Data processing system

Abstract

The present application relates to a system for making the selection of the installation site of a wireless base station efficient and advanced, having the following structure: a topographic data acquisition unit, a population density data acquisition unit, a radio wave propagation characteristic data acquisition unit collect various data from multiple information sources, a generative AI processing unit normalizes and weights these data to calculate a comprehensive score, and a candidate site selection unit extracts the best installation candidate site. In addition, the AR visualization unit and the VR visualization unit visually display the candidate site information in three-dimensional space, and the site manager can compare and discuss in the field or virtual space. The site information or the performance is also continuously learned and reflected by the generative AI, and the present application simultaneously realizes the optimization of the installation plan and the rapidity of the decision-making.

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] The problem this invention aims to solve is that, when selecting locations for wireless base stations, although it is necessary to comprehensively and objectively evaluate multiple factors such as terrain, population density, and radio wave propagation characteristics, previous methods required the separate collection and analysis of this information, and manual comparison and discussion on spreadsheet software or maps. This necessitates a significant amount of time and labor, and demands specialized knowledge and experience, making it difficult to achieve efficient and optimized deployment plans. Consequently, problems arise such as reliance on personal judgment or omissions in site selection, increased costs, and uneven communication quality after deployment. Furthermore, previous methods for exploring candidate sites only involved abstract comparisons using maps or numerical data, failing to intuitively grasp the terrain or surrounding environment, the distribution of communication demand, and the ease of radio wave propagation. Therefore, these methods also fail to fully achieve rapid decision-making or sophisticated deployment plans. Therefore, the objective of this invention is to use generative AI to comprehensively analyze various information such as terrain data, population density data, and radio wave propagation characteristics data, automatically extract the best candidate sites for deployment, and utilize AR or VR-based 3D visualization capabilities to provide managers with an intuitive and efficient way to compare and discuss the environment of deployment sites, thereby simultaneously achieving efficiency in wireless base station deployment planning, cost reduction, uniformity of communication quality, rapid decision-making, and elimination of reliance on personal judgment. Summary of the Invention

[0005] The means to solve the problem involved in this invention is to construct a system comprising a terrain data acquisition unit, a population density data acquisition unit, a radio wave propagation characteristic data acquisition unit, a generative AI processing unit, a candidate site selection unit, an AR visualization unit, and a VR visualization unit. Specifically, the terrain data acquisition unit acquires terrain information such as elevation or slope from a geographic information system or external database; the population density data acquisition unit acquires population density information for each location from statistical data or sensor data; and the radio wave propagation characteristic data acquisition unit acquires information about radio wave attenuation or propagation characteristics at each location from simulated data or measured values. This acquired data is input into the generative AI processing unit, which normalizes the terrain data, population density data, and radio wave propagation characteristic data, weights each element, and calculates a comprehensive score. The candidate site selection unit automatically extracts candidate sites suitable for wireless base station installation based on the comprehensive score calculated by the generative AI processing unit. In addition, the AR visualization department and VR visualization department can intuitively and visually display information such as coordinates, altitude, population density, radio wave propagation characteristics, and comprehensive scores of the candidate sites extracted by the candidate site selection department in three-dimensional space. This allows the person in charge to easily compare and discuss the on-site conditions or various factors, eliminating the complicated comparisons based on manual operation or relying on personal judgment in the past. This is a means to achieve efficiency and optimization of wireless base station deployment plans. 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] Figure Labels 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) The embodiments for implementing this invention will be described in more detail below. The system of this invention is implemented through the collaborative operation of a server and a terminal. On the server side are installed a terrain data acquisition unit, a population density data acquisition unit, a radio wave propagation characteristic data acquisition unit, a generative AI processing unit, and a candidate site selection unit; on the terminal side are installed an AR visualization unit and a VR visualization unit. The structure, function, processing flow, generative AI utilization method, role division between the server and terminal, specific utilization examples, and extension examples of each unit will be described comprehensively and from multiple perspectives below. Specifically, this system adopts a distributed computing architecture. This system connects the parallel computing cluster using GPUs on the server side to the edge processor on the terminal side via a broadband network. This system performs large-scale tensor operations on the server side. This system performs low-latency rendering processing on the terminal side. This system thus optimally distributes the computational load. This system configures multiple data acquisition modules on the server side. This system containers each module as an independent microservice. This system thus improves scalability. This system, as the generative AI processing unit, installs a large-scale language model based on Transformer and a graph neural network. This system, serving as a candidate site selection unit, possesses computational circuitry capable of executing heuristic search algorithms. The system's terminal side is equipped with a depth sensor and inertial measurement unit for spatial identification. Real-time 3D coordinate transformation processing is performed in both AR and VR visualization units. Through the synergy of these hardware and software components, this system completely eliminates the need for manual map verification, a task previously performed by humans. This system calculates base station placement locations as an optimization problem in a high-dimensional data space. This results in a significant increase in the speed of the selection process. Simultaneously, the system reduces selection accuracy errors. As a specific application area, this system is utilized in the high-density deployment plan of millimeter-wave band base stations for next-generation communication standards in urban areas. This system is also applied to the rapid deployment and determination of temporary base stations during disaster recovery.

[0040] First, the terrain data acquisition unit runs on a server, acquiring terrain data from various information sources, including geographic information systems, external map databases, satellite imagery services, UAV aerial photography data, and publicly available data from autonomous regions. For example, it can change the resolution to acquire elevation grid data in 1-meter, 5-meter, and 10-meter units, allowing for differentiation based on the accuracy requirements of candidate sites. Furthermore, it can acquire tilt angle data, land use classification data (residential areas, commercial areas, industrial areas, farmland, forests, rivers, etc.), foundation strength data, soil classification data, and flood and debris flow disaster risk data. For instance, it can also incorporate automatic exclusion of areas with weak foundations and areas with a history of past disasters. Specifically, the terrain data acquisition unit performs asynchronous communication with multiple external API endpoints. It converts the acquired heterogeneous spatial data into a common three-dimensional voxel grid format. The elevation grid data is expanded in memory as a two-dimensional floating-point array. The unit dynamically downsamples or upsamples the array resolution based on the breadth of the target area. The terrain data acquisition unit automatically calculates tilt angle data through spatial differentiation operations on elevation arrays. It uses land use classification data as a categorical variable and performs one-hot encoding. It assigns foundation strength data and disaster risk data as attribute vectors for each voxel. The unit constructs these preprocessed data into a multidimensional tensor. This tensor is then output to the subsequent generative AI processing unit. The unit incorporates rule-based filtering circuitry. It applies masking to voxels with foundation strength below a predetermined threshold. It sets the computational weight of valid coordinate regions with disaster history markers to zero. This masking process suppresses unnecessary computational resource consumption during subsequent AI inference, thereby improving the overall system processing speed. In specific applications, the unit is used for surveying the feasibility of constructing tower-type base stations in mountainous areas. It is also used in the design of secure communication infrastructure that avoids flood-prone areas along rivers.

[0041] The population density data acquisition unit operates on a server, acquiring various population and pedestrian flow data, including census data, basic resident register data from municipalities, base station connection logs from portable terminals, traffic flow sensor data, social media post location information, Wi-Fi access point connection data, store visitor data from commercial facilities, and attendance data from event organizers. For example, it can grasp population density changes in each time period or situation, such as weekdays and weekends, day and night, seasons, and event times, and reflect these changes in the evaluation of candidate sites. Furthermore, it acquires population forecast data for future urban development plans or redevelopment areas, enabling site planning based on anticipated future needs. Specifically, the population density data acquisition unit is equipped with a streaming data processing engine. It accumulates real-time incoming base station connection logs or sensor data into a time-series database. It performs statistical analysis on the acquired raw data for each time range. The unit structures the statistical results into time-series vectors for each spatial grid. For example, it generates a 24-dimensional numerical vector representing population changes every hour. This population density data acquisition unit applies a natural language processing model to the location information of posts on social networking sites (SNS). It determines whether activity has occurred based on the post content. This determination result is then incorporated as a binary flag into a time-series vector. The unit performs linear interpolation on future population projection data. The interpolated predictions are stored in memory as a future demand tensor. These time-series vectors and the future demand tensor are then fed into a recurrent neural network. This network is used to predict communication traffic demand scores for each spatial grid. Through this predictive processing, the unit fundamentally improves upon previous evaluation methods based solely on static demographics. This enables dynamic and predictive demand assessment, thereby preventing over-investment in base stations. As a specific application area, the unit is used to address temporary traffic surges around large stadiums. It is also used in simulating phased base station additions accompanying new city development.

[0042] The radio wave propagation characteristic data acquisition unit runs on a server, acquiring 3D map data, building / tree / terrain obstacle data, communication quality logs of existing base stations, output data from radio wave propagation simulation tools, meteorological data (rainfall, snow cover, wind direction and speed, temperature, humidity, etc.), and obstacle data from moving bodies such as vehicles or railways. For example, it performs radio wave propagation simulations on a 3D map, taking into account the height and density of buildings or trees, calculating the radio wave attenuation rate, propagation loss, the impact of obstructions, and the probability of multipath or fading at various locations. Furthermore, it can also consider changes in propagation characteristics caused by meteorological conditions (e.g., increased attenuation during rainfall, reflection and absorption caused by snow cover, and obstruction changes caused by tree swaying in strong winds). As measured values ​​from existing base stations, it also acquires communication speed, connection stability, disconnection rate, and the number of complaints from users, enabling evaluations that reflect the actual situation on-site. Specifically, this radio wave propagation characteristic data acquisition unit has a dedicated processing processor for executing ray tracing algorithms. This unit constructs 3D map data and obstacle data as polygonal meshes in virtual space. The radio wave propagation characteristic data acquisition unit calculates the virtual radiation path of the radio wave within the virtual space. Based on material properties, it calculates the reflection coefficient and diffraction coefficient for each path. Based on the calculation results, it derives the received power intensity of each voxel within the space as a floating-point number. Meteorological data is input into the simulation model as environmental parameters. An attenuation coefficient corresponding to rainfall is dynamically added. Communication quality logs from existing base stations are used as reference data. The propagation model parameters are optimized using gradient descent to minimize the error between the simulation results and measured values. This optimization process continuously improves the simulation accuracy. The calculated radio wave attenuation rate or multipath probability is formatted into a multidimensional feature vector. This feature vector is then fed into the subsequent AI model. This radio wave propagation characteristic data acquisition unit eliminates the risk of communication quality degradation after setup by providing high-precision propagation prediction. This significantly improves the reliability of communication networks. As a specific application area, this unit is used in the utilization of complex millimeter-wave reflected waves in densely populated urban areas with high-rise buildings. It is also applied to predict changes in communication coverage during snowfall in blizzard-prone areas.

[0043] These diverse data are input into the generative AI processing unit on the server. This unit combines, for example, large-scale language models, deep learning models, image generation AI, time series analysis AI, and multimodal AI. The generative AI normalizes terrain data, population density data, and radio wave propagation characteristic data. For instance, it subtracts the normalized value of the terrain data from 1 when lower altitudes are considered more suitable for site selection; it directly uses the normalized value of the population density data when higher population density better predicts communication needs; and it subtracts the normalized value of the radio wave propagation characteristic data from 1 when less radio wave attenuation is considered higher in evaluation. Evaluation benchmarks are set for each element. Furthermore, the generative AI weights these elements, dynamically adjusting the weights of factors such as population density emphasis, terrain stability emphasis, and radio wave quality emphasis based on intended use or operational policies. The generative AI also imports past site selection performance or communication quality data, as well as feedback from on-site personnel, as learning data, continuously improving the accuracy of candidate site selection. For example, it can learn the communication quality or failure rate after deployment based on the operational performance of previously deployed base stations and reflect this in the extraction of candidate deployment sites under similar conditions. Furthermore, generative AI can automatically generate optimal evaluation logic based on multiple deployment objectives (e.g., expanding coverage in urban areas, eliminating dead zones in mountainous areas, temporary addition of event venues, etc.). Specifically, the generative AI processing unit is equipped with a multimodal neural network based on the Transformer architecture. As input data, the generative AI processing unit receives a three-dimensional tensor representing terrain data, a time-series vector representing population density data, and a feature vector representing radio wave propagation characteristics. The generative AI processing unit feeds these input data into dedicated encoder layers. The generative AI processing unit combines the latent representation vectors output from each encoder layer. The generative AI processing unit applies a self-attention mechanism to this combined vector. Through this self-attention mechanism, the generative AI processing unit extracts complex nonlinear relationships such as the interaction between terrain and radio wave propagation. The generative AI processing unit receives, for example, numerical data such as "sloping land at an altitude of 50 meters," "population inflow of 1000 people per hour," and "radio wave attenuation rate of 20% during rainfall." In a normalization layer, the unit scales these values ​​to a range from zero to one. As output data, the unit outputs a suitability score for each spatial coordinate as a floating-point number from zero to one hundred. It also outputs an importance mapping representing the basis of this score as an image tensor in heatmap format. The unit further outputs a recommendation text based on natural language. For example, it generates text such as, "Due to the high population density and minimal building obstruction at this location, the reach of millimeter waves is maximized." This output is then sent to a subsequent candidate site selection unit.During the learning phase, this generative AI processing unit uses past performance data as labels for correct answers. It calculates the mean squared error between predicted scores and actual communication quality metrics as a loss function. The unit uses backpropagation to update the network's weight parameters. Through this learning process, it establishes a data-driven evaluation benchmark that does not rely on human intuition. This ensures the objectivity and reproducibility of the evaluation. In specific applications, it is used to fully automate the deployment plans for thousands of base stations annually by nationwide telecommunications operators. It is also applied to generate optimal route configurations for mobile base station vehicles used in specific events.

[0044] The comprehensive score calculated by the generative AI processing unit is sent to the candidate site selection unit on the server. The candidate site selection unit flexibly extracts and sets candidate sites according to the application policy, such as the top 10 locations or all locations exceeding a predetermined threshold. For the extracted candidate sites, detailed data such as location coordinates, altitude, population density, radio wave propagation characteristics, comprehensive score, and surrounding environmental information are compiled and sent to the terminal side. Specifically, the candidate site selection unit has a computation module that executes ranking and clustering algorithms. The candidate site selection unit receives tens of thousands of comprehensive score arrays for each spatial coordinate from the generative AI processing unit. The candidate site selection unit performs a fast sorting of the score array in descending order. To prevent top-ranked locations from being geographically too close, the candidate site selection unit performs spatial clustering. Within each cluster, the candidate site selection unit selects only the representative location with the highest score. Through this spatial filtering, the candidate site selection unit minimizes the duplication of the base station's coverage area. For each extracted candidate site, the candidate site selection unit generates a structured data object in JSON format. The candidate site selection unit stores floating-point data of latitude and longitude, integer data of altitude, time-series arrays of population density, radar vectors of radio wave propagation characteristics, and a comprehensive score within the target site. This structured data is transmitted to the terminal via an encrypted communication protocol. Through this automatic extraction process, the candidate site selection unit eliminates the need for human intervention to visually narrow down the candidates from a large list. This dramatically reduces the selection process time from weeks to seconds. As a specific application area, the candidate site selection unit is used for efficient configuration decisions of base stations in rural areas. It is also used for optimal configuration planning of IoT sensor gateways in smart city projects.

[0045] AR and VR visualization units are installed on the terminal side. For example, when the on-site manager uses a tablet terminal to travel to the location, the location or recommendation level of the candidate site and various scores can be overlaid on the camera image of the terminal. For example, the AR visualization unit uses pins or 3D icons to display the location of the candidate site based on the actual local scenery. When each candidate site is clicked, detailed information such as altitude, population density, radio wave propagation characteristics, and comprehensive score pops up. In addition, multiple candidate sites can be compared and displayed simultaneously, allowing for intuitive selection of the best location. The VR visualization unit, for example, uses a goggle-type terminal to overlay terrain models or building models, population distribution heat maps, and radio wave propagation simulation results on a 3D virtual space. While the manager moves freely in the virtual space, they can confirm the surrounding environment or communication needs and radio wave quality of each candidate site from multiple angles. For example, the radio wave obstruction situation in high-rise building clusters in urban areas, radio wave dead zones in mountain valleys, and temporary population concentrations around the event venue can all be visualized in 3D. Specifically, the AR visualization unit processes RGB image tensors acquired from the terminal's camera and point cloud data acquired from the depth sensor in real time. The AR visualization unit executes a simultaneous localization and mapping (SLT) algorithm. Using this algorithm, the AR visualization unit calculates the terminal's absolute coordinates and attitude angles in 3D space with high precision. The AR visualization unit converts the latitude and longitude coordinates of candidate locations received from the server into the terminal's local coordinate system. Based on these converted coordinates, the AR visualization unit renders 3D polygonal objects at the corresponding pixel positions in the camera image. The AR visualization unit detects user clicks as 2D vectors of screen coordinates. The AR visualization unit performs a raycasting determination between this vector and the 3D object. If the determination is true, the AR visualization unit overlays a text layer displaying detailed information. The VR visualization unit uses a graphics processing unit to depict a high-resolution 3D virtual space. The VR visualization unit maps the results of radio wave propagation simulations as color information for each voxel. The VR visualization unit depicts the intensity of radio waves as a semi-transparent gradient from red to blue in space. The VR visualization unit detects user head movements using a gyroscope sensor. Based on the detection data, the VR visualization department updates the virtual camera's viewpoint at a rate of over 90 frames per second. Through these advanced rendering techniques, the system allows users to intuitively understand the spatial propagation of radio waves or the impact of physical obstacles that are impossible to grasp on a two-dimensional map. This significantly improves the accuracy of on-site decision-making. As a specific application area, this system has been explored for the placement of indoor antennas in complex indoor environments such as underground streets or large shopping malls. The system is also used for remote collaborative work where experts share VR spaces and issue instructions to on-site operators.

