A multi-agent based mobile core network operation and maintenance management method and device

Abstract

The application provides a kind of mobile core network operation and maintenance management method and device based on multi-agent, applied to communication technology and artificial intelligence cross technical field.The method comprises: receiving the first dialogue message input by operation and maintenance personnel;First dialogue information is received by first agent, for representing the first operation and maintenance intention for mobile network;Based on the first operation and maintenance intention, determine the second agent;The second agent sends access core network request to mobile core network;Based on the query result returned by mobile core network, the second agent returns the result corresponding to the first operation and maintenance intention.Such, it solves the problem that operation and maintenance is complex and professional threshold is high, operation and maintenance personnel not only master mobile core network architecture and each network element function, but also memorize a large number of special query commands, and the learning cost is huge.Secondly, directly communicate with natural language, improve efficiency.Enhance system flexibility and adaptability, can cope with complex and changeable combination query scene.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of communication technology and artificial intelligence, and in particular to a method and apparatus for operation and maintenance management of mobile core network based on multi-agent technology. Background Technology

[0002] Therefore, there is an urgent need for a method to enhance the cross-linguistic capabilities of large language models to address the shortcomings in their processing capabilities for low-resource languages. Summary of the Invention

[0003] This invention describes a method and apparatus for operation and maintenance management of mobile core network based on multi-agent systems, which can solve the above-mentioned technical problems.

[0004] According to a first aspect, a multi-agent-based mobile core network operation and maintenance management method is provided. The method includes: receiving a first dialogue message input by an operation and maintenance personnel; the first dialogue message being received by a first intelligent agent and used to represent a first operation and maintenance intent for the mobile network; determining a second intelligent agent based on the first operation and maintenance intent, wherein the second intelligent agent is determined through reasoning using a large language model; the second intelligent agent sending a request to access the core network to the mobile core network; based on the query result returned by the mobile core network, the second intelligent agent returning a result corresponding to the first operation and maintenance intent; the result corresponding to the first operation and maintenance intent being obtained through reasoning using the large language model.

[0005] In some embodiments, the method further includes: receiving a second dialogue message input by an operations and maintenance (O&M) personnel; the second dialogue message is received by a first intelligent agent and is used to characterize a second O&M intent for the mobile network; determining a third intelligent agent based on the second dialogue message, wherein the third intelligent agent is determined through reasoning using a large language model; sending a request to the second intelligent agent based on the second O&M intent; scoring the target network element based on the target network element query result returned by the second intelligent agent; and returning a result corresponding to the second O&M intent based on the scoring data; wherein the result corresponding to the second O&M intent is obtained through reasoning using the large language model.

[0006] In some embodiments, the method further includes: receiving a third dialogue message input by an operations and maintenance (O&M) personnel; the third dialogue message being received by a first intelligent agent and used to characterize a third O&M intent for the mobile network; determining a fourth intelligent agent based on the third dialogue message, the fourth intelligent agent being determined through reasoning using a large language model; the fourth intelligent agent sending a first query request to a second intelligent agent; obtaining a query result returned by the second intelligent agent; based on the query result, the fourth intelligent agent inputting the features of the target network element into a second and / or a third large language model for processing, obtaining a first evaluation result and / or a second evaluation prediction result for the target network element; based on the first evaluation result and / or the second evaluation prediction result, the fourth intelligent agent returning a result corresponding to the third O&M intent.

[0007] In some embodiments, the fourth agent includes a first machine learning model and / or a second machine learning model, wherein the first machine learning model is a classification model and the second machine learning model is a classification model; based on the first machine learning model, it outputs numerical point estimates of future key performance indicators, indicating the specific degree to which the fault may worsen; based on the second machine learning model, it outputs the probability of an anomaly occurring within a future time interval, indicating whether there is an overall risk of the fault occurring; based on the output results of the first machine learning model and / or the second machine learning model, it obtains numerical predictions and / or probability predictions.

[0008] In some embodiments, data preparation for training the first machine learning model and / or the second machine learning model includes: collecting historical operational data, including time-series performance index data of the target network element and alarm log data aligned with it in time; the time-series performance index data of the target network element includes CPU utilization, memory utilization, number of sessions, and bandwidth usage; the data is aligned according to a uniform time granularity; the historical data is processed in a sliding time window manner, and for each historical moment, a multi-dimensional performance index sequence within a preset time range before the historical moment is extracted to construct a feature base reflecting the operational trend.

[0009] In some embodiments, training the first machine learning model and / or the second machine learning model includes: for the first machine learning model, constructing a supervision label as the true key performance indicator value after the prediction step size at the current time; for the second machine learning model, constructing a supervision label as the binary classification result of whether a target abnormal event occurs within a preset time interval in the future; dividing the training set and the test set according to the time order, and training the first machine learning model and / or the second machine learning model.

[0010] In some embodiments, constructing input features for the first model and / or the second model includes: calculating multiple time-series statistical features for key performance indicators, wherein the key performance indicators include statistical values ​​of the most recent time window, overall fluctuation level, changing trends between adjacent windows, extreme value features, and positional features relative to the historical distribution of the same period, to characterize the current state and changing trends; classifying and statistically analyzing alarm data within the same time window according to type and severity to generate alarm count class and anomaly density class features; and concatenating the multiple time-series statistical features and the alarm count class and anomaly density class features to obtain an input feature vector. In the methods and systems provided in the embodiments of this specification,

[0011] In some embodiments, the second intelligent agent scores the target network element, including scoring the target network element based on weight score, historical performance score, resource health score, and error penalty item.

