A method and system for tracking communication network operation and maintenance faults

Abstract

The application discloses a communication network operation and maintenance fault tracking method and system: a communication network operation and maintenance knowledge graph is constructed; a graph algorithm and communication field rules are used to optimize the knowledge graph; a PCST algorithm is used to generate an optimized subgraph with a correlation greater than a first threshold with a fault query, and the optimized subgraph is converted into a first natural language text; a large language model is used to judge the integrity of the first natural language text; when it is judged to be incomplete, entities and relationships that are not in the optimized subgraph are extracted from the first natural language text as supplementary query words; based on the supplementary query words, a supplementary subgraph is iteratively generated on the optimized knowledge graph to obtain a complete natural language text; a large language model is used to reason about the fault of the complete natural language text to obtain a reasoning answer; and four-layer reverse traceability links are constructed according to the reasoning answer, the first natural language text and a second natural language text. The accuracy of communication network fault tracking is improved.

Description

Technical Field

[0001] This application relates to the field of operation and maintenance management, and in particular to a method and system for tracing communication network operation and maintenance faults. Background Technology

[0002] With the rapid development of 5G networks, the Internet of Things (IoT), and the Industrial Internet, communication networks, as the core infrastructure supporting the digital economy, are experiencing exponential growth in scale and complexity. Current communication networks face operational challenges, including the deployment of over a million base stations, the coexistence of diverse equipment types (switches, routers, transmission equipment, etc.), and increasingly complex service types (real-time services such as high-definition video, telemedicine, and autonomous driving). This places higher demands on the accuracy and efficiency of fault tracing and location.

[0003] However, existing methods for tracing faults in communication network operation and maintenance generally suffer from the following technical shortcomings:

[0004] First, incomplete knowledge graph data leads to low fault location accuracy. Traditional methods for constructing knowledge graphs often use static data storage, lacking comprehensive integration of multi-dimensional operational data such as device status data, link performance data, and alarm log data. This results in a large number of isolated nodes, redundant relationships, and abnormal data in the knowledge graph, failing to accurately reflect the true relationships between network devices. Simultaneously, existing methods lack professional rule constraints specific to the communications field (such as device calibration cycle rules, alarm priority rules, and link latency rules), making it difficult to identify and filter invalid or erroneous information in the knowledge graph, thus affecting the accuracy of subsequent fault reasoning.

[0005] Secondly, the fault reasoning results lack credibility assessment and evidence tracing mechanisms. Most existing knowledge graph-based fault location methods employ a one-way reasoning process, only outputting the root cause of the fault, but failing to provide a complete evidence chain and data source tracing to support the result. This makes it difficult for operations and maintenance personnel to judge the reliability of the reasoning answer. Especially in complex fault scenarios (such as cascading faults across devices and regions), the lack of quantitative assessment of the completeness of the evidence path, the validity of the data source, and the accuracy of the tracing relationship leads to insufficient credibility of the fault location results, increasing the risk of operational decisions.

[0006] Furthermore, knowledge graphs lack dynamic optimization and closed-loop feedback mechanisms. In the fault tracing process, traditional methods cannot automatically identify the specific location and type of missing information when the knowledge graph data is incomplete or the confidence of the inference results is low. They also cannot supplement and update the knowledge graph in a targeted manner, resulting in repeated troubleshooting by users when similar faults occur repeatedly, leading to low operational efficiency and low knowledge reuse rate.

[0007] Therefore, there is an urgent need for a communication network operation and maintenance fault tracing method that can achieve knowledge graph quality optimization, reliable quantification of fault reasoning, and dynamic knowledge completion and updating, so as to improve the accuracy and reliability of fault location. Summary of the Invention

[0008] To address the issue of low fault location accuracy caused by incomplete knowledge graph data in the existing communication network operation and maintenance fault tracing process, this application provides a method and system for tracing communication network operation and maintenance faults, thereby improving the accuracy of communication network fault tracing.

[0009] One aspect of this application provides a method for tracking communication network operation and maintenance faults, comprising: S1, acquiring full-dimensional data on communication network operation and maintenance, and constructing a communication network operation and maintenance knowledge graph containing entity sets, relation sets, and attribute sets; S2, optimizing the knowledge graph using graph algorithms and communication domain rules; based on the optimized knowledge graph and a fault query input by the user, generating an optimized subgraph with a relevance greater than a first threshold to the fault query using the PCST algorithm, and converting the optimized subgraph into first natural language text; S3, performing a completeness judgment on the first natural language text using a large language model; when the judgment is incomplete, extracting data from the first natural language text. Entities and relationships not appearing in the optimized subgraph are taken as supplementary query terms; supplementary subgraphs are iteratively generated on the optimized knowledge graph based on the supplementary query terms, and the supplementary subgraphs are converted into second natural language text; the first and second natural language texts are merged to obtain the complete natural language text; S4, fault reasoning is performed on the complete natural language text using a large language model to obtain a structured reasoning answer containing the root cause of the fault, evidence path, data source, and processing suggestions; S5, based on the reasoning answer, the first and second natural language texts, a four-layer reverse tracing link containing the answer layer, evidence layer, data layer, and tracing layer is constructed.

[0010] Among them, graph algorithms refer to a set of algorithms for processing graph structured data, and in this application, they specifically refer to breadth-first search, depth-first search, and minimum spanning tree algorithms for knowledge graph optimization.

[0011] The PCST algorithm, or Prize-Collecting Steiner Tree algorithm, is an optimization algorithm for finding the optimal connected subgraph in a graph. In this application, the PCST algorithm uses the nodes corresponding to the key entities and key relationships extracted from the fault query as seed nodes. It searches the optimized knowledge graph for a connected subgraph that maximizes the node and edge rewards while minimizing the structural cost, ensuring that the generated optimized subgraph is highly relevant to the fault query and has a compact structure.

[0012] The first natural language text refers to the natural language description text obtained after transforming the optimized subgraph generated by the PCST algorithm.

[0013] The second natural language text refers to the natural language description text obtained after transforming the supplementary subgraph generated based on supplementary query terms. Supplementary query terms are entities and relationships extracted from the first natural language text that are not present in the optimized subgraph but are necessary for fault reasoning. By iteratively searching the supplementary subgraph on the optimized knowledge graph and converting it into natural language text, the missing information in the first natural language text is made up for. After being merged with the first natural language text, a complete fault analysis context is formed.

