safety and convenience in use cases such as city-wide mobility management during large-scale events like international sporting events or logistics optimization for large-scale exhibitions.

[0102] From a disaster response perspective, in the event of a sudden rainstorm or earthquake, the city's data collection and integration department instantly gathers rapid reports from meteorological observation systems or seismometers, the capacity of shelters, information from road flooding sensors or building damage, and notification data from social media. The city's digital twin generation and updating department then reflects the disaster area or shelter congestion, impassable sections, and isolated risk areas in real time onto the virtual space. The city operations optimization department uses generative AI to automatically formulate and execute disaster expansion predictions, evacuation guidance plans, infrastructure restoration priorities, and material distribution plans, providing residents or those arriving at the scene with the best evacuation instructions or safety information in real time. Specifically, this system processes data in a distributed computing environment consisting of multiple computing nodes. The city's data collection and integration department converts raw data from various sensors into multi-dimensional time-series tensors. The city's digital twin generation and updating department maps the converted tensor data into 3D spatial coordinates. The processing department uses graph neural networks to model the spatial dependencies of urban infrastructure. The City Operations Optimization Department employs a composite AI architecture combining a large-scale language model and a reinforcement learning agent. Input data to this AI includes time-series arrays of meteorological data, graph-structured data representing infrastructure damage, or embedded vectors of SNS text. Specific examples of input data include time-series tensors of 50 mm / hour precipitation, cutoff markers for specific road nodes, or capacity values ​​for shelter nodes. The AI ​​processes this input data in a high-dimensional space. Output data from this AI includes impassable probability scores for each road link, congestion prediction probability distributions for each shelter, or a ranking list of recovery priorities. Specific examples of output data include an impassable probability of 0.85 for road A, a predicted capacity of 95% for shelter B, or a recovery priority score of 0.92. The system then passes the output results to subsequent processing. If the output score exceeds a pre-set threshold, the subsequent processing department performs a branching process that excludes that route from the calculation of evacuation routes. Unlike existing rule-based judgments performed by humans, this system autonomously constructs a nonlinear disaster impact model. This system analyzes wide-area disaster conditions at an unprecedented speed through parallel computation in high-dimensional space. This process dramatically improves the computational efficiency and prediction accuracy of evacuation guidance. Specific applications of this system include dynamic evacuation guidance and optimal deployment of rescue forces during localized torrential rains during large-scale outdoor festivals.

[0103] The Information Sharing and Visualization Department provides real-time visualization and sharing of city status, traffic, pedestrian flow, environment, disaster risks, and infrastructure recovery progress with city operators, infrastructure operators, event organizers, residents, and security and emergency personnel through dashboards or notifications. For example, dashboards for city operators display maps or charts showing congestion levels around major intersections or stations, public transportation operations, garbage collection progress, disaster damage or shelter capacity, and infrastructure recovery progress, enabling stakeholders to stay informed and make informed decisions. For residents or visitors, smartphone applications automatically provide notifications of optimal routes or ways to avoid congestion, energy-saving alerts, and evacuation instructions during disasters. Specifically, this Information Sharing and Visualization Department utilizes parallel processing of an image processing processor for rendering. The processing unit uses a neural network with an attention mechanism to dynamically optimize the displayed content for each user. The input data to this AI consists of profile vectors representing user attributes and multidimensional tensors representing the city's status. Specific examples of input data could include user vectors with administrator privilege indicators, congestion scores of 0.9 for specific intersections, or infrastructure recovery progress rates of 0.5. This AI extracts features from the input data and determines their importance. The output data from this AI consists of structured data defining the display layout and weighted scores for the importance of each user interface component. Specific examples of output data could include warning banner display indicator 1, a list of red highlighted coordinates on a map, or a generated string of notification text. This system passes the output structured data to subsequent processing. This subsequent processing unit calls the client terminal's application programming interface (API) based on the structured data to asynchronously update the screen display. This system is not simply a list of information, but automatically performs information filtering based on user conditions within the computer. Through this processing, this system achieves the technical effect of significantly reducing network communication load and preventing delays in depiction processing during disasters when communication bandwidth is easily limited. Specific application areas of this system include the use of daily traffic management dashboards in smart cities and personalized emergency information distribution during disasters.