[0012] In some embodiments, the target network element is scored based on a weighted score, a historical performance score, a resource health score, and an error penalty item, including: constructing a scoring formula:

[0013]

[0014] N() represents the normalization process applied to the original scores of this dimension within the range [0,1]. Weight coefficients , , , It can be dynamically configured via policy templates;

[0015] The current_load can be extracted from metrics such as CPU utilization and number of sessions in resourceUsage, and max_capacity is the maximum load threshold preset for the instance.

[0016]

[0017] Where success_rate is the transaction success rate and avg_delay is the average response time. and Adjustable weights;

[0018]

[0019] in Let k be the current utilization rate of the resource. This is the alarm threshold for the corresponding resource;

[0020]

[0021] in, To assign different weights based on error type, the error types include configuration error, runtime exception, and communication failure.

[0022] According to a second aspect, this application provides a mobile core network operation and maintenance management device based on a multi-agent system, applied to a mobile core network, characterized in that the device comprises: a receiving module, configured to receive a first dialogue message input by an operation and maintenance personnel; the first dialogue message is received by a first agent and is used to represent a first operation and maintenance intention for the mobile network; a determining module, configured to determine a second agent based on the first operation and maintenance intention, wherein the second agent is determined by reasoning through a large language model; a result acquisition module, configured to send an access request to the core network to the mobile core network using the second agent; based on the query result returned by the mobile core network, the second agent returns a result corresponding to the first operation and maintenance intention; the result corresponding to the first operation and maintenance intention is obtained by reasoning through the large language model. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of a 5G core network structure;

[0025] Figure 2 A schematic diagram of a mobile core network operation and maintenance management architecture based on multi-agent provided in this application embodiment;

[0026] Figure 3 This is a schematic diagram of the structure of an intelligent agent module 120 provided in an embodiment of this application;

[0027] Figure 4 A flowchart illustrating a mobile core network operation and maintenance management method based on multi-agent systems, provided for an embodiment of this application;

[0028] Figure 5 This is a schematic diagram of the structure of another intelligent agent module 120 provided in an embodiment of this application;

[0029] Figure 6 A flowchart illustrating yet another mobile core network operation and maintenance management method based on multi-agent systems is provided for embodiments of this application.

[0030] Figure 7 This is a schematic diagram of the structure of another intelligent agent module 120 provided in an embodiment of this application;

[0031] Figure 8 This is a schematic diagram of the input and output of a fault prediction agent provided in an embodiment of this application;

[0032] Figure 9 A flowchart illustrating yet another mobile core network operation and maintenance management method based on multi-agent systems is provided for embodiments of this application.

[0033] Figure 10 This application provides a method for fault prediction based on multiple agents.

[0034] Figure 11 This application provides a schematic diagram of a mobile core network operation and maintenance management device based on multiple agents. Detailed Implementation

[0035] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0036] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be described below with reference to the accompanying drawings.

[0037] In the description of the embodiments of this application, the words "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the words "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a specific manner.

[0038] In the description of the embodiments of this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, B existing alone, and A and B existing simultaneously. Furthermore, unless otherwise stated, the term "multiple" means two or more.

[0039] The terms “including,” “comprising,” “having,” and variations thereof all mean “including but not limited to,” unless otherwise specifically emphasized.

[0040] As mobile communication technology evolves from 5G to 6G, the core network architecture is undergoing unprecedented changes. The 5G Core Network (5GC), through the introduction of Service-Based Architecture (SBA), achieves deep integration of control plane and user plane separation, Network Function Virtualization (NFV), and Software Defined Networking (SDN), laying the foundation for flexible, efficient, and scalable network deployment. Facing the 6G era, the core network will face more stringent performance requirements and new development opportunities. 6G networks aim to achieve integrated air, space, and sea coverage, supporting Tbps-level peak transmission rates, sub-millisecond end-to-end latency, and near 100% service reliability, placing higher demands on the core network's architecture, capabilities, and intelligence.

[0041] The 5G core network is the most crucial part of the 5G architecture. It is responsible for user access management, data transmission, network control, and the provision of various services. It serves as the brain of the entire 5G network, undertaking the vital mission of handling diverse communication scenarios and providing efficient services to users. Compared to the 4G core network (Evolved Packet Core, EPC), the 5G core network introduces a Service-Based Architecture (SBA), achieving higher data rates, lower latency, and more flexible network function deployment.

[0042] Figure 1 This is a schematic diagram of a 5G core network structure. Figure 1As shown, the 5G core network adopts a modular design, consisting of multiple Network Functions (NFs). These NFs communicate with each other through standardized service interfaces, ensuring architectural flexibility and facilitating future expansion. Among these NFs, the Access and Mobility Management Function (AMF) is primarily responsible for signaling interaction, authentication, and mobility management for terminal access; the Session Management Function (SMF) handles session establishment, modification, and release, and is also responsible for allocating user data plane resources and IP addresses. The User Plane Function (UPF) is responsible for user data forwarding and data path selection. In addition, the Unified Data Management Function (UDM) manages user subscription information and policy data; the Policy Control Function (PCF) provides policy formulation and distribution services such as traffic control and charging rules; and the Network Slice Selection Function (NSSF) helps users allocate suitable network slices. For scenarios requiring third-party application access, the Network Exposure Function (NEF) provides the relevant interfaces. Regarding network security, the Authentication Server Function (AUSF) is responsible for performing user authentication. Finally, the Network Function Repository Function (NRF) is responsible for registering and discovering network element services, ensuring efficient collaboration across the entire core network.

