Agentic AI-based Water Supply Network Crisis Management Apparatus

Abstract

The present invention is a crisis management device that provides the establishment and execution of autonomous incident response plans by analyzing the physical condition of a water supply network in real time based on a knowledge graph and agentic artificial intelligence. It derives an optimal recovery scenario through inference combining real-time measurement data and expert knowledge, and ensures the reliability of the response through a self-correcting loop that automatically adjusts the plan upon the detection of physical inconsistencies.

Description

Technology Field

[0001] The present invention relates to an agentic AI-based water supply network crisis management device, and more specifically, to a technology that analyzes the physical condition of a water supply network based on a knowledge graph and agentic AI to provide the establishment and execution of an accident response plan. Background Technology

[0002] Practical Limitations and Operational Problems in Water Supply Network Management Field

[0003] As water supply networks are a core national infrastructure maintaining the lifeline of cities, the introduction of smart network management systems combining sensor data and Geographic Information Systems (GIS) is accelerating due to recent advancements in Information and Communication Technology (ICT). However, the following practical limitations exist in actual water supply network operations and accident response sites.

[0004] First, there is the problem of data fragmentation and a lack of real-time consistency. While GIS-based facility data is limited to static physical information, measurement data collected from sensors possesses dynamic time-series characteristics. Existing systems lack organic integration between these heterogeneous data, resulting in significant delays in comprehensively understanding the phase structure of the pipe network and real-time hydraulic conditions in the event of an accident; this is a major cause of difficulty in securing the initial golden time.

[0005] Second, there is a lack of an intelligent judgment system for complex decision-making. Accidents in water supply networks vary widely, ranging from simple leaks to complex pressure anomalies; responding to them requires both sophisticated hydraulic analysis simulations and the judgment of skilled experts based on past accident history. However, current systems are limited to providing simple threshold alarms or numerical data from individual models, leaving the majority of the process of interpreting collected information and establishing concrete recovery plans dependent on the operators' experiential judgment.

[0006] Third, there is a lack of validation and execution capability regarding accident response scenarios. It is difficult to accurately predict in advance the secondary damage (e.g., widespread water cutoff or expansion of low-pressure zones) that may occur when response measures derived from a crisis situation are applied to actual waterworks. In particular, as specialized knowledge—such as unstructured data like repair manuals or accident reports—is not integrated with real-time analysis models, issues are continuously being raised, such as response plans being established that are detached from the actual situation on the ground or unexpected physical inconsistencies occurring during the execution process.

[0007] Therefore, there is an urgent need to develop advanced crisis management technology capable of recognizing the complex physical conditions of water supply networks in real time, integrating expert response knowledge, and autonomously establishing and verifying optimal response plans.

[0008] Limitations of Patent Literature and Technical Gaps

[0009] Patent documents 1 to 3 described as prior art are intended for real-time accident detection, water quality tracking, data correction, etc., in relation to water supply network management. However, these patent documents contain technical limitations as follows.

[0010] Patent documents 1 and 2 have a fragmentary analysis structure centered on unit models. Patent document 1 primarily concerns accident detection through pressure prediction in specific sections, while patent document 2 focuses on individual analysis algorithms, such as tracking points of water quality anomalies. Consequently, there are limitations in organically selecting and linking various analysis tools to derive comprehensive conclusions regarding complex crisis situations occurring within a pipe network, such as accidents involving a combination of leakage sounds and water pressure anomalies. In other words, there is a lack of a higher-level intelligent system capable of orchestrating multiple models according to the situation.

[0011] Patent Document 2 exhibits a discrepancy between static data modeling and real-time dynamic states. As a system based on an existing network model, Patent Document 2 primarily defines the connection relationships of facilities and lacks a structure capable of dynamically processing real-time changing measurement data, such as water pressure, flow rate, and vibration, within a physical phase structure by mapping them to node attributes of a knowledge graph in real time. This reduces the system's ability to perceive the current hydraulic state of the pipe network in three dimensions.

[0012] Patent Document 3 lacks a system for utilizing unstructured operational knowledge. Patent Document 3 relates to water data preprocessing or anomaly data correction and is specialized only in the processing of structured data; it does not provide a Search Augmented Generation (RAG)-based inference function that immediately calls up and reflects unstructured data, such as field repair manuals or past accident response reports, into the real-time accident response process. Consequently, a technical disconnect occurs in which the tacit knowledge of experts cannot be reflected in the system's execution plan.

[0013] Furthermore, the patent documents describe passive control methods lacking self-correction capabilities. Existing systems merely output results based on predefined logic or fixed scenarios; they lack an active response system in which an agent independently plans an execution strategy and, if the execution result violates a physical threshold, traces back the inference path to modify or roll back the plan.

[0014] In conclusion, while conventional methods contribute to the monitoring of individual data or the prediction of specific models, they show distinct limitations in providing autonomous and complete crisis management solutions by integrating the complex physical characteristics of water supply networks with expert response logic. Prior art literature

[0015] Korean Registered Patent No. 10-2493362, Korean Published Patent No. 10-2025-0070522, Korean Published Patent No. 10-2025-0006676 The problem to be solved

[0016] To address the aforementioned problems, the present invention aims to resolve the fragmentation between static facility information generated at water supply network management sites and real-time dynamic measurement data, and to organically combine them to secure the golden time for accident response. It seeks to improve the speed and accuracy of response to complex crisis situations by transitioning the existing manual decision-making system, which relied entirely on the experiential judgment of individual operators, into an AI-based intelligent orchestration system.

[0017] The present invention aims to ensure the completeness of crisis management by implementing an active self-correction function that verifies the validity of an execution plan and traces back the inference path to correct or roll back the plan, thereby preventing secondary damage that may occur when a response scenario derived by artificial intelligence violates the physical laws of an actual water supply network. means of solving the problem

[0018] An agentic AI-based water supply network crisis management device according to an embodiment of the present invention for the above-mentioned problem to be solved comprises: a graph engine (110) that receives facility coordinates and pipeline data from a GIS (300), structures the topology of the water supply network by setting facilities as nodes and connections between facilities as edges through spatial join operations, and generates a knowledge graph in which physical connection information and dynamic state values ​​of facilities are integrated by mapping real-time measurement data to the nodes; an agent controller (120) that analyzes the physical state of the water supply network by searching the topology structure of the knowledge graph to obtain a range of hydraulic influence between facilities and extracting a dynamic state value mapped to at least one node included within the range of hydraulic influence; and a tool interface (130) that performs at least one of hydraulic analysis and accident diagnosis according to an execution plan generated for each unit task by analyzing the physical state, and is characterized by linking the topology structure of the knowledge graph with the dynamic state value to limit the range of hydraulic influence and providing the establishment and execution of an accident response plan based on the physical state analysis within the limited range.

[0019] The agent controller may include: a task planning unit (121) that generates multiple unit tasks by exploring a knowledge graph through a large-scale language model; a sequence determination unit (122) that sets the execution order in response to an accident scenario requiring the combined use of a hydraulic analysis model that simulates the physical flow of the water supply network and an intelligent diagnostic model that analyzes data patterns; a parameter conversion unit (123) that converts the unit tasks into control signals containing API call parameters corresponding to each input protocol and outputs them for technical compatibility with different input specifications for each hydraulic analysis model and intelligent diagnostic model; and a validity verification unit (124) that verifies validity by comparing the result returned from the tool interface with a mathematical threshold for each node defined in the knowledge graph, and, if a contradiction is found, feeds back the cause of the deficiency to the task planning unit to correct the API call parameters. The agent controller may be characterized by performing a self-correcting loop that integrates and controls the mutually heterogeneous hydraulic analysis model and intelligent diagnostic model based on API call parameters and secures the physical reliability of the execution plan through threshold-based feedback of the knowledge graph.

[0020] The above tool interface may be characterized by identifying accident information by operating an intelligent diagnostic model composed of supervised learning that classifies accident types by learning past leakage accident patterns and unsupervised learning that detects abnormal signs based on the reconstruction error of input data by learning the vibration frequency of a normal state, linking the accident information as an input condition for a hydraulic analysis model to predict water cutoff zones and low-pressure zones by simulating changes in water pressure within the pipe network due to virtual valve operation or pump operation, and the sequence determination unit may be characterized by mutually comparing the range of water cutoff damage according to multiple simulation results returned from the tool interface to select a recovery scenario in which water cutoff damage is minimized, and finally confirming the control target and operation sequence according to the selected scenario as an execution plan. Effects of the invention

[0021] The present invention organically integrates fragmented static information of facilities and real-time dynamic measurement data within a knowledge graph, thereby enabling three-dimensional recognition of the status of the observation network in the event of an accident and significantly shortening the golden time for initial response.

[0022] The present invention reduces the amount of computation that artificial intelligence must process by logically limiting the range of hydraulic influence through the topological structure of a knowledge graph instead of the entire water supply network, and enables the establishment of an accurate response plan within seconds even in large-scale water networks.

[0023] The intelligent orchestration through the work planning unit and sequence determination unit of the present invention moves away from the existing method that relied on the empirical judgment of experts and realizes object-oriented decision-making based on data and physical laws. In particular, by simulating multiple recovery scenarios in a virtual space through a tool interface and comparing and selecting a method that minimizes damage, it is possible to prevent secondary damage caused by human error and autonomously derive the optimal recovery solution.

[0024] In the event that a plan proposed by artificial intelligence violates actual mathematical thresholds, the validity verification unit and the state management unit immediately intervene to identify the error and perform a self-correcting loop to roll back the system to a normal state or adjust parameters. This active response system corrects logical errors that agentic AI may commit using physical formulas, ensuring the completeness of the final incident response plan and serving as a foundation for maintaining the stability of water supply operations at all times.

[0025] The combination of the RAG module and the language user interface of the present invention converts complex mathematical analysis results and inference paths into natural language reports and intuitive visualization data, thereby significantly reducing the workload of operators and enhancing the explainability of the system. Managers can converse with artificial intelligence without specialized data processing skills to verify precise analytical grounds in real time and perform highly reliable response tasks that can be immediately applied in the field. Brief explanation of the drawing

[0026] FIG. 1 is a block diagram illustrating a water supply network system according to an embodiment of the present invention. Figure 2 is a block diagram illustrating the water supply network crisis management device of Figure 1 in detail. Figure 3 is an example illustrating the types and meanings of nodes. Figure 4 is an example illustrating the types and meanings of edges. Figure 5 is an example illustrating the configuration of detailed attributes of a flow direction edge. Figure 6 is an example illustrating the classification of detailed attributes held by a measurement point node and their meanings. Figure 7 is an example illustrating an ontology-based knowledge graph. Figure 8 is a block diagram illustrating the agent controller of Figure 2 in detail. Figure 9 is an example illustrating orchestration and scenario optimization methods. Figure 10 is an example of making a final decision by combining the decision results of supervised and unsupervised learning models. FIG. 11 is an example illustrating the underlying mechanism of a state-aware self-correcting loop and domain-specific readjustment rules adopted by the agent of the present invention to flexibly respond to crisis situations. Figure 12 illustrates an example of the detailed configuration of additional self-correcting triggers implemented in the form of multilayer middleware to address five failure modes that are not resolved by the underlying mechanism, the readjustment loop alone. Figure 13 is an example illustrating five types of predefined guard rules explicitly included within natural language descriptions possessed by each expert tool, and the corresponding technical features. FIG. 14 is an example illustrating a surveillance crisis management agentic AI architecture according to the implementation of the present invention. Figure 15 is an example illustrating the specific workflow of an agentic artificial intelligence for water supply network crisis management designed based on Langraff. Figure 16 illustrates an example of a data integration management system in which a graph engine and an agent controller are organically combined through search, augmentation, and generation technologies to derive a final execution plan. Figure 17 illustrates an example of an intelligent anomaly detection model based on leakage sound applied to a water supply network agentic artificial intelligence system. FIG. 18 illustrates an example of an autoencoder-based unsupervised learning model that creates a reference model based on data from a normal operating state and then detects abnormal phenomena that deviate from it. Figure 19 illustrates an example of the analysis process of agentic artificial intelligence that identifies abnormal signs based on water pressure data and estimates suspected leakage sections. Figure 20 illustrates a case in which the agentic artificial intelligence system for water supply network crisis management of the present invention was applied on a trial basis in an actual field. Figure 21 illustrates the actual screen configuration of the language user interface of the present invention through a pilot application example. Figure 22 illustrates an example of four key analysis screens implemented through the language user interface of the water supply network crisis management agentic artificial intelligence. Figure 23 illustrates a specific pilot application case for detecting a suspected leak section based on water pressure analysis. Specific details for implementing the invention

[0027] Embodiments of the present invention will be described in detail below with reference to the attached drawings and the contents described therein, but the present invention is not limited or restricted by the embodiments.

