Knowledge graph-based communication network operation and maintenance method and related device

Abstract

The embodiment of the application provides a communication network operation and maintenance method based on a knowledge graph and related devices. In the method, based on a dynamic ontology evolution model, a network new entity type is identified by using concept drift detection, and an implicit association is mined by using a meta-path reasoning algorithm. A time window-based incremental update strategy is constructed for high-frequency change business link data. In response to a fault signal, false associations are stripped from the knowledge graph by a causal graph model, and time sequence feature extraction is performed on a fault propagation subgraph by using a graph convolution network (GCN). A root cause confidence ranking is output by combining a root cause ordering model trained by reinforcement learning. A double-target game model of business SLA and network resources is constructed based on a positioning result. A resource scheduling scheme that satisfies constraints is solved by Nash equilibrium. After digital twin simulation verification, strategy execution bias data is fed back to the knowledge graph for model parameter correction. The application significantly improves the intelligence level and stability of communication network operation and maintenance.

Description

Technical Field

[0001] This application belongs to the field of communication network operation and maintenance technology, and particularly relates to communication network operation and maintenance methods and related devices based on knowledge graphs. Background Technology

[0002] Communication network operation and maintenance services cover the operation management, maintenance and fault handling of various aspects such as data center environment, base station equipment, and transmission lines. Its importance is self-evident, as it directly affects whether end users can enjoy stable and high-quality network services.

[0003] Currently, with the continuous upgrading and expansion of communication networks, different network standards are layered and intertwined, making network infrastructure and networking environments increasingly complex. Against this backdrop, intelligent operation and maintenance has become a key development direction for the industry, and technologies such as artificial intelligence and big data are widely used in communication network operation and maintenance services, driving its transformation towards digitalization, intelligence, and refinement.

[0004] Currently, existing operation and maintenance technologies typically construct knowledge graphs by collecting network device operation logs and topology data. In terms of fault location, simple rule matching or basic reasoning is used to determine possible causes of faults by utilizing device relationships and historical fault cases in the graph. However, in the fault location stage, facing large-scale and complex communication networks, traditional rule-based and simple reasoning methods are difficult to quickly and accurately locate the root cause of faults. When multiple concurrent faults or complex cascading faults occur in the network, the reasoning speed and accuracy drop sharply, failing to effectively meet the operation and maintenance requirements for timely fault handling. Summary of the Invention

[0005] The purpose of this application is to provide a knowledge graph-based communication network operation and maintenance method and related apparatus, which aims to solve the technical problems mentioned in the background art.

[0006] To achieve the above objectives, the embodiments of this application provide the following technical solutions.

[0007] According to one embodiment of this application, a knowledge graph-based communication network operation and maintenance method is provided, including the following steps: Based on a collaborative acquisition architecture deployed in a communication network, data desensitization and feature alignment are performed through a federated learning mechanism, and entities and relations are jointly extracted from unstructured data using a BERT pre-trained model. Based on the dynamic ontology evolution model, concept drift detection is used to identify new entity types in the network, and meta-path reasoning algorithm is used to mine implicit associations. For business link data with high frequency changes, an incremental update strategy based on time window is constructed to make the knowledge graph update latency lower than the target value. In response to fault signals, false associations are removed from the knowledge graph through a causal graph model, and temporal features are extracted from the fault propagation subgraph using a graph convolutional network (GCN). The root cause ranking model trained by reinforcement learning is then used to output the root cause confidence ranking. Based on the location results, a dual-objective game model of business SLA and network resources is constructed. The resource scheduling scheme that satisfies the constraints is solved by Nash equilibrium. After verification by digital twin simulation, the policy execution deviation data is fed back to the knowledge graph for model parameter correction.

[0008] Preferably, the step of using a BERT pre-trained model to jointly extract entities and relations from unstructured data includes: Unstructured data from communication networks is preprocessed and categorized into equipment maintenance documents, fault alarm description texts, and business interaction logs based on data type. Fine-tuned corpora containing a domain terminology dictionary are generated for each type of data. The fine-tuned corpus is input into the BERT model pre-trained based on corpus from the communication domain. By setting a dual task head in the output layer, which includes an entity recognition head and a relation classification head, and adding association labels to the entity recognition head, entity boundary labels and relation labels between entities are obtained. The output results are cross-validated. If the confidence of the joint extraction of entities and relations is lower than the preset threshold, the existing entity-relation pairs in the knowledge graph are called to correct the results, and the corrected samples are returned to the BERT model for incremental training.

