Data processing system

Abstract

The present application relates to a kind of electric wave countermeasure case automatic processing system, for the electric wave countermeasure case from legal customer or business performer, with case information acquisition unit, generative AI analysis unit, case determination unit, case classification unit, construction company selection unit, contract processing unit, main performer distribution unit etc., obtain building information or field photo, drawing, customer requirement and other various information, use generative AI and past similar case database comparison, analysis, so that the information of each case is not perfect determination, the inference of best countermeasure method, case classification, automatic selection of construction company and contract, automatic allocation of main performer, progress management, history management is automatically integrated until, realize business efficiency, best allocation of human resources, cost reduction and customer satisfaction improvement.

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, in the large number of radio wave countermeasures cases from corporate customers or sales representatives, the past practice of having the responsible party confirm information for each case (either excessive or insufficient) and assigning a lead responsible party based on the source department or corporation, followed by manual discussion of countermeasures or subcontracting to construction companies, resulted in significant time and labor costs for case processing. This prevented sufficient resources from being allocated to key or complex cases, necessitating the outsourcing of some judgment or assignment work. Consequently, this led to increased business costs, processing delays, and decreased customer satisfaction. Furthermore, regarding the standard radio wave countermeasures cases, which constitute the majority of cases, although patterns could be established based on past countermeasure history or on-site conditions, the reliance on human judgment or manual operation hindered efficiency and automation, and also presented challenges in effectively utilizing human resources. Additionally, the lengthy delivery time from application to subcontracting due to incomplete application information or the time required for case content confirmation further exacerbated the problem of failing to meet customers' demands for rapid response. To address these challenges, this invention aims to automate a series of business processes with high speed and precision by leveraging generative AI, including automatic acquisition of case information, automatic determination of incomplete information, automatic comparison with past history, automatic case classification, automatic selection and subcontracting of construction companies, and automatic allocation of the main responsible party. This will significantly improve business efficiency, optimize the allocation of human resources, reduce costs, and enhance customer satisfaction. Summary of the Invention

[0005] The means to solve the problem involved in this invention is to construct an automated radio countermeasure case processing system comprising a case information acquisition unit, a case determination unit, and a case processing unit. First, the case information acquisition unit electronically receives case information from customers or sales personnel, including building information, floor information, surrounding base station information, drawings, and personnel information, and stores this information in a database. Next, the case determination unit uses generative AI to compare the received case information with a historical radio countermeasure database, analyzing multiple factors such as the building or floor, the number of surrounding base stations, past countermeasure patterns, and the radio countermeasure status of surrounding buildings. This automatically determines whether the application information is incomplete and automatically categorizes the case as a simple case that can be resolved by setting up a standard amplifier or small base station, or a complex case requiring a special amplifier or a newly installed base station.

[0006] Based on the judgment results of the case assessment department, the case handling department uses generative AI to automatically select a construction company and process the contracting for simple cases; it automatically assigns a principal responsible party for complex cases; and it automatically sends a rejection notice to the applicant if the case information is incomplete. This high-speed and high-precision automation of the entire business process from case acceptance to judgment, allocation, and contracting improves business efficiency, optimizes human resource allocation, reduces costs, and enhances customer satisfaction. Attached Figure Description

[0007] Figure 1 This is a conceptual diagram illustrating an example of the configuration of a data processing system according to the first embodiment.

[0008] 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.

[0009] Figure 3 This is a conceptual diagram illustrating an example of the data processing system configuration in the second embodiment.

[0010] 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.

[0011] Figure 5 This is a conceptual diagram illustrating an example of the data processing system configuration in the third embodiment.

[0012] 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.

[0013] Figure 7 This is a conceptual diagram illustrating an example of the data processing system configuration in the fourth embodiment.

[0014] Figure 8 This 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.

[0015] Figure 9 It represents an emotion graph that maps multiple emotions.

[0016] Figure 10 It represents an emotion graph that maps multiple emotions.

[0017] Explanation of reference numerals in the attached figures Data processing systems 10, 210, 310, and 410 12 Data processing device 14 Smart devices 214 Smart Glasses 314 Head-mounted terminal 414 Robot. Detailed Implementation

[0018] Hereinafter, an example of an implementation of the system involved in this disclosure will be described with reference to the accompanying drawings.

[0019] First, let's explain the terms used in the following description.

[0020] 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.

[0021] 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.

[0022] 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.

[0023] 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).

[0024] 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.

[0025] [First Implementation] Figure 1 An example of the configuration of the data processing system 10 according to the first embodiment is shown.

[0026] 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.

[0027] 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).

[0028] 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.

[0029] 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.

[0030] 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.

[0031] 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.

[0032] Figure 2 An example of the main functions of the data processing device 12 and the smart device 14 is shown.

[0033] 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.

[0034] 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).

[0035] 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.

[0036] In addition, other devices besides the data processing device 12 may also have a data generation model 58.

[0037] For example, a server device (e.g., a generation server) may have a data generation model 58.

[0038] 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.

[0039] (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".

[0040] (Implementation methods for carrying out the present invention) The embodiments of the present invention will be described in more specific and detailed terms.

[0041] The automatic case processing system for radio wave countermeasures of this invention consists of a server and a terminal. The terminal is a personal computer, tablet, smartphone, etc., used for customers or sales personnel to input case information. The server is a high-performance computer system responsible for core processing such as summarizing, analyzing, judging, processing, historical management, and executing generative AI models of case information. The terminal and server can communicate via a network, and information input from the terminal is sent to the server in real time. Specifically, this system adopts a distributed computing architecture. The server consists of multiple computing nodes. Each computing node has dedicated computing circuits for parallel execution of large-scale language models or image recognition models. The terminal has a preprocessing unit that converts input unstructured data into structured data. This preprocessing unit converts input text into word vectors. This preprocessing unit converts images into multidimensional tensors. The terminal sends the converted data to the server via a secure communication protocol. The server temporarily stores the received data in an in-memory database. This in-memory database enables high-speed data read and write. The server uses an asynchronous processing queue to distribute data to each processing module. Through this structure, the system minimizes processing latency even with a large number of concurrent accesses. This system effectively reduces communication load. This system is capable of real-time case processing. It can be applied to use cases such as communication infrastructure maintenance or equipment management in large-scale facilities.

[0042] The Case Information Acquisition Department receives case information from customers or business operators via a dedicated application on a terminal or through the input screen of a web browser. The input information is diverse, including building name, building address, building type (e.g., office building, commercial facility, factory, hospital, school, etc.), floor number, floor area, number of surrounding base stations, models of surrounding base stations, drawing data (PDF, CAD files, image files, etc.), on-site photos, operator information (name, contact information, department), information of the commissioning department, customer attributes (large corporation, small corporation, public organization, etc.), electromagnetic wave intensity measurements (e.g., -95dBm, -80dBm, etc.), desired response period, history of past countermeasures, special circumstances on-site (e.g., obstructions, basement, roof, special structures, etc.), and customer requests (e.g., "desired short delivery time," "desired nighttime construction," "desired simultaneous handling of multiple floors," etc.). For example, consider a scenario where a sales representative inputs detailed information from a terminal: "7th floor of office building A in Tokyo, floor area 500 square meters, three base stations nearby, attached site photos and drawings, representative Mr. B, client is a corporate sales department, customer is a large corporation, on-site electromagnetic field strength is -90dBm, request for completion within one month, located on a basement floor with many obstructions, previous experience installing amplifiers on the 5th floor of the same building, nighttime installation preferred." The terminal sends this information to a server, where the case information acquisition department stores it in a database. Specifically, the case information acquisition department uses natural language processing to parse the input text data. It uses named entity recognition algorithms to extract keywords from the text data. The extracted keywords are mapped to a multi-dimensional vector space. A convolutional neural network is applied to the image data. A feature map is generated from the image. Line segment information in vector form is extracted from the drawing data. The extracted line segment information is converted into graph structure data. The case information acquisition department integrates these heterogeneous data to generate a multimodal feature vector. This multimodal feature vector is, for example, represented as a 512-dimensional floating-point array. The case information acquisition department stores the generated vector in the storage area of ​​this server. Through this process, the system fully automates information collection operations that previously relied on manual visual verification. This system delivers a significant improvement in data input accuracy. This system detects missing or inconsistent input data at an early stage. This system is extremely useful in use cases such as pre-construction site investigations or real estate property appraisals.

[0043] The case assessment department operates on a server, utilizing generative AI to analyze case information. This generative AI combines multiple AI technologies, including natural language processing, image recognition, machine learning, and deep learning models. For example, regarding blueprints or site photos, image recognition AI automatically analyzes the content, extracting information such as building structure, obstructions, floor size, and available space. Regarding textual information, natural language processing AI detects omissions or inconsistencies, automatically identifying incomplete information such as "no attached blueprints," "no assigned personnel information," or "unknown number of surrounding base stations." Furthermore, the generative AI takes multiple factors as input, including building name, floor, number of surrounding base stations, past countermeasures history, radio wave countermeasures status of surrounding buildings, radio wave strength, floor height, customer attributes, client department information, special circumstances at the scene, and customer requirements, and retrieves similar past cases from the database, performing pattern recognition or statistical inference. For example, referencing historical examples such as "the problem was solved in the past by installing an amplifier on the 6th floor of the same building," "there are 3 surrounding base stations and many similar cases," "the customer attributes are the same and the problem was solved in the past by installing a small base station," and "the site is on a basement floor and there are cases requiring a special amplifier," the system accurately determines whether a case can be solved using standard countermeasures. Specifically, this case determination unit possesses a large-scale language model based on the Transformer architecture. This case determination unit receives multimodal feature vectors as input data. This input data is, for example, a 512-dimensional numerical vector. This input data contains an image tensor representing the building structure. This input data contains a text embedding vector representing the customer's requirements. This case determination unit uses a self-attention mechanism to comprehensively parse these input data. This case determination unit uses weight parameters pre-learned from past case history data to perform inference. This case determination unit outputs probability distribution data representing the success probability of each countermeasure as output. This output data is, for example, an array of values ​​such as amplifier set to 0.85, small base station set to 0.12, and new base station set to 0.03. Based on this probability distribution data, the case determination unit determines the optimal countermeasure. The case determination department calculates a reliability score for the determined countermeasures. This reliability score is then passed to subsequent classification processing. Through this process, the system eliminates fuzzy judgments based on human experience rules. This system leverages the technical effect of achieving objective and high-precision judgments through complex pattern recognition in high-dimensional space. This system can be applied to use cases such as optimizing communication network services or preventing and protecting infrastructure equipment.

[0044] The case assessment department categorizes cases into "simple cases" and "complex cases" based on the inference results of generative AI. For example, cases deemed to require only the installation of amplifiers or small base stations are classified as simple cases, while those requiring special amplifiers, new base stations, or complex on-site investigations are classified as complex cases. The generative AI automatically extracts patterns such as "the same countermeasures have been effective multiple times in the same building, on the same floor, with the same number of base stations, and with the same customer attributes," and records the basis for the judgment. Furthermore, case priority is also automatically determined by AI, considering conditions such as "the customer is a large corporation, and the on-site radio wave intensity is extremely low," "there is an urgent response request from the customer," and "the same customer has frequently filed complaints in the past," extracting high-priority cases to support optimal allocation of human resources. Specifically, this case assessment department uses machine learning algorithms such as support vector machines or random forests to perform classification processing. The department receives reliability scores and probability distribution data calculated in the previous stage as input. It plots the input data on a multi-dimensional feature space. The department uses pre-learned decision boundaries to classify cases into simple and complex cases. The case determination unit outputs binary labels representing the classification results. This output data, for example, is a value of 0 for simple cases or 1 for complex cases. The case determination unit also calculates a priority score. This priority score is output as a continuous value from 0.0 to 1.0. The case determination unit calculates the priority score by weighted summation of the customer's attribute vector or radio wave intensity numerical data. The case determination unit compares the calculated priority score with a threshold. The case determination unit registers cases with priority scores exceeding the threshold into a priority processing queue. Through this process, the system automates case classification. This system brings the technical effect of optimizing the allocation of computing and human resources. This system is extremely effective in use cases such as disaster recovery operations requiring emergency response or communication countermeasures during large-scale events.

[0045] The case handling department operates on a server and automatically executes processing based on the judgment results of the case adjudication department. In the case of simple cases, generative AI automates the selection of construction companies and the contracting process. When selecting a construction company, the generative AI comprehensively evaluates multiple parameters such as the construction company's past performance (e.g., "10 past projects in the same building", "experienced in setting up the same equipment"), service area (e.g., "serves the entire Tokyo area", "only serves the Kanto region"), availability (e.g., "can be handled this week", "can only be handled next month or later"), construction period (e.g., "standard construction period of 2 weeks", "can handle short delivery periods"), cost (e.g., "standard price", "additional fees"), quality evaluation (e.g., "high past customer reviews", "low complaint rate"), past dispute history (e.g., "many construction delays", "past construction errors"), and customer feedback (e.g., "attentive handling", "prompt reporting"), and automatically selects the best construction company. For example, AI analyzes information such as "Construction Company A has 10 past projects in the same building, with short project durations and high ratings" and "Construction Company B currently has available work, but has a history of delays," to determine the best construction company. Contracting data is automatically generated and sent to the construction companies' systems. For example, a contract is automatically sent to a construction company stating, "7th floor of Building A, amplifier setup, drawings included, project duration within 2 weeks, contractor Mr. B, nighttime work preferred." Furthermore, after contracting to the construction companies, progress management is also automatically handled by AI, including functions such as "automatically collecting progress reports from construction companies," "automatically detecting construction completion reports," and "automatically issuing alerts in case of delays." Specifically, the case processing department uses reinforcement learning algorithms to select construction companies. The department receives a profile vector representing the characteristics of each construction company as input. This profile vector contains numerical data such as the number of past construction projects, average project duration, and quality evaluation scores. The department calculates the cosine similarity between the requirement vector representing the project's specifications and the profile vector. The department selects the construction company with the highest similarity as the best candidate. This case processing department generates contracting data for the selected construction company. This contracting data is output as structured data in JSON format. It contains key-value pairs such as construction content, desired construction period, and URLs of attachments. The case processing department sends the generated contracting data to the construction company's system via a secure API. Upon completion of transmission, the case processing department receives a status code. Based on the status code, the case processing department updates the case status in the database. This process completely eliminates the need for manual contracting operations. This system effectively reduces contracting errors and significantly improves processing speed. This system can be widely applied to use cases such as client selection in supply chain management or logistics distribution industries.

[0046] On the other hand, in cases involving complexity, generative AI considers the lead specialist's expertise (e.g., "expert in amplifier setup," "experienced technician in base station setup"), workload (e.g., "number of cases currently being handled," "overtime hours"), past handling history (e.g., "experience handling cases in the same building," "performance on similar complex cases"), applicable region (e.g., "Kanto region handling," "nationwide handling"), and suitability in technical field (e.g., "specialist in indoor radio wave countermeasures," "specialist in outdoor base stations") to automatically assign the best lead specialist. For example, the AI ​​analyzes information such as "lead specialist C has handled many similar complex cases in the past and currently has a low workload" and "lead specialist D is not included in the assignment due to a different technical field" to select the best lead specialist. The case details are automatically communicated to the lead specialist's terminal, allowing them to conduct detailed on-site investigations, request assistance from the technical department, or discuss special amplifiers or new base station installations. For example, it can also handle complex business processes such as "adjusting on-site visit schedules for on-site investigations," "discussions with the technical department," "discussion of special amplifier specifications," and "administrative procedures for new base station installations." Specifically, this case handling department uses a graph neural network to perform the assignment of the lead person. The department represents the skill sets or current workload of technical personnel as nodes. It also represents the required skills for the case as another node. The department generates graph structure data, defining the relationships between nodes as edges. This graph structure data is then input into the graph neural network. The department outputs a fitness score for each technical personnel node. This fitness score is a floating-point number from 0.0 to 1.0. The department selects the technical personnel with the highest fitness score as the lead person. An identifier for the selected lead person is generated as output data. This identifier is used to retrieve contact information from the company's internal systems. The department automatically sends notification messages based on the retrieved contact information. Through this process, the system instantly performs optimal resource allocation that satisfies complex constraints. This system eliminates the personalization of tasks and improves the overall productivity of the organization. This system can be applied to use cases such as troubleshooting IT systems requiring high levels of expertise or assigning specialists in medical institutions.

[0047] Furthermore, when the case information is deemed incomplete, the generative AI automatically extracts the incomplete content, such as specific correction instructions like "insufficient drawing data," "no responsible person information entered," "unclear record of surrounding base station numbers," "unclear on-site photos," and "customer requirements not filled in," and returns them to the terminal. The applicant can make corrections on the terminal and reapply. For example, they can "re-attach drawing files," "enter responsible person information," and "retake and upload on-site photos." Specifically, this generative AI executes a rule-based verification algorithm and a deep learning-based anomaly detection algorithm in parallel on the input case information. This generative AI determines whether there are any required input items through logical operations. This generative AI uses context embedding vectors to evaluate the semantic consistency of the input text data. This generative AI applies an image quality evaluation model to the image data to calculate resolution or contrast values. This generative AI integrates these evaluation results and outputs a flag indicating the presence or absence of incompleteness. When incompleteness is detected, this generative AI generates an array of error codes representing the type and location of the incompleteness. This generative AI inputs this array of error codes into a natural language generation model. This generative AI outputs specific correction instructions that are easy for users to understand. The output text might be a string like, "Please re-upload due to low resolution of the drawing data." This generative AI sends this text to the terminal, where it is displayed as a warning on the user interface. Through this process, the system guarantees the quality of input data at the system entry point. This system dramatically reduces the error rate in subsequent processing. This system is extremely useful in use cases such as electronic application systems for administrative agencies or account opening procedures for financial institutions.

[0048] Furthermore, generative AI automatically records and accumulates the handling history, judgment basis, reasons for selecting the construction company, reasons for assigning the main responsible party, progress status, customer feedback, construction completion reports, and dispute occurrence history for each case. This data can also be utilized for future AI model relearning, business improvement, and audit response. For example, it can perform trend analysis such as "all cases in the same building and on the same floor were resolved by setting up amplifiers in the past year," "Construction company A's project delay rate is showing a downward trend," and "Main responsible party C has high customer satisfaction in handling cases." Furthermore, AI can use this historical data to automatically predict future case occurrences, conduct risk analysis, and propose business process optimizations. Specifically, this generative AI stores the accumulated historical data in a time-series database. It periodically initiates batch processing to relearn the model using historical data. It uses a loss function to calculate the error between past predictions and actual business results. It uses backpropagation to update the weight parameters of the neural network. It uses gradient descent for optimization. This generative AI executes a time-series forecasting model that uses a learned model to predict the number of future cases. This time-series forecasting model receives past case frequency or seasonal variations as input tensors. The output is a numerical array representing the predicted number of cases per week for the next month. Based on this forecast data, the generative AI pre-allocates the necessary computational resources. Through this process, the system autonomously adapts to environmental changes. This system achieves the technical effect of consistently maintaining optimal processing performance. This system can be widely applied to use cases such as call volume forecasting in call centers or demand forecasting in the retail industry.