[0028] FIG. 1 is a block diagram illustrating a water supply network system according to an embodiment of the present invention, wherein the water supply network system (10) includes a water supply network crisis management device (100), an IoT sensor (200), a GIS (300), and a user terminal (400).

[0029] The water supply network crisis management device (100) is the core entity of the present invention and corresponds to a management server or computing device that integrates heterogeneous data received from the outside to generate a water supply network knowledge graph and autonomously establishes, executes, and verifies an accident response plan based on Agentic AI technology.

[0030] The IoT sensor (200) is a device installed at key points (pipes, valves, pumps, etc.) of the water supply network to collect measurement data such as water pressure, flow rate, vibration, and water quality in real time. Time series data collected from the IoT sensor (200) is transmitted to the water supply network crisis management device (100) and mapped to the dynamic state values ​​of individual nodes within the knowledge graph, thereby reflecting the current hydraulic state of the water supply network in real time.

[0031] The Geographic Information System (GIS) (300) is a database that stores and manages static spatial data including the coordinates, pipe diameter, pipe type, year of installation, and connection information between facilities of water supply network facilities. The water supply network crisis management device (100) can form a basic topology of the water supply network based on data received from the GIS (300), and based on this, can establish a foundation for exploring the range of physical impact.

[0032] The remote monitoring and control system (SCADA) may include information for monitoring and controlling the overall operational status of the system, such as the operating status of pumps at major pump stations or reservoirs, valve opening rates, and large flow meter readings. The combined information from these two systems (GIS and SCADA) is transmitted to a water supply network crisis management device (100) and can be used as basic data to define the physical structure of the network and analyze the range of influence. For convenience of explanation, in the present invention, these two systems (GIS and SCADA) will be described as GIS (300).

[0033] The user terminal (400) is an interface device for a user or manager to monitor the status of the water supply network and to approve or control a response plan established by AI, and may include a smartphone, a tablet PC, or a workstation in a control room.

[0034] The user terminal (400) is internally equipped with a language user interface (LUI) and can integrate and output analysis results and reasoning grounds transmitted from the water supply network crisis management device (100) in the form of map-based visualization data, multi-charts, and natural language-based detailed reports.

[0035] FIG. 2 is a block diagram illustrating in detail the water supply network crisis management device of FIG. 1, wherein the water supply network crisis management device (100) includes a graph engine (110), an agent controller (120), and a tool interface (130).

[0036] The graph engine (110) can process vast amounts of data flowing in from the GIS (300) to form the basis of an intelligent knowledge graph. First, by analyzing the latitude and longitude-based facility coordinates of the geographic information system (GIS), individual facilities such as pipes, valves, pumps, and reservoirs can be defined as nodes, which are vertices of the knowledge graph. At this time, by analyzing the physical connection type between adjacent facilities and connecting them with edges, the three-dimensional topological structure of the entire water supply network can be logically modeled.

[0037] The graph engine (110) can combine dynamic data, such as water pressure, flow rate, valve opening rate, and pump operating status collected from IoT sensors (200) and a remote monitoring and control system (SCADA), with the attributes of each node of the knowledge graph in real time. Through this dynamic mapping process, the knowledge graph functions as a dynamic intelligent entity that goes beyond a simple facility map and reflects the hydraulic energy and physical conditions currently flowing within the network in real time.

[0038] In particular, the graph engine (110) can independently and organically operate a graph database (191) that manages the connection relationships of facilities and a time series database (192) that manages measurement data in a time series.

[0039] Additionally, the graph engine (110) can generate an up-to-date knowledge graph necessary for analysis by immediately mapping the latest measurements from the time-series database (192) to the nodes at the time when the agent controller (120) requests analysis of a specific area. This data integration system enables the complex physical connectivity and real-time repair status of the water supply network to be understood within a single graph structure, thereby serving as a basis for subsequent steps such as limiting the scope of impact and establishing an accident response plan.

[0040] Figure 3 is an example illustrating the types and meanings of nodes. The knowledge graph nodes of the present invention may include water source, pump, valve, junction point, water supply tap, flow meter, and water pressure gauge nodes according to the hydraulic role of the water supply network.

[0041] Each facility node can possess physical attributes necessary for hydraulic analysis, such as material, specifications, capacity, installation elevation, and coordinates. For example, a pump node includes pump specifications and performance curve information, while a valve node manages valve diameter and opening / closing characteristic information as attribute values. These physical attributes can be utilized as basic input values ​​for hydraulic analysis simulations performed by the agent controller.

[0042] The hydraulic roles of each node are examined in detail as follows. Water source nodes supply energy to the pipeline network system, pump nodes add hydraulic energy, and valve nodes can regulate flow or dissipate energy. Junction nodes maintain mass conservation at the branching and joining points of the pipelines, while water tap nodes refer to the points where water flows out from the final point of consumption.

[0043] Flow meter and pressure gauge nodes can function as measurement point nodes to measure real-time physical quantities at the site and update the dynamic state of the knowledge graph. In particular, measurement point nodes additionally possess semantic vector embeddings, making them targets for natural language search. This configuration enables the agent controller to quickly search for and identify the measurement point within the knowledge graph that is semantically most similar to the query when analyzing a user's natural language query.

[0044] In the present invention, the node classification system combines mathematical meaning and intelligent search functions with the physical connection relationships of a water supply network, thereby providing a data basis that enables agentic AI to accurately recognize and respond to complex hydraulic environments.

[0045] FIG. 4 illustrates an example of the types and meanings of edges, wherein a spatially included edge is an edge that defines which block polygon a specific node is located within. This edge connects the geographical zone to which the facility belongs and can hold the result of a spatial operation as an attribute. Through this, the graph engine (110) can spatially recognize which administrative zone or drainage zone the facility of the water supply network is included in.

[0046] The operational affiliation edge is an edge that defines which operational entity a specific block belongs to. It hierarchically connects the management entities or operational units of the water supply network and serves to clarify the affiliation relationship between the facility and the management entity. The agent controller (120) can utilize this edge to identify the operational entity responsible for the area in the event of a specific accident and derive customized response guidelines.

[0047] A flow direction edge is an edge that defines the physical direction in which water flows from an upstream node to a downstream node. It represents the most critical dynamic relationship of a water supply network and can hold attributes such as distance-based hydraulic damping rates and distance information between points. Beyond simply representing the state of connected pipes, this edge serves as a foundation for sophisticatedly combining simulation results of hydraulic analysis models with knowledge graphs by digitizing and managing the movement paths and damping characteristics of hydraulic energy.

[0048] As such, the three types of edges illustrated in Fig. 4 can connect the water supply network three-dimensionally from three perspectives: space (including space), organization (operational affiliation), and physical (flow direction). Through this multi-faceted edge configuration, the water supply network crisis management device can spatially determine the scope of impact of an accident, transmit information to the appropriate operating entity, and accurately infer hydraulic changes based on the flow of water.

[0049] Weights can be calculated for flow direction edges to quantify the change in hydraulic energy that occurs when fluid moves from an upstream node to a downstream node. The calculation of weights is based on a mathematical model that combines the physical length of the pipeline with the hydraulic damping characteristics occurring per unit distance.

[0050] The weighting of the flow direction edge does not simply represent a physical straight-line distance, but is a numerical value representing the energy lost as the fluid passes through the pipe. The distance-based hydraulic damping rate may include physical resistance determined by the pipe material, diameter, roughness coefficient, and gradient.

[0051] The graph engine (110) can store the calculated result value as edge attribute data. The agent controller (120) can analyze this weight when establishing an accident response plan to infer, in three dimensions, the magnitude of the pressure drop and the change in flow rate transmitted to the downstream node when an accident occurs at a specific node.

[0052] This weight calculation method enables agentic AI to rapidly predict logically valid impact ranges even before undergoing complex hydraulic analysis simulations by encoding the complex physical environment of the water supply network as data within a knowledge graph.

[0053] Figure 5 illustrates an example of the detailed attribute configuration of a flow-direction edge, wherein the flow-direction edge has a multidimensional attribute system including distance-based weights and distance values ​​to precisely represent the physical characteristics and hydraulic changes of the water supply network.

[0054] Distance-based weighting is an attribute that reflects the physical phenomenon in which hydraulic pressure gradually decreases as the distance between nodes increases. This attribute includes a hydraulic damping coefficient that indicates the characteristic of energy loss as the distance increases. The agent controller (120) can use this coefficient to mathematically infer the degree of pressure drop according to distance when an accident occurs at a specific point.

[0055] The distance value refers to the actual physical separation distance between two connected nodes on the knowledge graph. This data represents the distance across the Earth's surface calculated based on the latitude and longitude coordinates of each facility. This information can serve as basic data for determining the actual length of a pipeline and for calculating the time required for a fluid to pass through a section or the friction loss within the pipeline.

[0056] In addition to such distance-related information, the edge can hold status information containing critical water pressure, which indicates the physical limits of the section, and real-time operational status. It can also be configured with a flow direction vector, which mathematically defines the direction of fluid movement, and connection time information to determine the sequence of data.

[0057] This configuration of detailed attributes enables the knowledge graph constructed by the graph engine (110) to go beyond a simple network and perfectly simulate the physical constraints of an actual water supply network. Through this, the water supply network crisis management device can organically combine real-time measurement data and simulation results to establish a foundation for autonomously deriving physically valid and highly reliable crisis response scenarios.

[0058] Figure 6 illustrates an example of the detailed attribute classification and meanings possessed by a measurement point node. The measurement point node has an attribute system including an identifier, spatial, topographical, hydraulic reference value, semantic embedding, and incremental key to recognize the dynamic state of the water supply network and communicate organically with the agentic AI.

[0059] An identifier is unique information used to uniquely distinguish individual measurement points within a system. It includes node IDs and facility names, and serves as the basis for clearly defining the location and role of specific sensors within a complex water supply network where numerous facilities are intricately intertwined.

[0060] Spatial attributes refer to coordinate information, such as latitude and longitude, received from the GIS (300). This information determines the physical location of the corresponding node on the knowledge graph and is used as basic geographical data to calculate the distance from the accident site or the range of hydrological impact.

[0061] Topographic attributes include information on the installation elevation or sea level of the location where the measurement point is situated. In hydraulic analysis of water supply systems, elevation is a critical factor in determining positional head and calculating water pressure; therefore, topographic attributes can play a role in ensuring consistency between actual measurements and the simulation results of the hydraulic analysis model.

[0062] The repair threshold value is threshold data that defines the range of normal water pressure or flow rate that must be maintained at the point. The agent controller (120) can detect abnormal signs or determine the physical validity of the established response plan by comparing real-time measurement data or the execution results of an analysis tool to see if they fall outside the range of this threshold value.

[0063] Semantic embeddings are numerical data that transforms text information, such as node names, roles, and surrounding environments, into a high-dimensional vector space. This attribute forms the basis of intelligent search, enabling large-scale language models to rapidly explore and identify the most semantically relevant measurement points within a knowledge graph when processing user natural language queries or commands.

[0064] The incremental key is an identification key that manages the inflow order or timestamp information of time-series data collected from the IoT sensor (200). Through this key, the graph engine (110) accurately synchronizes the vast measurement data of the time-series database and the node state of the knowledge graph in chronological order, and can ensure data consistency during the rollback process of restoring the observation state at a specific point in time.

[0065] This systematic configuration of attributes enables measurement point nodes to function as intelligent objects that comprehensively possess the physical environment and operational knowledge of the water supply network, going beyond merely transmitting numerical values.

[0066] Figure 7 illustrates an example of an ontology-based knowledge graph, which provides a logical data structure designed to enable an artificial intelligence agent to systematically understand the physical configuration and water flow of the entire water supply network system. This goes beyond simply displaying geographical locations to hierarchically define the hydraulic relationships and operational systems between facilities.

[0067] The nodes of the knowledge graph consist of key elements constituting the water supply network, such as reservoirs, sub-blocks (DMA), pumping stations, pressure reducing valves, and IoT sensors (200). Each node holds physical specifications and real-time status data of the corresponding facility and functions as a unit object that allows artificial intelligence to immediately identify the role of a specific point within the network.

