Method for constructing core network operation and maintenance system based on large model and core network operation and maintenance system

Abstract

The application provides a large model-based core network operation and maintenance system construction method and a core network operation and maintenance system. The method comprises the following steps: creating a subgraph structure of a plurality of task execution agents and a graph structure of a supervisor agent; establishing a calling relationship between the graph structure and / or the subgraph structure and a large language model (LLM); compiling the graph structure and the subgraph structure to obtain a graph instance; and constructing a core network operation and maintenance system based on the graph instance. The graph instance is used to define the operation process of the core network operation and maintenance system. The operation process at least comprises the following steps: the supervisor agent responds to a user instruction, calls the LLM to analyze the user instruction to identify a task intention, and dispatches a corresponding task execution agent based on the task intention; and the task execution agent responds to the dispatching and calls the LLM to generate a task execution process, and executes the operation and maintenance task based on the task execution process. In this way, the core network operation and maintenance system can be constructed, the intelligentization and automation of the core network operation and maintenance can be realized, and the operation and maintenance efficiency and accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a method for constructing a core network operation and maintenance system based on a large model and the core network operation and maintenance system. Background Technology

[0002] A mobile communication system consists of multiple parts, including the access network, core network, user equipment, and transmission network. Through their cooperation, the mobile communication system can achieve efficient, reliable, and secure operation. The core network is a key part of the mobile communication system, used for data exchange, routing, and network management.

[0003] With the rapid development of technologies such as 5G, IoT, and cloud computing, network architecture is gradually becoming cloud-based and virtualized, the types and number of network elements have increased significantly, and the complexity of interfaces has also increased, making core network operation and maintenance more difficult.

[0004] Traditional core network operation and maintenance methods mainly rely on manual monitoring and troubleshooting. This results in a heavy workload, low automation, and poor accuracy, making it difficult to meet the high requirements of Service Level Agreements (SLAs) and failing to meet current demands for efficient and precise operation and maintenance. Therefore, it is clear that traditional core network operation and maintenance methods cannot be used to build automated core network operation and maintenance tools, and thus cannot achieve automated operation and maintenance. Summary of the Invention

[0005] This application provides a method for constructing a core network operation and maintenance system based on a large model, as well as a core network operation and maintenance system, to solve the technical problem that traditional core network operation and maintenance methods cannot achieve automated operation and maintenance.

[0006] Firstly, embodiments of this application provide a method for constructing a core network operation and maintenance system based on a large model. The method includes: based on the operation and maintenance tasks of the core network, creating a subgraph structure of multiple task execution agents using the LangGraph framework, and creating a graph structure of a supervisor agent using the LangGraph framework; wherein the supervisor agent is used to communicate with each task execution agent to schedule the task execution agents, and different task execution agents correspond to different operation and maintenance tasks; the task execution agents include at least one of the following agents: a network monitoring agent, a fault detection and location agent, an alarm handling agent, a complaint handling agent, a configuration management agent, a network optimization agent, a data analysis agent, and a knowledge base agent; the network monitoring agent is used to monitor the operating status of the core network; the fault detection and location agent is used to locate the cause and location of faults in the core network; the alarm handling agent is used to process alarm information in the core network; the complaint handling agent is used to process user complaints; and the configuration management agent is used to manage configurations. The management agent is used to configure and manage the core network, the network optimization agent is used to optimize the core network, the data analysis agent is used to analyze the network data of the core network, and the knowledge base agent is used to manage and update the core network operation and maintenance knowledge base. The system establishes the calling relationship between the graph structure and / or subgraph structure and the Large Language Model (LLM). The graph structure and subgraph structure are compiled to obtain graph instances. The core network operation and maintenance system is constructed based on the graph instances. The graph instances define the operation flow of the core network operation and maintenance system. When the core network operation and maintenance system is running, the graph instances run to execute the operation flow. The operation flow includes at least the following: when a user terminal issues a user instruction to the core network operation and maintenance system to execute an operation and maintenance task, the supervisor agent responds to the user instruction, calls the LLM to parse the user instruction to identify the task intent, and schedules one or more corresponding task execution agents based on the task intent. The operation flow also includes: the task execution agent responds to the scheduling and calls the LLM to generate a task execution flow, and executes the operation and maintenance task based on the task execution flow.

[0007] In one possible implementation, the steps of creating a subgraph structure with multiple task execution agents using the LangGraph framework include: creating a subgraph structure based on the LangGraph framework for each task execution agent; creating one or more child agent nodes in the subgraph structure; defining a graph state object for each child agent node, which is used to update and view the graph state object, and the number of child agent nodes is determined based on the task execution flow corresponding to the task execution agent; and / or creating task executor and / or tool nodes in the subgraph structure, the number and type of task executor and tool nodes being determined based on the operation and maintenance task corresponding to the task execution agent; and creating an end node in the subgraph structure.

[0008] In one possible implementation, the steps of creating a subgraph structure with multiple task execution agents using the LangGraph framework further include: setting an entry point for each subgraph structure, where the entry point is a sub-agent node, representing the starting point for the task execution agent to perform operational tasks; and connecting the sub-agent node, task executor, tool node, and / or end node using edge components, where the edge components include conditional edges and unconditional edges, and the conditional edges have routing conditions, which at least include that after the sub-agent node calls the LLM, if the LLM's return message includes a tool call field, then the route is to the task executor and / or tool node; if the return message does not include a tool call field, then the route is to the end node.

[0009] In one possible implementation, after the step of creating task executor and / or tool nodes in the subgraph structure, the method further includes: creating human review nodes in the subgraph structure; and connecting the human review nodes between the sub-agent nodes and the task executor and / or tool nodes using conditional edges.

[0010] In one possible implementation, the steps for creating the graph structure of the supervisor agent include: creating the graph structure based on the LangGraph framework; creating a supervisor agent node in the graph structure; defining a graph state object for the supervisor agent node, which is used to update and view the graph state object; connecting the supervisor agent's graph structure to each subgraph structure using edge components; and further, the method includes: creating an LLM and an external toolset, the external toolset including multiple callable tools, the tools being determined based on the operational tasks; and binding each tool in the external toolset to the LLM through the LLM's interface.

[0011] In one possible implementation, the steps of creating the subgraph structure of the network monitoring agent include: creating a real-time network monitoring agent node, a data acquisition tool node, an anomaly detection tool node, an alarm triggering tool node, and an end node in the subgraph structure of the network monitoring agent; drawing unconditional edges from the real-time network monitoring agent nodes to the data acquisition tool nodes, drawing unconditional edges from the data acquisition tool nodes to the anomaly detection tool nodes, drawing conditional edges from the anomaly detection tool nodes to the alarm triggering tool nodes and the end node respectively, and drawing unconditional edges from the alarm triggering tool nodes to the end node; the steps of creating the subgraph structure of the fault detection and localization agent include: in the fault... In the subgraph structure of the fault detection and localization agent, fault detection and localization agent nodes, fault detection agent nodes, fault delimitation and localization agent nodes, fault report generation agent nodes, fault detection task executor nodes, and an end node are created. Conditional edges are drawn from the fault detection and localization agent nodes pointing to the fault detection agent nodes, fault delimitation and localization agent nodes, fault report generation agent nodes, and the end node, respectively. Unconditional edges are drawn from the fault detection agent nodes, fault delimitation and localization agent nodes, and fault report generation agent nodes pointing to the fault detection and localization agent nodes, respectively. Unconditional edges are also drawn from the fault detection agent nodes pointing to the fault detection task executor nodes. The fault detection task executor node extends an unconditional edge pointing to the fault detection agent node; the steps of creating the subgraph structure of the alarm processing agent include: creating an alarm processing agent node, an alarm processing tool node, an alarm information feedback agent node, an alarm feedback tool node, and an end node in the subgraph structure of the alarm processing agent; extending conditional edges from the alarm processing agent node to the alarm processing tool node, the alarm information feedback agent node, and the end node respectively; extending unconditional edges from the alarm processing agent node and the alarm information feedback agent node to the alarm processing agent node respectively; and extending conditional edges from the alarm information feedback agent node to the alarm feedback tool node, and The steps for creating the subgraph structure of the data analysis agent include: creating a data analysis agent node, a device status query agent node, a device status query tool node, a performance index analysis agent node, a performance index analysis tool node, an alarm data query agent node, an alarm data query tool node, a log analysis agent node, a log analysis tool node, and an end node in the subgraph structure of the data analysis agent; and drawing conditional edges from the data analysis agent node to the status query agent node, the performance index analysis agent node, the alarm data query agent node, the log analysis agent node, and the end node, respectively.Furthermore, unconditional edges are drawn from the device status query agent node, performance indicator analysis agent node, alarm data query agent node, and log analysis agent node, respectively, pointing to the data analysis agent node; and conditional edges are drawn from the device status query agent node to the device status query tool node; and unconditional edges are drawn from the device status query tool node to the device status query agent node; and conditional edges are drawn from the performance indicator analysis agent node to the performance indicator analysis tool node; and conditions are drawn from the performance indicator analysis tool node to the performance indicator analysis agent node. Unconditional edges; and conditional edges leading from the alarm data query agent node to the alarm data query tool node; and unconditional edges leading from the alarm data query tool node to the alarm data query agent node; and conditional edges leading from the log analysis agent node to the log analysis tool node; and unconditional edges leading from the log analysis tool node to the log analysis agent node; the steps for creating a knowledge base agent include: creating a knowledge question-answering agent node and an end node in the subgraph structure of the knowledge base agent; and unconditional edges leading from the knowledge question-answering agent node to the end node.

[0012] Secondly, embodiments of this application provide a core network operation and maintenance system based on a large model. The system is constructed based on the aforementioned first aspect and its various implementations, using a method for building a core network operation and maintenance system based on a large model. The system includes at least a supervisory agent and task execution agents. The supervisory agent communicates with each task execution agent to schedule the task execution agents. Different task execution agents correspond to different operation and maintenance tasks. The supervisory agent is configured to: respond to a user instruction to execute an operation and maintenance task issued by a user terminal, call an LLM to parse the user instruction to identify the task intent, and schedule one or more corresponding task execution agents based on the task intent. The task execution agents include at least one of the following agents: a network monitoring agent and a fault detection and location agent. The system includes an alarm handling agent, a complaint handling agent, a configuration management agent, a network optimization agent, a data analysis agent, and a knowledge base agent. The network monitoring agent monitors the core network's operational status; the fault detection and location agent locates the cause and location of faults in the core network; the alarm handling agent processes alarm information in the core network; the complaint handling agent handles user complaints; the configuration management agent manages the core network's configuration; the network optimization agent optimizes the core network; the data analysis agent analyzes the core network's network data; and the knowledge base agent manages and updates the core network's operation and maintenance knowledge base. The task execution agent is configured to respond to scheduling and invoke the LLM to generate a task execution flow, and then execute operation and maintenance tasks based on this flow.

[0013] In one possible implementation, the supervisor agent is further configured to: generate maintenance task information based on task intent and preset maintenance knowledge, and update the maintenance task information in the graph state object to pass the maintenance task information to the task execution agent through the graph state object; the task execution agent is further configured to: view the graph state object to obtain maintenance task information, and based on the maintenance task information and task execution flow, call its internal nodes to execute the maintenance task, the nodes including sub-agent nodes, task executors, tool nodes, end nodes and / or manual review nodes; after the maintenance task is completed, update the task execution result in the graph state and route it to the supervisor agent or the user; the task execution agent is further configured to: when the task execution flow includes routing to the end node, route to the end node to end the task execution flow; and / or, when the task execution flow includes routing to the manual review node, respond to the user input received by the manual review node, and route to its internal nodes, another task execution agent or the supervisor agent based on the user input.