[0043] Currently, the status monitoring and troubleshooting of core network elements mainly rely on the following two methods: Command Line Interface (CLI) query: Operations and maintenance personnel need to log in to the server or container where each network element resides and enter specific, complex command-line instructions to obtain status information. This method requires operations and maintenance personnel to memorize a large number of commands and parameters, resulting in high learning costs and low efficiency. Traditional script automation: Writing scripts (such as Python and Shell) to automatically execute query tasks. However, scripts lack natural language interaction capabilities, have fixed functions, and cannot understand the user's ambiguous or complex intentions. Each time a new query requirement is encountered, the script needs to be modified and tested, resulting in poor flexibility. The above work requires operations and maintenance personnel to be proficient in various network management technologies and tools, familiar with the architecture and characteristics of mobile networks, and possess rich practical experience and the ability to flexibly handle various complex situations. At the same time, it is also necessary to establish a sound operations and maintenance management process and management system, a robust monitoring system and automation tools, continuously optimize network maintenance and management work, and improve operations and maintenance management efficiency and network service quality.

[0044] In summary, the current operation and maintenance management system faces several prominent challenges. First, the operation is complex and requires a high level of expertise. Operation and maintenance personnel must not only be proficient in the mobile core network architecture and the functions of each network element, but also memorize a large number of specialized query commands, resulting in a significant learning curve. Second, the interaction method is not intuitive enough; direct communication using natural language is impossible, and each query requires translating the problem into machine instructions, severely slowing down efficiency in time-sensitive troubleshooting. Third, the system lacks flexibility and adaptability; fixed scripts are insufficient to handle complex and ever-changing combined query scenarios. More challenging is that when cross-network element faults occur, existing methods are mostly limited to reactive responses, lacking proactive risk prediction and fault analysis and early warning mechanisms, leading to frequent and inefficient manual intervention.

[0045] To address the problems of existing technologies, this invention proposes an operation and maintenance management method for mobile core networks. By constructing a comprehensive operation and maintenance management system based on intelligent agents, a first intelligent agent receives dialogue messages input by operation and maintenance personnel, performs intent parsing, and sends them to a large language model for processing. Based on the dialogue messages, the intelligent agent responsible for subsequent message processing is determined. When the dialogue message indicates that a target needs to be evaluated, the first intelligent agent invokes a second intelligent agent, which accesses the mobile core network. After the mobile core network returns the query results for the target network element, the second intelligent agent sends these results to the large language model for processing, obtains a judgment and analysis of the target network element's status, and then returns the final query result. This solves the problem of complex and highly specialized operation and maintenance management operations. Operation and maintenance personnel do not need to be proficient in the mobile core network architecture and the functions of various network elements, nor do they need to memorize a large number of specialized query commands, nor do they need to incur learning costs. Through intuitive interaction and communication using natural language, operation and maintenance efficiency is improved, and various complex and ever-changing combined query scenarios can be handled.

[0046] Figure 2 This is a schematic diagram of a multi-agent-based mobile core network operation and maintenance management architecture provided for an embodiment of this application. Figure 2 As shown, the architecture includes a large language model 110 and an intelligent agent module 120. The intelligent agent module 120 receives dialogue messages sent by maintenance personnel 130, parses the dialogue messages, obtains the target network element status information from the mobile core network, performs inference using the large language model 110, and sends the results back to the maintenance personnel 130. The intelligent agent module 120 may include a graph analysis intelligent agent (first intelligent agent), a network element evaluation intelligent agent (second intelligent agent), a network element screening intelligent agent (third intelligent agent), and a fault prediction intelligent agent (fourth intelligent agent). Through a multi-agent division of labor and cooperation mechanism, core functions such as natural language intent parsing, network element status perception, intelligent screening and evaluation, and fault prediction analysis are integrated to build an automated intelligent operation and maintenance system, providing comprehensive support for the intelligent operation and maintenance management of the mobile core network. The multi-agent cluster includes four types of intelligent agents: intent parsing, network element screening, fault prediction, and network element evaluation. Each agent focuses on a single core capability and cooperates with the others. The mobile core network data layer provides the system with basic data such as the full network element operating status, performance indicators, and alarm logs.

[0047] In one possible implementation, the agent module 120 includes an intent parsing agent (first agent) and a network element evaluation agent (second agent). Figure 3 This is a schematic diagram of the structure of an intelligent agent module 120 provided in an embodiment of this application. Figure 3As shown, the intent parsing agent is mainly used to receive natural language input commands from operation and maintenance personnel 130, perform semantic understanding and task parsing on the input commands, generate structured task descriptions or execution plans, and drive the corresponding agent to complete subsequent operations. For example, the natural language input command is passed as complete semantic information to the intent parsing agent for processing.