[0068] Edges defining the relationship between nodes clearly define water flow and ownership and inclusion relationships. Water flow relationships represent a hydraulic path starting from a reservoir or pumping station and extending along a pipeline to the final consumer. Ownership and inclusion relationships represent a logical topology indicating which sub-block (DMA) area a specific pressure reducing valve or IoT sensor (200) belongs to, or which operating entity manages that area.

[0069] This ontology-based structuring enables the agent controller (120) to perceive the entire complex water supply network as a single organic intelligent entity. For example, if abnormal water pressure is detected by an IoT sensor within a specific sub-block, the artificial intelligence can search the knowledge graph to immediately identify upstream pumping stations or pressure reducing valves supplying water to that area and establish an optimal shut-off or adjustment plan.

[0070] As a result, the knowledge graph exemplified in Fig. 7 integrates static information and dynamic states of the water supply network into a form that artificial intelligence can infer, thereby serving as a core knowledge base that enables agentic AI to autonomously generate physically valid and sophisticated crisis response scenarios.

[0071] The graph engine (110) receives facility coordinates and pipeline data from the GIS (300), structures the topology of the water supply network by setting the facilities as nodes and the connections between facilities as edges through spatial join operations, and generates a knowledge graph in which physical connection information and dynamic state values ​​of the facilities are integrated by mapping real-time measurement data to the nodes.

[0072] A spatial join operation is an analysis technique that combines two or more spatial datasets based on geographical location relationships. In this invention, the process involves spatially overlapping point data of facilities provided by a Geographic Information System (GIS) with line data of pipelines to determine which facilities a specific pipeline is connected to and to logically connect them.

[0073] Water Pipe Network Topology is a mathematically and logically structured representation of the physical connectivity and arrangement of facilities and pipelines that constitute a water supply network. Going beyond simply displaying facilities on a map, it is the network framework defined to enable artificial intelligence to understand which paths water can flow through and how changes at one point are transmitted to other points.

[0074] Physical Connection Information refers to the static specifications and connection status of the facilities constituting the pipeline network. This may include the unique physical attributes of the facilities, such as pipe material (pipe type), diameter, length, and installation year. This information serves as the basic data for calculating fluid frictional resistance or energy loss during hydraulic analysis.

[0075] The dynamic status value is variable measurement data that changes in real time during the operation of the water supply network. This includes current water pressure, flow rate, and water quality concentration collected from IoT sensors (200) and the SCADA system, as well as the open / closed status of valves and the operating frequency of pumps. Through this value, the system can recognize the current physical phenomena of the network in real time.

[0076] A Knowledge Graph is a data model structured in the form of a semantic network that organically connects individual facilities (nodes) and the relationships (edges) between them, combining physical information and dynamic states. In this invention, by integrating mathematical knowledge of the water supply domain with real-time field data, it can function as an intelligent database that enables agentic AI to infer the scope of impact in the event of an accident and autonomously generate recovery scenarios.

[0077] The agent controller (120) searches the topological structure of the knowledge graph to obtain the range of hydraulic influence between facilities and extracts a dynamic state value mapped to at least one node included within the range of hydraulic influence to analyze the physical state of the water supply network.

[0078] The Hydraulic Impact Range is a logical and spatial area where hydraulic fluctuations (pressure drop, flow rate change, damage, etc.) occurring at a specific node are physically propagated along the phase structure of the water supply network, causing significant changes. This does not simply mean the entire range where pipes are connected, but is the result of calculating the effective distance over which energy is transmitted by considering the fluid flow characteristics and the weights (distance-based hydraulic damping rate) defined on the flow direction edges. By limiting this range, the agent controller (120) can identify key analysis target nodes directly linked to accidents among the vast network data.

[0079] The physical state is the overall status of the water network, combining static attributes—the inherent design specifications of the water supply network facilities—and dynamic attributes—the operational data that changes in real time.

[0080] Static attributes are physical entity information of the facility, such as pipe diameter, pipe type, year of burial, and terrain elevation. Dynamic attributes are real-time water pressure, flow rate, vibration, valve opening rate, etc., measured through IoT sensors and SCADA.

[0081] The agent controller (120) analyzes these physical conditions to diagnose whether the current observation network is operating within a normal range and can use this as basic data to verify whether the established execution plan does not violate actual physical laws (hydraulic mechanics).

[0082] The tool interface (130) analyzes the physical state and performs at least one of hydraulic analysis and intelligent diagnosis according to the execution plan generated for each unit task. The hydraulic analysis model is a physical analysis AI tool that simulates changes in fluid flow, water pressure distribution, and water quantity within a water supply network using mathematical equations based on fluid dynamic principles (e.g., the Hazen-Williams formula). It utilizes static attributes such as pipe diameter, material, roughness coefficient, and facility elevation as basic data and can simulate hydraulic changes according to scenarios such as accidents at specific points or valve operations. Through this, it can quantitatively calculate predicted water cutoff zones or low-pressure areas and perform the role of numerically verifying the physical validity of the response plan established by the agent controller (120).

[0083] The Intelligent Diagnosis Model is a data analysis AI tool that utilizes artificial intelligence and machine learning algorithms to detect abnormal patterns in measurement data of a water supply network and classify types of accidents. It can determine whether there is a deviation from a normal state by analyzing the characteristics of time-series data, such as water pressure, flow rate, and vibration, flowing in from IoT sensors (200). Based on past accident learning data, it identifies the causes of accidents such as leakage, pipe breakage, and sensor malfunction, and provides the function of detecting subtle abnormal signs that are difficult to find using only physical formulas within complex unstructured data at an early stage and transmitting them to an agent.

[0084] The agent controller (120) links the topological structure of the knowledge graph with dynamic state values ​​to limit the range of mathematical influence and provides the establishment and execution of an accident response plan based on the analysis of physical state within the limited range.

[0085] The agent controller (120) of the present invention can perform an intelligent orchestration function that helps artificial intelligence autonomously recognize and determine the complex physical environment of a water supply network.

[0086] The agent controller (120) explores the knowledge graph constructed by the graph engine (110) and organically links the topological structure, which is the physical connection relationship between facilities, with the real-time dynamic state values ​​received from the sensors. Through this, the path through which hydraulic energy propagates from a specific accident point can be calculated, thereby logically limiting the range of hydraulic influence, which is the area requiring analysis. This can serve as a filtering function to increase the concentration of analysis by selecting only the core nodes and edges directly related to the accident from the entire network data.

[0087] The agent controller (120) can precisely analyze the physical condition of the water supply network by integrating the physical attributes (elevation, pipe diameter, pipe type, etc.) possessed by nodes within a limited range with current measurements. Based on the analyzed condition, it can determine the type of accident through a reasoning process based on a large-scale language model (LLM) and establish an accident response plan by designing an optimal response scenario through a tool interface (130). Finally, by providing the execution of control commands applicable to the actual field, such as valve operation or pump control, according to the established plan, it can perform an integrated control function that autonomously leads the entire process from situation awareness to action.

[0088] The operation of such an agent controller (120) can improve the speed of crisis response and computational efficiency. Unlike the conventional method of performing analysis on the entire water supply network, by limiting the range of hydraulic impact in advance through a knowledge graph, the amount of data computation that artificial intelligence must process can be drastically reduced. Through this, a golden time is secured to identify the range of impact and establish a response plan within seconds when an accident occurs, even in large-scale water networks.

[0089] In addition, the operation of the agent controller (120) can improve the reliability of accident diagnosis and response plans. Because it analyzes by combining the physical topography of the water supply network with actual terrain information rather than simply detecting numerical changes in the sensor, it can reduce the possibility of misjudgment due to data errors. Since it derives only physically executable plans based on real-time dynamic state values, it minimizes the discrepancy between the field situation and the simulation results and enables sophisticated crisis response.

[0090] In addition, the operation of the agent controller (120) can realize intelligent and autonomous operation management. By autonomously performing the determination of the scope of an accident and the creation of response scenarios, which previously required expert intervention, the workload of the manager can be significantly reduced. In particular, even in situations where management personnel are scarce, such as at night or on holidays, the system analyzes the physical condition on its own and presents an optimal execution plan, thereby preventing secondary damage caused by human error and maintaining the stability of water supply operations at all times.

[0091] FIG. 8 is a block diagram illustrating the agent controller of FIG. 2 in detail, wherein the agent controller (120) may include a work planning unit (121), a sequence determination unit (122), a parameter conversion unit (123), a validity verification unit (124), a state management unit (125), and an inference analysis unit (126).

[0092] The work planning unit (121) can generate multiple unit tasks by exploring a knowledge graph through a large language model. The large language model (LLM) is an artificial intelligence neural network model that learns a vast amount of text data and can understand context and reason at a level similar to humans.

[0093] The large-scale language model in this invention goes beyond merely generating natural language; it performs the role of an 'inference engine' that combines specialized knowledge of the water supply domain with real-time network conditions to analyze the causes of accidents and logically design solutions. In particular, by converting the topological relationships and dynamic attributes of facilities stored in a knowledge graph into text-based prompts and receiving them as input, it makes intelligent decisions that comply with physical constraints.

[0094] A unit task is the smallest execution unit of sub-tasks that must be performed to complete the entire accident response process. It refers to individual processes broken down into smaller parts to solve a single complex problem. For example, given a macroscopic goal of "resolving water pressure drop in a specific area," tasks such as "extracting a list of valves adjacent to the accident location," "executing a repair analysis simulation during a virtual shutdown," and "generating user notification messages based on the predicted water cutoff area" become individual unit tasks. Each unit task has a self-contained input / output structure and takes the form of being matched with a specific analysis tool and executed through a tool interface (130).

[0095] The work planning unit (121) can operate as follows by utilizing this large-scale language model and unit task concept. After exploring the physical topography of the water supply network and currently measured abnormal signs through the knowledge graph, it can apply the 'Chain of Thought' reasoning technique of the large-scale language model. Through this process, it can derive the optimal work scenario to be performed in the current situation and structure it into multiple unit tasks that can be executed sequentially or in parallel.

[0096] Unlike conventional methods that rely on predetermined scenarios, this method allows artificial intelligence to flexibly design response procedures in accordance with real-time field conditions reflected in the knowledge graph. As a result, the work planning unit (121) can serve as a starting point that enables autonomous crisis response of the system by converting complex water supply accident situations into logical and concrete execution steps.

[0097] The sequence determination unit (122) can set the execution sequence in response to accident scenarios requiring the combined use of a hydraulic analysis model that simulates the physical flow of a water supply network and an intelligent diagnostic model that analyzes data patterns.

[0098] An accident scenario refers to the entire process from the occurrence of a crisis situation outside the scope of normal operation within a water supply network—such as leakage, pipe breakage, water pollution, or sensor malfunction—to recovery. In this invention, the accident scenario is not limited to a single event but may include complex situations combining abnormal patterns found in time-series data from IoT sensors (data domain) with resulting changes in water pressure and flow rate within the network (physical domain). Therefore, to resolve such scenarios, it is necessary to simultaneously identify the cause using an intelligent diagnostic model and predict the impact using a hydraulic analysis model.

[0099] The execution order is a sequence of tasks determined by considering data dependencies and mathematical precedence among multiple unit tasks. This can reflect a logical structure where the output of a specific analysis tool must be used as input for another tool. For example, since hydraulic analysis cannot be performed before the location and scale of an accident are determined, the task of identifying the accident site through intelligent diagnosis is configured to have an execution order that precedes the hydraulic analysis simulation task.

[0100] FIG. 9 illustrates an example of an orchestration and scenario optimization method, and the sequence determination unit (122) can operate as follows based on the concept of such accident scenarios and execution sequence.

[0101] First, the necessity of utilizing a complex model is determined. The plan derived from the work planning unit (121) is analyzed to determine whether it is a simple monitoring plan or a complex accident requiring both precise diagnosis and damage spread prediction. In the case of a scenario where a physical contradiction is found within the range of mathematical influence, the sequence determination unit (122) can determine a complex execution plan that calls both a mathematical analysis model and an intelligent diagnosis model.

[0102] Second, the order can be optimized based on guard sentences. By analyzing the constraints and execution conditions included in the descriptions of each analysis tool using a large-scale language model, logical guards such as "repair analysis is possible only after the accident location information has been confirmed" are identified. Through this, the execution order can be arranged so that an intelligent diagnostic model is deployed first to acquire accident information, and the repair analysis model is executed by receiving that information as a parameter.

[0103] Third, it performs a feedback-based order resetting function. If a contradiction in the execution result is found by the validity verification unit (124), the order determination unit (122) can provide flexible orchestration, such as changing the execution order or inserting a new unit task in the middle by reflecting the backtracking result of the inference analysis unit (126).