[0104] The Federated Learning Evolution Unit enables AI to learn from performance data of current event operations or disaster response, the effectiveness of traffic congestion mitigation measures, the performance of energy supply and demand adjustments, the effectiveness of environmental policies, and the utilization of residential services. This is to allow for more accurate predictions and optimizations in future large-scale events or disaster responses, thus evolving the generative AI model. Specifically, this Federated Learning Evolution Unit performs federated learning across multiple edge devices and a central server. The system utilizes servers in each autonomous region or residents' portable terminals as edge nodes. These edge nodes use locally accumulated performance data to perform learning using a neural network-based error backpropagation method. Each edge node calculates gradients to minimize the mean squared error and cross-entropy error between the predicted and performance values. The input data for this AI during learning consists of a tensor of past prediction results and a tensor representing the correct answer label of the actual result. Specific examples of input data include predicted congestion length vectors, actual congestion length vectors, or numerical data on predicted energy consumption. The output data from this AI during learning is a gradient vector representing the weight updates of each neural network layer. Specific examples of output data could include the updated difference of the weight matrix or the updated difference of the bias term in Layer 1, or the change in the attention weights. Each edge node only sends the calculated gradient vectors to the central server. The central server aggregates the received gradient vectors and updates the parameters of the global model. The system then redistributes the updated parameters to each edge node as subsequent processing. This subsequent processing unit synchronizes the local model of each edge node with the latest global model. This system employs a distributed processing architecture fundamentally different from existing methods that reduce raw data to a central server. Through this processing, the system achieves the technical effect of protecting highly confidential personal information or detailed infrastructure data while reducing communication overhead in a distributed computing environment and improving the generalization performance of the model. Specific application areas of this system include the sharing of proprietary disaster response technologies among multiple autonomous systems and the collaborative cultivation of AI models through wide-area collaboration.

[0105] Specific examples of prompts that can be read by generative AI include: "Based on current traffic volume data, public transportation operation status, event information, pedestrian flow sensor data, and SNS notification data, predict the risk of congestion and traffic jams around major intersections and stations in the next hour, and propose the best traffic mitigation measures such as signal control, increasing public transportation frequency, park-and-ride recommendations, and notifications to avoid congestion routes," "Based on the information from garbage bin fullness sensors and garbage truck location information around the event venue, automatically generate the best garbage collection route to maximize cleaning efficiency," "Based on data from meteorological observation systems and air quality sensors, automatically notify and remind users of precautions and countermeasures when the risk of heatstroke or air pollution increases," and "Based on the occupancy status of shelters during a disaster, road flooding sensors, building damage information, and SNS notification data, automatically formulate disaster expansion predictions, evacuation guidance plans, infrastructure restoration priorities, and material distribution plans, and notify residents or visitors of the best evacuation instructions in real time." Specifically, this system has a prompt generation department that automates the prompting process for natural language processing models. The prompt generation unit converts numerical arrays of various sensor data collected by the system into a natural language labeled sequence. The processing unit generates embedding vectors from these converted labeled sequences. The input data to this AI is a high-dimensional embedding vector containing contextual information. Specific examples of input data include a labeled vector representing traffic volume data of 500 vehicles per hour, a context vector containing the flag value 1 from a garbage bin full sensor, or a numerical vector representing meteorological data of 35 degrees Celsius. This AI uses a self-attention mechanism based on the Transformer architecture to resolve dependencies between input vectors. The output data from this AI consists of the text-generated results of the proposed countermeasures and a structured set of control commands. Specific examples of output data include extending the green light by 15 seconds, a list of garbage truck patrol nodes, or polygonal coordinate data representing the distribution area for evacuation instructions. This system passes the output control command set to subsequent processing. The subsequent processing unit performs syntax parsing on the control command set, converts it into communication requests for each infrastructure control system, and sends them. This system comprehensively interprets both unstructured and structured data, automatically generating dynamic control parameters for complex events that are difficult to handle using rule-based methods. Through this process, the system achieves a significant leap in the efficiency of urban infrastructure utilization, realizing autonomous and optimized control without human intervention. Specific applications of this system include dynamic traffic control during large-scale sporting events and automated infrastructure control based on environmental changes.