[0009] Preferably, the step of using the meta-path reasoning algorithm to mine implicit associations includes: Based on the cross-domain data features aligned by federated learning, an initial meta-path candidate set containing device-fault-business class meta-paths is generated; Network topology constraints and business rule constraints are imposed on the initial meta-path candidate set to eliminate paths that do not conform to the physical connection logic. At the same time, entity link state constraints are introduced. When the link packet loss rate is greater than the packet loss threshold, the priority of the meta-path related to the current link is reduced. We employ a weighted meta-path reasoning algorithm to incorporate implicit associations obtained through reasoning into the knowledge graph after passing a temporal correlation test.

[0010] Preferably, in the step of employing the weighted meta-path reasoning algorithm, the weights are jointly determined by the entity relationship confidence level extracted by the BERT model and the frequency of historical fault associations, specifically including: The initial weights are calculated by weighting the entity relationship confidence scores extracted from the BERT model and the frequency of historical fault associations, wherein the frequency of historical fault associations is assigned differentiated weights according to the fault level; the initial weights are expressed as follows: In the formula, Indicates the confidence level of the entity relationship. This represents the standardized value of the historical fault association frequency; Indicates the fault level coefficient. Indicates the initial weight allocation coefficients; Real-time monitoring of network load rate, a network status parameter, and dynamic adjustment of weight allocation are represented as follows: In the formula, Indicates network load rate. This represents the dynamic adjustment coefficient for the weights. Indicates the initial weight allocation coefficients. This represents the dynamically adjusted weighting coefficients; when the network load rate exceeds the preset threshold, the weighting percentage of historical fault association frequency is increased, while the weighting percentage of entity relationship confidence is decreased; when a new service type is added to the network, the weighting percentage of entity relationship confidence is increased, while the weighting percentage of historical fault association frequency is decreased. At each preset interval, the basic weight ratio is adjusted based on the accuracy of implicit association reasoning within that interval, forming a dynamic weight optimization mechanism.

[0011] Preferably, the step of constructing an incremental update strategy based on a time-series window for frequently changing business link data, so that the knowledge graph update latency is lower than the target value, includes: Based on the real-time load fluctuation coefficient and historical change patterns of the business links, a dynamic time-series window generation model is constructed, and the window duration dynamically expands and contracts with the load fluctuation rate of the links. A predictive update triggering mechanism is introduced, which uses an LSTM neural network to predict the trend of business link changes in the future period. If the predicted change exceeds the threshold, incremental update preprocessing is initiated. A layered update architecture is adopted, with edge nodes used for incremental updates of basic attributes of the business link, and the cloud used for the reconstruction of implicit relationships across links. Simultaneously, the update operation logs of each node are stored on the blockchain. When anomalies occur in update latency, the bottleneck node is located based on the logs, and computing resources are dynamically allocated to ensure that both the update latency of the business link and the average update latency across the entire link are below the target value. Preferably, a predictive update triggering mechanism is introduced, which uses an LSTM neural network to predict the trend of business link changes in the future, including the following steps: Construct an LSTM input layer that integrates multi-dimensional features. The input features include real-time load data of the business link, cross-domain environmental interference data aligned by federated learning, and business scheduling instruction entities extracted by the BERT model. Configure dynamic weights for different features. Attention mechanisms are introduced into the forget gate and input gate of the LSTM network. The forget gate dynamically adjusts the forget coefficient by calculating the matching degree between historical change data and the current business type, while the input gate filters redundant features based on the implicit association strength obtained from the meta-path reasoning, making the LSTM network more focused on key influencing factors. A prediction error feedback loop is established, inputting the deviation between each prediction result and the actual changed data into the network. Reinforcement learning is used to adjust the number of hidden layer neurons in the LSTM, where the number of hidden layer neurons is expressed as: In the formula, Indicates the prediction error. Indicates the error threshold. Indicates the adjustment factor. This indicates the adjusted number of neurons. This indicates the number of neurons before adjustment; Simultaneously, the feature weights are corrected by combining a dynamic weight optimization mechanism. The corrected feature weights are updated in real time to the feature weighting module of the LSTM input layer. When new real-time data of service links, cross-domain environmental interference data, and service scheduling instruction entities enter the input layer, the features are weighted and fused according to the updated weights, and prediction is performed using the input vector of the generated current network state.