[0104] This system can serve as a key control mechanism that prevents artificial intelligence lacking specialized domain knowledge in the water supply sector from executing actions contrary to the laws of physics, and helps find the fastest and most accurate analysis path even in complex accident situations.

[0105] Accordingly, the operation method of the sequence determination unit may include: a step of identifying the call priority for each model, the exclusivity between tools, and interdependencies by searching tool descriptions containing natural language guard sentences assigned to each of the hydraulic analysis model and the intelligent diagnostic model; a tool matching step of selecting the hydraulic analysis model when the technical objective of the unit task is to predict physical changes in water pressure or flow rate within the pipeline, and selecting the intelligent diagnostic model when the technical objective is to identify anomalies based on data patterns; and a step of generating an execution sequence by placing the hydraulic analysis model as a preceding process when physical validity verification must precede according to the accident scenario, and placing the intelligent diagnostic model as a preceding process when hydraulic impact analysis based on the diagnostic results is required. The sequence determination unit can dynamically determine the combination of tools and the execution order that align with the purpose of the unit task based on declarative rules specified in the tool descriptions of each model.

[0106] The tool interface (130) can identify accident information by running an intelligent diagnostic model consisting of supervised learning that learns past leakage accident patterns to classify accident types and unsupervised learning that learns normal vibration frequencies to detect abnormal signs based on the reconstruction error of input data, and can predict water cutoff zones and low-pressure zones by linking the accident information to input conditions of a hydraulic analysis model to simulate changes in water pressure within the pipe network due to virtual valve operation or pump operation.

[0107] Supervised learning is a machine learning method that predicts or classifies results for new data by training pairs of input data and their corresponding labels. In this invention, waveforms or water pressure data patterns of various past leakage accidents are learned along with a label called an 'accident type.' Through this, it is possible to clearly classify which previously defined accident type the currently incoming data resembles.

[0108] Unsupervised learning is a machine learning method that identifies similarities between data or detects outliers by learning the structure or patterns of the input data itself without label information. In this invention, a reference model can be generated by learning only the vibration frequencies of a normal operating state. Subsequently, a reconstruction error is calculated, which indicates how much the input data deviates from the normal model; if this error is large, it is determined to be an unknown anomaly sign for which there is no correct answer data, thereby enabling the detection of an accident.

[0109] The tool interface (130) organically links data-based intelligent diagnosis and physical law-based mathematical analysis, and can perform detailed operations in the following steps.

[0110] First, there is a step of identifying accidents through a complex intelligent diagnostic model. The tool interface (130) classifies known leakage patterns through a supervised learning model and, at the same time, detects exceptional vibration patterns outside the normal range through an unsupervised learning model. By combining these two models, specific accident information such as the location, type, and severity of the accident is derived, which can be processed into quantitative data necessary for repair analysis, going beyond a mere signal that 'there is an abnormality'.

[0111] The second step is the information linkage and simulation condition setting stage. Identified accident information is immediately converted into input conditions for the hydraulic analysis model. For example, if it is confirmed that an accident has occurred at a specific location, the pipeline rupture rate or outflow volume at that point can be set as input parameters for the hydraulic analysis model. This is a process of information fusion where diagnostic results in the data domain naturally lead to analysis conditions in the physical domain.

[0112] Third is the virtual scenario execution and damage prediction stage. Based on the established input conditions, virtual valve operations or pump drive commands are applied to the simulation. The hydraulic analysis model calculates changes in water pressure within the pipe network through fluid dynamic calculations and can compare this with hydraulic thresholds defined in the knowledge graph. Finally, it can provide visualized results by precisely predicting low-pressure zones where water pressure falls below the standard and water cutoff zones where supply is interrupted.

[0113] Consequently, the tool interface (130) can implement an integrated diagnostic environment that encompasses everything from identifying the cause of an accident to analyzing the effectiveness of response action scenarios by integrating the pattern recognition ability of artificial intelligence and the physical prediction ability of repair analysis. Through this, the manager can secure a technical means to review the appropriateness of recovery scenarios in a virtual space in advance before taking actual on-site action.

[0114] The sequence determination unit (122) can compare the single damage ranges based on multiple simulation results returned from the tool interface (130) to select an optimal recovery scenario in which single damage is minimized, and can finally determine the control target and operation sequence based on the selected scenario as an execution plan.

[0115] Mutual Comparison is a process of comparing multiple simulation results (Scenario A, B, C...) executed in parallel on the tool interface (130) by placing them on the same evaluation indicators. It refers to an analysis process that determines superiority by converting the predicted number of households without water supply, the area of ​​low water pressure, the time required for restoration, and the impact on major facilities such as hospitals or factories into quantitative figures, rather than simply checking whether water supply is cut off.

[0116] Optimal Recovery Scenario Selection is a decision-making process that determines the single best course of action—based on quantitative indicators obtained through mutual comparison—that can restore the health of the entire water supply network system most rapidly while minimizing social damage (water outages and public complaints). This refers to the selection of the best path, which artificial intelligence deems logically the most valid, to achieve the objective function of 'damage minimization.'

[0117] The sequence determination unit (122) can finalize a complex accident response plan through the following three-step process.

[0118] ① Multidimensional Intercomparison of Simulation Results

[0119] The tool interface (130) performs simulations for each of the various response measures (e.g., upstream valve shut-off vs. downstream valve control) proposed by the work planning unit (121). The ordering unit (122) can extract and compare the range of single-water damage from the returned result data. At this time, by utilizing the spatial properties of the knowledge graph to perform a combined analysis of population density, consumer characteristics, etc., within the range, it is possible to compare which scenario causes the least physical and economic loss.

[0120] ② Selection of optimal scenarios based on weights

[0121] It is not simply selected as the optimal plan just because the area of ​​the water cutoff region is small. The order determination unit (122) can assign weights to each scenario according to a predefined priority (e.g., maintaining supply of core infrastructure > maintaining water pressure in general residential areas). For example, if Plan A has a small water cutoff area but includes a large hospital, and Plan B has a somewhat larger water cutoff area but is mainly general green space, the system can select Plan B as the optimal recovery scenario through weight calculation.

[0122] ③ Final determination of control targets and operation sequence

[0123] Once the optimal scenario is determined, a specific execution plan is drawn up to realize that scenario. This specifies control targets such as "how percent to close which valve" or "how much to adjust the rotational speed of which pump." In particular, to prevent hydraulic shock (such as water hammer), the final control sequence can be completed by arranging the operation sequence along the time axis, such as closing valve 1 first and adjusting valve 2 later.

[0124] Through this mechanism, the present invention can provide object-oriented decision-making based on data and physical laws, moving away from the manager's subjective judgment or simple rules of thumb. In particular, by simulating multiple alternatives in an instant and comparing them, it is possible to autonomously derive the most sophisticated recovery solution capable of minimizing water supply damage even within complex observation structures that are difficult for humans to perceive. Consequently, the sequence determination unit (122) performs the role of a control tower that ultimately guarantees the 'realism' and 'efficiency' of the judgment made by the agentic AI.

[0125] The parameter conversion unit (123) can convert unit tasks into control signals containing API call parameters corresponding to each input protocol and output them for technical compatibility with different input specifications for each hydraulic analysis model and intelligent diagnosis model.

[0126] An input protocol is a set of agreed-upon data formats and transmission protocols to facilitate the smooth exchange of data between different systems or models. Specific file formats required by hydraulic analysis models in water supply network management systems, or data array structures required by intelligent diagnostic models, are examples of respective input protocols.

[0127] API call parameters are a set of detailed variables passed as input values ​​to an interface (API) to execute specific analysis functions. Parameters may include pipe roughness coefficients for hydraulic analysis, demand per node, or sensor identification numbers and time ranges for operating a diagnostic model.

[0128] A control signal is a final execution command packet that instructs an analysis tool to start an actual operation or perform a specific action. This is in the form of a combination of API call parameters processed into an appropriate protocol and the driving command of the tool, and can be transmitted to an external analysis module through the tool interface (130).

[0129] The parameter conversion unit (123) plays the role of translating the agent's 'inference result' into the analysis model's 'input data' and can perform detailed operations through the following process.

[0130] First, there is the stage of ensuring technical compatibility. Since the mathematical analysis model and the intelligent diagnostic model have different production bases or purposes, they require different data specifications. The parameter conversion unit (123) can prevent technical conflicts between models by identifying the characteristics of the analysis tool to be called and reconstructing the data to fit the input protocol that the tool can accept.

[0131] Second is the data mapping and parameter generation step. Real-time measurement data and physical attributes extracted from the knowledge graph are accurately mapped to each item of the API call parameters in accordance with the purpose of the unit task derived from the work planning unit (121). Through this process, the qualitative execution plan is converted into a quantitative numerical dataset.

[0132] Third is the final control signal output stage. A control signal in a completed form is generated, including configured API call parameters, so that the analysis tool can immediately perform calculations. The generated control signal is transmitted to each model via the tool interface (130), and through this, the agent controller (120) can organically control complex and heterogeneous external tools as if they were a single system.

[0133] Consequently, the parameter conversion unit (123) functions as a data conversion mechanism that enables the intelligent judgment of the agentic AI to lead to the operation of the actual water supply network analysis tool. Through this, the system can secure scalability that allows it to freely integrate and use analysis software of various manufacturers or standards.

[0134] The validity verification unit (124) verifies the validity of the results returned from the tool interface (130) by comparing them with the numerical thresholds for each node defined in the knowledge graph, and if a contradiction is found, feeds back the cause of the deficiency to the work planning unit (121) to correct the API call parameters.

[0135] A hydraulic threshold is a physical limit or operational standard value that individual facilities (nodes) must adhere to for the safe operation of a water supply network. This may include the maximum internal pressure (maximum water pressure) that a pipeline can withstand, the minimum water pressure that does not disrupt water supply, and the flow rate range that does not exceed the pipeline's capacity. It is stored as an attribute for each node within the knowledge graph and can be utilized as an absolute reference point for verification.

[0136] Validity / Feasibility refers to a state in which the results of a repair analysis or intelligent diagnosis performed through the tool interface (130) do not contradict the physical laws and the operating rules of the actual system defined in the knowledge graph. In other words, it is an indicator that the accident response plan proposed by the artificial intelligence is not only theoretically possible but is also safe and effective when applied to the actual field.

[0137] The Cause of Deficiency is the specific reason for physical or logical inconsistencies that occur when analysis results fail to pass validity verification. For example, causes of deficiency may include phenomena where the predicted water pressure at a downstream node drops below a hydraulic threshold (minimum pressure) when a specific valve is shut off, or computational errors caused by input parameters being set longer than the physical length of the pipeline.

[0138] The validity verification unit (124) ensures that the plan generated by artificial intelligence can operate in the ‘field’ beyond the ‘virtual’ space and can perform detailed operations in the following steps.

[0139] First, there is a data comparison and result verification step. The hydraulic analysis simulation results (predicted water pressure, flow rate, etc.) returned from the tool interface (130) are received. Then, the hydraulic thresholds of the nodes within the corresponding range in the knowledge graph are called to precisely compare the two data. If the predicted result value falls outside the threshold range, the execution plan can be determined to lack physical feasibility.

[0140] The second step is the analysis and identification of the causes of the deficiency. If validity is confirmed to be lacking, the cause is not simply dismissed but traced to determine which parameters or logic caused the contradiction. For example, specific violations are identified, such as "the currently set valve opening rate parameters are insufficient to prevent water cutoff in the downstream area."

[0141] Third is the feedback and parameter correction step. The identified cause of deficiency is immediately transmitted to the task planning unit (121). Based on this feedback, the task planning unit (121) can restart the inference process of the large-scale language model to generate new API call parameters that can resolve contradictions or adjust the order of the execution plan.

[0142] Through this feedback process, the present invention corrects logical errors that agentic AI may commit using physical formulas, and the finally derived incident response plan can support the complete assurance of the physical integrity of the water supply network.

[0143] The agent controller can integrate and control mutually heterogeneous mathematical analysis models and intelligent diagnostic models based on API call parameters, and perform a self-correction loop that ensures the physical reliability of the execution plan through threshold-based feedback of the knowledge graph.

[0144] Figure 10 illustrates an example of making a final decision by combining the judgment results of supervised and unsupervised learning models. If the supervised learning model determines it to be a leak and the unsupervised learning model determines it to be an anomaly, it can be confirmed as a high-reliability leak. This implies that since the analysis results of the two models match, the probability of an accident occurring is very high, and therefore the site manager must take immediate action.