[0046] In addition, the on-site supervisor uploads supplementary information, local photos, and on-site observation data from the terminal to the generative AI processing unit on the server side. The generative AI analyzes and learns from this new data in real time, and can also update the evaluation or extraction results of candidate sites instantly. For example, the generative AI can flexibly import and reflect on site-specific information that cannot be grasped by past statistical data, such as newly discovered obstacles, seasonal population changes, or sudden events. Specifically, this system installs a site information collection module on the terminal side. This module acquires high-resolution image data captured by the terminal's camera as JPEG format byte arrays. The module also assigns location information and timestamps obtained from the GPS module as metadata to the image. The module sends this image file with metadata to the server via a RESTful API. The generative AI processing unit on the server side receives the image file and inputs the image data into a convolutional neural network. Using this network, the generative AI processing unit uses object detection algorithms to identify buildings under construction or trees that are new obstacles within a specific image. The generative AI processing unit outputs the bounding box coordinates and category labels of detected objects. Based on this output, it dynamically updates the corresponding voxels of the 3D map data on the server. Using the updated map data, it re-executes the radio wave propagation simulation. The recalculated radio wave attenuation rate is then used as a new feature to recalculate the overall score. Based on this recalculated score, the system immediately pushes a new list of candidate locations to the terminal. This real-time feedback loop eliminates the discrepancy between the static database and the dynamic real world. The system thus maintains an optimal setup plan based on the latest site conditions. As a specific application area, this system is used for agile network construction that tracks daily environmental changes in rapidly developing new residential areas. It is also applied to emergency response systems that import terrain changes after disasters from UAV imagery to establish optimal temporary communication routes.

[0047] In this embodiment of the invention, by centrally collecting, analyzing, and processing large-scale data on the server side, and achieving intuitive and interactive visualization and comparative analysis on the terminal side, the selection of wireless base station installation sites can be optimized efficiently and effectively. Generative AI goes beyond simple data analysis; it plays an indispensable role in automatically generating and optimizing the logic for extracting candidate installation sites by learning from various factors such as past installation performance, on-site information, and differences in application objectives. For example, it can automatically switch evaluation benchmarks between urban and mountainous areas, reflect on-site feedback, or continuously learn from post-installation performance to improve accuracy. Furthermore, the AR / VR visualization function on the terminal side makes three-dimensional, intuitive comparative analysis, previously difficult in map or spreadsheet software, possible, simultaneously achieving efficient installation planning, cost reduction, uniform communication quality, and rapid decision-making. Specifically, this system constructs a hybrid processing architecture that tightly integrates the cloud computing environment with edge devices. This system uses server-side tensor processing units to execute inferences from deep learning models with hundreds of millions of parameters at high speed. This system utilizes mobile GPUs on the terminal side to perform 3D rendering of millions of polygons per second. Through this role-sharing, the system minimizes network bandwidth consumption while providing users with lag-free, interactive operability. The generative AI processing unit incorporates reinforcement learning algorithms. This unit defines the configured communication quality metrics as a reward function. It autonomously adjusts the hyperparameters of the extraction logic to maximize this reward. This autonomous learning mechanism completely replaces the manual parameter tuning work performed by human engineers. This technical architecture maximizes the efficiency of computing resource utilization. The system also mathematically solves complex multi-objective optimization problems as gradient search in a multi-dimensional space. This elevates the selection of services based on past heuristics and rules of thumb to a fully data-driven and reproducible engineering process. As a specific application area, this system is utilized by global telecommunications operators when building standardized base station deployment platforms suitable for different geographical conditions or legal regulations in various countries. This system has also been applied to large-scale projects that optimize the low-latency communication infrastructure for autonomous vehicles along the national road network.

[0048] Furthermore, for example, in selecting candidate sites, generative AI can automatically propose site selection plans that minimize communication disruptions during disasters, taking into account factors such as the location of evacuation routes or shelters, emergency vehicle routes, and the location of critical infrastructure. Alternatively, by importing future population dynamics forecasts, urban development plans, and new transportation infrastructure plans, long-term site optimization with a focus on the next few years is possible. Additionally, importing site selection or maintenance costs, power supply status, landowner information, and relevant laws, regulations, and ordinances allows for the extraction of candidate sites while conducting comprehensive cost and risk assessments. Specifically, this system connects to external disaster prevention information databases and urban planning databases via a dedicated API gateway. The system acquires the location coordinates of shelters or emergency vehicle route information as graph-structured data. This graph data is then input into a graph neural network. The system uses this network to identify important nodes in a specific network topology. The system assigns a new feature vector, called a disaster importance score, to the spatial grid surrounding each important node. This system also acquires land price data, construction unit price data, and commercial power supply distance data for each candidate site as a numerical matrix. Using this matrix, the system calculates the initial investment cost and operating cost for each candidate site. In its generative AI processing unit, the system executes a multi-objective optimization algorithm with inverse objective functions: maximizing communication quality and minimizing total cost. This generative AI processing unit calculates multiple optimal solutions on the Pareto boundary. The system outputs these results as multiple configuration schemes. For example, it generates results based on different parameter sets such as "cost-focused scheme," "disaster-resistance-focused scheme," and "future-scalability-focused scheme." Through this multi-faceted evaluation process, the system automates the formulation of comprehensive infrastructure investment plans on a computer, going beyond purely technical radio wave simulations. This powerfully supports management decision-making. As a specific application area, the system is used for robust communication infrastructure design in smart cities built collaboratively by autonomous regions and telecommunications operators. It is also applied to optimal universal service delivery plans within limited budgets in sparsely populated areas.

[0049] Furthermore, for example, in situations where multiple telecom operators jointly deploy base stations, the generative AI also considers the communication methods or frequency bands of each operator, interference risks with existing equipment, interoperability, etc., and can automatically generate the optimal shared deployment scheme. In addition, during the post-deployment phase, the generative AI continuously learns from communication quality logs or user feedback, and can automatically re-evaluate the deployment site, add or remove deployment proposals as needed. Specifically, this system adopts a multi-tenant data management architecture. The system stores the frequency allocation information of each telecom operator, the antenna directivity pattern, and the power transmission settings as separate configuration file data in a secure database. The generative AI processing unit reads this configuration file data. The generative AI processing unit simulates intermodulation distortion or adjacent channel interference between different frequency bands at the physical layer level. The generative AI processing unit incorporates the interference simulation results as a penalty score into the comprehensive evaluation function. The generative AI processing unit uses a genetic algorithm to explore the optimal combination of height or angle when configuring antennas from multiple operators on the same tower to minimize this function. During the operational phase, this system accumulates performance metrics such as throughput, latency, and packet loss rate from periodically transmitted data from each base station into a time-series database. The generative AI processing unit applies anomaly detection algorithms to this time-series data. When the system detects a consistent downward trend in communication quality, it outputs an alert signal recommending the addition of more base stations to the area. Through this continuous cycle of monitoring and re-evaluation, the system achieves autonomous optimization throughout the entire network lifecycle. This significantly reduces maintenance and operational costs. As a specific application area, this system is used by infrastructure-sharing operators to provide optimal tower leasing plans to multiple telecommunications operators. It is also used in systems that automatically plan the consolidation and decommissioning of existing base stations in response to changes in the radio wave environment accompanying urban redevelopment.

[0050] In this way, the embodiments of the present invention possess a highly complex and multi-faceted structure, including server and terminal collaborative operation, multi-layered utilization of generative AI, comprehensive analysis of massive and diverse data, three-dimensional and intuitive visualization, real-time feedback from the field, comprehensive future prediction or cost evaluation, support for multiple operators, and continuous optimization during the application phase. It is a system capable of flexibly and effectively addressing all challenges in wireless base station setup planning. Specifically, this system manages these multiple technical components as a loosely coupled group of microservices on an orchestration platform. The system employs the gRPC protocol for data communication between modules, thereby minimizing the overhead of inter-process communication. The system uses asynchronous task queues to schedule the generative AI inference pipeline, thus maintaining stable throughput even under bursts of computational requests. The system uses a cross-platform game engine to develop terminal-side applications, providing the same high level of three-dimensional visualization capabilities across various operating systems and hardware architectures. Through this comprehensive system architecture, the system fully digitizes the entire process from data collection to analysis, visualization, decision-making, and application maintenance. This fundamentally transforms previously siloed business processes. In its specific application area, this system is utilized as an autonomous deployment and management platform for ultra-high-density terahertz wave base stations for next-generation 6G communication networks. It is also used for optimizing and simulating global communication coverage by integrating low-Earth orbit satellite constellations and terrestrial networks.

[0051] (System Composition) The system described in this embodiment includes a terrain data acquisition unit, a population density data acquisition unit, a radio wave propagation characteristic data acquisition unit, a generative AI processing unit, a candidate site selection unit, an AR visualization unit, a VR visualization unit, a data collaboration unit, a user interface unit, and a field information collection unit. The terrain data acquisition unit acquires terrain data from various information sources, such as geographic information systems, external map databases, satellite imagery services, UAV aerial photography data, and publicly available data from autonomous regions. For example, it can change the resolution to acquire elevation grid data in 1-meter, 5-meter, and 10-meter units, and can differentiate the use based on the accuracy requirements of the candidate sites. Furthermore, it can acquire tilt angle data, classification information for land use classification such as residential areas, commercial areas, industrial areas, farmland, forests, and rivers, foundation strength data, soil classification data, and flood and debris flow disaster risk data. For example, it can also incorporate automatic exclusion of areas with weak foundations or areas with a history of past disasters. The Population Density Data Acquisition Department gathers a variety of population and pedestrian flow data, including census data, basic resident registers of municipalities, base station connection logs from portable devices, traffic flow sensor data, social media post location information, Wi-Fi access point connection data, foot traffic data from commercial facilities, and attendance data from event organizers. This allows them to track population density changes over various time periods and conditions, such as weekdays and weekends, day and night, seasons, and event times, and reflect these changes in the evaluation of candidate sites. Furthermore, they acquire population projection data for future urban development plans or redevelopment areas, enabling them to develop site selection plans that anticipate future demand. The radio wave propagation characteristic data acquisition unit acquires 3D map data, building / tree / terrain obstacle data, communication quality logs of existing base stations, output data from radio wave propagation simulation tools, meteorological data such as rainfall, snow accumulation, wind direction and speed, temperature, and humidity, and data on obstacles such as vehicles or railways. For example, it performs radio wave propagation simulations on 3D maps that take into account the height and density of buildings or trees, calculating the radio wave attenuation rate, propagation loss, the impact of obstructions, and the probability of multipath or fading at various locations. Furthermore, it can consider changes in propagation characteristics caused by meteorological conditions, such as increased attenuation during rainfall, reflection and absorption caused by snow accumulation, and obstruction changes caused by tree swaying in strong winds. It also acquires measured values ​​from existing base stations, including communication speed, connection stability, disconnection rate, and the number of user complaints, enabling evaluations that reflect the actual situation on-site.The Generative AI Processing Department receives terrain data, population density data, and radio wave propagation characteristic data as input. It combines large-scale language models, deep learning models, image generation AI, time series analysis AI, and multimodal AI, among others. It normalizes each data point based on minimum and maximum values. For example, it subtracts the normalized value of the terrain data from 1 when lower altitudes are considered more suitable for site selection; it directly uses the normalized value of the population density data when higher population density allows for better prediction of communication needs; and it subtracts the normalized value of the radio wave propagation characteristic data from 1 when less radio wave attenuation is considered higher in evaluation. It sets evaluation benchmarks for each element and then weights these elements. The weights of factors such as population density emphasis, terrain stability emphasis, and radio wave quality emphasis are dynamically adjusted according to the intended use or application strategy. Past site selection performance or communication quality data, as well as feedback from on-site personnel, are also imported as learning data to continuously improve the accuracy of extracting candidate sites. For example, it can learn from the operational performance of previously deployed base stations, assess communication quality or failure rates after deployment, and reflect this in the extraction of candidate deployment sites under similar conditions. Alternatively, it can automatically generate optimal evaluation logic based on multiple deployment objectives, such as expanding coverage in urban areas, eliminating dead zones in mountainous regions, or temporarily adding venues for events. Specific examples of prompts for the generative AI processing unit could include: "Comprehensively analyze the following terrain data, population density data, and radio wave propagation characteristic data. Calculate a comprehensive score based on the principle that lower altitude, higher population density, and less radio wave attenuation result in a higher evaluation. Extract the top 10 locations as candidate deployment sites. Output coordinates, altitude, population density, radio wave propagation characteristics, and comprehensive score for each location." or "Considering past deployment performance data and feedback from on-site personnel, prioritize areas where future population growth is anticipated." The candidate site selection unit is based on a comprehensive score calculated by the generative AI processing unit. For example, it flexibly extracts the top 10 locations based on the application strategy, or all locations exceeding a predetermined threshold, as candidate sites. It compiles detailed data such as coordinates, altitude, population density, radio wave propagation characteristics, comprehensive score, and surrounding environmental information for each location and sends it to the terminal. For instance, it can also consider the radio wave obstruction in urban high-rise buildings, radio wave dead zones in mountain valleys, and temporary population concentrations around event venues, taking into account the specific circumstances of each location. The AR visualization unit overlays the location or recommendation level and various scores of the candidate sites onto the terminal's camera image. The location of the candidate sites is displayed relative to the actual local scenery using pins or 3D icons. Clicking on a candidate site displays detailed information such as altitude, population density, radio wave propagation characteristics, and comprehensive score. Simultaneously comparing multiple candidate sites allows for intuitive selection of the best location.For example, when a site manager uses a tablet terminal to travel to the location, they can simultaneously confirm the location relationships of local buildings, roads, and obstacles in real time while comparing and discussing candidate sites. The VR visualization department uses goggle-type terminals to overlay terrain or building models, population distribution heatmaps, and radio wave propagation simulation results onto a 3D virtual space. The manager can move freely within the virtual space while simultaneously confirming the surrounding environment, communication requirements, and radio wave quality of each candidate site from multiple angles. For example, the radio wave obstruction in high-rise building clusters in urban areas, radio wave dead zones in mountain valleys, and temporary population concentrations around event venues can all be visualized in 3D. The data collaboration department is responsible for the real-time transmission and reception of terrain data, population density data, radio wave propagation characteristic data, candidate site information, site information, and visualization data between the server and terminals, ensuring data consistency and timeliness. For instance, when a site manager uploads local photos or observation data from the terminal, the generative AI processing department on the server side analyzes the data instantly, updating the evaluation or extraction results of candidate sites in real time. The user interface unit provides a screen on the terminal for the person in charge to view detailed information on candidate sites, compare and discuss them, input feedback, or add information. For example, it enables various interactions such as displaying detailed information for each candidate site, parallel comparison of multiple candidate sites, attaching on-site photos, and inputting evaluations / comments on candidate sites. The on-site information collection unit uses the terminal's camera, GPS, various sensors, and user input to collect the latest local information or additional data and send it to the server. Examples include collecting photos of newly discovered obstacles, observational data on seasonal population changes, information on sudden events, and subjective evaluations or comments from the on-site person in charge. Generative AI analyzes and learns this new data in real time, enabling it to update the evaluation or extraction results of candidate sites instantly. In this way, the system described in this embodiment, through the collaborative operation of each unit, efficiently and highly optimizes the selection of wireless base station installation sites, flexibly addressing various issues such as actual on-site conditions or future predictions, cost / risk assessment, support for multiple operators, and continuous optimization during the operational phase. Specifically, this system interconnects these ten functional modules via a system bus and shared memory space. This system uses a distributed message queue as a data collaboration unit. This system uses a queue to asynchronously publish and subscribe to event messages across modules. The system provides a web application based on a reactive programming paradigm as the user interface. Through this interface, user actions are reflected immediately in the state management store. In the generative AI processing unit, the system receives "altitude data tensor," "population density time series matrix," and "3D voxel data of radio wave propagation simulation results" as input data. This generative AI processing unit feeds this data into a neural network with a multi-head attention mechanism.The generative AI processing unit outputs the following data for each coordinate: "Recommendation Score (floating-point number from zero to one)," "Predicted Communication Coverage Radius (value in meters)," and "Category Code of the Main Recommendation Reason." This output is then fed into the threshold determination circuit of the candidate location selection unit. The system filters out coordinates with scores exceeding a predetermined threshold and forwards them to the subsequent visualization unit. Through this series of data streams, the system automatically extracts meaningful features from massive input data, deriving the optimal solution with speed and accuracy exceeding the limits of human cognition. This system thus maximizes information processing efficiency. As a specific application area, this system is provided as a cloud-based SaaS platform for the design of fully automated communication infrastructure. It is also used in the optimal configuration planning of data collection concentrators for smart meters in smart grids.