[0012] Preferably, the step of extracting temporal features from the fault propagation subgraph using a graph convolutional network (GCN) includes: A hierarchical GCN structure with bidirectional time-series awareness is constructed, and the fault propagation subgraph is divided into three levels of dynamic subgraphs according to the causal chain of alarm triggering: the first level subgraph corresponds to the state characteristics of the fault source candidate node and the initial alarm time, the second level subgraph corresponds to the characteristics of the intermediate nodes affected within t seconds of fault propagation, and the third level subgraph corresponds to the characteristics of the terminal alarm node within 2t seconds of fault propagation. A node importance weight matrix is ​​introduced in the spatial feature extraction layer. This matrix is ​​dynamically generated based on the device association strength obtained from the meta-path inference, where the spatial weight of the i-th node... Represented as: In the formula, This represents the basic importance of node i. This represents the strength of the meta-path association between node i and its neighboring node j. Let n represent the set of adjacent nodes of node i, and n represent the total number of nodes in the subgraph. A temporal difference convolution module is introduced to perform temporal dimension fusion of features from the three-level subgraphs: the feature change rate between adjacent subgraphs is calculated using a first-order difference operator, and the change rate is expressed as: , The subgraph eigenvalues ​​at time t are represented; the acceleration of the eigenvalue changes is captured using a second-order difference operator, and the acceleration is expressed as: T represents the time interval; and the differential features and the original spatial features are input into the gated recurrent unit (GRU), and the GRU's reset gate and update gate dynamically filter key time nodes; The output enhanced feature vector includes spatial association weights of the three-level subgraphs, temporal change rates of adjacent subgraphs, and acceleration features. Furthermore, an attention mechanism dynamically weights each feature dimension. Represented as: In the formula, This represents the spatial correlation weight at time t. Represents the rate of change of the time series at time t. This represents the spatial association weight of the i-th level subgraph. This represents the temporal rate of change of the i-th level subgraph.

[0013] Preferably, the step of constructing a bi-objective game model of service SLA and network resources based on the positioning results includes: Convert the root cause confidence ranking in the location results into game priority coefficients; A dynamic compliance threshold is introduced into the business SLA objective function. The dynamic compliance threshold is determined based on the business type entity extracted by the BERT model and the fault impact duration feature extracted by the GCN. The network resource objective function incorporates the resource recovery cost coefficient. At the same time, combined with the real-time resource utilization rate updated by the time series window, a dynamic strategy space for the game participants is constructed. The fault propagation subgraph in the location result is used as the boundary. Nodes inside the subgraph are set as players, and nodes outside the subgraph are set as auxiliary players. The strategy execution authority of the players increases linearly with the root cause confidence ranking. A dual-objective balancing mechanism is set up. When the difference between the business SLA compliance rate and the resource utilization rate exceeds a preset threshold, the dynamic optimization mechanism of the weight is invoked to adjust the weight of the objective function, giving priority to the objective corresponding to a positive difference, and the adjustment result is synchronized to the strategy rule layer of the knowledge graph.

[0014] Secondly, according to another embodiment of this application, a knowledge graph-based communication network operation and maintenance device is provided, comprising the following modules: The data processing module is used to perform data desensitization and feature alignment through a federated learning mechanism based on a collaborative acquisition architecture deployed on a communication network, and to jointly extract entities and relations from unstructured data using a BERT pre-trained model. The knowledge graph dynamic construction module is used to identify new entity types in the network based on the dynamic ontology evolution model, utilize concept drift detection, and use the meta-path reasoning algorithm to mine implicit associations. It also constructs an incremental update strategy based on time-series windows for frequently changing business link data, so that the knowledge graph update latency is lower than the target value. The fault localization module is used to respond to fault signals by removing false associations from the knowledge graph through a causal graph model, extracting temporal features from the fault propagation subgraph using a graph convolutional network (GCN), and outputting the root cause confidence ranking by combining a root cause ranking model trained by reinforcement learning. The resource scheduling module is used to construct a dual-objective game model of service SLA and network resources based on the location results. It solves the resource scheduling scheme that satisfies the constraints through Nash equilibrium. After verification by digital twin simulation, the strategy execution deviation data is fed back to the knowledge graph for model parameter correction.

[0015] Thirdly, according to another embodiment of this application, a readable storage medium is provided, including a stored program, wherein, when the program is executed, it controls the device where the readable storage medium is located to execute a knowledge graph-based communication network operation and maintenance method.