[0014] Completeness assessment refers to analyzing the first natural language text using a large language model to determine whether it contains all the key information required for fault reasoning. The assessment process relies on the semantic understanding capabilities of the large language model to identify issues such as missing entities, broken relationships, or incomplete logical chains within the text. When the assessment result is incomplete, a supplementary subgraph generation process is triggered; when the assessment result is complete, the process directly proceeds to the fault reasoning stage.

[0015] Fault reasoning refers to using large language models to perform deep semantic analysis and logical reasoning on complete natural language text. Based on knowledge and causal relationships in the field of communication network operation and maintenance, it identifies the root cause of the fault, sorts out the evidence path from the fault phenomenon to the root cause, locates the data source supporting the evidence, and generates targeted handling suggestions.

[0016] Furthermore, it also includes: S6, calculating the credibility score of the reasoning answer based on the association relationship between nodes at each layer in the reverse tracing link; when the credibility score is lower than the second threshold, identifying nodes along the reverse tracing link that have interrupted evidence paths, missing data sources, or incomplete tracing relationships; using the information corresponding to the nodes as missing information; and using a large language model to generate candidate supplementary data based on the missing information. Here, candidate supplementary data refers to the candidate data automatically generated by the large language model to complete the knowledge graph when the credibility score of the reasoning answer is lower than the second threshold, based on the identified missing information. Missing information includes three categories: interrupted evidence paths (missing logical relationships between reasoning steps), missing data sources (the knowledge graph entity or relationship corresponding to the evidence does not exist), and incomplete tracing relationships (data source identifier not recorded or invalid). The large language model, based on a dictionary in the field of communication operation and maintenance (such as the "5G Base Station Operation and Maintenance Terminology Collection") and industry standards (such as the "Communication Link Fault Handling Standard"), combined with the context of the missing information, generates candidate supplementary data including newly added entities, newly added relationships, newly added attribute values, or newly added tracing node identifiers.

[0017] S7. Update the knowledge graph using candidate supplementary data. Repeat steps S1 to S6 until the credibility score is greater than or equal to the second threshold, and output the fault tracking results.

[0018] Furthermore, the communication network operation and maintenance data includes: device data, link data, alarm data, and operation and maintenance data; device data includes: device model, deployment location, hardware parameters, and calibration records; link data includes: link length, latency, packet loss rate, and connection relationships; alarm data includes: alarm type, trigger time, priority, and associated devices; and operation and maintenance data includes: fault handling procedures and historical cases.

[0019] Furthermore, a knowledge graph for communication network operation and maintenance, comprising entity sets, relationship sets, and attribute sets, is constructed. This includes: extracting device entities, link entities, alarm entities, and knowledge entities from the full-dimensional data of communication network operation and maintenance to obtain entity sets; determining the relationships between entities based on the connection relationships in device data, link data, and associated devices in alarm data to construct relationship sets; using device signals, deployment locations, hardware parameters, and calibration records from device data; link bandwidth, latency, and packet loss rate from link data; alarm types, trigger times, and priorities from alarm data; and fault handling processes and historical cases from operation and maintenance data as attributes of the corresponding entities to construct attribute sets; and constructing the knowledge graph for communication network operation and maintenance based on the entity sets, relationship sets, and attribute sets.

[0020] Furthermore, the rules in the communications field include: equipment calibration cycle rules, alarm priority rules, and link latency rules. Equipment calibration cycle rules: the calibration cycle of the measurement terminal is less than or equal to 3 months. Alarm priority rules: when the number of link nodes affected by an alarm exceeds a preset node threshold, the alarm priority is set to high or emergency level; where the preset node threshold is a positive integer greater than or equal to 3. Link latency rules: the one-way propagation latency of the core transmission link is less than or equal to 100ms. The core transmission link refers to a high-priority link in the communication network topology that carries critical business data transmission and connects core network nodes. Core transmission links typically have the following characteristics: connecting backbone network equipment such as core routers and core switches; carrying large-scale business traffic and multi-user concurrent access; and having a decisive impact on network service availability and business continuity. In the knowledge graph, the core transmission link, as a type of link entity, is labeled as a core transmission link in terms of its link-level attributes. Specifically, a latency of less than 100ms ensures the needs of real-time services such as high-definition video and industrial control.

[0021] Furthermore, the knowledge graph is optimized using graph algorithms and communication domain rules, including: traversing the knowledge graph using a breadth-first search algorithm, counting the in-degree and out-degree of each node, and marking nodes with both in-degree and out-degree zero as isolated nodes; calculating the semantic similarity between relation pairs based on cosine similarity, and marking relations with semantic similarity greater than a preset similarity threshold as redundant relations; detecting loop structures in the knowledge graph using a depth-first search algorithm; extracting calibration records of measurement terminal entities in the knowledge graph based on device calibration cycle rules, calculating the difference between the current timestamp and the previous calibration timestamp, and marking measurement terminal entities with a difference greater than 3 months as abnormal entities; extracting the priority and associated devices of alarm entities in the knowledge graph based on alarm priority rules, obtaining the number of affected link nodes through association relationships, and marking alarm entities with the number of affected link nodes greater than a preset node threshold and alarm priority lower than higher levels as abnormal entities; and extracting the latency and link level of link entities in the knowledge graph based on link latency rules, and marking link entities with a core transmission link level and a one-way propagation latency greater than 100ms as abnormal entities.

[0022] In a knowledge graph, in-degree refers to the number of edges pointing to a given node in a directed graph structure. In a communication network operation and maintenance knowledge graph, nodes represent entities (such as device entities, link entities, and alarm entities), and edges represent relationships between entities (such as device connection relationships and alarm association relationships). In-degree reflects the number of times a node is referenced or depended upon by other nodes as the target of a relationship. For example, an in-degree of 5 for a device entity means that there are 5 relationships pointing to that device, i.e., the device is associated with 5 other entities. An in-degree of zero indicates that the node is not referenced by any other node, which may be isolated data or redundant information.