[0145] If a supervised learning model classifies a situation as leakage but an unsupervised learning model determines it to be normal, it can be defined as low-confidence leakage. Since there is a possibility of misjudgment in the AI's decision at this stage, it is recommended to collect additional evidence to reinforce the basis for the judgment before taking definitive action.

[0146] If an unsupervised learning model detects an anomaly even though a supervised learning model judged it to be normal, it may be classified as an unknown anomaly. Since this is highly likely to be a novel anomaly pattern that did not exist in past accident data, it is managed by designating it as an active learning candidate and a labeling target to enhance the performance of artificial intelligence in the future.

[0147] If both models determine that the condition is normal, it is ultimately recognized as a normal state, and no separate additional measures are taken. This fused decision-making method serves as the technical basis for reducing false alarms in leak diagnosis and ensuring the overall reliability of the system by complementarily combining the pattern classification capabilities of supervised learning with the anomaly detection capabilities of unsupervised learning. Agentic AI can update the state of nodes within the knowledge graph and autonomously determine the optimal timing for response based on these predefined fusion rules.

[0148] FIG. 11 is an example illustrating the underlying mechanism of a state-aware self-correcting loop and domain-specific readjustment rules adopted by the agent of the present invention to flexibly respond to crisis situations.

[0149] The agent of the present invention utilizes a react loop that repeats observation and re-evaluation operations as its underlying mechanism. Immediately after a tool call, the agent adds the result to a message history and observes it during the next inference step. If the result of the tool execution does not align with user intent or physical validity, the agent can perform a process of recalculating the argument values ​​of the same tool and re-calling it. A feature of the present invention is the combination of a unique domain-specific self-correcting trigger derived from a water supply network operation environment on top of this underlying mechanism.

[0150] The domain-specific rebalancing rule of Fig. 11 is designed to respond to five major failure patterns. First, if hydrostatic pressure data is not available within a specified period, the period parameter can be extended from a short-term standard to a long-term standard. This is based on hydraulic grounds to increase the reliability of the analysis by securing a statistically valid minimum sample size.

[0151] Secondly, if no suspected leak candidates are found and the result is an empty set, a re-search can be performed by relaxing the water pressure threshold from a strict standard to a sensitive standard. This measure reflects the characteristic that indicators of suspected leaks and indicators of water pressure anomalies are in a continuous spectral relationship of the same physical indicators.

[0152] Thirdly, in situations where a vector search for a facility name fails, the top key indicating the search range can be expanded to extend the list of candidates to nearby points. This is a logical mechanism designed to overcome the impossibility of accurate name matching that occurs when operators refer to facilities using idioms or abbreviations.

[0153] Fourth, if the analysis results of a single point are insufficient to determine the range of influence, the flow tracking step can be extended from one hop to two hops. This takes into account the hydraulic characteristics that the range of influence can be chained downstream as hydraulic damping propagates with the accumulation of distance.

[0154] Finally, if no results are obtained within the reference block, the scope of analysis is extended to the higher-level operating entity or adjacent blocks. This reflects the incorporation of the specific field characteristics into the self-correcting loop, where there may be sections where sub-block boundaries do not fully guarantee mathematical isolation.

[0155] All of these rules enable the hydraulic characteristics of the water supply network to function as prior knowledge for artificial intelligence. As a result, the present invention can dramatically improve the success rate of responding to water supply accidents through domain-specific reorganization decision-making that cannot be derived by general artificial intelligence models.

[0156] Figure 12 illustrates an example of the detailed configuration of additional self-correcting triggers implemented in the form of multilayer middleware to address five failure modes that are not resolved by the underlying mechanism, the readjustment loop alone.

[0157] First, the tool execution exception trigger targets errors that occur when calling the analysis tool. When an error occurs, it undergoes a process of converting it into a user-friendly message, which can induce the agent to modify the input value according to predefined rules and re-call the tool.

[0158] Secondly, the context state change trigger can detect when a user has changed the reference point for an analysis. In a water supply network operation environment, time point transitions occur frequently, such as changing the analysis reference date from yesterday to a week ago; however, existing large-scale language models have limitations in that they retain previous analysis results in the context, leading to errors caused by time point confusion. The present invention can support the agent in establishing a completely new plan under the new context by forcibly removing the message history used in the previous analysis when a new reference point is given.

[0159] Thirdly, the call count accumulation trigger can operate to prevent the AI ​​from falling into an infinite loop. If the number of calls exceeds a preset maximum during the execution of a unit task, the system can be forcibly terminated to prevent resource waste and abnormal behavior.

[0160] Fourth, context burst triggers can address situations where past tool execution results excessively occupy tokens. By replacing outdated tool execution data with compressed markers, the valid context space that large-scale language models can process can be managed efficiently.

[0161] Finally, semantic inconsistency triggers in tool results can detect cases where the tool's execution outcome does not match the user's intent. A large-scale language model observes the results to apply domain-specific rules or readjust argument values, and can establish a new execution plan using alternative tools if necessary.

[0162] This multi-layered middleware-based trigger configuration can explicitly resolve domain-specific failure patterns that are difficult to address with general agent operations alone. Through this, the water supply network crisis management device can secure the technical completeness to derive accurate and stable response results without timing confusion or logical errors, even in complex operational scenarios.

[0163] Figure 13 is an example illustrating five types of predefined guard rules explicitly included within natural language descriptions possessed by each expert tool, and the corresponding technical features.

[0164] The large-scale language model of the present invention reads the entire description of each candidate tool during the tool selection phase and retains only the tools that meet these guard rules in the final candidate set. Referring to FIG. 13, the guard rules consist of entry keywords, sibling tool avoidance, call order guards, exclusivity guards, and recall guards.

[0165] The entry keyword guard can perform the function of controlling the tool so that it emerges as a candidate only when operator vocabulary, such as suspected leaks or leak sections, is matched. This serves to ensure that the tool operates only when the practitioner's intent aligns with its purpose.

[0166] Sibling tool avoidance guards allow tools to declare situations where semantically adjacent tools should be used, such as at anomaly points or water pressure anomalies. This clarifies the boundaries between tools with similar functions, preventing incorrect tool routing.

[0167] Call order guards can specify a precedence order to ensure that a specific ID resolution tool is executed before an analysis tool. A logical execution sequence can be guaranteed by directly embedding dependency rules in the documentation that require block IDs or point IDs to be looked up first.

[0168] Exclusivity guards can declare exclusion rules that prohibit parallel calls between specific tools. They can maintain system consistency by blocking the concurrent execution of pairs of tools that might cause semantic conflicts or mathematical contradictions if called simultaneously.

[0169] Finally, the re-call guard encodes patterns requiring a re-call of the same tool with modified parameters into natural language. In the event of a user's request for additional detail, it prompts the existing tool to be re-run with specific parameters, thereby allowing for control over the depth of analysis.

[0170] In this invention, since a large-scale language model directly interprets these sentences during the tool selection phase to identify candidates, there is no need to modify router code when adding new tools. By simply writing the description of the tool, the decision tree automatically expands, which can dramatically increase the scalability of the system. Consequently, this system is consistently applied during the observation phase of the React loop and can support artificial intelligence in acquiring sophisticated tool utilization capabilities based on knowledge of water supply network operations.

[0171] The state management unit (125) can record and manage operational data including the thought process of a large-scale language model, input / output data of each unit task, and the execution status of the tool interface (130) as time-series snapshots.

[0172] The inference analysis unit (126) can generate an inference path in which the logical flow until the final parameter is calculated is structured by calling operational data and mapping the logical steps of the thought process and the execution results of the unit tasks.

[0173] Operational data is integrated log information that includes the chain of thought generated by the large-scale language model for problem solving, the input and output data of each unit task, and the execution status of the tool interface (130). This serves as source data proving what the system determined and what tools were used at a specific point in time.

[0174] A snapshot is a record that saves an exact copy of all operational data at a specific point in time. Stored in a time-series database, this record preserves the system's past state and serves as a reference point for rollback, which restores the system to a normal state if an error is detected during validation.

[0175] The inference path is a structured lineage of logical steps traversed from the moment the initial signs of an accident are recognized until the final control parameters are calculated. It is configured to match individual thought stages with the execution results of actual tools one-to-one, allowing for the visualization or tracking of the flow and the rationale on which the agent made its final decision.

[0176] The state management unit (125) acts as a black box that records all activities of the agent and operates through the following process. First, it can collect the results of all inference steps and unit tasks performed by the agent controller (120) in real time. At this time, it can manage not only text results but also the detailed status of data packets exchanged with an external model through the tool interface (130) by including them in the category of operational data.

[0177] Collected operational data is recorded in a time-series repository in the form of snapshots over time. This snapshot management system can provide a physical foundation for the immediate restoration of the system to its safest past state in situations involving infinite loops or logical contradictions that may arise due to the non-deterministic nature of agentic AI.

[0178] The inference analysis unit (126) performs the function of post-analyzing or verifying the agent's thought process based on fragmented data stored in the state management unit (125). This component calls the stored operational data to organically connect the logical steps passed by the large-scale language model and the results of the unit tasks executed at each step. Through this, an inference path can be generated that shows how the final control parameter was derived through logical validity.

[0179] The generated inference path can be used as a map to trace back where a logical error occurred when the validity verification unit (124) discovers a physical contradiction. In particular, it functions as a key technological device that transparently discloses the decision-making process of artificial intelligence and ensures explainability in fields where safety is critical, such as water supply network operation.

[0180] Collaboration between the state management unit (125) and the inference analysis unit (126) can generate the effect of converting a series of processes performed by the agentic AI into data assets, thereby simultaneously securing the stability and reliability of the system. This means completing a structure in which the artificial intelligence not only produces results but also manages and takes responsibility for the process itself.

[0181] When a mathematical contradiction is determined by the validity verification unit, the inference analysis unit (126) can identify a specific unit task or parameter that caused the mathematical contradiction by comparing the error data calculated by the validity verification unit (124) with the intermediate output values ​​at each step of the inference path.

[0182] When a mathematical contradiction is determined by the validity verification unit (124), the inference analysis unit (126) can identify a specific unit task that caused the mathematical contradiction by comparing the error data calculated by the validity verification unit (124) with the intermediate values ​​of each step on the inference path.

[0183] The state management unit (125) can perform a rollback to restore the device state by loading a valid snapshot corresponding to the point immediately before the execution of a specific unit task identified. The inference analysis unit (126) can generate correction input information based on the identified error causes and error data, thereby inducing the task planning unit (121) to reconstruct the execution plan based on the correction input information.

[0184] A mathematical contradiction refers to a state in which response plans or simulation results proposed by artificial intelligence violate the physical laws of the water supply system or predefined operational limits. For example, this includes situations of physical non-conformity, such as when closing a specific valve causes the pressure in adjacent pipelines to exceed the design internal pressure, or conversely, when the water pressure in essential supply areas drops below the standard level.

[0185] Error data is a quantitative indicator calculated by the validity verification unit (124) during the process of determining mathematical contradictions. It represents the numerical difference between the simulation result value and the actual mathematical threshold value, and serves as a key basis for analysis indicating how large a physical deviation occurred at a given point.

[0186] Rollback is an action that restores the logical state of the device to a safe point immediately before the error occurred by calling a past valid snapshot stored in the state management unit (125) when an error or contradiction occurs in the system. This is a safety mechanism that ensures the agent is not bogged down in incorrect reasoning results and can resume the accident from the correct path.

[0187] Correction input information is feedback data generated based on error causes and error data identified by the inference analysis unit (126). It includes specific constraints or modified parameter guidelines that the work planning unit (121) should refer to when re-establishing the execution plan, and serves to guide the large-scale language model to derive an alternative with physical validity.

[0188] When a mathematical contradiction is detected by the validity verification unit (124), the inference analysis unit (126) can intervene immediately. The inference analysis unit (126) can compare the error data received from the validity verification unit with the step-by-step intermediate outputs on the inference path of the current system. Through this process, the source of which specific unit task or parameter in the complex execution plan caused the mathematical contradiction can be accurately identified.

[0189] When the point that caused the error is identified, the state management unit (125) can load a valid snapshot immediately before the unit task is executed. This allows the state of the agent controller (120) to be rolled back to a clean state before the inconsistency occurred. This serves as a physical basis for preventing incorrect data from spreading throughout the system and increasing the accuracy of recalculations.

