A power communication network operation and maintenance management method based on a large model and an agent drive

Abstract

The application discloses a power communication network operation and maintenance management method based on a large model and an agent driver, which comprises the following steps: acquiring a multi-source standardized original data set; performing edge calculation based on the multi-source standardized original data set by using an edge perception agent to obtain general feature vectors and key event structured data corresponding to various operation and maintenance tasks, and uploading the general feature vectors and the key event structured data to a cloud core decision-making agent; performing core research and judgment based on the general feature vectors and the key event structured data by using a power communication special large model of the cloud core decision-making agent to determine an optimal operation and maintenance decision scheme, performing general confidence evaluation on the optimal operation and maintenance decision scheme, and determining a comprehensive confidence evaluation result; performing decision distribution based on the comprehensive confidence evaluation result and the optimal operation and maintenance decision scheme, and performing operation and maintenance based on a decision distribution result; and performing closed-loop self-optimization on an edge side light-weight small model, a cloud special large model and a knowledge base based on operation and maintenance execution feedback data.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent operation and maintenance of power communication networks and application of industrial large models, and more specifically, to a method and system for operation and maintenance management of power communication networks based on large models and agent-driven systems. Background Technology

[0002] The power communication network is the "neural network" of the new power system, and its operational reliability directly determines the safe and stable operation of the power grid. Currently, the power communication network has been built into a four-tiered network covering the entire country: backbone network, provincial network, municipal network, and distribution network. The total length of optical fiber cables exceeds 2 million kilometers, and there are over 10 million communication devices of various types. The number and complexity of the services it carries are growing exponentially. Traditional operation and maintenance models are no longer suitable for the "high reliability, high resilience, and high intelligence" operational requirements of the new power system, mainly due to three core pain points: Fragmented operation and maintenance scenarios and numerous system silos: Different operation and maintenance scenarios (fault handling, performance monitoring, resource management, emergency power supply) are built with independent digital systems, resulting in data not being interconnected, processes not being connected, and standards not being unified, forming a large number of information silos and making it impossible to achieve unified management and control of operation and maintenance across the entire domain; Low level of intelligence and high dependence on manual labor: The existing system only has basic data collection and fixed threshold alarm capabilities. The core links such as anomaly identification, root cause analysis, decision generation, and disposal execution all rely on the professional experience of operation and maintenance personnel, resulting in low operation and maintenance efficiency, slow response speed, and high risk of human error. Lack of a universal intelligent architecture and serious duplication of construction: All kinds of intelligent operation and maintenance applications adopt a "chimney-style" development model, with each scenario building its own data access, model training, and decision execution modules. This results in low technology reuse, long development cycles, high maintenance costs, and an inability to quickly respond to the operation and maintenance needs of new businesses.

[0003] While intelligent agent technology has achieved basic applications in general industrial equipment operation and maintenance scenarios, it is completely detached from the business attributes of power communication networks. It fails to integrate core elements such as power operation and maintenance standards, power grid-communication network coupling mechanisms, and power-specific system interfaces, making it unsuitable for the power industry's operational requirements of "high reliability, strong compliance, and high security." Therefore, there is an urgent need to develop a universal intelligent operation and maintenance closed-loop management method for all scenarios in power communication networks, enabling intelligent and autonomous management of the entire process of operation and maintenance for all types of equipment.

[0004] The first existing technology is a multi-system integrated digital operation and maintenance platform for power communication networks, which is currently the mainstream operation and maintenance solution in the power industry. This technology uses a data middleware platform to centrally collect and standardize data from multiple independent operation and maintenance systems, such as the TMS transmission network management system, the integrated communication network management system, and the dispatch data network management system. This constructs a unified communication resource database and a large-scale operation monitoring dashboard, enabling visualized display of all network communication resources, centralized monitoring of operational status, and unified presentation of fault alarms. Some platforms integrate basic work order management functions, enabling online workflow from fault alarms to maintenance work orders.

[0005] Existing technology one only solves the basic problems of "data centralization" and "visual display", but does not achieve true intelligent operation and maintenance, nor does it build a generalized closed-loop management and control architecture. Its core shortcomings are as follows: the technology has no autonomous decision-making ability and is still a passive operation and maintenance mode; the technology process is fragmented and does not form a closed-loop management and control; the technology's intelligent capabilities are fragmented and lack general algorithm support; at the same time, the technology is disconnected from power grid business and cannot support targeted operation and maintenance.

[0006] Existing technology two is a general-purpose intelligent operation and maintenance platform for industrial equipment, mainly applied in general industrial fields such as manufacturing and petrochemicals. This technology adopts a three-tier "cloud-edge-device" architecture, constructing functional modules such as equipment data acquisition, edge-side anomaly detection, cloud-based big data analysis, and predictive maintenance. Some platforms introduce large-scale model technology to achieve natural language interaction and simple root cause analysis of faults. This technology is designed with universality and versatility as its goals, and can be adapted to the operation and maintenance needs of industrial equipment in different industries.

[0007] Existing technology 2, as a smart operation and maintenance solution for general industrial fields, completely fails to consider the industry-specific characteristics of power communication networks and cannot be directly applied to power communication operation and maintenance scenarios. Its core shortcomings are as follows: the technology is not adapted to the business coupling characteristics of power communication networks; the technology has extremely poor compatibility with the existing power production system; the technology lacks a dedicated knowledge base and rule base for the power industry; and the technology lacks a hierarchical control and compliance guarantee mechanism.

[0008] Therefore, in response to the core shortcomings of existing technologies, there is an urgent need for a power communication network operation and maintenance management method based on large models and agent-driven approaches. Summary of the Invention

[0009] This invention proposes a method and system for operation and maintenance management of power communication networks based on large models and agent-driven approaches, in order to solve the problem of how to efficiently manage and maintain power communication networks.

[0010] To address the aforementioned problems, according to one aspect of the present invention, a method for operation and maintenance management of power communication networks based on large-scale models and agent-driven approaches is provided, the method comprising: Acquire multi-source operation and maintenance data output by various production systems in the power communication network, and preprocess the multi-source operation and maintenance data to obtain a multi-source standardized raw dataset; The edge-aware agent performs edge computing based on the multi-source standardized raw dataset to obtain general feature vectors and key event structured data corresponding to various operation and maintenance tasks, and uploads the general feature vectors and key event structured data to the cloud core decision-making agent; The cloud-based core decision-making intelligent agent utilizes a dedicated large model for power communication, based on the general feature vectors and key event structured data, to perform core analysis according to the type of operation and maintenance task, determine the optimal operation and maintenance decision scheme, and conduct a general confidence assessment of the optimal operation and maintenance decision scheme to determine the comprehensive confidence assessment result. Decision routing is performed based on the comprehensive confidence assessment results and the optimal operation and maintenance decision scheme, and operation and maintenance are carried out based on the decision routing results.

[0011] Preferably, the multi-source operation and maintenance data includes: basic resource static data, real-time dynamic operation data, and operation and maintenance knowledge and management data.

[0012] Preferably, the preprocessing of the multi-source operation and maintenance data includes: The multi-source operation and maintenance data are processed for format unification, time sequence alignment, metadata annotation, and quality verification.

[0013] Preferably, the method further includes: before using the edge-aware agent to perform edge computing based on the multi-source standardized original dataset, performing dynamic outlier removal, intelligent missing value completion, data smoothing and noise reduction, and normalization and standardization processing on the multi-source standardized original dataset in sequence.

[0014] Preferably, the method further includes: Local low-latency processing is performed on the aforementioned multi-source standardized original dataset to obtain local low-latency handling execution logs, which are used for local fault tracing and evidence collection and offline self-supervised learning training samples for edge small models. The local low-latency handling execution logs include: records of control command issuance triggered autonomously by edge nodes based on local judgment and local anomaly marking records. The control command issuance records include: action type, execution time, target device, and execution result status code. The local anomaly marking records include: anomaly type, feature value, and a summary of the judgment criteria.

[0015] Preferably, based on the general feature vector and key event structured data, the core tasks are analyzed according to the type of operation and maintenance tasks to determine the optimal operation and maintenance decision-making scheme, including: RAG retrieval enhancement is performed based on the general feature vector to obtain contextual enhancement information; Based on the contextual enhancement information and key event structured data, the corresponding algorithm plugins are automatically invoked according to different operation and maintenance task types to complete the special analysis and judgment under the current operation and maintenance scenario and generate the preliminary analysis results of the corresponding scenario. Based on the preliminary assessment results, a coupling analysis of the power grid and communication network is performed. A quantitative assessment model is established based on the four-level hierarchical topology relationship of "communication equipment - network link - service bearer - power grid security". The objective risk level and the list of affected power grid services are determined based on the quantitative assessment model. The quantitative assessment model is persistently stored in the knowledge graph triple format to support bidirectional inference of forward and reverse mapping. Based on the enhanced contextual information, preliminary assessment results, objective risk levels, and a list of affected businesses, a large model is used for analysis, and the decision generation algorithm plugin for the corresponding operation and maintenance scenario performs structured constraint verification to output the optimal operation and maintenance decision solution.