[0023] Out-degree refers to the number of edges originating from a given node and pointing to other nodes in a directed graph structure of a knowledge graph. Out-degree reflects the number of times a node acts as the starting point for a relationship, establishing connections with other nodes. For example, an alarm entity with an out-degree of 3 indicates that the alarm is associated with 3 other entities (such as 3 affected devices). An out-degree of zero indicates that the node does not point to any other node; it may be a terminal node or isolated data. When both the in-degree and out-degree of a node are zero, it means that the node is neither referenced by other nodes nor references other nodes, placing it in a completely isolated state within the knowledge graph. It does not participate in any knowledge associations and is therefore marked as an isolated node and deleted during the optimization process.

[0024] A loop structure refers to a closed path in a directed graph structure of a knowledge graph that starts from a node, passes through several edges and intermediate nodes, and eventually returns to the starting node. Loop structures lead to circular dependencies in the knowledge graph, affecting the termination of graph traversal algorithms and the convergence of inference. For example, device A connects to device B, device B connects to device C, and device C connects back to device A, forming a loop A→B→C→A. In communication network operation and maintenance scenarios, reasonable loops may reflect network redundancy and backup design, but abnormal loops may indicate data entry errors or logical relationship conflicts. This application uses a depth-first search algorithm to detect loop structures and uses a minimum spanning tree algorithm to delete the edge with the smallest weight in the loop, eliminating redundant connections, ensuring the tree-like or directed acyclic graph structure of the knowledge graph, and improving the rationality of the graph structure.

[0025] Furthermore, it also includes: based on communication network topology rules and device connection rules, detecting whether the connection relationships between device entities in the knowledge graph meet the requirements of port type, interface protocol and signal rate, and marking non-compliant relationships as abnormal relationships; deleting isolated nodes, redundant relationships, abnormal entities and abnormal relationships; and using the minimum spanning tree algorithm to delete the edge with the smallest weight in the loop structure to obtain the optimized knowledge graph.

[0026] Among them, communication network topology rules refer to the constraints that regulate the physical or logical connection methods between device nodes in a communication network, defining the legality requirements of the network architecture. Based on industry standards (such as ITU-T recommendations and 3GPP specifications) and network design principles, communication network topology rules specify the connection modes between devices at different levels. For example, access layer devices (base stations) must be connected to aggregation layer devices (aggregation switches), and aggregation layer devices must be connected to core layer devices (core routers); there should be no direct connection between devices at the same level, etc.

[0027] Device connection rules refer to the technical compatibility requirements for physical interface connections between devices in a communication network, including three dimensions: port type matching, interface protocol consistency, and signal rate compatibility. Port type matching requires that the port types of the two connected devices be the same or compatible (e.g., optical port to optical port, electrical port to electrical port); interface protocol consistency requires that both parties support the same communication protocol (e.g., Ethernet protocol, SDH protocol, OTN protocol, etc.); signal rate compatibility requires that the transmission rates of both parties match or support rate negotiation (e.g., 1Gbps, 10Gbps, 100Gbps, etc.).

[0028] The minimum spanning tree algorithm is a graph algorithm that finds a spanning tree in a weighted connected graph that contains all nodes and has the minimum sum of edge weights. Commonly used algorithms include Prim's algorithm and Kruskal's algorithm. In this application, the minimum spanning tree algorithm is used to process the loop structure detected in the knowledge graph.

[0029] Furthermore, the PCST algorithm is used to generate optimized subgraphs with a relevance greater than a first threshold to the fault query. This includes: segmenting and tagging the user-input fault query into words, extracting nouns as key entities, and extracting verbs and prepositions as key relations; in the optimized knowledge graph, nodes matching the entity names of key entities and nodes connected by edges matching the relation types of key relations are marked as seed nodes; starting from the seed nodes, the PCST algorithm is used to search for candidate subgraphs in the optimized knowledge graph, and the relevance score of each candidate subgraph to the fault query is calculated; the objective function of the PCST algorithm is: Constraints: Let S be a connected subgraph; where S* is the candidate subgraph with the highest relevance score. Let Vs be the k-th candidate subgraph. The set of nodes in Candidate subgraph The set of edges in the graph, where n is the candidate subgraph. A single node in the graph, where e is a candidate subgraph. A single edge in; The entity reward value for node n is determined when the semantic similarity between the device entity, link entity, alarm entity, or knowledge entity corresponding to node n and the key entity is greater than a preset entity similarity threshold. Set to 0.9, when the semantic similarity is less than or equal to the preset entity similarity threshold, Set it to 0.1; The relation reward value for edge e is calculated when the semantic similarity between the association corresponding to edge e and the key relation is greater than a preset relation similarity threshold. Set to 0.8, when the semantic similarity is less than or equal to the preset relation similarity threshold, Set it to 0.3; Candidate subgraph The structural cost is calculated using the following formula: ,in, Candidate subgraph The number of middle edges, Set the preset cost parameters for each edge; output S* as the optimized subgraph.

[0030] Specifically, this application achieves a dynamic balance between fault query relevance and subgraph structure compactness by simultaneously optimizing three dimensions—node reward value, edge reward value, and structural cost—in the objective function. Specifically, the node reward value... Edge reward value Semantic similarity metric was used to quantify the relevance between knowledge graph entities and relationships and fault queries, ensuring that optimized subgraphs contain core evidence for fault localization; while structural cost By penalizing candidate subgraphs with an excessive number of edges, redundant subgraphs containing a large amount of weakly related or irrelevant information are prevented, avoiding the information overload problem caused by blind expansion in traditional graph search algorithms. This reward-penalty trade-off mechanism ensures that optimizing subgraphs guarantees the integrity of fault-related information while minimizing the introduction of noise data. This provides high-quality, high-density knowledge context for subsequent fault reasoning in large language models, fundamentally solving the problem of low fault location accuracy caused by the coexistence of incomplete knowledge graph data and information redundancy, and significantly improving the accuracy of fault tracking and reasoning efficiency.