[0190] The inference analysis unit (126) can process the analysis results regarding why the error occurred into the form of correction input information, rather than simply restoring the state. This information is transmitted to the task planning unit (121), and the task planning unit (121) can restart the inference of the large-scale language model based on this. As a result, the system can autonomously reconstruct a new physically valid execution plan by reflecting the corrected conditions without repeating the same mistake.

[0191] This series of processes is a key mechanism that enables agentic AI to independently learn and adhere to the rules of the precise physical world of water supply. The process of discovering mathematical contradictions and resolving them through comparison with error data and inference paths enhances the explainability of the system, while reconstruction through rollback and correction input can dramatically improve the completeness of the incident response plan. Through this, the present invention secures a technical foundation that enables the completion of an optimal recovery scenario with guaranteed physical validity without manual intervention by an administrator.

[0192] Referring again to FIG. 2, the agent controller (120) may further include a language user interface (140), a communication module (170), a RAG module (180), and a database (190).

[0193] The Language User Interface (140) is an input / output window that helps the user interact with the system in natural language without using specialized commands or complex code. It converts the user's natural language request into a logical command that the agent can understand and transmits it to the task planning unit (121), and conversely, it can perform the function of reconstructing the system's analysis results or control suggestions into a language that is easy for the user to understand, and visualizing or outputting them.

[0194] The communication module (170) is a hardware and software module that safely transmits and receives data with external devices and servers, such as IoT sensors (200), geographic information systems (300), and remote monitoring and control systems (SCADA) installed at the water supply network site. It can serve as a physical channel to reflect real-time measurement data in a knowledge graph or to transmit the final control signal determined by the agent to the pump or valve driving device at the actual site.

[0195] The RAG module (Retrieval-Augmented Generation, 180) is a module that utilizes real-time retrieval from external data sources for inference, such as the latest water supply operation manuals, special accident response guidelines for specific regions, or historical accident response reports, which large-scale language models have not previously learned.

[0196] This prevents the hallucination phenomenon, where artificial intelligence generates information different from the facts, and enables sophisticated inference based on literature-based evidence regarding unstructured crisis situations that are difficult to judge solely from the structured data of the knowledge graph.

[0197] The database (190) is a repository that systematically stores and manages a knowledge graph, which is a digital model of the water supply network, time-series snapshots recorded by the state management unit (125), and the agent's chain of thought logs. It is an infrastructure that immediately provides all the data resources necessary for the agent to review past response cases or to roll back the system state to a safe previous point in time during the validity verification process.

[0198] These components enable the agent controller (120) to function not merely as simple analysis software, but as an intelligent operating system that actively communicates with field managers and autonomously manages the water supply network, which is a physical entity.

[0199] The database (190) may include a graph database (191), a time-series database, and an expert knowledge repository (193). The graph database (191) is a database management system that stores and manages data in a connection structure of nodes and edges. It is a space that effectively stores an ontology knowledge graph built based on the facility coordinates and pipeline information of the water supply network system. Unlike a simple table-type database, it can intuitively express the hydraulic connectivity or inclusion relationships between facilities and can provide overwhelming computational speed and efficiency when artificial intelligence searches the range of influence along the direction of water flow from a specific accident point.

[0200] The time series database (192) is a storage specialized for sequentially recording measurement data and system state information that occur over time. It stores dynamic state values ​​such as water pressure, flow rate, and vibration flowing in from IoT sensors at the site, as well as snapshot data of the accident process generated by the state management unit (125), in a time series. It can supply temporal basis data that can learn abnormal patterns based on past data during the inference process of a large-scale language model, or perform a rollback function to restore the system state to a specific past point in time in the event of a failure in validity verification.

[0201] The expert knowledge repository (193) is a repository that manages unstructured expert knowledge, such as water supply operation manuals, technical guidelines, accident response guidelines for specific regions, and past accident response reports. It includes complex operational scenarios or administrative measures that are difficult to judge based solely on quantified data. It is linked with the RAG module (180) to support real-time searching of legal and technical grounds required for inference by a large-scale language model, thereby preventing errors in the artificial intelligence generating information different from the facts and helping to derive a highly reliable response plan that can be immediately applied to practical work.

[0202] The structure in which these three types of databases are organically combined forms the foundation for agentic AI to perfectly understand and manage the complex physical system of the water supply. Graphs containing topological information, time-series data containing operational records, and a knowledge repository containing expert wisdom create synergy to further refine the digital twin of the water supply network.

[0203] The graph engine (110) manages the storage system through a graph database (191) that manages physical connection information of facilities and a time-series database (192) that manages measurement data of IoT sensors (200) in a time-series manner, and can generate a knowledge graph by dynamically mapping real-time measurement values ​​to nodes of facilities at the time of a request by the agent controller (120).

[0204] The graph engine (110) can perform the role of completing an intelligent map that can be understood by artificial intelligence by combining the physical framework of the water supply network and the real-time operational status. First, it manages physical connection information and phase relationships between facilities through the graph database (191), and records water pressure and flow rate measurement data flowing in from the IoT sensors (200) at the site in chronological order through the time series database (192).

[0205] The graph engine (110) merges the two databases in real time at a specific point in time when the agent controller (120) requests analysis. By dynamically mapping live real-time measurements to static facility nodes, it generates a knowledge graph that reflects the latest status of where and how much water is flowing in the current network. This serves as a physical foundation that enables the artificial intelligence to make decisions based on the current, live field conditions rather than past data.

[0206] The work planning department (121) can refer to an expert knowledge repository (193) in which unstructured repair manuals and past accident reports are stored in the form of vector data, and can convert the real-time repair status analyzed in the knowledge graph into a search query of the RAG module (180) to extract response guidelines that correspond to the current situation from the expert knowledge repository (193).

[0207] The work planning unit (121) can be responsible for establishing response strategies by adding the refined knowledge of experts to the reasoning ability of artificial intelligence. In this process, the expert knowledge repository (193) is referenced, in which unstructured repair manuals in text form and past accident reports are stored in the form of vector data with high search efficiency.

[0208] The advanced operation of the work planning unit (121) occurs in the stage where the current abnormal hydraulic state analyzed in the knowledge graph is automatically converted into a search query that the RAG module (180) can understand. For example, if a graph analysis result indicates that the water pressure in a specific area has dropped sharply, this is converted into a specific search term such as "optimal valve shut-off sequence in case of pipe damage in the area" to extract response guidelines that best fit the current situation from the expert knowledge repository (193). This prevents logical errors that artificial intelligence might commit and allows the response method of actual experts to be reproduced exactly.

[0209] The sequence determination unit (122) can generate an execution plan that combines the physical phenomena of observation and the response logic of experts by specifying the execution conditions of the unit tasks based on the extracted response guidelines.

[0210] The sequence determination unit (122) can perform the role of concretizing theoretical guidelines into detailed processes that are feasible in the actual field based on the extracted response guidelines. The extracted guidelines include logical instructions such as "gradually closing the upstream valve and performing drainage work." Based on these guidelines, the sequence determination unit (122) defines physical constraints under which individual unit tasks must be performed. Finally, an accident response execution plan can be generated by combining hydraulic physical phenomena, such as changes in water pressure in the pipe network, and operational logic accumulated by experts over decades into a single sequence. As a result, the system autonomously derives an optimal recovery scenario in which physical feasibility and operational efficiency are simultaneously secured.

[0211] The language user interface (140) can receive the execution plan of the agent controller and the execution results of the tool interface and provide them to the user in a visualized manner. The language user interface (140) can display map-based results that visualize the accident location and the range of mathematical impact based on the spatial coordinates of the knowledge graph, generate a multi-chart that analyzes the fluctuation trend of measurement data extracted from a time-series database, and generate and provide a detailed report in natural language form by analyzing the inference path and final execution results through a large-scale language model.

[0212] The language user interface (140) can extract topological information and spatial coordinates of facilities stored in a knowledge graph and implement them on a map-based dashboard. Beyond simply marking the location where an accident occurred as a point, the range of mathematical impact inferred by the agent can be visualized as a spatial area.

[0213] This allows operators to immediately identify the topographical potential for the spread of current water pressure drops or pipeline ruptures. By displaying the locations of key valves near the accident site and the direction of water flow on a map, this serves as foundational data that helps managers gain a three-dimensional understanding of the physical conditions at the site.

[0214] A language user interface (140) linked to a time-series database (192) can precisely visualize the variability of measurement data collected from IoT sensors (200) at the site. It can automatically generate and provide a multi-chart that allows for the comparison of water pressure and flow rate trends before and after an accident from various angles.

[0215] Multi-charts refer to a visualization technique that arranges various types of data series or different indicators side-by-side on a single screen or based on a shared axis. Going beyond simply listing multiple graphs, this can be utilized as an analytical tool to gain intuitive insight into correlations between data and changes in patterns.

[0216] Specific implementation methods include the overlay type, which displays different indicators such as water pressure and flow rate within a single chart area, and the partition type, which identifies minute differences between multiple points by arranging charts of the same scale in a grid pattern. Additionally, the multi-axis chart method, which places different axes on the left and right sides to plot data with completely different units on a single chart, is also a major form of multi-chart.

[0217] In the water supply network management system (10), multi-charts play an important role in helping to recognize the context of a situation that is difficult to grasp with only a single indicator. For example, by comparing a phenomenon of falling water pressure with a phenomenon of a sudden surge in flow rate on the same time axis, it is possible to confirm that the accident is caused by actual pipe damage rather than a simple measurement error. In addition, by comparing real-time data with past normal operation patterns on a single chart, it is possible to immediately determine whether the current value falls within the risk category.

[0218] Consequently, the multi-chart serves as visual evidence proving the validity of the agent's reasoning results to the operator. Through this, the operator can confirm that various sensor data, such as water pressure drop or vibration levels, uniformly point to the same incident pattern, thereby securing grounds to trust and approve the system's control commands.

[0219] Rather than simply displaying current figures, it overlays past normal driving patterns with current abnormal patterns, thereby providing operators with a basis to objectively assess the severity of accidents. This multi-chart analysis can serve as a decisive indicator during the operator's final review of the technical validity of abnormal signs detected by the intelligent diagnostic model.

[0220] The present invention has a report generation function that translates a complex data analysis process into human language by utilizing a large-scale language model. The language user interface (140) can analyze the inference path generated by the inference analysis unit (126) and explain the logical basis for why the artificial intelligence established such a response plan.

[0221] In addition to analyzing the final execution results, it provides detailed reports in natural language containing specific action suggestions, such as "It is predicted that shutting off the valve in Zone A will stabilize the water pressure in Zone B." This enables even operators lacking professional mathematical analysis expertise to clearly understand the AI's expert judgment and promptly implement actual measures.

[0222] This multifaceted visualization and reporting system enhances the transparency and reliability of the water supply management system. Operators can perform optimal crisis response tasks by organically collaborating with Agentic AI through a process of spatial awareness via maps, data verification via charts, and logical understanding via reports.

[0223] FIG. 14 illustrates an example of an AI architecture for monitoring crisis management according to the embodiment of the present invention. The system (10) is based on large-scale language models (LLM) and LangGraph technology and has a structure that organically combines a graph database (192), a relational database (time-series database), a vector database (expert instruction repository), and various expert tools. The AI ​​agent can perform the role of monitoring the monitoring status in real time, and when an abnormal sign occurs, immediately recognize the problem and initiate inference and analysis to identify the cause and establish a response plan.

[0224] Numerous IoT sensors installed at the water supply network site are positioned on the right side of the system to collect water pressure and leakage sound data in real time. At the same time, flow rate, water pressure, and geospatial data from the Supervisory Control and Data Acquisition (SCADA) system and the Geographic Information System (GIS) are fed into the server and stored in the database (190).

[0225] The collected data is combined with an ontology-based knowledge graph shown at the bottom to construct a knowledge base that reflects the physical topography of the observation network. In addition, unstructured document data such as manuals, facility documents, GIS guidelines, and maintenance history are stored in the form of vector data in a specialized knowledge repository (193). When a complex crisis situation occurs, the agent can refer to this knowledge base to obtain the logical basis that best fits the current situation.

[0226] At the heart of the architecture lie AI agents and large-scale language models that interact through prompt control. The agents monitor incoming instrumentation data, and upon detecting anomalies, they can leverage the reasoning capabilities of the large-scale language models to define the current situation and formulate a work plan for resolution.

[0227] During the problem-solving process, the agent calls upon and utilizes the expert tools depicted in the upper left corner for each unit task. These expert tools can provide detailed functions such as discrimination models, prediction models, anomaly detection, data retrieval, problem recognition, hydraulic analysis, planning, and screen display. The agent can organically link and operate these tools to perform hydraulic analysis simulations or derive optimal recovery scenarios.