[0049] In this way, the automatic case processing system for radio wave countermeasures of the present invention, through its structure of implementing a case information acquisition unit, a case judgment unit, and a case processing unit on a server, and with the terminal mainly undertaking user interface functions such as information input, notification, correction, confirmation, and acknowledgment, can use generative AI to automate a series of business processes from case acceptance to judgment, allocation, contracting, notification, progress management, and historical management at high speed and with high precision. For example, the time from case acceptance to contracting to a construction company can be shortened from several days to within a few hours, enabling effective utilization of human resources, reduction of business costs, improvement of customer satisfaction, standardization of business processes, and improvement of quality. Furthermore, through the utilization of generative AI, it can flexibly handle various patterns or exceptions of each case, making future business expansion or application to other fields easier. For example, it is possible to apply it to a wide range of fields, not only for indoor radio wave countermeasures, but also for the construction or enhancement of outdoor base stations, emergency response during disasters, and the development of public facilities or transportation infrastructure. In addition, by utilizing AI-based automatic learning functions, the accuracy or efficiency of case processing is continuously improved, enabling a high degree of overall business automation and self-discipline. Specifically, this system deploys its processing modules as microservices. It uses container orchestration technology to dynamically control the scaling of each module. The system automatically increases or decreases the number of instances in the case determination or processing department based on the inflow of cases. This system distributes the overall system load and eliminates single points of failure. The system performs generative AI inference processing on dedicated tensor processing units. This system reduces the computation time required for inference to milliseconds. This system completes the entire data pipeline in memory, from the vectorization of input data to the generation of final packet data. This system eliminates bottlenecks caused by disk I / O. Through this close collaboration of hardware and software, this system achieves a high degree of automation previously unattainable in computer systems. This system delivers overwhelming improvements in processing speed and significant reductions in operating costs. This system provides innovative solutions for use cases such as integrated infrastructure management in smart cities or automated production line control in large-scale manufacturing.

[0050] (System Composition) The system described in this embodiment includes a case information acquisition department, a case database management department, a generative AI analysis department, a case judgment department, a case classification department, a construction company selection department, a contracting processing department, a main responsible party allocation department, a progress management department, a notification and feedback department, a history management and learning department, and a user interface department.

[0051] The Case Information Acquisition Department has the following functions: It receives a wide variety of case information from terminals such as personal computers, tablets, and smartphones used by customers or sales representatives. This information includes building name, building address, building type, floor number, floor area, number of surrounding base stations, models of surrounding base stations, drawing data, on-site photos, representative information, client department, customer attributes, electromagnetic field strength measurements, desired response time, past countermeasure history, special circumstances at the site, and customer requests. This information is then sent to the server. For example, if a sales representative enters information such as, "On the 7th floor of office building A in Tokyo, there are 3 surrounding base stations, with attached on-site photos and drawings. The representative is Mr. B, the client is a corporate sales department, the customer is a large corporation, the electromagnetic field strength is -90dBm, a response is desired within one month, it is on a basement floor with many obstructions, and nighttime construction is preferred," the Case Information Acquisition Department accurately receives all this information and coordinates with various departments on the server side. Furthermore, the case information acquisition department also possesses functions such as input information format conversion, file encryption, data standardization, and real-time input assistance. For example, it automatically converts drawing data files with different formats and displays warnings when the quality of on-site photos is low, thus improving the quality of input information. Specifically, the case information acquisition department uses an asynchronous communication protocol to receive data streams sent from the terminal. It segments the received data stream into data packets and temporarily stores them in a buffer area in memory. The department executes a data format discrimination algorithm to parse the file extensions or header information of each file. It preprocesses image data by applying noise reduction filters or edge enhancement. It performs morphological parsing on text data, removing unnecessary symbols or stop words. The department converts the preprocessed data into a structured JSON object. It then sends this JSON object to a message queue system. Through this process, the system standardizes diverse input data formats into a unified data model. This system delivers a significant improvement in computational efficiency during subsequent AI analysis and processing. This system can be applied to use cases such as the comprehensive management of electronic medical records in medical institutions or the collection of sensor data from multiple locations.

[0052] The Case Database Management Department is capable of organizing, storing, and managing acquired case information by time sequence or attributes in a database. For example, it can categorize data by building, floor, customer, and case progress, efficiently managing past case history, ongoing cases, completed cases, and returned cases. Furthermore, the department also performs tasks such as saving large files like drawings or scene photos, data backup, access control management, and encrypted data storage to reduce the risk of information leakage or data loss. For instance, by immediately retrieving the history of cases that occurred in the same building and providing it to the AI ​​analysis department, similar cases can be quickly extracted. Specifically, the Case Database Management Department employs a hybrid storage architecture combining relational and NoSQL databases. Structured data is stored in tables within the relational database. Unstructured data such as images or drawings is stored in object storage. A unique hash value is generated for each data point and used as an index. A built-in vector search engine is included. A multi-dimensional index structure is used to manage the feature vectors of cases. The database management department of this case uses an approximate nearest neighbor search algorithm to quickly retrieve the past case vectors that are closest to the query vector. Through this process, the system identifies similar cases from millions of historical data entries in milliseconds. This system achieves the technical effect of minimizing data retrieval latency. This system is extremely effective in use cases such as prior art investigation systems for patent documents or product recommendation engines in e-commerce.

[0053] The generative AI analysis department runs on a server, taking a variety of data as input, including case information, past case history, scene photos, blueprint data, and customer requirements. It combines multiple AI technologies, such as natural language processing models, image recognition models, machine learning models, and deep learning models, for analysis. For example, regarding blueprint data or scene photos, image recognition AI automatically analyzes the content, extracting information such as the building structure, the presence or absence of obstructions, the size of the floors, and available space. Regarding text information, natural language processing AI detects omissions or inconsistencies, automatically identifying incomplete information such as "no attached blueprint data," "no information on the responsible party," or "unknown number of surrounding base stations." Furthermore, the generative AI analysis department takes multiple factors as input, including building name, floor, number of surrounding base stations, past countermeasures history, radio wave countermeasures status of surrounding buildings, radio wave strength, floor height, customer attributes, client department, special circumstances at the scene, and customer requirements, and retrieves similar past cases from the database, performing pattern recognition or statistical inference. For example, referencing historical examples such as "a previous case on the 6th floor of the same building was resolved by installing an amplifier," "there are 3 surrounding base stations and many similar cases," "the customer attributes are the same and the case was previously resolved by installing a small base station," and "the case is on a basement floor and requires a special amplifier," the system can accurately determine whether a case can be resolved using standard countermeasures. As a specific example of the prompt text read by the generative AI analysis unit, it could be: "Based on the following case information, retrieve similar past cases and infer the best radio wave countermeasure method. Case information: Building name = Office Building A, Floor = 7th floor, Number of surrounding base stations = 3, Attached site photos, Attached drawings, Customer attributes = Large corporation, On-site radio wave intensity = -90dBm, Special circumstances = Basement floor, Many obstructions, Requirement = Desire for nighttime construction," etc. Specifically, this generative AI analysis unit executes a Transformer model with a multi-layered self-attention mechanism. This generative AI analysis unit tokenizes the input case information and generates a tokenized embedding vector. This generative AI analysis unit adds positional encoding to construct the input tensor. The input tensor is, for example, a third-order tensor with batch size, sequence length, and hidden layer dimension. This generative AI parsing unit forward-propagates this input tensor through each layer of the neural network. This generative AI parsing unit applies a visual Transformer model to the image data to extract the feature vector of each image patch. This generative AI parsing unit uses a cross-attention mechanism to fuse textual and image features. This generative AI parsing unit inputs the fused feature vector into a fully connected layer. As the final output, this generative AI parsing unit generates an array of logarithmic values ​​representing the fitness of each strategy. This generative AI parsing unit applies a Softmax function to convert the logarithmic values ​​into a probability distribution. The output data is, for example, numerical data such as an amplifier setting probability of 0.88 and a small base station setting probability of 0.09.This system comprehensively interprets heterogeneous data through processing, uncovering hidden correlations that are difficult for humans to perceive. The system improves analytical accuracy and ensures objective judgment. It can be widely applied to use cases such as ambient environment recognition in autonomous vehicles or medical image diagnostic support systems.

[0054] Based on the analysis results of the generative AI analysis unit, the case determination department automatically determines the appropriateness or excess / deficiency of case information, classifying cases as simple cases that can be resolved by setting up standard amplifiers or small base stations, or complex cases requiring special amplifiers or new base stations. For example, if incomplete information such as "no attached drawing data," "no information of the responsible party entered," or "unclear on-site photos" is detected, a rejection notice is automatically sent to the applicant. The case determination department also records the inference basis of the generative AI, automatically extracting patterns such as "the same countermeasures have been effective multiple times in the same building, on the same floor, with the same number of base stations, and with the same customer attributes," ensuring the transparency of the determination. Furthermore, it automatically prioritizes cases, considering conditions such as "the customer is a large corporation and the on-site radio wave intensity is extremely low," "there is an urgent response request from the customer," and "the same customer has frequently complained in the past," extracting high-priority cases. Specifically, this case determination department receives probability distribution data and feature vectors output from the generative AI analysis unit as input. This case determination department uses a decision tree algorithm to apply classification rules. The case determination department determines whether the maximum value of the probability distribution exceeds a predetermined threshold. If the maximum value exceeds the threshold, the department outputs the proposed countermeasure as the final result. If the maximum value is below the threshold, the department marks the case as a complex case. The department also executes a priority calculation algorithm. It constructs an input vector based on the importance score of customer attributes, the reciprocal of the radio wave intensity, and the number of past complaints. The department calculates the inner product of this input vector with a pre-defined weight vector. This inner product result is input into a sigmoid function, and the output is a priority score normalized to the range of 0 to 1. For example, this output might represent a high-priority value such as 0.92. The department writes this priority score as case metadata into the database. Through this process, the system quantitatively evaluates the importance of cases and dynamically optimizes the processing order. This system achieves the most efficient allocation of limited computing resources. This system can be applied to use cases such as improper transaction detection systems in financial institutions or threat alert classification in cybersecurity.

[0055] Based on the judgment results of the case determination department, the case classification department categorizes cases into simple and complex cases, automatically branching subsequent processing flows. For example, cases determined to require only the installation of amplifiers or small base stations are classified as simple cases, while those requiring special amplifiers, new base stations, or complex on-site investigations are classified as complex cases. The case classification department also automatically determines the processing priority or resource allocation for each case, aiming to optimize human resources. Specifically, this case classification department receives classification tags and priority scores output from the case determination department as input. This case classification department manages multiple processing queues built in memory. When the classification tag indicates a simple case, the case classification department enqueues the case's data pointer into the queue for automatic packet dispatch processing. When the classification tag indicates a complex case, the case classification department enqueues the case's data pointer into the queue for allocation by professional technicians. This case classification department applies a priority-based queuing algorithm based on priority scores when enqueuing. This case classification department dynamically reorganizes the data structure so that cases with higher priority scores are placed at the beginning of the queue. This case classification department constantly monitors the length of each queue. When cases are delayed in a specific queue, the case classification department sends a warning signal to the system administrator. Through this process, the system automatically selects the optimal processing path corresponding to the nature of the case. This system maximizes overall system throughput and prevents bottlenecks from occurring in advance. This system is extremely useful in use cases such as task scheduling in cloud computing environments or call distribution systems in large-scale call centers.

[0056] In cases deemed simple, the construction company selection department uses generative AI to comprehensively evaluate multiple parameters, including past performance, service area, availability, construction period, cost, quality evaluation, past dispute history, and customer feedback, to automatically select the best construction company. For example, the AI ​​analyzes information such as "Construction company A has 10 past projects in the same building, with short construction periods and high ratings" and "Construction company B currently has available time but has a history of delays," to determine the best company. The department also records the selection rationale and past selection history for future business improvement or AI model relearning. Specifically, the department uses a multi-objective optimization algorithm to perform the selection process. It retrieves attribute data for each construction company from the database to construct a candidate matrix. This candidate matrix is ​​a two-dimensional numerical array where rows represent construction companies and columns represent evaluation indicators (cost, construction period, quality score, etc.). Based on the case requirements, the department sets weight coefficient vectors for each evaluation indicator. The department calculates the matrix product of the candidate matrix and the weight coefficient vectors. Finally, the department calculates the comprehensive evaluation score for each construction company. The construction company selection department identifies the construction company with the highest comprehensive evaluation score as the best candidate. This department generates the ID of the selected construction company as output data. The department uses interpretable AI technology to analyze the contribution of each indicator in the selection process. The analysis results are stored in memory as log data. Through this process, the system mathematically handles multiple opposing conditions to derive the optimal solution. This system achieves a fair and reasonable selection process that eliminates human bias. This system can be applied to use cases such as automated bidding systems for public works projects or real-time bidding platforms for online advertising.

[0057] The Contracting Processing Department automatically generates contracting data for construction companies selected by the Construction Company Selection Department and automatically sends it to the construction company's system. For example, it automatically sends a contracting message such as "7th floor of Building A, amplifier setup, attached drawings, construction period within 2 weeks, contractor Mr. B, nighttime construction preferred." The Contracting Processing Department also handles post-contracting progress or delivery management, automatically collects progress reports from the construction company, automatically detects construction completion reports, and sends alarms when delays occur. Specifically, this Contracting Processing Department uses a template engine to generate structured contracting data. It extracts necessary field values ​​from the case database and dynamically constructs payloads in XML or JSON format. It assigns digital signatures to the constructed payloads to prevent data tampering. The Contracting Processing Department sends the payloads to the construction company's external system via a RESTful API endpoint. It implements timeout handling and retry mechanisms during transmission. In the event of a communication error, it uses an exponential backoff algorithm to control the retransmission interval. The Contracting Processing Department receives an HTTP status code 200 indicating acceptance confirmation from the construction company's system. This contracting department uses this status code as a trigger to update the case status in the database to "contracted". This department periodically performs polling to obtain progress data from the construction company. Through this process, the system reliably and securely executes data linkage between heterogeneous systems. This system provides the technical benefits of ensuring transaction integrity and preventing business disruptions. This system is extremely effective in use cases such as settlement gateway linkage in e-commerce or automatic inventory replenishment systems in supply chains.

[0058] When a case is deemed complex, the lead agent allocation department automatically assigns the best lead agent based on factors such as the lead agent's expertise, workload, past experience, applicable area, and suitability in the technical field. For example, AI analyzes information such as "Leader C has handled many similar complex cases in the past and currently has a low workload" and "Leader D is excluded from the allocation pool due to a different technical field," selecting the optimal lead agent. The lead agent allocation department automatically notifies the lead agent's terminal of the case details, allowing the lead agent to conduct a detailed on-site investigation, request assistance from the technical department, or discuss the use of specialized amplifiers or new base stations. Specifically, this lead agent allocation department uses a constraint satisfaction problem-solving algorithm to explore the best lead agent. The department stores all technical personnel's current task status, skill sets, and work schedules as a multidimensional array in memory. It calculates the logical multiplication between the required skills for the case and the technical personnel's existing skills, generating a list of candidates who meet the skill requirements. The department then applies a penalty function based on the current workload score to this candidate list. The assignor allocation department determines the technician with the lowest penalty score as the best assignor. The department outputs the ID of the selected technician. Using this ID, the department sends a request to the push notification server. The push notification server sends an encrypted notification payload containing case details to the technician's mobile terminal. Through this process, the system solves resource allocation problems under complex constraints in polynomial time. This system effectively prevents business concentration on specific individuals and levels out the overall organizational workload. This system can be widely applied to use cases such as automated taxi dispatch systems or driver allocation for on-demand delivery services.

[0059] The Progress Management Department monitors the progress of cases in real time, automatically collects progress reports from construction companies or lead contractors, and visualizes the progress. For example, it records the progress of each step, such as "Construction Start," "Site Investigation Completed," "Amplifier Setup Completed," and "Construction Completion Report," and automatically sends alerts in case of delays. The Progress Management Department also automatically notifies customers or sales representatives of progress updates, helping to improve customer satisfaction. Specifically, this Progress Management Department uses an event-driven architecture to manage state transitions. It asynchronously receives status update events from external systems or terminals via a message broker. It extracts the case ID and a new status code from the received event data. While maintaining exclusive control over the corresponding records in the database, it updates the status value. It records timestamps during the update process. It calculates the difference between the predefined date and time of each milestone and the actual timestamp. If this difference exceeds an acceptable threshold, it generates an alert event indicating a delay. This alert event is then input into the rules engine to determine the appropriate notification destination and method. This progress management department uses email or SMS to automatically send messages to the designated notification destinations. Through this process, the system tracks the status of most concurrently running cases in real time and accurately. This system offers the technical advantage of early detection of process delays and the ability to implement rapid corrective actions. This system can be applied to use cases such as project management tools in large-scale software development or process management systems in manufacturing.

[0060] The Notification & Feedback Department automatically generates the necessary notifications and feedback for each stage of a case's processing and sends them to relevant personnel such as customers, sales representatives, construction companies, and lead personnel. For example, it automatically generates various notifications such as "Case Acceptance Completion Notification," "Return Notification Due to Incomplete Information," "Contract Submission Notification to Construction Company," "Construction Completion Notification," and "Customer Questionnaire Survey Request." The Notification & Feedback Department also collects feedback and survey results from customers and utilizes them for AI model relearning or business improvement. Specifically, this Notification & Feedback Department receives state transition triggers from each processing department via an event-driven message queue. Based on the received state transition triggers, the Notification & Feedback Department initiates a natural language generation model. This natural language generation model is a large-scale language model based on the Transformer architecture. The Notification & Feedback Department converts the case's attribute data and status codes into multi-dimensional vectors. The Notification & Feedback Department inputs these multi-dimensional vectors into the natural language generation model. The natural language generation model generates context-optimized notification text from the input vector. For example, the input to this natural language generation model is tensor data containing a case identifier, customer attribute vectors, and progress status flags. The output of this natural language generation model is the generated text string and a sending priority score. When the output sending priority score exceeds a predetermined threshold, the notification / feedback department sends a notification via an instant messaging communication protocol. The notification / feedback department parses the text data of questionnaire responses from customer terminals using a natural language processing model. This natural language processing model extracts sentiment scores and keyword vectors from the text data. The notification / feedback department stores the extracted sentiment scores and keyword vectors as teacher data in a database. Using the accumulated teacher data, the system updates the weight parameters of the neural networks in the case determination department and the construction company selection department using backpropagation. Thus, the system optimizes notifications and continuously improves the model without human intervention. Through this structure, the system achieves the technical effect of reducing communication bandwidth waste and improving the overall processing efficiency of the system. This technology can be applied to minimize information transmission delays between relevant personnel in the maintenance and management system of large-scale communication infrastructure.

[0061] The History Management and Learning Department automatically records and accumulates the processing history, judgment basis, reasons for selecting the construction company, reasons for assigning the lead responsible party, progress status, customer feedback, construction completion reports, and fault occurrence history for each case. This information is then utilized for future AI model relearning, business improvement, and audit response. For example, it can perform trend analyses such as "all cases in the same building and on the same floor were resolved by setting up amplifiers in the past year," "Construction company A's project delay rate is showing a downward trend," and "Lead responsible party C has high customer satisfaction in handling cases." The History Management and Learning Department utilizes AI-based automatic learning capabilities to continuously improve the accuracy and efficiency of case processing. Specifically, this department has time-series data storage built on a distributed database system. It collects structured and unstructured data output from various processing departments in real time. The department maps the collected data to a high-dimensional feature space and stores it in a vector database. It periodically performs batch processing, using the accumulated vector data for machine learning model relearning. In this relearning process, the History Management and Learning Department calculates the error between past judgments and actual construction results as a loss function. The department uses stochastic gradient descent to optimize the weights of the neural network. A clustering algorithm is executed for trend analysis. This clustering algorithm, for example, groups the performance indicators of construction companies in a multi-dimensional space. The input data to the History Management and Learning Department is a numerical vector containing the processing time, cost, and customer satisfaction score for each case. The output data from the department is the predicted performance scores of each construction company or responsible party, as well as a set of parameters for model updates. The system uses the output predicted scores as weighting coefficients in subsequent construction company selections. Thus, unlike static rule-based systems, this system achieves a self-evolving architecture that dynamically adapts to environmental changes. Through this structure, the system achieves the technical effects of high-speed data retrieval and continuously improved prediction accuracy. This technology is suitable for dynamically changing resource allocation problems or large-scale field service management systems.