[0016] Preferably, the optimal operation and maintenance decision scheme is subjected to a general confidence assessment to determine the comprehensive confidence assessment result, including: The optimal operation and maintenance decision scheme is automatically compared with the power communication operation and maintenance specifications and safety procedures to check whether the decision scheme meets the compliance requirements. When the optimal operation and maintenance decision scheme meets the compliance requirements, the comprehensive credibility assessment of the optimal operation and maintenance decision scheme is carried out through the ConfidenceNet general confidence network; Among them, a multi-dimensional input feature system is composed of multiple core confidence factors, including: alarm confidence, data quality confidence, model confidence, scenario matching confidence, expert rule matching confidence, and operating environment confidence; Based on Bayesian neural networks, uncertainty modeling is performed on multi-dimensional input features, and comprehensive confidence scores, uncertainty estimates and attribution data are output. Based on the comprehensive confidence score, uncertainty estimate, and attribution data, the comprehensive confidence assessment result is determined.

[0017] Preferably, the method further includes: Based on the feature distribution of the current operation and maintenance scenario and historical feedback data, the fusion weights of each core confidence factor are dynamically generated through an attention mechanism; wherein, the attention network uses the original input feature vector As input, output the dynamic fusion weights of the six types of factors. All weights satisfy the normalization constraint ; the weighted feature vector Input the Bayesian confidence inference engine to determine the fusion weights of each core confidence factor.

[0018] Preferably, the decision-making process based on the comprehensive confidence assessment result and the optimal operation and maintenance decision scheme includes: Based on the comprehensive confidence assessment results, dynamic thresholds are determined to complete the decision-making process, including: matching scenario-based dynamic thresholds based on the optical cable carrying service level, power grid operation scenario, and risk level to divide the assessment results. The automated processing flow is executed. When the confidence level is greater than or equal to the scenario dynamic threshold and it is not a high-risk scenario, the specific measures in the optimal operation and maintenance decision scheme are used as the execution instructions. The existing operation and maintenance system is directly connected through the standardized API interface to complete the fully automated execution of the process, including: automatic push of graded early warning information to the corresponding operation and maintenance personnel, automatic generation and circulation of maintenance work orders, connection of maintenance plan system and optimization of production plan, and full tracking and data feedback of operation and maintenance handling process. The manual review process is executed. When the confidence level is less than the scene dynamic threshold or it is a high-risk scene, the full analysis results and the optimal operation and maintenance decision plan are pushed to professional operation and maintenance personnel for manual review and optimization, and the review opinions and corrected data are output. Output the results of the implementation of operation and maintenance decisions, the data for manual review and correction, the feedback data of the entire process, and the comprehensive confidence assessment results.

[0019] Preferably, the method further includes: A dynamic threshold is determined based on preset constraints; wherein, the preset constraints include: Lower bound constraint: The minimum confidence threshold for automated execution branches in any scenario must not be lower than 0.75; Scenario-based risk weighted constraints: Dynamic thresholds are positively correlated with scenario risk levels—power supply scenario threshold >= 0.95, routine maintenance scenario threshold >= 0.80, and daily monitoring scenario threshold >= 0.75; Data quality constraints: When the source data integrity rate is below 85% or the key feature missing rate exceeds 10%, the threshold will automatically increase by 0.05-0.10.

[0020] Preferably, the method further includes: Based on the implementation results of operation and maintenance decisions, manual review and correction of data, feedback data from the entire process of handling, and comprehensive confidence assessment results, the lightweight small model corresponding to the pluggable algorithm plugin in the edge-side perception agent and the cloud-dedicated large model are collaboratively optimized, and the knowledge base is updated to complete the closed-loop optimization of the model and the knowledge base.

[0021] According to another aspect of the present invention, a power communication network operation and maintenance management system is provided, the system comprising: The data acquisition unit is used to acquire multi-source operation and maintenance data output by various production systems in the power communication network, and to preprocess the multi-source operation and maintenance data to obtain a multi-source standardized raw dataset. The edge sensing unit is used to perform edge computing based on the multi-source standardized raw dataset using the edge sensing agent, to obtain general feature vectors and key event structured data corresponding to various operation and maintenance tasks, and to upload the general feature vectors and key event structured data to the cloud core decision-making agent; The core decision-making unit is used by the cloud-based core decision-making intelligent agent to utilize a dedicated large model for power communication, based on the general feature vector and key event structured data, to perform core analysis according to the type of operation and maintenance task, determine the optimal operation and maintenance decision scheme, and perform a general confidence assessment on the optimal operation and maintenance decision scheme to determine the comprehensive confidence assessment result. The decision-making and routing unit is used to perform decision-making and routing based on the comprehensive confidence assessment results and the optimal operation and maintenance decision scheme, and to perform operation and maintenance based on the decision-making and routing results.

[0022] This invention provides a power communication network operation and maintenance management method based on a large model and agent-driven approach, comprising: acquiring multi-source operation and maintenance data output by various production systems in the power communication network, and preprocessing the multi-source operation and maintenance data to obtain a multi-source standardized raw dataset; utilizing an edge-aware agent to perform edge computing based on the multi-source standardized raw dataset to obtain general feature vectors and key event structured data corresponding to various operation and maintenance tasks, and uploading the general feature vectors and key event structured data to a cloud-based core decision-making agent; the cloud-based core decision-making agent using a power communication-specific large model to perform core analysis based on the general feature vectors and key event structured data according to the operation and maintenance task type, determining the optimal operation and maintenance decision scheme, and performing a general confidence assessment on the optimal operation and maintenance decision scheme to determine a comprehensive confidence assessment result; performing decision diversion based on the comprehensive confidence assessment result and the optimal operation and maintenance decision scheme, and performing operation and maintenance based on the decision diversion result. This invention is the first to construct a universal intelligent operation and maintenance closed-loop architecture for power communication networks across all scenarios: it breaks down the scenario barriers and data silos of traditional operation and maintenance systems, realizes unified management and control of operation and maintenance scenarios, and significantly reduces the cost of intelligent construction and maintenance; the collaboration between edge small models and cloud large models takes into account both online real-time performance and deep prediction capabilities, and is suitable for the deployment needs of communication networks at different levels; it establishes a quantitative assessment model of "communication anomaly - business impact - power grid risk", realizing a deep binding between operation and maintenance decisions and power grid security requirements, ensuring that limited operation and maintenance resources prioritize core power grid services; through a full-link closed-loop self-optimization mechanism, the system can continuously iterate and optimize based on actual operation and maintenance data, constantly improving the accuracy of judgment and the rationality of decision-making, and adapting to the dynamically changing operational needs of new power systems. Attached Figure Description

[0023] Exemplary embodiments of the present invention can be more fully understood by referring to the following figures: Figure 1 A flowchart of a power communication network operation and maintenance management method 100 according to an embodiment of the present invention; Figure 2 This is an overall architecture diagram according to an embodiment of the present invention; Figure 3 A flowchart illustrating the multi-source data unified access process according to an embodiment of the present invention; Figure 4 This is a block diagram of the internal structure of the edge-aware general-purpose intelligent agent according to an embodiment of the present invention; Figure 5 This is a flowchart of the cloud-based core decision-making general intelligent agent reasoning process according to an embodiment of the present invention; Figure 6 A detailed flowchart of the decision-making and implementation process according to an embodiment of the present invention; Figure 7 This is a flowchart illustrating the entire closed-loop self-optimization iterative process according to an embodiment of the present invention. Figure 8 This is a schematic diagram of the structure of a power communication network operation and maintenance management system 800 according to an embodiment of the present invention. Detailed Implementation

[0024] Exemplary embodiments of the invention will now be described with reference to the accompanying drawings. However, the invention may be embodied in many different forms and is not limited to the embodiments described herein. These embodiments are provided to fully and completely disclose the invention and to fully convey its scope to those skilled in the art. The terminology used in the exemplary embodiments illustrated in the drawings is not intended to limit the invention. In the drawings, the same units / elements are referred to by the same reference numerals.

[0025] Unless otherwise stated, the terms used herein (including technical terms) have their common meaning as understood by one of ordinary skill in the art. Furthermore, it is understood that terms defined in commonly used dictionaries should be understood to have a meaning consistent with the context of their relevant field, and not to be interpreted as having an idealized or overly formal meaning.