[0228] The final analysis results and response plans are visualized and provided to operators through a central map-based monitoring dashboard. This integrated architecture enables rapid and accurate responses in water supply network crisis situations, as artificial intelligence autonomously leads the entire process from data collection to cause analysis and final decision-making.

[0229] Figure 15 illustrates a specific workflow of agentic AI for water supply network crisis management designed based on Langraff, allowing one to grasp at a glance how the entire workflow architecture and expert tools are organically connected for the successful implementation of agentic AI.

[0230] When a user submits a query in natural language through the system, the system first clearly identifies the type of question. The types of questions are subdivided and classified into categories such as water pressure inquiry, anomaly detection, microblock (DMA) analysis, chart viewing, map display, identification of facility anomalies, and other utilities. This process serves as the first gateway for a Large Language Model (LLM) to interpret the user's intent in accordance with the water supply operation domain.

[0231] The Large Language Model (LLM) at the top center manages the entire workflow and makes real-time decisions. Based on the type of question identified, the most suitable FastAPI-based expert tools can be selected and executed. The tools executed may include sub-block and branch search, water pressure and leak analysis, chart and map viewing, and dedicated AI models for anomaly detection.

[0232] During the operation of the selected tool, queries are performed simultaneously on relational databases (MariaDB), graph databases, and vector databases (Qdrant). Ontology-based knowledge maps and multimodal search augmented generative (RAG) technologies are utilized to maximize analysis accuracy. The collected data is then passed to a large-scale language model to generate final responses, which are visualized and displayed on the integrated monitoring screen.

[0233] All generated responses and the agent's decision-making processes are systematically recorded through Langfuse. This history management goes beyond simple log storage and can be utilized as key feedback data to analyze the system's inference paths and continuously train and advance artificial intelligence models in the future.

[0234] This Langgraph-based intelligent workflow becomes a powerful tool that improves the accuracy of incident response and supports efficient decision-making by managers by logically connecting vast amounts of data in water supply networks.

[0235] Figure 16 illustrates an example of a data integration management system in which a graph engine and an agent controller are organically combined through search, augmentation, and generation technologies to derive a final execution plan.

[0236] The present invention can maximize the crisis response capability of a water supply network by fusing static facility information, dynamic measurement data, and unstructured knowledge of experts into one. First, the graph engine (110) located on the left side of the drawing can operate based on a graph database (191) that manages the static connection structure of facilities and a time series database (192) that manages dynamic measurement data streams.

[0237] When a request is made by the agent controller (120), these two databases can be dynamically mapped to generate a knowledge graph that reflects the current hydraulic status of the pipe network in real time. The knowledge graph generated in this way serves as basic data that provides the agent with key information about what physical abnormalities are occurring in the current pipe network.

[0238] The work planning unit (121) located inside the central agent controller (120) can generate unit tasks by utilizing the search function of a large-scale language model. At this time, based on the real-time repair status provided by the knowledge graph, it can formulate the question most appropriate for the current situation and convert it into a search query of the RAG module (180).

[0239] The RAG module (180) can use the converted query to search the expertise repository (193) on the right. The expertise repository (193) stores unstructured repair manuals or past accident reports in a vector data format that is easy to search, and the RAG module (180) can extract expert response guidelines that best suit the current situation from this and return them to the work planning department.

[0240] Finally, the device (100) can combine the real-time physical phenomena of the water network identified through the knowledge graph and the expert response logic obtained through the RAG module (180) into one. The execution plan generated through this fusion process includes specific unit tasks and precise execution conditions, and has a level of completeness that allows a manager to immediately apply it to the field. Consequently, the system of FIG. 16 can be described as an architecture that supports artificial intelligence in making judgments at the same level as water supply experts through the separate storage of data and intelligent integrated search.

[0241] Figure 17 illustrates an example of an intelligent anomaly detection model based on leak sounds applied to a water supply network agentic artificial intelligence system. This model is implemented as a supervised learning model based on a Convolutional Neural Network (CNN) that learns from data clearly classified by experts as leak sounds and non-leak sounds. The central part of the figure shows the process of converting collected acoustic data into a spectrogram-shaped image combining time and frequency components and using it as an input value for the neural network. The convolutional neural network can perform the role of extracting leak-specific features from these visualized acoustic patterns to ultimately determine the current state as either a leak or no leak.

[0242] At the bottom of the diagram, waveform data, which is a raw acoustic signal in the audible frequency range collected from on-site sensors, and spectrum data analyzed from it in the frequency domain are exemplified. This acoustic analysis data serves as a key physical basis for intelligent diagnostic models to determine whether an accident has occurred and can be utilized by agents to identify accident information as a detailed component of expert tools.

[0243] The model in Fig. 17 can perform the function of precisely detecting signs of micro-leaks or complex accidents that are difficult to detect with simple changes in water pressure alone through acoustic pattern recognition. This serves as a technical foundation that, in conjunction with the tool interface described earlier, provides reliable basic diagnostic information for an agent to confirm abnormal situations at the site and initiate subsequent simulations.

[0244] FIG. 18 illustrates an example of an autoencoder-based unsupervised learning model that creates a reference model based on data from a normal operating state and then detects abnormal phenomena that deviate from it.

[0245] This model learns from a large amount of normal data without ground truth labels to efficiently extract features and restore them to their original form. Examining the architecture in the center of the diagram reveals a structure where input data is converted into compressed features through an encoder and then reconstructed into output data through a decoder.

[0246] The core operating principle of unsupervised learning models is the utilization of reconstruction error. Since the model learns only normal leakage sound patterns during the training process, it produces output values ​​nearly identical to the input when normal data is input. However, if abnormal signs or new types of leakage patterns that have never been learned are input, the model fails to accurately reconstruct them and can generate a reconstruction error resulting in a significant difference between the input and output values.

[0247] The contrast of spectrogram images on the left and right sides of the diagram visually demonstrates this process. By comparing the magnitude of the error occurring as input acoustic data passes through the model and is output with a predefined threshold, the system precisely detects even novel, unlabeled anomaly patterns.

[0248] Consequently, the unsupervised learning model of Fig. 18 performs a complementary function of detecting unknown accidents that the supervised learning model might miss. This serves as the technical basis for agentic artificial intelligence to maximize the ability to identify accident information by statistically monitoring changes in the field at all times, rather than relying solely on past accident cases.

[0249] Figure 19 illustrates an example of the analysis process of agentic artificial intelligence that identifies abnormal signs based on water pressure data and estimates suspected leakage sections. This figure shows how time-series diagnosis, which analyzes the temporal change pattern of water pressure, and intelligent knowledge graph technology, which analyzes the connection relationships between facilities, are combined.

[0250] On the left side of the diagram, a multi-chart is presented showing the trends in water pressure fluctuations collected from various points. The system utilizes Isolation Forest technology to analyze abnormal change patterns hidden within this water pressure data in real time. Isolation Forest is an artificial intelligence algorithm that identifies isolated data points significantly deviating from normal ranges as outliers based on data density or distance; through this, it immediately triggers an alarm if the water pressure at a specific point exceeds the normal operating range.

[0251] The right side of the diagram shows the results of analyzing hydraulic relationships between pressure points using a knowledge graph and visualizing them on a map. It goes beyond simply tracking information that water pressure has dropped at a specific point; it tracks the connection status of pipelines and the direction of water flow defined in the knowledge graph. By analyzing the correlations between adjacent pressure sensors, it identifies the path through which the impact of an accident propagates and ultimately precisely estimates the suspected section with the highest probability of leakage.

[0252] The system in Fig. 19 integrates statistical analysis, which detects anomalies in individual data, with physical analysis, which understands the overall structure of the network. Through this three-dimensional analysis, agentic AI clearly distinguishes between simple water pressure drop phenomena and actual accidents. Additionally, suspected leak areas are highlighted with red markers on a map-based interface, providing an intuitive basis for decision-making that enables managers to establish plans for immediate on-site inspections or valve operation.

[0253] FIG. 20 illustrates a case of pilot application of the agentic artificial intelligence system for water supply network crisis management of the present invention in an actual field. This demonstration case covers four sub-block areas in Cheongyang-gun, Chungcheongnam-do, and an environment was established to collect real-time measurement data by installing a total of 50 IoT water pressure sensors in the area. It is characterized by implementing a field-oriented agentic artificial intelligence service by combining the collected sensor data with knowledge graph data containing physical information of water supply network facilities.

[0254] Looking at the specific facility status of the pilot application area, three reservoirs, including Jeongsan, Cheongyang 1, and Cheongyang 2, are operating as core water supply sources. The sub-blocks subject to analysis consist of a total of four blocks: Hwaseong, CY 1, Jeongsan 1, and Jeongsan 2. Additionally, six booster stations and two pressure reducing valves are installed for water pressure control, and all of these facilities are defined as nodes within the knowledge graph and become subjects of analysis for the agent.

[0255] The map visualization results show 50 water pressure sensors evenly distributed within the sub-blocks to monitor the status of the water network in real time. Water pressure data from each point is recorded in a time-series database, and agentic artificial intelligence maps this to facility information in a knowledge graph to evaluate the hydraulic health of each block in real time.

[0256] Consequently, FIG. 20 provides important technical grounds demonstrating that the present invention can successfully perform AI-based intelligent crisis management functions by organically fusing vast amounts of sensor data and facility information even in a complex actual water supply operation environment. Through this, the operator can three-dimensionally identify the operating status of major facilities within the pipe and the water pressure distribution by sub-block, and take quick and accurate measures based on the agent's inference path in the event of an accident.

[0257] FIG. 21 illustrates the actual screen configuration of the language user interface of the present invention through a pilot application example. This interface can provide an integrated control environment designed to allow an operator to understand and respond to the status of a water supply network through natural language without having to learn complex system commands.

[0258] The overall screen is broadly divided into a main function button and statistics area, a map-based results display area, and an agent conversation area for Q&A. In the button and statistics area on the left, control buttons for hydraulic analysis or unit conversion are displayed alongside a circular statistical chart that allows for a quick overview of water pressure status within the network. Through this, operators can immediately recognize the ratio of normal to abnormal conditions among all measurement points, using this as a basis for rapid situational assessment.

[0259] The central map-based results display area is linked to an intelligent knowledge graph to show the water pressure status of each node in combination with geographical information. Points where water pressure deviates from the normal range are highlighted with red markers to provide a visual warning, allowing operators to intuitively identify the specific geographical location where an incident occurred.

[0260] The conversation and response area on the right functions as a key communication channel between the AI ​​agent and the operator. In the example screen, when an operator requests the status of minimum nighttime water pressure on a specific date, the agent presents a summary of the analysis results for a total of 87 locations. Rather than simply listing numerical values, the agent enhances the specificity of corrective actions by providing the names and detailed addresses of the locations where minimum and maximum water pressure occurred.

[0261] The bottom section features an FAQ area where users can select frequently asked questions and a query input window for entering new commands, ensuring conversational continuity. Additionally, various detailed charts, data tables, and report templates are displayed in separate pop-up windows, allowing users to selectively view in-depth analysis information as needed while maintaining the visibility of the main screen.

[0262] The language user interface of Fig. 21 is a technical result that reduces the workload of operators and enables data-driven scientific decision-making by integrating and providing vast amounts of water supply data into maps, statistics, and natural language reports.

[0263] Figure 22 illustrates an example of four key analysis screens implemented through the language user interface of an agentic AI for water supply network crisis management, showing in detail the process in which, when a user enters a query in natural language into a question window, the agentic AI identifies the type of question, executes the optimal expert tool, and visualizes the results.

[0264] The first screen in the top left shows the execution of functions for analyzing abnormal water pressure locations and estimating suspected leakage sections. Based on connection information between facilities defined in the intelligent knowledge graph, the flow of pipelines connected to the accident site is visualized on a map using arrows. This allows operators to intuitively understand where hydraulic impacts are propagating and quickly identify the locations of shut-off valves for accident recovery.

[0265] The second screen in the upper right corner displays the visualization results of analyzing factors such as minimum nighttime water pressure. Water pressure data from multiple points within a specific zone is displayed in the form of a bar graph, allowing for a quick comparison of temporal trends and numerical differences between locations. This serves to assist operators in objectively assessing the severity of a situation by presenting hydraulic anomalies analyzed by the agent alongside statistical evidence.