[0016] Compared with existing technologies, this application utilizes a collaborative acquisition architecture deployed in a communication network. Data anonymization and feature alignment are performed through a federated learning mechanism. A BERT pre-trained model is used to jointly extract entities and relations from unstructured data. Based on a dynamic ontology evolution model, concept drift detection is used to identify new entity types in the network, and a meta-path reasoning algorithm is employed to mine implicit associations. An incremental update strategy based on a time-series window is constructed for frequently changing business link data, ensuring that the knowledge graph update latency is lower than the target value. In response to fault signals, a causal graph model is used to remove false associations from the knowledge graph, and a graph convolutional network (GCN) is used to perform time-series feature extraction on the fault propagation subgraph. This paper extracts and combines the root cause ranking model trained by reinforcement learning to output the root cause confidence ranking; based on the location results, it constructs a bi-objective game model of service SLA and network resources, solves the resource scheduling scheme that satisfies the constraints through Nash equilibrium, and verifies it through digital twin simulation. After verification, the policy execution deviation data is fed back to the knowledge graph for model parameter correction. This application solves the problem of difficult cross-domain data fusion and improves data utilization; by dynamically updating the knowledge graph, it ensures timeliness and accuracy; improves the accuracy and efficiency of fault location; realizes dynamic optimization scheduling of resources, balances service SLA and resource utilization, forms an adaptive iterative mechanism, and significantly improves the intelligence level and stability of communication network operation and maintenance. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] In the attached diagram: Figure 1 This is a flowchart illustrating the implementation of the knowledge graph-based communication network operation and maintenance method in this application. Figure 2 This is a sub-flowchart of the knowledge graph-based communication network operation and maintenance method according to an embodiment of this application; Figure 3 This is another sub-flowchart of the knowledge graph-based communication network operation and maintenance method according to an embodiment of this application; Figure 4 This is a structural block diagram of a knowledge graph-based communication network operation and maintenance device according to an embodiment of this application; Figure 5 This is a structural block diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0021] In this invention, unless otherwise stated, directional terms such as "upper," "lower," "top," and "bottom" are generally used in relation to the direction shown in the accompanying drawings, or in relation to the vertical, perpendicular, or gravitational direction of the component itself; similarly, for ease of understanding and description, "inner" and "outer" refer to the inner and outer contours of each component itself, but the above directional terms are not intended to limit this invention.

[0022] like Figure 1 As shown, according to one embodiment of this application, a knowledge graph-based communication network operation and maintenance method is provided, which includes the following steps: S101: Based on a collaborative acquisition architecture deployed in a communication network, data desensitization and feature alignment are performed through a federated learning mechanism, and entities and relations are jointly extracted from unstructured data using a BERT pre-trained model. Specifically, in the collaborative acquisition architecture of step S101, a three-level distributed deployment mode of edge nodes, regional gateways and cloud platform is adopted; In this embodiment, edge nodes are deployed at the network periphery, such as base stations and equipment rooms, and integrate multiple types of sensors and data acquisition interfaces to collect equipment operating parameters, link performance indicators and environmental data in real time. The collected data is also preliminarily cleaned, including filtering out null values ​​and correcting abnormal fluctuation values. In this embodiment, the regional gateway aggregates data from its subordinate edge nodes according to geographical partitions, performs data compression and format conversion through a lightweight protocol, and performs data desensitization preprocessing based on preset security rules, retaining only key feature fields; In this embodiment, the cloud platform serves as the core of collaboration, receiving standardized data uploaded by gateways in various regions, constructing a global data index library, and dynamically allocating computing resources through a load balancing mechanism to support subsequent data alignment in federated learning and entity extraction tasks in the BERT model, thereby achieving efficient collaborative collection and centralized management of cross-domain and multi-source data. Furthermore, in step S101 of this embodiment, the step of data anonymization and feature alignment through the federated learning mechanism is used to solve the problem of difficulty in achieving cross-domain sharing and collaborative analysis of communication networks. In addition, data anonymization is used to avoid data leakage problems caused by cross-domain transmission of network data. Specifically, the steps for data anonymization and feature alignment using a federated learning mechanism include: The edge nodes in the collaborative acquisition architecture are divided according to network domains as follows: The core network node group includes backbone network routers, core switches, mobile core network and other equipment. These devices are deployed in the computer room and are used for key performance indicators of the core network. Access network node groups, including 4G / 5G base stations, are distributed in buildings, sites, etc., and are used to collect access layer data such as user access rate and access failure reasons; Transmission network node groups, including optical cable lines, are deployed along the transmission link to collect transmission layer indicators such as fiber attenuation value, bit error rate (BER), and number of link interruptions. Furthermore, in this embodiment, each node group performs desensitization processing on local data based on a preset homomorphic encryption rule, generating an encrypted feature vector containing node physical location information, so that the feature vector has spatial topological attributes; wherein, in the implementation of the homomorphic encryption rule for desensitizing local data, a partial homomorphic encryption algorithm is used to achieve localized desensitization; Furthermore, the cloud-coordinated node distributes the feature alignment benchmark to each node group, and each node group completes the matching and correction of the feature vector with the benchmark locally, only uploading the corrected feature mapping relationship to the cloud; wherein, in this embodiment, the node group completes the alignment of the feature vector with the benchmark locally, avoiding the uploading of the original data; The cloud constructs a global feature alignment matrix based on the feature mapping relationship of each node group, and corrects the matrix weight by combining the transmission network link delay parameter. The matrix parameters are aggregated by the federated averaging algorithm and fed back to each node group to achieve semantic alignment of cross-domain data. The cloud uses the federated averaging algorithm to aggregate the matrix parameters to ensure global consistency. In addition, the aggregation process is iterated once per hour to keep the semantic alignment error of cross-domain features within 3%.