[0106] In this way, the system of this invention collects and integrates various types of urban data in real time, generates and updates a digital twin of the city with high precision, and uses generative AI as a necessary component to optimize urban operations and automatically respond to disasters. It visualizes and shares information, and evolves the model through federated learning, thereby achieving continuous optimization and improved resilience of the entire city. Specifically, this system adopts a multi-agent architecture with multiple different neural network models working together. The system comprehensively performs image analysis through convolutional neural networks, time series prediction through recurrent neural networks, and optimization control through reinforcement learning. The data input to this AI is multimodal data of the entire city. Specific examples of input data include RGB image tensors obtained from surveillance cameras, time series numerical matrices representing electricity consumption over the past year, and embedding vectors from residents' feedback text. This AI projects these heterogeneous data into a common latent space to calculate the optimization indicators for the entire city. The output data from this AI is the optimization indicator score for the entire city and the allocation ratio of computing resources to each subsystem. Specific examples of output data include a city resilience score of 0.88, a 40% allocation of computing resources to the traffic control module, and a 60% allocation to the disaster prediction module. This system transmits the output resource allocation ratio to subsequent processing. This subsequent processing unit controls the container orchestration tools on the cloud infrastructure to dynamically scale computing nodes horizontally. The system dynamically reallocates computing resources based on changes in the city's conditions. Through this processing, the system achieves the technical effect of minimizing processing latency during emergencies while suppressing overall system power consumption, thereby improving the availability and processing efficiency of the computer system itself. Specific application areas for this system include a seamless transition from routine smart city applications to emergency response modes during large-scale disasters, and the complete automation of energy management across the entire city.

[0107] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the smart device 14. In the smart device 14, the control unit 46A causes the output device 40 to output the result of the specific processing. The microphone 38B acquires voice representing the user's input on the result of the specific processing. The control unit 46A sends the voice data representing the user's input acquired by the microphone 38B to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0108] Data generation model 58 is what is known as generative AI (Artificial Intelligence).

[0109] One example of data generation model 58 includes ChatGPT (registered trademark) (Internet search).<URL: https: / / openai.com / blog / chatgpt> Generative AI, etc.

[0110] The data generation model 58 is obtained by performing deep learning on a neural network. A prompt containing instructions is input to the data generation model 58, along with inference data such as speech data representing speech, text data representing text, and image data representing images (e.g., data from still images or moving images). The data generation model 58 infers the input inference data based on the instructions shown in the prompt and outputs the inference result in one or more data forms, such as speech data, text data, and image data. The data generation model 58 may include, for example, text generation AI, image generation AI, and multimodal generation AI. Here, inference refers to, for example, analysis, classification, prediction, and / or summarization. The specific processing unit 290 performs the aforementioned specific processing while using the data generation model 58. The data generation model 58 may also be a fine-tuned model to output inference results from prompts without instructions; in this case, the data generation model 58 can output inference results from prompts without instructions. The data processing device 12, etc., includes multiple data generation models 58, including AI other than generative AI. AI other than generative AI includes, but is not limited to, linear regression, logistic regression, decision trees, random forests, support vector machines (SVM), k-means clustering, convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks (GAN), or Naive Bayes. Furthermore, AI can also act as an AI agent. Additionally, when the processing described above is performed by AI, this processing may be partially or entirely performed by AI, but is not limited to this example. Moreover, processing performed by AI, including generative AI, can be replaced by rule-based processing, and vice versa.

[0111] Furthermore, the processing performed by the aforementioned data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the smart device 14, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the smart device 14. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the smart device 14 or external devices, and the smart device 14 acquires or collects information required for processing from the data processing device 12 or external devices.

[0112] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the smart device 14.

[0113] [Second Implementation] Figure 3 An example of the configuration of the data processing system 210 according to the second embodiment is shown.

[0114] like Figure 3 As shown, the data processing system 210 includes a data processing device 12 and smart glasses 214. One example of the data processing device 12 is a server.