[0048] The network element evaluation agent (the second agent) is primarily used to automatically query target network element information in the mobile core network and analyze the input network element instance status data using LLM. Finally, it outputs a report generated by the large model. For example, the generated report includes key indicator data and analytical conclusions and actionable recommendations described in natural language. For example, network element evaluation agent 122 automatically queries target network element information in the mobile core network, including automatically locating the instance address, port, and other information from the target network element instance's configuration file according to predefined rules and configurations, and combining this information into a complete health check interface link. For example, the health check interface URL is: protocol + : / / + domain name or IP + : + port + / + health check endpoint path. Subsequent access to this URL retrieves raw operational status data, including utilization rate and current session count, and then cleans, integrates, and formats this data.

[0049] In one possible implementation, please refer to [link / reference needed]. Figure 3 Step 1: During the runtime phase, a preset startup mechanism is triggered to activate the system. This mechanism initializes the runtime environment and coordinates subsequent execution processes. Step 2: Natural language is input from the command line. Step 3: The natural language is transmitted to the large language model. Step 4: Inference results are obtained from the large language model. Step 5: The intent agent invokes the network element evaluation agent based on the inference results from the large language model. Step 6: The network element evaluation agent automatically accesses the core network. Step 7: The core network returns the query results for the target network element. Step 8: The network element evaluation agent sends the query results for the target network element to the large language model for inference. Step 9: The large language model infers and obtains the analysis results of the query results for the target network element. Step 10: The network element evaluation agent returns an analysis report.

[0050] For example, the query results for the target network element obtained in step 8 include: the running status information of each network element instance, including configuration parameters, signaling indicators, and resource load. Key data such as status.resourceUsage (resource usage) and error (fault information) are returned to the large language model.

[0051] For example, the analysis results are output in a combination of natural language and structured data.

[0052] This application embodiment deploys intelligent agents in the mobile core network to acquire real-time operational status information of each network element instance and returns this new information to a large language model. Based on the returned status information, the large language model, combined with the network element configuration rule base and fault detection logic, analyzes and judges the operational status of the network elements. The entire process realizes a closed loop from natural language intent understanding, automatic addressing query to intelligent analysis and result output, significantly reducing the reliance on operation and maintenance personnel's command line memorization and manual data parsing.

[0053] Figure 4 This is a flowchart illustrating a multi-agent-based mobile core network operation and maintenance management method provided in an embodiment of this application. Figure 4 As shown, the method includes:

[0054] S410, Receive the first dialogue message input by the operation and maintenance personnel; the first dialogue information is received by the first intelligent agent and is used to represent the first operation and maintenance intention for the mobile network.

[0055] As mentioned earlier, maintenance personnel 130 can input their maintenance intentions using natural language through the front-end interactive page, without needing dedicated data query commands. For example, the first maintenance intention is a request for information about the target network element.

[0056] For example, operations and maintenance personnel activate the intelligent agent through a preset startup mechanism and input natural language commands through an interactive interface. These natural language commands are then fully transmitted to the intent-parsing intelligent agent for subsequent semantic understanding and task parsing. The intent-parsing intelligent agent uses LLM to perform semantic parsing on the natural language commands, identifies the user's intent, and generates a structured query task description based on the parsing results. This structured task serves as a trigger signal, invoking the network element evaluation intelligent agent to execute the subsequent data acquisition process.

[0057] S420. Based on the first operation and maintenance intention, determine the second intelligent agent, which is determined by reasoning through a large language model.

[0058] For example, the second agent is the network element evaluation agent.

[0059] S430, the second intelligent agent sends a request to access the core network to the mobile core network.

[0060] In one possible implementation, the network element evaluation agent, combining predefined configuration rules or registration information, automatically locates the network address, access port, and interface path of the target network element instance, generating corresponding access path planning information to guide subsequent status query operations. Following the planned path, the network element evaluation agent initiates a status query request to the instance in the core network to obtain raw operational data. The returned data undergoes unified format conversion, field mapping, and standardization processing to generate structured status data results.

[0061] S440. Based on the query results returned by the mobile core network, the second intelligent agent returns a result corresponding to the first operation and maintenance intention; the result corresponding to the first operation and maintenance intention is obtained by reasoning from the large language model.

[0062] In one possible implementation, standardized structured data is submitted to an LLM (Local Management Module). The LLM, combined with state assessment logic and acquired network element state data, performs comprehensive reasoning and diagnostic analysis to generate structured or text-based reports tailored to operation and maintenance management scenarios. The final analysis results are returned to operation and maintenance management personnel via an interactive interface, completing the entire process from natural language input to diagnostic report generation.

[0063] exist Figure 3 Based on this, the embodiments of this application add a network element screening agent to the agent module 120. Figure 5 This is a schematic diagram of the structure of another intelligent agent module 120 provided in an embodiment of this application. (See attached diagram.) Figure 5 As shown, the network element screening agent performs a multi-dimensional comprehensive analysis of the target network element's real-time load, historical performance indicators, resource health, and the priority of its corresponding slice, constructing a dynamic scoring model. For example, if an operations and maintenance personnel inputs a natural language intent indicating a need for load balancing, the intent parsing module performs semantic analysis using a large language model. The large language model then returns a standardized response. By recognizing the response and invoking the network element evaluation agent, the agent receives real-time operational data (including indicators such as status, resource usage, and error) from each instance obtained from the network element evaluation agent and feeds it into the network element screening agent for multi-dimensional health assessment. Finally, the results are returned to the large language model for analysis and decision-making. The specific network element evaluation agent is determined based on one or more of the following: load, historical performance, resource health score, and error penalty.