[0031] Furthermore, a four-layer reverse tracing link is constructed, comprising an answer layer, an evidence layer, a data layer, and a tracing layer. This includes: using the root causes and suggested solutions in the reasoning answer as answer layer nodes; parsing the evidence path in the reasoning answer, extracting the reasoning steps and logical relationships within the evidence path, using each reasoning step as an evidence layer node, and establishing directed edges from answer layer nodes to evidence layer nodes as support relationships, with the edge weights of the support relationships determined based on the contribution of the reasoning steps to the root causes or suggested solutions; extracting entities and relationships corresponding to each evidence layer node from the first and second natural language texts, and then... Relationships are used as data layer nodes, and directed edges from evidence layer nodes to data layer nodes are established as reference relationships. The data sources in the reasoning answer are analyzed to obtain the original data source identifiers corresponding to each data layer node. The original data source identifiers are used as tracing layer nodes, and directed edges from data layer nodes to tracing layer nodes are established as source relationships. The original data source identifiers include the collection timestamp, collection device identifier, and data table identifier of device data, link data, alarm data, or operation and maintenance data. Based on the supporting relationships, reference relationships, and source relationships, a four-layer reverse tracing link is constructed from the answer layer to the evidence layer, data layer, and tracing layer.

[0032] Specifically, this application establishes a multi-layered traceability system—from the answer layer to the evidence layer, data layer, and source tracing layer—to achieve a fully credible quantification transmission mechanism from the fault reasoning result to the original data source. Specifically, the edge weights of supporting relationships quantify the contribution of the reasoning steps to the root cause of the fault, citation relationships clarify the correspondence between evidence and knowledge graph entities, and source relationships trace back to source metadata such as the original data collection timestamp, collection device identifier, and data table identifier. This layered source tracing architecture enables precise location along the reverse source tracing link of a specific layer and node where the evidence path is interrupted, the data source is missing, or the source relationship is incomplete, when the credibility score of the reasoning answer is below a threshold, rather than simply judging the reasoning result as unreliable in traditional methods. By transforming abstract credibility assessment into operable node-level missing information localization, this application fundamentally solves the problem of low fault localization accuracy caused by incomplete knowledge graph data. It provides precise target guidance for subsequent targeted generation of candidate supplementary data using large language models and iterative optimization of knowledge graphs, forming a closed-loop optimization mechanism of reasoning-source tracing-evaluation-completion-re-reasoning, which significantly improves the reliability and adaptability of communication network fault tracking.

[0033] Furthermore, a large language model is used to generate candidate supplementary data based on missing information, including: traversing the four-layer reverse tracing link, extracting the evidence layer nodes associated with the answer layer nodes, calculating the data quality level of the original data corresponding to each evidence layer node, assigning values ​​according to the data quality level: 1.0 for valid data quality, 0.5 for pending review, and 0 for invalid data quality. The average score of the core evidence quality is calculated as the evidence quality confidence level, using the following formula: Where Q is the confidence level of evidence quality, m is the number of core pieces of evidence, and qi is the quality score of the i-th core piece of evidence; the integrity of the nodes in the answer layer, evidence layer, data layer, and tracing layer of the four-layer reverse tracing link is checked. When all four layers of nodes are complete, the link integrity confidence level C is set to 1.0; when one layer of nodes is missing, the link integrity confidence level C is reduced by 0.25; the credibility score of the reasoning answer is calculated using the formula: Credibility score = Q × C; where the credibility score ranges from [0, 1]; when the credibility score is lower than the second threshold, the missing levels of the four-layer reverse tracing link are detected, the answer layer nodes of the missing evidence layer nodes are marked as evidence path interruption nodes, the evidence layer nodes of the missing data layer nodes are marked as data source missing nodes, and the data layer nodes of the missing tracing layer nodes are marked as tracing relationship incomplete nodes; the root cause or handling suggestion corresponding to the evidence path interruption node, the reasoning steps corresponding to the data source missing node, and the entity and relationship corresponding to the tracing relationship incomplete node are extracted as missing information; the missing information and fault query are input into the big language model, prompting the big language model to generate candidate supplementary data required for supplementing evidence path interruption, data source missing or tracing relationship incomplete.

[0034] Preferably, in step S7, the knowledge graph is updated using candidate supplementary data. Steps S1 to S6 are repeated until the credibility score is greater than or equal to the second threshold, and the fault tracing result is output, including: designing data integrity rules, semantic consistency rules, and source tracing validity rules; the data integrity rules require that the core evidence for fault location must cover at least two of the following: device status data, link performance data, and alarm log data; the semantic consistency rules require that the semantic association between evidence must conform to the communication network logic; the source tracing validity rules require that the data acquisition node must meet the following requirements: the monitoring terminal has been calibrated within 3 months, the operation and maintenance platform qualification is within the validity period, and the deviation between the data acquisition time and the fault time is less than or equal to 24 hours; the data integrity is verified layer by layer along the four-layer reverse source tracing link, deducing from the reasoning answer. The system assesses the satisfaction of integrity rules, semantic consistency rules, and source traceability validity rules to identify the types of missing information in the knowledge graph. These missing information types include data missing, knowledge missing, and source traceability missing. Data missing refers to the absence or invalidity of maintenance data corresponding to core evidence; knowledge missing refers to the absence of knowledge elements required for fault handling procedures; and source traceability missing refers to the failure to record or invalidity of evidence collection node information. The system loads a communication maintenance domain dictionary and industry standards, inputs candidate supplementary data into a large language model, and generates candidate supplementary information based on the missing information type. After user review, the candidate supplementary information is marked as updatable. This updatable information is then integrated into the knowledge graph to update it. The updated knowledge graph Gnew is represented by the formula: Gnew = G ∪ {nnew, enew, vnew} ∪ {idt-new};where G is the knowledge graph before the update, nnew is the newly added entity, enew is the newly added relation, vnew is the newly added attribute value, and idt-new is the newly added traceability node identifier; based on the updated knowledge graph, repeat steps S1 to S6 to calculate the updated credibility score; when the updated credibility score is greater than or equal to the second threshold, or the number of iterations reaches the preset number of iterations, output the current reasoning answer as the fault tracking result, and output the updated knowledge graph.