[0115] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0116] The smart glasses 214 include a computer 36, a microphone 238, a speaker 240, a camera 42, and a communication I / F 44. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, and communication I / F 44 are also connected to the bus 52.

[0117] Microphone 238 receives instructions from user 20 by receiving the sounds emitted by user 20. Microphone 238 captures the sounds emitted by user 20, converts the captured sounds into speech data, and outputs it to processor 46. Speaker 240 outputs sound according to the instructions from processor 46.

[0118] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, used to capture the area around user 20 (e.g., the shooting range defined by an angle of view equivalent to the field of vision of an average healthy person).

[0119] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F 44 and 26 is performed in a secure state.

[0120] Figure 4 An example of the main functions of the data processing device 12 and the smart glasses 214 is shown. Figure 4 As shown, in the data processing device 12, specific processing is performed by the processor 28. The memory 32 stores a specific processing program 56.

[0121] The specific processing program 56 is an example of a "program" as disclosed herein. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0122] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290. The specific processing unit 290 can use the emotion-specific model 59 to infer the user's emotions and perform specific processing based on the user's emotions. The emotion inference function using the emotion-specific model 59 (emotion-specific function) includes various inferences and predictions related to the user's emotions, but is not limited to this example. Furthermore, emotion inference and prediction also include, for example, emotion analysis (analysis).

[0123] In the smart glasses 214, the processor 46 performs the receive / output processing. The memory 50 stores the receive / output program 60. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on the RAM 48. The receive / output processing is implemented by the processor 46 operating as a control unit 46A based on the receive / output program 60 executed on the RAM 48. Furthermore, the smart glasses 214 may also have the same data generation model and emotion-specific model as the data generation model 58 and the emotion-specific model 59, and use these models to perform the same processing as the specific processing unit 290.

[0124] Next, the specific processing performed by the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the smart glasses 214. In the following description, the data processing device 12 will be referred to as the "server" and the smart glasses 214 will be referred to as the "terminal".

[0125] (Example 1) Since the specific processing flow is the same as that in Embodiment 1 described in the first embodiment above, the description is omitted.

[0126] (Application Example 1) Since the specific processing flow is the same as that in Application Example 1 described in the first embodiment above, the description is omitted.

[0127] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the smart glasses 214. In the smart glasses 214, the control unit 46A causes the speaker 240 to output the result of the specific processing. The microphone 238 acquires voice representing the user's input on the specific processing result. The control unit 46A sends the voice data representing the user's input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0128] Furthermore, the processing performed by the aforementioned data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the smart glasses 214, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the smart glasses 214. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the smart glasses 214 or external devices, and the smart glasses 214 acquires or collects information required for processing from the data processing device 12 or external devices.

[0129] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the smart glasses 214.

[0130] [Third Implementation] Figure 5 An example of the configuration of the data processing system 310 according to the third embodiment is shown.

[0131] like Figure 5 As shown, the data processing system 310 includes a data processing device 12 and a head-mounted terminal 314. An example of the data processing device 12 is a server.

[0132] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0133] The head-mounted terminal 314 includes a computer 36, a microphone 238, a speaker 240, a camera 42, a communication I / F 44, and a display 343. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, display 343, and communication I / F 44 are also connected to the bus 52.

[0134] Microphone 238 receives instructions from user 20 by receiving the sounds emitted by user 20. Microphone 238 captures the sounds emitted by user 20, converts the captured sounds into speech data, and outputs it to processor 46. Speaker 240 outputs sound according to the instructions from processor 46.

[0135] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, used to capture the area around user 20 (e.g., the shooting range defined by an angle of view equivalent to the field of vision of an average healthy person).

[0136] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F 44 and 26 is performed in a secure state.

[0137] Figure 6 An example of the main functions of the data processing device 12 and the head-mounted terminal 314 is shown. Figure 6 As shown, in the data processing device 12, specific processing is performed by the processor 28. The memory 32 stores a specific processing program 56.

[0138] The specific processing program 56 is an example of a "program" as disclosed herein. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0139] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290.

[0140] In the head-mounted terminal 314, the processor 46 performs the receiving and output processing.