[0026] Figure 1 This is a flowchart of a power communication network operation and maintenance management method 100 according to an embodiment of the present invention. Figure 1As shown, the power communication network operation and maintenance management method provided by the embodiments of the present invention constructs for the first time a universal intelligent operation and maintenance closed-loop architecture for the entire power communication network scenario: breaking down the scenario barriers and data silos of traditional operation and maintenance systems, realizing unified management and control of operation and maintenance scenarios, and significantly reducing the cost of intelligent construction and maintenance; the collaboration between edge small models and cloud large models takes into account both online real-time performance and deep prediction capabilities, and is suitable for the deployment needs of communication networks at different levels; a quantitative assessment model of "communication anomaly - business impact - power grid risk" is established, realizing the deep binding of operation and maintenance decisions with power grid security requirements, ensuring that limited operation and maintenance resources prioritize core power grid services; through a full-link closed-loop self-optimization mechanism, the system can continuously iterate and optimize based on actual operation and maintenance data, continuously improving the accuracy of judgment and the rationality of decision-making, and adapting to the dynamic changes in the operation requirements of new power systems. The power communication network operation and maintenance management method 100 provided by the embodiments of the present invention starts from step 101. In step 101, multi-source operation and maintenance data output by each production system in the power communication network are obtained, and the multi-source operation and maintenance data are preprocessed to obtain multi-source standardized raw datasets.

[0027] Preferably, the multi-source operation and maintenance data includes: basic resource static data, real-time dynamic operation data, and operation and maintenance knowledge and management data.

[0028] Preferably, the preprocessing of the multi-source operation and maintenance data includes: The multi-source operation and maintenance data are processed for format unification, time sequence alignment, metadata annotation, and quality verification.

[0029] Combination Figure 2 As shown, in this invention, a four-layer general intelligent agent closed-loop architecture of "perception-decision-execution-reflection" is constructed, forming a general management and control process of unified access of multi-source data → lightweight preprocessing and feature extraction of edge perception intelligent agents → in-depth analysis and decision generation of cloud core decision-making intelligent agents → decision diversion and execution implementation → closed-loop self-optimization iteration. All modules adopt standardized interface design to support rapid expansion and customized configuration for different operation and maintenance scenarios.

[0030] In this invention, the various production systems of the power communication network are first connected through a unified access link for multi-source data to complete the lightweight standardization and labeling of seven types of operation and maintenance data, and output a standardized original multi-source dataset.

[0031] In this invention, this step is the system's only data entry point, combined with... Figure 3 As shown, a standardized end-to-end API interface gateway connects to all existing production systems in the power communication network, receiving operation and maintenance data output from each system. The core access points are three main categories of general operation and maintenance data: Static data of basic resources: basic information that does not change frequently over time, forming the foundation for operation and maintenance; Real-time dynamic operation data: Status data that changes in real time over time is the core of the sensing network operation; Operations and maintenance knowledge and management data: The knowledge and process data that support decision-making and management are the core inputs for intelligent operations and maintenance.

[0032] For example, the multi-source operation and maintenance data specifically includes, but is not limited to, the following seven categories: Equipment operation data: Real-time operating parameters, performance indicators, and alarm information of various communication devices; Network topology data: physical topology, logical topology, service topology, and protection routing topology of the entire communication network; Service-bearing data: Configuration information, security level, bearer links, and redundancy backup status of various power grid services; Power grid operating data: load status, operating mode, fault information, and power supply protection task information of the primary power grid; Environmental and geographic data: meteorological information, geological information, natural disaster early warning, and information on the surrounding environment of the route; Operations and maintenance management data: Operations and maintenance personnel information, maintenance plans, historical work orders, spare parts inventory, and operations and maintenance procedures; Knowledge base data: power communication operation and maintenance standards, maintenance procedures, historical fault cases, and expert experience and knowledge.

[0033] This invention also requires lightweight normalization and annotation of multi-source data, including: ① format unification (converting raw data from different sources and with different protocol formats into a unified standard data format within the system); ② time sequence alignment (timestamp calibration and synchronization of multi-source data according to a unified time base); ③ metadata annotation (adding structured metadata information such as source system identifier, device ID, data type label, and collection timestamp to each data entry); ④ quality verification (identifying and marking missing fields, aberration fields, and timestamp breakpoints in the data to form data quality labels, without modifying or filling in the original numerical content). After processing, a standardized raw multi-source dataset is output, providing unified and compliant data input for subsequent end-to-end processing.

[0034] In step 102, the edge-aware agent performs edge computing based on the multi-source standardized raw dataset to obtain general feature vectors and key event structured data corresponding to various operation and maintenance tasks, and uploads the general feature vectors and key event structured data to the cloud core decision agent.

[0035] Preferably, the method further includes: before using the edge-aware agent to perform edge computing based on the multi-source standardized original dataset, performing dynamic outlier removal, intelligent missing value completion, data smoothing and noise reduction, and normalization and standardization processing on the multi-source standardized original dataset in sequence.

[0036] Preferably, the method further includes: calling a pluggable lightweight algorithm plugin based on the edge-aware agent, wherein the loaded algorithm plugin completes the front-end perception, feature extraction and low-latency local execution of various operation and maintenance tasks to obtain the general feature vector and key event structured data; wherein the pluggable lightweight algorithm plugin is a modular component for the edge-aware agent to realize perception and local processing, and is loaded or unloaded as needed according to the computing power conditions and operation and maintenance task requirements of each edge node to support the flexible expansion of operation and maintenance capabilities across all scenarios.

[0037] Preferably, the method further includes: Local low-latency processing is performed on the aforementioned multi-source standardized original dataset to obtain local low-latency handling execution logs, which are used for local fault tracing and evidence collection and offline self-supervised learning training samples for edge small models. The local low-latency handling execution logs include: records of control command issuance triggered autonomously by edge nodes based on local judgment and local anomaly marking records. The control command issuance records include: action type, execution time, target device, and execution result status code. The local anomaly marking records include: anomaly type, feature value, and a summary of the judgment criteria.

[0038] like Figure 2 As shown, in this invention, the edge-aware intelligent agent uses the standardized original multi-source dataset as the sole input, completes data preprocessing at the edge node, loads the corresponding algorithm plugin according to the type of operation and maintenance task, completes front-end perception, feature extraction and low-latency local execution of various operation and maintenance tasks, encrypts and uploads the generated multi-dimensional general feature vector and key event structured data to the cloud core decision-making intelligent agent, and retains the local low-latency processing execution log at the edge node for closed-loop self-optimization.

[0039] This step is deployed at the edge nodes and site monitoring terminals of the municipal communication network, such as... Figure 4 As shown, a pluggable lightweight algorithm framework is adopted, which supports loading algorithm plugins corresponding to any operation and maintenance task on demand according to the computing power conditions and operation and maintenance needs of different edge nodes, realizing flexible expansion of operation and maintenance capabilities across all scenarios. The core completes three major general functions: (1) Real-time data preprocessing: Using the standardized original multi-source dataset output in step 101 as the sole input, deep data cleaning and refinement for feature extraction are completed at the edge nodes: ① Dynamic outlier removal (using sliding window filtering or the 3σ principle to identify and remove outliers exceeding the limit); ② Intelligent missing value completion (filling missing fields with values ​​based on time-series interpolation or neighbor-device correlation regression); ③ Data smoothing and noise reduction (suppressing acquisition noise and improving the signal-to-noise ratio); ④ Normalization and standardization (mapping different dimensional parameters to the same numerical range to eliminate the influence of dimensionality). After processing, a high-quality, clean time-series data stream is output as the direct data base for subsequent algorithm plugin feature extraction and local discrimination.

[0040] (2) Full-scene perception and local processing: Load the corresponding algorithm plugin to complete the front-end perception, feature extraction and low-latency local execution of various operation and maintenance tasks. After processing, three types of data are output in parallel: ① Local low-latency handling execution log: including the control command issuance record (action type, execution time, target device, execution result status code) triggered autonomously by the edge node based on local judgment and the local anomaly marking record (anomaly type, feature value, judgment basis summary). This log is stored in the local storage medium of the edge node and is not actively uploaded to the cloud. It is used for local fault tracing and evidence collection and offline self-supervised learning training samples of edge small models; ② Multi-dimensional general feature vector (a high-dimensional semantic feature array formed by dimensionality reduction mapping of the standardized original multi-source dataset output in step 1, used for similarity matching of RAG retrieval in the cloud); ③ Key event structured data (including standardized fields such as device identification, alarm type, parameter limit value, topological relationship, etc., used for inference context filling of large models in the cloud).