[0052] (Implementation steps) Step 1: Terrain Data Collection The server-side terrain data acquisition department collects elevation grid data, tilt angle data, land use classification data, foundation strength data, soil classification data, and flood and debris flow hazard risk data from various information sources, including Geographic Information Systems (GIS), external map databases, satellite imagery services, UAV aerial photography data, and publicly available data from local governments. For example, high-resolution elevation data is used in urban areas, while tilt or foundation strength data is acquired in mountainous regions. Specifically, in this step, the system sends HTTP requests to multiple external data sources in parallel. The system temporarily stores the acquired raw data in GeoJSON or TIFF format in a memory buffer. Based on a spatial reference system, the system reprojects this data onto a common geodesic coordinate system. The system reconstructs the reprojected data as a three-dimensional numerical tensor in memory. For each element of this tensor, the system stores the elevation, tilt angle, and foundation strength values ​​as vectors representing channel directions. By unifying this data structure, the system standardizes the input format for subsequent machine learning models. This reduces the computational overhead of data preprocessing. This system, as a specific application area, is used in the foundational data collection process for constructing digital twins of the entire national territory. It is also applied in the initial stages of a pipeline for generating high-precision 3D maps for autonomous driving.

[0053] Step 2: Collection of population density and pedestrian flow data The server-side population density data acquisition department gathers census data, basic resident register data from municipalities, base station connection logs from portable terminals, traffic flow sensor data, SNS post location information, Wi-Fi access point connection data, store visitor data from commercial facilities, and attendance data from event organizers, etc., to grasp the changes in population density or pedestrian flow in each time period or situation, such as weekdays and weekends, day and night, seasons, and event times. It also acquires population forecast data for future urban development plans or redevelopment areas, and addresses long-term planning. Specifically, in this step, the system initiates a streaming data processing framework. The system parses location information logs, which are inflows of tens of thousands of transactions per second, in real time. The system uses a spatial hash function to map the parsed data to specific geographic grid IDs. For each grid ID, the system counts the number of unique users for each time range. The system writes this count result as a time-series numerical array into the database. For future forecast data, the system applies an autoregressive moving average model to imputate missing values. Through this dynamic data statistical processing, the system quantifies subtle changes in pedestrian flow that static statistical data cannot fully capture. This system thus accurately models the spatiotemporal deviations of communication needs. As a specific application area, this system is used to track the movement trajectories of evacuees during large-scale disasters in real time and dynamically allocate communication resources. This system is also applied in smart retail to predict the number of potential customers around stores.

[0054] Step 3: Collection of radio wave propagation characteristic data The server-side radio wave propagation characteristic data acquisition unit acquires 3D map data, building / tree / terrain obstacle data, communication quality logs of existing base stations, output data from radio wave propagation simulation tools, meteorological data (rainfall, snow cover, wind direction and speed, temperature, humidity, etc.), and data on obstacles such as vehicles or railways. For example, it can gain a detailed understanding of the impact of special terrain or obstacles such as high-rise building clusters in urban areas, valleys in mountainous regions, or the perimeter of event venues. Specifically, in this step, the system drives a parallelized ray tracing engine on a GPU cluster. The system converts the 3D map data into a hierarchical bounding box structure and loads it into memory. The system radiates millions of rays from a virtual transmitting antenna. The system uses hardware acceleration to quickly calculate the intersection determination of each ray with the polygon. The system calculates the intensity of the reflected and diffracted waves at the intersection points based on physical laws. The system outputs the calculation results as a floating-point matrix of the received power in each voxel within space. Through this physics-based simulation, the system replaces on-site radio wave measurement work with computation on a computer. This system thereby reduces the significant cost and time spent on measurement. This system has been used, in specific applications, to model the complex metallic reflection environment in the construction of a private 5G network within a factory. It has also been applied to simulate overhead radio interference in communication networks between drones.

[0055] Step 4: Data Synthesis and Normalization As preprocessing for input into the generative AI processing unit, terrain data, population density data, and radio wave propagation characteristic data acquired on the server side are normalized based on minimum and maximum values, enabling comparison of data at different units or scales using the same benchmark. For example, transformations are performed according to evaluation benchmarks, such as subtracting the normalized value from 1 for altitude data, directly using the normalized value for population density data, and subtracting the normalized value from 1 for radio wave attenuation data. Specifically, in this step, the system executes a data preprocessing pipeline. The system scans and determines the minimum and maximum values ​​for each dimension of each data tensor. The system uses the determined values ​​to apply the min-max scaling algorithm. The system converts all values ​​into single-precision floating-point numbers ranging from zero to one. For altitude data, the system performs a uniform "one minus normalized value" operation on a vector basis, as a value inversion process. The system applies Z-score-based pruning to outliers, eliminating the influence of noise. This rigorous normalization process prevents gradient vanishing or gradient explosion during subsequent neural network learning. This improves the convergence speed and inference accuracy of the AI ​​model. In its specific application area, this system is used as a preprocessing module for all IoT platforms that integrate environmental data of different scales acquired from heterogeneous sensors and input it into machine learning models. This system is also applied to calculate risk scores by integrating economic indicators from different units in the financial market.

[0056] Step 5: Setting up candidate site scoring based on generative AI The generative AI processing unit receives normalized terrain data, population density data, and radio wave propagation characteristic data as input, weights each element, and calculates a comprehensive score. The generative AI also imports past setup performance data or feedback from on-site managers, future population growth predictions, disaster risk information, and cost information as learning data to continuously improve the accuracy of candidate site extraction. For example, it dynamically adjusts weights based on the intended use or application strategy, such as expanding urban coverage, eliminating blind spots in mountainous areas, or temporarily adding event venues. Specific examples of prompts for the generative AI to read could include: "Comprehensively analyze the following terrain data, population density data, and radio wave propagation characteristic data, calculating a comprehensive score based on the principle that lower altitude, higher population density, and less radio wave attenuation result in higher evaluations, and extracting the top 10 locations as candidate sites. Output coordinates, altitude, population density, radio wave propagation characteristics, and comprehensive score for each location." or "Considering past setup performance data and feedback from on-site managers, prioritize extracting areas where future population growth is predicted." Specifically, in this step, the system executes a hybrid AI architecture integrating a large-scale language model and a graph convolutional network. The system receives a normalized 3D tensor group and a natural language prompt string as input data. It then converts the prompt string into an array of integer tag IDs using a tokenizer. This tag array is fed into the embedding layer to generate a context vector. This context vector is used as a query for the attention mechanism. The system performs cross-attention operations using feature maps extracted from the tensor group as keys and values. Through this operation, the system dynamically changes the weight parameters of each feature based on the prompt's content (e.g., "prioritize future population growth"). As output data, the system outputs a probability distribution of the comprehensive score for each spatial grid. It also outputs a feature contribution vector representing the basis for calculating this score. This output is then passed to the subsequent ranking processing module. Through this dynamic weighting mechanism, the system flexibly adapts to various business requirements that fixed algorithms cannot handle. Thus, this system covers countless use cases with a single system. In specific applications, this system is used in the logistics industry for optimal configuration simulation of distribution centers, enabling dynamic evaluation of network locations based on seasonal variations or changes in fuel costs. It is also applied in the selection of wind turbine installation sites in the renewable energy sector, incorporating a score that integrates wind condition data and environmental regulations.

[0057] Step 6: Set up candidate site extraction and detailed data generation The candidate site selection unit extracts locations with high scores or those exceeding a predetermined threshold as candidate sites based on the comprehensive score calculated by the generative AI processing unit. For each candidate site, detailed data such as coordinates, altitude, population density, radio wave propagation characteristics, comprehensive score, surrounding environmental information, disaster risk, and cost information are compiled and sent to the terminal. For example, it can also extract data considering the radio wave obstruction in high-rise building clusters in urban areas, radio wave dead zones in mountain valleys, and temporary population concentrations around event venues, considering the specific circumstances of each site. Specifically, in this step, the system executes a Top-K extraction algorithm using a priority queue data structure. The system scans the comprehensive scores across millions of grids, inserting only elements with scores exceeding a threshold into the queue. The system retrieves the top K elements from the queue. The system inversely calculates the original latitude and longitude coordinates from the indexes of each retrieved element. For each coordinate, the system uses SQL queries to retrieve text and numerical data from a relational database, combining surrounding environmental information or cost information. The system serializes this combined data as a JSON payload. This system sends the payload to the terminal's IP address via HTTPS. Through efficient extraction and serialization processing, it instantly extracts only the necessary information from massive computational results. This allows for high-speed data transmission to the terminal while minimizing network bandwidth consumption. In specific applications, this system is used in real estate technology as a high-speed retrieval and detailed data generation API for properties meeting customer requirements. It is also used in ride-hailing applications to provide drivers with real-time updates on optimal waiting locations.

[0058] Step 7: Discussion on 3D visualization and on-site comparison based on AR / VR The AR and VR visualization departments on the terminal side use detailed data of candidate sites sent from the server to overlay the location, recommendation level, and various scores of the candidate sites onto the camera footage or 3D virtual space on-site. On-site supervisors can confirm the location of candidate sites using pins or 3D icons within the actual scenery or virtual space. By clicking or selecting each candidate site, they can view detailed information such as altitude, population density, radio wave propagation characteristics, and overall score in pop-up displays. Multiple candidate sites can be compared and displayed simultaneously, allowing for intuitive selection of the best location. Specifically, in this step, the system drives the graphics pipeline on the terminal side. The system deserializes the received JSON payload and expands the coordinate data of each candidate site into the terminal's local memory. In AR mode, the system applies an optical flow algorithm to track feature points from the 60 frames per second image stream acquired from the terminal's camera. Based on this feature point group, the system detects a plane in the real space and uses alpha blending to render 3D cylindrical objects representing candidate sites on this plane in a semi-transparent manner. In VR mode, this system performs stereoscopic rendering, generating two images with parallax for the left and right eyes. The system detects raycast input from the user's controller. When a ray intersects an object, the system renders a 2D UI panel containing detailed information as a billboard in 3D space. Through advanced spatial computing technology, the system converts numerical data into spatial information that can be directly processed by the human visual cortex. This significantly reduces the cognitive load on field operators and eliminates the risk of setup errors caused by misidentification. Specifically, this system is used in AR verification systems for underground piping routes in construction sites. It is also applicable in the medical field as a surgical support system that overlays patient MRI data onto the affected area.

[0059] Step 8: Collection and feedback of on-site information The on-site supervisor uses the terminal's camera, GPS, various sensors, and user input to collect the latest or additional information from the site and send it to the server. For example, this can include photos of newly discovered obstacles, observational data on seasonal population changes, information on unexpected events, and subjective evaluations or comments from the on-site supervisor. This new data is input into the generative AI processing unit in real time, updating the evaluation or extraction results of candidate sites instantly. A specific example of a prompt for the generative AI to read could be, "Considering the additional information obtained on-site, recalculate the evaluation of candidate sites. Reflecting new obstacles, population changes, or event information, re-extract the best candidate sites." Specifically, this system utilizes an asynchronous communication protocol between edge devices and the cloud server. The on-site information collection unit preprocesses the RGB image tensors acquired from the terminal's camera. This preprocessing unit normalizes the image resolution and applies a noise reduction filter. The system encodes the latitude and longitude coordinates obtained from GPS into floating-point vectors. The system formats environmental data obtained from various sensors into time-series arrays. This generative AI processing unit employs a multimodal architecture incorporating a large-scale language model and convolutional neural networks. As input data, it receives image tensors, numerical vectors, and embedding vectors from natural language text. For example, the input data includes a 3-channel 224×224 pixel image tensor, two-dimensional coordinate vectors, and a 512-dimensional text embedding vector. The AI ​​processing unit projects this heterogeneous data into a common high-dimensional feature space. It uses a self-attention mechanism to calculate the correlations between features. As output, it generates updated suitability scores for each candidate location as continuous values ​​from 0.0 to 1.0. For example, the output data includes numerical data such as a score of 0.85 for candidate location A and a score of 0.62 for candidate location B. The AI ​​processing unit also outputs structured text data representing the basis for re-evaluation. A subsequent processing unit compares the output scores with a predetermined threshold. This subsequent processing unit only extracts candidate locations exceeding the threshold and updates the database. Through this real-time recalculation, the system improves its adaptability to dynamic changes in the environment. This system transforms human intuition into quantitative parameters and integrates them into an objective evaluation process performed by a computer. This eliminates the need for manual re-evaluation, dramatically increasing processing speed. The system is applicable not only to the deployment of communication infrastructure but also to various use cases such as the rapid deployment of temporary base stations during disasters. It employs backpropagation to periodically fine-tune the model based on feedback data from the field. The system updates weight parameters in a manner that minimizes the cross-entropy error function.

[0060] Step 9: Finalize and continuously optimize the plan Based on the results of comparisons and discussions within the AR / VR visualization department or on-site information, the on-site manager or supervisor ultimately determines the optimal deployment candidate location. After deployment, communication quality logs, user feedback, and operational cost data are continuously collected on the server side. Through a generative AI processing unit, these operational performance metrics are continuously learned, enabling automatic future deployment planning, re-evaluation of existing base stations, addition of new base stations, and withdrawal of proposals. Specifically, this system implements a continuous learning pipeline. The data collection unit periodically obtains communication quality logs from the deployed base stations. These logs are structured as time-series tensors containing numerical data such as throughput, latency, and packet loss rate. The system processes operational cost data as a normalized floating-point array. The generative AI processing unit possesses a hybrid model combining recurrent neural networks and reinforcement learning algorithms. As input data, the AI ​​processing unit receives past prediction scores, actual communication quality tensors, and operational cost vectors. For example, the input data includes a one-dimensional time-series array representing 24-hour communication traffic and a three-dimensional cost vector. The AI ​​processing unit calculates the discrepancy between predicted and actual performance values. The AI ​​processing unit inputs the deviation as a reward signal into the reinforcement learning agent. As output, the AI ​​processing unit generates classification labels representing the continued use, parameter adjustment, additional settings, or removal of existing base stations. For example, the output data includes a "parameter adjustment" label for base station X and a "removal" label for base station Y. The AI ​​processing unit outputs a probability distribution based on the confidence score for each proposal. The follow-up processing unit automatically sends proposals with confidence scores above a set value to the operation management system. The follow-up processing unit generates an alert to the administrator when the confidence score is low. This system continuously improves the model's prediction accuracy through this feedback loop. This system prevents performance degradation due to environmental changes, achieving long-term optimization of computing resources. This system enables data-driven infrastructure management that does not rely on human rules of thumb. This system is applicable to a wide range of fields, such as energy grid optimization in smart cities or dynamic reconfiguration of traffic control systems. This system uses online learning techniques, updating the model weights sequentially whenever new data arrives. This system uses gradient descent to optimize the loss function.