[0023] Furthermore, in step S101 of this embodiment, the step of jointly extracting entities and relations from unstructured data using a BERT pre-trained model enhances the quality of the corpus through a domain dictionary and employs dual task heads to achieve collaborative learning of entities and relations, thus solving the problem of extracting unstructured text from communication networks. Specifically, this step includes: Unstructured data from communication networks is preprocessed and categorized into equipment maintenance documents, fault alarm description texts, and service interaction logs. For each data type, a fine-tuned corpus containing a domain-specific terminology dictionary is generated. Specifically, in corpus construction, equipment maintenance documents include equipment model parameters, hardware structure, equipment operating procedures, and historical maintenance records; fault alarm description texts include alarm trigger time, alarm level, fault symptoms, and preliminary fault location results, providing crucial clues for root cause analysis; and service interaction logs include user service requests, service quality indicators, and service scheduling records, used to assist in analyzing whether service interruptions are caused by unreasonable resource scheduling strategies. When generating the fine-tuned corpus, each data type is segmented into sentences, and terms from the domain-specific dictionary are used as entity annotation seeds to generate tagged fine-tuned corpora. After obtaining the fine-tuned corpus, the corpus is input into a BERT model pre-trained based on communication domain corpus. By setting a dual-task head in the output layer, which includes an entity recognition head and a relation classification head, and adding association labels to the entity recognition head, entity boundary labels and entity relation labels are obtained. The association labels added in the entity recognition head contain contextual association information. In the relation classification head, a softmax classifier is used to output a predefined relation type for entity pairs. The dual-task head in this embodiment is used to achieve collaborative learning of entity boundaries and entity relations. Furthermore, a cross-validation mechanism is used for the output results. If the confidence of the joint extraction of entities and relations is lower than a preset threshold, the existing entity-relationship pairs in the knowledge graph are called to make corrections, and the corrected samples are returned to the BERT model for incremental training to update the parameters of the BERT model, so that the BERT model can be continuously optimized in actual operation and maintenance scenarios.

[0024] Please continue to refer to Figure 1 In one embodiment of this application, the knowledge graph-based communication network operation and maintenance method further includes step S102, wherein, in this step, based on the dynamic ontology evolution model, concept drift detection is used to identify new network entity types, and a meta-path reasoning algorithm is used to mine implicit associations. An incremental update strategy based on time-series windows is constructed for frequently changing business link data, so that the knowledge graph update latency is lower than the target value. In this embodiment, new devices, new services, and new alarm types will continuously emerge in the communication network. This embodiment dynamically incorporates them into the knowledge graph ontology system through concept drift detection. First, the core entity types and features of the initial ontology are defined. For example, device types include core features such as manufacturer, model, network domain, and number of interfaces; alarm types include alarm code, level, trigger threshold, and associated device features; the distribution range of each feature is calculated through historical data to form a feature baseline. For the determination of new entities, for unstructured data that comes in in real time, candidate entities that do not match the existing ontology are extracted. The JS divergence is used to measure the difference in feature distribution between the candidate entities and the existing types. When the JS divergence is greater than a preset threshold, it can be determined as concept drift, that is, the candidate entity is a new entity type. For the identified new entity type, the new entity type is integrated into the ontology through semantic association verification, feature stability verification and business adaptation verification.

[0025] For further details, please refer to Figure 2 This embodiment provides a method for mining implicit associations using the meta-path reasoning algorithm, which is achieved through the following steps: S201: Based on the cross-domain data features aligned by federated learning, generate an initial meta-path candidate set containing device-fault-service meta-paths; where device entities include routers, base stations, etc., and device attributes include network domain, vendor, deployment location, etc.; fault entities include alarm codes and fault types, and fault attributes include fault level and fault trigger threshold. In this embodiment, explicit relationships between entities are extracted, and initial candidate paths are generated based on the explicit relationships between entities such as devices, faults, and services. For example, the path centered on the fault entity is: Fault → Impact → Service → Dependence → Device; the path centered on the service entity is: Service → Association → Fault → Occurred at → Device; the path centered on the device entity is: Device → Connection → Device → Occurrence → Fault. Each candidate path contains a sequence of entity types and a sequence of relationship types, and the initial set size is controlled to be between 50 and 100 (to avoid path redundancy).

[0026] S202: Apply network topology constraints and business rule constraints to the initial meta-path candidate set, eliminate paths that do not conform to physical connection logic, and introduce entity link state constraints. When the link packet loss rate is greater than the packet loss threshold, reduce the priority of the current link-related meta-path. In step S202 of this embodiment, invalid paths are eliminated and priorities are dynamically adjusted through three layers of filtering: network topology constraints, business rule constraints, and link state constraints, to ensure that candidate paths conform to the physical characteristics and business logic of the communication network. S203: Employ a weighted meta-path reasoning algorithm to incorporate implicit associations obtained through reasoning into the knowledge graph after passing a temporal correlation test.