[0014] In one possible implementation, the network monitoring agent includes: a real-time network monitoring agent node, a data acquisition tool node, an anomaly detection tool node, an alarm triggering tool node, and a termination node. The real-time network monitoring agent node is configured to: receive real-time monitoring task instructions, plan monitoring tasks, and sequentially call the data acquisition tool node, anomaly detection tool node, and alarm triggering tool node to execute the monitoring tasks. The monitoring tasks include data acquisition monitoring tasks, anomaly detection monitoring tasks, and alarm triggering monitoring tasks. The data acquisition tool node is configured to: integrate external data acquisition tools and execute data acquisition monitoring tasks. The anomaly detection tool node is configured to: integrate external anomaly detection tools and execute data acquisition monitoring tasks. The system integrates external alarm notification tools and executes alarm triggering monitoring tasks. If the external alarm detection tool detects an alarm event, the route is routed to the alarm triggering tool node; otherwise, it is routed to the termination node. The alarm triggering tool node is configured to integrate external alarm notification tools and execute alarm triggering monitoring tasks. The fault detection and localization agent includes: a fault detection and localization agent node, a fault detection agent node, a fault delimitation and localization agent node, a fault report generation agent node, a fault detection task executor node, and a termination node. The fault detection and localization agent node is configured to receive alarm information triggered by the network monitoring process and dispatch tasks to the fault detection agent. The system includes nodes for fault identification and localization, and fault report generation. The fault detection agent node is configured to: plan detection tasks according to business rules and output these tasks to the fault detection task executor node for execution. Detection tasks include at least one of the following: device status check, performance indicator check, alarm check, and CHR log check. The fault detection task executor node is configured to: execute detection tasks cyclically. The fault identification and localization agent node is configured to: output identification and localization conclusions based on fault investigation results and fault identification and localization rules. These conclusions include the fault cause and handling suggestions. The fault report generation agent node... The alarm processing agent is configured to: generate a fault analysis report based on the fault investigation results and the delimitation conclusions. The fault analysis report should include at least the equipment's operating status, fault statistics, and performance trends. The alarm processing agent includes: an alarm processing agent node, an alarm processing tool node, an alarm information feedback agent node, an alarm feedback tool, and a termination node. The alarm processing agent node is configured to: receive alarm information and processing suggestions, and dispatch alarm tasks to the alarm processing tool node. The alarm processing tool node is configured to: execute alarm tasks according to alarm classification and alarm level. The alarm information feedback agent node is configured to: feed back the processing results and related information to maintenance personnel or relevant systems.The data analysis agent includes: a data analysis agent node, a device status query agent, a device status query tool, a performance indicator analysis agent, a performance indicator analysis tool, an alarm data query agent, an alarm data query tool, a log analysis agent, a log analysis tool, and an end node. The data analysis agent node is configured to: detect user queries and dispatch tasks to the device status query agent node, performance indicator analysis agent node, alarm data query agent node, and / or log analysis agent node. The device status query agent node is configured to: integrate external tools for device status queries and perform device status data queries. The performance indicator analysis agent node is configured to: integrate external tools for performance indicator analysis and perform performance indicator data query analysis and indicator prediction. The alarm data query agent node is configured to: integrate external tools for alarm data queries and perform alarm data queries and statistics. The log analysis agent node is configured to: integrate external tools for log data queries and perform log data analysis. The knowledge base agent includes: a knowledge question-answering agent node and an end node.

[0015] Thirdly, embodiments of this application provide a core network operation and maintenance system construction device based on a large model. The device includes: a first construction module, used to create a subgraph structure of multiple task execution agents using the LangGraph framework based on the operation and maintenance tasks of the core network, and to create a graph structure of a supervisor agent using the LangGraph framework; wherein, the supervisor agent is used to communicate with each task execution agent to schedule the task execution agents, and different task execution agents correspond to different operation and maintenance tasks; the task execution agents include at least one of the following agents: network monitoring agent, fault detection and location agent, alarm handling agent, complaint handling agent, configuration management agent, network optimization agent, data analysis agent, and knowledge base agent; the network monitoring agent is used to monitor the operating status of the core network; the fault detection and location agent is used to locate the cause and location of faults in the core network; the alarm handling agent is used to process alarm information in the core network; the complaint handling agent is used to process user complaints; and the configuration management agent is used to... The system comprises a configuration management core network, a network optimization agent for optimizing the core network, a data analysis agent for analyzing core network data, and a knowledge base agent for managing and updating the core network operation and maintenance knowledge base. A relationship establishment module establishes the calling relationships between the graph structure and / or subgraph structure and the Large Language Model (LLM). A compilation module compiles the graph structure and subgraph structure to obtain graph instances. A second construction module builds the core network operation and maintenance system based on the graph instances. The graph instances define the operation flow of the core network operation and maintenance system. When the core network operation and maintenance system is running, the graph instances run to execute the operation flow. The operation flow includes at least the following: when a user terminal issues a user instruction to the core network operation and maintenance system to execute an operation and maintenance task, the supervisor agent responds to the user instruction, calls the LLM to parse the user instruction to identify the task intent, and schedules one or more corresponding task execution agents based on the task intent. The operation flow also includes: the task execution agent responds to the scheduling and calls the LLM to generate a task execution flow, and executes the operation and maintenance task based on the task execution flow.

[0016] As can be seen from the above, this application provides a method for constructing a core network operation and maintenance system based on a large model, and a core network operation and maintenance system. The method includes: based on the operation and maintenance tasks of the core network, creating a subgraph structure of multiple task execution agents using the LangGraph framework, and creating a graph structure of a supervisor agent using the LangGraph framework; wherein, the supervisor agent is used to communicate with each task execution agent to schedule the task execution agents, and different task execution agents correspond to different operation and maintenance tasks; the task execution agents include at least one of the following agents: a network monitoring agent, a fault detection and location agent, an alarm handling agent, a complaint handling agent, a configuration management agent, a network optimization agent, a data analysis agent, and a knowledge base agent; the network monitoring agent is used to monitor the operating status of the core network; the fault detection and location agent is used to locate the cause and location of faults in the core network; the alarm handling agent is used to process alarm information in the core network; and the complaint handling agent is used to process user complaints. The system employs a user complaint handling mechanism, a configuration management agent for configuring and managing the core network, a network optimization agent for optimizing the core network, a data analysis agent for analyzing core network data, and a knowledge base agent for managing and updating the core network operation and maintenance knowledge base. It establishes the calling relationship between a graph structure and / or subgraph structure and a Large Language Model (LLM). The graph structure and subgraph structures are compiled to obtain graph instances. A core network operation and maintenance system is constructed based on these graph instances. The graph instances define the operational flow of the core network operation and maintenance system. When the system is running, the graph instances execute the operational flow. The operational flow includes at least the following: when a user issues a user instruction to the core network operation and maintenance system to perform an operation and maintenance task, the supervisor agent responds to the user instruction, calls the LLM to parse the user instruction to identify the task intent, and schedules one or more corresponding task execution agents based on the task intent. The operational flow also includes: the task execution agent responds to the scheduling and calls the LLM to generate a task execution flow, and executes the operation and maintenance task based on the task execution flow. This allows for the construction of a core network operation and maintenance system, enabling intelligent and automated core network operation and maintenance, improving efficiency and accuracy. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the core network operation and maintenance system provided in the embodiments of this application;

[0018] Figure 2 A flowchart illustrating the method for constructing a core network operation and maintenance system based on a large model, as provided in this application embodiment;

[0019] Figure 3 This is a schematic diagram of the graph structure of the supervisory agent provided in an embodiment of this application;

[0020] Figure 4This is a schematic diagram of the network monitoring intelligent agent subgraph structure provided in the embodiments of this application;

[0021] Figure 5 This is a schematic diagram of the subgraph structure of the fault detection and localization intelligent agent provided in the embodiments of this application;

[0022] Figure 6 This is a schematic diagram of the subgraph structure of the alarm processing agent provided in the embodiments of this application;

[0023] Figure 7 This is a schematic diagram of the subgraph structure of the data analysis agent provided in the embodiments of this application;

[0024] Figure 8 This is a schematic diagram of the subgraph structure of the knowledge base agent provided in the embodiments of this application;

[0025] Figure 9 A schematic diagram of a subgraph structure including tool review nodes provided for embodiments of this application;

[0026] Figure 10 This is a schematic diagram of a multi-turn dialogue architecture provided in an embodiment of this application;

[0027] Figure 11 This is a schematic diagram of the core network operation and maintenance system based on a large model provided in an embodiment of this application;

[0028] Figure 12 A schematic diagram of the structure of the core network operation and maintenance system construction device based on a large model provided in the embodiments of this application. Detailed Implementation

[0029] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of this application.

[0030] Before introducing the technical solutions of the embodiments of this application, the terminology involved in the embodiments of this application will be introduced by way of example.

[0031] 1. Core Network: An important component of a mobile communication system, it connects to external network components such as the access network, transmission network, and radio access network, and is responsible for carrying and managing the control, signaling, and data processing of the communication system, as well as connecting and exchanging data between different networks.

[0032] 2. Agent: A system or entity that can autonomously perceive its environment, make decisions, and perform actions to achieve a specific goal.

[0033] To improve the accuracy and timeliness of core network operation and maintenance, this application provides a method for constructing a core network operation and maintenance system based on a large model, as well as a core network operation and maintenance system.

[0034] 3. Language Graph Framework: A model or tool for the structured representation and processing of language information, used to build stateful, multi-participant applications based on large language models (LLMs) by modeling steps as edges and nodes in a graph.

[0035] Specifically, the construction method provided in this application combines large model technology and LangGraph technology to build a core network intelligent operation and maintenance framework. This framework leverages the language understanding and generation capabilities of large models to accurately parse user operation and maintenance needs, and utilizes the intelligent proxy function of the LangGraph framework to automate the execution and real-time monitoring of operation and maintenance tasks. This improves operation and maintenance efficiency and accuracy, reduces operation and maintenance costs, and enhances user experience.

[0036] Figure 1 A schematic diagram of the core network operation and maintenance system provided in the embodiments of this application.

[0037] like Figure 1 As shown, the core network operation and maintenance system built based on the construction method provided in this application embodiment can include multiple architectural layers. Specifically, the core network operation and maintenance system provided in this application embodiment uses infrastructure or computer as the underlying framework, and the infrastructure and computer can be used to provide computing and storage resources. Furthermore, the core network operation and maintenance system provided in this application embodiment can build a multi-Agent orchestration and scheduling framework based on the LangGraph framework. This multi-Agent orchestration and scheduling framework can communicate with the API layer and the Large Language Model (LLM) or a collection of multiple Large Language Models (LLMs) to integrate the intelligent question answering, text analysis and other functions of the Large Language Model (LLM). Furthermore, based on the multi-agent orchestration and scheduling framework, a supervisor agent and specialized agents can be formed. The supervisor agent is used to receive instructions to schedule specialized agents. Specialized agents, also known as task execution agents, can include network monitoring agents, detection and location agents, alarm handling agents, complaint handling agents, fault management agents, network optimization agents, knowledge base agents, etc. Each specialized agent is specifically related to the operation and maintenance tasks performed by the core network operation and maintenance system.

[0038] Furthermore, the core network operation and maintenance system framework integrates multiple layers, including external toolsets, data management, and large-scale model training. The external toolset includes operation and maintenance automation tools, such as status monitoring toolsets, performance monitoring toolsets, alarm monitoring toolsets, and log analysis toolsets, used to acquire core network status information and handle operation and maintenance tasks. The data management layer is used for data storage, retrieval, and management, providing data support for intelligent agents. The large-scale model training layer is used to train and optimize large language models.

[0039] As can be seen, the core network operation and maintenance system provided in this application, under its framework, can realize the collaborative work of intelligent agents and toolsets at different levels, thereby achieving efficient operation and maintenance and management of the core network. For example, it can be applied to scenarios such as knowledge-based question answering, monitoring and troubleshooting, and performance optimization.

[0040] It is worth noting that the core network operation and maintenance system provided in this application embodiment may include the following parts:

[0041] ①Large Language Model (LLM): This includes a large language model, an embedding model, and a multimodal model. In this application, a unified large model access API layer can be constructed, which can realize unified access to multiple models and flexible switching.

[0042] ② Data Management: Access to various intermediate and persistent data during Agent management and operation, including structured and unstructured knowledge documents, vector databases, analytical data, message history, log data, etc., and provide necessary data maintenance and management tools, such as cleaning, vectorization, import and export of private knowledge data.