[0064] For example, a model implementation based on the following scoring formula is provided:

[0065]

[0066]

[0067] Where N() represents the normalization process performed on the original scores of this dimension within the range [0,1]. Weight coefficients , , , It can be dynamically configured via policy templates and supports temporary adjustments via natural language commands (e.g., "prioritize the instance with the least available resources" can be increased accordingly). The scoring definitions for each dimension are as follows:

[0068] (1) Load fraction:

[0069]

[0070] The current_load can be extracted from metrics such as CPU utilization and number of sessions in resourceUsage, and max_capacity is the maximum load threshold preset for the instance.

[0071] (2) Historical performance score:

[0072] Calculated based on the success rate and average response time of the instance over the past T time window:

[0073]

[0074] Where success_rate is the transaction success rate and avg_delay is the average response time. and The weights are adjustable.

[0075] (3) Resource health score:

[0076] Overall resource utilization metrics include memory, I / O, and available bandwidth.

[0077]

[0078] in Let k be the current utilization rate of the resource. This is the alarm threshold for the corresponding resource.

[0079] (4) Error Penalty Items:

[0080]

[0081] Different weights are assigned based on the type of error (such as configuration error, runtime error, communication failure, etc.).

[0082] In one possible implementation, such as Figure 5As shown, Step 1: During the runtime phase, a preset startup mechanism is triggered to activate the system. This mechanism initializes the runtime environment and coordinates subsequent execution processes. Step 2: Natural language is input from the command line. Step 3: The natural language is transmitted to the large language model. Step 4: Inference results are obtained from the large language model. Step 5: The intent agent calls the network element screening agent based on the inference results of the large language model. Step 6: The network element screening agent calls the network element evaluation agent to query core network time-series data. Step 7: The network element evaluation agent automatically accesses the core network. Step 8: The core network returns the query results for the target network element. Step 9: The network element evaluation agent sends the inspection results of the target network element to the large language model for inference. Step 10: The large language model infers and obtains the analysis results of the inspection results of the target network element. Step 11: The network element evaluation agent returns the analysis results to the network element screening agent. Step 12: The network element screening agent inputs the scoring results into the LLM for judgment and analysis. Step 13: The LLM returns the analysis results to the network element screening agent. Step 14: Output the results of the network element selection agent.

[0083] Figure 6 This application provides a flowchart illustrating yet another mobile core network operation and maintenance management method based on multi-agent systems. For example... Figure 6 As shown, the method includes:

[0084] S610, Receive the second dialogue message input by the operation and maintenance personnel; the first dialogue information is received by the first intelligent agent and is used to represent the second operation and maintenance intention for the mobile network;

[0085] In one possible implementation, operations and maintenance personnel activate the intelligent agent through a preset startup mechanism and input natural language commands through an interactive interface. These natural language commands are then fully transmitted to the intent-parsing intelligent agent for subsequent semantic understanding and task parsing.

[0086] S620. Based on the first dialogue message, determine a third intelligent agent, which is determined through reasoning using a large language model.

[0087] In one possible implementation, the intent parsing agent uses LLM to perform semantic parsing on the natural language instructions, identify the user's intent, and generate a structured task description based on the parsing results. This structured task serves as a trigger signal to invoke the network element filtering agent (the third agent) to execute the subsequent data acquisition process.

[0088] S630. Based on the second maintenance intention, the third intelligent agent sends a request to the second intelligent agent.

[0089] In one possible implementation, the network element screening agent calls the network element evaluation agent. After obtaining the data, the network element evaluation agent checks the data and then returns the data in the correct format.

[0090] The network element evaluation agent comprehensively scores the target network element, as described above, for example, using the formula...

[0091]

[0092] In one possible implementation, the network element selection agent returns the candidate instance with the highest score and its corresponding rating to the LLM. The LLM generates a decision result. Finally, the decision result is output to the operation and maintenance management personnel, realizing intelligent network element selection based on multi-dimensional evaluation.

[0093] S640. Based on the query results of the target network element returned by the second intelligent agent, the third intelligent agent scores the target network element.

[0094] In one possible implementation, the network element screening agent selects a set of candidate network element instances that meet the conditions based on network topology relationships, slice matching rules and other constraints, and the comprehensive score of the target network element by the network element evaluation agent.

[0095] In one possible implementation, the network element screening agent returns the candidate instance with the highest score and its corresponding score result to the LLM.

[0096] S650. Based on the scoring data, the third agent returns a result corresponding to the second operation and maintenance intention; the result corresponding to the second operation and maintenance intention is obtained by reasoning from the large language model.

[0097] In one possible implementation, LLM generates the decision results. Finally, the decision results are output to the operation and maintenance management personnel by the network element selection agent, realizing intelligent network element selection based on multi-dimensional evaluation.