[0041] (3) Data intelligent filtering and upload scheduling: After the ② multi-dimensional general feature vector and ③ key event structured data generated in step (2) are compressed and encrypted, they are uploaded to the core decision-making intelligent body in the cloud as needed, so that different functional modules can use them as needed; the ① local low-latency processing execution log generated in step (2) is retained in the local storage medium of the edge node according to the preset strategy and does not enter the upload scheduling process.

[0042] For example, the algorithm plugins described in this section are divided into multiple categories according to the type of operation and maintenance tasks, including but not limited to: fault diagnosis plugins (anomaly detection, root cause localization), performance analysis plugins (deterioration trend extraction, service quality monitoring), resource scheduling plugins (resource load monitoring, equipment status collection), security assurance plugins (power supply link monitoring, configuration compliance check), etc., which can cover all mainstream operation and maintenance scenarios of power communication networks.

[0043] In step 103, the cloud-based core decision-making agent utilizes a dedicated large model for power communication, based on the general feature vector and key event structured data, to perform core analysis according to the type of operation and maintenance task, determine the optimal operation and maintenance decision scheme, and conduct a general confidence assessment of the optimal operation and maintenance decision scheme to determine the comprehensive confidence assessment result.

[0044] Preferably, based on the general feature vector and key event structured data, the core tasks are analyzed according to the type of operation and maintenance tasks to determine the optimal operation and maintenance decision-making scheme, including: RAG retrieval enhancement is performed based on the general feature vector to obtain contextual enhancement information; Based on the contextual enhancement information and key event structured data, the corresponding algorithm plugins are automatically invoked according to different operation and maintenance task types to complete the special analysis and judgment under the current operation and maintenance scenario and generate the preliminary analysis results of the corresponding scenario. Based on the preliminary assessment results, a coupling analysis of the power grid and communication network is performed. A quantitative assessment model is established based on the four-level hierarchical topology relationship of "communication equipment - network link - service bearer - power grid security". The objective risk level and the list of affected power grid services are determined based on the quantitative assessment model. The quantitative assessment model is persistently stored in the knowledge graph triple format to support bidirectional inference of forward and reverse mapping. Based on the enhanced contextual information, preliminary assessment results, objective risk levels, and a list of affected businesses, a large model is used for analysis, and the decision generation algorithm plugin for the corresponding operation and maintenance scenario performs structured constraint verification to output the optimal operation and maintenance decision solution.

[0045] Preferably, the optimal operation and maintenance decision scheme is subjected to a general confidence assessment to determine the comprehensive confidence assessment result, including: The optimal operation and maintenance decision scheme is automatically compared with the power communication operation and maintenance specifications and safety procedures to check whether the decision scheme meets the compliance requirements. When the optimal operation and maintenance decision scheme meets the compliance requirements, the comprehensive credibility assessment of the optimal operation and maintenance decision scheme is carried out through the ConfidenceNet general confidence network; Among them, a multi-dimensional input feature system is composed of multiple core confidence factors, including: alarm confidence, data quality confidence, model confidence, scenario matching confidence, expert rule matching confidence, and operating environment confidence; Based on Bayesian neural networks, uncertainty modeling is performed on multi-dimensional input features, and comprehensive confidence scores, uncertainty estimates and attribution data are output. Based on the comprehensive confidence score, uncertainty estimate, and attribution data, the comprehensive confidence assessment result is determined.

[0046] Preferably, the method further includes: Based on the feature distribution of the current operation and maintenance scenario and historical feedback data, the fusion weights of each core confidence factor are dynamically generated through an attention mechanism; wherein, the attention network uses the original input feature vector As input, output the dynamic fusion weights of the six types of factors. All weights satisfy the normalization constraint ; the weighted feature vector Input the Bayesian confidence inference engine to determine the fusion weights of each core confidence factor.

[0047] like Figure 2 As shown, in this invention, the cloud-based core decision-making intelligent agent, based on a dedicated large-scale model for power communication, receives multi-dimensional general feature vectors and key event structured data uploaded from the edge side. The RAG retrieval engine uses the multi-dimensional general feature vectors as query keys to retrieve contextual enhancement information from the general knowledge base for power communication. The large-scale model inference engine uses structured data as the main body to fuse the retrieval results to complete intelligent judgment and multi-dimensional decision generation for the entire scenario, and outputs an objective risk level through joint analysis of the power grid and communication network. After compliance verification, the decision scheme is input into the ConfidenceNet confidence network, which outputs a comprehensive confidence assessment result including the comprehensive confidence score C, uncertainty estimate σ, and attribution reports of the contribution of each factor.

[0048] This stage is the core decision-making center of the system, built upon the "General Large Model of Power Communication," such as... Figure 5 As shown, this supports unified analysis and decision generation for various operation and maintenance scenarios. The input to this stage consists of two parts, strictly corresponding to the two types of data packets output from the edge: one is a multi-dimensional general feature vector, input into the RAG retrieval engine to complete semantic similarity matching; the other is structured data of key events, used as core slots to fill in the inference prompts of the large model for scenario-based decision generation. This stage essentially completes six general steps: ① RAG Retrieval Enhancement: Using the multi-dimensional general feature vector produced in step 102 as the query key, relevant operation and maintenance procedures, historical fault cases, and expert experience knowledge in the general knowledge base of power communication are retrieved, and context enhancement information (relevant case fragments, procedure clauses, and knowledge graph association triples) is output to serve subsequent intelligent judgment in all scenarios, provide professional background knowledge for the construction of large model prompts, and avoid the general large model from producing "illusions" in power-specific scenarios; Among them, the RAG retrieval enhancement mechanism deeply integrates the exclusive knowledge system of the power communication field, specifically including the following key designs: (1) Exclusive vectorization processing of power operation and maintenance knowledge: For documents such as operation and maintenance specifications, maintenance procedures, and historical fault cases, according to the power document format structure, they are decomposed into independent semantic units according to "chapter-section-article-clause", and the structured metadata such as equipment type and business level are simultaneously labeled; the normalization of synonyms is completed based on the power standard terminology library, and the knowledge vector is generated by the vector model specially fine-tuned by the power operation and maintenance corpus, and stored in layers according to procedures, cases, and equipment mechanisms to solve the problems of professional semantic breakage and low matching accuracy caused by general block cutting. (2) Precise retrieval guided by power ontology: Construct a power operation and maintenance ontology library and knowledge graph covering five major categories of entities: equipment, fault, business, procedures, and resources, and define core association relationships such as equipment-fault, fault-handling, and business-level. The retrieval adopts a two-stage model of "ontology localization + targeted recall": first, the knowledge scope corresponding to the current event is locked through entity links, and irrelevant categories are filtered out; then, semantic matching is completed within the targeted subset, and knowledge graph reasoning is used to supplement the knowledge of rare scenarios, and the results are reordered according to the power grid business security level, which greatly improves the accuracy and business adaptability of professional knowledge recall.

[0049] ② Intelligent Analysis Across All Scenarios: The input for this step comes from two parts—the key event structured data produced in step 2 (core input, providing factual information about the current operation and maintenance event) and the contextual enhancement information output in step ① (auxiliary input, providing historical cases and procedure references). Based on different operation and maintenance task types (including but not limited to fault handling, performance degradation analysis, resource load scheduling, power supply link assurance, configuration compliance check, business quality monitoring, etc.), the corresponding algorithm plugins deployed in the cloud are automatically invoked (this plugin uses a large model as the inference engine, encapsulates exclusive inference templates and thought chains for different operation and maintenance scenarios, and drives the large model to complete the special analysis in the scenario), complete the special analysis in the current operation and maintenance scenario, generate the preliminary analysis results of the corresponding scenario, and serve ③ power grid-communication network joint analysis (providing analysis objects) and ④ multi-dimensional decision generation (providing decision basis); ③ Power Grid-Communication Network Joint Analysis: Based on the preliminary assessment results generated in ②, a coupling analysis of the power grid and communication network is performed, specifically including: Forward mapping: Communication link anomaly → trace the power grid services it carries → locate the affected primary power grid equipment (such as relay protection and stability control devices), and quantify the scope and severity of the service impact; Reverse mapping: Changes in power grid operating conditions (such as peak loads and power supply guarantee tasks) → Increase the monitoring priority of corresponding communication services → Dynamically adjust the scheduling weight of operation and maintenance resources.