[0061] (specific use case) For example, when developing a plan to install new wireless base stations in an urban area, the terrain data acquisition department first obtains elevation grid data from a geographic information system and collects land use classification data or foundation strength data from the municipality's public database. The population density data acquisition department obtains census data or base station connection logs from portable terminals, traffic flow sensor data, and data on the number of people entering commercial facilities, tracking population density or pedestrian flow changes for each time period, such as weekdays and weekends, day and night, and during events. The radio wave propagation characteristics data acquisition department obtains 3D map data or building / tree distribution data, communication quality logs from existing base stations, output data from radio wave propagation simulation tools, and meteorological data such as rainfall or snow cover, wind direction, and wind speed, to understand in detail the impact of special terrain or obstacles such as high-rise buildings, underground streets, and complex road networks in the urban area. This data is integrated on the server side and normalized by the generative AI processing department. The generative AI processing department weights each element, awarding higher scores based on lower elevation, higher population density, and less radio wave attenuation, and calculates a comprehensive score. For example, when prioritizing the expansion of urban coverage, a high weight is set for population density, while the weight of terrain or radio wave propagation characteristics is adjusted. Generative AI also imports past site performance data, feedback from on-site managers, future population growth predictions, disaster risk information, and site cost information as learning data to improve the accuracy of candidate site extraction. The candidate site selection unit, based on the comprehensive score calculated by the generative AI processing unit, extracts locations with high scores or exceeding predetermined thresholds as site candidates. Detailed data such as coordinates, altitude, population density, radio wave propagation characteristics, comprehensive score, surrounding environmental information, disaster risk, and cost information for each candidate site are compiled and sent to the terminal. The AR visualization unit on the terminal overlays the location or recommendation level and various scores of the site candidate sites onto the on-site camera footage. On-site managers can simultaneously confirm the positional relationships of buildings, roads, and obstacles while comparing and discussing the site candidate sites. The VR visualization department overlays terrain or building models, population distribution heat maps, and radio wave propagation simulation results onto a 3D virtual space. The person in charge can move freely within the virtual space while simultaneously assessing the surrounding environment, communication needs, and radio wave quality of each candidate site from multiple angles. On-site supervisors use cameras, GPS, various sensors, and user input to collect the latest or additional data from the site and send it to the server. For example, this includes collecting photos of newly discovered obstacles, observational data on seasonal population changes, information on unexpected events, and subjective evaluations or comments from the on-site supervisor. This new data is input into the generative AI processing department in real time, instantly updating the evaluation of candidate sites and extracting results.After setup, the system continuously collects communication quality logs, user feedback, and operational cost data on the server side. Through generative AI processing, it continuously learns from these performance metrics, automatically enabling future setup plans, re-evaluation of existing base stations, addition of new locations, and removal of proposals. Specific examples of prompts for the generative AI include: "Comprehensively analyze the following terrain data, population density data, and radio wave propagation characteristic data. Calculate a comprehensive score based on the principle that lower altitude, higher population density, and less radio wave attenuation result in higher evaluations. Extract the top 10 locations as deployment candidates. Output coordinates, altitude, population density, radio wave propagation characteristics, and comprehensive score for each location." or "Considering additional information obtained on-site, recalculate the evaluation of deployment candidates. Reflecting new obstacles, population changes, or event information, re-extract the best deployment candidates." Specifically, the system includes a data integration module that synchronizes multiple data streams. This module aligns time-series data with different sampling rates on the time axis through interpolation. The system maps spatial data to a three-dimensional voxel grid based on a common geodesic system. The generative AI processing unit is equipped with a deep learning model based on the Transformer architecture. As input data, this unit receives multi-dimensional feature vectors corresponding to each voxel. For example, the input data includes a scalar value representing altitude, a 24-dimensional vector representing 24-hour population variation, and a tensor representing radio wave attenuation rate. The unit uses a multilayer perceptron to map these feature vectors into a high-dimensional space. The unit calculates the interactions between spatially separated voxels using a self-attention mechanism. As output, the unit generates a probability distribution representing the suitability of each voxel's location. For example, the output data includes a suitability score of 0.92 at a specific coordinate and the variance of the confidence interval for that score. The unit also outputs the contribution of the features used to calculate the score as attention weights. A non-maximum suppression algorithm is applied to the output probability distribution. This subsequent processing unit eliminates spatially close, redundant candidate locations and extracts an independent group of optimal candidate locations. Through spatial analysis at this altitude, this system discovers locally optimal locations that were missed in previous methods. This system leverages parallel processing by computers to significantly shorten the parsing time for wide-area urban data. It is widely applicable to spatially constrained infrastructure configuration problems such as charging station deployment for autonomous vehicles or delivery port selection for drones. In model learning, the system employs a contrastive learning approach, using past successful setup examples as positive examples and failed examples as negative examples. By minimizing the interval ranking loss function, the system maximizes the separation between positive and negative examples in the feature space. As a data augmentation technique, a small amount of Gaussian noise is added to the input data to improve the model's robustness.

[0062] (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".

[0063] (Implementation methods for carrying out the present invention) The embodiments used to implement this invention will be described in more specific and detailed, and with arbitrarily increased thickness. This embodiment, in order to support the optimal configuration of urban infrastructure, has the following structure: Multiple data are collected and integrated in units of a grid that subdivides the entire urban area; generative AI is used to extract optimal candidate sites for the installation; and the results are visually displayed, compared, and discussed using 3D visualization technologies such as AR or VR. Specifically, this system uses a spatial segmentation algorithm to convert the entire urban area into a hierarchical grid structure. This system employs a quadtree data structure to segment high-density data areas with finer resolution. The data collection unit assigns a unique identifier to each grid cell. This system uses a distributed database to manage the attribute data of each cell in parallel. The generative AI processing unit has an inference engine based on a graph neural network. The AI ​​processing unit receives graph structure data as input data, with each grid cell as a node and adjacency relationships as edges. For example, the input data includes a feature matrix for each node and an adjacency matrix representing the connection relationships of the edges. The AI ​​processing unit uses a message passing algorithm to propagate feature information between adjacent nodes. The AI ​​processing unit outputs latent representation vectors for each node and a final suitability score. For example, the output data includes a 64-dimensional embedding vector for each grid cell and a score normalized to 0 to 100. The subsequent processing unit clusters the grid cells based on the output scores. This subsequent processing unit identifies densely packed high-scoring cells as macroscopic candidate regions. This system accurately models spatial continuity or network effects between infrastructures through graph-based analysis. This system is not limited to evaluating single cells but achieves global optimization considering the topology of the entire city. This system thus saves computational resources and significantly improves the processing efficiency for large-scale urban data. This system is applicable to network-structured topics such as route design for public transportation or configuration optimization of wide-area surveillance camera networks. This system uses unsupervised learning techniques to extract hidden patterns within the graph structure. This system uses a graph autoencoder to optimize the network weights in a way that minimizes the reconstruction error of the input graph.

[0064] The Urban Infrastructure Data Acquisition Department divides the entire city into grids with arbitrary resolutions such as 10 meters, 25 meters, 50 meters, and 100 meters, collecting various data for each grid. For example, elevation data is acquired from geographic information systems or satellite imagery, drone aerial photography data, topographic laser scanning data, and foundation subsidence observation data, used for assessing flood risk, foundation subsidence risk, landslide risk, and tsunami inundation risk. For example, detailed elevation or foundation information is acquired for low-lying areas, riverbanks, reclaimed land, and sloping areas within the city, and utilized in disaster risk assessment. Population density data is acquired from census data, basic resident registers, location information from portable devices, traffic flow sensors, location information from social media posts, data on the number of people entering commercial facilities, attendance data from event organizers, Wi-Fi connection data, and public transportation usage data, also capturing population fluctuations or peak traffic flows on weekdays and weekends, day and night, during events, and in each season. For example, the system meticulously captures population dynamics for each time period or situation, such as daytime high population density in office districts, nighttime high population density in residential areas, and rapid population increases around tourist attractions or event venues during specific periods. Specifically, the urban infrastructure data acquisition unit possesses a multi-scale spatial resolution processing module. This module resamples the input raw data according to a specified grid size. The system applies bilinear interpolation to elevation data to estimate missing values. For population density data, the system uses kernel density estimation to convert discrete point data into a continuous probability density function. The generative AI processing unit is equipped with a spatiotemporal convolutional network. As input data, this AI processing unit receives a 4-dimensional tensor with spatial and temporal dimensions. For example, the input data is a tensor consisting of width, height, time step, and feature channels. The AI ​​processing unit uses a 3D convolutional filter to simultaneously extract spatial features and temporal variations. As output, the AI ​​processing unit generates predicted future population changes and disaster risk scores for each grid. For example, the output data includes arrays of predicted population densities for each time period in the next 24 hours and flood risk probabilities ranging from 0.0 to 1.0. The post-processing unit uses the output predictions to generate a dynamic infrastructure demand map. This unit identifies time periods and regions where demand exceeds specific thresholds and formulates dynamic resource allocation plans. This spatiotemporal analysis overcomes the limitations of previous planning methods based on static data. The system anticipates environmental changes over time, enabling proactive infrastructure management. This allows the system to prevent overinvestment while efficiently meeting peak demand. The system is best suited for use cases with drastic time fluctuations, such as the deployment of emergency medical facilities or the dynamic reconfiguration of shared bicycles. The system trains its model using autoregressive learning based on past time-series data.This system uses the mean squared error function to calculate the error between the predicted value and the performance value, and updates the parameters using the gradient descent method.

[0065] Traffic flow data is acquired from road sensors, intersection traffic counts, vehicle GPS data, public transportation passenger boarding and alighting counts, bus stop or station user counts, shared mobility service usage data, road congestion data, and traffic accident history. This data is used to identify areas with high pedestrian and vehicle traffic, such as major roads or intersections, areas around stations, and commercial facilities; areas prone to traffic congestion; and areas with high accident risk. For example, it also tracks traffic flow changes during morning and evening rush hours, during events, and during traffic control measures implemented due to construction or accidents. Specifically, this system employs an event-driven architecture for real-time processing of traffic data streams. The data acquisition unit stores data packets asynchronously sent from various sensors into a message queue. The system uses a stream processing engine and applies filtering processes such as moving averages or exponential smoothing. The generative AI processing unit has a time-series prediction model incorporating a long short-term memory network. This AI processing unit receives a multivariate time-series matrix representing traffic flow over a past period as input data. For example, the input data is a 60x3 matrix containing the number of vehicles passing through, average speed, and congestion index for each minute over the past 60 minutes. The AI ​​processing unit processes this matrix at each time step, updating the hidden state vector. As output, the AI ​​processing unit generates the future traffic congestion probability and accident risk score for the target area. For example, the output data includes a congestion probability of 85% within the next 30 minutes and an accident risk score of 0.75. If the output probability exceeds a predetermined safety threshold, the post-processing unit automatically initiates the detour route generation module. The post-processing unit then publishes the generated route information to a collaborating navigation system. This system uses this predictive method to pre-identify traffic infrastructure bottlenecks and optimize the overall system processing capacity. This system eliminates the need for reactive measures by human monitors, achieving automated and high-speed traffic control. This system is applicable to real-time applications such as dynamic optimization of signal control in smart cities or priority routing for emergency vehicles. This system uses supervised learning to train the model from pairs of past traffic data with actual congestion / accident records. This system uses the backward propagation method to calculate the gradient in the time direction and optimize the network weights.

[0066] Disaster risk data is obtained from past disaster records, disaster risk maps, meteorological data, foundation information, soil classification data, flood and debris flow disaster risk data, earthquake prediction data, tsunami inundation scenario data, fire occurrence history, shelter location information, and emergency vehicle route information, etc., to evaluate the risks of earthquakes, floods, debris flows, fires, tsunamis, storm surges, storms, and lightning strikes from multiple perspectives. For example, it extracts detailed information on areas with river flooding history, slopes that have experienced landslides in the past, reclaimed land with high liquefaction risk during earthquakes, tsunami inundation scenario areas, and densely populated urban areas with a high number of fires, etc., indicating areas with high disaster risk. Specifically, this system has a semantic data integration module for integrating disaster data in heterogeneous formats. This module converts vector-format polygonal data and raster-format grid data into a common spatial reference frame. This system generates independent risk layers for each disaster factor. The generative AI processing unit adopts an ensemble learning architecture that integrates multiple expert models. This AI processing unit receives multidimensional feature tensors extracted from each risk layer as input data. For example, the input data includes scalar values ​​representing the shear wave velocity of the foundation, a time-series vector of rainfall over the past 50 years, and a matrix representing the elevation gradient. The AI ​​processing unit uses a random forest algorithm and gradient boosting decision trees to calculate the individual risk for each disaster type. As output, the AI ​​processing unit generates a comprehensive disaster vulnerability index for each location and the contribution of key risk factors. For example, the output data includes a comprehensive vulnerability score of 0.88, and structured data such as a flood risk contribution of 60% and a liquefaction risk contribution of 30%. The post-processing unit inputs the output vulnerability index as a constraint on infrastructure configuration into the optimization solver. The post-processing unit performs a masking process that automatically excludes areas with vulnerability scores exceeding the allowable upper limit from the candidate sites. This system, through this multi-faceted risk assessment, accurately quantifies the risks of complex disasters that are missed in a single disaster risk map. This system delivers the technical benefits of improved durability throughout the entire lifecycle of infrastructure and significantly reduced maintenance and management costs. This system is applicable to use cases requiring high robustness, such as data center site selection or the configuration planning of backup facilities for critical infrastructure. This system uses cross-validation to optimize the hyperparameters of each base learner in the ensemble model. The system learns the branching conditions of each decision tree in a manner that minimizes the logarithmic loss function.

[0067] The power grid stability data is obtained from power company supply network information, past outage history, renewable energy adoption status, the presence or absence of distributed power sources, supply stability during peak demand periods, transmission line and substation location information, and the aging status of power infrastructure. This data is used to evaluate locations where stable power supply can be predicted, locations with low outage risk, and locations where renewable energy can be locally generated and consumed. For example, detailed information such as distance from transmission lines or substations, past outage frequency, and the adoption status of solar power generation or battery storage is obtained. Specifically, this system includes a graph processing module that parses the topology of the power network. This module models the connection relationships of transmission lines or substations as a directed graph. The system calculates the inflow and outflow of power to each node and executes a power flow calculation algorithm. The generative AI processing unit has an anomaly detection model using a variational autoencoder. As input data, the AI ​​processing unit receives a multivariate time series vector representing the power supply status of each node. For example, the input data is a 12-dimensional vector containing voltage variability, frequency deviation, and past outage durations. The AI ​​processing unit uses an encoder network to compress the input data into a low-dimensional latent space. The AI ​​processing unit reconstructs the original data from latent variables using a decoder network. As output, this unit generates a grid instability score based on the reconstruction error and the probability of future power outages. For example, the output data includes an instability score of 0.15 for node A and a 5% probability of a power outage within the next year. The subsequent processing unit normalizes the inverse of the output instability score as a grid stability index. This index is then used as a weighted parameter in the evaluation function for infrastructure setup. This system mathematically guarantees the reliability of energy supply during infrastructure operation through this advanced grid analysis. It minimizes system downtime caused by power outages, significantly improving service continuity. This system can be applied to use cases with high power consumption and unreliable stable supply, such as the configuration of fast charging stations for electric vehicles and the selection of edge computing sites. The system pre-learns the model using only normal-time power data. The system defines the sum of the Kullback-Leibler divergence and the reconstruction error as the loss function and updates the network parameters by minimizing this sum.

[0068] The system obtains data on existing infrastructure from urban infrastructure management ledgers, on-site survey data, on-site image analysis based on drones or autonomous vehicles, operational status data from sensor networks, and resident reports, and determines whether the same type of infrastructure has already been installed, whether there are aging or faulty facilities, utilization rates, and maintenance status. For example, it provides a detailed understanding of the location and operational status of existing disaster prevention sites, public charging stations, shared mobility terminals, surveillance cameras, and environmental sensors, and utilizes this information to identify redundant or excessive installations and areas without infrastructure. Specifically, the system has a spatial database engine that integrates infrastructure information from multiple sources. This engine links text-based management ledger data with image-based on-site survey data. The system applies semantic segmentation to high-resolution images acquired from drones. The generative AI processing unit is equipped with a deep learning model based on a single-stage convolutional neural network architecture as an object detection algorithm. This AI processing unit receives on-site image tensors and operational status vectors from sensors as input data. For example, the input data includes a 3-channel 1024×1024 pixel image tensor and a 1D array representing operational rate. The AI ​​processing unit extracts spatial feature maps from the image using convolutional layers. As output, this AI processing unit generates bounding box coordinates, category labels, and degradation scores representing the degree of aging of existing infrastructure within the image. For example, the output data includes a list of coordinates, the category "surveillance camera," and structured data such as a degradation score of 0.7. The subsequent processing unit converts the output coordinate data into geospatial coordinates and updates the distribution map of existing infrastructure. This subsequent processing unit uses the Thiessen polygon generation algorithm to calculate the coverage of existing infrastructure and geometrically determine any gaps in coverage. This system achieves automated infrastructure inventory without relying on manual surveys through the fusion of image recognition and spatial analysis. This system delivers the technical effect of completely eliminating redundant investment and maximizing the efficiency of infrastructure deployment within a limited budget. This system can be applied to fields requiring effective utilization of existing assets, such as developing inspection plans for aging bridges and optimizing the configuration of smart streetlights. The system fine-tunes the model using a labeled infrastructure image dataset. This system updates weights by minimizing the sum of the bounding box regression error and the cross-entropy error of the class classification.