[0098] exist Figure 2 Based on this, the embodiments of this application add a fault prediction agent to the agent module 120. Figure 7 This is a schematic diagram of the structure of another intelligent agent module 120 provided in an embodiment of this application. (See attached diagram.) Figure 7As shown, the fault prediction agent receives results from the network element evaluation agent, including real-time performance metrics of network element instances, historical alarm logs, and network topology relationships—all from multiple heterogeneous sources. This data collectively forms the core foundation for the machine learning model's dynamic learning and pattern recognition. For example, this machine learning model is LightGBM, which will be detailed later. The fault prediction agent outputs a quantitative assessment of the probability of fault occurrence for each network element within a specific future time period, providing directly executable input for intelligent operation and maintenance management. Figure 8 This is a schematic diagram of the input and output of a fault prediction agent provided in an embodiment of this application.

[0099] In one possible implementation, such as Figure 7 As shown, Step 1: During the runtime phase, a preset startup mechanism is triggered to activate the system. This mechanism initializes the runtime environment and coordinates subsequent execution processes. Step 2: Natural language is input from the command line. Step 3: The natural language is transmitted to the large language model. Step 4: Inference results are obtained from the large language model. Step 5: The intent agent calls the fault prediction agent based on the inference results from the large language model. Step 6: The fault prediction agent calls the network element evaluation agent to query core network time-series data. Step 7: The network element evaluation agent automatically accesses the core network. Step 8: The core network returns the query results for the target network element. Step 9: The network element evaluation agent sends the inspection results of the target network element to the large language model for inference. Step 10: The large language model infers and obtains the analysis results of the inspection results of the target network element. Step 11: The network element evaluation agent returns the analysis results to the fault prediction agent. Step 12: The fault prediction agent inputs the prediction results into the LLM for judgment and analysis. Step 13: The LLM returns the analysis results to the fault prediction agent. Step 14: The fault prediction agent outputs the results.

[0100] Figure 9 This application provides a flowchart illustrating yet another mobile core network operation and maintenance management method based on multi-agent systems. For example... Figure 9 As shown, the method includes:

[0101] S910, Receive a third dialogue message input by the operation and maintenance personnel; the third dialogue information is received by the first intelligent agent and is used to represent a third operation and maintenance intention for the mobile network;

[0102] In one possible implementation, operations and maintenance personnel activate the intelligent agent through a preset startup mechanism and input natural language commands through an interactive interface. These natural language commands are then fully transmitted to the intent-parsing intelligent agent for subsequent semantic understanding and task parsing.

[0103] S920. Based on the third dialogue message, determine the fourth intelligent agent, which is determined through reasoning using a large language model.

[0104] In one possible implementation, the intent parsing agent uses LLM to perform semantic parsing of natural language instructions, identifies the user's intent, and generates a structured task description based on the parsing results. This structured task serves as a trigger signal to invoke the fault prediction agent to execute the subsequent data acquisition process.

[0105] S930, the fourth intelligent agent sends a first query request to the second intelligent agent; and obtains the query result returned by the second intelligent agent.

[0106] In one possible implementation, the fault prediction agent calls the network element evaluation agent. After obtaining the data, the network element evaluation agent checks the data and then returns the data in the correct format.

[0107] S940. Based on the query results, the fourth agent inputs the features of the target network element into the first machine learning model and / or the second machine learning model for processing, and obtains the first evaluation result and / or the second evaluation prediction result of the target network element.

[0108] In one possible implementation, the fault prediction agent acquires real-time multidimensional data of network elements, and further calculates the statistical characteristics of each indicator along the time axis based on this sequence (e.g., the mean, variance, variability, and historical percentile of the most recent window), and combines various alarm counts within the past M minutes as contextual features. Subsequently, the above features are input into a machine learning model for inference to obtain coarse-grained and / or fine-grained evaluation results.

[0109] In one possible implementation, pre-defined qualitative and quantitative analysis is used to predict failures using coarse-grained and / or fine-grained information. `p_risk` is coarse-grained information (predicting the probability of a failure event occurring within a future time interval), describing whether a failure will occur; `y_pred` is fine-grained information (numerical predictions of future key performance indicators), characterizing the extent to which the failure will worsen, supporting refined decision-making. Relying solely on point estimates of `y_pred` based on a fixed threshold leads to black-and-white risk jumps near the threshold, failing to capture the actual business logic of risk continuously increasing as the value approaches the threshold. Coarse- and fine-grained information complement each other, better supporting intelligent failure prediction.

[0110] In one possible implementation, during the data preparation phase, historical operational data is first collected, including time-series performance metrics of network elements (such as CPU utilization, memory usage, number of sessions, bandwidth usage, etc.) and alarm log data aligned with these metrics. All data is aligned to a uniform time granularity to ensure comparability of metrics and alarms on the same timeline. Subsequently, the historical data is processed using a sliding time window approach. For each historical moment, a series of multi-dimensional performance metrics within a preset time range is extracted to construct a feature base reflecting operational trends.

[0111] In one possible implementation, during the feature construction process, multiple time-series statistical features are calculated for each key performance indicator, including the statistical value of the most recent time window, the overall fluctuation level, the changing trend between adjacent windows, extreme value features, and positional features relative to the historical distribution of the same period, to characterize the current state and changing trends. Simultaneously, alarm data within the same time window are classified and statistically analyzed according to type and severity, generating alarm count and anomaly density features. These performance features and alarm features are concatenated to form the original feature vector, which is then standardized based on pre-saved statistical parameters to obtain a model input feature vector X_t with a uniform scale.