[0062] The user interface unit runs on the terminal, providing operation screens for customers, sales representatives, lead representatives, construction company representatives, etc., to input case information, confirm progress, receive notifications, and provide feedback. For example, it includes various screens such as a case information input screen, a drawing or site photo upload screen, a progress timeline display, a notification history list display, and a feedback input form. The user interface unit also includes input assistance functions, error checking functions, and real-time notification functions to improve user convenience. Specifically, this user interface unit has a dynamic rendering engine running on the client terminal's processor. This user interface unit receives structured data in JSON format sent from the server side. Based on the received data, this user interface unit constructs a virtual DOM tree, only redrawing the differences in the screen. This user interface unit executes a lightweight convolutional neural network running on the client side during image upload. This convolutional neural network receives the uploaded image tensor as input. This convolutional neural network analyzes the image's resolution, brightness, and noise level. The output from this convolutional neural network is the image quality score and the bounding box coordinates of unclear areas. If the output image quality score falls below a specified threshold, this user interface unit retains the data transmission to the server and displays a warning on the screen prompting the user to retake the image. In the text fields of the input form, this user interface unit performs real-time input completion processing using a natural language processing model. This input completion process takes the user's keystroke history as a time-series array input and outputs the probability distribution of the next entered word. This user interface unit displays the word with the highest probability as a suggestion on the screen. Therefore, this system proactively prevents invalid data from being sent to the server, significantly reducing network bandwidth consumption. Through this structure, the system effectively reduces the processing load on the server side and improves the overall system response speed. This technology can be applied to mobile applications operating in low-bandwidth communication environments or data input systems requiring real-time performance.

[0063] As described above, the system involved in this embodiment includes a case information acquisition department, a case database management department, a generative AI analysis department, a case judgment department, a case classification department, a construction company selection department, a contracting processing department, a principal responsible person allocation department, a progress management department, a notification and feedback department, a history management and learning department, and a user interface department. The linkage of these departments can automate and optimize a series of business processes from the acceptance of radio wave countermeasure cases to judgment, classification, allocation, contracting, progress management, history management, feedback collection, and AI learning. As specific examples of prompts read by the generative AI, one could say, "Based on the following case information, retrieve similar past cases and infer the best radio wave countermeasure method. Case information: Building name = Office Building A, Floor = 7th floor, Number of surrounding base stations = 3, Attached on-site photos, Attached drawings, Customer attributes = Large-scale corporation, On-site radio wave intensity = -90dBm, Special circumstances = Basement floor, Numerous obstructions, Requirement = Request for nighttime construction" or "From the following construction company candidates, comprehensively evaluate past performance, response area, availability, construction period, cost, quality evaluation, failure history, and customer evaluation to select the best construction company. Candidates: Construction Company A, Construction Company B, Construction Company C," etc. Specifically, this system containers each processing unit as multiple independent microservices, executing them in parallel on the orchestration platform. This system uses a high-speed remote procedure call protocol for data communication between the microservices. The generative AI parsing unit converts the input prompt text into an array of integer tag IDs using a tokenizer. This generative AI parsing unit inputs the array of tag IDs into a Transformer model with a multi-layered self-attention mechanism. The Transformer model calculates the dependencies between the input tags, generating a context vector. The output of this Transformer model is a probability distribution of the text tags representing the optimal radio wave countermeasure method, and a normalized score vector representing the fitness of each candidate construction company. For example, the output score vector might be numerical data such as 0.85 for construction company A, 0.12 for construction company B, and 0.03 for construction company C. The system then inputs the output score vector into a subsequent ranking algorithm, automatically selecting the construction company with the highest score. Based on the selection result, the system performs database transactions and updates the contract status. Thus, this system eliminates human subjective judgment, enabling objective and high-speed decision-making based on multi-dimensional evaluation indicators on a computer. Through this structure, the system leverages the technical effect of integrating complex conditional branching business logic into a single inference process and optimizing computational resource consumption. This technology can be widely applied to supply chain management or dynamic resource allocation systems that require simultaneous satisfaction of multiple constraints.

[0064] (Implementation steps) Step 1: Inputting and Retrieving Case Information Customers or sales representatives use a dedicated application or web browser on their terminals to input a variety of case information, including building name, building address, building type, floor number, floor area, number of surrounding base stations, models of surrounding base stations, drawing data (PDF, CAD files, image files, etc.), on-site photos, representative information, client department information, customer attributes, electromagnetic field strength measurements, desired response time, past countermeasure history, special circumstances on-site, and customer requests. This information is then sent to the server via the Case Information Acquisition Department. For example, a sales representative might input information such as, "On the 7th floor of office building A in Tokyo, there are 3 surrounding base stations. Attached are on-site photos and drawings. The representative is Mr. B, the client is a corporate sales department, the customer is a large corporation, the electromagnetic field strength is -90dBm, a response is desired within one month, it's in a basement with many obstructions, and nighttime construction is preferred." Specifically, the Case Information Acquisition Department receives HTTP requests from the terminals and extracts various data from the payload. The extracted text data is then input into a natural language processing pipeline. This natural language processing pipeline includes a morpheme parser and a named entity extraction model. The case information acquisition unit extracts entities such as "Tokyo," "Office Building A," and "-90dBm" from text data and converts them into structured JSON objects. The unit then expands attached image files and PDF files as binary data into memory. This binary data is input into a preprocessing module for resizing and normalization, generating a multi-dimensional image tensor. For example, the input image tensor is 224×224 resolution, 3-channel RGB data. The generated image tensor is then input into a convolutional neural network for feature extraction. The output of this convolutional neural network is a 512-dimensional feature vector representing the visual features within the image. The unit combines the structured text data with the extracted feature vector to generate a comprehensive case vector. This comprehensive case vector is then forwarded to the subsequent case database management unit and generative AI parsing unit. Thus, this system enables the processing of both unstructured and structured data in a unified vector space. This architecture enables a significant improvement in computational efficiency and a reduction in data processing latency in subsequent AI models. This technology can be applied to any information gathering system or automated help desk system that processes multimodal data.

[0065] Step 2: Database registration and management of case information All case information received from the Case Information Acquisition Department is categorized by the Case Database Management Department according to the progress of each building, floor, customer, and case, and stored and managed in the database. Large files, including drawings and scene photographs, are stored, and security measures such as data backup, access control management, and encrypted storage are implemented. Specifically, the Case Database Management Department employs a hybrid storage architecture combining relational databases and distributed object storage. The department inserts the received comprehensive case vector and structured metadata into indexed tables in the relational database. It divides the large volumes of binary data (drawings or scene photographs) into blocks and writes them in parallel to the distributed object storage. During writing, the department applies a hash function to each block to generate a unique object ID. The department links the generated object IDs to the corresponding records in the relational database. The department performs block encryption using Advanced Encryption Standard (AES) algorithms on all stored data. The department manages the encryption keys through a dedicated hardware security module. To ensure data redundancy, the case database management department performs asynchronous replication in the background across multiple geographically separated data centers. Upon receiving an access request, the department verifies the authentication token of the requesting source and compares it against a role-based access control list to determine access permissions. Inputs to the case database management department include a comprehensive case vector forwarded from the case information acquisition department and a binary data stream. Outputs from the case database management department include a status code indicating a successful write to the database and a unique record ID. The system uses the output record ID as a reference key for data in subsequent processing steps. This allows the system to efficiently separate and manage large volumes of multimedia data and metadata requiring high-speed retrieval. Through this structure, the system achieves scalable data storage without reducing database retrieval speed. This technology is suitable for large-scale electronic medical record systems or project management systems in the construction and civil engineering industries.

[0066] Step 3: AI Analysis of Case Information - Incomplete Information Determination The generative AI analysis department takes a variety of data as input, including case information, past case history, scene photos, drawing data, and customer requirements. It combines multiple AI technologies, such as natural language processing, image recognition, machine learning, and deep learning, to analyze the data, detecting omissions or inconsistencies. It automatically identifies incomplete information such as "no attached drawing data," "no information on the responsible party," and "unknown number of surrounding base stations." For example, a prompt text for the generative AI could read might read: "Based on the following case information, detect omissions or inconsistencies in the input information, and point out any incompleteness. Case information: Building name = Office Building A, Floor = 7th floor, Number of surrounding base stations = 3, Attached scene photos, Attached drawing documents, Customer attribute = Large-scale legal entity, On-site electromagnetic field strength = -90dBm, Special circumstances = Basement floor, numerous obstructions, Request = Request for nighttime construction." Specifically, this generative AI analysis department has an inference engine equipped with a multimodal neural network architecture. It combines comprehensive case vectors, image feature vectors, and text prompts obtained from the case database management department to construct a single input tensor. This generative AI analysis unit inputs the input tensor into an incompleteness judgment model equipped with a multilayer perceptron and a self-attention mechanism. The incompleteness judgment model calculates the correlations between the various features of the input and detects missing or logically contradictory items. For example, the input to this incompleteness judgment model is a 1280-dimensional tensor formed by concatenating a 768-dimensional embedding vector extracted from text information with a 512-dimensional feature vector extracted from an image. The output from this incompleteness judgment model is an array representing the output values ​​of the sufficiency probability of each item using the sufficiency function, and a latent vector representing the reason for the incompleteness for text generation. This generative AI analysis unit identifies items whose sufficiency probability is lower than a pre-set threshold (e.g., 0.95). Based on the identified incomplete items, this generative AI analysis unit uses a decoder network to generate a specific warning message (e.g., "Insufficient model information for surrounding base stations"). This generative AI analysis unit sends the generated warning message along with an error flag to the subsequent notification / feedback unit. Therefore, this system effectively detects inconsistencies in context-dependent information that cannot be detected through simple blank checks (e.g., situations where an obstruction is captured in a site photograph but is recorded as unobstructed in the text information). Through this structure, the system proactively prevents rework in subsequent processes and significantly reduces the overall processing time of the business process. This technology can be applied to automated review systems for complex application documents or automated insurance claims assessment systems.

[0067] Step 4: AI-based search for similar cases and deduction of countermeasures and methods The generative AI analysis department takes multiple factors as input, including building name, floor, number of surrounding base stations, past countermeasure history, radio wave countermeasure status of surrounding buildings, radio wave intensity, floor height, customer attributes, client department information, special circumstances on site, and customer requirements. It retrieves similar past cases from the database, performs pattern recognition or statistical inference, and accurately determines whether a case can be resolved using standard countermeasures. A specific example of a prompt for the generative AI could be: "Based on the following case information, retrieve similar past cases and infer the best radio wave countermeasure method. Case information: Building name = Office Building A, Floor = 7th floor, Number of surrounding base stations = 3, Attached site photos, Attached drawings, Customer attributes = Large corporation, On-site radio wave intensity = -90dBm, Special circumstances = Basement floor, Numerous obstructions, Requirement = Desire for nighttime construction." Specifically, this generative AI analysis department has a dedicated computation module for executing vector similarity retrieval algorithms. This generative AI analysis department converts the input case information into a multi-dimensional query vector. This generative AI analysis unit calculates the cosine similarity between the query vector and a group of past case vectors stored in the database. It then extracts the top N past cases whose similarity exceeds a predetermined threshold. The unit inputs the extracted historical data of past cases and current case information into a graph neural network. This network represents the relationships between cases through nodes and edges, inferring the success probability of a strategy. For example, the input to this network is a graph structure containing a matrix of node features and an adjacency matrix. The output of this network is an array of success probability scores for multiple strategy candidates (e.g., setting up a small amplifier, adding a directional antenna, establishing a new base station, etc.). The unit selects the strategy candidate with the highest success probability score as the optimal solution. It then determines whether the selected strategy candidate is included in the standard strategy list. Finally, the unit outputs the determination result and the inferred strategy method as structured data to the subsequent case classification unit. Thus, this system achieves objective and statistically based strategy inference based on massive amounts of past data, without relying on human rules of thumb. Through this structure, the system effectively reduces the risk of rework due to mismatched countermeasures and dramatically improves the success rate of countermeasures. This technology can be widely applied in the medical field to recommend treatment methods based on past cases or in manufacturing to infer repair steps for mechanical failures.

[0068] Step 5: Case Classification and Prioritization. Based on the inferences from the generative AI analysis department, the case classification and determination departments classify cases as either simple cases that can be resolved by setting up standard amplifiers or small base stations, or complex cases requiring special amplifiers or new base stations. They then consider customer attributes or requirements, the urgency of the situation, and past complaint history to automatically determine case priority. For example, considering conditions such as "the customer is a large corporation, and the radio wave intensity at the scene is extremely low" and "there is an urgent response requirement from the customer," high-priority cases are extracted. Specifically, the case classification department executes a gradient boosting model based on decision tree ensemble learning. This department combines the success probability scores of the countermeasure candidates output by the generative AI analysis department with the case's metadata (radio wave intensity, customer attribute codes, complaint history markers, etc.) as feature vectors. This feature vector is then input into the gradient boosting model. The gradient boosting model performs a weighted average of the outputs of multiple decision trees to calculate the case's complexity score and priority score. For example, the input to this gradient boosting model is a 128-dimensional vector containing values ​​such as radio wave intensity (-95dBm), customer size (level 5), and number of complaints (2). The output from this gradient boosting model is a complexity score (e.g., 0.2) and a priority score (e.g., 0.85) normalized to a range of 0.0 to 1.0. The case classification department marks a case as a "simple case" if the output complexity score is less than a pre-set threshold (e.g., 0.5). The case classification department marks a case as a "complex case" if the complexity score is above the threshold. Based on the calculated priority score, the case classification department dynamically rearranges the processing order within the case queue. The case classification department sends a control signal containing the classification results and priority scores to the subsequent construction company selection department or main responsible party allocation department. Thus, this system simultaneously evaluates multivariate conditions, quantifying and reflecting minor risk factors that are easily overlooked by human intuition in the classification. Through this structure, the system achieves the technical effect of optimal resource allocation and preventing delays in processing highly urgent cases. This technology can be applied to automated incident triage systems in call centers or dynamic prioritization systems for delivery routes in logistics networks.

[0069] Step 6: Automatic Selection and Contracting of Construction Companies for Simple Cases. When a case is determined to be simple, the construction company selection department uses generative AI to comprehensively evaluate multiple parameters, including past performance, service area, available space, construction period, cost, quality evaluation, past dispute history, and customer feedback, to automatically select the best construction company. The contracting department automatically generates contracting data for the selected construction companies and sends it to the construction companies' systems. A specific example of a prompt message for the generative AI could be: "From the following construction company candidates, we comprehensively evaluate past performance, service area, available space, construction period, cost, quality evaluation, dispute history, and customer feedback to select the best construction company. Candidates: Construction Company A, Construction Company B, Construction Company C. Case Information: 7th floor of Building A, amplifier installation, attached drawings, construction period within 2 weeks, responsible person Mr. B, nighttime construction preferred." Specifically, this construction company selection department has an optimization engine based on reinforcement learning algorithms. This department retrieves dynamic parameters (current operating rate, real-time current location, recent quality score, etc.) from the database for each construction company and constructs a state vector. The construction company selection department combines the case requirements data with the feature vectors of each construction company and inputs them into a deep Q-network. This deep Q-network calculates the expected return value (Q-value) for each selected construction company. For example, the input to this deep Q-network is a multidimensional tensor containing the case's time constraints, cost ceiling, and each construction company's past delay or cost overrun rates. The output from this deep Q-network is an array of Q-values ​​for each candidate construction company (e.g., Company A: 85.4, Company B: 62.1, Company C: 91.2). The construction company selection department determines the construction company with the highest output Q-value (Company C) as the best candidate. The contracting processing department receives the identifiers of the selected construction companies and automatically generates structured data (XML or JSON format) for contracting. This contracting processing department asynchronously sends the generated contracting data to the target construction company's external system via a secure API gateway. After transmission, the contracting processing department updates the database status to "Contracted". Therefore, this system is not a static, rule-based allocation system, but rather a dynamic resource optimization system that reflects the constantly changing external environment or the status of construction companies in real time. Through this structure, the system effectively shortens the lead time for contracting tasks to the limit and mathematically optimizes the trade-off between cost and quality. This technology can be applied to the dynamic allocation of drivers in ride-hailing services or the optimal configuration algorithm for virtual machines in cloud computing.

[0070] Step 7: Automatic Assignment and Notification of Lead Personnel for Complex Cases. When a case is determined to be complex, the lead personnel assignment department uses generative AI to comprehensively evaluate the lead personnel's expertise, workload, past experience, suitable area, and technical skill level, automatically assigning the best lead personnel and automatically notifying the lead personnel's terminal of the case details. A specific example of the prompt message read by the generative AI could be: "From the following lead personnel candidates, based on a comprehensive evaluation of their expertise, workload, past experience, suitable area, and technical skill level, the best lead personnel has been selected. Candidates: Personnel X, Personnel Y, Personnel Z. Case Information: 7th floor of Building A, discussion on special amplifier setup, on-site investigation required, customer attribute = large-scale corporation, preference for nighttime construction." Specifically, this lead personnel assignment department executes a hybrid inference module combining a constraint satisfaction problem solver and a neural network. This department retrieves each lead personnel's skill matrix, current workload, and calendar availability from the company's personnel database. The primary assignee allocation department normalizes this information and generates a profile vector for each assignee. The department calculates the inner product of the case complexity vector and each assignee's profile vector to determine the skill fit score. The department incorporates the calculated fit score, the reciprocal of the workload, and a penalty term for movement distance into the objective function. The department then explores assignees that maximize this objective function using gradient descent. For example, the input to this inference module is a matrix containing the required skill vector for the case (e.g., special amplifier knowledge: 0.9, large-scale corporate response: 0.8) and the retained skill vectors of each assignee. The output from this inference module is a list of comprehensive evaluation scores for each assignee and the identification ID of the best assignedee. Based on the output identification ID, the department encrypts the detailed case data and sends it to the assignee's mobile terminal via a push notification server. Before receiving an acceptance response from the assignee, the department initiates an upgrade timer for alternative candidates. Therefore, this system treats the allocation of tasks based on human intuition or experience as a mathematical optimization problem in a multidimensional space. Through this structure, the system effectively prevents tasks from concentrating on specific personnel and maximizes the overall processing throughput of the organization. This technology is suitable for application in standby allocation systems for specialist physicians in large-scale hospitals or dispatch systems for senior engineers in IT infrastructure maintenance.

[0071] Step 8: Progress Management, Notification, and Feedback Collection. The Progress Management Department monitors the progress of cases in real time, automatically collects progress reports from construction companies or lead agents, and visualizes the progress status. The Notification and Feedback Department automatically generates the necessary notifications or feedback for each stage of case processing and sends them to relevant personnel such as customers, sales representatives, construction companies, and lead agents. For example, it automatically generates various notifications such as "Case Acceptance Completion Notification," "Return Notification Due to Incomplete Information," "Contract Submission Notification to Construction Company," "Construction Completion Notification," and "Customer Questionnaire Survey Request," and collects feedback or survey results from customers. Specifically, this Progress Management Department has an event tracking engine based on a state machine architecture. This Progress Management Department asynchronously receives Webhook events sent from external systems or mobile terminals. This Progress Management Department extracts timestamps, location information, and status codes from the received event data and updates the case records in the database. Based on the updated status, this Progress Management Department generates JSON data for depicting Gantt charts or Kanban boards and provides it to the User Interface Department. This Notification and Feedback Department is triggered by status update events and uses a rules engine to determine the notification template to be sent. This notification and feedback department embeds dynamic case data into a determined template and uses a natural language generation model to generate personalized message text. For example, the input to this natural language generation model is a context vector containing the customer's name, the name of the completed process, and the scheduled date and time of the next process. The output from this natural language generation model is a text string with an appropriate tone and attitude matching the customer's attributes. This notification and feedback department sends the generated text as an SMS, email, or push notification via a dedicated application. The department collects access logs for the URLs of sent questionnaires and input response data, feeding them into a sentiment analysis model to quantify customer satisfaction scores. Thus, this system eliminates information asymmetry among multiple stakeholders and ensures project transparency through computer control. Through this structure, the system achieves the technical effect of reducing unnecessary communication traffic or telephone response time for progress confirmation, and improving the overall system efficiency. This technology can be applied to complex supply chain tracking systems or large-scale software development defect tracking systems.