[0141] The memory 50 stores a receive / output program 60. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on the RAM 48. The receive / output processing is implemented by the processor 46 operating the control unit 46A based on the receive / output program 60 executed on the RAM 48.

[0142] Next, the specific processing performed by the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the head-mounted terminal 314. In the following description, the data processing device 12 will be referred to as the "server" and the head-mounted terminal 314 will be referred to as the "terminal".

[0143] (Example 1) Since the specific processing flow is the same as that in Example 1 described in the first embodiment above, the description is omitted.

[0144] (Application Example 1) Since the specific processing flow is the same as that in Application Example 1 described in the first embodiment above, the description is omitted.

[0145] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the head-mounted terminal 314. In the head-mounted terminal 314, the control unit 46A causes the speaker 240 and the display 343 to output the result of the specific processing. The microphone 238 acquires voice representing the user's input on the specific processing result. The control unit 46A sends the voice data representing the user's input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0146] Furthermore, the processing performed by the aforementioned data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the head-mounted terminal 314, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the head-mounted terminal 314. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the head-mounted terminal 314 or external devices, and the head-mounted terminal 314 acquires or collects information required for processing from the data processing device 12 or external devices.

[0147] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the head-mounted terminal 314.

[0148] [Fourth Implementation] Figure 7 An example of the configuration of the data processing system 410 according to the fourth embodiment is shown.

[0149] like Figure 7 As shown, the data processing system 410 includes a data processing device 12 and a robot 414. An example of the data processing device 12 is a server.

[0150] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0151] Robot 414 includes a computer 36, a microphone 238, a speaker 240, a camera 42, a communication I / F 44, and a control object 443. Computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, control object 443, and communication I / F 44 are also connected to the bus 52.

[0152] Microphone 238 receives instructions from user 20 by receiving the sounds emitted by user 20. Microphone 238 captures the sounds emitted by user 20, converts the captured sounds into speech data, and outputs it to processor 46. Speaker 240 outputs sound according to the instructions from processor 46.

[0153] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, used to capture the area around user 20 (e.g., the shooting range defined by an angle of view equivalent to the field of vision of an average healthy person).

[0154] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F 44 and 26 is performed in a secure state.

[0155] The controlled object 443 includes a display device, LEDs for the eyes, and motors for driving the arms, hands, and feet. The posture and movements of the robot 414 are controlled by controlling the motors for the arms, hands, and feet. Some of the robot 414's emotions can be expressed by controlling these motors. In addition, facial expressions of the robot 414 can also be expressed by controlling the illumination state of the LEDs for the robot 414's eyes.

[0156] Figure 8 An example of the main functions of the data processing device 12 and the robot 414 is shown. For example... Figure 8 As shown, in the data processing device 12, specific processing is performed by the processor 28. The memory 32 stores a specific processing program 56.

[0157] The specific processing program 56 is an example of a "program" as disclosed herein. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0158] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the special processing unit 290.

[0159] In robot 414, the processor 46 performs the receive / output processing. The memory 50 stores the receive / output program 60. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on RAM 48. The receive / output processing is implemented by the processor 46 operating the control unit 46A based on the receive / output program 60 executed on RAM 48.

[0160] Next, the specific processing performed by the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the robot 414. In the following description, the data processing device 12 will be referred to as the "server" and the robot 414 will be referred to as the "terminal".

[0161] (Example 1) Since the specific processing flow is the same as that in Example 1 described in the first embodiment above, the description is omitted.

[0162] (Application Example 1) Since the specific processing flow is the same as that in Application Example 1 described in the first embodiment above, the description is omitted.

[0163] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the robot 414. In the robot 414, the control unit 46A causes the speaker 240 and the controlled object 443 to output the result of the specific processing. The microphone 238 acquires voice representing the user's input on the specific processing result. The control unit 46A sends the voice data representing the user input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0164] Furthermore, the processing performed by the data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the robot 414, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the robot 414. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the robot 414 or external devices, and the robot 414 acquires or collects information required for processing from the data processing device 12 or external devices.

[0165] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the robot 414.