[0043] ③ Large Model Operation and Maintenance Management: This includes the operation and maintenance management of LLM and various intelligent agents built on top of LLM, specifically including management work at different stages of the application lifecycle.

[0044] ④ Development and orchestration framework: The LangGraph framework is a low-level library designed for building applications with complex state management and multi-role interaction capabilities. Based on the LangGraph framework, the embodiments of this application can simplify the complexity and workload of building upstream agents and reduce risks.

[0045] ⑤ Core network operation and maintenance intelligent agents (including supervisory intelligent agents and task execution intelligent agents): These are intelligent agents built upon the development and orchestration framework. Core network operation and maintenance intelligent agents collaborate with external tools through APIs or code interpreters to complete tasks. Task execution intelligent agents specifically include:

[0046] Network monitoring intelligent agent: Real-time monitoring of the status and performance indicators of the core network and equipment, detection of abnormal states and errors in the network, and triggering alarms.

[0047] Fault detection and location intelligence agent: Locates the cause and location of the fault, outputs solutions or handling suggestions, and guides maintenance personnel to carry out repair work.

[0048] Alarm handling agent: Receives alarm information and triggers corresponding processing procedures to quickly restore the normal operating status of the device.

[0049] Intelligent Complaint Handling Agent: This agent is primarily used to handle user complaints. It can automatically complete complaint classification, diagnosis, signaling analysis, and work order completion in a one-stop process. By automating the process, it reduces manual intervention and improves the speed and accuracy of complaint handling.

[0050] Configuration Management Agent: The configuration management agent is responsible for the configuration management of the core network, including device configuration and network topology configuration. Through automated configuration management, it reduces human error and improves configuration efficiency and accuracy.

[0051] Network optimization agent: The network optimization agent is responsible for optimizing the core network, including performance optimization and resource optimization. Through intelligent analysis of network data, it identifies potential performance bottlenecks and resource waste, proposes optimization suggestions, and implements them to improve network performance and resource utilization.

[0052] Data analytics agent: It analyzes and mines various types of network data to discover network operation trends and potential problems, helping operations and maintenance personnel to better understand the network's operating status.

[0053] Knowledge Base Agent: Stores and manages knowledge and cases related to core network operation and maintenance, and provides knowledge question-and-answer interfaces for other agents.

[0054] Task scheduling agent: Identifies and senses the needs of operations and maintenance personnel, and schedules the operations and maintenance needs to the relevant agents according to the agent's capabilities.

[0055] Operations Supervisor Agent (hereinafter referred to as Supervisor Agent): The Operations Supervisor Agent acts as a bridge between humans and machines, enabling operations personnel to easily communicate with intelligent assistants through natural language processing interfaces, and to delegate the operations tasks proposed by users to various professional agents for collaborative processing. After the task is completed, the results are returned to the user.

[0056] ⑥ External toolset: including equipment status monitoring and query tools, alarm monitoring and query tools, log analysis tools, performance monitoring and query tools, etc., used to obtain core network status information and handle operation and maintenance tasks.

[0057] Figure 2 A flowchart illustrating the method for constructing a core network operation and maintenance system based on a large model, as provided in this application embodiment.

[0058] See Figure 1and Figure 2 This application provides a method for constructing a core network operation and maintenance system based on a large model, which may include steps S100-S400.

[0059] S100: Operation and maintenance tasks based on the core network, using the LangGraph framework to create a subgraph structure of multiple task execution agents, and a graph structure of a supervisor agent; wherein, the supervisor agent is used to communicate with each task execution agent to schedule the task execution agents, and different task execution agents correspond to different operation and maintenance tasks.

[0060] The steps of creating the subgraph structure of the task execution agent and the graph structure of the supervisor agent refer to constructing the graph structure of the agent. Specifically, this may include defining components such as nodes, edges, and states, and determining the connection relationships and interaction methods between the agents. Nodes can represent specific steps or operations in the operation and maintenance task, edges represent logical relationships or data flows between steps, and states represent globally shared information during the process of execution.

[0061] Task execution agents are used to perform specific operational tasks, such as network monitoring, alarm handling, complaint processing, configuration management, and network optimization, thereby improving the professionalism and accuracy of task execution. This ensures that each task is completed by a specific agent and avoids performance degradation caused by overloading a single agent.

[0062] The operations and maintenance supervisor intelligent agent acts as a bridge between humans (operations and maintenance personnel) and machines, enabling operations and maintenance personnel to communicate with the operations and maintenance system through a natural language processing interface, and to delegate the operations and maintenance tasks proposed by users to various professional intelligent agents for collaborative processing, and return the results to the user after the task is completed.

[0063] It is understandable that the specific composition of the task execution intelligent agent subgraph structure and the supervisor intelligent agent graph structure can be determined based on actual needs. In this embodiment, the task execution intelligent agent includes at least one of the following intelligent agents: network monitoring intelligent agent, fault detection and location intelligent agent, alarm handling intelligent agent, complaint handling intelligent agent, configuration management intelligent agent, network optimization intelligent agent, data analysis intelligent agent, and knowledge base intelligent agent. The network monitoring intelligent agent is used to monitor the operating status of the core network; the fault detection and location intelligent agent is used to locate the cause and location of faults in the core network; the alarm handling intelligent agent is used to process alarm information in the core network; the complaint handling intelligent agent is used to process user complaints; the configuration management intelligent agent is used to configure and manage the core network; the network optimization intelligent agent is used to optimize the core network; the data analysis intelligent agent is used to analyze the network data of the core network; and the knowledge base intelligent agent is used to manage and update the core network operation and maintenance knowledge base. The specific structure of each task execution intelligent agent will be detailed below and will not be repeated here.

[0064] S200: Establish the calling relationship between the graph structure and / or subgraph structure and the large language model LLM.

[0065] It is understandable that a Large Language Model (LLM) can understand natural language instructions, enabling it to comprehend the requirements of operational tasks and generate corresponding task execution flows. In this embodiment, the LLM can be connected to graph structures and / or subgraph structures via an API interface. Thus, during the core network operational system application phase, the LLM can interact with the supervisor agent and / or task execution agent to guide the agent in completing complex operational tasks.

[0066] S300: Compile the graph structure and its subgraph structure to obtain a graph instance.

[0067] Understandably, compiling the graph structure and its subgraphs transforms them into executable graph instances, which represent the operational flow of the core network operations and maintenance (O&M) system. Thus, when the O&M system receives a user's O&M task request, it can begin executing the task based on the compiled graph instances. For example, by inputting an initial state (such as user instructions, the current state of the core network, etc.), the graph instances can automatically schedule various task execution agents to perform O&M tasks according to the defined relationships between nodes and edges.

[0068] S400: A core network operation and maintenance system built based on graph instances.

[0069] When the core network operation and maintenance system is running, the graph instance can run to execute the operation process.

[0070] Furthermore, the steps for deriving the core network operation and maintenance system based on graph instances include at least constructing the user terminal of the core network operation and maintenance system. Specifically, this may include constructing a user interface, which serves as the front-end entry point for users to interact with the core network operation and maintenance system, allowing users to input user commands. Further, the interface between the user interface and the graph instances can be defined, such as defining the data transmission format, protocol, and method to ensure accurate data transmission and stable interaction, thereby integrating the user interface with the graph instances.

[0071] Furthermore, graph instances encapsulate the logic of the operation and maintenance system, including task scheduling, state management, and agent collaboration. After receiving user commands through the user interface, the graph instances execute them, obtain the operation results, and then the operation results can be fed back to the user through the user interface.

[0072] In some implementations, step S400 may also include deploying the core network operation and maintenance system to a suitable server and performing configuration and optimization steps to ensure that the system can operate stably. This application embodiment does not specifically limit this.

[0073] It is worth noting that, based on the graph instance, the core network operation and maintenance system operation process includes at least the following: when the user terminal issues a user instruction to the core network operation and maintenance system to execute an operation and maintenance task, the supervisor agent calls the LLM to parse the user instruction to identify the task intent, and schedules one or more corresponding task execution agents based on the task intent; and the operation process also includes: the task execution agent responds to the scheduling and calls the LLM to generate a task execution process, and executes the operation and maintenance task based on the task execution process.

[0074] In other words, the compiled graph instance can be used to implement the following process:

[0075] ① Responding to user input commands: When a user inputs a command, the supervisor agent can parse the command through the LLM node and identify the task intent.

[0076] ② Task scheduling: The supervisor agent can schedule the corresponding task execution agents to perform operation and maintenance tasks based on the task intent parsed by the LLM.

[0077] ③ Task execution parsing: Each task execution agent can call LLM to further parse the task execution process and obtain the execution steps or execution flow.

[0078] ④ Task execution: The task execution agent can execute tasks based on the LLM feedback process and finally return the result or complete the task.

[0079] As can be seen from the above, this application provides a method for constructing a core network operation and maintenance system based on a large model. This method includes: creating a subgraph structure of multiple task execution agents using the LangGraph framework based on the core network's operation and maintenance tasks; and creating a graph structure of a supervisor agent using the LangGraph framework. The supervisor agent communicates with each task execution agent to schedule the task execution agents, and different task execution agents correspond to different operation and maintenance tasks. The task execution agents include at least one of the following agents: a network monitoring agent, a fault detection and location agent, an alarm handling agent, a complaint handling agent, a configuration management agent, a network optimization agent, a data analysis agent, and a knowledge base agent. The network monitoring agent monitors the core network's operating status; the fault detection and location agent locates the cause and location of faults in the core network; the alarm handling agent processes alarm information in the core network; and the complaint handling agent processes user complaints. A configuration management agent is used to configure and manage the core network; a network optimization agent is used to optimize the core network; a data analysis agent is used to analyze the network data of the core network; and a knowledge base agent is used to manage and update the core network operation and maintenance knowledge base. The system establishes the calling relationship between the graph structure and / or subgraph structure and the Large Language Model (LLM); compiles the graph structure and subgraph structure to obtain graph instances; and constructs the core network operation and maintenance system based on the graph instances. The graph instances define the operation flow of the core network operation and maintenance system. When the core network operation and maintenance system is running, the graph instances run to execute the operation flow. The operation flow includes at least the following: when a user terminal issues a user instruction to the core network operation and maintenance system to execute an operation and maintenance task, the supervisor agent responds to the user instruction, calls the LLM to parse the user instruction to identify the task intent, and schedules one or more corresponding task execution agents based on the task intent; and the operation flow also includes: the task execution agent responds to the scheduling and calls the LLM to generate a task execution flow, and executes the operation and maintenance task based on the task execution flow. In this way, a core network operation and maintenance system can be constructed, and based on this system, intelligent and automated core network operation and maintenance can be achieved, improving operation and maintenance efficiency and accuracy.

[0080] Furthermore, in the embodiments of this application, the task execution intelligent agent may specifically include the following types:

[0081] ① Network monitoring intelligent agent: Used to monitor the operational status of the core network in real time, including key indicators such as device status, network performance, network traffic, and number of user connections. Once an anomaly is detected, it immediately triggers the alarm handling process to ensure that the problem is resolved in a timely manner.

[0082] ② Fault detection and location intelligence agent: used to detect network status, performance, alarms, and logs, locate the cause and location of faults, output solutions or handling suggestions, and guide operation and maintenance personnel to carry out repair work.

[0083] ③ Alarm Processing Intelligent Agent: Used to process alarm information in the core network. Specifically, it can interface with the work order system to achieve one-click alarm knowledge Q&A, case recommendation, and intelligent diagnosis. By intelligently analyzing alarm data, it can quickly locate the root cause of the problem, reduce manual intervention, shorten alarm processing time, and improve operation and maintenance efficiency.

[0084] ④ Intelligent Complaint Handling Agent: This agent handles user complaints, performing tasks such as complaint classification, diagnosis, signaling analysis, and work order completion. By automating processes, it reduces manual intervention and improves the speed and accuracy of complaint handling.

[0085] ⑤ Configuration Management Agent: Used for core network configuration management, including device configuration and network topology configuration. Automated configuration management reduces human error and improves configuration efficiency and accuracy.