[0069] In addition, data on urban development plans, land use classification, foundation strength, event organization, new public transportation plans, urban redevelopment areas, disaster shelter locations, critical infrastructure locations, emergency vehicle routes, landowner information, laws and regulations, setup costs, maintenance costs, resident survey results, user feedback, and post-implementation performance data are also collected and utilized for multi-faceted optimization of urban infrastructure configuration. Specifically, this system adopts a data lake architecture that integrates structured and unstructured data. The system uses a natural language processing module to extract keywords and sentiment polarities from legal texts or resident surveys. Cost data is formatted as a floating-point matrix and input into the economic evaluation module. The generative AI processing unit is a hybrid system combining a knowledge graph embedding model and a multi-objective optimization algorithm. As input data, this AI processing unit receives triples of entities and relationships extracted from various data sources, as well as vectors of numerical data. For example, the input data includes a text triple "Area A - Use - Commercial Area" and a scalar value representing the initial setup cost. The AI ​​processing unit uses a translation embedding technique to configure entities within the knowledge graph into a continuous vector space. As output, this AI processing unit generates a multi-dimensional evaluation index vector for each candidate site and a set of Pareto optimal solutions. For example, the output data is a vector containing a cost efficiency score of 0.8, a resident acceptance score of 0.9, and a legal suitability indicator of "true". The subsequent processing unit calculates a weighted linear sum based on decision-makers' preferences for the output Pareto optimal solutions. This subsequent processing unit calculates the final comprehensive evaluation score and generates a ranking list of candidate sites. This system mathematically solves the trade-offs between opposing objective functions through the integrated analysis of this multi-dimensional data. This system rapidly derives optimal solutions that simultaneously satisfy complex constraints beyond the limits of human cognition. This system can be applied to complex issues involving multiple stakeholders, such as the master planning of large-scale urban redevelopment projects and the optimization of tenant allocation for mixed-use commercial facilities. This system uses a genetic algorithm to explore the Pareto frontier. This system uses the weighted sum of various evaluation indicators as the fitness function and repeatedly applies crossover and mutation operators to evolve the solution population.

[0070] These diverse data are input from the urban infrastructure data acquisition department to the generative AI processing department. The generative AI processing department is composed of, for example, a combination of large-scale language models, deep learning models, time series analysis AI, image generation AI, multimodal AI, and reinforcement learning AI. Generative AI normalizes each data point based on minimum and maximum values, for example, by adjusting the evaluation based on factors such as lower altitude, higher population density, greater traffic flow, lower disaster risk, more stable power supply, and the absence of existing infrastructure. Furthermore, the weighting can be dynamically adjusted for each purpose or category of urban infrastructure installation. For example, in the installation of disaster prevention bases, the weight of disaster risk or power grid stability can be increased; in the installation of shared mobility terminals, the weight of traffic flow or population density can be increased; in the installation of surveillance cameras, the weight of crime history or pedestrian data can be increased; and in the installation of environmental sensors, the weight of air pollution risk or factory and traffic volume data can be increased. This can be flexibly set according to the purpose of urban planning or the category of infrastructure. Generative AI incorporates past installation performance data, feedback from on-site managers, future population growth projections, disaster risk information, installation cost information, operational cost data, maintenance and management history, number of user complaints, post-installation communication quality data, resident survey results, and user feedback data as learning data to continuously improve the accuracy of candidate site extraction. For example, it can learn from the operational performance of past infrastructure installations to assess post-installation utilization rates or failure rates, maintenance and management costs, resident satisfaction, and improvements in user convenience, and reflect these improvements in candidate site extraction under similar conditions. Furthermore, it considers population projections from urban development plans or redevelopment areas, future transportation infrastructure construction plans, disaster evacuation routes or shelter locations, the location of critical infrastructure facilities, and emergency vehicle routes, enabling long-term installation optimization looking several years into the future. Specifically, this system executes a feature engineering pipeline to unify the dimensions of the input data. This pipeline applies min-max scaling to each feature, mapping values ​​to the range of 0.0 to 1.0. The system converts categorical data into binary vectors using one-hot encoding. This generative AI processing unit is equipped with a multi-task learning network featuring an attention mechanism. As input data, the AI ​​processing unit receives normalized feature vectors and context vectors representing infrastructure categories. For example, the input data includes a 256-dimensional environmental feature vector and a 64-dimensional embedding vector representing a "disaster preparedness site." Based on the context vectors, the AI ​​processing unit calculates dynamic attention weights for each feature. As output, the AI ​​processing unit generates an optimized overall suitability score for each infrastructure category and a heatmap tensor representing the contribution of each feature. For example, the output data includes a suitability score of 0.95 for a disaster preparedness site and an attention weight of 0.6 for disaster risk features.The post-processing unit combines the output scores with spatial coordinates to construct a three-dimensional fitness distribution matrix. This matrix is ​​then input into a clustering algorithm to extract the optimal setting regions as polygon data. This system, through its dynamic weighting mechanism, flexibly addresses the optimization requirements of various infrastructure categories with a single model. Because the system can adapt to new infrastructure requirements without model relearning, it significantly reduces computational costs and development timelines. This system can be applied to scalable use cases such as complex infrastructure development plans in smart cities and the formulation of strategies for the simultaneous deployment of multiple services. The system uses a weighted sum of the losses from each task as the loss function for multi-task learning. Homoscedastic uncertainty is used to automatically adjust the weights of each task during the learning process.

[0071] The comprehensive score calculated by the generative AI processing unit is sent to the candidate site selection unit. This unit selects candidate sites flexibly based on the application policy or urban planning objectives, such as the top 10 locations or all locations exceeding a predetermined threshold. For each selected candidate site, detailed data is compiled, including coordinates, altitude, population density, traffic flow, disaster risk, power grid stability, presence or absence of existing infrastructure, comprehensive score, surrounding environment information, disaster risk, cost information, future population growth projections, integration with urban development plans, relationship with disaster evacuation routes, landowner information, legal and regulatory information, maintenance and management costs, resident satisfaction, and user convenience improvement effects. Specifically, the candidate site selection unit uses an in-memory database to perform high-speed sorting and filtering. The selection unit applies a quicksort algorithm to the input comprehensive score array, rearranging the array in descending order. The system dynamically reads the number of extractions or score thresholds set as application policy parameters. The selection unit uses an R-tree as a spatial index to calculate the physical distances between the extracted locations. The selection unit receives the score array of all grids and the attribute data dictionary of each grid as input data. For example, the input data is a floating-point array with 1 million features and a structured text object containing metadata corresponding to each feature. The selection unit excludes lower-scoring locations using a penalty function if nearby candidate locations do not meet the specified minimum distance constraint. As output, the selection unit generates a structured dataset containing a list of the finally selected candidate locations and detailed attributes of each candidate location. For example, the output data is a list containing candidate location IDs, latitude and longitude coordinates, a comprehensive score of 0.91, and initial cost estimates. The post-processing unit serializes the output dataset into a structured text format and sends it to the visualization unit via the network. This system, through a selection algorithm that considers spatial constraints, prevents infrastructure bias and maximizes coverage across the entire city. Mathematically, this system avoids the local over-configuration problem that may arise in simple score-order extraction. This system can be applied to use cases requiring maximum coverage and minimum self-cannibalization, such as new store opening plans for chain stores and optimal configuration of logistics warehouses. This system aims to solve constrained optimization problems by employing a heuristic algorithm based on the Lagrange multiplier method. Through iterative exploration, the system efficiently discovers a set of solutions that maximizes the objective function while satisfying the constraints.

[0072] The visualization department uses candidate site data sent from the candidate site selection department to overlay and display the location or recommendation level, various scores, and detailed attributes of candidate sites in a 3D AR or VR space. For example, the location of candidate sites can be displayed as pins on the camera image of a tablet terminal. When a candidate site is clicked, a pop-up window displays detailed information such as altitude, population density, traffic flow, disaster risk, power grid stability, presence or absence of existing infrastructure, overall score, cost information, resident satisfaction, and user convenience improvement effect. If VR glasses are used, the distribution of candidate sites or their surrounding environment, disaster risk distribution, population density heatmap, traffic flow distribution, existing infrastructure distribution, urban development plan areas, evacuation routes, the location of important infrastructure facilities, and emergency vehicle routes can be confirmed in multiple layers overlaid on a 3D model of the entire city. In addition, the simultaneous comparison and display of multiple candidate sites allows for an intuitive selection of the best location. For example, the setup of disaster prevention outposts in high-rise building complexes in urban areas, shared mobility terminals in mountainous transportation hubs, surveillance cameras or environmental sensors in temporary population concentration areas around event venues, public charging stations in redevelopment areas, and infrastructure around shelters to reduce disaster risk—each site's unique circumstances can be visualized in three dimensions. Specifically, this visualization unit is equipped with a hardware-accelerated rendering engine using a graphics processing unit. The visualization unit calculates a projection matrix to convert the received candidate site data into a three-dimensional coordinate system. In AR mode, the system uses the terminal's gyroscope sensor and the camera's optical flow to simultaneously perform self-position estimation and mapping. The visualization unit receives the candidate site's three-dimensional coordinate vectors, attribute data, and the camera's pose matrix as input data. For example, the input data includes (X, Y, Z) vectors in the world coordinate system, scalar values ​​representing scores, and a 4×4 transformation matrix. The visualization unit uses a shader program to dynamically generate pin colors or particle effects based on the scores. As output, the visualization unit generates a framebuffer that overlays virtual objects onto the real-world camera image. For example, the output data is an RGBA image tensor updated at 60 frames per second. The post-processing unit handles user gaze input or raycasting decisions from the controller and renders detailed data of the selected candidate locations on the UI layer. This system uses advanced 3D visualization technology to present complex multidimensional data in a format optimized for human visual processing. It facilitates understanding of spatial context difficult to grasp within numerical data sets, dramatically improving the speed and accuracy of decision-making. This system can be applied to essential use cases for intuitive information sharing, such as consensus-building tools in resident briefings for urban planning and on-site signage systems from distant locations. The system uses a level-of-detail (LHD) control algorithm to dynamically adjust the polygon count of the 3D model based on the distance from the viewpoint. This optimizes the rendering load and enables smooth real-time rendering even on mobile devices.

[0073] Furthermore, if the on-site supervisor inputs new obstacles or population changes, event information, resident requests, on-site observation data, on-site photos, and user action logs from the terminal, this information is reflected in real time in the generative AI on the server side, and the evaluation or extraction results of candidate sites are updated instantly. For example, by collecting photos of newly discovered obstacles on-site, observation data on seasonal population changes, sudden event information, subjective evaluations or comments from the on-site supervisor, resident requests or complaints, and user action logs, the generative AI analyzes and learns this new data in real time, enabling it to update the evaluation or extraction results of candidate sites instantly. Specifically, this system utilizes the WebSocket protocol to establish bidirectional real-time communication between edge devices and the cloud. The on-site information collection unit transmits the input text data to the natural language processing pipeline and performs morphological parsing and named entity extraction. The system inputs uploaded on-site photos into a convolutional neural network and extracts image feature vectors. The generative AI processing unit has a dynamically updated module with an online learning algorithm installed. The AI ​​processing unit receives newly acquired image feature vectors, text embedding vectors, and time-series arrays of observation data as input. For example, the input data might include a 512-dimensional image feature vector, a 768-dimensional text embedding vector, and timestamps representing the start and end times of an activity. The AI ​​processing unit compares this new data with the existing knowledge base and locally updates the feature tensors of the relevant grid cells. As output, the AI ​​processing unit generates a difference vector of new suitability scores based on the updated feature quantities. For example, the output data might include an update command indicating that the score for a specific grid cell has increased by +0.15. The subsequent processing unit uses this difference vector to incrementally update the database and pushes re-rendering commands for the changed areas to the visualization unit. This system, through this real-time feedback loop, instantly bridges the gap between static pre-processing data and dynamic on-site conditions. The system eliminates latency caused by batch processing, reducing the system's responsiveness to changes in conditions to milliseconds. This system can be applied to time-constrained use cases such as dynamic revision of emergency infrastructure recovery plans during disasters and real-time optimization of crowd guidance systems during large-scale events. This system uses stochastic gradient descent to update the model weights for new data with a small learning rate. To prevent catastrophic forgetting, the system uses a replay buffer that holds key past data for learning.

[0074] Furthermore, considering factors such as the location of evacuation routes or shelters during disasters, emergency vehicle routes, the location of critical infrastructure, urban development plans, redevelopment areas, new public transportation plans, landowner information, laws and regulations, setup or maintenance costs, resident satisfaction, and improvements in user convenience, generative AI can automatically propose setup plans that easily maintain urban functions during disasters, flexibly adapt to future urban growth or population dynamics, and maximize cost efficiency or resident satisfaction. Moreover, when simultaneously optimizing the configuration of multiple urban infrastructure categories (such as disaster preparedness sites, public charging stations, shared mobility terminals, surveillance cameras, and environmental sensors), generative AI can automatically generate optimal composite configuration schemes by setting individual evaluation benchmarks or weights for each infrastructure category, considering factors such as mutual configuration balance or interference risks, the possibility of shared setups, and the synergistic effects of composite infrastructure. Furthermore, during the post-installation operation phase, the generative AI continuously learns from performance or operational costs, resident satisfaction, user feedback, maintenance and management history, and the number of user complaints. It can also automatically re-evaluate or add installations, withdraw proposals, and propose improvements as needed. Specifically, this system installs an evolutionary computation engine that simultaneously optimizes multiple objective functions. This engine encodes the configuration patterns of each infrastructure category into chromosomes and executes a genetic algorithm. The system formulates physical or radio interference between infrastructures as penalty functions. The generative AI processing unit possesses a multi-agent system based on deep reinforcement learning. As input data, the AI ​​processing unit receives a multi-dimensional state tensor representing the current state of the city and action history vectors of each agent. For example, the input data includes a 3-dimensional tensor representing the attributes of the city grid and a list of coordinates representing the current location of each infrastructure. Each agent in the AI ​​processing unit is responsible for configuring specific infrastructure categories, and the agents learn to coordinate their actions. As output, the AI ​​processing unit generates a set of optimal configuration coordinates for each infrastructure and a prediction of the expected future returns resulting from that configuration. For example, the output data includes coordinate data representing the shared setup of disaster preparedness points and charging stations, as well as a utility score of 0.94 for the entire system. The post-processing unit inputs the output configuration scheme into a simulation environment and automatically performs a stress test during a disaster. Based on the simulation results, the post-processing unit verifies the robustness of the configuration scheme and only uses the scheme that meets the benchmark as the final output. This system, through multi-agent reinforcement learning, derives the optimal solution as a system within a system of systems for the entire city, which is unattainable through optimization of a single infrastructure. This system generates significant economic benefits by preventing infrastructure silos and eliminating duplication of equipment investment.This system can be applied to the design of extremely complex composite systems, such as the overall design of smart cities and the optimal allocation of limited resources within space stations. The system models the environment based on Markov decision processes and uses the Q-learning algorithm to approximate the action value function. It employs an experience replay memory to sample past transition data, stabilizing the neural network's learning process.

[0075] In this way, the embodiments of the present invention, through the coordinated operation of the urban infrastructure data acquisition unit, the generative AI processing unit, the candidate site selection unit, and the visualization unit, comprehensively analyze various data across the entire city, extract the best candidate sites for each installation purpose or infrastructure category, and provide an environment that allows for intuitive comparison and discussion in three-dimensional space. This achieves profound and multifaceted technical effects, including efficient and high-level urban infrastructure configuration planning, cost reduction, improved disaster response capabilities, future prediction and response, real-time on-site feedback, composite optimization of multiple infrastructure categories, increased resident satisfaction, optimized operating costs, and flexible response to urban growth. Specifically, this system uses a microservice architecture to ensure high scalability and fault tolerance between modules. The system controls data communication between units via an API gateway through asynchronous message passing. The system uses container orchestration tools to dynamically scale instances of the generative AI processing unit horizontally according to the computing load. The system receives metric vectors representing the throughput, latency, and resource utilization of the entire system as input data. For example, the input data includes values ​​such as CPU utilization of 85%, memory usage of 16 gigabytes, and average response time of 120 milliseconds. This system collects these metrics in real time through a monitoring agent and inputs them into anomaly detection algorithms. As output, the system generates resource allocation change commands or automatic failover commands in case of failure. For example, the output data includes control signals instructing the startup of new containers and traffic routing change rules. The post-processing unit sends the output commands to the infrastructure management API and autonomously optimizes the system architecture. Through this autonomous system management, the system maintains stable performance even in large-scale urban data processing. This system maximizes the efficiency of hardware resource utilization and significantly reduces the cost of cloud computing. It can be applied to large-scale, mission-critical systems such as national-scale wide-area infrastructure management platforms and global logistics network optimization systems. The system encodes its entire architecture as software-defined infrastructure and integrates with a version control system to achieve continuous integration and continuous deployment. This enables rapid deployment of new features and rapid rollback in case of failure.