[0166] Furthermore, the emotion-specific model 59, serving as an emotion engine, can determine the user's emotion based on a specific mapping. Specifically, the emotion-specific model 59 can determine the user's emotion based on an emotion graph that serves as a specific mapping (see...). Figure 9 The system determines the user's emotions. In addition, the emotion-specific model 59 can also determine the robot's emotions in the same way, and the specific processing unit 290 can also perform specific processing using the robot's emotions.

[0167] Figure 9This is a diagram representing an emotion map 400 that maps various emotions. In the emotion map 400, emotions are arranged radially from the center in concentric circles. The closer to the center of the concentric circles, the more primitive the emotion is. Further out on the concentric circles, emotions are arranged representing states or actions arising from mood. Emotion is a concept that includes both feelings and mental states. To the left of the concentric circles, emotions generated by reactions occurring in the brain are arranged roughly. To the right of the concentric circles, emotions guided by situational judgments are arranged roughly. Above and below the concentric circles, emotions generated by reactions occurring in the brain and guided by situational judgments are arranged roughly. Furthermore, the emotion of "pleasure" is arranged above the concentric circles, and the emotion of "unpleasantness" is arranged below. Thus, in the emotion map 400, various emotions are mapped according to the structure of emotion generation, while easily generated emotions are mapped nearby.

[0168] These emotions are distributed at the 3 o'clock position on the Emotion Chart 400, and usually fluctuate between peace and unease. In the right half of the Emotion Chart 400, because situational awareness is more dominant than internal feelings, it gives a sense of calm.

[0169] The inner side of the emotion diagram 400 represents the mind, and the outer side of the emotion diagram 400 represents actions. Therefore, the further you go to the outer side of the emotion diagram 400, the more the emotion can be seen (manifested in actions).

[0170] Here, human emotions are based on a balance of various factors such as posture and blood sugar levels. When these balances deviate from the ideal, it indicates unhappiness; when they approach the ideal, it indicates pleasure. In robots, cars, and motorcycles, emotions can also be created based on a balance of factors such as posture and remaining battery power. When these balances deviate from the ideal, it indicates unhappiness; when they approach the ideal, it indicates pleasure. Emotion maps can be generated, for example, based on Dr. Mitsuyoshi's emotion map (Research on Speech Emotion Recognition and Brain Physiological Signal Analysis Systems for Emotion, Tokushima University, Doctoral Dissertation: https: / / ci.nii.ac.jp / naid / 500000375379). In the left half of the emotion map, emotions belonging to the "Reaction" domain, where sensation is dominant, are arranged. Furthermore, in the right half of the emotion map, emotions belonging to the "Situation" domain, where situational cognition is dominant, are arranged.

[0171] The emotion map defines two types of emotions that promote learning. One is a negative emotion located near the middle of "repentance" or "reflection" on the situation side. That is, when the robot experiences negative emotions such as "I never want to feel this way again" or "I never want to be scolded again." The other is a positive emotion located near "desire" on the response side. That is, when the robot experiences positive feelings such as "wanting more" or "wanting to know more."

[0172] The emotion-specific model 59 feeds user input into a pre-learned neural network to obtain emotion values ​​representing each emotion shown in the emotion graph 400, and determines the user's emotion. This neural network is pre-learned based on multiple learning data sets that combine user input with emotion values ​​representing each emotion shown in the emotion graph 400. Furthermore, this neural network is learned to... Figure 10 As shown in sentiment graph 900, sentiment values ​​in nearby configurations are similar to each other. Figure 10 Examples show that multiple emotions such as "peace of mind", "stability", and "reassurance" have similar emotional values.

[0173] The above description primarily focuses on the functions of the data processing device 12 to illustrate the system disclosed herein. However, the system disclosed herein may not necessarily be implemented on a server. The system disclosed herein can be implemented as a general information processing system. For example, it can be implemented as a software program running on a personal computer, an application running on a smartphone, etc. The methods disclosed herein can be provided to users in the form of SaaS (Software as a Service).