[0072] Step 9: Historical Management and Continuous Learning of AI Models. The Historical Management and Learning Department automatically records and accumulates the processing history, judgment basis, reasons for selecting the construction company, reasons for assigning the main responsible party, progress status, customer feedback, construction completion reports, dispute occurrence history, etc., for each case, and utilizes this information for future AI model relearning, business improvement, and audit response. For example, it can also perform trend analysis such as "all cases in the same building and on the same floor were resolved through amplifier settings in the past year," "the construction delay rate of construction company A has shown a downward trend," and "customer satisfaction is high in the cases handled by the main responsible party C." By utilizing AI-based automatic learning functions, the accuracy or efficiency of case processing can be continuously improved. Specifically, this Historical Management and Learning Department adopts a data lake architecture to accumulate structured and unstructured data in an unprocessed state. This Historical Management and Learning Department regularly performs ETL (Extract, Transform, Load) processing on the accumulated data to update the feature storage used for machine learning. This Historical Management and Learning Department is equipped with an AutoML pipeline, which automatically triggers the model relearning process whenever a certain amount of new data is accumulated. In this relearning process, the History Management and Learning Department defines the difference between past inference results and actual outcomes (e.g., on-time compliance rate or customer satisfaction score) as a loss function. The History Management and Learning Department uses backpropagation to fine-tune the weight parameters of the neural network. For example, the input to this relearning pipeline is a feature vector of 10,000 past cases paired with corresponding correct answer labels (success, failure, delay days, etc.). The output from this relearning pipeline is the updated model's weight file (e.g., HDF5 format) and accuracy metrics (precision, recall, F1 score, etc.) for the validation dataset. When the new model's accuracy surpasses that of the existing model, the History Management and Learning Department seamlessly updates the inference endpoints in the production environment through shadow deployment. Thus, this system autonomously learns from environmental changes or new trends (e.g., performance improvements of specific construction companies) and adapts its inference logic without the intervention of human data scientists. Through this structure, the system effectively prevents concept drift, which causes the system's predictive accuracy to deteriorate over time, ensuring long-term operational stability. This technology can be widely used in algorithmic trading systems in financial markets or in the continuous updating of anomaly detection systems in cybersecurity.

[0073] (specific use case) For example, imagine a corporate client's sales representative, located on multiple floors of an office building in an urban area, receives a request from a business wireless terminal indicating weak radio waves. The representative inputs case information from the terminal's input screen. The sales representative inputs detailed information including: building name, building address, building type (office building), target floor numbers (3rd and 7th floors), area of ​​each floor, number and type of nearby base stations, site photos, building drawings (PDF file), representative information, client department information, large-scale corporate entity, measured radio wave strength, desired response period, history of previous countermeasures, special site conditions (e.g., underground floors with many obstructions), and a client request for nighttime construction. This information is then sent to the server via the case information acquisition unit. Specifically, the terminal's processor buffers the input text data, numerical data, and binary files (JPEG data of site photos or PDF data of building drawings) in memory. The terminal encodes this heterogeneous data into a multi-part form data format, establishes a secure TLS communication channel, and sends it to the server's case information acquisition unit. The case information acquisition unit parses the received data stream and extracts each data field. The case information acquisition unit converts extracted electromagnetic wave intensity measurements (e.g., -95dBm) or floor area data into floating-point arrays. It then inputs text data (e.g., "in the basement with many obstructions," "desiring nighttime construction") into a natural language processing model to generate word embedding vectors. The unit extracts text and vector graphics from PDF files and inputs them along with image files into an image processing pipeline. For example, the input to this pipeline consists of multiple image tensors that have been reduced in resolution and normalized in pixel values. The output from this pipeline is a high-dimensional feature vector representing the structural features within the image (wall material, column configuration, etc.). The unit combines the numerical arrays, word embedding vectors, and image feature vectors to generate a single, comprehensive case vector representing the case. This comprehensive case vector is then passed to subsequent processing modules. Thus, the system converts human-perceived report content into a quantitative, multi-dimensional data structure that can be processed by a computer. This structure allows the system to achieve the technical effect of homogenizing the quality of input data in subsequent AI inferences and suppressing biases in inference accuracy. This technology can be applied to systems that automatically process complex fault reports from field engineers, or to multimodal data collection systems in insurance damage investigations.

[0074] The Case Database Management Department categorizes received information by building, floor, customer, and progress status, storing all data, including drawings and site photos, in the database. The Generative AI Analysis Department, based on the input case information, uses a natural language processing model to detect omissions or inconsistencies in the text information and an image recognition model to analyze the drawings or site photos for obstructions or the feasibility of setting up spaces. For example, if drawings are missing, site photos are unclear, or the responsible party's information is missing, a rejection notice is automatically sent to the applicant as incomplete information. Specifically, the Case Database Management Department distributes the received comprehensive case vectors across a composite cluster of key-value stores and relational databases, enabling high-speed query execution. The Generative AI Analysis Department reads text vectors and image tensors from the stored data and executes multiple inference models simultaneously on a parallel processing architecture. The image recognition model of the Generative AI Analysis Department uses deep layers of convolutional neural networks to extract edge or texture feature maps from the pixel arrays of site photos. This image recognition model, based on the extracted feature map, segments the probability of the presence of occlusions (e.g., concrete walls or metal racks) and the area of ​​blank space where machines can be placed, pixel by pixel. For example, the input to this image recognition model is a normalized RGB image tensor. The output of this image recognition model is a heatmap matrix representing the probability that each pixel belongs to an occlusion or a space where machines can be placed, and a Laplacian variance score representing the image's sharpness. If the output Laplacian variance score is below a threshold, this generative AI analysis unit determines the image to be "unsharp." Simultaneously, the natural language processing model of this generative AI analysis unit performs sequence labeling tasks using text vectors as input, outputting the probability of the presence of required entities (name of the person in charge, contact information, etc.). This generative AI analysis unit comprehensively evaluates the output results of image analysis and text analysis through logical AND operations. If any required requirement is not met, this generative AI analysis unit generates a JSON payload containing error codes and detailed information on missing items, which is immediately sent to the applicant's terminal via the notification and feedback unit. Therefore, this system automates the quality inspection of drawings or photographs that previously relied on human visual verification through pixel-level mathematical operations. This structure physically prevents incomplete data from entering subsequent processing, dramatically reducing the overall system error rate. This technology is suitable for image review in online self-identity verification (eKYC) systems or automated pre-approval systems for building verification applications.

[0075] With complete case information, the generative AI analysis department takes the building name, floor, number of surrounding base stations, past countermeasure history, radio wave countermeasure status of surrounding buildings, radio wave intensity, floor height, customer attributes, client department information, special circumstances on site, and customer requirements as input. It then retrieves similar past cases from the database and performs pattern recognition or statistical inference. For example, if multiple cases exist on other floors of the same building that were resolved using amplifiers, and the number of surrounding base stations or customer attributes are similar, the AI ​​infers that this case can also be resolved using amplifiers. A specific example of a prompt for the generative AI is: "Based on the following case information, retrieve similar past cases and infer the best radio wave countermeasure method. Case information: Building name = office building, Floors = 3rd and 7th floors, Number of surrounding base stations = 2, Attached site photos, Attached drawings, Customer attributes = large corporation, On-site radio wave intensity = -92dBm, Special circumstances = basement floors, numerous obstructions, Requirement = Desire for nighttime construction." Specifically, when the generative AI analysis department receives a signal indicating that the case information is complete, it initiates the inference process. This generative AI analysis unit normalizes the diverse case parameters input to generate a query vector in a high-dimensional feature space. Using an approximate nearest neighbor search algorithm (e.g., HNSW), it extracts the top K similar cases with the smallest Euclidean distance to the query vector from a vector database storing millions of past case vectors, measured in milliseconds. The unit then aggregates the metadata of the extracted similar cases (countermeasures used, post-construction radio wave intensity improvement, etc.) and inputs it into a Transformer-based inference model. This inference model uses a self-attention mechanism to weight the nonlinear relationships between current case parameters (e.g., "radio wave intensity = -92dBm", "most obstructions") and past successful cases. For example, the input to this inference model is a tensor concatenating the current case vector and the top K past case vectors. The output from this inference model is a probability distribution vector processed by a Softmax function for multiple countermeasure categories (amplifier setup, antenna addition, new base station construction, etc.). When the generative AI analysis unit determines that the probability of "amplifier setting" is the highest (e.g., 0.92) in the output probability distribution vector, it identifies the strategy as the optimal solution. This generative AI analysis unit then passes a list of past case IDs, along with attention weight values, to the Explainable AI (XAI) module to generate textual data explaining the reasoning. Thus, this system does not simply perform keyword retrieval, but rather executes advanced pattern recognition based on geometric distance calculations in multidimensional space. Through this structure, the system leverages the technical effect of reproducing the tacit knowledge of skilled engineers as a mathematical model, maximizing the probability of success for the strategy.This technology can be applied to troubleshooting support systems for complex IT infrastructure or parameter optimization systems for improving yield in manufacturing.

[0076] Based on AI's inference results, the case determination and classification departments categorize the case as a simple case that can be resolved using standard amplifier settings and assign it a high priority. The construction company selection department uses AI to comprehensively evaluate construction companies' past performance, target area, availability, construction period, cost, quality evaluation, past dispute history, and customer reviews, automatically selecting the best construction company. For example, it might select a construction company in the same building with a rich track record, short construction period, and high customer reviews. A specific example of the prompt text for the generative AI to read is: "From the following construction company candidates, comprehensively evaluate past performance, target area, availability, construction period, cost, quality evaluation, dispute history, and customer reviews to select the best construction company. Candidates: Construction Company A, Construction Company B, Construction Company C. Case Information: 3rd and 7th floors of office building, amplifier settings, attached drawings, construction period within 2 weeks, responsible person information, preferred nighttime construction." Specifically, the case classification department receives the identification code of the inferred countermeasure category and the calculated probability score as input, and uses a decision tree algorithm to determine the case's routing path. When a case classification department identifies a case as having a "simplified case" status flag if its countermeasure category is "amplifier setting" and its probability score exceeds a threshold, it inserts the case at the top of the priority queue. The construction company selection department, triggered by the "simplified case" flag, retrieves dynamic profile data of candidate construction companies from the database. This department normalizes the retrieved profile data (past construction frequency in the same building, average construction period in days, average customer evaluation scores, etc.) to construct a feature matrix for each construction company. The department then inputs this feature matrix and the case's requirement vector (nighttime construction flag, time constraints, etc.) into a multilayer perceptron-based scoring model. This scoring model applies a nonlinear activation function to the weighted sum of each feature to calculate the overall fitness score for each construction company. For example, the input to this scoring model is a combined tensor of each construction company's 100-dimensional feature vector and case requirement vector. The output from this scoring model is a fitness score for each construction company ranging from 0 to 1 (e.g., Company A: 0.88, Company B: 0.45, Company C: 0.95). The construction company that selects the company with the highest score (Company C) is locked as the best candidate, and the resource allocation table in the database is updated. Thus, this system calculates multiple constraints and changing external factors in real time, completely eliminating manual competitive bidding or schedule adjustments by humans. Through this structure, the system dramatically accelerates the entire project by reducing the lead time selected by the client from days to milliseconds. This technology can be applied to task allocation algorithms in on-demand gig worker platforms or to optimal node configuration systems for containers in cloud environments.

[0077] The Contracting Processing Department automatically generates contracting data for selected construction companies and sends it to their systems. For example, it automatically sends contracting details such as "Office building 3rd and 7th floors, amplifier setup, attached drawings, construction period within 2 weeks, nighttime construction preferred." The Progress Management Department automatically collects progress reports from construction companies, recording the progress of each stage, including construction start, site survey completion, amplifier setup completion, and construction completion report, and automatically sends alerts in case of delays. The Notification and Feedback Department automatically generates various notifications, including case acceptance completion notifications, contracting notifications to construction companies, construction completion notifications, and customer questionnaire survey requests, and collects customer feedback and questionnaire results. Specifically, based on the selected construction company's identification ID, the Contracting Processing Department extracts detailed case data and binary data of attached documents from the database. The extracted data is serialized into a structured JSON payload and sent to the construction company's external system via an HTTP POST request through a RESTful API. To prevent network errors during transmission, the Contracting Processing Department has installed an automatic retry mechanism using an exponential backoff algorithm. This progress management department provides a dedicated endpoint for receiving status updates from the construction company's system via webhooks. The department extracts the work order code and timestamp from the received webhook payload and updates the progress tracking table in the database. The department periodically executes a background process that compares the scheduled completion dates and actual timestamps for each work order. Inputs to the progress management department are the current timestamps and the scheduled dates for each project. Outputs include a list of project IDs that have exceeded their scheduled dates and calculated delay times. Upon detecting a delay, the progress management department triggers an alert and relays it to the notification feedback department. Upon receiving an alert, the notification feedback department uses a natural language generation model to generate an escalation message and sends push notifications to the management dashboard and relevant personnel's mobile devices. Upon receiving a construction completion event, the notification feedback department automatically generates a message containing a questionnaire form URL and sends it to the customer to collect feedback data. Thus, this system fully automates data linkage between systems, enabling real-time and accurate monitoring of project progress. This structure enables the system to prevent human error in reporting or delays in confirmation, and to accelerate initial response to failures to the maximum extent possible. This technology is preferably applied to management systems for large-scale construction projects involving multiple subcontractors, or to global logistics tracking systems.

[0078] Furthermore, in cases deemed complex, the lead agent allocation department uses AI to comprehensively evaluate the lead agent's professionalism, workload, past experience, applicable area, and technical suitability, automatically assigning the best lead agent and automatically notifying the lead agent's terminal of the case details. As a specific example of a prompt message read by the generative AI, it might be: "From the following lead agent candidates, comprehensively evaluate their professionalism, workload, past experience, applicable area, and technical suitability, and select the best lead agent. Candidates: Agent X, Agent Y, Agent Z. Case information: 3rd and 7th floors of office building, discussion of special amplifier settings, need for on-site investigation, customer attribute = large-scale corporation, desire for nighttime construction." Specifically, upon receiving the complex case marker from the case classification department, the lead agent allocation department initiates the inference engine of an advanced matching algorithm. The lead agent allocation department retrieves the skill set vectors of all engineers, the current task queue length, and their current location coordinates based on GPS from the personnel database in real time. The lead agent allocation department encodes the case requirements (knowledge of special amplifiers, skills for handling large-scale corporations, etc.) into a multi-dimensional requirement vector. The primary task allocation unit calculates the cosine similarity between the requirement vector and the skill set vector of each engineer, quantifying the skill suitability. Furthermore, the primary task allocation unit calculates a load penalty score based on the task queue length of each engineer and a distance penalty score based on the travel time from the current location to the site. For example, the input to this inference engine is a feature matrix containing the requirement vector, the skill vectors of each engineer, the load score, and the distance score. The output from this inference engine is the calculated value of the comprehensive evaluation function for each engineer (e.g., engineer X: 0.75, engineer Y: 0.92, engineer Z: 0.60). The primary task allocation unit selects the engineer (engineer Y) with the largest calculated output value as the best candidate and updates the engineer ID field in the database. The primary task allocation unit sends a push notification containing detailed case data and the best route information to the site to the selected engineer's mobile terminal via an encrypted communication channel. Thus, this system instantly solves a multi-objective optimization problem simultaneously optimizing skills, load, and geographical conditions on a computer for tasks requiring high levels of expertise. Through this architecture, the system effectively avoids concentrating business on specific experts, maximizing the resource utilization rate of the entire organization. This technology can be applied to dynamic scheduling systems for widely distributed field service engineers, or to automated allocation systems for maintenance support requiring emergency response.

[0079] The Historical Management and Learning Department automatically records and accumulates the processing history, judgment basis, reasons for selecting the construction company, reasons for assigning the lead responsible party, progress status, customer feedback, construction completion reports, and fault occurrence history for each case. This information is then utilized for future AI model relearning, business improvement, and audit response. For example, it can perform trend analyses such as "all cases in the same building and on the same floor were resolved using amplifier settings in the past year," "Construction company A's project delay rate is showing a downward trend," and "Lead responsible party C has high customer satisfaction in handling cases," leveraging AI-based automatic learning capabilities to continuously improve the accuracy and efficiency of case processing. Specifically, this Historical Management and Learning Department possesses a data warehouse built on a distributed file system, persistently storing all transaction logs and inference logs as time-series data. The department applies natural language processing and statistical feature extraction to the accumulated log data, automatically generating datasets for machine learning model training. Through regular Cron jobs, the department uses the generated datasets to fine-tune deep learning models in the background. In this fine-tuning process, the Historical Management Learning Department calculates the mean squared error between past inferences (e.g., the selection of construction companies) and actual performance metrics (e.g., actual construction period or customer evaluation scores) as the loss function. The department uses the Adam optimizer to repeatedly update the weight parameters of the neural network. For example, the input to this learning process is tensor data containing vectors of past case features and continuous values ​​of their actual success / failure labels or delay days. The output from this learning process is the optimized new weight matrix and a graph showing the shift in loss values ​​relative to the validation dataset. After confirming that the loss value of the new model is lower than that of the existing model, the department continuously updates the model files on the inference server via hot-swapping. The department performs principal component analysis or k-means clustering on the accumulated data to extract hidden patterns such as the performance shifts of construction companies or seasonal trends in electromagnetic interference in specific buildings, and visualizes these patterns on a dashboard. Thus, the system forms an autonomous feedback loop that refines its inference logic the more it is used. Through this structure, the system achieves the technical effect of adapting to changes in the external environment or new obstacle patterns without human intervention, maintaining and improving long-term forecast accuracy. This technology can be widely applied to continuous integration systems for demand forecasting models or automatic update systems for anomaly detection models in manufacturing production lines.

[0080] (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".

[0081] (Implementation methods for carrying out the present invention) In this invention, the automatic fault handling system for urban infrastructure targets a variety of IoT or communication devices installed in the city, including streetlights, traffic lights, surveillance cameras, environmental sensors, public Wi-Fi access points, smart parking meters, smart trash cans, meteorological observation devices, disaster alarm terminals, digital signs, electric bus station terminals, smart water meters, and smart electricity meters. It addresses all faults, anomalies, and performance degradation occurring in these devices, such as communication interruptions, power outages, equipment malfunctions, performance degradation, data loss, abnormal actions, malfunctions, video interruptions, abnormal sensor values, firmware failures, inability to remotely operate, abnormal overheating, abnormal sounds, external damage, misalignment, battery malfunctions, network latency, inability to restart, and software crashes. The system uniformly receives fault reports or automatic detection information from residents, field operators, city managers, remote monitoring systems, and periodic inspection systems. Specifically, this system adopts a distributed computing architecture. The system acquires multidimensional tensors from time-series operational log data sent by each IoT device. The system preprocesses the acquired multidimensional tensors. This system inputs preprocessed data into an anomaly detection model. The system uses an autoencoder-type neural network as the anomaly detection model. The system calculates the error between the input data and the reconstructed data. The system determines an anomaly if the calculated error exceeds a specified threshold. The system receives, as input, a one-dimensional numerical array representing, for example, temperature, humidity, voltage, and communication latency over the past 24 hours. The system outputs a floating-point number ranging from 0 to 1, representing the anomaly probability of each device. For example, the output might be a value with an anomaly probability of 0.92. This output is then sent to the subsequent fault reporting information acquisition department. The system uses a cross-entropy error function to train the model. The system updates the weight parameters using gradient descent. Thus, the system can detect minute performance degradations in real time that are impossible for human visual surveillance to detect. This system effectively minimizes downtime for urban infrastructure. This system can be applied to wide-area infrastructure monitoring in smart cities.