[0086] ⑥ Network Optimization Intelligent Agent: Used to optimize the core network, including performance optimization and resource optimization. By intelligently analyzing network data, it identifies potential performance bottlenecks and resource waste, proposes optimization suggestions, and implements them to improve network performance and resource utilization.

[0087] ⑦ Knowledge Base Agent: Used to manage and update the internal core network operation and maintenance knowledge base, ensuring that all agents can access the latest and most accurate operation and maintenance knowledge.

[0088] It is worth noting that if the graph structure or subgraph structure of the embodiments of this application is visualized, the graph structure or subgraph structure can be a set of interconnected nodes and edges, where nodes represent supervisor agents, task execution agents, and LLMs, and edges represent the communication and control relationships between nodes.

[0089] Figure 3 This is a schematic diagram of the graph structure of the supervisory agent provided in an embodiment of this application.

[0090] like Figure 3 As shown, the graph structure of the supervisor agent provided in this embodiment includes a supervisor agent node and multiple task execution agent nodes. Conditional edges pointing to each task execution agent node can be derived from the supervisor agent node. Figure 3 In the diagram, dashed lines with arrows represent conditional edges. Each task-executing agent can generate unconditional edges pointing to the supervisor agent. Figure 3 In the diagram, unconditional edges are represented by arrows. Conditional and unconditional edges are edge components, with conditional edges having routing conditions. Based on conditional edges, the supervisor agent node can route to various task execution agent nodes to perform operational tasks. The diagram only shows a portion of the task execution agent nodes; the actual task execution agents in the graph structure can be added or removed based on actual needs.

[0091] The supervisory agent can be called a Supervisor node. In this embodiment, prompts and routes can be designed for the supervisory agent to manage the various task execution agents.

[0092] The code for designing a prompt can be:

[0093] members = ["Real-time monitoring agent", "Fault detection and location agent", "Alarm handling agent", "Data analysis agent", "Knowledge-based question answering agent"]

[0094] system_prompt =

[0095] "You are an Agent Manager, responsible for managing the conversations between the following Agents: {members}. Given the following user requests, the agents respond together to take the next step. Each agent will perform a task and respond with its result and status. Upon completion, a 'FINISH' response is given."

[0096] The routing design is used to define the results of Supervisor node routing, so that Supervisor nodes can dispatch user queries to the next Agent node by calling llm.with_structured_output.

[0097] For example, the code for route design can be:

[0098]

[0099]

[0100] Based on this code, an LLM-based task routing controller can be implemented to build a supervisor node in a multi-agent collaborative system, enabling the supervisor agent to perform the following functions: dynamically determine which sub-agent (task execution agent) should be called next or end the task based on the current dialogue state.

[0101] Figure 4 The example illustrates a visual representation of the subgraph structure of a network monitoring intelligent agent. Specifically, the network monitoring intelligent agent may include network real-time monitoring Agent nodes, data acquisition tool nodes, anomaly detection tool nodes, alarm triggering tool nodes, and end nodes, as well as start nodes. The specific composition of the subgraph structure can be designed based on actual needs, and this application embodiment does not impose specific limitations on it.

[0102] In this embodiment of the application, the agent construction process may include four key steps: LangGraph graph structure design, large model integration, external tool interface development, and state management implementation.

[0103] ① LangGraph Graph Structure Design: Constructing the graph structure of the agent, including defining components such as nodes, edges, and states, and determining the connections and interaction methods between them. Nodes can represent specific steps or operations in the operation and maintenance task, edges represent logical relationships or data flows between steps, and states represent globally shared information during the process execution.

[0104] ② Large Model Integration: In the graph structure, large language models (LLMs) are integrated into the agent as key nodes. LLMs are responsible for processing natural language instructions, understanding the requirements of operational tasks, and generating corresponding execution plans. By interacting with other nodes in the graph structure, LLMs can guide the agent to complete complex operational tasks.

[0105] ③ External Tool Interface Development: To automate the execution of operation and maintenance tasks, the agent needs to interact with external tools. Therefore, corresponding interfaces need to be developed to connect the agent and external tools. These interfaces can be customized according to the type and function of the external tools to ensure that the agent can correctly call the external tools and execute the corresponding operation and maintenance operations.

[0106] ④ State Management Implementation: The state management function based on the LangGraph framework ensures that the agent maintains consistent contextual understanding and decision-making capabilities when performing operational tasks. By defining state graphs and state transition rules, the agent can perform different operations in different states and dynamically adjust its workflow as needed.

[0107] Furthermore, in order to build a core network operation and maintenance system based on a large model, the method provided in this application embodiment can first initialize the large model and tools. Initializing the large model can refer to instantiating an LLM object, and initializing the tools can refer to each tool in the external toolset and binding it to the LLM through the LLM interface.

[0108] Furthermore, prior to step S200, the method provided in this application embodiment may also include the following steps S501-S502.

[0109] S501: Create an LLM and external toolset. The external toolset includes multiple callable tools, which are determined based on the operation and maintenance tasks.

[0110] In this embodiment, the core network operation and maintenance assistant can integrate multiple external tools to complete specific operation and maintenance tasks. These external tools are integrated into the graph structure, and corresponding calling code is written, enabling the operation and maintenance assistant to utilize these tools to complete specific operation and maintenance tasks. For example, a database can be used to store and query core network status information, log analysis tools can be used to analyze device log information, and performance monitoring tools can be used to monitor core network performance metrics. These tools can be integrated and invoked through LangGraph's API.

[0111] In this embodiment of the application, the step of instantiating an LLM object to create an LLM may include the step of loading a pre-trained model. The pre-trained model may be an open-source model, and this embodiment of the application does not specifically limit it.

[0112] S502: Binds various tools from an external toolset to the LLM through the LLM interface.

[0113] The interface for LLM is, for example, the bind_tools() interface, which is not specifically limited in this embodiment of the application.

[0114] Furthermore, in the embodiments of this application, step S100 may specifically include steps S101-S102.

[0115] S101: For each task, an agent is executed to create a subgraph structure based on the LangGraph framework.

[0116] Understandably, based on the LangGraph framework, a subgraph structure is a graph.

[0117] S102: Create one or more child agent nodes in the subgraph structure, and define a graph state object for the child agent node. The child agent node is used to update and view the graph state object. The number of child agent nodes is determined based on the task execution flow corresponding to the task execution agent.

[0118] For example, such as Figure 3 As shown, the subgraph structure of the network monitoring agent can include only one sub-agent node, namely the network real-time monitoring agent.

[0119] Figure 5 This is a schematic diagram of the subgraph structure of the fault detection and localization agent provided in the embodiments of this application.

[0120] like Figure 5As shown, the subgraph structure of the alarm processing agent may include four sub-agent nodes, namely, the fault detection and location agent node, the fault detection agent node, the fault delimitation and location agent node, and the fault report generation agent node, and may include a start node and an end node. This application embodiment does not specifically limit this.

[0121] It should also be noted that defining a graph state object for a sub-agent node refers to initializing the subgraph structure using the LangGraph's built-in MessagesState object. Specifically, the graph state object can be used to maintain state information related to the execution of operational tasks, such as the current progress of the task, data input and output, and processing results.

[0122] For example, the State design of a network monitoring agent can be:

[0123] By inheriting from the MessagesState class of the LangGraph framework, three new properties are added: collect_data, abnormal_info, and alarm_info. collect_data is used to collect data, abnormal_info is used to record abnormal information, and alarm_info is used to process alarm information.

[0124] collect_data:Annotated[List[Tuple],operator.add] / / Collect data;

[0125] abnormal_info:Annotated[List[Tuple],operator.add] / / Abnormal information;

[0126] alarm_info: Annotated[List[Tuple], operator.add] / / Alarm information;

[0127] And / or, step S100 may also include the following steps S103-S104.

[0128] S103: Create task executors and / or tool nodes in the subgraph structure. The number and type of task executors and tool nodes are determined based on the operation and maintenance tasks corresponding to the task execution agent.

[0129] The executor can refer to a collection of tools, such as... Figure 5As shown, the fault detection and localization agent can include a fault detection Executor. This Executor can integrate external tools such as device status checks, performance indicator checks, alarm checks, and chr log checks, and execute detection tasks cyclically. It is understood that these tools can be integrated and invoked through the APIs provided by the LangGraph architecture.

[0130] S104: Create the end node.

[0131] Furthermore, after step S104, the following steps S105-S106 may also be included.

[0132] S105: For each subgraph structure, set its entry point, which is the sub-agent node.

[0133] The entry point refers to the starting point where the task execution agent begins to perform the operation and maintenance task.

[0134] S106: Connect sub-agent nodes, task executors, tool nodes, and / or end nodes using edge components. Edge components include conditional edges and unconditional edges. Conditional edges have routing conditions. The routing conditions include at least the following: after the sub-agent node calls the LLM, if the LLM's return message includes a tool call field, then the route is to the task executor and / or tool node; if the return message does not include a tool call field, then the route is to the end node.

[0135] Understandably, the tool call field is used to indicate that further calls to tools or task executors are needed.

[0136] It is understood that routing conditions can be designed based on the actual task objectives of the operation and maintenance tasks, and this application embodiment does not specifically limit this.

[0137] The following is an example of how to create the subgraph structure for each task-executing agent.

[0138] See also Figure 4 In the core network operation and maintenance assistant, the network real-time monitoring intelligent agent plays the role of real-time network monitoring and status feedback. It monitors the operating status and performance parameters of core network devices in real time, and promptly detects and reports equipment faults or abnormal states. Core network devices include routers, switches, gateway devices, server devices, controller devices, and security devices.

[0139] The steps for creating a subgraph structure for a network monitoring agent may include:

[0140] S100-a: Create network real-time monitoring agent nodes, data acquisition tool nodes, anomaly detection tool nodes, alarm triggering tool nodes, and termination nodes in the subgraph structure of the network monitoring agent;

[0141] S100-b: An unconditional edge from the network real-time monitoring agent node to the data acquisition tool node, an unconditional edge from the data acquisition tool node to the anomaly detection tool node, a conditional edge from the anomaly detection tool node to the alarm triggering tool node and the termination node respectively, and an unconditional edge from the alarm triggering tool node to the termination node.

[0142] In this way, the anomaly detection tool can be routed to the alarm triggering tool when an alarm event exists to trigger an alarm, and can be routed to the end node to terminate the task process when no alarm event exists.

[0143] See also Figure 5 The fault detection and location intelligence agent is used to detect network status, performance, alarms, and logs, locate the cause and location of faults, output solutions or handling suggestions, and guide operation and maintenance personnel to carry out repair work.

[0144] The steps for creating a subgraph structure for a fault detection and localization agent may include:

[0145] S100-c: Create a fault detection and localization agent node, a fault detection agent node, a fault delimitation and localization agent node, a fault report generation agent node, a fault detection task executor node, and an end node in the subgraph structure of the fault detection and localization agent.

[0146] S100-d: Conditional edges extending from the fault detection and localization agent node to the fault detection agent node, the fault delimitation and localization agent node, the fault report generation agent node, and the termination node, respectively; and unconditional edges extending from the fault detection agent node, the fault delimitation and localization agent node, and the fault report generation agent node to the fault detection and localization agent node, respectively; and unconditional edges extending from the fault detection agent node to the fault detection task executor node; and unconditional edges extending from the fault detection task executor node to the fault detection agent node.

[0147] Figure 6 This is a schematic diagram of the subgraph structure of the alarm processing agent provided in the embodiments of this application.

[0148] like Figure 6As shown, the alarm processing agent may include alarm processing agent nodes, alarm processing tool nodes, alarm information feedback agent nodes, alarm feedback tools, and termination nodes, and may also include a start node. The alarm processing agent receives alarm information sent by the network monitoring agent and processing suggestions sent by the fault diagnosis and location agent, triggers corresponding processing flows, such as restarting the device or adjusting the configuration, and feeds back the processing results to maintenance personnel or relevant systems.