[0076] (System Composition) The system described in this implementation includes an urban infrastructure data acquisition unit, a generative AI processing unit, a candidate site selection unit, a visualization unit, a field information collection unit, a data linkage unit, and a user interface unit. The urban infrastructure data acquisition unit collects various types of data from each grid cell after the entire city is divided into grids at arbitrary resolutions such as 10 meters, 25 meters, 50 meters, and 100 meters. For example, it acquires elevation data from geographic information systems, satellite imagery, drone aerial photography data, topographic laser scanning data, and surface subsidence observation data, and uses this data for assessing flood risk, surface subsidence risk, landslide risk, and tsunami inundation risk. Furthermore, it acquires population density data from census data, resident basic records, location information from portable terminals, traffic flow sensors, location information from SNS posts, data on the number of people entering commercial facilities, data on the number of attendees from event organizers, Wi-Fi access point connection data, and public transportation usage data. It also tracks population changes and peak traffic flows during weekdays and weekends, day and night, event times, and each season. Traffic flow data is obtained from road sensors, intersection traffic counts, vehicle GPS data, public transportation passenger boarding and alighting counts, bus stop and station user counts, shared mobility service usage data, road congestion data, and traffic accident history. This data identifies areas with high pedestrian and vehicle traffic, such as major roads and intersections, areas around stations, and areas around commercial facilities, as well as locations with frequent traffic congestion and high traffic accident risk. Disaster risk data is also obtained from past disaster history, disaster risk maps, meteorological data, foundation information, soil classification data, flood and debris flow disaster risk data, earthquake prediction data, tsunami inundation scenario data, fire history, shelter location information, and emergency vehicle route information. This allows for a multi-faceted evaluation of risks related to earthquakes, floods, debris flows, fires, tsunamis, storm surges, storms, and lightning strikes. Data on grid stability is obtained from power company supply network information, past power outage records, renewable energy adoption status, the presence of distributed power sources, supply stability during peak demand periods, transmission line and substation locations, and the aging status of power infrastructure. This data is used to evaluate locations where stable power supply can be foreseen, locations with low outage risk, and locations where renewable energy can be locally generated and consumed. Data on the availability of existing infrastructure is obtained from urban infrastructure management records, on-site survey data, analysis of on-site images from drones or autonomous vehicles, operational data from sensor networks, and reports from residents. This data also determines whether similar infrastructure has already been installed, its aging or malfunction status, utilization rate, and maintenance management status.In addition, data such as urban development plans, land use classification, foundation strength, event information, new public transportation plans, urban redevelopment area information, disaster shelter location information, critical infrastructure location information, emergency vehicle route information, landowner information, laws and regulations, setup cost data, maintenance and management cost data, resident survey results, user feedback data, and post-implementation performance data are also acquired and utilized for multi-faceted optimization of urban infrastructure configuration. Specifically, this system installs a distributed message queue that integrates asynchronous streams from various data sources. The data linkage unit functions as a protocol conversion gateway, converting different communication specifications into a unified internal format. This system uses an in-memory data grid to cache frequently updated sensor data with low latency. The urban infrastructure data acquisition unit uses a spatial hashing algorithm to map the coordinates of each grid to a one-dimensional index key. This system receives binary payloads sent from each sensor node as input data. For example, the input data is a 128-byte data packet containing the sensor ID, timestamp, and measurement value. This system decodes the payload and updates the attribute table of the corresponding grid. This system outputs a multidimensional state tensor reflecting the latest state of the entire grid. For example, the output data is a tensor with a spatial resolution of 1000×1000 and 64 channels of features covering the entire city. The subsequent processing unit writes this state tensor into the shared memory area of ​​the generative AI processing unit at regular intervals. Through this advanced data pipeline, the system gains the ability to process petabyte-level city data in real time. This system eliminates data silos, significantly improving the overall data consistency and availability of the system. This system can be applied to use cases requiring ultra-large-scale and low-latency data processing, such as wide-area control systems for autonomous vehicles and real-time supply and demand adjustment of smart grids. This system uses data compression algorithms to minimize network bandwidth consumption while ensuring data integrity. This system also adds parity checks and error correction codes to automatically repair data gaps in the communication path.

[0077] The generative AI processing department analyzes various data input from the urban infrastructure data acquisition department, such as combined large-scale language models, deep learning models, time series analysis AI, image generation AI, multimodal AI, and reinforcement learning AI. It normalizes the data based on minimum and maximum values, and assigns higher evaluations based on factors like lower altitude, higher population density, greater traffic flow, lower disaster risk, more stable power supply, and the absence of existing infrastructure. The weighting is dynamically adjusted according to the purpose and category of urban infrastructure deployment. For example, in setting up disaster prevention bases, the weight of disaster risk or power grid stability is increased; in setting up shared mobility stations, the weight of traffic flow or population density is increased; in setting up surveillance cameras, the weight of crime history or pedestrian data is increased; and in setting up environmental sensors, the weight of air pollution risk or factory and traffic volume data is increased. This can be flexibly set according to the purpose of urban planning or the category of infrastructure. Furthermore, past performance data, feedback from on-site personnel, future population growth forecasts, disaster risk information, setup cost information, operational cost data, maintenance and management history, number of user complaints, post-operation communication quality data, resident survey results, and user feedback data are also imported as learning data to continuously improve the accuracy of candidate site extraction. For example, based on the operational performance of previously installed infrastructure, the system can learn the post-installation utilization rate or failure rate, maintenance and management costs, resident satisfaction, and user convenience improvement effects, and reflect these in the extraction of candidate sites under similar conditions. Specific examples of prompts for generative AI to read could include: "Comprehensively analyze the following city map containing elevation data, population density data, traffic flow data, disaster risk data, power grid stability data, and existing infrastructure availability data to extract the top 10 most suitable locations for setting up disaster prevention bases. Output coordinates, elevation, population density, traffic flow, disaster risk, power grid stability, existing infrastructure availability, and comprehensive score for each location." or "Focusing on traffic flow and population density, extract the most suitable candidate sites for setting up shared mobility stations." Specifically, the generative AI processing unit possesses an orchestration engine that integrates multiple neural network models. This engine dynamically routes appropriate sub-models based on the modality of the input data. The system converts natural language prompts into integer sequences via a word segmenter and inputs them into the embedding layer. The AI ​​processing unit receives tokenized prompt vectors and multi-dimensional feature tensors of a city map as input data. For example, the input data contains a 512-dimensional prompt embedding vector and a 256×256×64 spatial feature tensor. The AI ​​processing unit uses a cross-attention mechanism to calculate the correlation between the text's indication content and spatial features. Based on the conditions specified in the prompt, such as "disaster prevention base" or "shared mobility station," the AI ​​processing unit dynamically changes the activation function parameters of each feature channel.The AI ​​processing unit outputs a suitability score map for each grid and a structured list of the top-ranked locations. For example, the output data is a 256×256 probability map and an array of structured text containing the coordinates and scores of the top 10 locations. The subsequent processing unit applies a Gaussian smoothing filter to the output probability map to remove noise. This subsequent processing unit detects local maxima from the smoothed map to determine the final candidate site coordinates. This system, through the fusion of a natural language interface and deep learning, enables urban planners without programming knowledge to perform advanced AI analysis. This system eliminates the need for manual parameter adjustments, significantly reducing the delivery time from requirement definition to result output. This system can be applied to use cases requiring flexible interfaces, such as conversational decision support systems during disasters and citizen-participatory urban planning simulators. This system uses the policy gradient method from reinforcement learning, using user feedback as a reward signal to fine-tune the model. This system employs a proximal policy optimization algorithm, balancing learning stability and convergence speed.

[0078] The candidate site selection unit is based on a comprehensive score calculated by a generative AI processing unit, such as the top 10 locations or all locations exceeding a predetermined threshold. It flexibly extracts candidate sites according to the application policy or urban planning objectives. For each extracted candidate site, detailed data is compiled, including coordinates, altitude, population density, traffic flow, disaster risk, power grid stability, presence or absence of existing infrastructure, comprehensive score, surrounding environment information, disaster risk, cost information, future population growth projections, consistency with urban development plans, relationship with disaster evacuation routes, landowner information, legal and regulatory information, maintenance and management costs, resident satisfaction, and user convenience improvement effects. Specifically, this candidate site selection unit uses an online analytical processing engine based on a multidimensional data cube. The selection unit stores the calculated comprehensive score in a fact table and combines the attribute data as dimension tables. The system dynamically converts user-defined application policy parameters into queries using a structured query language. The selection unit receives the full-grid score dataset and extraction condition parameters as input data. For example, the input data contains millions of rows of records and a logical expression for "score above 0.8 and initial cost below 10 million". The selection unit drives a parallel query execution engine on an in-memory database, rapidly scanning records that meet the criteria. As output, the selection unit generates a filtered and sorted data frame of candidate locations. For example, the output data is a tabular data structure with the attribute values ​​of the top 10 locations organized as columns. The post-processing unit serializes the output data frame and forwards it to the visualization unit or an external report generation system. This system uses advanced database processing technology to instantly extract the best subset from a large pool of candidate locations. This system reduces the computational order of retrieval processing accompanied by complex constraints, dramatically improving system responsiveness. This system can be applied to use cases that frequently involve conditional retrieval from large-scale data, such as automated real estate matching systems and optimal warehouse selection in wide-area supply chains. This system uses Bloom filters to minimize disk I / O overhead by early exclusion of data from the retrieval targets. This system caches the execution plans of frequently executed queries, further accelerating subsequent retrieval processes.

[0079] The visualization unit uses candidate site data sent from the candidate site selection unit to overlay and display the location or recommendation level, various scores, and detailed attributes of candidate sites in a 3D AR or VR space. For example, the location of candidate sites can be displayed as pins on a tablet's camera image. When a candidate site is clicked, detailed information such as altitude, population density, traffic flow, disaster risk, power grid stability, availability of existing infrastructure, overall score, cost information, resident satisfaction, and user convenience improvement effect can be displayed. Using VR glasses, the distribution of candidate sites or their surrounding environment, disaster risk distribution, population density heatmaps, traffic flow distribution, existing infrastructure distribution, urban development plan areas, evacuation routes, the location of important infrastructure, and emergency vehicle routes can be confirmed in multiple layers overlaid on a 3D model of the entire city. Furthermore, multiple candidate sites can be compared and displayed simultaneously, allowing for intuitive selection of the best location. Specifically, this visualization unit uses a dedicated graphics processing unit on an edge device in conjunction with a neural network accelerator. This visualization unit goes beyond simple 3D graphics rendering, performing dynamic rendering optimization processing using artificial intelligence. Existing augmented reality and virtual reality systems suffer from computational resource exhaustion and frame rate degradation due to their constant rendering of massive polygon and texture data at the highest resolution. To address this issue, this visualization department has implemented a rendering-optimized AI architecture that integrates multiple deep learning models specialized for gaze prediction and spatial recognition. The visualization department first acquires environmental data from a depth sensor and camera mounted on the terminal. The input data for the spatial recognition AI model consists of a 1024x1024 resolution RGB image tensor, a depth map tensor of the same resolution, and a six-DOF time-series array of data obtained from an inertial measurement unit. Specific examples of this input data include depth data representing the arrangement of walls or furniture indoors, image data representing the outlines of buildings or roadside trees outdoors, and head acceleration data accompanying the user's walking. The spatial recognition AI model uses a convolutional neural network to process this input data. The output data from this spatial recognition AI model includes semantic labels for each voxel in three-dimensional space, a probability distribution map of occlusion boundaries, and an inferred vector of ambient light. Specific examples of this output data include occlusion mask values ​​representing areas of virtual objects obscured by real-world buildings, numerical values ​​for the direction and intensity of light sources applied to the virtual objects, and category identifiers for identifying bed surfaces or walls. The visualization unit inputs this output into subsequent occlusion culling and lighting calculation modules. The visualization unit accurately omits the rendering of virtual objects hidden behind real-world objects, pixel by pixel. This drastically reduces the unnecessary rendering load on the graphics processing unit. Furthermore, the visualization unit executes a gaze prediction AI model to predict the user's gaze direction.The input data for this gaze prediction AI model consists of a time-series array of two-dimensional pupil coordinates obtained from eye-tracking sensors, and quaternion data representing head posture. Specific examples of this input data include gaze movement trajectory data over the past 500 milliseconds, dwell time data when gazing at a specific virtual object, and angular velocity data representing sharp head rotation. The gaze prediction AI model employs a Transformer architecture. It uses a self-attention mechanism to predict the destination of gaze movement. The output data from this gaze prediction AI model consists of a gaze probability distribution map for each region on the screen, and a numerical vector of importance scores for each virtual object. Specific examples of this output data include the center coordinates of areas where the user is more than 90% likely to gaze next, low importance scores for objects located in the periphery of the visual field, and high priority values ​​for pop-up information. The visualization unit inputs this output to the subsequent detail level control module. The visualization unit renders virtual objects in areas with high gaze probability using high-resolution polygon models and high-fine textures. For areas in the peripheral visual field with low gaze probability, the visualization unit applies a simplified model with reduced polygon count. This visualization unit significantly reduces the computational cost of rendering without compromising visual quality through its gaze-based rendering technique. In the learning process of these AI models, the visualization unit performs supervised learning using a large amount of gaze tracking data and 3D environment data. This learning process updates the network weights in a way that minimizes the mean squared error between the predicted gaze position and the actual gaze position. As a data augmentation technique, the visualization unit applies random noise additions to the input image or temporal offsets to the gaze data. This improves the model's generalization performance for unknown users or diverse environments. Through these techniques, the system maintains a stable frame rate of over 90 frames per second. The visualization unit suppresses rendering latency to below 20 milliseconds. As a result, the visualization unit effectively prevents motion sickness in users in virtual reality environments. When simultaneously comparing and displaying multiple candidate locations, the visualization unit also dynamically adjusts the amount of information based on the user's gaze. This reduces the user's cognitive load. A specific application area for this visualization unit is the simulation of numerous communication infrastructure placement candidate locations on a large 3D model of a city by urban planning officials. This visualization unit can also be applied to disaster response use cases where emergency vehicle routes or evacuation routes are overlaid with the operational status of communication infrastructure during a disaster. In augmented reality displays using tablet terminals, this visualization unit also integrates self-positioning algorithms for camera self-positioning and environment mapping with artificial intelligence models. The visualization unit aligns the coordinate systems of real and virtual spaces with millimeter-level precision.This visualization provides natural, landscape-like information prompts that blend seamlessly into reality. It goes beyond a simple listing of information; instead, it offers an adaptive interface that understands the user's intent and the context of the environment. This visualization fundamentally improves the efficiency of computer graphics processing and the operability of user interfaces.

[0080] The on-site information collection unit is responsible for receiving data from the on-site manager, including new obstacles or population changes, event information, resident requests, on-site observation data, photos, and user activity logs. This data is then fed back to the generative AI on the server side in real time, enabling the immediate updating of evaluations and extraction results for potential site selections. For example, it collects photos of newly discovered obstacles, seasonal population change observation data, sudden event information, subjective evaluations or comments from the on-site manager, resident requests or complaints, and user activity logs. The generative AI analyzes and learns from this new data in real time, updating the evaluations and extraction results for potential site selections accordingly. Specifically, this on-site information collection unit is installed as a dedicated application running on an edge device. It acquires multimedia data from the terminal's camera module or GPS sensor. The unit converts the acquired image data into, for example, 224-pixel square RGB image tensors. It converts text input data into word embedding vectors for natural language processing. Finally, the unit sends this preprocessed data to the generative AI processing unit on the server side via a secure communication protocol. This generative AI processing unit employs a multimodal Transformer architecture. It maps input image tensors and text vectors to the same high-dimensional feature space. For example, it receives an image tensor of a building under construction and a text vector of an obstacle 50 meters high as input. Based on these inputs, it outputs an environmental variation vector containing the obstacle's 3D coordinates and attenuation coefficients. This vector is then added to the weight parameters of an existing urban environment model. The system recalculates the evaluation score of candidate sites using these parameters. If the recalculated score falls below a predetermined threshold, the system automatically excludes the candidate site. This dynamic update process significantly improves the immediacy of evaluation compared to existing batch processing that relies on manual human input. The system reflects the latest on-site conditions in the model with a latency of milliseconds. This prevents a decrease in accuracy in communication infrastructure deployment plans. The system demonstrates high technical effectiveness in use cases such as temporary base station deployment during disasters or mobile base station configuration during events.

[0081] The Data Linkage Department is responsible for real-time data exchange between the Urban Infrastructure Data Acquisition Department, Generative AI Processing Department, Candidate Site Selection Department, Visualization Department, Field Information Collection Department, and User Interface Department, ensuring data consistency and timeliness. When a field manager uploads field photos or observation data from a terminal, the server-side Generative AI Processing Department immediately parses the data and updates the evaluation or extraction results of the candidate sites in real time. Specifically, this Data Linkage Department utilizes a distributed message queue system and an in-memory database. It establishes asynchronous and high-throughput data stream communication between the various processing departments. The Data Linkage Department employs a publish-subscribe model. It publishes image tensors or text vectors sent from the Field Information Collection Department to specific topics. The Generative AI Processing Department subscribes to these topics and immediately acquires new data. The Data Linkage Department assigns timestamps and spatial coordinate metadata to data packets. Using this metadata, it synchronizes asynchronous data from multiple sources spatiotemporally. For example, it combines an image tensor of location A at time T1 with a sensor data vector of location A at time T2 within the same time window. The data linkage unit inputs the combined multidimensional tensor into the generative AI processing unit. The generative AI processing unit parses the multidimensional tensor and outputs the latest environmental assessment score for location A. The data linkage unit forwards this assessment score to the candidate location selection unit and the visualization unit in milliseconds. This system reduces the coupling between modules through its asynchronous message passing architecture. This prevents overall system latency even when the load is concentrated on a specific processing unit. The system achieves large-scale parallel processing while maintaining data consistency. In use cases such as real-time traffic control systems in smart cities or dynamic reconfiguration of wide-area monitoring networks, this system delivers the technical benefits of reduced communication load and increased processing speed.