[0082] The Fault Reporting Information Acquisition Department has multiple input channels, including smartphone applications, web portals, management terminals, tablets for field operators, remote monitoring system APIs, and periodic inspection robots. This allows applicants to input equipment type, location (address, latitude and longitude, landmark, floor information, installation height, etc.), fault details (selective / free description), on-site photos or videos (JPEG, PNG, MP4, etc.), applicant information (name, contact information, affiliation, access level, etc.), date and time of occurrence, detailed description of the fault (e.g., occurs only at night, only in rainy weather, periodic, sudden, simultaneous occurrence of multiple devices, etc.), on-site environmental information (e.g., surrounding construction, traffic volume, weather conditions, surrounding radio environment, etc.), past fault history, special on-site conditions (e.g., underground installation, rooftop installation, outdoor installation, densely populated area installation, hazardous area installation, etc.), and requests (e.g., request for emergency response, request for nighttime operation, request for simultaneous operation of multiple devices, request for on-site witnessing, etc.). This information is then sent to the server in real time. For example, residents may report malfunctions such as "the smart trash can in Block B, 3-chome cannot retrieve remaining trash information due to communication interruption," "the surveillance camera in Block C, 1-chome repeatedly interrupts video feeds," "the smart parking timer in Block D, 5-chome malfunctions and causes billing errors," "the digital signage in Block E, 2-chome causes display abnormalities," and "the disaster alarm terminal in Block F, 4-chome emits abnormal sounds." They can attach photos or videos of the scene and provide a detailed description of the situation or requests. Specifically, the malfunction report information acquisition unit converts received text and image data into structured data. The text data is then input into a natural language processing (NLP) model. This NLP model uses a Transformer architecture. It converts the input text into word embedding vectors. It uses a self-attention mechanism to parse the context. The image data is then input into a convolutional neural network. This convolutional neural network extracts feature maps from the image tensor. The malfunction report information acquisition unit receives, for example, a 224-pixel square RGB image tensor or a 512-dimensional text embedding vector as input data. The fault report information acquisition unit generates numerical data representing the urgency score of the fault or the probability of equipment damage as output. For example, the output might be an urgency score of 85 and a damage probability of 0.78. This fault report information acquisition unit integrates these output scores. The integrated scores are then sent to the subsequent case database management unit. This fault report information acquisition unit uses data augmentation techniques to increase learning data. Thus, this system achieves an objective and quantitative evaluation of the reported content. This system provides a high-precision initial triage function that is independent of human subjectivity. This system can be applied to citizen reporting systems in municipalities, etc.

[0083] The Case Database Management Department categorizes and tags received fault reports by multiple dimensions, including equipment category, location, fault content, report date and time, progress status, reporter attributes, site environmental conditions, fault frequency, past fault history, response performance of the operating company or technicians, site photos or videos, drawings, site environmental sensor values, meteorological data, and traffic volume data, and stores this information in the database. For example, it manages and links fault history for the same equipment category and location, frequency of the same fault content, multiple reports from the same reporter, simultaneous faults of multiple devices at the same site, fault trends in each season, correlation with meteorological conditions, correlation with surrounding construction, past response history, response performance of the operating company or technicians, customer satisfaction evaluations, work completion reports, and fault history, and utilizes this information for fault trend analysis, preventative maintenance plan development, and AI model relearning. Specifically, the Case Database Management Department uses a graph database to store information. Each fault report is defined as a node. Equipment category or location is defined as an edge. The database management department calculates the similarity between nodes using cosine similarity. Based on the calculated similarity, the department performs clustering. The K-means algorithm is used to form clusters. As input, the department receives multi-dimensional numerical vectors representing the characteristics of each fault report. For example, the input data is a 128-dimensional vector containing equipment category ID, latitude and longitude, and a timestamp of the date and time of occurrence. As output, the department generates a list of cluster IDs for each node. For example, the output is an integer value with cluster ID 5. This cluster ID is provided to the subsequent fault determination department for inference processing. The department extracts fault occurrence patterns from the formed clusters. Thus, this system achieves high-speed retrieval of similar events from massive historical data. This system significantly reduces the computational load of data retrieval. This system can be applied to large-scale equipment maintenance management systems.

[0084] The fault assessment department uses generative AI to analyze fault report information. Generative AI uses a natural language processing model to detect omissions or inconsistencies in text information, and an image recognition model to automatically extract the equipment's status or abnormal parts (e.g., damage, corrosion, broken wires, LCD cracks, display abnormalities, setting misalignment, foreign object intrusion, external damage, etc.) from on-site photos or videos. It also uses an anomaly detection model to extract abnormal patterns or abnormal time series from sensor values ​​or log data. For example, in cases such as "setting location not filled in," "fault content unclear," "on-site photos unclear," "reporter information not entered," "insufficient on-site environmental information," "fault occurrence date and time unknown," or "unclear correlation between multiple simultaneous equipment failures," an automatic rejection notification is sent to the reporter as incomplete information. With complete information, generative AI takes multiple elements as input, including equipment category, installation location, fault content, on-site photos, videos, past fault history, fault frequency, on-site environmental conditions, complainant attributes, date and time of occurrence, detailed description of the fault, on-site environmental sensor values, meteorological data, traffic volume data, surrounding construction information, past response history, response performance of the operating company or technicians, customer satisfaction evaluation, work completion report, and fault occurrence history. It then retrieves similar past fault cases from the database and performs pattern recognition or statistical inference, anomaly detection, time series analysis, clustering, causal inference, and risk assessment. For example, referencing past incidents such as "a communication interruption occurred in the smart trash can in Block B, 3-chome, which was restored by restarting," "a video interruption occurred in the surveillance camera in Block C, 1-chome, which was resolved by replacing the equipment," "a billing error occurred in the smart parking meter in Block D, 5-chome, which was improved by a firmware update," "a display malfunction occurred in the digital signage in Block E, 2-chome, which was restored by replacing the display module," and "an abnormal sound occurred in the disaster alarm terminal in Block F, 4-chome, which was found to be caused by external damage through on-site investigation," the system accurately classifies faults into simple faults that can be handled by past solutions, or complex faults that require on-site investigation or more complex handling. Furthermore, the AI ​​also performs fault occurrence trend analysis, identifies equipment or locations with high fault frequency, automatically extracts abnormal patterns, detects early signs of abnormalities, predicts fault occurrence risks, assesses the risk of fault ripple effects, prioritizes fault occurrences, and optimizes resource allocation for fault response. Specifically, this fault determination unit executes an integrated model that combines multiple specialized neural networks. As input data, this fault determination unit acquires image tensors, text embedding vectors, and time-series arrays of sensor values. For example, the image tensor is 3-channel RGB image data with a square of 224 pixels. The text embedding vector is a 768-dimensional numerical vector. The time-series array of sensor values ​​is a 1-dimensional array representing the temperature or voltage over the past 24 hours. This fault determination unit combines these input data in the combining layer. This fault determination unit then inputs the combined data into the fully connected layer.This fault determination unit generates a Softmax distribution of the probabilities of simple and complex faults as output. For example, the output might be a simple fault probability of 0.85 and a complex fault probability of 0.15. This unit compares these output probabilities with a predetermined threshold. If the simple fault probability exceeds the threshold, a simple fault flag is set. This flag is then sent to the subsequent fault response processing unit. The unit uses a cross-entropy error function for learning. It updates the weight parameters using gradient descent. Thus, this system achieves high-precision automatic classification based on multimodal data. This system demonstrates the technical advantage of objective fault determination that does not rely on human experience rules. This system can be applied to network fault isolation services in telecommunications operators, etc.

[0085] Based on the fault assessment results, the Fault Response and Handling Department uses generative AI to automatically select and subcontract work companies or field technicians in cases of simple faults. The AI ​​comprehensively evaluates multiple parameters, including the work company or technician's past performance (e.g., number of cases handled within the same equipment category and location, response speed, response quality, customer satisfaction ratings, fault incidence rate, project deadline compliance rate, cost performance, etc.), response area (e.g., city-wide response, specific area response, remote area response, etc.), availability (e.g., same-day response, next-day response, weekend response, nighttime response, etc.), project duration (e.g., standard duration, short-term response, long-term response, etc.), cost (e.g., standard price, surcharges, discount response, etc.), quality evaluation (e.g., customer surveys, on-site audits, third-party evaluations, etc.), past fault history (e.g., project delays, operational errors, reporting delays, on-site faults, etc.), and feedback from the complainant (e.g., thorough response, rapid reporting, easy-to-understand on-site explanations, etc.), to automatically select the best work company or technician. For example, based on information such as "operator_A has a proven track record of handling smart trash cans in Block B, 3-chome, and is currently available," "operator_B has a short lead time for the same fault," and "operator_C has low cost but low quality evaluation," the system automatically generates outsourcing data and sends it to the operating company. The outsourcing content includes specific instructions such as "smart trash cans in Block B, 3-chome, communication interruption, restart response, attach on-site photos, request urgent response." The operating company or technicians respond on-site based on the received instructions and input progress or work completion reports into the system. Specifically, this fault response handling department uses reinforcement learning algorithms to select the best operating company. This fault response handling department models each operating company as an agent. This fault response handling department defines the operating company's past performance or availability status as a state vector. This fault response handling department defines outsourcing as an action. This fault response handling department sets adherence to the schedule or cost reduction as the reward function. This fault response handling department uses a Q-learning algorithm to update the action value function. This fault response handling department uses the evaluation score array of each operating company or the coordinate data of the manageable area as input data. For example, the evaluation score array is a 3D vector representing quality, speed, and cost. This fault response processing unit outputs a probability distribution for selecting each work company. For example, the output might be the probability values ​​for work company A (0.7), work company B (0.2), and work company C (0.1). The fault response processing unit selects the work company with the highest probability. Based on this selection result, the fault response processing unit calls the outsourcing API. Thus, this system achieves optimal resource allocation that adapts to dynamically changing field conditions. This system dramatically reduces the computational processing time for outsourcing tasks. This system can be applied to automated scheduling systems for field engineers.

[0086] In cases of complex faults, the fault response department uses AI to comprehensively evaluate the expertise of technical personnel (e.g., fault analysis of surveillance cameras, smart parking timers, IoT communication, firmware, network, power supply, sensor anomalies, and on-site environmental anomalies), workload (e.g., number of cases currently being handled, overtime hours, available timeframes), past response history (e.g., performance in handling the same equipment category and fault type, on-site investigation experience, fault analysis experience), applicable area (e.g., city-wide response, specific area response, remote area response), and suitability of technical field (e.g., specialization in indoor equipment, outdoor equipment, communication equipment, power supply equipment, etc.) to automatically assign the best technical personnel. For example, based on information such as "Technician X is proficient in analyzing surveillance camera faults and has handled many similar complex faults in the past," "Technician Y is currently under high workload and not within the assigned scope," and "Technician Z has extensive experience in handling outdoor equipment faults," the fault details are automatically notified to the technician's terminal. Technicians conduct on-site investigations or detailed fault analysis, equipment replacement or complete system redesign, firmware analysis, network reconstruction, sensor recalibration, on-site environmental improvement, and discussion of fault spillover prevention strategies, among other advanced responses, and input progress or work completion reports into the system. Specifically, this fault response processing unit uses graph neural networks to match qualified technicians. This unit represents the technician's skill set and the required skills for the fault as multi-dimensional vectors. This unit maps these vectors to a high-dimensional feature space. This unit calculates the Euclidean distance between vectors in the feature space. This unit then identifies the technician with the smallest calculated distance. As input data, this unit uses textual data or categorical data representing the technician's past response history or qualification information. For example, the input data could be an array of the technician's qualification IDs or integer values ​​representing the number of past response cases. As output, this unit generates a list of suitability scores for each technician. For example, the output could be numerical data where technician X scores 95 and technician Y scores 40. This unit sorts the score list in descending order. This fault response department automatically sends notification data packets to the highest-level technical personnel. As a result, the system instantly assigns the best technical personnel to meet complex constraints. This system effectively improves the accuracy of assignment for tasks requiring high levels of expertise. This system can be applied to high-level maintenance and support operations for IT infrastructure.

[0087] The Progress Management Department automatically collects progress reports from operating companies or technical personnel (arrival at site, work started, restart completed, equipment replacement completed, recovery confirmed, site investigation completed, fault analysis completed, firmware update completed, network reconstruction completed, sensor recalibration completed, site environment improvement completed, fault spillover prevention measures implemented, etc.) and notifies managers or the applicant in real time. For example, it automatically visualizes progress information such as "Operating company arrived at site," "Restart operation implemented," "Equipment replacement completed," "Fault recovered," "Site investigation completed," "Fault analysis completed," "Firmware update completed," "Network reconstruction completed," "Sensor recalibration completed," "Site environment improvement completed," and "Fault spillover prevention measures implemented." It automatically sends alerts in case of delays, prompting managers or the applicant to respond quickly. The Progress Management Department also includes a timeline display of progress status, automatic detection of progress delays, automatic summarization of progress reports, a progress status dashboard display, and automatic generation of progress reports. Specifically, this Progress Management Department adopts an event-driven architecture. It asynchronously receives status update data packets sent from various terminals. It parses the timestamps of the received data packets. This schedule management department calculates the difference between the planned work schedule and the actual timestamp. If this difference exceeds an acceptable threshold, the department generates a delay flag. The delay flag triggers the alarm generation module. As input data, the department uses UNIX timestamps representing the start or end time of the work. For example, the input data is an integer value of 1625097600. As output, the department generates numerical data representing the probability of schedule delay or the predicted completion time. For example, the output is data with a delay probability of 0.8 and a predicted completion time of 1625101200. The department sends this output data to the dashboard's rendering engine. Thus, the system tracks the status of tens of thousands of concurrently running cases without delay. This system achieves real-time status synchronization without straining network bandwidth. This system can be applied to the schedule management of large-scale construction projects.

[0088] The History Management and Learning Department automatically records and accumulates the handling history, judgment basis, reasons for selecting the work company, reasons for assigning technical personnel, progress status, feedback from the complainant, work completion report, dispute history, customer satisfaction evaluation, on-site environmental data, meteorological data, traffic volume data, surrounding construction information, fault occurrence tendency, fault occurrence precursors, fault ripple risk, fault response resource allocation, fault response cost, fault response quality, fault response speed, fault response efficiency, fault response accuracy, fault response automation rate, fault response standardization rate, and fault response optimization rate. This information is then applied to future AI model relearning or business improvement, audit response, fault tendency analysis, preventative maintenance plan development, fault occurrence risk assessment, fault ripple risk assessment, optimal allocation of fault response resources, fault response cost reduction, fault response quality improvement, fault response speed improvement, fault response efficiency improvement, fault response accuracy improvement, fault response automation rate improvement, fault response standardization rate improvement, and fault response optimization rate improvement. For example, it can also perform trend analyses such as "all communication interruption faults of the same streetlights in the past year were resolved by restarting", "the project delay rate of operator_A shows a decreasing trend", "the complainant satisfaction rate is high in the cases handled by technician X", "frequent communication interruption faults in the smart trash cans in Block B, 3-chome", "the video interruption faults in the surveillance cameras in Block C, 1-chome show an increasing trend", "the billing error faults in the smart parking meters in Block D, 5-chome show a decreasing trend", "the abnormal display faults in the digital signs in Block E, 2-chome are related to seasonal changes", and "the abnormal sound faults in the disaster alarm terminals in Block F, 4-chome are related to surrounding construction". Utilizing AI-based automatic learning functions, the accuracy and efficiency of fault response are continuously improved. Specifically, the History Management and Learning Department constructs a continuous learning pipeline. The History Management and Learning Department periodically batch processes the accumulated historical data. The History Management and Learning Department extracts feature vectors from the historical data. The History Management and Learning Department uses these feature vectors to fine-tune the existing AI model. The History Management and Learning Department uses the error backpropagation method to fine-tune the model weights. This History Management & Learning Department uses textual data from job completion reports or numerical scores of customer satisfaction as input data. For example, input data might be integer values ​​from 1 to 5 representing customer satisfaction, or numerical values ​​of job time in minutes. This department generates an updated weight parameter matrix for the model as output. For example, the output might be a multidimensional tensor containing millions of floating-point numbers. This weight parameter matrix is ​​then deployed to the inference server. To prevent overfitting, the department applies Dropout. Thus, the system autonomously adapts to environmental changes or new failure modes. This system leverages the technical effect of improving the model's inference accuracy over time. This system can be applied to self-evolving system management foundations.

[0089] Furthermore, in this embodiment, to flexibly address the diverse equipment or failure modes of urban infrastructure, the AI ​​model can be expanded or the database broadened, the user interface customized, and it can be integrated with external systems (urban management systems, remote monitoring systems, disaster response systems, traffic control systems, meteorological information systems, construction management systems, resident reporting applications, etc.). For example, even when a large number of failure reports are concentrated during a disaster, the AI ​​can automatically determine the priority of failures and prioritize responses starting with critical infrastructure. Additionally, it can visualize the failure trends of the entire city in real time and propose optimal times for preventative maintenance or equipment upgrades to city operators. Furthermore, it can achieve AI-based failure precursor detection or failure ripple risk assessment, optimal allocation of failure response resources, cost reduction in failure response, improved failure response quality, speed, efficiency, accuracy, automation rate, standardization rate, and optimization rate in failure response. Specifically, this system adopts a microservice architecture to loosely couple the various functions. The system provides RESTful API endpoints for integration with external systems. This system automatically scales up computing nodes horizontally using container orchestration tools when traffic surges during disasters. It uses precipitation data from meteorological information systems or congestion length data from traffic control systems as input. For example, precipitation data is a floating-point number expressed in millimeters per hour. The system generates heatmap data showing the predicted probability of fault occurrence for each area as output. For example, the output is a two-dimensional array representing the predicted probability for each grid. This heatmap data is then distributed to the terminals of city managers. Thus, the system maintains stable processing performance even under sudden load fluctuations. This system achieves a significant improvement in the availability and fault tolerance of the entire system. This system can be applied to the integrated operation centers of smart cities.

[0090] Thus, the automated urban infrastructure fault response system of this invention will be consistently automated from fault report acceptance to incomplete information determination, automatic comparison with past data, fault classification, automatic selection and subcontracting of operating companies or technicians, progress management, historical management, AI learning, fault tendency analysis, fault occurrence precursor detection, fault ripple risk assessment, optimal allocation of fault response resources, cost reduction, quality improvement, speed improvement, efficiency improvement, accuracy improvement, automation rate improvement, standardization rate improvement, and optimization rate improvement. This will achieve efficient use of urban infrastructure, rapid fault response, optimal allocation of human resources, cost reduction, improved service quality for urban residents, enhanced urban resilience, and higher levels of urban operation and self-discipline. Specifically, this system possesses a workflow engine that fully automates a series of processing steps. This workflow engine executes data transfer between processing modules at high speed via a message queue in memory. By eliminating human intervention, the system reduces data processing latency to milliseconds. The system uses all transaction logs from fault report to report completion as input data. For example, the input data is structured log data containing the start and end times of each processing step. The system calculates a KPI score representing the overall system processing efficiency as output. For example, the output is a numerical value in milliseconds representing the average processing time. Based on this KPI score, the system automatically identifies bottlenecks. Thus, the system completely eliminates the inherent waiting time or communication errors in previous manual processes. This system leverages the technical effectiveness of ensuring the continuity and completeness of information processing. This system can be applied to fully automated solutions for all infrastructure management.

[0091] (System Composition) The system involved in this embodiment includes a fault reporting information acquisition department, a case database management department, an AI analysis generation department, a fault judgment department, a fault classification department, an operation company selection department, a subcontracting processing department, a professional and technical personnel allocation department, a progress management department, a history management and learning department, and a user interface department.