[0149] The steps for creating a subgraph structure for an alarm handling agent may include:

[0150] S100-e: Create alarm processing agent nodes, alarm processing tool nodes, alarm information feedback agent nodes, alarm feedback tool nodes, and an end node in the subgraph structure of the alarm processing agent;

[0151] S100-f: Conditional edges originating from the alarm processing agent node pointing to the alarm processing tool node, the alarm information feedback agent node, and the termination node, respectively; and unconditional edges originating from the alarm processing agent node and the alarm information feedback agent node pointing to the alarm processing agent node, respectively; and conditional edges originating from the alarm information feedback agent node pointing to the alarm feedback tool node, and unconditional edges originating from the alarm feedback tool node pointing to the alarm information feedback agent node.

[0152] Figure 7 This is a schematic diagram of the subgraph structure of the data analysis agent provided in the embodiments of this application.

[0153] like Figure 7 As shown, a data analysis agent may include a data analysis agent node, a device status query agent, a device status query tool, a performance indicator analysis agent, a performance indicator analysis tool, an alarm data query agent, an alarm data query tool, a log analysis agent, a log analysis tool, and an end node, and may also include a start node. The data analysis agent is used to analyze and mine collected network data, identify potential network faults or performance bottlenecks, provide optimization suggestions, or predict future network trends.

[0154] The steps for creating a subgraph structure for a data analytics agent may include:

[0155] S100-g: Creates the following nodes in the subgraph structure of the data analysis agent: data analysis agent node, device status query agent node, device status query tool node, performance index analysis agent node, performance index analysis tool node, alarm data query agent node, alarm data query tool node, log analysis agent node, log analysis tool node, and end node.

[0156] S100-h: Conditional edges originating from the data analysis agent node, pointing to the status query agent node, performance indicator analysis agent node, alarm data query agent node, log analysis agent node, and termination node, respectively; and unconditional edges originating from the device status query agent node, performance indicator analysis agent node, alarm data query agent node, and log analysis agent node, respectively, pointing to the data analysis agent node; and conditional edges originating from the device status query agent node, pointing to the device status query tool node; and conditions originating from the device status query tool node, pointing to the device status query agent node. Unconditional edges of nodes; and conditional edges leading from the performance index analysis agent node to the performance index analysis tool node; and unconditional edges leading from the performance index analysis tool node to the performance index analysis agent node; and conditional edges leading from the alarm data query agent node to the alarm data query tool node; and unconditional edges leading from the alarm data query tool node to the alarm data query agent node; and conditional edges leading from the log analysis agent node to the log analysis tool node; and unconditional edges leading from the log analysis tool node to the log analysis agent node.

[0157] Figure 8 A schematic diagram of the subgraph structure of the knowledge base agent provided in the embodiments of this application.

[0158] like Figure 8 As shown, the knowledge base agent can include knowledge question-answering agent nodes and an end node, and may also include a start node. The knowledge base agent is used to store and manage knowledge and experience related to core network operation and maintenance, and provides other agents with interfaces for querying operation and maintenance knowledge and cases.

[0159] The steps for creating a subgraph structure for a knowledge base agent may include:

[0160] S100-i: Creates a knowledge-answering agent node and an end node in the subgraph structure of the knowledge base agent;

[0161] S100-j: An unconditional edge originating from a knowledge-answering agent node and pointing to the end node.

[0162] It should be added that, Figures 4-8 In the diagram, solid lines with arrows represent unconditional edges, and dashed lines with arrows represent conditional edges.

[0163] Furthermore, in this embodiment of the application, a human-machine collaborative working mode can also be set for the core network operation and maintenance system. The "Human-in-the-Loop" human-machine collaborative working mode based on the LangGraph framework integrates human (operation and maintenance personnel) input into the automated process through the embedded memory mechanism.

[0164] Specifically, step S102 may be followed by steps S107-S108.

[0165] S107: Create a human_review node in the subgraph structure.

[0166] It's worth noting that the human review node interacts with operations personnel through the LLM, specifically using natural language to facilitate this interaction. The human review node can be applied to reviewing tool invocation scenarios, where operations personnel review, edit, or approve tool invocations requested by the LLM before execution. It can also be applied to multi-turn dialogue scenarios, involving multiple back-and-forth interactions between the agent and the user, allowing the agent to collect information through dialogue, make informed decisions, and ultimately hand over the dialogue to another agent or other parts of the system.

[0167] Figure 9 This is a schematic diagram of a subgraph structure including a tool review node, provided for an embodiment of this application.

[0168] like Figure 9 As shown, the tool review node can be used to review, edit, or approve tool calls requested by the LLM before tool execution. Specifically, a human_review node can be added between the Agent and tools nodes. If the Agent node route returns "END", the session ends; otherwise, it proceeds to the human_review node. At this point, the session pauses and waits for user input. The user can perform several different actions: approve the tool call and continue; manually modify the tool call and continue, such as correcting or supplementing the tool call parameters; provide natural language feedback and then pass the feedback to the agent, rejecting the tool call.

[0169] S108: Use conditional edges to connect the human review node between the sub-agent node and the task executor and / or tool node.

[0170] Alternatively, the manual review node can be connected only to the sub-agent node; this embodiment of the application does not specifically limit this.

[0171] This enables human-machine collaboration.

[0172] Furthermore, the steps for creating the graph structure of the supervisor agent may specifically include steps S109-S111.

[0173] S109: Creating graph structures based on the LangGraph framework;

[0174] S110: Create a master agent node in the graph structure, and define a graph state object for the master agent node, which is used to update and view the graph state object.

[0175] It is understandable that the graph structure of the supervisory agent is a graph.

[0176] S111: Use edge components to connect the graph structure of the supervisor agent to each subgraph structure.

[0177] In this way, a graph representing the complete operation and maintenance process can be formed.

[0178] As can be seen, the core network operation and maintenance system construction method based on a large model provided in this application can decompose the core network operation and maintenance system into multiple smaller independent intelligent agents, combine these agents into a multi-agent system, and use a supervisory intelligent agent to manage the various task execution intelligent agents, achieving centralized scheduling. Furthermore, based on the method provided in this application, the intelligent agents can use LLM to determine the control flow, resulting in high operation and maintenance efficiency and accuracy.

[0179] It should also be noted that the embodiments of this application can also build a multi-turn dialogue architecture for the core network operation and maintenance system.

[0180] Figure 10 This is a schematic diagram of a multi-turn dialogue architecture provided in an embodiment of this application.

[0181] like Figure 10 As shown, a multi-turn dialogue architecture can involve multiple back-and-forth interactions between an agent and a user. This allows the agent to invoke the LLM (Local Management Model) to gather more information from the user in a conversational manner in order to make the right decisions, until the agent decides to hand the dialogue over to another agent or other part of the system. One or more agent nodes can engage in multi-turn dialogues with the user, where the user provides input or feedback at different stages of the dialogue. The agent node determines the next action based on the execution result: wait for user input to continue the dialogue; route the request to another agent (or return to itself, for example, in a loop).

[0182] In some implementations, the fault report generating agent can adopt the multi-turn dialogue mode described above. Users provide feedback on the generated fault report, and through multiple interactions with the large model, a high-quality fault analysis report is finally generated.

[0183] Figure 11 This is a schematic diagram of the core network operation and maintenance system based on a large model, provided in an embodiment of this application.

[0184] like Figure 11As shown, this application provides a core network operation and maintenance system based on a large model, which can also be called a core network operation and maintenance assistant based on a large model. This operation and maintenance system can be built based on the aforementioned core network operation and maintenance system construction method based on a large model.

[0185] In this embodiment of the application, the core network operation and maintenance system based on the large model includes at least a supervisory agent, namely... Figure 11 The operations and maintenance supervisor agent shown may also include multiple task execution agents, which may include network monitoring agents, fault detection and location agents, alarm handling agents, complaint handling agents, configuration management agents, network optimization agents, and knowledge base agents. The supervisor agent communicates with each task execution agent to schedule the task execution agents, with different task execution agents corresponding to different operations and maintenance tasks.

[0186] Furthermore, the supervisory agent is specifically configured to perform the following steps S600.

[0187] S600: In response to user commands issued by the user terminal to execute operation and maintenance tasks, it calls the LLM to parse the user commands to identify the task intent, and schedules one or more corresponding task execution agents based on the task intent. The task execution agents include at least one of the following agents: network monitoring agent, fault detection and location agent, alarm handling agent, complaint handling agent, configuration management agent, network optimization agent, data analysis agent, and knowledge base agent. The network monitoring agent monitors the operating status of the core network; the fault detection and location agent locates the cause and location of faults in the core network; the alarm handling agent processes alarm information in the core network; the complaint handling agent handles user complaints; the configuration management agent manages the core network configuration; the network optimization agent optimizes the core network; the data analysis agent analyzes the network data of the core network; and the knowledge base agent manages and updates the core network operation and maintenance knowledge base.

[0188] As the interface between humans (operations and maintenance personnel) and intelligent assistants, the supervisory intelligent agent can receive operations and maintenance instructions expressed by users in natural language. Then, the supervisory intelligent agent can use its large model language perception capabilities, combined with its own operations and maintenance expertise, to accurately analyze the operations and maintenance needs of operations and maintenance personnel, identify their intentions, and dispatch the operations and maintenance instructions to the relevant professional intelligent agents.

[0189] The task execution agent is specifically configured to perform the following steps S700.

[0190] S700: Responds to scheduling and invokes LLM to generate task execution flow, and executes operation and maintenance tasks based on the task execution flow.

[0191] In this way, multi-agent collaboration can be used to achieve operation and maintenance task perception, operation and maintenance task scheduling, operation and maintenance task execution, and return of execution results.

[0192] Furthermore, the supervisory agent is also configured to perform the following steps S800.

[0193] S800: Generates operation and maintenance task information based on task intent and preset operation and maintenance knowledge, and updates the operation and maintenance task information in the graph state object so as to pass the operation and maintenance task information to the task execution agent through the graph state object.

[0194] For example, the operation and maintenance task information can be the name of a network element device in the core network, or the operation and maintenance task information can be used to represent instruction information, such as: perform fault diagnosis on a certain network element.

[0195] Furthermore, the task execution agent is also configured to perform the following steps S901-S902.

[0196] S901: View the graph state object to obtain operation and maintenance task information. Based on the operation and maintenance task information and task execution process, call its internal nodes to execute the operation and maintenance task. The nodes include sub-agent nodes, task executors, tool nodes, end nodes and / or manual review nodes.

[0197] Understandably, the graph state object is constantly updated to track and manage the flow of messages or state with the agent.

[0198] S902: After the operation and maintenance task is completed, update the task execution result in the graph state and route it to the supervisor agent or user terminal.

[0199] It is understandable that routing to the supervisory agent means returning the control flow to the supervisory agent, while routing to the user means returning the control flow to the operations and maintenance personnel.

[0200] It should also be noted that, in the embodiments of this application, after the task execution agent obtains the operation and maintenance task information, if other agents need to cooperate, the control flow return value will be sent to the supervisor agent, which will then schedule the task again.

[0201] Furthermore, the task execution agent is also configured to perform the following steps S903.

[0202] S903: When the task execution flow includes routing to the end node, route to the end node to end the task execution flow.

[0203] And / or, the task execution agent is also configured to perform the following steps S904.

[0204] S904: When routing to a human review node is included in the task execution process, respond to the user input received by the human review node and route to a node within it, another task execution agent, or a supervisor agent based on the user input.

[0205] In this context, routing to a node within it, another task execution agent, or a supervisory agent refers to transferring control flow to a node within it or to another task execution agent or supervisory agent.

[0206] For example, in a tool call review scenario, user input could be approving the tool call and continuing, manually modifying the tool call and continuing, manually modifying such as correcting or supplementing the tool call parameters, or providing natural language feedback to refuse to execute the tool call.

[0207] In some implementations, such as Figure 11 As shown, the Assistant node represents a manual review node. Regular tools may not require review, while important tools must be reviewed. This application embodiment does not impose specific limitations on this.

[0208] Alternatively, in multi-turn dialogue scenarios, user input can be user feedback. For example, in response to a fault report generated by a task execution agent, the user provides feedback, and through multiple interactions with the LLM, a high-quality fault analysis report is ultimately generated.

[0209] It should be added that, Figure 11 The interactive process shown includes at least the following three aspects.

[0210] ① Interaction between operations and maintenance personnel and operations and maintenance supervisor intelligent agents: Operations and maintenance personnel send messages to operations and maintenance supervisor intelligent agents, which then process the messages and provide responses.