[0112] In one possible implementation, during the online execution phase, the real-time generated standardized feature vector X_t is input into a pre-trained first machine learning model and / or a second machine learning model for parallel inference. The first machine learning model is a regression model used to predict the numerical point estimate y_pred of future key performance indicators, characterizing the specific degree to which the failure may worsen. The second machine learning model is a classification model used to predict the probability p_risk of an anomaly occurring within a future time interval, characterizing the overall risk of a failure occurring. The outputs of the two models complement each other, forming a fusion result combining numerical and probabilistic predictions.

[0113] In one possible implementation, during the result processing stage, the uncertainty of the predicted value y_pred is first quantified based on the error statistics parameters obtained during the model validation stage, constructing a corresponding confidence interval to characterize the range of credibility of the prediction result. Subsequently, according to the interval in which the risk probability p_risk falls, it is converted into the corresponding risk level according to a preset and configurable mapping rule, realizing the conversion from continuous probability values ​​to discrete risk levels. By combining the dual information of the predicted value and the risk probability, the system can avoid the critical jump problem caused by traditional fixed threshold judgments, making risk assessment more continuous, smooth, and consistent with actual business logic.

[0114] In one possible implementation, during the risk assessment and early warning generation phase, the fault prediction agent receives the predicted value y_pred and the corresponding risk probability p_risk returned after LLM inference.

[0115] S950. Based on the first evaluation result and / or the second evaluation prediction result, the fourth agent returns a result corresponding to the third operation and maintenance intention.

[0116] In one possible implementation, the fault prediction agent returns the results to the LLM. The LLM generates a judgment result. Finally, the judgment result is output to the operation and maintenance management personnel through the agent module, which can be used for reference and early warning, and relevant suggestions can be given.

[0117] In one possible implementation, the agent module can proactively invoke the fault prediction service during the analysis process to obtain fault risk predictions for specified network elements or slices; simultaneously, the fault prediction agent will also run automatically periodically. In implementation, the fault prediction agent constructs a complete data pipeline, including real-time data ingestion, feature engineering, and model inference, forming a closed loop of monitoring-prediction-decision-execution. Figure 10 This application provides a method for fault prediction based on multi-agent systems. For example... Figure 10 As shown, the method includes:

[0118] S1010. Based on the time-series performance index data and alarm log data of the target network element, calculate the time-series statistical features and alarm count and anomaly density features for several performance indicators; the time-series statistical features include the statistical value of the most recent time window, the overall fluctuation degree, the changing trend between adjacent windows, extreme value features, and the positional features relative to the historical distribution of the same period; the alarm count and anomaly density features are obtained by classifying and statistically analyzing alarm data within the same time window according to type and severity.

[0119] S1020. The time-series statistical features of the calculated performance indicators and the alarm count and anomaly density features are concatenated to obtain the output feature vector.

[0120] S1030. Based on the input feature vector, the first machine learning model outputs a numerical point estimate of the predicted key performance indicators for the future, indicating the specific degree to which the fault may deteriorate; the second machine learning model outputs the probability of an anomaly occurring within a future time interval, indicating whether there is an overall risk of a fault occurring; the first machine learning model is a classification model, and the second machine learning model is a classification model.

[0121] S1040. Based on the output results of the first machine learning model and / or the second machine learning model, obtain numerical prediction and / or probability prediction.

[0122] In summary, the intelligent operation and maintenance management method for mobile core networks provided in this application constructs a comprehensive operation and maintenance management system that integrates mobile core network status detection, intelligent element selection, and dynamic fault prediction, relying on multi-agent collaboration. This improves operation and maintenance management efficiency and reduces manual maintenance costs. Specifically, firstly, the network element status detection function automatically acquires the status information of each network element instance and analyzes whether the network element has experienced a fault through an intelligent agent. Secondly, intelligent element selection is performed based on the network element instance information. The intelligent agent sets a score based on the current load, historical performance, and other information of the network element and selects network element instances with higher scores for connection. Finally, a dynamic fault prediction method relying on a machine learning model is provided to minimize network element instance crashes and improve the stability of the mobile core network.

[0123] Figure 11 This application provides a schematic diagram of a mobile core network operation and maintenance management device based on multi-agent systems. (See the attached diagram.) Figure 11 As shown, the device includes: a receiving module for receiving a first dialogue message input by maintenance personnel; the first dialogue message is received by a first intelligent agent and is used to represent a first maintenance intention for the mobile network; a determining module for determining a second intelligent agent based on the first maintenance intention, wherein the second intelligent agent is determined by reasoning through a large language model; a result acquisition module for sending an access request to the core network to the mobile core network using the second intelligent agent; based on the query result returned by the mobile core network, the second intelligent agent returns a result corresponding to the first maintenance intention; the result corresponding to the first maintenance intention is obtained by reasoning through the large language model.

[0124] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-agent-based operation and maintenance management method for mobile core networks, applied to mobile core networks. Its features are, The method includes: Receive the first dialogue message input by the operation and maintenance personnel; the first dialogue information is received by the first intelligent agent and is used to represent the first operation and maintenance intention for the mobile network; Based on the first operational intent, a second intelligent agent is determined, which is determined through reasoning using a large language model. The second intelligent agent sends a request to access the core network to the mobile core network; Based on the query results returned by the mobile core network, the second agent returns a result corresponding to the first operation and maintenance intention; the result corresponding to the first operation and maintenance intention is obtained by reasoning from the large language model.