[0266] The third screen at the bottom left is a multi-chart view function that performs precise analysis of specific abnormal water pressure points. By displaying high-resolution time-series data collected from sensors as linear graphs, it provides an environment for precisely examining minute water pressure fluctuations or abnormal vibration patterns. This can be utilized as an essential analytical tool during the final visual verification stage where operators verify the physical validity of the agent's inference results.

[0267] The final screen in the bottom right corner shows the result of an extensive visualization of the overall water pressure distribution within the network. By distinguishing water pressure status at the sub-block level using the size and color of circular markers and displaying them on the map, it enables the immediate identification of areas within the entire system where water pressure imbalances or low pressure issues are occurring.

[0268] Consequently, the interface configuration illustrated in FIG. 22 practically demonstrates the process in which agentic artificial intelligence autonomously operates complex hydraulic analysis tools and reconstructs the analysis results into an expert-level visual report. This technical device enables managers to fully control complex crisis situations in water supply networks solely through conversation with artificial intelligence, without requiring professional data processing skills.

[0269] FIG. 23 illustrates a specific pilot application case for detecting suspected leakage sections based on water pressure analysis. This case covers the Hwaseong sub-block in Cheongyang-gun and demonstrates the process of estimating the location of leakage by precisely analyzing the hydraulic relationship between the change in maximum water pressure at night and the pressure points. The agentic artificial intelligence of the present invention performs inference that considers the physical connectivity between facilities, going beyond simple numerical comparison.

[0270] Looking at the map area on the left side of the drawing, a black, explosion-shaped marker is located in the center of the circular zone indicated by a dotted line; this signifies the estimated location of the leak ultimately determined by the system. The changes in water pressure at surrounding nodes centered on this point are displayed as numerical values ​​inside the red circular marker. For example, values ​​such as -2.08 or -1.25 indicate how much the water pressure has dropped compared to the normal state, while the blue arrow indicates the direction of pressure propagation calculated based on the pipeline connection status defined in the knowledge graph.

[0271] In the time-series chart area on the right side of the drawing, real-time trends in water pressure fluctuations at each key point can be monitored. Patterns in which water pressure drops sharply to a constant level and remains stable from a specific point in time are combined with spatial analysis results on the map to serve as decisive physical evidence for agents to confirm leakage incidents.

[0272] This analysis system demonstrates technical excellence, in which agents trained on various abnormal water pressure patterns leverage expert-level analytical capabilities to precisely locate accident sites within complex pipe networks. Through these visual analysis results, operators can immediately assess the scale and location of accidents and make optimal decisions for on-site recovery operations.

[0273] The agentic AI-based water supply network crisis management device according to the present invention organically integrates fragmented facility information and real-time measurement data within a single intelligent topological structure called a knowledge graph. This data integration system enables three-dimensional recognition of the complex physical connectivity and real-time repair status of the water supply network, thereby serving as a solid foundation for limiting the scope of impact and establishing response plans in the event of an accident.

[0274] The agent controller orchestrates the reasoning capabilities of large-scale language models and various expert tools through a Langgraph-based workflow to precisely diagnose the causes of accidents and autonomously generate unit tasks for resolution. In particular, the sequencing unit identifies the precedence and exclusivity between the mathematical analysis model and the intelligent diagnostic model through guard statements in the tool description, and can dynamically determine the optimal execution sequence that does not violate physical laws.

[0275] The tool interface runs a diagnostic model that fuses supervised and unsupervised learning to detect known and unknown signs of accidents at an early stage, and performs virtual simulations by linking this with a mathematical analysis model. The validation and state management units can complete a self-correcting loop that rolls back the system state to a safe point in time if simulation results violate mathematical thresholds, and autonomously reconstructs the execution plan using correction input information.

[0276] Finally, the language user interface reorganizes complex analysis data into map-based visualizations, multi-charts, and detailed natural language reports, providing them to operators in an integrated manner. This enables data-driven scientific decision-making even in situations where specialized personnel are scarce, prevents secondary damage caused by human error, and allows for the constant maintenance of stability in water supply operations.

[0277] "Real-time" as used in this specification refers to a physical period (about 10-minute intervals in the embodiment of the invention) during which measurement data transmitted by the IoT sensor (200) at the site is accumulated in a remote time-series database (192). This operates independently of the period during which static data, such as the facility connection structure of the knowledge graph, is updated, and the fact that the system infers the state of the observation network in real time defines that the latest time-series data flowing in at such short intervals is used as the subject of analysis.

[0278] In addition, 'Mapping' refers to a system that logically corresponds data existing in different storage locations on a one-to-one basis using the unique identifier of a facility node as a medium, rather than forcibly matching the physical storage location of the data. The agent controller (120) combines the structural information of the graph database (191) and the dynamic measurements of the time series database (192) on-demand at the time of calling the analysis tool, thereby providing the user (agent) with an integrated interface identical to one in which dynamic attributes are directly mapped to the node, even in a physically separated storage environment.

[0279] The work planning unit (121) may possess meta-rules (L1 rules) at the system prompt level to prevent information hallucination and maintain domain consistency during the inference process through a large-scale language model. Identifier hallucination prevention rules can control the system so that numbers included in facility names (e.g., 'Gokseong 6-4') are not mistaken for unique IDs within the system. Instead of arbitrarily estimating an ID from a name, the agent is forced to call a search tool (search_*) first and use only the returned identifier as an argument for a subsequent tool. Entity type branching rules can detect whether keywords such as 'block', 'DMA', and 'zone' are included in the user's query and classify whether to process the analysis target in sub-block units or individual measurement point units. Domain terminology and unit fixing rules interpret specialized water supply terms such as MNP (Most Nighttime Pressure), Inflow / Outflow, and FLOWS_TO (Flow Relationship) into a fixed logical structure, and can provide consistent analysis results by automatically converting internal storage units (e.g., PSI) to user display units (e.g., bar).

[0280] The statistical leak suspicion estimation module among the intelligent diagnostic models of the tool interface (130) can perform analysis by setting a predefined percentile (75 percentile in this embodiment) of the daily nighttime lowest water pressure time series as the normal baseline. This utilizes the hydraulic characteristic that water pressure gradually decreases during periods of low nighttime usage when a leak occurs, and determines that a leak is suspected when the deviation from the daily measurement value exceeds a threshold.

[0281] This statistical analysis technique overcomes the limitation of a short detection window of about one day associated with a simple comparison method compared to the previous day, thereby providing a technical effect that allows for the continuous identification of signs of an accident for about 20 to 25 days after a leak occurs. This extends the leak detection time by more than 20 times compared to conventional technology, serving as a basis for early detection of micro-leaks hidden within the pipe network and significantly improving the water flow rate.

[0282] The agentic AI-based water supply network crisis management device according to the present invention overcomes the technical limitations of the prior art, Patent Documents 1 to 3, and has the following differentiated technical features.

[0283] Conventional patent documents 1 and 2 have a unit model-centered structure that performs individual analysis algorithms, such as pressure prediction or water quality tracking, in a fragmentary manner. On the other hand, the present invention provides an intelligent orchestration function in which an agent controller (120) based on a large-scale language model (LLM) independently recognizes a complex crisis situation combining a leak sound and an abnormal water pressure, and organically selects and links multiple expert tools (a hydraulic analysis model, an intelligent diagnostic model, etc.) suitable for the situation. Through this, it is possible to precisely identify the causes of complex accidents that are difficult to solve with a single model and derive a comprehensive conclusion.

[0284] Patent Document 2 is limited to static network modeling that defines the connection relationships of facilities, and thus has limitations in reflecting real-time changing field conditions. However, the present invention constructs an intelligent knowledge graph that dynamically maps real-time measurement data (water pressure, flow rate, vibration, etc.) as node attributes to the physical topological structure of a water supply network through a graph engine (110). This structure enables the system to spatially and three-dimensionally perceive the hydraulic state of the network beyond simple numerical values, thereby logically limiting the range of impact in the event of an accident and maximizing the accuracy of the analysis.

[0285] Patent Document 3 is specialized only in the preprocessing or correction of structured data and cannot utilize unstructured data, such as on-site repair manuals or past accident response reports, for inference. However, the present invention is equipped with a Search Augmented Generation (RAG) module (180) to call up unstructured documents containing expert tacit knowledge in real time during the accident response process and immediately reflect them in the execution plan. Through this, the system can generate an advanced response guideline in which quantitative data and expert experiential knowledge are fused.

[0286] Conventional systems remain in a passive control method that outputs result values ​​according to predefined fixed scenarios. On the other hand, the present invention performs an active self-correction loop in which an agent establishes an execution plan on its own and, if the result violates a physical threshold, traces back the inference path to modify or roll back the plan. By identifying physical contradictions in the simulation results through the validity verification unit (124) and the state management unit (125) and generating correction input information to reconstruct the plan, the reliability and completeness of the accident response are autonomously secured.

[0287] Therefore, the present invention possesses an inventive step that is clearly distinguishable from conventional technology in that it overcomes the limitations of fragmented data and passive analysis models and provides an intelligent crisis management solution in which the physical characteristics of the water supply network and expert operational logic are perfectly integrated.

[0288] Although embodiments of the present invention have been described as such, the present invention is not limited to the above embodiments, and various modifications and variations are possible within the scope of the technical spirit of the present invention and the equivalent scope of the claims described below by those skilled in the art to which the present invention belongs. Explanation of the symbols

[0289] 10: Water supply network system 100: Water supply network crisis management device 110: Graph Engine 120: Agent Controller 121: Work Planning Department 122: Sequence Determination Department 123: Parameter conversion unit 124: Validity verification unit 125: State Management Department 126: Inference Analysis Department 130: Tool Interface 140: Language User Interface 170: Communication Module 180: RAG Module 190: Database 191: Graph Database 192: Time-series databases 193: Expertise repositories 200: IoT Sensor 300: GIS 400: User terminal

Claims

Claim 1 A graph engine (110) that receives facility coordinates and pipeline data from a GIS (300), structures the topology of a water supply network by setting facilities as nodes and connections between facilities as edges through spatial join operations, and generates a knowledge graph in which physical connection information and dynamic state values ​​of facilities are integrated by mapping real-time measurement data to the nodes; an agent controller (120) that analyzes the physical state of a water supply network by exploring the topology structure of the knowledge graph to obtain the hydraulic influence range between facilities and extracting dynamic state values ​​mapped to at least one node included within the hydraulic influence range; and a tool interface (130) that performs at least one of hydraulic analysis and accident diagnosis according to an execution plan generated for each unit task by analyzing the physical state, and limits the hydraulic influence range by linking the topology structure and dynamic state values ​​of the knowledge graph, and provides the establishment and execution of an accident response plan based on the physical state analysis within the limited range, wherein the agent controller includes a work planning unit (121) that creates a plurality of unit tasks by exploring the knowledge graph through a large-scale language model; a hydraulic analysis model that simulates the physical flow of the water supply network and an intelligent diagnosis that analyzes data patterns. A sequence determination unit (122) that sets the execution order in response to accident scenarios requiring complex utilization of the model;An agentic AI-based water supply network crisis management device comprising: a parameter conversion unit (123) that converts unit tasks into control signals containing API call parameters corresponding to each input protocol and outputs them for technical compatibility with different input specifications for each of the above-mentioned mathematical analysis model and intelligent diagnostic model; and a validity verification unit (124) that verifies validity by comparing the result returned from the tool interface with a mathematical threshold for each node defined in the knowledge graph, and, if a contradiction is found, feeds back the cause of the deficiency to the work planning unit to correct the API call parameters; wherein the agent controller performs a self-correcting loop that integrates and controls mutually heterogeneous mathematical analysis models and intelligent diagnostic models based on API call parameters and secures the physical reliability of the execution plan through threshold-based feedback of the knowledge graph. Claim 2 delete Claim 3 The agentic AI-based water supply network crisis management device according to claim 1, wherein the tool interface identifies accident information by operating an intelligent diagnostic model composed of supervised learning that classifies accident types by learning past leakage accident patterns and unsupervised learning that detects abnormal signs based on the reconstruction error of input data by learning the vibration frequency of a normal state, and predicts water cutoff zones and low-pressure zones by simulating changes in water pressure within the pipe network due to virtual valve operation or pump operation by linking the accident information as input conditions of a hydraulic analysis model, and the sequence determination unit selects a recovery scenario in which water cutoff damage is minimized by comparing the range of water cutoff damage according to a plurality of simulation results returned from the tool interface, and finally confirms the control target and operation sequence according to the selected scenario as an execution plan.

Images