[0211] ② Task allocation by the Operations Supervisor: The Operations Supervisor assigns specific tasks to different task execution agents, such as network monitoring agents and fault detection and location agents, through delegation.

[0212] ③ Direct Interaction: There is also direct interaction between the task execution agent and the operation and maintenance personnel, which facilitates collaborative work.

[0213] In this embodiment, the network monitoring agent includes: a real-time network monitoring agent node, a data acquisition tool node, an anomaly detection tool node, an alarm triggering tool node, and a termination node. The real-time network monitoring agent node is configured to: receive real-time monitoring task instructions, plan monitoring tasks, and sequentially call the data acquisition tool node, anomaly detection tool node, and alarm triggering tool node to execute monitoring tasks. The monitoring tasks include data acquisition monitoring tasks, anomaly detection monitoring tasks, and alarm triggering monitoring tasks. The data acquisition tool node is configured to: integrate external data acquisition tools and execute data acquisition monitoring tasks. The anomaly detection tool node is configured to: integrate external anomaly detection tools and execute anomaly detection monitoring tasks. If the external anomaly detection tool detects an alarm event, it routes to the alarm triggering tool node; if the external anomaly detection tool does not detect an alarm event, it routes to the termination node. The alarm triggering tool node is configured to: integrate external alarm notification tools and execute alarm triggering monitoring tasks.

[0214] The fault detection and localization intelligent agent includes: a fault detection and localization intelligent agent node, a fault detection intelligent agent node, a fault delimitation and localization intelligent agent node, a fault report generation intelligent agent node, a fault detection task executor node, and a termination node. The fault detection and localization intelligent agent node is configured to: receive alarm information triggered by the network monitoring process and dispatch tasks to the fault detection intelligent agent node, the fault delimitation and localization intelligent agent node, and the fault report generation intelligent agent node. The fault detection intelligent agent node is configured to: plan the detection tasks according to business rules and output the detection tasks to the fault detection task executor node for execution; the detection task package... Includes at least one of the following tasks: device status check task, performance indicator check task, alarm check task, and chr log check task; the fault detection task executor node is configured to: execute the detection task cyclically; the fault delimitation and localization agent node is configured to: output delimitation and localization conclusions based on the fault investigation results and fault delimitation and localization rules, the delimitation and localization conclusions including the fault cause and handling suggestions; the fault report generation agent node is configured to: generate a fault analysis report based on the fault investigation results and delimitation and localization conclusions, the fault analysis report including at least the device's operating status, fault statistics, and performance trends.

[0215] The alarm handling intelligence agent includes: alarm handling intelligence agent nodes, alarm handling tool nodes, alarm information feedback intelligence agent nodes, alarm feedback tools, and termination nodes. The alarm handling intelligence agent nodes are configured to: receive alarm information and handling suggestions, and dispatch alarm tasks to alarm handling tool nodes. The alarm handling tool nodes are configured to: execute alarm tasks according to alarm categories and alarm levels. The alarm information feedback intelligence agent nodes are configured to: feed back the processing results and related information to maintenance personnel or relevant systems. The data analysis intelligence agent includes: data analysis intelligence agent nodes, device status query intelligence agent, device status query tool, performance indicator analysis intelligence agent, performance indicator analysis tool, alarm data query intelligence agent, alarm data query tool, log analysis intelligence agent, and daily... The system includes analysis tools and termination nodes. Data analysis agent nodes are configured to: detect user queries and dispatch tasks to device status query agent nodes, performance metric analysis agent nodes, alarm data query agent nodes, and / or log analysis agent nodes; device status query agent nodes are configured to: integrate external device status query tools and perform device status data queries; performance metric analysis agent nodes are configured to: integrate external performance metric analysis tools and perform performance metric data query analysis and metric prediction; alarm data query agent nodes are configured to: integrate external alarm data query tools and perform alarm data queries and statistics; and log analysis agent nodes are configured to: integrate external log data query tools and perform log data analysis.

[0216] The knowledge base agent includes: knowledge question answering agent nodes and end nodes.

[0217] As can be seen, the core network operation and maintenance system (also known as the core network operation and maintenance assistant) provided in this application embodiment can realize multi-agent communication and collaboration design. The core network operation and maintenance assistant is divided into multiple agents according to roles, including real-time network monitoring, fault detection and location, alarm handling, data analysis, knowledge base, etc. These agents perceive the environment, make decisions and communicate and collaborate with other agents to jointly complete core network operation and maintenance tasks, and achieve efficient and accurate management of core network operation and maintenance.

[0218] In summary, the core network operation and maintenance system construction method and the core network operation and maintenance system based on a large model provided in this application, by integrating large model and LangGraph technology, construct an intelligent operation and maintenance framework, improving operation and maintenance efficiency and accuracy. Specifically, it can achieve: ① using large model and LangGraph to realize operation and maintenance requirement parsing and automated task execution; ② performing intelligent analysis based on multimodal data (operation and maintenance knowledge) to support decision-making; ③ optimizing operation and maintenance task execution through real-time monitoring and dynamic adjustment; ④ in a multi-agent system, agent collaboration and task scheduling improve work efficiency, realizing intelligent and automated operation and maintenance.

[0219] In summary, this application provides a method for constructing a core network operation and maintenance system based on a large model, and a core network operation and maintenance system based thereon. This method achieves intelligent and automated core network operation and maintenance work by constructing a LangGraph graph structure with a large language model and external tool integration. It features modularity, specialization, and controllability. Modularity refers to dividing the system into multiple intelligent agents, making it easier to develop, test, and maintain these agents. Specialization refers to creating intelligent agents focused on specific domains, which helps improve overall system performance. Controllability refers to the ability to explicitly control the communication methods of the intelligent agents.

[0220] This application constructs an intelligent operation and maintenance framework for the core network by integrating large-scale models and LangGraph technology. This framework leverages the powerful language understanding and generation capabilities of large-scale models to accurately analyze user operation and maintenance needs, and through the intelligent agent function of the LangGraph framework, it achieves automated execution and real-time monitoring of operation and maintenance tasks. This not only improves operation and maintenance efficiency but also significantly enhances the accuracy and reliability of operation and maintenance. Furthermore, this application can perform intelligent analysis based on multi-dimensional data. Addressing the complexity of core network operation and maintenance data, this application can utilize large-scale models to perform fusion analysis of various types of data, such as network performance indicators, logs, and alarm events, providing a scientific basis for operation and maintenance decisions. Simultaneously, this application combines the planning capabilities of the LangGraph framework to optimize operation and maintenance strategies and improve operation and maintenance effectiveness. Further, this application can adaptively adjust operation and maintenance tasks dynamically. During the operation and maintenance process, this application monitors the execution status of operation and maintenance tasks and the system status in real time, and uses the dynamic adjustment function of the LangGraph framework to adaptively adjust the operation and maintenance tasks. This adaptive adjustment mechanism ensures efficient execution of operation and maintenance tasks while reducing operation and maintenance costs and improving efficiency. Furthermore, this application provides a multi-agent communication and collaboration mechanism. Core network operation and maintenance assistant agents exchange information and share resources through a shared graph state, using standardized communication protocols and data formats to ensure accurate information transmission and parsing. A task scheduling agent (which can be integrated with the supervisor agent as a single agent) assigns tasks to each agent based on the needs of the operation and maintenance tasks and the agents' capabilities. Each agent then collaborates according to its assigned tasks to jointly achieve the task objectives. In summary, the method provided in this application can realize the intelligent and automated operation and maintenance of the core network, improving operational efficiency and accuracy.

[0221] Figure 12 A schematic diagram of the structure of the core network operation and maintenance system construction device based on a large model provided in the embodiments of this application.

[0222] like Figure 12 As shown in the figure, this application embodiment provides a core network operation and maintenance system construction device based on a large model. The device may include:

[0223] The first construction module 100 is used for core network operation and maintenance tasks. It creates a subgraph structure of multiple task execution agents using the LangGraph framework, and a graph structure of a supervisor agent using the LangGraph framework. The supervisor agent communicates with each task execution agent to schedule them, and different task execution agents correspond to different operation and maintenance tasks. The task execution agents include at least one of the following agents: network monitoring agent, fault detection and location agent, alarm handling agent, complaint handling agent, configuration management agent, network optimization agent, data analysis agent, and knowledge base agent. The network monitoring agent monitors the core network's operating status; the fault detection and location agent locates the cause and location of faults in the core network; the alarm handling agent processes alarm information in the core network; the complaint handling agent processes user complaints; the configuration management agent manages the core network's configuration; the network optimization agent optimizes the core network; the data analysis agent analyzes the core network's network data; and the knowledge base agent manages and updates the core network operation and maintenance knowledge base.

[0224] The relationship establishment module 200 is used to establish the calling relationship between the graph structure and / or subgraph structure and the large language model LLM;

[0225] Compiler module 300 is used to compile graph structures and subgraph structures to obtain graph instances;

[0226] The second construction module 400 is used to construct a core network operation and maintenance system based on graph instances. The graph instances define the operation flow of the core network operation and maintenance system. When the core network operation and maintenance system is running, the graph instances run to execute the operation flow. The operation flow includes at least the following: when a user terminal issues a user instruction to the core network operation and maintenance system to execute an operation and maintenance task, the supervisor agent responds to the user instruction, calls the LLM to parse the user instruction to identify the task intent, and schedules one or more corresponding task execution agents based on the task intent; and the operation flow also includes: the task execution agent responds to the scheduling and calls the LLM to generate a task execution flow, and executes the operation and maintenance task based on the task execution flow.

[0227] In some implementations, the first building module 100 is specifically used for: creating a subgraph structure based on the LangGraph framework for each task execution agent; creating one or more child agent nodes in the subgraph structure, and defining a graph state object for the child agent node, which is used to update and view the graph state object, the number of child agent nodes being determined based on the task execution flow corresponding to the task execution agent; and / or creating task executor and / or tool nodes in the subgraph structure, the number and type of task executor and tool nodes being determined based on the operation and maintenance task corresponding to the task execution agent; and creating an end node in the subgraph structure.

[0228] In some implementations, the first building module 100 is specifically used to: set an entry point for each subgraph structure, the entry point being a sub-agent node, the entry point representing the starting point for the task execution agent to execute the operation and maintenance task; and connect the sub-agent node, task executor, tool node, and / or end node using edge components, the edge components including conditional edges and unconditional edges, the conditional edges having routing conditions, the routing conditions including at least that after the sub-agent node calls the LLM, if the LLM's return message includes a tool call field, then the route is to the task executor and / or tool node, if the return message does not include a tool call field, then the route is to the end node.

[0229] In some implementations, the first building module 100 is also used to: create human review nodes in the subgraph structure; and connect the human review nodes between the sub-agent nodes and the task executor and / or tool nodes using conditional edges.

[0230] In some implementations, the first building module 100 is also used to: create a graph structure based on the LangGraph framework; create a master agent node in the graph structure, and define a graph state object for the master agent node, which is used to update and view the graph state object; connect the master agent's graph structure with each subgraph structure using edge components; and the first building module 100 is also used to: create an LLM and an external toolset, the external toolset including multiple callable tools, the tools being determined based on the operation and maintenance tasks; and bind each tool in the external toolset to the LLM through the LLM's interface.

[0231] In some implementations, the first construction module 100 is further configured to: create network real-time monitoring agent nodes, data acquisition tool nodes, anomaly detection tool nodes, alarm triggering tool nodes, and termination nodes in the subgraph structure of the network monitoring agent; draw unconditional edges from the network real-time monitoring agent nodes pointing to the data acquisition tool nodes, draw unconditional edges from the data acquisition tool nodes pointing to the anomaly detection tool nodes, draw conditional edges from the anomaly detection tool nodes pointing to the alarm triggering tool nodes and termination nodes respectively, and draw unconditional edges from the alarm triggering tool nodes pointing to the termination nodes.