[0035] Another aspect of this application provides a tracking system for communication network operation and maintenance faults, comprising: a data acquisition module for acquiring full-dimensional data on communication network operation and maintenance; a graph optimization module for optimizing a knowledge graph using graph algorithms and communication domain rules; generating optimized subgraphs with a relevance greater than a first threshold to the fault query using the PCST algorithm based on the optimized knowledge graph and the fault query input by the user, and converting the optimized subgraphs into first natural language text; an integrity judgment module for performing integrity judgment on the first natural language text using a large language model; when the judgment is incomplete, extracting entities and relations not appearing in the optimized subgraph from the first natural language text as supplementary query terms; iteratively generating supplementary subgraphs on the optimized knowledge graph based on the supplementary query terms, and converting the supplementary subgraphs into second natural language text; merging the first natural language text and the second natural language text to obtain complete natural language text; and a fault reasoning module for performing fault reasoning on the complete natural language text using a large language model. The fault reasoning module obtains a structured reasoning answer containing the root cause of the fault, evidence path, data source, and processing suggestions. The source tracing link construction module constructs a four-layer reverse source tracing link, including an answer layer, evidence layer, data layer, and source tracing layer, based on the reasoning answer, the first natural language text, and the second natural language text. The credibility assessment module calculates the credibility score of the reasoning answer based on the relationships between nodes in each layer of the reverse source tracing link. When the credibility score is lower than a second threshold, nodes with interrupted evidence paths, missing data sources, or incomplete source tracing relationships are identified along the reverse source tracing link. The information corresponding to the nodes is treated as missing information. Candidate supplementary data is generated based on the missing information using a large language model. The iterative optimization module updates the knowledge graph using the candidate supplementary data, repeatedly executing the processing flow by calling the data acquisition module, graph optimization module, integrity judgment module, fault reasoning module, source tracing link construction module, and credibility assessment module until the credibility score is greater than or equal to the second threshold, and then outputs the fault tracking result.

[0036] Compared to existing technologies, the advantages of this application are:

[0037] This application constructs a four-layer reverse tracing link comprising an answer layer, an evidence layer, a data layer, and a tracing layer. Based on the correlation between nodes in each layer, it quantitatively calculates the credibility score of the reasoning answer. When the credibility score is below a threshold, it can accurately locate specific nodes in the knowledge graph where the evidence path is interrupted, the data source is missing, or the tracing relationship is incomplete. It also uses a large language model to generate candidate supplementary data to update the knowledge graph, forming a closed-loop optimization mechanism of fault reasoning, credibility assessment, missing data identification, and knowledge completion. This effectively solves the problem of low fault location accuracy caused by incomplete knowledge graph data and significantly improves the accuracy and reliability of communication network fault tracking. Attached Figure Description

[0038] This application will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0039] Figure 1 This is an exemplary flowchart illustrating the tracking of communication network operation and maintenance faults according to some embodiments of this application;

[0040] Figure 2 This is an exemplary flowchart of constructing a knowledge graph according to some embodiments of this application;

[0041] Figure 3 This is an exemplary flowchart of an optimized knowledge graph according to some embodiments of this application;

[0042] Figure 4 This is an exemplary flowchart illustrating the generation of extended subgraphs according to some embodiments of this application;

[0043] Figure 5 This is an exemplary flowchart illustrating the generation of the final location result according to some embodiments of this application. Detailed Implementation

[0044] The methods and systems provided in the embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0045] Example 1

[0046] like Figure 1 As shown, the process involves acquiring comprehensive data on communication network operation and maintenance, constructing a knowledge graph of communication network operation and maintenance that includes entity sets, relation sets, and attribute sets; optimizing the knowledge graph using graph algorithms and communication domain rules; generating optimized subgraphs with a relevance greater than a first threshold based on the optimized knowledge graph and user-input fault queries using the PCST algorithm, and converting the optimized subgraphs into first natural language text; performing a completeness check on the first natural language text using a large language model; extracting entities and relations not appearing in the optimized subgraphs from the first natural language text as supplementary query terms based on the supplementary query terms; iteratively generating supplementary subgraphs on the optimized knowledge graph based on the supplementary query terms, and converting the supplementary subgraphs into second natural language text; merging the first and second natural language texts to obtain a complete natural language text; performing fault reasoning on the complete natural language text using a large language model to obtain a structured reasoning answer containing fault root causes, evidence paths, data sources, and processing suggestions; and constructing a four-layer reverse tracing link containing an answer layer, an evidence layer, a data layer, and a tracing layer based on the reasoning answer, the first natural language text, and the second natural language text.

[0047] Specifically, step S1 is as follows:

[0048] like Figure 2 As shown, S1.1 acquires full-dimensional data on communication network operation and maintenance and constructs a knowledge graph.

[0049] (1) Collect data from all dimensions, including: equipment data (equipment model, deployment location, hardware parameters, calibration records), link data (link bandwidth, latency, packet loss rate, connection relationship), alarm data (alarm type, trigger time, priority, associated equipment), and operation and maintenance knowledge data (fault handling solutions, historical fault cases, and the "Communication Equipment Operation and Maintenance Procedures" standard).

[0050] (2) Construct a knowledge graph for communication network operation and maintenance. ,in, It is an entity set, which includes "device entities" (such as router R1, base station B2), "link entities" (such as link L3), "alarm entities" (such as alarm A4), and "operation and maintenance knowledge entities" (such as solution S5). This is a set of relationships, including "device-link" associations (such as "connection"), "device-alarm" associations (such as "trigger"), and "alarm-solution" associations (such as "corresponding processing"). It is a set of attributes for nodes and edges, including entity attributes (such as device "calibration time" and alarm "priority") and relationship attributes (such as link "delay value" and association "confidence").

[0051] S1.2 employs a combination of graph algorithms and rule engines to filter invalid knowledge for quality control.

[0052] (1) Graph algorithm verification: Identify isolated entities (such as devices without associated links) and redundant relationships (such as "device-alarm" association with duplicate records) through depth-first search (DFS); detect logically conflicting entity clusters (such as the same device being marked as "normal operation" and "fault shutdown") using community detection algorithms (such as Louvain).

[0053] (2) Rule engine verification in the field of communication: Based on the predefined rules in the "Technical Standard for Operation and Maintenance of Communication Networks", such as "Equipment calibration cycle ≤ 3 months (measurement terminal)", "Alarm priority matches the scope of fault impact (core link alarm priority ≥ high)", "Link delay ≤ 100ms (backbone network)", verify the consistency of entity attributes and the rationality of relationships, and filter out erroneous knowledge that does not conform to the rules (such as equipment data that has not been calibrated within the time limit, alarm association with priority mismatch).