[0092] The Fault Reporting Information Acquisition Department accepts fault reports from residents, on-site operators, city managers, remote monitoring systems, and regular inspection systems for a unified system covering a wide range of IoT or communication devices in the city, including streetlights, traffic lights, surveillance cameras, environmental sensors, public Wi-Fi access points, smart parking meters, smart trash cans, meteorological observation devices, disaster alarm terminals, digital signs, electric bus station terminals, smart water meters, and smart electricity meters. For example, it has a variety of input channels, such as smartphone applications, web portals, management terminals, tablets for field operators, remote monitoring system APIs, and regular inspection robots, which enable applicants to input equipment type, location (address, latitude and longitude, landmark, floor information, setting height, etc.), fault details (selective or free description), on-site photos or videos (JPEG, PNG, MP4, etc.), applicant information (name, contact information, affiliation, permission level, etc.), date and time of occurrence, detailed description of the fault, on-site environmental information (whether there is surrounding construction, traffic volume, weather conditions, surrounding radio environment, etc.), past fault history, special on-site conditions (underground setting, roof setting, outdoor setting, dense setting, dangerous area setting, etc.), and requirements (wish for emergency response, wish for night operation, wish for multiple devices to operate simultaneously, wish for on-site witnessing, etc.), and send this information to the server in real time. For example, residents can report faults such as "the smart trash can in Block B, 3-chome cannot retrieve remaining information due to communication interruption," "the surveillance camera in Block C, 1-chome repeatedly interrupts video," "the smart parking meter in Block D, 5-chome malfunctions and causes billing errors," "the digital signage in Block E, 2-chome causes display abnormalities," and "the disaster alarm terminal in Block F, 4-chome emits abnormal sounds." They can attach photos or videos of the scene and provide a detailed description of the situation or requests. Specifically, this fault report information acquisition unit utilizes edge computing technology. This unit performs initial data filtering on edge devices configured at each input channel. The unit receives GPS coordinate data or time-series data from a gyroscope sensor sent from a smartphone as input data. For example, GPS coordinate data is a floating-point pair representing latitude and longitude. The unit generates a noise-removed normalized data stream as output. For example, the output is a smoothed numerical array after removing outliers. The unit encrypts this normalized data stream and sends it to the central server. This significantly reduces the communication bandwidth consumption of the central server. This system effectively eliminates the bottleneck in processing a large number of simultaneous applications. It can be applied to the data collection infrastructure of widely distributed sensor networks.

[0093] The Case Database Management Department categorizes and tags received fault reports across multiple dimensions, including equipment category, location, fault content, report date and time, progress status, reporter attributes, site environmental conditions, fault frequency, past fault history, response performance of the operating company or technicians, site photos or videos, drawings, site environmental sensor values, meteorological data, and traffic volume data, and stores this information in the database. For example, it manages and links together past fault history for the same equipment category and location, frequency of the same fault content, multiple reports from the same reporter, simultaneous faults of multiple devices at the same site, fault trends in different seasons, correlation with meteorological conditions, relevance to surrounding projects, past response history, response performance of the operating company or technicians, customer satisfaction evaluations, work completion reports, and dispute history. This information is used for fault trend analysis, preventative maintenance plan development, and AI model relearning. Specifically, the Case Database Management Department uses a hybrid distributed relational database and NoSQL database. Structured data is stored in the relational database. Images and videos, which are considered unstructured data, are stored in object storage. The database management department of this case receives an array of multi-dimensional metadata tags as input data. For example, the input data is JSON formatted data containing device IDs, fault codes, and priority flags. The database management department generates a unique hash index for each data record as output. For example, the output is a 256-bit hexadecimal string. The database management department uses this hash index to achieve high-speed pointer referencing. Therefore, even with petabyte-scale data, this system maintains millisecond-level retrieval response times. This system effectively optimizes the IOPS performance of data storage. This system can be applied to comprehensive data lakes based on big data analytics.

[0094] The generative AI analysis department takes a variety of data as input, including fault report information, past fault history, on-site photos and videos, on-site environmental data, meteorological data, traffic volume data, surrounding engineering information, the performance of the operating company or technicians, customer satisfaction evaluations, work completion reports, and dispute history. It then combines multiple AI technologies, such as natural language processing models, image recognition models, anomaly detection models, time series analysis models, clustering models, causal inference models, and risk assessment models, for analysis. For example, it automatically extracts the status or abnormal parts of equipment (damage, corrosion, broken wires, LCD cracks, display abnormalities, setting misalignment, foreign object intrusion, external damage, etc.) from on-site photos or videos; extracts abnormal patterns or abnormal time series from sensor values ​​or log data; detects omissions or inconsistencies in text information; and performs fault occurrence trend analysis, anomaly warning detection, fault ripple risk assessment, and optimal allocation of fault response resources. As a specific example of a prompt text that generative AI can read, one could be: "Based on the following fault report information and on-site photos, retrieve past similar fault cases and infer the best countermeasure method. Fault report information: Equipment type = smart trash can, Location = B-zone 3-chome, Fault content = communication interruption, attached on-site photos, Report date and time = June 1, 2024, On-site environment = surrounding construction, Weather = rain, Request = urgent response requested." Specifically, this generative AI analysis unit executes a multimodal inference engine based on a large-scale language model. This generative AI analysis unit receives text prompts, image tensors, and matrices of time-series sensor data as input data. For example, the text prompt is a tokenized array of integers. The image tensor is a multidimensional array of normalized pixel values. This generative AI analysis unit projects these heterogeneous data into a common latent space vector. This generative AI analysis unit uses the Transformer's self-attention mechanism to calculate the correlations between modalities. This generative AI analysis unit outputs a fault cause probability distribution, a text generation result of recommended countermeasure steps, and a list of IDs of relevant past cases. For example, the output might be numerical data indicating an 88% probability that the cause is a communication module failure, or natural language text representing countermeasures. This generative AI analysis unit then passes this output data to the subsequent fault determination unit. Thus, this system identifies complex fault causes that cannot be found in a single data source. This system achieves a significant improvement in the accuracy and comprehensiveness of inference. This system can be applied to fault diagnosis systems for complex industrial machinery.

[0095] Based on the analysis results of the generative AI analysis unit, the fault determination department automatically determines the appropriateness or excess / deficiency of fault report information, classifying the fault as a simple fault that can be handled through past solutions or a complex fault requiring on-site investigation or more stringent handling. For example, in cases such as "location not filled in," "fault details unclear," "unclear on-site photos," "no information entered by the applicant," "insufficient on-site environmental information," "fault occurrence date and time unknown," or "unclear correlation between multiple simultaneous equipment failures," the information is considered incomplete, and a rejection notification is automatically sent to the applicant. When the information is complete, referring to similar past fault cases, the department accurately classifies the fault as a simple fault that can be handled through past solutions or a complex fault requiring on-site investigation or more stringent handling. Specifically, this fault determination department executes a hybrid decision logic combining decision tree algorithms and neural networks. The department receives the causal probability distribution or a numerical vector representing the missing information rate output from the generative AI analysis unit as input data. For example, the input data is an array of causal probabilities or a binary array representing the presence or absence of required input items. The department first uses a decision tree algorithm to determine the excess / deficiency of information through hard rules. When sufficient information is determined, this fault diagnosis unit uses a neural network to calculate a complexity score. The unit generates a continuous value between 0 and 1 (the complexity score) and a binary flag indicating simplicity or complexity as output. For example, the output might be a dataset with a complexity score of 0.3 and a flag of 0 indicating simplicity. Based on this flag, the unit determines the subsequent processing route. Thus, this system combines the certainty of rule-based systems with the flexibility of machine learning. This system effectively reduces the rework rate caused by misjudgments to a minimum. This system can be applied to automated, responsive customer support triage systems.

[0096] Based on the fault determination results, the fault classification department categorizes faults into simple and complex faults and automatically branches subsequent processing flows. For example, faults that can be resolved through rebooting, firmware updates, component replacement, setting changes, remote operation, or on-site cleaning are classified as simple faults, while faults requiring on-site investigation, equipment replacement, system redesign, network reconstruction, sensor recalibration, on-site environmental improvement, and investigation of fault spillover prevention measures are classified as complex faults. Specifically, this fault classification department functions as a software switch that executes a dynamic routing algorithm. As input data, this fault classification department receives a binary flag and fault metadata sent from the fault determination department. For example, the input data is a flag value of 1 and header information containing the fault ID. This fault classification department reads the bit value of the binary flag. When the bit value is 0, this fault classification department pushes the data to the processing queue for simple faults. When the bit value is 1, this fault classification department pushes the data to the processing queue for complex faults. As output, this fault classification department generates a status code indicating successful enqueueing to each queue. For example, the output is an integer value of 200 indicating success. The fault classification department sends the status code to the log management module. This distributes the load across subsequent processing modules, maximizing parallel processing efficiency. The system effectively improves the overall processing throughput. This system can be applied as a load balancer in large-scale transaction processing systems.

[0097] When a fault is determined to be simple, the company selection department uses generative AI to comprehensively evaluate multiple parameters, including the past performance of the company or field technicians, their assigned area, available time, project duration, cost, quality evaluation, past dispute history, and evaluations from applicants, to automatically select the best company or technician. For example, selection is based on information such as "operator_A has a proven track record of handling smart trash cans in Block B, 3-chome, and is currently available," "operator_B has a shorter project duration for the same fault," and "operator_C has low cost but low quality evaluation." The department automatically generates contracting data and sends it to the companies. Specifically, the company selection department executes an evolutionary algorithm for solving constrained optimization problems. As input, the department receives a multidimensional matrix containing parameter groups for each company. For example, this matrix might have rows representing companies and columns representing numerical data for cost, project duration, and quality score. The department represents combinations of companies as chromosomes. It repeatedly performs crossover and mutation operations to explore optimal solutions. As output, the department generates the ID of the company with the highest fitness function value and detailed evaluation scores. For example, the output might be data for company ID 1024 with a comprehensive score of 92.5. The company selection department then passes this ID to the contracting department. Thus, the system rapidly discovers the globally optimal solution from a large pool of combinations. This system effectively achieves optimal matching while minimizing computational resource consumption. This system can be applied to automated selection systems for delivery providers in the logistics industry.

[0098] The Outsourcing Processing Department automatically generates outsourcing content (fault description, on-site photos, work instructions, response deadline, requirements, etc.) for work companies or technicians selected by the Work Company Selection Department, and automatically sends it to the system of the work company or technicians. For example, it might include specific instructions such as "Smart trash can in Block B, 3-chome, communication interrupted, restart response, attached on-site photos, requesting urgent response." The work company or technicians respond on-site based on the received instructions and input the progress status or work completion report into the system. Specifically, this Outsourcing Processing Department uses smart contract technology to record outsourcing transactions on the blockchain. As input data, this Outsourcing Processing Department receives the work company ID and text data containing the fault description from the Work Company Selection Department. For example, the input data is an integer value of the work company ID and a string of work instructions. This Outsourcing Processing Department generates a cryptographic hash value from this input data. This Outsourcing Processing Department generates electronic outsourcing document data containing this hash value. As output, this Outsourcing Processing Department generates a digitally signed outsourcing data packet. For example, the output is binary data signed using the RSA algorithm. This outsourcing processing department sends the outsourcing data packet to the operating company's system via a secure communication protocol. This physically makes tampering with the outsourcing content impossible. The system effectively guarantees the integrity and non-repudiation of data in inter-company transactions. This system can be applied to automated outsourcing platforms in supply chain management.

[0099] When a fault is determined to be complex, the professional technician allocation department uses AI to comprehensively evaluate the professional expertise, workload, past experience, applicable area, and suitability for the technical field of the technicians, automatically assigning the best professional technician. For example, selection is based on information such as "Technician X is proficient in fault analysis of surveillance cameras and has handled many similar complex faults in the past," "Technician Y currently has a high workload and is outside the assigned category," and "Technician Z has extensive experience in handling outdoor equipment faults." The fault details are automatically notified to the technician's terminal. The technicians then conduct on-site investigations or detailed fault analysis, equipment replacement or complete system redesign, firmware analysis, network reconstruction, sensor recalibration, on-site environmental improvement, and discussion of fault spillover prevention strategies, providing progress status or work completion reports to the system. Specifically, this professional technician allocation department uses support vector machines to perform multi-class classification tasks. The department receives feature vectors of complex faults and skill vectors of each technician as input data. For example, the feature vectors are arrays representing one-hot encoded technical fields of the fault. The department uses kernel functions to map these vectors to a high-dimensional space. This professional technician allocation department uses a hyperplane to separate the optimal technician category. As output, it generates a list of suitable allocation probabilities for each technician. For example, the output might be probability data for technician X (85%) and technician Z (60%). The department then sends push notifications to the terminals of the technicians with the highest probabilities. Thus, the system achieves high-precision matching even for complex skill requirements with non-linear relationships. This system maximizes the utilization rate of highly skilled personnel. It can also be applied to the automated consultation allocation of specialist physicians in the medical field.

[0100] The Progress Management Department automatically collects progress reports from operating companies or technicians (arrival at site, commencement of work, restart completion, equipment replacement completion, recovery confirmation, site investigation completion, fault analysis completion, firmware update completion, network reconstruction completion, sensor recalibration completion, site environment improvement completion, fault impact prevention measure implementation completion, etc.) and notifies managers or the reporting party in real time. It also includes a timeline display of progress status, automatic detection of progress delays, automatic summarization of progress reports, a progress status dashboard display, and automatic generation of progress reports. Specifically, this Progress Management Department uses a stream processing engine to summarize real-time data. As input data, this department receives a continuous stream of status codes sent from the terminals of various technicians. For example, status codes are integer values ​​such as 200 indicating commencement of work and 201 indicating restart completion. This department uses a sliding window method to calculate the progress rate at regular intervals. As output, this department generates time-series statistical data representing the progress status of the entire system. For example, the output is structured data showing a completion rate of 75% and 3 delayed cases. This statistical data is cached in an in-memory database and responds immediately to queries from the dashboard. Therefore, this system avoids the database's disk I / O bottleneck. This system achieves the technical effect of visualizing the status of thousands of simultaneously running tasks with millisecond-level latency. This system can be applied to the operational monitoring of large-scale manufacturing production lines.

[0101] The History Management and Learning Department automatically records and accumulates the handling history, judgment criteria, reasons for selecting the work company, reasons for assigning technical personnel, progress status, feedback from the complainant, work completion reports, dispute history, customer satisfaction evaluations, on-site environmental data, meteorological data, traffic volume data, surrounding construction information, fault occurrence trends, fault occurrence precursors, fault ripple risks, fault response resource allocation, fault response costs, fault response quality, fault response speed, fault response efficiency, fault response accuracy, fault response automation rate, fault response standardization rate, and fault response optimization rate for each fault. This data is then utilized for future AI model relearning or business improvement, audit responses, fault trend analysis, preventative maintenance plan development, fault occurrence risk assessment, fault ripple risk assessment, optimal allocation of fault response resources, fault response cost reduction, improvement of fault response quality, improvement of fault response speed, improvement of fault response efficiency, improvement of fault response accuracy, improvement of fault response automation rate, improvement of fault response standardization rate, and improvement of fault response optimization rate. Specifically, this History Management and Learning Department structures the diverse data collected and stores it in a large-scale distributed database system. This system utilizes this accumulated data as input to a continuous machine learning pipeline. The History Management & Learning Department extracts data from a specific period through periodic batch processing. The processing department then divides the extracted data into a learning dataset and a validation dataset. The system performs relearning of a deep learning model. This model, for example, employs a neural network architecture with multilayer perceptrons or self-attention mechanisms. The History Management & Learning Department generates numerical vectors containing environmental sensor values ​​or meteorological data at the time of the failure, using this data as input. This input data is, for example, a 128-dimensional floating-point vector composed of elements such as temperature, humidity, precipitation, and wind speed. The processing department applies word embedding techniques to text data, converting it into a 512-dimensional dense vector. The system combines these vectors to generate a comprehensive feature tensor. This comprehensive feature tensor is received as input by the model. The model performs forward propagation and outputs the future failure probability or the optimal resource allocation ratio. This output data is, for example, a floating-point score representing the failure probability of each device within 24 hours, ranging from 0.0 to 1.0. Additionally, this output data includes an array of integer values ​​representing the recommended personnel allocation for each work shift. The Historical Management and Learning Department calculates the error between the output predictions and the actual application results. The processing department uses cross-entropy error or mean squared error as the error function. The system uses backpropagation to calculate the gradient of the weight parameters within the network. The processing department updates the weight parameters using an optimization algorithm. The system saves the updated weight parameters as a new inference model. By automating this relearning process, the system continuously improves the model's adaptability to environmental changes or new failure modes. The processing department sends the output failure probability score to the preventative maintenance module.This preventative maintenance module automatically issues inspection instructions for equipment whose scores exceed pre-set thresholds. Therefore, this system achieves high-precision preventative maintenance through data-driven methods, without relying on human rules of thumb. By preventing failures before they occur, this system minimizes downtime for urban infrastructure and dramatically increases the overall system uptime.

[0102] The user interface provides interactive screens for residents, site workers, city managers, technical professionals, and project managers to input case information, confirm progress, receive notifications, and provide feedback. For example, it includes screens for inputting case information, uploading site photos or videos, displaying a timeline of progress, listing notification history, providing feedback forms, displaying a fault tendency analysis dashboard, and providing preventative maintenance proposals. The user interface also incorporates input assistance or error checking functions, real-time notifications, and AI-based auto-completion of input content to enhance user convenience. Specifically, this user interface uses an asynchronous communication protocol between the front-end application and the back-end server. The system sends dynamically generated screen description data to the web browser or dedicated application running on the user's terminal. Whenever a user enters text in the input form, the user interface sends the input string sequence to the back-end inference server in real time. This inference server is equipped with a lightweight natural language processing model. This model tokenizes the input string sequence and converts it into an embedding vector. This input data is, for example, an integer array with a maximum length of 64 tokens. Based on this input data, the model predicts the most likely words or phrases to be entered next. This model outputs the probability distribution of each token in the vocabulary dictionary. This output data is, for example, a 30,000-dimensional probability vector. The inference server extracts the top 5 candidate texts with the highest probability values. The inference server returns the extracted candidate texts to the front end in structured data format. The user interface immediately renders the received candidate texts as a dropdown list on the screen. When the user selects a candidate, the system automatically completes the remaining input fields. This auto-completion function significantly reduces the number of physical keystrokes required for input. This system reduces the incidence of input errors and improves the quality of information stored in the database. Furthermore, the user interface performs preprocessing on the client side for the uploaded image data. This processing unit adjusts the image resolution to a standard size and normalizes the pixel values. The system sends the preprocessed image tensor to the backend image recognition module. Through these techniques, the system reduces the computational load on the server side while suppressing network bandwidth consumption.

[0103] As a specific example of a prompt message that can be read by generative AI, one could be: "Based on the following fault report information and on-site photos, search for similar past fault cases and infer the best countermeasure. Fault report information: Equipment type = smart trash can, Location = B-3-chome, Fault content = communication interruption, Attached on-site photos, Report date = June 1, 2024, On-site environment = construction in the surrounding area, Weather = rain, Request = urgent response needed." Another example is: "From the following candidate companies, comprehensively evaluate their past performance, response area, availability, construction period, cost, quality evaluation, dispute history, and applicant evaluation to select the best company." Company. Candidates: operator_A, operator_B, operator_C. Case information: Smart trash can in Block B, 3-chome, communication interrupted, restart required, attached photos of the scene, requesting urgent response. "From the following professional and technical personnel candidates, comprehensively evaluate their professionalism, workload, past response history, applicable area, and suitability in the technical field to select the best professional and technical personnel. Candidates: Technician X, Technician Y, Technician Z. Case information: Surveillance camera in Block C, 1-chome, video interrupted, on-site investigation required, attached photos of the scene, request = requesting nighttime operation," etc. Specifically, this system has a prompt generation module that dynamically constructs these prompts. This prompt generation module embeds structured data obtained from the database into a natural language template. This system formats the generated prompts in conjunction with the input format of a large-scale language model. This system divides the prompts into three parts: system instructions, context information, and user requirements. This processing unit tokenizes the text data in units of sub-words. This processing unit projects the token column onto a multi-dimensional embedding vector space. The input data is, for example, an integer array with a maximum length of 2048 tokens. This large-scale language model uses a self-attention mechanism to compute dependencies between input tokens. The model performs matrix operations using billions of parameters in parallel. As output data, the model generates a probability distribution of text tokens representing the inference results. The system samples the most probable token columns from the probability distribution to generate the final response text. This output data may be, for example, structured text data containing a specific company ID, a list of recommended countermeasures, and the basis for their inferences. The system performs regular expression parsing on the output text data. This processing unit extracts necessary parameters from the parsing results and passes them to subsequent automatic packet sending or notification modules. Through this series of processes, the system generates highly accurate, computer-executable structured commands from natural language instructions presented as unstructured data. The system automatically executes complex decision-making logic without human intervention. By leveraging the high contextual understanding capabilities of the large-scale language model, the system can flexibly and appropriately handle exceptional fault conditions or ambiguous requirements that are difficult to handle in previous rule-based systems.