[0232] In some implementations, the first construction module 100 is further configured to: create a fault detection and localization agent node, a fault detection agent node, a fault delimitation and localization agent node, a fault report generation agent node, a fault detection task executor node, and an end node in the subgraph structure of the fault detection and localization agent; extend conditional edges from the fault detection and localization agent node to the fault detection agent node, the fault delimitation and localization agent node, the fault report generation agent node, and the end node, respectively; extend unconditional edges from the fault detection agent node, the fault delimitation and localization agent node, and the fault report generation agent node to the fault detection and localization agent node, respectively; extend unconditional edges from the fault detection agent node to the fault detection task executor node; and extend unconditional edges from the fault detection task executor node to the fault detection agent node.

[0233] In some implementations, the first construction module 100 is further configured to: create an alarm processing agent node, an alarm processing tool node, an alarm information feedback agent node, an alarm feedback tool node, and an end node in the subgraph structure of the alarm processing agent; extend conditional edges from the alarm processing agent node to the alarm processing tool node, the alarm information feedback agent node, and the end node, respectively; extend unconditional edges from the alarm processing agent node and the alarm information feedback agent node to the alarm processing agent node, respectively; extend conditional edges from the alarm information feedback agent node to the alarm feedback tool node, and extend unconditional edges from the alarm feedback tool node to the alarm information feedback agent node.

[0234] In some implementations, the first construction module 100 is further configured to: create, in the subgraph structure of the data analysis agent, a data analysis agent node, a device status query agent node, a device status query tool node, a performance index analysis agent node, a performance index analysis tool node, an alarm data query agent node, an alarm data query tool node, a log analysis agent node, a log analysis tool node, and an end node; conditional edges pointing to the status query agent node, performance index analysis agent node, alarm data query agent node, log analysis agent node, and end node are respectively derived from the data analysis agent node; and conditional edges pointing to the data analysis agent node are respectively derived from the device status query agent node, performance index analysis agent node, alarm data query agent node, and log analysis agent node. Unconditional edges; and conditional edges leading from the device status query agent node to the device status query tool node; and unconditional edges leading from the device status query tool node to the device status query agent node; and conditional edges leading from the performance metric analysis agent node to the performance metric analysis tool node; and unconditional edges leading from the performance metric analysis tool node to the performance metric analysis agent node; and conditional edges leading from the alarm data query agent node to the alarm data query tool node; and unconditional edges leading from the alarm data query tool node to the alarm data query agent node; and conditional edges leading from the log analysis agent node to the log analysis tool node; and unconditional edges leading from the log analysis tool node to the log analysis agent node.

[0235] In some implementations, the first building module 100 is also used to: create knowledge question-answering agent nodes and end nodes in the subgraph structure of the knowledge base agent; and draw unconditional edges from the knowledge question-answering agent nodes to the end nodes.

[0236] This application embodiment also provides a core network operation and maintenance method based on a large model, including the following steps S1001-S1002:

[0237] S1001: The supervisor agent responds to the user command issued by the user terminal to execute the operation and maintenance task, calls the LLM to parse the user command to identify the task intent, and schedules one or more corresponding task execution agents based on the task intent;

[0238] S1002: The task execution agent responds to the schedule and calls the LLM to generate a task execution process, and executes operation and maintenance tasks based on the task execution process.

[0239] The supervisor agent communicates with each task execution agent, and different task execution agents correspond to different operation and maintenance tasks.

[0240] In some implementations, step S1002 may specifically include the following steps.

[0241] S1002-1: The supervisor agent generates operation and maintenance task information based on the task intent and preset operation and maintenance knowledge, and updates the operation and maintenance task information in the graph state object so as to pass the operation and maintenance task information to the task execution agent through the graph state object.

[0242] S1002-2: The task execution agent views the graph state State object to obtain operation and maintenance task information. Based on the operation and maintenance task information and task execution process, it calls the internal nodes to execute the operation and maintenance task. The nodes include sub-agent nodes, task executors, tool nodes, end nodes and / or manual review nodes.

[0243] S1002-3: After the task execution is completed, the task execution agent updates the task execution result in the graph state and routes it to the supervisor agent or the user terminal.

[0244] In some implementations, step S1002 may also include the following steps.

[0245] S1002-4: When the task execution agent includes routing to the end node in the task execution process, it routes to the end node to end the task execution process;

[0246] And / or, S1002-5: When the task execution agent is routed to a human review node in the task execution process, it responds to the user input received by the human review node and routes to a node within it, another task execution agent, or a supervisor agent based on the user input.

[0247] In some implementations, the network monitoring agent includes: a real-time network monitoring agent node, a data acquisition tool node, an anomaly detection tool node, an alarm triggering tool node, and a termination node. The real-time network monitoring agent node is configured to: receive real-time monitoring task instructions, plan monitoring tasks, and sequentially call the data acquisition tool node, anomaly detection tool node, and alarm triggering tool node to execute monitoring tasks. These monitoring tasks include data acquisition monitoring tasks, anomaly detection monitoring tasks, and alarm triggering monitoring tasks. The data acquisition tool node is configured to: integrate external data acquisition tools and execute data acquisition monitoring tasks. The anomaly detection tool node is configured to: integrate external anomaly detection tools and execute anomaly detection monitoring tasks. If the external anomaly detection tool detects an alarm event, it routes to the alarm triggering tool node; if the external anomaly detection tool does not detect an alarm event, it routes to the termination node. The alarm triggering tool node is configured to: integrate external alarm notification tools and execute alarm triggering monitoring tasks.

[0248] The fault detection and localization intelligent agent includes: a fault detection and localization intelligent agent node, a fault detection intelligent agent node, a fault delimitation and localization intelligent agent node, a fault report generation intelligent agent node, a fault detection task executor node, and a termination node. The fault detection and localization intelligent agent node is configured to: receive alarm information triggered by the network monitoring process and dispatch tasks to the fault detection intelligent agent node, the fault delimitation and localization intelligent agent node, and the fault report generation intelligent agent node. The fault detection intelligent agent node is configured to: plan the detection tasks according to business rules and output the detection tasks to the fault detection task executor node for execution; the detection task package... Includes at least one of the following tasks: device status check task, performance indicator check task, alarm check task, and chr log check task; the fault detection task executor node is configured to: execute the detection task cyclically; the fault delimitation and localization agent node is configured to: output delimitation and localization conclusions based on the fault investigation results and fault delimitation and localization rules, the delimitation and localization conclusions including the fault cause and handling suggestions; the fault report generation agent node is configured to: generate a fault analysis report based on the fault investigation results and delimitation and localization conclusions, the fault analysis report including at least the device's operating status, fault statistics, and performance trends.

[0249] The alarm handling intelligence agent includes: alarm handling intelligence agent nodes, alarm handling tool nodes, alarm information feedback intelligence agent nodes, alarm feedback tools, and termination nodes. The alarm handling intelligence agent nodes are configured to: receive alarm information and handling suggestions, and dispatch alarm tasks to alarm handling tool nodes. The alarm handling tool nodes are configured to: execute alarm tasks according to alarm categories and alarm levels. The alarm information feedback intelligence agent nodes are configured to: feed back the processing results and related information to maintenance personnel or relevant systems. The data analysis intelligence agent includes: data analysis intelligence agent nodes, device status query intelligence agent, device status query tool, performance indicator analysis intelligence agent, performance indicator analysis tool, alarm data query intelligence agent, alarm data query tool, log analysis intelligence agent, and daily... The system includes analysis tools and termination nodes. Data analysis agent nodes are configured to: detect user queries and dispatch tasks to device status query agent nodes, performance metric analysis agent nodes, alarm data query agent nodes, and / or log analysis agent nodes; device status query agent nodes are configured to: integrate external device status query tools and perform device status data queries; performance metric analysis agent nodes are configured to: integrate external performance metric analysis tools and perform performance metric data query analysis and metric prediction; alarm data query agent nodes are configured to: integrate external alarm data query tools and perform alarm data queries and statistics; and log analysis agent nodes are configured to: integrate external log data query tools and perform log data analysis.

[0250] The knowledge base agent includes: knowledge question answering agent nodes and end nodes.

[0251] In a specific implementation, the present invention also provides a computer storage medium, wherein the computer storage medium may store a program, which, when executed, may include some or all of the steps in the various embodiments of the core network operation and maintenance system construction method based on a large model and / or the core network operation and maintenance method based on a large model provided by the present invention. The storage medium may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0252] It is readily understood that, based on the several embodiments provided in this application, those skilled in the art can combine, split, or reorganize the embodiments of this application to obtain other embodiments, none of which exceed the protection scope of this application.

[0253] The above detailed embodiments further illustrate the purpose, technical solution, and beneficial effects of the embodiments of this application. It should be understood that the above are merely specific embodiments of the embodiments of this application and are not intended to limit the protection scope of the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solutions of the embodiments of this application should be included within the protection scope of the embodiments of this application.

Claims

1. A method for constructing a core network operation and maintenance system based on a large model, characterized in that, The method includes: Based on the core network's operation and maintenance tasks, a subgraph structure of multiple task execution agents is created using the LangGraph framework, and a graph structure of a supervisor agent is also created using the LangGraph framework. The supervisor agent communicates with each task execution agent to schedule them, and different task execution agents correspond to different operation and maintenance tasks. Each task execution agent includes at least one of the following agents: a network monitoring agent, a fault detection and location agent, an alarm handling agent, a complaint handling agent, a configuration management agent, a network optimization agent, a data analysis agent, and a knowledge base agent. The network monitoring agent monitors the core network's operational status; the fault detection and location agent locates the cause and location of faults in the core network; the alarm handling agent processes alarm information in the core network; the complaint handling agent processes user complaints; the configuration management agent manages the core network's configuration; the network optimization agent optimizes the core network; the data analysis agent analyzes the core network's network data; and the knowledge base agent manages and updates the core network's operation and maintenance knowledge base. Establish the calling relationship between the graph structure and / or the subgraph structure and the Large Language Model (LLM); The graph structure and the subgraph structure are compiled to obtain a graph instance; The core network operation and maintenance system is constructed based on the graph instance; wherein, the graph instance is used to define the operation flow of the core network operation and maintenance system, and when the core network operation and maintenance system is running, the graph instance runs to execute the operation flow; the operation flow includes at least: when the user terminal issues a user instruction to the core network operation and maintenance system to execute the operation and maintenance task, the supervisor agent responds to the user instruction, calls the LLM to parse the user instruction to identify the task intent, and schedules one or more corresponding task execution agents based on the task intent; and the operation flow further includes: the task execution agent responds to the scheduling and calls the LLM to generate a task execution flow, and executes the operation and maintenance task based on the task execution flow.

2. The method for constructing a core network operation and maintenance system based on a large model according to claim 1, characterized in that, The steps for creating a subgraph structure with multiple task execution agents using the LangGraph framework include: For each of the aforementioned tasks, an intelligent agent creates the subgraph structure based on the LangGraph framework; One or more child agent nodes are created in the subgraph structure, and a graph state object is defined for the child agent node. The child agent node is used to update and view the graph state object. The number of child agent nodes is determined based on the task execution flow corresponding to the task execution agent. And / or, create task executors and / or tool nodes in the subgraph structure, wherein the number and type of task executors and tool nodes are determined based on the operation and maintenance task corresponding to the task execution agent; Create an end node in the subgraph structure.

3. The method for constructing a core network operation and maintenance system based on a large model according to claim 2, characterized in that, The steps for creating a subgraph structure with multiple task execution agents using the LangGraph framework also include: For each of the subgraph structures, an entry point is set, which is the sub-agent node, and the entry point represents the starting point for the task execution agent to execute the operation and maintenance task; The sub-agent node, the task executor, the tool node, and / or the end node are connected using an edge component. The edge component includes conditional edges and unconditional edges. The conditional edges have routing conditions. The routing conditions include at least the following: after the sub-agent node calls the LLM, if the LLM's return message includes a tool call field, then the route is to the task executor and / or the tool node; if the return message does not include a tool call field, then the route is to the end node.

4. The method for constructing a core network operation and maintenance system based on a large model according to claim 3, characterized in that, After the step of creating task executors and / or tool nodes in the subgraph structure, the method further includes: Create a manual review node in the subgraph structure; The conditional edge is used to connect the human review node between the sub-agent node and the task executor and / or the tool node.