[0050] This invention establishes a quantitative assessment model based on a four-level hierarchical topology relationship of "communication equipment - network link - service bearer - power grid security". The quantitative assessment model is persistently stored in the format of knowledge graph triples (head entity, relation, tail entity) to support bidirectional deduction of forward mapping (communication anomaly → service impact → power grid risk) and reverse mapping (power grid risk → communication tracing → protection and hardening).

[0051] Output objective risk levels (high risk / medium risk / low risk) and a list of affected power grid businesses with priority adjustment suggestions, serving ④ multi-dimensional decision generation (providing decision weight basis, prioritizing core power grid businesses in high-risk scenarios).

[0052] ④ Multi-dimensional decision generation: This step integrates three input sources—① enhanced contextual information from the output, ② preliminary assessment results from the output, and ③ the objective risk level and list of affected businesses from the output. Combining the current operating conditions of the power grid, the availability of maintenance resources, and task requirements, an initial handling suggestion is generated using a large model as a general inference engine. Then, the decision generation algorithm plugin for the corresponding maintenance scenario completes the structured constraint verification and outputs the optimal maintenance decision solution (including specific handling measures, execution priority, time window, and resource allocation), serving ⑤ automatic compliance verification (as the object to be verified). ⑤ Automatic Compliance Verification: The input to this step comes from the optimal operation and maintenance decision scheme output from step ④. This scheme is automatically compared with power communication operation and maintenance specifications and safety procedures to check whether the decision scheme meets compliance requirements. The output includes a compliance verification pass indicator (pass / pass with warning / fail) and a verification log summary. Schemes that pass verification proceed to step ⑥ ConfidenceNet for confidence assessment; schemes that fail verification are returned to step ④ with a list of non-compliance items, where they are regenerated by the decision generation step without violating compliance constraints. ⑥ General Confidence Assessment: The input for this step comes from two parts—⑤ the optimal operation and maintenance decision plan that has passed verification (the object to be evaluated) and the metadata of each step in the entire process (including the quantitative values ​​of six confidence factors such as alarm historical accuracy, source data quality labels, model historical performance statistics, scenario feature distribution, expert rule consistency score, and edge node operating environment status). Through the ConfidenceNet general confidence network, a comprehensive confidence assessment of the optimal operation and maintenance decision plan is completed.

[0053] The ConfidenceNet general confidence network adopts a multi-dimensional dynamic weighted confidence evaluation architecture based on Bayesian Neural Network (BNN), and its specific components and calculation logic are as follows: (1) Multi-dimensional input feature system: including six core confidence factors—① Alarm confidence (historical alarm accuracy, false alarm rate, false alarm rate); ② Data quality confidence (source data integrity rate, missing value ratio, outlier ratio, timestamp accuracy); ③ Model confidence (historical model prediction accuracy, F1-score, area under the AUC curve); ④ Scene matching confidence (cosine similarity between the current scene and historical similar scenes, KL divergence of feature distribution); ⑤ Expert rule matching confidence (consistency score between the judgment result and the power communication operation and maintenance procedures and historical cases); ⑥ Operating environment confidence (edge ​​node computing power utilization rate, network transmission delay, sensor health status); the six factors together constitute the input feature vector. This provides full-dimensional quantitative input for confidence calculation.

[0054] (2) Bayesian confidence inference engine: Based on Bayesian neural network, it models the uncertainty of input features, performs after-a priori prediction after Monte Carlo sampling, and simultaneously outputs two core quantitative indicators: Overall confidence score (mean of prediction distribution): in For the number of samples taken in Monte Carlo, This score represents the confidence prediction value corresponding to a single weighted sampling, and it provides the core basis for subsequent decision-making and triage.

[0055] Uncertainty estimation (standard deviation of the predicted distribution): A larger standard deviation indicates a lower reliability of the model's assessment of the current scenario. Confidence intervals can be generated based on the mean and standard deviation, providing a reference for determining decision boundaries. This avoids the overconfidence problem of traditional point estimation in unknown scenarios and is suitable for the high reliability requirements of power operation and maintenance.

[0056] (3) Dynamic weight adaptive adjustment: The fusion weights of each confidence factor are not fixed and preset, but are dynamically generated based on the feature distribution of the current operation and maintenance scenario and historical feedback data through the attention mechanism. The attention network uses the original input feature vector As input, output the dynamic fusion weights of the six types of factors. All weights satisfy the normalization constraint ; the weighted feature vector Input the Bayesian confidence inference engine to achieve adaptive adjustment of automatic increase of model confidence weight in high data quality scenarios and automatic increase of expert rule confidence weight in low data quality scenarios; (4) Explainable confidence attribution data output: ConfidenceNet outputs the comprehensive confidence score C and uncertainty estimate. Simultaneously, interpretable attribution data for the contribution of each confidence factor is output. This attribution data is a structured set of fields, including: ① dynamic fusion weights of each factor. ② Weighted contribution of each factor ③ Low confidence factor labeling (automatically marking factors whose contribution ranking is at the bottom and below the preset contribution threshold).

[0057] The above attribution data, combined confidence score C, and uncertainty estimation Together, these data constitute a complete comprehensive confidence assessment result data package, which is then output to the decision-making process. Its core value lies in the following: During decision-making, if the low-confidence bottleneck factor involves "data quality confidence" or "operational environment confidence," the system automatically lowers the dynamic threshold compensation coefficient for the corresponding scenario, prioritizing manual review; during manual review, the maintenance interface highlights the bottleneck factor and its contribution, assisting personnel in quickly locating the root cause of unreliable decisions; during closed-loop optimization, attribution data serves as a weighted label for feedback samples, driving the model to perform targeted incremental training on low-confidence bottleneck factors.

[0058] In step 104, decision routing is performed based on the comprehensive confidence assessment results and the optimal operation and maintenance decision scheme, and operation and maintenance are performed based on the decision routing results.

[0059] Preferably, the decision-making process based on the comprehensive confidence assessment result and the optimal operation and maintenance decision scheme includes: Based on the comprehensive confidence assessment results, dynamic thresholds are determined to complete the decision-making process, including: matching scenario-based dynamic thresholds based on the optical cable carrying service level, power grid operation scenario, and risk level to divide the assessment results. The automated processing flow is executed. When the confidence level is greater than or equal to the scenario dynamic threshold and it is not a high-risk scenario, the specific measures in the optimal operation and maintenance decision scheme are used as the execution instructions. The existing operation and maintenance system is directly connected through the standardized API interface to complete the fully automated execution of the process, including: automatic push of graded early warning information to the corresponding operation and maintenance personnel, automatic generation and circulation of maintenance work orders, connection of maintenance plan system and optimization of production plan, and full tracking and data feedback of operation and maintenance handling process. The manual review process is executed. When the confidence level is less than the scene dynamic threshold or it is a high-risk scene, the full analysis results and the optimal operation and maintenance decision plan are pushed to professional operation and maintenance personnel for manual review and optimization, and the review opinions and corrected data are output. Output the results of the implementation of operation and maintenance decisions, the data for manual review and correction, the feedback data of the entire process, and the comprehensive confidence assessment results.

[0060] Preferably, the method further includes: A dynamic threshold is determined based on preset constraints; wherein, the preset constraints include: Lower bound constraint: The minimum confidence threshold for automated execution branches in any scenario must not be lower than 0.75; Scenario-based risk weighted constraints: Dynamic thresholds are positively correlated with scenario risk levels—power supply scenario threshold >= 0.95, routine maintenance scenario threshold >= 0.80, and daily monitoring scenario threshold >= 0.75; Data quality constraints: When the source data integrity rate is below 85% or the key feature missing rate exceeds 10%, the threshold will automatically increase by 0.05-0.10.

[0061] Combination Figure 2 As shown, in this invention, the decision-making process is streamlined, and a two-dimensional combination matrix of comprehensive confidence score C (0~1 range, representing the model's confidence in the correctness of the current decision) and objective risk level is used to make the following judgments: high confidence (C≥dynamic threshold) and low risk (non-power supply / non-security control scenarios) are transferred to automated execution; low confidence (C<dynamic threshold) or high risk (power supply / security control scenarios) are transferred to manual review.

[0062] This stage is the core of the system's decision-making process, such as... Figure 6 As shown, its core positioning is to ensure that the input plan has passed the compliance verification, use the comprehensive confidence score C and the objective risk level as the basis for the diversion decision, use the optimal operation and maintenance decision plan as the core object of execution or review, and use the preliminary judgment conclusion and interpretable attribution data as the background support for manual review to complete the diversion and handling of the decision plan, achieve seamless linkage with the existing power operation and maintenance system, and complete the automated implementation or manual review and optimization of operation and maintenance decisions.