[0082] The user interface (UI) unit provides a screen on the terminal for administrators to view detailed information about candidate sites, compare and discuss them, input feedback, or add information. For example, it enables various interactions such as displaying detailed information for each candidate site, parallel comparison of multiple candidate sites, attaching on-site photos, inputting evaluations and comments on candidate sites, inputting resident survey results, and collecting user feedback. Specifically, the UI utilizes the edge terminal's graphics processing unit to perform dynamic rendering. The UI parses the evaluation scores or 3D coordinate data of candidate sites received from the server as structured data in JSON format. Based on this structured data, the UI generates a polygonal mesh in 3D space. The UI applies color mapping corresponding to the evaluation scores to the generated polygonal mesh. For example, the UI renders high-scoring locations as red heatmaps. The UI acquires touch or gesture input from the user as 2D screen coordinate vectors. The UI converts these screen coordinate vectors into raycast vectors in 3D space. The UI performs an intersection determination between the raycast vectors and the polygonal mesh. Based on this intersection determination, the UI identifies the specific candidate site selected by the user. The user interface overlays detailed attribute vectors for specific candidate locations onto the screen. It inputs user-defined text comments into a natural language processing (NLP) model. The NLP model outputs a sentiment polarity score or importance weight based on the text. The user interface then sends this output score as feedback data to the server. This system significantly reduces human cognitive load through its intuitive 3D interaction. It simplifies the comparative analysis of complex multidimensional data. In use cases such as urban planning simulators or infrastructure maintenance inspection applications, this system provides technical benefits in reducing operational errors and accelerating decision-making.

[0083] Thus, the system involved in this embodiment operates in tandem with a city infrastructure data acquisition unit, a generative AI processing unit, a candidate site selection unit, a visualization unit, a field information collection unit, a data linkage unit, and a user interface unit. It comprehensively analyzes various data across the entire city, extracts the optimal candidate sites for each installation purpose or infrastructure category, and provides an environment for intuitive comparison and discussion in three-dimensional space. This achieves profound and multifaceted technical effects, including improved efficiency and sophistication in urban infrastructure configuration planning, cost reduction, enhanced disaster response capabilities, future prediction and response, real-time on-site feedback, composite optimization of multiple infrastructure categories, increased resident satisfaction, optimized utilization costs, and flexible response to urban growth. Specifically, this system constructs a distributed computing environment where each processing unit operates independently as a microservice architecture. The system integrates the petabyte-level heterogeneous mixed data collected from the entire city into a high-dimensional tensor. This high-dimensional tensor is input into a deep neural network of the generative AI processing unit. The generative AI processing unit extracts latent feature vectors from the input high-dimensional tensor. Using these latent feature vectors, the generative AI processing unit outputs the infrastructure installation suitability probability distribution for each grid. This system determines the optimal candidate locations based on this probability distribution. It completely eliminates existing human-heuristic, rule-based selection methods. This system employs a data-driven, non-intuitive process based on machine learning models. Through this non-intuitive process, the system identifies non-linear correlations between multiple variables that are indistinguishable to humans. For example, the system learns the hidden correlation between subtle changes in traffic flow under specific weather conditions and local peaks in communication traffic. Using the learning results, the system predicts future traffic demand and calculates the optimal base station configuration. This process saves on wasted computational resources, significantly improving the overall computational efficiency of the system. This system is applicable to use cases such as optimizing roadside unit configurations for autonomous vehicles or planning the deployment of distributed power sources in smart grids. In these use cases, the system delivers a crucial technical advantage: maximizing coverage while mitigating over-investment in infrastructure.

[0084] (Implementation steps) Step 1: Urban Grid Segmentation and Basic Data Collection For each grid cell in a city that has been divided into grids at arbitrary resolutions such as 10 meters, 25 meters, 50 meters, and 100 meters, the system collects various data from sources including Geographic Information Systems (GIS), satellite imagery, drone aerial photography, traffic sensors, public transportation operation data, urban infrastructure management records, field survey data, resident reports, and various administrative databases. This data includes elevation data, population density data, traffic flow data, disaster risk data, power grid stability data, existing infrastructure data, land use classification data, foundation strength data, event information, public transportation utilization data, urban development plan data, laws, regulations, landowner information, setup cost data, maintenance and management cost data, resident survey results, user feedback data, and post-implementation performance data. Specifically, the system discretizes the city's geographic space into a three-dimensional voxel grid space. Each voxel is assigned a unique spatial index. The system associates the collected data with this spatial index to generate a multi-dimensional feature vector. For example, the system obtains elevation data as a floating-point number. This system acquires land use zoning data as a classification vector after one-hot encoding. It acquires traffic flow data as a time-series numerical array. Combining these heterogeneous data, it constructs, for example, a 256-dimensional initial feature vector for each voxel. It uses a convolutional neural network to extract building outlines or heights from drone aerial images and converts them into numerical data. It inputs legal and regulatory information described in natural language into a large-scale language model and converts it into semantic embedding vectors. It concatenates these converted data with the initial feature vector. In this way, the system transforms the physical and social characteristics of the city into a fully digital tensor representation. Through this tensor processing, it provides a data structure that subsequent machine learning models can directly compute. This system eliminates the need for existing human-based visual verification of maps. By automating the data collection and structuring process, it dramatically improves the efficiency of data management. In use cases such as wide-area urban development simulation or disaster prediction modeling, this system significantly reduces the computation time required for data preparation.

[0085] Step 2: Data Preprocessing and Normalization For the collected data, normalization is performed based on minimum and maximum values, with higher evaluations indicated by lower altitude, higher population density, greater traffic flow, lower disaster risk, more stable power supply, and absence of existing infrastructure. This process organizes data from different units or scales into a state that can be compared under the same benchmark. Specifically, the system applies a feature scaling algorithm to the multidimensional feature vectors associated with each voxel. The system analyzes the distribution of each feature and removes or prunes outliers using statistical methods. For example, the system applies minimum-maximum scaling to altitude data, linearly transforming the values ​​to the range of 0 to 1. For data that follows a log-normal distribution, such as population density or traffic flow data, the system standardizes the data after performing a logarithmic transformation. To unify the directionality of the evaluation, the system performs a process that inverts the values ​​of specific dimensions. For example, for disaster risk indicators, the system transforms the values ​​by subtracting them from 1, with higher values ​​indicating lower risk. For categorical data, the system uses an embedding layer to convert it into a dense vector of continuous values. After these preprocessing steps, the system generates a normalized feature tensor with all dimensions having the same scale and evaluation direction. This system significantly improves the convergence speed of machine learning models using gradient descent through its normalization process. It prevents overestimation of specific variables due to varying scales, thus stabilizing the model's inference accuracy. It guarantees the accuracy of distance calculations in high-dimensional spaces. In use cases such as environmental monitoring systems that comprehensively analyze heterogeneous mixed data acquired from multiple different sensor networks, or nursery evaluation systems in smart agriculture, this system delivers improved data quality and optimized computational efficiency.

[0086] Step 3: Setting up candidate site scoring and extraction based on generative AI Normalized data is input into the generative AI processing unit. While dynamically adjusting weights based on the purpose and category of urban infrastructure installation, a comprehensive score is calculated, and locations with high scores or exceeding predetermined thresholds are selected as candidate sites. The generative AI also imports past installation performance data, feedback from on-site personnel, future population growth forecasts, disaster risk information, installation cost information, operational cost data, maintenance and management history, number of user complaints, resident survey results, and user feedback data as learning data to continuously improve the accuracy of candidate site extraction. Specific examples of prompts for the generative AI could include: "Comprehensively analyze the following city map containing elevation data, population density data, traffic flow data, disaster risk data, power grid stability data, and data on the presence or absence of existing infrastructure to extract the top 10 most suitable locations for disaster prevention outposts. Output coordinates, elevation, population density, traffic flow, disaster risk, power grid stability, presence or absence of existing infrastructure, and comprehensive score for each location." or "Focusing on traffic flow and population density, extract the most suitable candidate sites for shared mobility hubs." Specifically, this generative AI processing unit is installed as a deep learning model with a Transformer-based encoder-decoder architecture. This generative AI processing unit receives input by combining a normalized feature tensor with the embedding vector of the natural language prompt. For example, it takes as input a prompt vector containing the top 10 most suitable locations for disaster preparedness site setup, along with a 256-dimensional feature tensor for each voxel. The generative AI processing unit uses a self-attention mechanism to calculate the correlation between the prompt's intent and each feature, generating dynamic attention weights. It then performs purpose-specific weighting by multiplying these attention weights onto the feature tensor. After passing through a multilayer perceptron, the generative AI processing unit outputs a site suitability score ranging from 0 to 1 for each voxel. As output, the generative AI processing unit generates structured JSON or list-formatted data, such as coordinates (XYZ), a comprehensive score of 0.95, and the main factors being low disaster risk and high population density. The system sorts this output score in descending order and extracts the top N voxels as site preparation candidates. This system forwards the extracted candidate site coordinate data to the subsequent visualization module. The generative AI processing unit continuously updates the model's weight parameters using comparative learning, treating past successful setups as positive examples and failed setups as negative examples. Through this dynamic attention mechanism, the system dramatically improves its versatility and adaptability for various setup purposes compared to existing methods using fixed weighting coefficients. This system enables data-driven decision-making that does not rely on human intuition. In use cases such as optimizing electric vehicle charging station configurations or maximizing sales volume in vending machines, this system delivers the technical benefits of maximizing return on investment and reducing planning time.

[0087] Step 4: Set up detailed data generation and organization for candidate sites For the extracted candidate sites, detailed data is compiled, including coordinates, altitude, population density, traffic flow, disaster risk, power grid stability, presence or absence of existing infrastructure, comprehensive score, surrounding environment information, cost information, future population growth forecasts, consistency with urban development plans, relationship with disaster evacuation routes, landowner information, laws and regulations, maintenance and management costs, resident satisfaction, and user convenience improvement effects. This data is then sent to the visualization department. Specifically, the system uses the spatial index of each extracted candidate site as the key to query a distributed database. This query retrieves all raw and predicted data associated with each candidate site. The system integrates the acquired data to construct a hierarchical data object for each candidate site. For example, the system inputs future population growth forecast data into an LSTM network used as a time series prediction model to generate a population shift vector for the next 10 years. The system uses a graph neural network to analyze the relationship with disaster evacuation routes, calculating the shortest path length or path safety score from the location to the nearest shelter. The system appends these high-level analytical results as attributes of the data objects. This system inputs textual data such as cost information or legal and regulatory information into a natural language generation model to generate human-readable summary text. This summary text is also stored in a data object. The system serializes the completed data object into JSON format. This serialized data is then sent to the visualization department via a high-speed RESTful API. Through this data integration and organization process, the system consolidates vast amounts of scattered information into a single structured format. This reduces the rendering processing load in the visualization department and minimizes data retrieval latency. In use cases such as land use evaluation systems in real estate development or support systems for large-scale commercial facility opening plans, this system delivers the technical benefits of reducing information collection costs and enabling multi-faceted evaluations.

[0088] Step 5: Discussion on 3D visualization and on-site comparison based on AR / VR The visualization department uses detailed data on candidate sites to overlay and display their location, recommendation level, various scores, and detailed attributes on a 3D AR or VR platform. This enables features such as pin displays or pop-up displays on camera images from tablets, multi-layered visualization on a 3D city model based on VR glasses, and simultaneous comparison of multiple candidate sites. This allows on-site managers or city planning officials to intuitively and comprehensively compare and discuss these sites. Specifically, the visualization department utilizes the GPU of edge devices to execute advanced graphics pipelines. It deserializes the received detailed JSON data and maps it to the attributes of virtual objects in 3D space. In AR mode, it applies the SLAM algorithm to real-time image frames acquired from the terminal's camera, simultaneously performing self-position estimation and environmental mapping. Using the estimated camera pose matrix, it overlays virtual pin objects or information panels onto the correct 3D coordinates in real space. In VR mode, it loads the city's overall 3D point cloud data or polygon model into memory and renders the scene from the perspective of a virtual camera. This visualization unit dynamically adjusts the luminescence intensity or particle effect density of virtual objects based on their overall scores. It calculates the user's gaze vector or the controller's ray projection to determine intersections with virtual objects. Upon detecting an intersection, the visualization unit displays the detailed attribute vectors of the candidate location as a pop-up window. Through this 3D visualization process, the system clarifies spatial context and positional relationships with the surrounding environment, which are difficult to grasp in 2D maps or spreadsheet software. This system expands human spatial cognitive abilities and facilitates the understanding of complex multidimensional data. In use cases such as underground piping planning simulation or landscape simulation, this system provides the technical benefits of early detection of design errors and rapid consensus-building among stakeholders.

[0089] Step 6: Collection and feedback of on-site information On-site supervisors input new obstacles, population changes, event information, resident requests, on-site observation data, photos, and user activity logs from terminals. This information is reflected in real-time to the generative AI on the server side, updating the evaluation or extraction results of candidate sites accordingly. A specific example of a prompt for the generative AI could be, "Consider the additional information obtained on-site and recalculate the evaluation of candidate sites. Reflecting new obstacles, population changes, event information, and resident requests, re-extract the best candidate sites." Specifically, this system receives unstructured data input through a client application running on an edge terminal. The system feeds the input on-site photos into a convolutional neural network, outputting bounding boxes and category labels for objects within the image. For example, the system obtains inference results such as category "temporary tent," confidence level 0.92, and XY coordinates. The system also feeds the input text comments into a natural language processing model for keyword extraction and sentiment analysis. Finally, the system converts these parsed results into structured feedback vectors. This system combines the feedback vector with a prompt vector that recalculates the evaluation of candidate sites based on additional information acquired on-site. The system sends this combined tensor to the generative AI processing unit on the server side. This generative AI processing unit receives this new tensor as input and dynamically updates the existing feature tensor. Using the updated feature tensor, the generative AI processing unit re-executes the forward pass and recalculates the suitability score for each voxel. Based on the recalculated scores, the generative AI processing unit outputs a new list of the top N candidate sites. This series of processes is completed within seconds, asynchronously updating the terminal display. Through this online learning and dynamic inference mechanism, the system eliminates latency caused by batch processing. The system demonstrates extremely high adaptability to dynamic environmental changes on-site. In use cases such as temporary infrastructure configuration during disaster recovery or security personnel configuration optimization during large-scale events, this system provides the technical effect of immediate responsiveness to changing conditions and refined planning.

[0090] Step 7: Finalize and continuously optimize the plan Based on the results of comparisons and discussions within the AR / VR visualization department or on-site information, the on-site manager or supervisor ultimately determines the optimal installation candidate site. After installation, the system continuously collects data on usage performance, resident feedback, maintenance and management costs, and improvements in user convenience. Through generative AI, this usage performance data is continuously learned, enabling automatic future installation plans, re-evaluation of existing infrastructure, additions, demolition proposals, and usage improvement proposals. Specifically, the system receives the coordinate data and confirmation markers of the final selected installation location via the user interface. This confirmation data is stored in the database as a correct answer label. After infrastructure installation, the system continuously collects time-series telemetry data such as uptime, power consumption, and fault history from the IoT sensor network installed at the target location. The system aggregates this telemetry data within a specific time window to generate an usage performance vector. The system uses a mean squared error function to calculate the error between this usage performance vector and the predicted score before installation. Using this error signal, the system fine-tunes the weight parameters of the neural network in the generative AI processing unit through backpropagation. Through this reinforcement learning method, the system autonomously improves the model's prediction accuracy. This system periodically performs batch re-evaluation of the entire city grid. When the performance vector of existing infrastructure falls below a predetermined threshold, the system outputs an alert signal proposing the demolition or relocation of that infrastructure. For areas where demand forecast vectors exceed supply capacity, the system outputs recommended coordinates for additional installations. Through this continuous learning cycle, the system optimizes infrastructure throughout its entire lifecycle. This system evolves existing maintenance plans that rely on rules of human experience into a data-driven predictive maintenance system. In use cases such as leak detection and upgrade programs in water supply networks or route reorganization for public transportation, this system delivers dramatic cost reductions and continuous improvements in service quality.