[0027] Specifically, in step S203, where the weighted meta-path reasoning algorithm is used, the weights are jointly determined by the entity relationship confidence level extracted by the BERT model and the frequency of historical fault associations. Specifically: The initial weights are calculated by weighting the entity relationship confidence extracted by the BERT model and the frequency of historical fault associations, wherein the frequency of historical fault associations is assigned differentiated weights according to the fault level. In one embodiment of this application, the initial weights are represented as: In the formula, Indicates the confidence level of the entity relationship. This represents the standardized value of the frequency of historical fault associations. Indicates the fault level coefficient. Indicates the initial weight allocation coefficients; Furthermore, this embodiment monitors the network load rate, a network status parameter, in real time and dynamically adjusts the weight allocation, as shown below: In the formula, Indicates network load rate. This represents the dynamic adjustment coefficient for the weights. Indicates the initial weight allocation coefficients. This represents the dynamically adjusted weighting coefficient; When the network load rate exceeds a preset threshold, the weighting of historical fault association frequency is increased, while the weighting of entity relationship confidence is decreased. When new business types are added to the network, the weighting of entity relationship confidence is increased, while the weighting of historical fault association frequency is decreased. Furthermore, at each preset interval, the basic weight ratio is adjusted based on the accuracy of implicit association reasoning within that interval, forming a dynamic weight optimization mechanism.

[0028] Please refer to Figure 3 In this embodiment, the step of constructing an incremental update strategy based on a time-series window for frequently changing business link data, so that the knowledge graph update latency is lower than the target value, includes: S301: Based on the real-time load fluctuation coefficient and historical change patterns of the business link, a dynamic time-series window generation model is constructed, and the window duration dynamically expands and contracts with the link load fluctuation rate. In this embodiment, the dynamic timing window is used to address high-frequency changes in service links, such as 5G slice bandwidth adjustments and VPN tunnel switching. This embodiment balances update timeliness and resource consumption through adaptive scaling of the window duration. Among them, for real-time load fluctuations, the degree of fluctuation of the current load of the business link is quantified and expressed as the real-time load volatility rate. The greater the volatility rate, the more severe the load fluctuation. In the step of dynamically scaling the window duration according to the link load volatility, the basic window duration is set to 60 seconds. When the load volatility is high, it indicates that the link status changes rapidly, and the window duration needs to be shortened. In this embodiment, the load volatility is dynamically compressed or stretched to match the update frequency with the actual speed of link changes, thus balancing timeliness and resource efficiency. S302: Introducing a predictive update triggering mechanism, which uses an LSTM neural network to predict the trend of business link changes in the future period. If the predicted change exceeds the threshold, incremental update preprocessing is initiated. S303: Adopts a layered update architecture. Edge nodes are used for incremental updates of basic attributes of business links, while the cloud is used for the reconstruction of implicit relationships across links. At the same time, the update operation logs of each node are stored on the blockchain. When an abnormal update latency occurs, the bottleneck node is located based on the logs and computing resources are dynamically allocated to ensure that the update latency of the business link and the average update latency of the entire link are both lower than the target value.

[0029] Furthermore, in step S302 of this embodiment, a predictive update triggering mechanism is introduced, which involves predicting the trend of business link changes in the future time period using an LSTM neural network. This includes: An LSTM input layer integrating multi-dimensional features is constructed. The input layer uses multi-dimensional feature fusion and dynamic weight allocation to enable the LSTM model to fully capture the key factors affecting changes in the business link and solve the problem of the one-sidedness of a single feature. The input features include real-time load data of the business link, cross-domain environmental interference data aligned by federated learning, and business scheduling instruction entities extracted by the BERT model. To avoid interference from irrelevant features, dynamic weights are configured for the above three types of features. For example, when configuring dynamic weights, the basic weight allocation is as follows: the weight of real-time load is configured as 0.4, the weight of environmental interference is configured as 0.2, and the weight of scheduling instructions is configured as 0.4. The adjustment rules include: when a high-priority emergency expansion instruction extracted by BERT is detected, the weight of the scheduling instruction is increased to 0.6, and the weight of real-time load is reduced to 0.3; when the neighboring cell link in the cross-domain environmental interference data is interrupted, the weight of environmental interference is increased to 0.3, squeezing out the weights of other features. In this embodiment, the weight adjustment is smoothly transitioned through the Sigmoid function to avoid abrupt changes that could cause model oscillations. Furthermore, this embodiment introduces an attention mechanism into the forget gate and input gate of the LSTM network. The forget gate dynamically adjusts the forget coefficient by calculating the matching degree between historical change data and the current business type, while the input gate filters redundant features based on the implicit association strength obtained from the meta-path reasoning, making the LSTM network more focused on key influencing factors and filtering out irrelevant features. Furthermore, this embodiment establishes a prediction error feedback loop, inputting the deviation between each prediction result and the actual changed data into the network, and adjusting the number of hidden layer neurons in the LSTM through reinforcement learning. In this embodiment, the number of hidden layer neurons is expressed as: In the formula, Indicates the prediction error. Indicates the error threshold. Indicates the adjustment factor. This indicates the adjusted number of neurons. This indicates the number of neurons before adjustment; Simultaneously, the feature weights are corrected by combining a dynamic weight optimization mechanism. The corrected feature weights are updated in real time to the feature weighting module of the LSTM input layer. When new real-time data of service links, cross-domain environmental interference data, and service scheduling instruction entities enter the input layer, the features are weighted and fused according to the updated weights, and prediction is performed using the input vector of the generated current network state.