5. The method for constructing a core network operation and maintenance system based on a large model according to claim 3, characterized in that, The steps to create a graph structure for a supervisor agent include: The graph structure is created based on the LangGraph framework; In the graph structure, a master agent node is created, and a graph state object is defined for the master agent node, which is used to update and view the graph state object; The edge component is used to connect the graph structure of the supervisory agent to each of the subgraph structures. Furthermore, the method further includes: Create the LLM and external toolset, which includes multiple callable tools, determined based on the operation and maintenance task; The tools in the external toolset are bound to the LLM through the LLM's interface.

6. The method for constructing a core network operation and maintenance system based on a large model according to claim 2, characterized in that, The steps for creating the subgraph structure of the network monitoring agent include: In the subgraph structure of the network monitoring agent, create a network real-time monitoring agent node, a data acquisition tool node, an anomaly detection tool node, an alarm triggering tool node, and an end node. An unconditional edge is drawn from the network real-time monitoring intelligent agent node to the data acquisition tool node, and an unconditional edge is drawn from the data acquisition tool node to the anomaly detection tool node, and a conditional edge is drawn from the anomaly detection tool node to the alarm triggering tool node and the termination node, respectively, and an unconditional edge is drawn from the alarm triggering tool node to the termination node. The steps for creating the subgraph structure of the fault detection and localization agent include: In the subgraph structure of the fault detection and localization agent, create a fault detection and localization agent node, a fault detection agent node, a fault delimitation and localization agent node, a fault report generation agent node, a fault detection task executor node, and an end node. Conditional edges are drawn from the fault detection and localization agent node, pointing to the fault detection agent node, the fault delimitation and localization agent node, the fault report generation agent node, and the termination node, respectively; unconditional edges are drawn from the fault detection agent node, the fault delimitation and localization agent node, and the fault report generation agent node, respectively, pointing to the fault detection and localization agent node; unconditional edges are drawn from the fault detection agent node, pointing to the fault detection task executor node; and unconditional edges are drawn from the fault detection task executor node, pointing to the fault detection agent node. The steps for creating the subgraph structure of the alarm handling agent include: In the subgraph structure of the alarm processing agent, create an alarm processing agent node, an alarm processing tool node, an alarm information feedback agent node, an alarm feedback tool node, and an end node; Conditional edges are drawn from the alarm processing agent node, pointing to the alarm processing tool node, the alarm information feedback agent node, and the termination node, respectively; and unconditional edges are drawn from the alarm processing agent node and the alarm information feedback agent node, respectively, pointing to the alarm processing agent node; and conditional edges are drawn from the alarm information feedback agent node, pointing to the alarm feedback tool node; and unconditional edges are drawn from the alarm feedback tool node, pointing to the alarm information feedback agent node. The steps for creating the subgraph structure of the data analysis agent include: In the subgraph structure of the data analysis agent, create a data analysis agent node, a device status query agent node, a device status query tool node, a performance index analysis agent node, a performance index analysis tool node, an alarm data query agent node, an alarm data query tool node, a log analysis agent node, a log analysis tool node, and an end node. Conditional edges originating from the data analysis agent node, pointing to the status query agent node, the performance indicator analysis agent node, the alarm data query agent node, the log analysis agent node, and the termination node, respectively; and unconditional edges originating from the device status query agent node, the performance indicator analysis agent node, the alarm data query agent node, and the log analysis agent node, respectively, pointing to the data analysis agent node; and conditional edges originating from the device status query agent node, pointing to the device status query tool node; and conditions originating from the device status query tool node, pointing to the device status query agent node. The following are examples of edges: unconditional edges; conditional edges leading from the performance index analysis agent node to the performance index analysis tool node; unconditional edges leading from the performance index analysis tool node to the performance index analysis agent node; conditional edges leading from the alarm data query agent node to the alarm data query tool node; unconditional edges leading from the alarm data query tool node to the alarm data query agent node; conditional edges leading from the log analysis agent node to the log analysis tool node; and unconditional edges leading from the log analysis tool node to the log analysis agent node. The steps for creating the knowledge base agent include: Create a knowledge question-answering agent node and an end node in the subgraph structure of the knowledge base agent; An unconditional edge is drawn from the knowledge-answering agent node, pointing to the end node.

7. A core network operation and maintenance system based on a large model, characterized in that, The system is constructed based on the core network operation and maintenance system construction method based on a large model as described in any one of claims 1-6. The system includes at least a supervisory agent and a task execution agent. The supervisory agent is used to communicate with each of the task execution agents to schedule the task execution agents. Different task execution agents correspond to different operation and maintenance tasks. The supervisor agent is configured to: respond to a user instruction issued by the user terminal to execute the operation and maintenance task, invoke the LLM to parse the user instruction to identify the task intent, and schedule one or more corresponding task execution agents based on the task intent; The task execution intelligence agent includes at least one of the following intelligence agents: network monitoring intelligence agent, fault detection and location intelligence agent, alarm handling intelligence agent, complaint handling intelligence agent, configuration management intelligence agent, network optimization intelligence agent, data analysis intelligence agent, and knowledge base intelligence agent. The network monitoring intelligence agent is used to monitor the operating status of the core network; the fault detection and location intelligence agent is used to locate the cause and location of faults in the core network; the alarm handling intelligence agent is used to process alarm information in the core network; the complaint handling intelligence agent is used to process user complaints; the configuration management intelligence agent is used to configure and manage the core network; the network optimization intelligence agent is used to optimize the core network; the data analysis intelligence agent is used to analyze network data of the core network; and the knowledge base intelligence agent is used to manage and update the core network operation and maintenance knowledge base. The task execution agent is configured to: respond to scheduling and invoke the LLM to generate a task execution process, and execute the operation and maintenance task based on the task execution process.

8. The core network operation and maintenance system based on a large model according to claim 7, characterized in that, The supervisory agent is also configured to: Based on the task intent and preset operation and maintenance knowledge, operation and maintenance task information is generated, and the operation and maintenance task information is updated in the graph state object so as to pass the operation and maintenance task information to the task execution agent through the graph state object; The task execution agent is also configured to: view the graph state object to obtain the operation and maintenance task information, and based on the operation and maintenance task information and the task execution process, call the internal nodes to execute the operation and maintenance task. The nodes include sub-agent nodes, task executors, tool nodes, end nodes and / or manual review nodes. After the operation and maintenance task is completed, the task execution result is updated in the graph state and routed to the supervisor agent or the user terminal. The task execution agent is also configured as follows: When the task execution flow includes routing to the end node, the task is routed to the end node to terminate the task execution flow. And / or, when the task execution process includes routing to the human review node, in response to user input received by the human review node, routing to the node within it, another task execution agent, or the supervisor agent based on the user input.

9. The core network operation and maintenance system based on a large model according to claim 8, characterized in that, The network monitoring agent includes: a real-time network monitoring agent node, a data acquisition tool node, an anomaly detection tool node, an alarm triggering tool node, and a termination node. The real-time network monitoring agent node is configured to: receive real-time monitoring task instructions, plan monitoring tasks, and sequentially call the data acquisition tool node, the anomaly detection tool node, and the alarm triggering tool node to execute the monitoring tasks. The monitoring tasks include data acquisition monitoring tasks, anomaly detection monitoring tasks, and alarm triggering monitoring tasks. The data acquisition tool node is configured to: integrate external data acquisition tools and execute the data acquisition monitoring tasks. The anomaly detection tool node is configured to: integrate external anomaly detection tools and execute the anomaly detection monitoring tasks. If the external anomaly detection tool detects an alarm event, the route is routed to the alarm triggering tool node; if the external anomaly detection tool does not detect the alarm event, the route is routed to the termination node. The alarm triggering tool node is configured to: integrate external alarm notification tools and execute the alarm triggering monitoring tasks. The fault detection and localization intelligent agent includes: a fault detection and localization intelligent agent node, a fault detection intelligent agent node, a fault delimitation and localization intelligent agent node, a fault report generation intelligent agent node, a fault detection task executor node, and a termination node; the fault detection and localization intelligent agent node is configured to: receive alarm information triggered by the network monitoring process, and dispatch tasks to the fault detection intelligent agent node, the fault delimitation and localization intelligent agent node, and the fault report generation intelligent agent node; the fault detection intelligent agent node is configured to: plan the detection task according to business rules, and output the detection task to the fault detection task executor node for execution; the detection task... The system includes at least one of the following tasks: device status check, performance indicator check, alarm check, and chr log check; the fault detection task executor node is configured to: cyclically execute the detection task; the fault delimitation and localization agent node is configured to: output a delimitation and localization conclusion based on the fault investigation results and fault delimitation and localization rules, the delimitation and localization conclusion including the fault cause and handling suggestions; the fault report generation agent node is configured to: generate a fault analysis report based on the fault investigation results and the delimitation and localization conclusion, the fault analysis report including at least the device's operating status, fault statistics, and performance trends. The alarm processing intelligence agent includes: an alarm processing intelligence agent node, an alarm processing tool node, an alarm information feedback intelligence agent node, an alarm feedback tool, and a termination node; the alarm processing intelligence agent node is configured to: receive alarm information and processing suggestions, and dispatch alarm tasks to the alarm processing tool node; the alarm processing tool node is configured to: execute the alarm task according to the alarm category and alarm level; the alarm information feedback intelligence agent node is configured to: feed back the processing results and related information to maintenance personnel or relevant systems; the data analysis intelligence agent includes: a data analysis intelligence agent node, a device status query intelligence agent, a device status query tool, a performance indicator analysis intelligence agent, a performance indicator analysis tool, an alarm data query intelligence agent, an alarm data query tool, a log analysis intelligence agent, and a log analysis... The data analysis agent node is configured to: sense user queries and dispatch tasks to the device status query agent node, the performance indicator analysis agent node, the alarm data query agent node, and / or the log analysis agent node; the device status query agent node is configured to: integrate external tools for device status queries and perform device status data queries; the performance indicator analysis agent node is configured to: integrate external tools for performance indicator analysis and perform performance indicator data query analysis and indicator prediction; the alarm data query agent node is configured to: integrate external tools for alarm data queries and perform alarm data queries and statistics; the log analysis agent node is configured to: integrate external tools for log data queries and perform log data analysis. The knowledge base agent includes: a knowledge question-answering agent node and an end node.

10. A device for constructing a core network operation and maintenance system based on a large model, characterized in that, The device includes: The first construction module is used for operation and maintenance tasks based on the core network. It creates a subgraph structure with multiple task execution agents using the LangGraph framework, and a graph structure with a supervisor agent using the LangGraph framework. The supervisor agent communicates with each task execution agent to schedule them, and different task execution agents correspond to different operation and maintenance tasks. The task execution agents include at least one of the following agents: a network monitoring agent, a fault detection and location agent, an alarm handling agent, a complaint handling agent, a configuration management agent, a network optimization agent, a data analysis agent, and a knowledge base agent. The network monitoring agent monitors the operating status of the core network; the fault detection and location agent locates the cause and location of faults in the core network; the alarm handling agent processes alarm information in the core network; the complaint handling agent processes user complaints; the configuration management agent manages the configuration of the core network; the network optimization agent optimizes the core network; the data analysis agent analyzes network data in the core network; and the knowledge base agent manages and updates the core network operation and maintenance knowledge base. The relationship establishment module is used to establish the calling relationship between the graph structure and / or the subgraph structure and the large language model LLM; A compilation module is used to compile the graph structure and the subgraph structure to obtain a graph instance; The second construction module is used to construct the core network operation and maintenance system based on the graph instance; wherein, the graph instance is used to define the operation flow of the core network operation and maintenance system, and when the core network operation and maintenance system is running, the graph instance runs to execute the operation flow; the operation flow includes at least: when the user terminal issues a user instruction to the core network operation and maintenance system to execute the operation and maintenance task, the supervisor agent responds to the user instruction, calls the LLM to parse the user instruction to identify the task intent, and schedules one or more corresponding task execution agents based on the task intent; and the operation flow further includes: the task execution agent responds to the scheduling and calls the LLM to generate a task execution flow, and executes the operation and maintenance task based on the task execution flow.

Images