[0174] In the above embodiments, an example of specific processing performed by a single computer 22 is given. However, the technology disclosed herein is not limited to this, and distributed processing of specific processes by multiple computers, including computer 22, is also possible. For example, the data generation model 58 may also be provided in an external device of the data processing apparatus 12, where data is generated based on input data.

[0175] In the above embodiments, an example of storing a specific processing program 56 in memory 32 is illustrated, but the technology disclosed herein is not limited thereto. For example, the specific processing program 56 may also be stored in a portable computer-readable non-transitory storage medium such as a USB (Universal Serial Bus) memory. The specific processing program 56 stored in the non-transitory storage medium is installed in the computer 22 of the data processing device 12. The processor 28 performs specific processing according to the specific processing program 56.

[0176] Alternatively, the specific processing program 56 can be stored in a storage device such as a server connected to the data processing device 12 via a network 54, and the specific processing program 56 can be downloaded and installed into the computer 22 upon request from the data processing device 12.

[0177] Furthermore, it is not necessary to store the entire specific process 56 in a storage device such as a server connected to the data processing device 12 via the network 54, nor is it necessary to store the entire specific process 56 in the memory 32; a portion of the specific process 56 may also be stored.

[0178] As a hardware resource for performing specific processing, various processors can be used. For example, a CPU can be listed as a processor; it is a general-purpose processor that functions as a hardware resource for performing specific processing by executing software, i.e., a program. Furthermore, as a processor, a dedicated circuit can be listed as a processor; it is a processor with a circuit structure specifically designed for performing specific processing, such as an FPGA (Field-Programmable Gate Array), a PLD (Programmable Logic Device), or an ASIC (Application Specific Integrated Circuit).

[0179] Every processor has built-in or connected memory, and every processor performs specific processes by using memory.

[0180] The hardware resources for performing a specific process can consist of one of these various processors, or a combination of two or more processors of the same or different types (e.g., a combination of multiple FPGAs, or a combination of a CPU and an FPGA). Furthermore, the hardware resources for performing a specific process can also be a single processor.

[0181] As an example of a single processor, the first type consists of a combination of one or more CPUs and software, which functions as a hardware resource to perform specific processing. The second type uses a processor, such as a System-on-a-chip (SoC), which implements the entire system functionality, including multiple hardware resources performing specific processing, using a single IC chip. In this case, the specific processing is implemented using one or more of the aforementioned processors that serve as hardware resources.

[0182] Furthermore, as the hardware architecture of these various processors, more specifically, circuits combining semiconductor elements and other circuit components can be used. Moreover, the specific process described above is merely an example. Therefore, it goes without saying that, without departing from the main point, unnecessary steps can be removed, new steps can be added, or the processing order can be changed.

[0183] The foregoing descriptions and illustrations are detailed explanations of the parts covered by this disclosure and are merely one example of this disclosure. For instance, the descriptions of the above-described structure, function, role, and effect are just one example of the structure, function, role, and effect of the parts covered by this disclosure. Therefore, it goes without saying that, without departing from the spirit of this disclosure, unnecessary parts can be deleted, new elements can be added, or replacements can be made to the foregoing descriptions and illustrations. Furthermore, to avoid confusion and facilitate understanding of the parts covered by this disclosure, explanations of technical common sense that does not require special explanation for implementing this disclosure have been omitted from the foregoing descriptions and illustrations.

[0184] All documents, patent applications and technical standards described in this specification are incorporated herein by reference as if they were specifically and individually described as incorporated by reference.

[0185] In addition, the following notes further explain the above explanation.