[0030] Please continue to refer to Figure 1 In some embodiments of this application, the knowledge graph-based communication network operation and maintenance method further includes the following steps: S103: In response to fault signals, false associations are removed from the knowledge graph through a causal graph model, and temporal features are extracted from the fault propagation subgraph using a graph convolutional network (GCN). The root cause ranking model trained by reinforcement learning is combined to output the root cause confidence ranking. When a fault signal occurs in the communication network (such as equipment alarm, service interruption notification, etc.), the causal graph model is used to filter the entity associations related to the fault in the knowledge graph, eliminating false associations that seem related but have no actual causal connection, thus retaining the associations with real causal logic. The causal graph model verifies the associations related to the current fault signal in the knowledge graph by introducing causal reasoning rules. Among them, associations that do not meet the causal sequence in time are marked as false associations, and associations that have no actual causal impact as proven by intervention analysis are removed. Furthermore, in step S103 of this embodiment, the step of extracting temporal features from the fault propagation subgraph using a graph convolutional network (GCN) includes: In addition, this application embodiment constructs a hierarchical graph convolutional network (GCN) with bidirectional time-series awareness to address the temporal characteristics and hierarchical relationships of communication network fault propagation. According to the causal logic chain triggered by fault alarms, the subgraph containing the fault propagation path is dynamically divided into three levels to accurately capture the propagation pattern of faults from the source to the terminal. Specifically, by constructing a hierarchical GCN structure with bidirectional time-awareness, the fault propagation subgraph is divided into three levels of dynamic subgraphs according to the causal chain of alarm triggering. In the three levels of dynamic subgraphs, the first level subgraph corresponds to the state features of the fault source candidate nodes and the initial alarm time; the second level subgraph corresponds to the features of the intermediate nodes affected within t seconds of fault propagation; and the third level subgraph corresponds to the features of the terminal alarm nodes within 2t seconds of fault propagation. The division of the three levels of dynamic subgraphs can clearly present the stages of fault propagation, allowing the hierarchical graph convolutional network (GCN) to extract the key features of each stage more efficiently. Furthermore, in this embodiment, a node importance weight matrix is ​​introduced in the spatial feature extraction layer. The node importance weight matrix is ​​dynamically generated based on the device association strength obtained from the meta-path reasoning. Wherein, the spatial weight of the i-th node Represented as: In the formula, This represents the basic importance of node i. This represents the strength of the meta-path association between node i and its neighboring node j. Let n represent the set of adjacent nodes of node i, and n represent the total number of nodes in the subgraph. Furthermore, in this embodiment, a temporal difference convolution module is introduced to perform temporal dimension fusion of the features of the third-level subgraphs: the feature change rate of adjacent subgraphs is calculated using a first-order difference operator, and the change rate is expressed as: , The subgraph eigenvalues ​​at time t are represented; the acceleration of the eigenvalue changes is captured using a second-order difference operator, and the acceleration is expressed as: T represents the time interval; and the differential features and the original spatial features are input into the gated recurrent unit (GRU), and the GRU's reset gate and update gate dynamically filter key time nodes; Furthermore, in this embodiment, the output enhanced feature vector includes spatial association weights of the three-level subgraphs, temporal change rates of adjacent subgraphs, and acceleration features. An attention mechanism is used to dynamically weight each feature dimension, with the feature dimension weights... Represented as: In the formula, This represents the spatial correlation weight at time t. Represents the rate of change of the time series at time t. This represents the spatial association weight of the i-th level subgraph. This represents the temporal rate of change of the i-th level subgraph.

[0031] Please continue to refer to Figure 1 In some embodiments of this application, the knowledge graph-based communication network operation and maintenance method further includes the following steps: S104: Based on the location results, construct a dual-objective game model of service SLA and network resources, solve the resource scheduling scheme that satisfies the constraints through Nash equilibrium, and after verification by digital twin simulation, feed the policy execution deviation data back to the knowledge graph to correct the model parameters.