[0104] (Implementation steps) Step 1: Input and Acquisition of Fault Reporting Information. Residents, field operators, city managers, remote monitoring systems, and periodic inspection systems can use various input channels, including smartphone applications, web portals, manager terminals, field operator tablets, remote monitoring system APIs, and periodic inspection robots, to report faults in equipment such as streetlights, traffic lights, surveillance cameras, environmental sensors, public Wi-Fi access points, smart parking meters, smart trash cans, meteorological observation devices, disaster alarm terminals, digital signage, electric bus stop terminals, smart water meters, and smart electricity meters. When reporting, users must input the equipment category, location, fault details, on-site photos or videos, reporter information, date and time of occurrence, detailed fault description, on-site environmental information, past fault history, special circumstances, and any requirements. This information is sent to the server in real time. Specifically, this input acquisition unit has an API gateway that comprehensively manages data reception from multiple endpoints. This system asynchronously receives data packets sent from various input channels. This processing unit parses the header information of the received data packets and extracts the payload. This system separates the text data, numerical data, and binary data contained in the payload. This processing unit performs preprocessing on the binary data of images or videos. It adjusts the image data, for example, to a resolution of 224 pixels square. It normalizes the RGB values ​​of each pixel to the range of 0 to 1. It converts the normalized image data into a 3-channel multidimensional tensor. The input data is, for example, a floating-point tensor with a shape of 224 x 224 x 3. For text data, this processing unit performs character encoding normalization or removes unwanted symbols. The system combines this preprocessed data to generate standardized declaration data objects. The output data is structured data objects defined by key-value pairs for each item. The system adds the generated declaration data objects to a message queue system. By using a message queue, this processing unit prevents server overload caused by a sudden surge in fault declarations. The system provides a stable data stream for subsequent database registration or AI parsing modules. Through this asynchronous processing architecture, the system maximizes overall throughput, maintaining real-time performance while achieving high fault tolerance. This system transforms input data in various formats into a unified internal representation, providing a foundation for subsequent machine learning models to efficiently process the data.

[0105] Step 2: Database Registration and Management of Fault Reporting Information. The Case Database Management Department categorizes and tags received fault reporting information according to multiple dimensions, including equipment category, location, fault content, reporting date and time, progress status, reporter attributes, site environmental conditions, fault frequency, past fault history, response performance of the operating company or technicians, site photos or videos, drawings, site environmental sensor values, meteorological data, and traffic volume data, and stores this information in the database. For example, it also manages and links together past fault history of the same equipment category and location, frequency of the same fault content, multiple reports from the same reporter, simultaneous faults of multiple devices at the same site, fault trends in each season, correlation with meteorological conditions, correlation with surrounding construction, past response history, response performance of the operating company or technicians, customer satisfaction evaluations, work completion reports, and dispute history. Specifically, this Case Database Management Department adopts a hybrid storage architecture combining relational databases and graph databases. This system retrieves the reported data objects from the message queue. This processing department extracts structured attribute data from the reported data objects and inserts it into the corresponding tables in the relational database. Simultaneously, to represent the complex relationships between data, this processing unit generates nodes and edges on the graph database. The system defines equipment, location, complainant, and fault details as independent nodes. These nodes are connected by edges representing relationships. To automate multidimensional classification and labeling, the processing unit executes clustering algorithms. For example, the system uses a density-based spatial clustering algorithm to group fault reports that are geographically and temporally close. The input data is a group of 3-dimensional numerical vectors containing the latitude, longitude, and timestamp of each report. The algorithm calculates the Euclidean distance between the vectors and classifies data points densely packed within a certain threshold into the same cluster. The output data is a unique cluster ID assigned to each report. The system appends this cluster ID as an index to the database. By utilizing the cluster ID, the processing unit reduces the execution speed of subsequent queries in similar fault retrieval or trend analysis to logarithmic time. By performing traversal queries of the graph database, the system instantly extracts hidden correlations between multiple devices failing simultaneously at the same location, or between specific weather conditions and fault occurrence. This system, through its highly structured and indexed data, maintains the ability to retrieve necessary information from massive historical datasets in milliseconds. By optimizing the physical storage methods, it eliminates storage I / O bottlenecks, dramatically improving the overall system's response performance.

[0106] Step 3: AI analysis of fault report information and determination of incomplete information. The generative AI analysis department takes a variety of data as input, including fault report information, on-site photos and videos, on-site environmental data, meteorological data, traffic volume data, the response performance of the operating company or technicians, customer satisfaction evaluations, work completion reports, and dispute history. It combines multiple AI technologies such as natural language processing models, image recognition models, anomaly detection models, time series analysis models, clustering models, causal inference models, and risk assessment models to analyze the data. It automatically detects omissions or inconsistencies in text information, unclear on-site photos or videos, insufficient on-site environmental information, unclear dates and times of fault occurrence, and unclear correlations between multiple equipment failures. If the information is incomplete, it automatically sends a rejection notice to the applicant.

[0107] As a specific example of a prompt message that can be read by generative AI, it could be: "Based on the following fault report information and on-site photos, detect any omissions or inconsistencies in the input information, and point out any incompleteness. Fault report information: Equipment type = smart trash can, Location = B-3-chome, Fault content = Communication interruption, Attached on-site photos, Report date and time = June 1, 2024, On-site environment = Construction is underway nearby, Weather = Rain, Request = Urgent response requested."

[0108] Specifically, this generative AI parsing unit implements a multimodal neural network architecture.

[0109] This system is built as a unified pipeline for processing both text and image data simultaneously.

[0110] This processing unit inputs text data into a natural language processing model and extracts context vectors.

[0111] This processing unit inputs image data into a convolutional neural network to extract spatial feature maps.

[0112] The input data is, for example, a 512-dimensional text context vector and a 2048-dimensional image feature vector.

[0113] This system integrates these vectors in the combination layer to generate a 2560-dimensional multimodal feature vector.

[0114] This model inputs the integrated vector into the fully connected layer to determine whether there is any incomplete information.

[0115] This model outputs incompleteness probability scores for each required input item.

[0116] The output data is, for example, an array of floating-point values ​​ranging from 0.0 to 1.0, such as "probability of blurry photo 0.85" and "probability of missing location information 0.12".

[0117] This system compares each output score with a pre-set threshold.

[0118] If any score exceeds the threshold, this processing unit will set the incomplete information flag to true.

[0119] If the incomplete information flag is true, this system will automatically generate a return message containing the specific name of the missing item and an instruction to re-enter it.

[0120] The system will send the generated messages to the applicant's terminal via API as push notifications.

[0121] This system, through its automated judgment process, completely eliminates the need for human managers to visually verify content.

[0122] This system prevents incomplete data from being mixed into subsequent processing, thus maintaining the overall inference accuracy and processing efficiency of the system.

[0123] This system uses multimodal analysis to complement information not recorded in the text but which can be read from the image, or its opposite, to achieve highly accurate incomplete judgment.

[0124] Step 4: AI-based fault content similar fault retrieval and countermeasure inference The generative AI analysis department takes multiple factors as input, including equipment category, installation location, fault content, on-site photos, videos, past fault history, fault frequency, on-site environmental conditions, complainant attributes, date and time of occurrence, detailed fault description, on-site environmental sensor values, meteorological data, traffic volume data, surrounding construction information, past response history, response performance of the operating company or technicians, customer satisfaction evaluation, work completion report, and history of disputes. It retrieves similar past fault cases from the database and performs pattern recognition, statistical inference, anomaly detection, time series analysis, clustering, causal inference, and risk assessment to accurately classify faults as simple faults that can be dealt with through past solutions or complex faults that require on-site investigation or highly demanding handling.

[0125] As a specific example of a prompt message that can be read by generative AI, one could be: "Based on the following fault report information and on-site photos, search for similar past fault cases and infer the best countermeasures. Fault report information: Equipment type = smart trash can, Location = B-3-chome, Fault content = communication interruption, Attached on-site photos, Report date = June 1, 2024, On-site environment = construction nearby, Weather = rain, Request = urgent response requested."

[0126] Specifically, this generative AI parsing unit performs a nearest neighbor search algorithm in a high-dimensional vector space.

[0127] This system inputs the current fault report data into the encoder model to generate a fixed-length query vector.

[0128] The input data is a 1024-dimensional dense vector that integrates text, numerical, and categorical variables.

[0129] This system calculates the cosine similarity between the vectors of past fault cases stored in a vector database and the query vectors.

[0130] This processing unit uses an approximate nearest neighbor search algorithm to extract the top 10 past cases with the highest similarity scores from millions of records, measured in milliseconds.

[0131] This system statistically extracts decision labels from past cases.

[0132] This model integrates similar case data and current environmental data, and inputs them into the classification network.

[0133] This model outputs a suitable probability distribution for each strategy category.

[0134] The output data is, for example, an array of probability values ​​such as "probability of recovery by restart 0.92", "probability of necessary component replacement 0.05", and "probability of necessary detailed on-site investigation 0.03".

[0135] This system selects the strategy category with the highest probability.

[0136] If the chosen countermeasure is one that can be completed remotely or through simple operation, this department will assign a simple fault label to the case.

[0137] This department assigns complex fault codes in cases involving professional diagnostics or physical equipment replacement.

[0138] The system will send the inference result and flag information to the subsequent priority determination module and resource allocation module.

[0139] This system achieves high-precision retrieval based on semantic similarity without relying on keyword consistency by using vector operations in high-dimensional space.

[0140] This system instantly applies a vast amount of past experience and knowledge to current judgments, determining the initial response strategy for fault handling more quickly and accurately than humans.

[0141] Step 5: Fault classification and priority determination Based on the inference results of the generative AI analysis department, the fault determination department and the fault classification department classify faults into simple faults that can be resolved by restarting, firmware updates, component replacement, setting changes, remote operation, and on-site cleaning, and complex faults that require on-site investigation, equipment replacement, system redesign, network reconstruction, sensor recalibration, on-site environmental improvement, and discussion of fault spread prevention measures. Furthermore, they take into account factors such as on-site urgency, fault occurrence frequency, criticality of infrastructure, past complaint history, complainant's requirements, on-site environmental conditions, and fault spread risk to automatically determine the priority of the fault.

[0142] Specifically, the fault determination unit and the fault classification unit execute the gradient boosting tree algorithm as a decision tree ensemble learning model.

[0143] This system receives simple or complex fault flag information output from the front-end modules.

[0144] In addition to the flag information, this processing unit also constructs a feature vector by combining various quantitative features obtained from the database.

[0145] The input data is, for example, a 64-dimensional numerical vector containing integer values ​​from 1 to 5 representing urgency, integer values ​​representing the number of failures in the past month, floating-point numbers representing the importance score of the target device, and probability values ​​representing the risk of impact on the surrounding area.

[0146] This model receives the feature vector as input and evaluates multiple decision trees in sequence.

[0147] This model calculates the final priority score by performing a weighted summation operation on the output values ​​of each decision tree.

[0148] This output data consists of continuous values ​​ranging from 0.0 to 100.0, representing priority scores.

[0149] This output data is mapped to categories such as "top priority" if the score is 85.5 and "normal response" if the score is 42.0.

[0150] The system will write the calculated priority score as an attribute of the case data object.

[0151] This processing department dynamically rearranges the order of cases in the waiting queue based on priority scores.

[0152] When an emergency case with an extremely high score is detected, this system bypasses the usual queue and assigns the case to a dedicated thread for immediate processing.

[0153] This system uses the priority score and classification result as parameters to pass to the subsequent work company selection module or the professional technician allocation module.

[0154] This system accurately assesses complex factors that human operators tend to overlook by conducting multi-faceted risk assessments based on machine learning models.

[0155] This system minimizes the overall damage to urban infrastructure by automatically and optimally focusing limited response resources on the most critical failures.

[0156] Step 6: The company responsible for handling simple malfunctions is automatically selected and outsourced for processing. In cases where a fault is determined to be simple, the work company selection department uses generative AI to comprehensively evaluate multiple parameters such as the work company's or field technicians' past performance, response area, availability, timeframe, cost, quality evaluation, past dispute history, and evaluations from the applicant, and automatically selects the best work company or technicians. The contracting department then automatically generates contracting details (fault details, on-site photos, work instructions, response deadline, requirements, etc.) for the selected work company or technicians and sends them to the work company's or technicians' system.

[0157] As a specific example of a prompt message that can be read by generative AI, it could be: "From the following candidate operating companies, comprehensively evaluate their past performance, service area, availability, timeframe, cost, quality assessment, dispute history, and applicant evaluation to select the best operating company. Candidates: operator_A, operator_B, operator_C. Case information: Smart trash can in 3-chome, B ward, communication interrupted, restart operation required, attached photos of the scene, requesting urgent response."

[0158] Specifically, the company selected a hybrid inference engine that combines constrained optimization algorithms and reinforcement learning models.

[0159] This system retrieves a list of currently operational companies and their respective attribute data from the database.

[0160] This processing department quantifies the attribute data of each operating company and generates a matrix of candidate feature quantities.

[0161] For example, the input data is an M-row, N-column floating-point matrix where each row represents a single operating company and each column represents parameters such as cost, distance, and evaluation score.

[0162] This model receives the matrix and the current case feature vector as input.

[0163] This model calculates a suitability score for each operating company.

[0164] As output data, this model outputs the inherent ID of the job company with the highest fitness score, and an array of weighted coefficients representing the reasons for its selection.

[0165] The output data is structured data such as "Best operating company ID: 789, Reason for selection: Distance score 0.8, Cost score 0.9".

[0166] The system will transmit the selected company ID to this outsourcing department.

[0167] This outsourcing processing department dynamically generates electronic outsourcing document objects in JSON format, containing detailed case information and work instructions.

[0168] This system will send the generated JSON payload to the external system of the operating company as an HTTP POST request via a RESTful API.

[0169] This system confirms that packet transmission is complete by receiving a response to status code 200.

[0170] Through this series of processes, this system can shorten the lead time from the occurrence of a fault to the completion of packet transmission to a few seconds.

[0171] This system achieves both high efficiency in outsourcing and improved quality by mathematically optimizing the trade-off between cost and quality.

[0172] Step 7: Automatic allocation and notification of specialized technical personnel for complex faults. In cases where a fault is determined to be complex, the professional technician allocation department uses generative AI to comprehensively evaluate the professional technician's expertise, workload, past response history, applicable area, and suitability in the technical field, automatically assigning the best professional technician and automatically notifying the technician's terminal of the fault details.

[0173] As a specific example of a prompt for generative AI to read, one could be: "From the following candidates of professional technicians, comprehensively evaluate their professionalism, workload, past handling history, applicable areas, and suitability in the technical field, and select the best professional technician. Candidates: Technician X, Technician Y, Technician Z. Case information: Surveillance camera in C-sector 1-chome, image interruption, on-site investigation required, attach on-site photos, requirement = nighttime operation preferred."

[0174] Specifically, the matching algorithm implemented by the technical personnel allocation department of this profession uses graph neural networks.

[0175] This system is constructed using the technical personnel's skill set, current location information, and schedule as nodes, and their relationships as edges in a graph structure.

[0176] This processing unit adds the elements of complex faults as new nodes to the diagram.

[0177] The input data consists of a matrix representing the feature values ​​of each node and an adjacency matrix representing the connection relationships between nodes.

[0178] This model repeatedly performs message passing processing to update the hidden state vector of each node.

[0179] This model calculates the inner product of the hidden state vectors of the faulty node and each technician node to determine the matching score.

[0180] This model outputs an array of fitness scores ranging from 0.0 to 1.0 for each professional technician.

[0181] The output data is, for example, a list of values ​​such as "Technician X: 0.95, Technician Y: 0.42, Technician Z: 0.68".

[0182] This system extracts the ID of the technician who achieved the highest score.

[0183] Based on the extracted technician ID, this processing department accesses the technician's schedule management database and automatically locks in the required work time.

[0184] This system generates a notification payload that includes detailed fault information, on-site latitude and longitude data, and recommended investigation steps.

[0185] This system uses the WebSocket communication protocol to send real-time push notifications to the mobile terminals of selected technical personnel.

[0186] When the system receives an acceptance response from a technician, it updates the case status to "Technical personnel assigned".

[0187] This system is designed for complex faults requiring a high level of expertise, and uses mathematical methods to instantly assign the most suitable personnel.

[0188] This system prevents mismatch of human resources and significantly shortens the time required to resolve complex faults.

[0189] Step 8: Progress Management, Notification, and Feedback Collection The Progress Management Department automatically collects progress reports from operating companies or technical personnel (arrival at site, commencement of work, restart completed, equipment replacement completed, recovery confirmed, site investigation completed, fault analysis completed, firmware update completed, network reconstruction completed, sensor recalibration completed, site environment improvement completed, fault impact prevention measures implemented, etc.) and notifies managers or applicants in real time. It also implements timeline display of progress status, automatic detection of progress delays, automatic statistics of progress reports, dashboard display of progress status, and automatic generation of progress status reports. The Notification and Feedback Department also collects feedback or questionnaire results from customers or applicants.

[0190] Specifically, the progress management department uses a finite state machine model to strictly manage the state transitions of each case.

[0191] This system continuously listens for status update events sent from the client's terminal via the API endpoint.

[0192] This input data is an event payload in JSON format containing the case ID, the new status code, and a timestamp.

[0193] This processing unit updates the status flags in the database based on the received events.

[0194] This system executes a time series anomaly detection algorithm to monitor progress delays.

[0195] This model employs a long short-term memory network learned from time data of similar past cases.

[0196] This model takes the current state and the elapsed time as input and calculates the predicted transition time to the next state.

[0197] The model outputs an anomaly score representing the deviation between the predicted time and the actual performance time.

[0198] This output data is, for example, a floating-point number in the range of 0.0 to 1.0.

[0199] If the abnormal score exceeds the threshold, the system will set the progress delay flag to true.

[0200] When this processing unit detects a delay flag, it automatically renders a warning alert on the administrator dashboard, triggering escalation processing.

[0201] Whenever the status is updated, this system uses a publish-subscribe messaging model to push and distribute status change notifications to the terminals of relevant administrators or applicants.

[0202] When this system receives a job completion status, it will initiate this notification and feedback process.

[0203] This notification and feedback department will send a message to applicants containing the URL of a dynamically generated questionnaire form.

[0204] This system receives input rating scores or text comments and performs sentiment analysis using a natural language processing model.

[0205] This system associates these quantitative and qualitative data with case records and stores them in a database.

[0206] This system combines real-time tracking and predictive monitoring to prevent operations from becoming a black box, thus dramatically improving the transparency and reliability of project management.

[0207] Step 9: Historical management, continuous learning of AI models, and trend analysis The historical management and learning department automatically records and accumulates the handling history of each fault, the basis for judgment, the reasons for selecting the operating company, the reasons for assigning technical personnel, the progress status, feedback from the complainant, the work completion report, the history of disputes, customer satisfaction evaluation, on-site environmental data, meteorological data, traffic data, surrounding construction information, fault occurrence trends, fault occurrence precursors, fault ripple risks, fault response resource allocation, fault response costs, fault response quality, fault response speed, fault response efficiency, fault response accuracy, fault response automation rate, fault response standardization rate, and fault response optimization rate. This information will be applied to future AI model relearning or business improvement, audit response, fault trend analysis, preventative maintenance plan development, fault occurrence risk assessment, fault ripple risk assessment, optimal allocation of fault response resources, fault response cost reduction, fault response quality improvement, fault response speed improvement, fault response efficiency improvement, fault response accuracy improvement, fault response automation rate improvement, fault response standardization rate improvement, and fault response optimization rate improvement.