[0054] S1.3 Fault Location Optimization Subgraph Generation: For user-input fault queries (such as "root cause of service interruption at base station B101"), based on the quality-controlled knowledge graph, the Prize-Collecting Steiner Tree (PCST) algorithm is applied to optimize the connectivity of candidate subgraphs, aiming to maximize the reward value of relevant nodes / edges and minimize the cost of the subgraph, thereby generating a highly relevant optimized subgraph. The objective function is:

[0055] ;in, , Candidate subgraphs Entity sets and relation sets; The reward value is set as follows: (the reward value for fault-related entities such as "base station B101" and "link L202" is set to 0.9, and the reward value for unrelated entities is set to 0.1). The relationship reward value is set as follows: core relationships such as "trigger" and "connection" are rewarded with 0.8, while secondary relationships are rewarded with 0.3. For the cost of the subgraph, A preset cost parameter (default value is 1) is set for each relationship to ensure that the subgraph is not excessively redundant.

[0056] S1.4 Subgraph Iterative Completion and Fault Reasoning Preparation.

[0057] (1) Optimize candidate subgraphs Convert to natural language text (e.g., “Entity: Base Station B101 (Attributes: Calibration Time 2024-03-15, Status: Service Interruption); Relationship: Base Station B101-Connection-Link L202 (Attributes: Delay 150ms); Alarm: Alarm A4 (Type: Link Congestion, Triggering Device: Link L202)”).

[0058] (2) Input LLM (such as GPT-4, Tongyi 1000 Questions), and combine it with the clapping query. ,judge Does it include "fault entity + core relationship + operation and maintenance data" (such as "base station B101 + connection + link L202 delay data")? If it does, proceed to the fault reasoning stage; otherwise, trigger iterative completion S1.5.

[0059] If S1.5 fails to generate a complete answer, the LLM first extracts new keywords from the query; then, based on the new keywords, the candidate subgraph selection and PCST optimization process of S1.3 is repeated to generate an expanded subgraph. ,Will Merge into the original optimized subgraph And update the subgraph text. Finally, repeat the above two steps until the subgraph meets the answer generation requirements or reaches the preset maximum number of iterations (default is 5).

[0060] like Figure 3 As shown, step S2 specifically involves:

[0061] S2.1 Based on the final subgraph text LLM is used to generate root cause reasoning and localization results for faults.

[0062] (1) First, conduct evidence path reasoning. LLM analysis The association of "fault phenomenon - associated entity - operation and maintenance data" is used to explore the reasoning path, such as "base station B101 service interruption (fault phenomenon) ← connection ← link L202 (entity) ← trigger ← alarm A4 (link congestion) (evidence) ← attribute ← latency 150ms (operation and maintenance data)".

[0063] (2) Secondly, for the multiple inference paths generated by multiple iterations, a "confidence-weighted" fusion method is used (confidence is calculated based on the quality level of evidence data, with "valid" evidence having a weight of 0.8 and "pending review" having a weight of 0.4), and the three paths with the highest confidence are selected as the fault query. Input LLM to generate a structured reasoning answer, which includes "root cause of the fault (e.g., link L202 congestion) + core evidence path + data source (e.g., link L202 latency monitoring log) + handling suggestions (e.g., restart link L202)".

[0064] S2.2 Text Based on Multi-round Iteration Subgraph In conjunction with the LLM inference log (including keyword extraction records and path inference logic), a four-layer reverse tracing link is constructed: "Answer - Evidence - Data - Source Node". The layers are defined as follows: the Answer layer refers to the inference answer generated by S2.1 (e.g., "Base station B101 service interruption root cause is link L202 congestion"); the Evidence layer refers to the core and auxiliary evidence supporting the answer (e.g., "Base station B101 ← connection ← link L202 ← trigger ← alarm A4"); the Data layer refers to the original data corresponding to the evidence (e.g., the latency of link L202 is 150ms, and the trigger time of alarm A4 is 2024-05-20); the Source tracing layer refers to the nodes where the original data is collected / processed (e.g., latency data comes from "monitoring terminal T301", and alarm data comes from "operation and maintenance platform P10").

[0065] S2.3 Based on the quality of evidence and the integrity of the link, the support confidence of the reasoning answer is quantitatively calculated.

[0066] (1) Confidence of evidence quality: For each core piece of evidence, assign values ​​according to data quality level ("Valid" = 1.0, "Pending review" = 0.5, "Invalid" = 0), and calculate the average quality score of the core evidence. ,in The number of core pieces of evidence, For the first The quality score of each core piece of evidence.

[0067] (2) Link integrity confidence: If there are no missing nodes in the four layers of the tracing link (answer-evidence-data-tracing), the integrity score is [not specified]. Each missing layer Deduct 0.25 (if the "source tracing layer" is missing). ).

[0068] (3) Support confidence level: confidence level (Value range [0, 1]); Set confidence threshold ,like If the reasoning is correct, the answer is plausible, and the reasoning is complete; if... Then proceed to S3.

[0069] like Figure 4 and Figure 5 As shown, step S3 specifically involves:

[0070] Based on the "Communication Network Operation and Maintenance Procedures" and fault location requirements, S3.1 designs three types of inversion verification rules to locate missing or erroneous information in the knowledge graph:

[0071] (1) Data integrity rules: The core evidence for fault location must cover at least two of the following categories: “device status data + link performance data + alarm log data”;

[0072] (2) Semantic consistency rules: The semantic association between evidence must conform to the communication network logic (e.g., "link congestion" alarm must match "delay exceeds threshold" data, and "equipment shutdown" status must match "no business traffic" data).

[0073] (3) Traceability validity rules: The data collection node must meet the following conditions: "the monitoring terminal has been calibrated within 3 months", "the operation and maintenance platform qualification is within the validity period", and "the deviation between the data collection time and the fault time is ≤24 hours", otherwise the traceability is invalid.

[0074] S3.2 Derive the answer from the inference path, tracing back along the reverse source link. The rules of S3.1 are verified layer by layer to locate the three types of missing information in the knowledge graph.

[0075] (1) Data missing: The operation and maintenance data corresponding to the core evidence is missing or invalid (e.g., there is no historical latency data for link L202, and the monitoring terminal T301 has not been calibrated over the time limit).

[0076] (2) Knowledge gap: The knowledge elements required for the entire fault handling process are missing (such as the handling scheme corresponding to "link congestion" is not recorded, and the relationship between "base station service interruption" and "link failure" is not defined).

[0077] (3) Lack of traceability: The information on the collection nodes of the evidence is not recorded or is invalid (e.g., the source of alarm A4 is not marked by the operation and maintenance platform, and the monitoring terminal T301 has no calibration record).