[0063] The input to this stage consists of three types of data packets, which are the output of step three: ① Full assessment results (including preliminary assessment conclusions, objective risk levels, and optimal operation and maintenance decision-making schemes); ② Overall confidence assessment results (including confidence score C, uncertainty estimate σ, and interpretable attribution data); ③ Compliance verification passed indicator.

[0064] Among them, compliance verification passes the identification as a prerequisite for traffic diversion (only those in the "pass" status enter the subsequent traffic diversion process); the comprehensive confidence score C and the objective risk level together constitute the two-dimensional basis for traffic diversion decision; the optimal operation and maintenance decision plan provides specific instructions for automated execution branches, and at the same time provides review objects for manual review branches; interpretable attribution data and preliminary judgment conclusions together provide background information and interpretable guidance for manual review.

[0065] Based on the above input, the core routing and execution process is divided into three parts: First, dynamic threshold judgment: based on the optical cable carrying service level, power grid operation scenario, and risk level, scenario-based dynamic thresholds are matched to complete the triage judgment of the analysis results. Second, automated execution branch: when the confidence level is greater than or equal to the scenario dynamic threshold and it is not a high-risk scenario, the specific measures (handling instructions, maintenance plan content, resource allocation plan, etc.) in the optimal operation and maintenance decision plan are used as execution instructions. The existing operation and maintenance system is directly connected through a standardized API interface to complete the fully automated execution of the process, including automatic push of graded early warning information to the corresponding operation and maintenance personnel, automatic generation and circulation of maintenance work orders, connection of maintenance plan system and optimization of production plan, and full tracking and data feedback of the operation and maintenance handling process. After automated execution, the optimal operation and maintenance decision plan is archived along with the execution results in step five. The third step is manual review. When the confidence level is insufficient or a high-risk scenario is identified, the full analysis results and the optimal operation and maintenance decision plan are pushed to professional operation and maintenance personnel for manual review and optimization. Review opinions and corrected data are output, and the original optimal operation and maintenance decision plan and the corrected decision plan are input as feedback data into the closed-loop self-optimization stage in step five. Finally, the output includes the implementation results of the operation and maintenance decisions, the manually reviewed and corrected data, the feedback data from the entire process, and the comprehensive confidence level assessment results.

[0066] The core design principle of the dynamic threshold judgment mechanism is as follows: the scenario-based dynamic threshold is not preset by a fixed mapping table, but is generated in real time by the cloud core decision-making big model based on the dynamic risk profile of the current power communication operation and maintenance scenario. The specific constraints include: (1) lower limit constraint: the minimum confidence threshold of the automated execution branch in any scenario shall not be lower than 0.75, so as to ensure the basic reliability of automated decision-making; (2) scenario risk weighting constraint: the dynamic threshold is positively correlated with the scenario risk level - the threshold for power supply scenario >= 0.95, the threshold for routine maintenance scenario >= 0.80, and the threshold for daily monitoring scenario >= 0.75. The higher the risk level, the stricter the threshold; (3) data quality constraint: when the source data integrity rate is lower than 85% or the key feature missing rate exceeds 10%, the threshold is automatically increased by 0.05-0.10 to compensate for the uncertainty of data quality with stricter confidence requirements.

[0067] Preferably, the method further includes: Based on the implementation results of operation and maintenance decisions, manual review and correction of data, feedback data from the entire process of handling, and comprehensive confidence assessment results, the lightweight small model corresponding to the pluggable algorithm plugin in the edge-side perception agent and the cloud-dedicated large model are collaboratively optimized, and the knowledge base is updated to complete the closed-loop optimization of the model and the knowledge base.

[0068] like Figure 2 As shown, in this invention, the closed-loop optimization of the model and knowledge base is finally completed based on the execution feedback data, thereby realizing the continuous self-evolution of the system capabilities.

[0069] This stage serves as the closed-loop self-optimization hub of the entire intelligent agent system. Its core function is to leverage feedback data from the execution stage to perform end-to-end iterative optimization of the model and knowledge base, enabling continuous self-learning and self-evolution of the system. Specifically, such as... Figure 7 As shown, the input to this step is the four types of data output from step 104: ① the implementation results of operation and maintenance decisions, ② the data of manual review and correction, ③ the feedback data of the entire process, and ④ the comprehensive confidence assessment results (including confidence score C, uncertainty estimate σ, and attribution data of the contribution of each factor, which provide a quantitative basis for the weight allocation of training samples in the iterative training of large and small models). The core optimization process in this stage consists of four steps: First, feedback data collection, which involves the standardized collection and aggregation of full-processing feedback data, manually corrected data, and actual fault and maintenance data; second, collaborative optimization of large and small models, which involves fine-tuning and iterating the lightweight small models corresponding to pluggable algorithm plugins in the edge-side perception agent based on feedback data to improve the accuracy of edge-side feature extraction and preliminary judgment, while incrementally optimizing the dedicated large models of the cloud-based core decision-making agent based on incremental feedback data to improve the scenario adaptability of the large model's judgment and decision-making; third, incremental updates of the knowledge base, which involves incrementally updating the historical case library, maintenance specification knowledge base, and knowledge graph based on actual maintenance cases, processing results, and manually corrected data to continuously optimize the retrieval effect of RAG; and fourth, feedback of optimization results, which involves synchronously feeding back the iteratively optimized small models, large models, and knowledge base to the preceding perception agent and core decision-making agent stages to complete the full-link closed-loop optimization of the entire system.

[0070] The power communication network operation and maintenance method provided by this invention can solve the following technical problems: To address the issues of fragmented operation and maintenance scenarios and siloed systems in existing technologies, a unified general intelligent agent closed-loop architecture is constructed to achieve unified data access, unified algorithm scheduling, unified decision execution, and unified closed-loop optimization across all operation and maintenance scenarios. To address the issues of low intelligence and heavy reliance on manual labor in existing technologies, large-scale model technology and pluggable algorithm engines are introduced to achieve autonomous and intelligent management and control of the entire operation and maintenance process, significantly improving operation and maintenance efficiency and response speed. To address the problem of existing technologies being disconnected from power grid operations and unable to support targeted operation and maintenance, a hierarchical linkage mechanism of "communication network - power grid operations - security risks" is established to achieve deep integration of operation and maintenance decisions with power grid security requirements; To address the issues of excessive duplication and poor scalability in existing technologies, we designed universal and standardized module interfaces to support rapid expansion across various operation and maintenance scenarios and on-demand loading of algorithm models, thereby significantly reducing the cost of intelligent upgrades.

[0071] The method of this invention can be deployed in a three-tier architecture of State Grid Control Cloud, provincial communication cloud, and municipal edge nodes, seamlessly connecting with existing production systems such as TMS transmission network management system, communication integrated network management system, dispatch data network management system, and communication power monitoring system. It is applicable to seven core operation and maintenance scenarios in power backbone communication network, distribution network communication system, and power Internet of Things, including equipment lifecycle management, network anomaly detection, intelligent fault handling, business impact assessment, intelligent resource scheduling, emergency power supply support, and automatic compliance verification. It has engineering promotion value at the network-wide and in all scenarios.

[0072] Taking the intelligent handling of optical cable faults in power communication networks as an example, the specific implementation process of this invention is as follows: (1) Multi-source data access: Through standardized API interface, connect to TMS transmission network management system to collect real-time operating data such as optical cable power, bit error rate, and alarm information, and at the same time access the power grid dispatch system to obtain carrying service information and power grid operation status data; (2) Edge perception processing: The perception agent deployed at the edge nodes of the city loads the anomaly detection algorithm plugin to monitor and extract features of the optical cable operation data in real time. When an abnormal degradation of optical power is detected, the fault feature vector is extracted and encrypted and uploaded to the core decision-making agent in the cloud. (3) Cloud-based decision-making and judgment: The cloud-based core decision-making intelligent agent uses RAG to search and match historical handling cases and operation and maintenance procedures of similar optical cable faults, calls the fault root cause analysis algorithm plug-in to locate the fault cause, and conducts joint analysis and evaluation of the impact range and risk level of the fault on core businesses such as relay protection and stability control through the power grid-communication network. It generates a complete decision plan including fault location, impact assessment and handling suggestions, and automatically performs compliance verification and confidence assessment. (4) Decision-making and execution: Based on the confidence assessment results and risk level, if it is a routine fault and the confidence level is higher than the threshold, a maintenance work order will be automatically generated and pushed to the corresponding maintenance personnel, and the maintenance plan will be connected simultaneously; if it is a high-risk fault or the confidence level is insufficient, the full assessment results will be pushed to professional maintenance personnel for manual review. (5) Closed-loop self-optimization: After the fault handling is completed, the handling results, manually corrected data and actual fault data are fed back to the closed-loop self-optimization link to complete the fine-tuning and optimization of the edge small model and the cloud large model. At the same time, the fault case is added to the knowledge base to continuously improve the fault judgment and handling capabilities of the system.