2. The method according to claim 1, characterized in that, The method further includes: The system receives a second dialogue message input by the operations and maintenance personnel; the second dialogue message is received by the first intelligent agent and is used to represent a second operations and maintenance intention for the mobile network. Based on the second dialogue message, a third intelligent agent is determined, which is determined by reasoning through a large language model. Based on the second maintenance intention, the third agent sends a request to the second agent; Based on the query results of the target network element returned by the second intelligent agent, the third intelligent agent scores the target network element; Based on the scoring data, the third agent returns a result corresponding to the second operation and maintenance intention; the result corresponding to the second operation and maintenance intention is obtained by reasoning from the large language model.

3. The method according to claim 1, characterized in that, The method further includes: Receive a third dialogue message input by the operation and maintenance personnel; the third dialogue information is received by the first intelligent agent and is used to represent a third operation and maintenance intention for the mobile network; Based on the third dialogue message, a fourth intelligent agent is determined, which is determined by reasoning through a large language model. The fourth agent sends a first query request to the second agent; and obtains the query result returned by the second agent. Based on the query results, the fourth agent inputs the features of the target network element into the second and / or third large language models for processing, and obtains the first evaluation result and / or the second evaluation prediction result of the target network element; Based on the first evaluation result and / or the second evaluation prediction result, the fourth agent returns a result corresponding to the third operation and maintenance intention.

4. The method according to claim 3, characterized in that, The fourth intelligent agent includes a first machine learning model and / or a second machine learning model, wherein the first machine learning model is a classification model and the second machine learning model is a classification model; Based on the first machine learning model, the system outputs numerical point estimates of future key performance indicators, indicating the specific extent to which the fault may worsen. Based on the second machine learning model, the output predicts the probability of an anomaly occurring within a future time interval, indicating the overall risk of a failure occurring. Numerical predictions and / or probabilistic predictions are obtained based on the outputs of the first machine learning model and / or the second machine learning model.

5. The method according to claim 3, characterized in that, Data preparation for training the first machine learning model and / or the second machine learning model includes: Collect historical operational data, including time-series performance metrics of the target network element and alarm log data aligned with the time; the time-series performance metrics of the target network element include CPU utilization, memory utilization, number of sessions, and bandwidth usage. The data is aligned according to a uniform time granularity. Historical data is processed using a sliding time window approach. For each historical moment, a multidimensional performance index sequence within a preset time range prior to that historical moment is extracted to construct a feature base reflecting the operating trend.

6. The method according to claim 3, characterized in that, Training the first machine learning model and / or the third major language model includes: For the first machine learning model, the supervision label is constructed as the true key performance indicator value after the prediction step size at the current time. For the second machine learning model, a binary classification result is constructed with the supervision label as whether the target abnormal event will occur within a future preset time interval; The training set and test set are divided in chronological order, and the first machine learning model and / or the second machine learning model are trained.

7. The method according to claim 3, characterized in that, Constructing input features for the first machine learning model and / or the second machine learning model includes: Multiple time-series statistical characteristics are calculated for key performance indicators, including statistical values ​​of the most recent time window, overall fluctuation degree, changing trends between adjacent windows, extreme value characteristics, and positional characteristics relative to the historical same period distribution, in order to characterize the current state and changing trends. Furthermore, alarm data within the same time window is classified and statistically analyzed according to type and severity, generating alarm count and anomaly density features; The various time-series statistical features and the alarm count and anomaly density features are concatenated to obtain the input feature vector.

8. The method according to claim 2, characterized in that, The second intelligent agent scores the target network element, including: The target network element is scored based on weighted scores, historical performance scores, resource health scores, and error penalty items.

9. The method according to claim 8, characterized in that, The target network element is scored based on weighted scores, historical performance scores, resource health scores, and error penalties, including: constructing a scoring formula: N() represents the normalization process applied to the original scores of this dimension within the range [0,1]. Weight coefficients , , , It can be dynamically configured via policy templates; The current_load can be extracted from metrics such as CPU utilization and number of sessions in resourceUsage, and max_capacity is the maximum load threshold preset for the instance. Where success_rate is the transaction success rate and avg_delay is the average response time. and Adjustable weights; in Let k be the current utilization rate of the resource. This is the alarm threshold for the corresponding resource; in, To assign different weights based on error type, the error types include configuration error, runtime exception, and communication failure.

10. A mobile core network operation and maintenance management device based on multi-agent systems, applied to mobile core networks. Its features are, The device includes: The receiving module is used to receive the first dialogue message input by the operation and maintenance personnel; the first dialogue information is received by the first intelligent agent and is used to represent the first operation and maintenance intention for the mobile network. The determination module is used to determine a second intelligent agent based on the first operation and maintenance intention, wherein the second intelligent agent is determined by reasoning through a large language model; The result acquisition module uses the second agent to send a request to access the core network to the mobile core network; based on the query result returned by the mobile core network, the second agent returns a result corresponding to the first operation and maintenance intention; the result corresponding to the first operation and maintenance intention is obtained by reasoning from the large language model.

Images