[0078] S3.3 Develop supplementary solutions based on the type of missing information, and achieve supplementation through "LLM assistance + human collaboration" to update the multivariate graph and form a closed loop.

[0079] (1) Load the dictionary of communication operation and maintenance field (such as the "5G base station operation and maintenance terminology collection") and industry standards (such as the "communication link fault handling standard"), and generate candidate supplementary information (such as missing link handling schemes and traceability node calibration records) by LLM; manually review the candidate information (focus on reviewing high-risk information such as core equipment parameters), and mark the verified information as "updatable";

[0080] (2) Integrate "updatable" information into the original knowledge graph to update the heterogeneous multi-dimensional knowledge graph. The updated map It is represented as follows:

[0081] ;

[0082] in, For the newly added entity, To establish a new relationship, To add a new attribute value, Add a new traceability node ID.

[0083] S3.4 Iterative optimization closes the loop, repeating the S1-S3 process, and re-performs fault location and tracing based on the updated knowledge graph until the location result has a certain confidence level. Alternatively, after reaching the preset number of iterations (default 5), the final fault location result and optimized knowledge graph will be output.

[0084] Example 2

[0085] Taking the operation and maintenance of a 5G base station network in a certain province as an example, this invention's method is used to locate and track faults in the case of "service interruption of base station B101 (located in the core business district of a certain city)".

[0086] Step S1: Knowledge Graph Construction and Quality Control

[0087] Data Acquisition and Modeling: Acquire equipment data (model: Huawei AAU5612, calibration time: 2024-03-15, status: service interrupted) from base station B101, link data (connection link L202, latency 150ms, bandwidth 10Gbps), alarm data (alarm A4, type: link congestion, trigger time: 2024-05-20 14:30), and operation and maintenance knowledge data (fault handling solution S1: link congestion restart equipment), and construct a knowledge graph. .

[0088] Quality control: Identify and delete unrelated redundant device entities (such as "Router R503") through DFS; the rule engine verifies that "Link L202 latency 150ms > backbone network threshold 100ms" and marks it as abnormal data; "Base station B101 calibration time 2024-03-15 (2 months from now)" meets the "≤3 months" rule and is marked as valid data.

[0089] Subgraph generation optimization: Applying the PCST algorithm, the reward value for core entities (base station B101, link L202, alarm A4) is set to 0.9, and the reward value for core relationships (connection, trigger) is set to 0.8, generating an optimized subgraph S^*; converting it into text. "Base station B101 (calibrated 2024-03-15, service interrupted) connected to link L202 (latency 150ms), link L202 triggered alarm A4 (congestion, 14:30)."

[0090] Iterative completion: LLM judgment The information for "upstream device of link L202" is missing. Extract the keyword "upstream of link L202", add "switch S305" (status: normal) as the upstream device, and update S^* and This satisfies the reasoning requirements.

[0091] Step S2: Fault Reasoning and Source Tracing Verification:

[0092] Reasoning and Result Generation: The LLM mining inference path "Base station B101 service interruption ← connection ← link L202 ← trigger ← alarm A4 (congestion) ← attribute ← latency 150ms" is weighted and fused to filter as the core path; the location result is generated as follows: "Root cause of the fault: link L202 congestion (latency 150ms exceeds the threshold); handling suggestion: restart link L202; data source: link L202 monitoring log".

[0093] Reverse link construction: Result layer (link L202 congestion) → Evidence layer (base station B101-connection-link L202-trigger-alarm A4) → Data layer (latency 150ms, alarm time 14:30) → Source tracing layer (monitoring terminal T301, calibration 2024-04-01).

[0094] Confidence Calculation: Quality of Evidence (All evidence is valid), link integrity (No missing layers in all four layers), the localization result is reliable.

[0095] Step S3: Knowledge Graph Update (Example Missing Scenario):

[0096] If the confidence level of the positioning result (Due to the lack of historical congestion data for link L202), then:

[0097] Rule verification: Triggering a "Data Integrity Rule" alert (core evidence only includes real-time data from the data link, with no historical data);

[0098] Additional information: LLM generated "Link L202 congestion records for the past month (April 10, 2024, May 5, 2024)," which was manually reviewed and matched with monitoring logs, and the verification was successful.

[0099] Map update: Add "Link L202-Historical Congestion-2024-04-10 / 2024-05-05" to G to form G_new; repeat S1-S2, new confidence. Output the final result.

[0100] The foregoing illustrative description of the present application and its embodiments is not restrictive and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. The accompanying drawings are only one embodiment of the present application, and the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present application, such designs should fall within the scope of protection of this application. Furthermore, the word "comprising" does not exclude other elements or steps, and the word "a" preceding an element does not exclude the inclusion of "a plurality" of that element. Terms such as "first," "second," etc., are used to indicate names and do not indicate any specific order.

Claims

1. A method for tracking a fault of operation and maintenance of a communication network, characterized in that, Comprise: S1, obtain communication network operation and maintenance full-dimensional data, and construct a communication network operation and maintenance knowledge graph containing an entity set, a relationship set and an attribute set; S2, optimize the knowledge graph using graph algorithms and communication field rules; According to the optimized knowledge graph and the user input fault query, the PCST algorithm is used to generate an optimized subgraph with a correlation degree greater than a first threshold value with the fault query, and the optimized subgraph is converted into a first natural language text; S3, using a large language model to judge the integrity of the first natural language text; When judged as not complete, extract the entities and relationships in the first natural language text that are not in the optimized subgraph as supplementary query words; Based on the supplementary query words, iteratively generate a supplementary subgraph on the optimized knowledge graph, and convert the supplementary subgraph into a second natural language text; Merge the first natural language text and the second natural language text to obtain a complete natural language text; S4, using a large language model to perform fault reasoning on the complete natural language text, and obtaining a structured reasoning answer containing fault root cause, evidence path, data source and processing suggestion; S5, according to the reasoning answer, the first natural language text and the second natural language text, a four-layer reverse traceability link containing answer layer, evidence layer, data layer and traceability layer is constructed; S6, based on the association relationship between nodes in each layer of the reverse traceability link, the credibility score of the reasoning answer is calculated; When the credibility score is lower than the second threshold value, the nodes with incomplete evidence path, missing data source or incomplete traceability relationship are identified along the reverse traceability link; the information corresponding to the nodes is taken as the missing information; Using a large language model to generate candidate supplementary data according to the missing information; S7, using the candidate supplementary data to update the knowledge graph, repeating steps S1 to S6 until the credibility score is greater than or equal to the second threshold value, and outputting the fault tracking result; Wherein, The PCST algorithm is used to generate an optimized subgraph with a correlation degree greater than a first threshold value with the fault query, comprising: Carry out word segmentation and part-of-speech tagging on the user input fault query, extract nouns as key entities, and extract verbs and prepositions as key relationships; In the optimized knowledge graph, nodes with entity names matching the key entities, and nodes connected by edges with relationship types matching the key relationships, are marked as seed nodes; Taking the seed nodes as the starting point, the PCST algorithm is used to search for candidate subgraphs in the optimized knowledge graph, and the correlation score of each candidate subgraph with the fault query is calculated; the objective function of the PCST algorithm is: ; wherein, is the candidate subgraph with the highest correlation score, is the kth candidate subgraph, is the candidate subgraph , n is a single node in the candidate subgraph , e is a single edge in the candidate subgraph , n is a single node in the candidate subgraph , e is a single edge in the candidate subgraph ; is the entity reward value for node n; is the relation reward value for edge e; is the structure cost of the candidate subgraph ; As an output of the optimization subgraph. As an output of the optimization subgraph.