[0073] Figure 8 This is a schematic diagram of the structure of a power communication network operation and maintenance management system 800 according to an embodiment of the present invention. Figure 8 As shown, the power communication network operation and maintenance management system 800 provided in this embodiment of the invention includes: a data acquisition unit 801, an edge sensing unit 802, a core decision-making unit 803, and a decision-distribution unit 804.

[0074] Preferably, the data acquisition unit 801 is used to acquire multi-source operation and maintenance data output by each production system in the power communication network, and to preprocess the multi-source operation and maintenance data to obtain a multi-source standardized raw dataset.

[0075] Preferably, the data acquisition unit 801 preprocesses the multi-source operation and maintenance data, including: The multi-source operation and maintenance data are processed for format unification, time sequence alignment, metadata annotation, and quality verification.

[0076] Preferably, the edge sensing unit 802 is used to perform edge computing based on the multi-source standardized raw dataset using the edge sensing agent, to obtain general feature vectors and key event structured data corresponding to various operation and maintenance tasks, and to upload the general feature vectors and key event structured data to the cloud core decision agent.

[0077] Preferably, the edge sensing unit 802 is further configured to: Before using the edge-aware agent to perform edge computing based on the multi-source standardized raw dataset, the multi-source standardized raw dataset is subjected to dynamic outlier removal, intelligent missing value completion, data smoothing and noise reduction, and normalization and standardization processing in sequence.

[0078] Preferably, the edge sensing unit 802 is further configured to: Based on the edge-aware intelligent agent, pluggable lightweight algorithm plugins are invoked. The loaded algorithm plugins complete the front-end perception, feature extraction, and low-latency local execution of various operation and maintenance tasks, obtaining the general feature vectors and key event structured data. The pluggable lightweight algorithm plugins are modular components for the edge-aware intelligent agent to realize perception and local processing. They are loaded or unloaded as needed according to the computing power conditions and operation and maintenance task requirements of each edge node, so as to support the flexible expansion of operation and maintenance capabilities in all scenarios.

[0079] Preferably, the edge sensing unit 802 is further configured to: Local low-latency processing is performed on the aforementioned multi-source standardized original dataset to obtain local low-latency handling execution logs, which are used for local fault tracing and evidence collection and offline self-supervised learning training samples for edge small models. The local low-latency handling execution logs include: records of control command issuance triggered autonomously by edge nodes based on local judgment and local anomaly marking records. The control command issuance records include: action type, execution time, target device, and execution result status code. The local anomaly marking records include: anomaly type, feature value, and a summary of the judgment criteria.

[0080] Preferably, the core decision-making unit 803 is used by the cloud-based core decision-making intelligent agent to utilize a dedicated large model for power communication, based on the general feature vector and key event structured data, to perform core analysis according to the type of operation and maintenance task, determine the optimal operation and maintenance decision scheme, and perform a general confidence assessment on the optimal operation and maintenance decision scheme to determine the comprehensive confidence assessment result.

[0081] Preferably, the core decision-making unit 803, based on the general feature vector and key event structured data, analyzes the core tasks according to the type of operation and maintenance tasks, and determines the optimal operation and maintenance decision-making scheme, including: RAG retrieval enhancement is performed based on the general feature vector to obtain contextual enhancement information; Based on the contextual enhancement information and key event structured data, the corresponding algorithm plugins are automatically invoked according to different operation and maintenance task types to complete the special analysis and judgment under the current operation and maintenance scenario and generate the preliminary analysis results of the corresponding scenario. Based on the preliminary assessment results, a coupling analysis of the power grid and communication network is performed. A quantitative assessment model is established based on the four-level hierarchical topology relationship of "communication equipment - network link - service bearer - power grid security". The objective risk level and the list of affected power grid services are determined based on the quantitative assessment model. The quantitative assessment model is persistently stored in the knowledge graph triple format to support bidirectional inference of forward and reverse mapping. Based on the enhanced contextual information, preliminary assessment results, objective risk levels, and a list of affected businesses, a large model is used for analysis, and the decision generation algorithm plugin for the corresponding operation and maintenance scenario performs structured constraint verification to output the optimal operation and maintenance decision solution.

[0082] Preferably, the core decision-making unit 803 performs a general confidence assessment on the optimal operation and maintenance decision scheme to determine the comprehensive confidence assessment result, including: The optimal operation and maintenance decision scheme is automatically compared with the power communication operation and maintenance specifications and safety procedures to check whether the decision scheme meets the compliance requirements. When the optimal operation and maintenance decision scheme meets the compliance requirements, the comprehensive credibility assessment of the optimal operation and maintenance decision scheme is carried out through the ConfidenceNet general confidence network; Among them, a multi-dimensional input feature system is composed of multiple core confidence factors, including: alarm confidence, data quality confidence, model confidence, scenario matching confidence, expert rule matching confidence, and operating environment confidence; Based on Bayesian neural networks, uncertainty modeling is performed on multi-dimensional input features, and comprehensive confidence scores, uncertainty estimates and attribution data are output. Based on the comprehensive confidence score, uncertainty estimate, and attribution data, the comprehensive confidence assessment result is determined.

[0083] Preferably, the core decision-making unit 803 further includes: Based on the feature distribution of the current operation and maintenance scenario and historical feedback data, the fusion weights of each core confidence factor are dynamically generated through an attention mechanism; wherein, the attention network uses the original input feature vector As input, output the dynamic fusion weights of the six types of factors. All weights satisfy the normalization constraint ; the weighted feature vector Input the Bayesian confidence inference engine to determine the fusion weights of each core confidence factor.

[0084] Preferably, the decision-making diversion unit 804 is used to perform decision diversion based on the comprehensive confidence assessment result and the optimal operation and maintenance decision scheme, and to perform operation and maintenance based on the decision diversion result.

[0085] Preferably, the decision-making triage unit 804 performs decision triage based on the comprehensive confidence assessment result and the optimal operation and maintenance decision scheme, including: Based on the comprehensive confidence assessment results, dynamic thresholds are determined to complete the decision-making process, including: matching scenario-based dynamic thresholds based on the optical cable carrying service level, power grid operation scenario, and risk level to divide the assessment results. The automated processing flow is executed. When the confidence level is greater than or equal to the scenario dynamic threshold and it is not a high-risk scenario, the specific measures in the optimal operation and maintenance decision scheme are used as the execution instructions. The existing operation and maintenance system is directly connected through the standardized API interface to complete the fully automated execution of the process, including: automatic push of graded early warning information to the corresponding operation and maintenance personnel, automatic generation and circulation of maintenance work orders, connection of maintenance plan system and optimization of production plan, and full tracking and data feedback of operation and maintenance handling process. The manual review process is executed. When the confidence level is less than the scene dynamic threshold or it is a high-risk scene, the full analysis results and the optimal operation and maintenance decision plan are pushed to professional operation and maintenance personnel for manual review and optimization, and the review opinions and corrected data are output. Output the results of the implementation of operation and maintenance decisions, the data for manual review and correction, the feedback data of the entire process, and the comprehensive confidence assessment results.

[0086] Preferably, the decision-making diversion unit 804 is further configured to: A dynamic threshold is determined based on preset constraints; wherein, the preset constraints include: Lower bound constraint: The minimum confidence threshold for automated execution branches in any scenario must not be lower than 0.75; Scenario-based risk weighted constraints: Dynamic thresholds are positively correlated with scenario risk levels—power supply scenario threshold >= 0.95, routine maintenance scenario threshold >= 0.80, and daily monitoring scenario threshold >= 0.75; Data quality constraints: When the source data integrity rate is below 85% or the key feature missing rate exceeds 10%, the threshold will automatically increase by 0.05-0.10.

[0087] Preferably, the system further includes: The closed-loop optimization unit is used to collaboratively optimize the lightweight small models and cloud-dedicated large models corresponding to pluggable algorithm plugins in the edge-side perception agent, based on the implementation results of operation and maintenance decisions, manual review and correction data, feedback data of the whole process handling process, and comprehensive confidence assessment results.

[0088] The power communication network operation and maintenance management system 800 of this invention corresponds to the power communication network operation and maintenance management method 100 of another embodiment of this invention, and will not be described again here.

[0089] The present invention has been described with reference to a few embodiments. However, it will be apparent to those skilled in the art that other embodiments besides those disclosed above fall equivalently within the scope of the present invention.