[0186] (Note 1) A system, characterized by comprising: a city data collection and integration unit, which collects and integrates in real time various types of city data, including traffic volume sensor data, public transportation operation data, energy consumption data, environmental sensor data, resident mobility records, disaster occurrence information, infrastructure operation status data, resident service utilization records, event information, information from SNS or notification applications, and the operation status of IoT home appliances; a city digital twin generation and updating unit, which uses the data obtained from the city data collection and integration unit to reproduce buildings, roads, traffic, energy infrastructure, environment, resident activities, etc., in virtual space with high precision, taking the entire city and each region as units, and updates the city's status in real time; and a city operation optimization unit, which uses generative AI as a mandatory component to analyze the data constructed in the city digital twin generation and updating unit. The data on the city's digital twin automatically simulates, proposes, and implements traffic congestion prediction and mitigation strategies, optimizes energy supply and demand balance, guides evacuation and controls infrastructure during disasters, proposes measures to reduce environmental impact, personalizes resident services, and optimizes urban events. The information sharing and visualization department visualizes and shares the city's status, predictions, optimization proposals, and disaster response plans in real time with city operators, infrastructure operators, residents, and other relevant personnel through dashboards or notifications. The federated learning and evolution department continuously learns from the city's operational performance data and new event data through federated learning, enabling the generative AI's prediction and optimization models to evolve. The city operation optimization department operates with generative AI as a mandatory component, achieving overall city operation optimization and improved disaster response capabilities.

[0187] (Note 2) As described in Appendix 1, the system is characterized in that the Urban Operation Optimization Department, in the field of transportation, uses generative AI to analyze traffic volume data, public transportation operation status, traffic light status, parking lot vacancy status, resident movement history, event information, etc., to predict the risk of future congestion, and simulate, propose, and execute signal control optimization, increase public transportation frequency, propose detour routes, and automatically release traffic control measures in real time, thereby achieving smooth urban traffic and reduced environmental burden.

[0188] (Note 3) As described in Appendix 1, the system is characterized in that, when a disaster occurs, the Urban Operations Optimization Department uses generative AI to analyze disaster occurrence information, shelter occupancy status, infrastructure damage status, resident location information, and notification data from SNS or notification applications, and automatically formulates and executes disaster expansion prediction, evacuation guidance plans, infrastructure restoration priority, and material distribution plans, and provides real-time visualization and notification of the best disaster response plan to urban operators and residents.

Claims

1. A system, characterized in that, include: The Data Collection and Integration Department collects and integrates various data in real time from autonomous regions and related institutions across the country, including data on mobile personnel, operational status of transportation infrastructure, number of people in shelters, inventory of supplies, and disaster information. The system includes a Digital Twin Generation and Update Department, which uses data obtained from the Data Collection and Integration Department to accurately reproduce the status of cities and regions in the real world in virtual space, on a national and regional basis. It also includes a Disaster Response Prediction and Simulation Department, which, when a disaster occurs, reflects the disaster event in real time in the model constructed by the Digital Twin Generation and Update Department. Using generative AI, based on past disaster data and real-time data, it accurately predicts and simulates the shift in the number of evacuees, the required quantity of supplies, the operational status of transportation infrastructure, and material distribution plans for each shelter, automatically generating the optimal material distribution plan. The Disaster Response Prediction and Simulation Department operates with generative AI as a mandatory component, using federated learning to continuously learn from the dispersed data across the country, enabling flexible responses to unknown disaster patterns and regionally specific challenges. Furthermore, the system includes an Information Sharing and Visualization Department, which shares and visualizes information such as the material needs, inventory status, distribution plans, and transportation progress of each shelter in real time among autonomous regions, the nation, and relevant agencies, achieving rapid and accurate support for disaster victims.

2. The system as described in claim 1, characterized in that, The disaster response prediction and simulation unit does not wait for requests from autonomous regions when a disaster occurs. As a push-type support provided by generative AI, it automatically arranges essential supplies in order of priority. Based on the needs of supplies, the inventory of each network point, road conditions, and the operation status of transportation vehicles, it automatically calculates the best delivery routes and methods, automatically generates delivery routes and sequences, the number of vehicles required, and the estimated arrival time, and optimizes the delivery to multiple shelters.

3. The system as described in claim 1, characterized in that, The information sharing and visualization unit displays the shortage of supplies in shelters across the country, the current location and estimated arrival time of delivery vehicles, the number of evacuees in each shelter, the consumption patterns of supplies, and the operational status of transportation infrastructure on a map in real time on the dashboard screen. This allows relevant personnel to grasp the situation instantly from terminals such as smartphones and personal computers, and make quick decisions such as additional arrangements and route changes. Furthermore, through federated learning, it continuously learns from disaster response performance data and new event data, thereby evolving the simulation accuracy and prediction model.

Images