[0208] Specifically, the History Management and Learning Department adopts a data lake architecture to integrate and store both structured and unstructured data.

[0209] This system initiates automated batch learning pipelines on a daily or weekly schedule.

[0210] This processing department extracts the most recent application performance data from the data lake and performs data cleaning and normalization.

[0211] This system converts the extracted data into pairs of feature matrices and correct answer label vectors.

[0212] This input data is, for example, a multidimensional tensor composed of tens of thousands of case records.

[0213] This system generates a clone of the currently running inference model and uses the input data to perform additional learning.

[0214] This model outputs predicted values ​​through forward propagation and calculates the loss function between the predicted values ​​and the correct answer labels, which are the actual application results.

[0215] This processing unit, for example, uses cross-entropy error or mean absolute error as the loss function.

[0216] This system uses the backpropagation method to calculate the gradient of each layer in the network and fine-tunes the weight parameters through optimization algorithms such as stochastic gradient descent.

[0217] This model serves as output data, producing a model file containing millions of updated weight parameter groups.

[0218] This system uses the validation dataset to evaluate the accuracy of the updated model.

[0219] When the accuracy metrics exceed the baseline, this processing unit automatically performs hot-swap deployment of the model for the inference server in the production environment.

[0220] This system uses an updated model to recalculate the probability of future failures for infrastructure equipment across the entire city.

[0221] Based on the calculation results, this system automatically generates a list of equipment requiring preventative maintenance and a recommended inspection schedule.

[0222] This system fully automates the continuous integration and continuous delivery cycle.

[0223] This system evolves its inference accuracy with each repeated use, permanently improving its adaptability to environmental changes over the years or unknown failure modes.

[0224] (specific use case) For example, imagine a scenario where a communication interruption occurs in a smart trash can located in a city's residential area, and the remaining amount information is no longer sent to the city's management system.

[0225] In this situation, residents or on-site workers use a smartphone app to enter fault reporting information.

[0226] The application screen requires the following information: smart trash can as the equipment category; "B-zone 3-chome" as the installation location; "communication interruption" as the fault description; a photo of the smart trash can's exterior as a site photo; the applicant's name and contact information; the date and time of the incident; a detailed description of the fault: "sudden communication interruption, unable to confirm remaining amount in the management system"; "road construction is underway in the surrounding area" as site environmental information; and "urgent response requested" as a requirement.

[0227] This information is sent to the server via the fault reporting information acquisition department.

[0228] Specifically, the fault reporting information acquisition department receives an HTTP POST request from a smartphone application at the endpoint.

[0229] This system parses multiple parts of form data from the raw data of the received request.

[0230] This processing unit extracts the values ​​from the text input fields and constructs a payload in JSON format.

[0231] This input data is, for example, a structured object that assigns the string "smart_trash_can" to the "device_type" key.

[0232] At the same time, this processing unit extracts the binary data of the additional appearance photos.

[0233] This processing department performs virus scans and file format verification checks on the image data.

[0234] This processing unit expands the image data into memory, converts it into a Base64 encoded string, or saves it to object storage and retrieves its reference URI.

[0235] This system integrates these text data and image reference URIs to generate a single event object.

[0236] This output data is a data record that has been assigned a UUID that can be uniquely identified within the system.

[0237] The system enqueues the generated event objects into a message queue in memory.

[0238] This processing unit uses the publish-subscribe mechanism of message queues to asynchronously distribute data to subsequent database management modules or AI analysis modules.

[0239] This system completes the data collection and initial processing from the edge device in milliseconds.

[0240] Even in the event of network delays or temporary disconnections, this system completely prevents data loss or duplication through retry control on the client side and idempotent reception processing on the server side.

[0241] This system effectively acquires events occurring in the physical world as high-precision digital data that can be processed by computers.

[0242] The Case Database Management Department categorizes the received information according to equipment type, installation location, fault details, reporting date and time, on-site environmental conditions, past fault history, on-site photos, and applicant attributes, and stores it in the database.

[0243] The generative AI analysis unit takes the received fault report information, on-site photos, and on-site environmental data as input. It uses a natural language processing model to detect omissions or inconsistencies in the text information, an image recognition model to analyze the on-site photos for any appearance abnormalities, setting offsets, or external damage, and an anomaly detection model to infer the tendency or cause of the fault from the on-site environmental data or past fault history.

[0244] For example, if there have been previous communication interruptions at the same location and there are many cases where the problem was resolved by restarting, the AI ​​infers that the issue can also be resolved by restarting this time.

[0245] As a specific example of a prompt message to be read by generative AI, it is set as follows: "Based on the following fault report information and on-site photos, search for similar past fault cases and infer the best countermeasure. Fault report information: Equipment type = smart trash can, Location = B-3-chome, Fault content = communication interruption, attached on-site photos, Report date = June 1, 2024, On-site environment = construction in the surrounding area, Request = urgent response is requested."

[0246] Specifically, the database management department and the generative AI analysis department of this case work closely together to execute a parallel processing pipeline.

[0247] The system launches the AI ​​inference engine in parallel with the write processing to the database.

[0248] This processing unit inputs the image tensors of the on-site photos into a convolutional neural network.

[0249] This model extracts feature maps of edges or textures within an image through multiple convolutional and pooling layers.

[0250] This model outputs the probability score of appearance anomalies and the coordinate data of the bounding boxes representing the abnormal parts.

[0251] Meanwhile, this processing unit inputs the text data into a Transformer-based language model to generate context embedding vectors.

[0252] The input data is a multidimensional tensor that combines image feature vectors and text embedding vectors.

[0253] This system uses the unified tensor to issue similarity search queries against the vector database.

[0254] This system retrieves the solution labels from similar past cases.

[0255] This model integrates this information to calculate the probability distribution of the final strategy.

[0256] This output data is, for example, the output value of the Softmax function, such as the probability of "reboot" being 0.88 and the probability of "module replacement" being 0.12.

[0257] This system adopts the category with the highest probability as the best strategy.

[0258] This processing department packages the deduced countermeasures and the list of IDs of similar cases on which they are based as a JSON-formatted deduction result object.

[0259] This system will pass the inference result to the subsequent fault determination department and the work company selection department.

[0260] This system significantly improves the reliability of inferences by integrating multimodal analysis of images and text, accurately identifying subtle anomalies or complex causes that might be overlooked in a single data source.

[0261] Based on AI's inference results, the Fault Judgment Department and Fault Classification Department categorized the fault as a simple fault that could be resolved by restarting. The Operations Company Selection Department automatically selected an operations company with a proven track record of handling smart trash cans in Block B, 3-chome, and currently available space, using AI. The Outsourcing Processing Department automatically generated an outsourcing request with details such as "Smart trash can in Block B, 3-chome, communication interruption, restart response, attached on-site photos, request for emergency response" and sent it to the operations company's system. The operations company dispatched operators to the site to carry out the restart operation and entered the operation completion report into the system. The Progress Management Department collected real-time progress data, including arrival at the site, commencement of work, restart completion, and restoration confirmation, and notified the manager or the applicant. The History Management and Learning Department recorded and accumulated the entire history from the fault report to the completion of the response, including the judgment criteria, reasons for selecting the operations company, progress status, operation completion report, and feedback from the applicant, and utilized this data for future AI model relearning, fault tendency analysis, and preventive maintenance plan development. Specifically, this system forms a data flow based on a microservice architecture among these modules. The fault determination department evaluates the probability score within the inference results received from the previous stage and sets the simple fault flag as true. The company selection department initiates a spatial search query to the database using this flag as a trigger. The input data consists of the latitude and longitude coordinates of the fault location and an identifier for the required skill combination. The processing department obtains a list of eligible companies and uses a reinforcement learning model to determine the best company. The output data is the unique ID of the selected company. The outsourcing processing department uses this ID to call the outsourcing API and send the electronic outsourcing document data. The system receives outsourcing acceptance callbacks from the company's system. The progress management department continuously monitors GPS coordinate data and status update events sent from the operator's mobile terminal. The input data is a JSON payload containing timestamps and status codes. The processing department overwrites the database record each time the status is updated and synchronizes the front-end display via WebSocket. When the system receives a job completion status, it closes a series of transactions. This historical management and learning department combines all input data related to the task, intermediate tensors of the inference process, output probability scores, and the final assignment results to generate a massive learning data record. This output data is a multi-dimensional vector labeled with the correct answer, which is then used in the next batch learning pipeline. Through end-to-end automation, this system completely eliminates delays or errors caused by human intervention, minimizing the cycle time from failure to recovery.

[0262] Additionally, for example, if a surveillance camera at a city center intersection experiences video interruption, and the surrounding environment is characterized by heavy rain or strong winds and large-scale construction, the on-site operator can report the fault via a tablet terminal, attaching photos or videos, weather data, and construction information. The AI, referencing past faults occurring under similar weather conditions or construction scenarios, infers countermeasures such as equipment replacement, cable reconnection, or improvements to the on-site environment. In cases deemed complex faults, the system automatically assigns a specialist expert in surveillance camera fault analysis and instructs them to conduct an on-site investigation and detailed analysis. A specific example of the prompt text read by the generative AI could be: "Based on the following fault report information, on-site photos, weather data, and construction information, retrieve past similar fault cases and infer the best countermeasures. Fault report information: Equipment type = Surveillance camera, Location = City center intersection, Fault content = Video interruption, Attached on-site photos, Weather = Heavy rain / strong winds, Construction information = Large-scale construction is underway nearby, Request = Nighttime operation preferred." Specifically, this system executes a highly inferential pipeline that comprehensively analyzes multimodal data including dynamically changing environmental factors. This processing unit extracts frame sequences from supplementary video data and inputs them into a 3D convolutional neural network to extract spatiotemporal features. It also inputs time-series data of meteorological or construction information into a recurrent neural network to extract dynamic features of environmental changes. Examples of input data include video frame tensors and time-series arrays of wind speed or precipitation. The system combines these feature vectors to generate a high-dimensional comprehensive context vector. This comprehensive vector is used by the model to explore nearest neighbors in the potential space of past complex fault modes. As output data, the model outputs the probability distribution of the root cause of the fault and complex fault indicators. Examples of output data include numerical data such as "probability of physical cable breakage 0.75" and "probability of short circuit due to water immersion 0.20". When a complex fault indicator is determined to be true, the system sends a trigger signal to the professional technician allocation module. Based on the inferred cause probability distribution, the processing unit determines the required professional skill code. The system retrieves a database of technicians with the corresponding skill codes and considers current location information and schedule to determine the best technician. This system automatically sends detailed work instructions, including the inferred causes and recommended investigation steps, to the terminals of selected technicians. By assembling not only static equipment information but also dynamic environmental or time-series data into its AI model, the system accurately diagnoses failures caused by complex factors that cannot be predicted by rule-based approaches alone. The system automates the deployment of appropriate personnel to the field requiring highly specialized knowledge, preventing prolonged downtime.

[0263] Thus, the automated urban infrastructure fault response system of this invention will automate the entire process for various IoT or communication devices within the city, from fault reporting to automatic determination of incomplete information, automatic comparison with past data, fault classification, automatic selection and subcontracting of operating companies or technicians, progress management, historical management, and AI learning. This achieves efficient use of urban infrastructure, rapid fault response, optimal allocation of human resources, reduced operating costs, improved service quality for urban residents, and enhanced overall urban resilience. Specifically, the system functions as a comprehensive architecture where multiple machine learning models and data processing modules operate asynchronously and coordinately in a distributed computing environment. The system maintains loose coupling between modules by passing standardized JSON-formatted data objects between processing steps. Based on the flow of input fault report data, the system dynamically scales up or down containerized inference nodes in the cloud. The input data consists of time-series indicators representing the overall system CPU utilization, memory consumption, and message queue backlog. The orchestration engine monitors these indicators and automatically optimizes resource allocation. The system outputs system performance metrics or logs of processing latency for each module. This system continuously analyzes log data to identify bottlenecks and performs self-healing and performance tuning. It completely replaces non-technical and error-prone processes such as manual data entry, visual verification, and communication adjustments via phone or email, performed by human operators, with rigorous computer processing based on high-dimensional vector operations and probabilistic inference. By continuously updating the weight parameters of the inference model using accumulated application data, the system autonomously improves judgment accuracy and processing efficiency over time. Through these technological improvements, the system minimizes downtime to mathematical limits in the face of failures within the complex modern urban infrastructure comprised of a vast number of IoT devices. This system most effectively allocates limited computing and human resources, providing a next-generation infrastructure management platform that robustly supports the continuity and stability of urban functions.

[0264] 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.

[0265] Data generation model 58 is what is known as generative AI (Artificial Intelligence).

[0266] One example of data generation model 58 includes ChatGPT (registered trademark) (Internet search).<URL: https: / / openai.com / blog / chatgpt> Generative AI, etc.

[0267] 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.

[0268] Furthermore, the processing performed by the aforementioned data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the smart device 14, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the smart device 14. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the smart device 14 or external devices, and the smart device 14 acquires or collects information required for processing from the data processing device 12 or external devices.

[0269] 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.

[0270] [Second Implementation] Figure 3 An example of the configuration of the data processing system 210 according to the second embodiment is shown.

[0271] 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.

[0272] 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).

[0273] 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.

[0274] 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.

[0275] 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).

[0276] 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.

[0277] 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.

[0278] 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.

[0279] 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).

[0280] 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.

[0281] 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".

[0282] (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.

[0283] (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.

[0284] 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.

[0285] 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.

[0286] 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.

[0287] [Third Implementation] Figure 5 An example of the configuration of the data processing system 310 according to the third embodiment is shown.

[0288] 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.

[0289] 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).

[0290] 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.

[0291] 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.

[0292] 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).

[0293] 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.

[0294] 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.

[0295] 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.

[0296] 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.

[0297] In the head-mounted terminal 314, the processor 46 performs the receiving and output processing.

[0298] 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.

[0299] 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".

[0300] (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.

[0301] (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.

[0302] 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.

[0303] 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.

[0304] 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.

[0305] [Fourth Implementation] Figure 7 An example of the configuration of the data processing system 410 according to the fourth embodiment is shown.

[0306] 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.

[0307] 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).

[0308] 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.

[0309] 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.

[0310] 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).

[0311] 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.

[0312] 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.

[0313] 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.

[0314] 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.

[0315] 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.

[0316] 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.

[0317] 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".

[0318] (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.

[0319] (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.

[0320] 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.

[0321] 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.

[0322] 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.

[0323] 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.

[0324] Figure 9This is a diagram representing an emotion map 400 that maps various emotions. In the emotion map 400, emotions are arranged radially from the center in concentric circles. The closer to the center of the concentric circles, the more primitive the emotion is. Further out on the concentric circles, emotions are arranged representing states or actions arising from mood. Emotion is a concept that includes both feelings and mental states. To the left of the concentric circles, emotions generated by reactions occurring in the brain are arranged roughly. To the right of the concentric circles, emotions guided by situational judgments are arranged roughly. Above and below the concentric circles, emotions generated by reactions occurring in the brain and guided by situational judgments are arranged roughly. Furthermore, the emotion of "pleasure" is arranged above the concentric circles, and the emotion of "unpleasantness" is arranged below. Thus, in the emotion map 400, various emotions are mapped according to the structure of emotion generation, while easily generated emotions are mapped nearby.

[0325] 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.

[0326] 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).

[0327] 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.

[0328] 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."

[0329] 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, nearby sentiment values ​​are similar to each other. Figure 10 Examples show how emotions such as "peace of mind," "stability," and "reassurance" can result in similar emotional values.

[0330] 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).

[0331] 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.

[0332] 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.

[0333] 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.

[0334] 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.

[0335] 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).

[0336] Every processor has built-in or connected memory, and every processor performs specific processes by using memory.

[0337] 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.

[0338] 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.

[0339] 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.

[0340] 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.

[0341] 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.

[0342] In addition, the following notes further explain the above explanation.

[0343] (Note 1) An automated system for handling urban infrastructure faults comprises a fault report information acquisition unit, a fault determination unit, and a fault response and processing unit. The fault report information acquisition unit receives fault reports from residents or managers regarding urban infrastructure equipment such as streetlights, traffic lights, surveillance cameras, environmental sensors, and public Wi-Fi access points. Information includes the equipment category, location, fault details, on-site photos and videos, reporter information, date and time of occurrence, and a detailed description of the fault. The fault determination unit uses generative AI to compare the received fault report information with a historical fault database, analyzing it from multiple perspectives. The system automatically determines whether there are similar cases with the same equipment category, the same location, and the same fault content, as well as the clarity of the solution. It also categorizes the fault into simple faults that can be handled by past solutions, and complex faults that require on-site investigation or more complex handling. Based on the determination results of the fault determination unit, the fault handling unit automatically selects and outsources the work company or on-site technicians using generative AI in the case of simple faults. In the case of complex faults, it automatically assigns professional technicians. Furthermore, it automatically sends a rejection notice to the applicant if the fault report information is determined to be incomplete.

[0344] (Note 2) As described in Appendix 1, the automatic processing system for urban infrastructure fault countermeasures is characterized in that the fault determination unit performs the following processing: inputting multiple elements such as equipment category, location, fault content, on-site photos, videos, past fault history, fault occurrence frequency, on-site environmental conditions, reporter attributes, date and time of occurrence, and detailed description of the fault into a machine learning model based on generative AI, automatically inferring the best countermeasure method for each fault, and classifying it with high precision as a simple fault or a complex fault.

[0345] (Note 3) As described in Appendix 1, the automatic handling system for urban infrastructure fault response is characterized in that the fault response and handling department performs the following processing: In the case of a simple fault, generative AI is used to comprehensively evaluate multiple parameters such as the past performance of the operating company or on-site technicians, the response area, availability, construction period, cost, quality evaluation, past dispute history, and evaluations from the applicant, to automatically select the best operating company or technicians and generate contracting data, and then automatically send the contracting content to the operating company or technicians. On the other hand, in the case of a complex fault, the professionalism, workload, past response history, response area, and suitability of the technical field of the professional technicians are considered to automatically assign professional technicians and notify them of the fault details.

Claims

1. An automated case processing system for radio wave countermeasures, comprising a case information acquisition unit, a case judgment unit, and a case processing unit, characterized in that, The case information acquisition unit receives case information from customers or business operators, including building information, floor information, surrounding base station information, drawings, and operator information. The case judgment unit uses generative AI to analyze the similarity between the input case information and past cases from multiple perspectives, referring to a historical database of radio wave countermeasures. It automatically determines whether there is any incomplete information and classifies the case as a simple case that can be solved by setting up a standard amplifier or a small base station, or a complex case that requires a special amplifier or a new base station. Based on the judgment result of the case judgment unit, the case processing unit automatically selects and contracts out construction companies in the case of simple cases using generative AI, automatically assigns a main operator in the case of complex cases, and automatically sends a rejection notice to the applicant in the case where the case information is determined to be incomplete.

2. The automatic case processing system for radio wave countermeasures as described in claim 1, characterized in that, The case determination department performs the following processing: inputting multiple elements such as building name, floor, number of surrounding base stations, past countermeasure history, radio wave countermeasure status of surrounding buildings, radio wave intensity, floor height, customer attributes, and information of the commissioning department into a machine learning model based on generative AI, automatically inferring the best countermeasure method for each case, and classifying it with high precision as a simple case or a complex case.

3. The automatic case processing system for radio wave countermeasures as described in claim 1, characterized in that, The case handling department performs the following processes: In cases determined to be simple, generative AI is used to comprehensively evaluate multiple parameters such as the construction company's past performance, response area, availability, construction period, cost, and quality evaluation, automatically selecting the best construction company and generating contracting data, and then automatically sending the contracting content to the construction company. On the other hand, in cases determined to be complex, the department considers the professionalism, workload, and past response history of the person in charge, automatically assigns a person in charge, and notifies them of the case details.

Images