2. The communication network operation and maintenance fault tracking method according to claim 1, wherein: The communication network operation and maintenance full-dimensional data comprises: device data, link data, alarm data and operation and maintenance data; The device data includes: device model, deployment location, hardware parameter and calibration record; The link data includes: link bandwidth, delay, packet loss rate and connection relationship; The alarm data includes: alarm type, trigger time, priority and associated equipment; The operation and maintenance data includes: fault handling process and historical cases.

3. The communication network operation and maintenance fault tracking method according to claim 2, wherein: Constructing a communication network operation and maintenance knowledge graph including an entity set, a relationship set and an attribute set, comprising: Extracting device entities, link entities, alarm entities and knowledge entities from communication network operation and maintenance multidimensional data to obtain the entity set; Determining the association relationship between entities based on the connection relationship in device data, the connection relationship in link data and the associated devices in alarm data to construct the relationship set; Taking device signals, deployment locations, hardware parameters and calibration records in device data, link bandwidth, latency and packet loss rate in link data, alarm types, triggering time and priority in alarm data, and fault handling processes and historical cases in operation and maintenance data as the attributes of the corresponding entities to construct the attribute set; Constructing the communication network operation and maintenance knowledge graph according to the entity set, the relationship set and the attribute set.

4. The communication network operation and maintenance fault tracking method according to claim 2, wherein: The communication field rules include: device calibration period rule, alarm priority rule and link latency rule; The device calibration period rule: the calibration period of the measurement terminal is less than or equal to 3 months; The alarm priority rule: when the number of link nodes affected by the alarm is greater than a preset node threshold, the alarm priority is set to high or emergency; The link latency rule: the one-way propagation latency of the core transmission link is less than or equal to 100 ms.

5. The communication network operation and maintenance fault tracking method according to claim 4, wherein: Optimizing the knowledge graph using graph algorithms and communication field rules, comprising: Using the breadth-first search algorithm to traverse the knowledge graph, counting the in-degree and out-degree of each node, and marking the nodes with both in-degree and out-degree of zero as isolated nodes; calculating the semantic similarity between relationship pairs based on cosine similarity, and marking the relationships with semantic similarity greater than a preset similarity threshold as redundant relationships; detecting the loop structure in the knowledge graph using the depth-first search algorithm; Based on the device calibration period rule, extracting the calibration records of the measurement terminal entities in the knowledge graph, calculating the difference between the current timestamp and the last calibration timestamp, and marking the measurement terminal entities with a difference greater than 3 months as abnormal entities; Based on the alarm priority rule, extracting the priority and associated devices of the alarm entities in the knowledge graph, obtaining the number of affected link nodes through the association relationship, and marking the alarm entities with the number of affected link nodes greater than the preset node threshold and the alarm priority lower than high as abnormal entities; Based on the link latency rule, extracting the latency and link level of the link entities in the knowledge graph, and marking the link entities with the link level as core transmission links and the one-way propagation latency greater than 100 ms as abnormal entities.

6. The communication network operation and maintenance fault tracking method according to claim 5, wherein: Optimizing the knowledge graph using graph algorithms and communication field rules further comprises: Based on the communication network topology rule and the device connection rule, detecting whether the connection relationship between device entities in the knowledge graph meets the requirements of port type, interface protocol and signal rate, and marking the association relationships that do not meet the requirements as abnormal relationships; Deleting isolated nodes, redundant relationships, abnormal entities and abnormal relationships; The minimum spanning tree algorithm is used to delete the edge with the minimum weight in the loop structure, and an optimized knowledge graph is obtained.

7. The method of claim 6, wherein the method further comprises: S5, constructing a four-layer reverse trace link including an answer layer, an evidence layer, a data layer, and a trace layer, comprising: taking the fault root cause and the processing suggestion in the reasoning answer as answer layer nodes; analyzing the evidence path in the reasoning answer, extracting reasoning steps and logical relationships in the evidence path, taking each reasoning step as an evidence layer node, and establishing a directed edge from the answer layer node to the evidence layer node as a support relationship, wherein the edge weight of the support relationship is determined according to the contribution of the reasoning step to the fault root cause or the processing suggestion; extracting entities and relationships corresponding to each evidence layer node from the first natural language text and the second natural language text, taking the entities and relationships as data layer nodes, and establishing a directed edge from the evidence layer node to the data layer node as a reference relationship; analyzing the data source in the reasoning answer, obtaining an original data source identifier corresponding to each data layer node, taking the original data source identifier as a trace layer node, and establishing a directed edge from the data layer node to the trace layer node as a source relationship, wherein the original data source identifier includes a collection time stamp, a collection device identifier, and a data table identifier of device data, link data, alarm data, or operation and maintenance data; constructing a four-layer reverse trace link from the answer layer to the evidence layer, the data layer, and the trace layer according to the support relationship, the reference relationship, and the source relationship.

8. A system for tracing a communication network operation and maintenance fault, configured to implement the method of any one of claims 1 to 7.

Images