[0090] Generally, all terms used in this invention are interpreted according to their ordinary meaning in the art, unless otherwise expressly defined herein. All references to “a / the / the [device, component, etc.]” ​​are openly interpreted as at least one instance of said device, component, etc., unless otherwise expressly stated. The steps of any method disclosed herein need not be performed in the exact order disclosed, unless explicitly stated otherwise.

[0091] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0093] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0094] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A power communication network operation and maintenance management method based on a large model and agent-driven approach, characterized in that, The method includes: Acquire multi-source operation and maintenance data output by various production systems in the power communication network, and preprocess the multi-source operation and maintenance data to obtain a multi-source standardized raw dataset; The edge-aware agent performs edge computing based on the multi-source standardized raw dataset to obtain general feature vectors and key event structured data corresponding to various operation and maintenance tasks, and uploads the general feature vectors and key event structured data to the cloud core decision-making agent; The cloud-based core decision-making intelligent agent utilizes a dedicated large model for power communication, based on the general feature vectors and key event structured data, to perform core analysis according to the type of operation and maintenance task, determine the optimal operation and maintenance decision scheme, and conduct a general confidence assessment of the optimal operation and maintenance decision scheme to determine the comprehensive confidence assessment result. Decision routing is performed based on the comprehensive confidence assessment results and the optimal operation and maintenance decision scheme, and operation and maintenance are carried out based on the decision routing results.

2. The method according to claim 1, characterized in that, The multi-source operation and maintenance data includes: basic resource static data, real-time dynamic operation data, and operation and maintenance knowledge and management data.

3. The method according to claim 1, characterized in that, Preprocessing of the multi-source operation and maintenance data includes: The multi-source operation and maintenance data are processed for format unification, time sequence alignment, metadata annotation, and quality verification.

4. The method according to claim 1, characterized in that, The method further includes: before using the edge-aware agent to perform edge computing based on the multi-source standardized original dataset, performing dynamic outlier removal, intelligent missing value completion, data smoothing and noise reduction, and normalization and standardization processing on the multi-source standardized original dataset in sequence.

5. The method according to claim 1, characterized in that, The method further includes: By utilizing an edge-aware agent to invoke a pluggable lightweight algorithm plugin, local low-latency processing is performed on the aforementioned multi-source standardized original dataset to obtain local low-latency handling execution logs. These logs are used for local fault tracing and evidence collection, as well as offline self-supervised learning training samples for edge mini-models. The local low-latency handling execution logs include: records of control command issuance autonomously triggered by edge nodes based on local judgment, and local anomaly marking records. The control command issuance records include: action type, execution time, target device, and execution result status code. The local anomaly marking records include: anomaly type, feature value, and a summary of the judgment criteria.

6. The method according to claim 1, characterized in that, Based on the aforementioned general feature vectors and structured data of key events, the core tasks are analyzed according to the type of operation and maintenance tasks to determine the optimal operation and maintenance decision-making scheme, including: RAG retrieval enhancement is performed based on the general feature vector to obtain contextual enhancement information; Based on the contextual enhancement information and key event structured data, the corresponding algorithm plugins are automatically invoked according to different operation and maintenance task types to complete the special analysis and judgment under the current operation and maintenance scenario and generate the preliminary analysis results of the corresponding scenario. Based on the preliminary assessment results, a coupling analysis of the power grid and communication network is performed. A quantitative assessment model is established based on the four-level hierarchical topology relationship of "communication equipment - network link - service bearer - power grid security". The objective risk level and the list of affected power grid services are determined based on the quantitative assessment model. The quantitative assessment model is persistently stored in the knowledge graph triple format to support bidirectional inference of forward and reverse mapping. Based on the enhanced contextual information, preliminary assessment results, objective risk levels, and a list of affected businesses, a large model is used for analysis, and the decision generation algorithm plugin for the corresponding operation and maintenance scenario performs structured constraint verification to output the optimal operation and maintenance decision solution.

7. The method according to claim 1, characterized in that, A general confidence assessment is performed on the optimal operation and maintenance decision scheme to determine the comprehensive confidence assessment result, including: The optimal operation and maintenance decision scheme is automatically compared with the power communication operation and maintenance specifications and safety procedures to check whether the decision scheme meets the compliance requirements. When the optimal operation and maintenance decision scheme meets the compliance requirements, the comprehensive credibility assessment of the optimal operation and maintenance decision scheme is carried out through the ConfidenceNet general confidence network; Among them, a multi-dimensional input feature system is composed of multiple core confidence factors, including: alarm confidence, data quality confidence, model confidence, scenario matching confidence, expert rule matching confidence, and operating environment confidence; Based on Bayesian neural networks, uncertainty modeling is performed on multi-dimensional input features, and comprehensive confidence scores, uncertainty estimates and attribution data are output. Based on the comprehensive confidence score, uncertainty estimate, and attribution data, the comprehensive confidence assessment result is determined.

8. The method according to claim 7, characterized in that, The method further includes: Based on the feature distribution of the current operation and maintenance scenario and historical feedback data, the fusion weights of each core confidence factor are dynamically generated through an attention mechanism; wherein, the attention network uses the original input feature vector As input, output the dynamic fusion weights of the six types of factors. All weights satisfy the normalization constraint ; the weighted feature vector Input the Bayesian confidence inference engine to determine the fusion weights of each core confidence factor.

9. The method according to claim 1, characterized in that, Decision routing is performed based on the comprehensive confidence assessment results and the optimal operation and maintenance decision scheme, including: Based on the comprehensive confidence assessment results, dynamic thresholds are determined to complete the decision-making process, including: matching scenario-based dynamic thresholds based on the optical cable carrying service level, power grid operation scenario, and risk level to divide the assessment results. The automated processing flow is executed. When the confidence level is greater than or equal to the scenario dynamic threshold and it is not a high-risk scenario, the specific measures in the optimal operation and maintenance decision scheme are used as the execution instructions. The existing operation and maintenance system is directly connected through the standardized API interface to complete the fully automated execution of the process, including: automatic push of graded early warning information to the corresponding operation and maintenance personnel, automatic generation and circulation of maintenance work orders, connection of maintenance plan system and optimization of production plan, and full tracking and data feedback of operation and maintenance handling process. The manual review process is executed. When the confidence level is less than the scene dynamic threshold or it is a high-risk scene, the full analysis results and the optimal operation and maintenance decision plan are pushed to professional operation and maintenance personnel for manual review and optimization, and the review opinions and corrected data are output. Output the results of the implementation of operation and maintenance decisions, the data for manual review and correction, the feedback data of the entire process, and the comprehensive confidence assessment results.

10. The method according to claim 9, characterized in that, The method further includes: A dynamic threshold is determined based on preset constraints; wherein, the preset constraints include: Lower bound constraint: The minimum confidence threshold for automated execution branches in any scenario must not be lower than 0.75; Scenario-based risk weighted constraints: Dynamic thresholds are positively correlated with scenario risk levels—power supply scenario threshold >= 0.95, routine maintenance scenario threshold >= 0.80, and daily monitoring scenario threshold >= 0.75; Data quality constraints: When the source data integrity rate is below 85% or the key feature missing rate exceeds 10%, the threshold will automatically increase by 0.05-0.

10.

11. The method according to claim 10, characterized in that, The method further includes: Based on the implementation results of operation and maintenance decisions, manual review and correction of data, feedback data from the entire process of handling, and comprehensive confidence assessment results, the lightweight small model corresponding to the pluggable algorithm plugin in the edge-side perception agent and the cloud-dedicated large model are collaboratively optimized, and the knowledge base is updated to complete the closed-loop optimization of the model and the knowledge base.

12. A power communication network operation and maintenance management system, characterized in that, The system includes: The data acquisition unit is used to acquire multi-source operation and maintenance data output by various production systems in the power communication network, and to preprocess the multi-source operation and maintenance data to obtain a multi-source standardized raw dataset. The edge sensing unit is used to perform edge computing based on the multi-source standardized raw dataset using the edge sensing agent, to obtain general feature vectors and key event structured data corresponding to various operation and maintenance tasks, and to upload the general feature vectors and key event structured data to the cloud core decision-making agent; The core decision-making unit is used by the cloud-based core decision-making intelligent agent to utilize a dedicated large model for power communication, based on the general feature vector and key event structured data, to perform core analysis according to the type of operation and maintenance task, determine the optimal operation and maintenance decision scheme, and perform a general confidence assessment on the optimal operation and maintenance decision scheme to determine the comprehensive confidence assessment result. The decision-making and routing unit is used to perform decision-making and routing based on the comprehensive confidence assessment results and the optimal operation and maintenance decision scheme, and to perform operation and maintenance based on the decision-making and routing results.

Images