[0091] (specific use case) For example, in the case of applying this invention to the establishment of new disaster prevention outposts in cities, the urban infrastructure data acquisition department divides the entire city into 25-meter square grids. For each grid, elevation data is acquired from geographic information systems, satellite imagery, and drone aerial photography data for the assessment of flood risk or foundation subsidence risk. Population density data is acquired from census data, location information from portable terminals, traffic flow sensors, data on the number of people entering commercial facilities, data on the number of attendees from event organizers, and data on the number of Wi-Fi access points connected, etc., to grasp population changes or peak traffic flows on weekdays and weekends, day and night, during event periods, and each season. Traffic flow data is acquired from road sensors, the number of vehicles passing through intersections, vehicle GPS data, the number of passengers getting on and off public transportation, the number of users at bus stops or stations, utilization data of shared mobility services, road congestion data, and traffic accident records, etc., to identify places with high traffic of people or vehicles, places with frequent traffic congestion, and places with high risk of traffic accidents, such as major roads or intersections, around stations, and around commercial facilities. Disaster risk data is obtained from past disaster records, disaster risk maps, meteorological data, foundation information, soil classification data, flood and debris flow disaster risk data, earthquake prediction data, tsunami inundation scenario data, fire occurrence records, shelter location information, and emergency vehicle route information, etc., to evaluate the risks of earthquakes, floods, debris flows, fires, tsunamis, storm surges, storms, and lightning strikes from multiple perspectives. Power grid stability data is obtained from power company supply network information, past power outage records, renewable energy adoption status, the presence or absence of distributed power sources, supply stability during peak electricity demand, transmission line and substation location information, and the aging status of power infrastructure, etc., to evaluate locations where stable power supply can be foreseen, locations with low power outage risk, and locations where localized consumption of renewable energy is possible. The system acquires data on existing infrastructure from sources such as urban infrastructure management ledgers, on-site survey data, image analysis from drones or autonomous vehicles, operational status data from sensor networks, and resident reports. This data is used to determine whether the same type of infrastructure has already been installed, the presence of aging or faulty infrastructure, utilization rates, and maintenance status. In addition, it acquires land use classification data, foundation strength data, event information, public transportation utilization data, urban development plan data, legal and regulatory information, landowner information, installation cost data, maintenance cost data, resident survey results, user feedback data, and post-implementation performance data, which are then used for multi-faceted optimization of urban infrastructure configuration. Specifically, the urban infrastructure data acquisition department sends asynchronous HTTP requests to multiple external API endpoints to collect petabyte-level raw data in parallel. This raw data is then stored in a data lake on a distributed file system.The city's infrastructure data acquisition department uses distributed processing frameworks such as Apache Spark to perform ETL processing on the raw data within the data lake. For example, it extracts loitering points from location logs of portable terminals using the DBSCAN clustering algorithm, generating population density tensors for each time period. It also generates traffic congestion probability maps from vehicle GPS trajectory data using Hidden Markov Models. Furthermore, it performs hydrodynamic simulations from past meteorological and topographic data to calculate flood depth prediction vectors for each grid cell. Using a 25-meter square spatial index as the key, the department combines this highly processed data to construct, for example, ultra-high-dimensional feature vectors of 1024 dimensions for each grid cell. This system generates datasets containing dynamic and predictive features rather than simply static statistical data. Through this data integration and processing, the system eliminates the silos of heterogeneous and mixed data, providing a unified format that machine learning models can directly interpret. This system significantly reduces the computational resources and time required for data preprocessing. In use cases such as the construction of integrated command centers in smart cities or the dynamic formulation of wide-area evacuation plans, this system enables data-driven, objective, and rapid decision-making.

[0092] These diverse data are input into the generative AI processing unit. For example, the unit uses a prompt to "comprehensively analyze the following city maps, including elevation data, population density data, traffic flow data, disaster risk data, power grid stability data, and data on the presence or absence of existing infrastructure, to extract the top 10 most suitable locations for setting up disaster prevention outposts. For each location, output coordinates, elevation, population density, traffic flow, disaster risk, power grid stability, presence or absence of existing infrastructure, and a comprehensive score." The weighting is dynamically adjusted based on factors such as lower elevation, higher population density, higher traffic flow, lower disaster risk, more stable power supply, and the absence of existing infrastructure, resulting in a higher evaluation. The generative AI also imports past setup performance data, feedback from on-site personnel, future population growth predictions, disaster risk information, setup cost information, operational cost data, maintenance and management history, number of user complaints, resident survey results, and user feedback data as learning data to continuously improve the accuracy of candidate site extraction. Specifically, this generative AI processing unit employs a hybrid architecture integrating large-scale language models and graph neural networks. This generative AI processing unit segments the input prompt text into a sequence of tokens using a word segmenter, and then converts it into a high-dimensional prompt vector through an embedding layer. The unit constructs a city spatial graph, using 1024-dimensional feature vectors of each grid as nodes and edges connecting adjacent grids. It applies graph convolution to this city spatial graph to generate node embedding vectors that consider spatial dependencies. The unit then implements a cross-attention mechanism, treating the prompt vector as a query and the node embedding vectors as keys and values. Through this cross-attention, the unit dynamically assigns high attention weights to the features that best match the prompt's content. This weighted vector is then input into a fully connected layer, where a sigmoid function outputs the overall score of each grid as a probability value between 0 and 1. The unit generates a structured JSON object as output, for example, with coordinates XY, an overall score of 0.98, and the main factors being high power grid stability and low disaster risk. This generative AI processing unit calculates cross-entropy error using the coordinates of past successes as positive example labels and updates the model weights using gradient descent. Through this hybrid model, the system achieves high-level inference by simultaneously understanding spatial context and complex linguistic indications. The system flexibly adapts to various setup conditions that fixed algorithms cannot handle. In use cases such as ambulance standby optimization or drone port configuration planning, this system delivers a significant leap in inference accuracy and automation of planning.

[0093] The candidate site selection department, based on the comprehensive score calculated by the generative AI processing department, extracts the top 10 locations as candidate sites. It then compiles detailed data on each location, including coordinates, altitude, population density, traffic flow, disaster risk, power grid stability, presence of existing infrastructure, comprehensive score, surrounding environment information, cost information, future population growth projections, consistency with urban development plans, relationship with disaster evacuation routes, landowner information, laws and regulations, maintenance and management costs, resident satisfaction, and user convenience improvements. The visualization department uses this detailed data to overlay and display the location, recommendation level, various scores, and detailed attributes of the candidate sites in AR or VR 3D space. This includes features such as pin-like or pop-up displays on tablet terminal camera images, multi-layered visualization on a VR-based 3D city model, and simultaneous comparison of multiple candidate sites. This allows on-site managers or urban planning officials to intuitively and comprehensively compare and discuss these sites. Specifically, the candidate site selection department applies a quicksort algorithm to the full-grid score array output from the generative AI processing department, rearranging it in descending order. From this sorting result, the department extracts the top 10 indices. The candidate site selection unit uses an index to retrieve detailed attribute vectors for each location from a distributed database. This unit serializes the attribute vectors into a JSON-formatted payload and sends it to the visualization unit. The visualization unit parses the received payload and inputs it into the 3D rendering engine. The visualization unit generates virtual anchor points on the city's 3D base map, corresponding to the coordinates of each candidate site. The visualization unit places cylindrical 3D virtual objects with heights corresponding to the overall score on these anchor points. The visualization unit generates polygons for indicators such as disaster risk or population density as radar charts and displays them floating above the virtual objects. The visualization unit acquires the user's head tracking data in the VR environment and updates the rendering matrix of the viewpoint camera at a rate of over 60 frames per second. When the user points to a specific candidate site with the controller, the visualization unit performs raycast-based collision detection and dynamically instantiates an information panel containing detailed text data. This system, through advanced 3D graphics processing, transforms massive amounts of numerical data into a visually intuitive format. This system supports human cognitive processes, making it easier to understand complex trade-offs. In use cases such as resident briefings or infrastructure investment meetings in urban redevelopment projects, this system brings about the technical benefits of increased efficiency in information sharing and faster decision-making.

[0094] On-site supervisors input new obstacles or population changes, event information, resident requests, on-site observation data, photos, and user activity logs from terminals. This information is reflected in real-time to the generative AI on the server side. For example, prompts such as "Consider the additional information obtained on-site and recalculate the evaluation of candidate sites. Reflect new obstacles or population changes, event information, and resident requests, and re-extract the best candidate sites" are used to update the evaluation or extraction results of candidate sites in real time. Ultimately, the on-site supervisor or manager determines the best candidate site based on the results of comparisons and discussions in the AR / VR visualization department or on-site information. After installation, the system continuously collects usage performance, resident feedback, maintenance and management costs, and improvements in user convenience. Through continuous learning of these usage performance data by the generative AI, it can automatically conduct future installation plans or re-evaluate existing infrastructure, add installations, propose demolition, and propose usage improvement. Specifically, this system sends new data obtained from the sensor clusters at the edge terminals to the server via a stream processing pipeline. This system, for example, inputs images captured on-site into a convolutional neural network, outputting a category label for "unregistered building under construction" and a 3D bounding box. The system then adds the coordinates of this new obstacle to the attribute vector of the corresponding node in the city spatial map. The updated city spatial map and recalculated cue vectors are input into the generative AI processing unit. This unit performs graph convolution operations and cross-attention mechanisms again, recalculating the suitability score of each node in milliseconds. Based on this recalculation result, the system asynchronously updates the 3D display in the visualization unit. When a manager selects a specific candidate location on the terminal and presses the confirmation button, the system submits the location's coordinates and feature vectors as the correct answer data to the learning database. The system inputs time-series data, such as operation rate or power consumption, collected from IoT sensors after infrastructure operation into a recurrent neural network to calculate a performance score. This performance score is used as a reward signal to apply a reinforcement learning algorithm to update the policy network of the generative AI processing unit. Through this closed-loop feedback system, the system autonomously improves the model's inference accuracy over time. This system demonstrates robust resilience against environmental changes over time and unexpected disturbances. In use cases such as optimal battery configuration in smart grids or base station lifecycle management in next-generation communication networks, it offers the crucial benefits of minimized operational costs and enhanced overall system resilience.

[0095] 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.

[0096] Data generation model 58 is what is known as generative AI (Artificial Intelligence).

[0097] One example of data generation model 58 includes ChatGPT (registered trademark) (Internet search).<URL: https: / / openai.com / blog / chatgpt> Generative AI, etc.

[0098] 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.

[0099] 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 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. In addition, 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.

[0100] 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.

[0101] [Second Implementation] Figure 3 An example of the configuration of the data processing system 210 according to the second embodiment is shown.

[0102] 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.

[0103] 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).

[0104] 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.

[0105] 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.

[0106] 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).

[0107] 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.

[0108] 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.

[0109] 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.

[0110] 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).

[0111] 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.

[0112] 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".

[0113] (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.

[0114] (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.

[0115] 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.

[0116] 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.

[0117] 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.

[0118] [Third Implementation] Figure 5 An example of the configuration of the data processing system 310 according to the third embodiment is shown.

[0119] 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.

[0120] 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).

[0121] 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.

[0122] 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.

[0123] 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).

[0124] 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.

[0125] 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.

[0126] 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.

[0127] 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.

[0128] In the head-mounted terminal 314, the processor 46 performs the receiving and output processing.

[0129] 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.

[0130] 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".

[0131] (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.

[0132] (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.

[0133] 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.

[0134] 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.

[0135] 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.

[0136] [Fourth Implementation] Figure 7 An example of the configuration of the data processing system 410 according to the fourth embodiment is shown.

[0137] 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.

[0138] 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).

[0139] 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.

[0140] 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.

[0141] 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).

[0142] 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.

[0143] 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.

[0144] 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.

[0145] 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.

[0146] 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.

[0147] 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.

[0148] 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".

[0149] (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.

[0150] (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.

[0151] 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.

[0152] 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.

[0153] 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.

[0154] 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.

[0155] Figure 9 This 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.

[0156] 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.

[0157] 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).

[0158] 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.

[0159] 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."

[0160] 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.

[0161] 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).

[0162] 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.

[0163] 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.

[0164] 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.

[0165] 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.

[0166] 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).

[0167] Every processor has built-in or connected memory, and every processor performs specific processes by using memory.

[0168] 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.

[0169] 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.

[0170] 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.

[0171] 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.

[0172] 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.

[0173] In addition, the following notes further explain the above explanation.

[0174] (Note 1) A system comprising an urban infrastructure data acquisition unit, a generative AI processing unit, a candidate site selection unit, and a visualization unit, characterized in that the urban infrastructure data acquisition unit collects various information for each grid within the city, including altitude data, population density data, traffic flow data, disaster risk data, power grid stability data, and data on the presence or absence of existing infrastructure, from sources such as geographic information systems, satellite imagery, drone aerial photography, population census data, traffic sensors, meteorological databases, and urban infrastructure management ledgers. The generative AI processing unit normalizes the acquired data and dynamically adjusts the weighting based on the purpose and category of urban infrastructure installation, calculating a comprehensive score based on factors such as lower altitude, higher population density, higher traffic flow, lower disaster risk, more stable power supply, and the absence of existing infrastructure. Furthermore, past performance data, feedback from on-site managers, future population growth forecasts, disaster risk information, and setup cost information are also imported as learning data. Generative AI, which continuously improves the extraction accuracy of candidate sites, is used to perform the processing. The candidate site selection unit extracts locations with high scores or those exceeding a predetermined threshold as candidate sites based on the comprehensive score calculated by the generative AI processing unit. Detailed data such as coordinates, altitude, population density, traffic flow, disaster risk, power grid stability, presence or absence of existing infrastructure, comprehensive score, surrounding environment information, disaster risk, and cost information are compiled for each candidate site. The visualization unit intuitively displays the extracted candidate site information in a three-dimensional space in AR or VR, so that on-site managers or urban planning managers can compare and discuss them on-site or in virtual space.

[0175] (Note 2) As described in Appendix 1, the urban infrastructure data acquisition unit collects various data for each grid after the city is subdivided, including not only altitude, population density, traffic flow, disaster risk, power grid stability, and the presence or absence of existing infrastructure, but also urban development plan data, land use classification data, foundation strength data, soil classification data, flood and debris flow disaster risk data, public transportation utilization data, renewable energy introduction status, past disaster history, power outage history, event information, and on-site survey data. These data are then input into the generative AI processing unit.

[0176] (Note 3) As described in Appendix 1, the visualization unit displays detailed information such as the three-dimensional coordinates, altitude, population density, traffic flow, disaster risk, power grid stability, presence or absence of existing infrastructure, comprehensive score, surrounding environment information, and cost information of each candidate location extracted by the candidate location selection unit in a three-dimensional space. This allows users to intuitively compare and discuss the multiple candidate locations. The system also has the function of displaying detailed information of any candidate location when the user selects it. Generative AI is used to implement this series of processes.

Claims

1. A system comprising: The system includes a terrain data acquisition department, a population density data acquisition department, and a radio wave propagation characteristics data acquisition department; and a generative AI processing department that comprehensively analyzes these various types of data. The system comprises a candidate site selection unit that extracts candidate sites for wireless base station deployment based on the analysis results; and an AR visualization unit and a VR visualization unit that visualize the extracted candidate site information in three-dimensional space. The system is characterized by: a terrain data acquisition unit acquiring terrain information such as altitude or slope from a geographic information system or external database; a population density data acquisition unit acquiring population density information for each location from statistical data or sensor data; a radio wave propagation characteristic data acquisition unit acquiring information about radio wave attenuation or propagation characteristics at each location from simulated data or measured values; a generative AI processing unit normalizing the acquired terrain data, population density data, and radio wave propagation characteristic data, calculating a comprehensive score after weighting each element; a candidate site selection unit extracting candidate sites based on the comprehensive score; and an AR visualization unit and a VR visualization unit visually displaying the coordinates, altitude, population density, radio wave propagation characteristics, comprehensive score, and other information of the extracted candidate sites in three-dimensional space.

2. The system as described in claim 1, characterized in that, The generative AI processing unit performs normalization processing based on minimum and maximum values ​​for each of the terrain data, population density data, and radio wave propagation characteristic data. For terrain data, the lower the altitude, the higher the evaluation; for population density data, the higher the evaluation; and for radio wave propagation characteristic data, the lower the evaluation, the higher the evaluation. It then performs individual weighting of each element and performs the process of summing these values ​​to calculate a comprehensive score.

3. The system as described in claim 1, characterized in that, The AR visualization unit and VR visualization unit display detailed information such as three-dimensional coordinates, altitude, population density, radio wave propagation characteristics, and comprehensive score of each candidate location in three-dimensional space for multiple candidate locations extracted by the candidate location selection unit, so that users can intuitively compare and discuss them. They also have the function of popping up and displaying detailed information of any candidate location when the user selects it. Generative AI is used to achieve this series of processes.

Images