[0032] In some embodiments of this application, the location result is the root cause location result of the fault, such as determining the root cause of service interruption caused by congestion of a certain transmission chain. This result is used to clarify the correlation between resource bottlenecks in the current network and the impact on services. Service SLA stands for Service Level Agreement, which is a service quality indicator agreed upon by the service provider and the network operator. In the bi-objective game model of this embodiment, the essence of the game is reflected in the service provider pursuing the maximization of SLA and the network side pursuing the maximization of resource utilization. The game model transforms resource scheduling into a dynamic balance problem of interests between the two parties by quantifying the goals of both parties. Furthermore, in this embodiment, the step of constructing a bi-objective game model of service SLA and network resources based on the location results includes: converting the root cause confidence ranking in the location results into game priority coefficients. Furthermore, a dynamic compliance threshold is introduced into the business SLA objective function. The dynamic compliance threshold is determined based on the business type entities extracted by the BERT model and the fault impact duration features extracted by the GCN. The network resource objective function incorporates the resource recovery cost coefficient. At the same time, combined with the real-time resource utilization rate updated by the time-series window, a dynamic strategy space for the game participants is constructed. Taking the fault propagation subgraph in the location result as the boundary, the nodes inside the subgraph are set as the game players, and the nodes outside the subgraph are set as auxiliary game players. Moreover, the strategy execution authority of the game players increases linearly with the root cause confidence ranking. Furthermore, in this embodiment, by setting a dual-objective balancing mechanism, when the difference between the business SLA compliance rate and the resource utilization rate exceeds a preset threshold, the weight dynamic optimization mechanism is invoked to adjust the weight of the objective function, prioritizing the objective corresponding to a positive difference, and the adjustment result is synchronized to the strategy rule layer of the knowledge graph.

[0033] In summary, this application solves the problem of difficult cross-domain data fusion and improves data utilization; ensures timeliness and accuracy by dynamically updating the knowledge graph; improves fault location accuracy and efficiency; realizes dynamic optimization and scheduling of resources, balances service SLA and resource utilization, forms an adaptive iteration mechanism, and significantly improves the intelligence level and stability of communication network operation and maintenance.

[0034] Please continue to refer to Figure 4 In another embodiment of this application, a knowledge graph-based communication network operation and maintenance device is provided. Specifically, the operation and maintenance device provided in this application embodiment includes the following modules: The data processing module 401 is used to perform data desensitization and feature alignment through a federated learning mechanism based on a collaborative acquisition architecture deployed in a communication network, and to jointly extract entities and relations from unstructured data using a BERT pre-trained model. The knowledge graph dynamic construction module 402 is used to identify new entity types in the network based on the dynamic ontology evolution model, utilize concept drift detection, and use the meta-path reasoning algorithm to mine implicit associations. It also constructs an incremental update strategy based on time-series windows for frequently changing business link data, so that the knowledge graph update latency is lower than the target value. The fault localization module 403 is used to respond to fault signals by removing false associations from the knowledge graph through a causal graph model, extracting temporal features from the fault propagation subgraph using a graph convolutional network (GCN), and outputting the root cause confidence ranking by combining a root cause ranking model trained by reinforcement learning. The resource scheduling module 404 is used to construct a dual-objective game model of service SLA and network resources based on the location results. It solves the resource scheduling scheme that satisfies the constraints through Nash equilibrium. After verification by digital twin simulation, the strategy execution deviation data is fed back to the knowledge graph for model parameter correction.

[0035] Figure 5 An internal structural diagram of a computer device in one embodiment is shown.

[0036] Specifically, such as Figure 5 As shown, the computer device includes a processor, memory, network interface, input device, and display screen connected via a system bus. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and may also store computer programs. When executed by the processor, these programs enable the processor to implement a knowledge graph-based communication network operation and maintenance method.

[0037] The internal memory may also store computer programs, which, when executed by the processor, enable the processor to perform knowledge graph-based communication network operation and maintenance methods.

[0038] Those skilled in the art will understand that Figure 5The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0039] Another embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the knowledge graph-based communication network operation and maintenance method described above.

[0040] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0041] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the above embodiments of the communication network operation and maintenance method based on knowledge graphs.

[0042] In the embodiments provided by this invention, any references to memory, storage, database, or other media may include both non-volatile and volatile memory. Non-volatile memory may include read-only memory, programmable ROM, electrically programmable ROM, electrically erasable programmable ROM, or flash memory. Volatile memory may include random access memory or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM, dynamic RAM, synchronous DRAM, dual data rate SDRAM, enhanced SDRAM, synchronous link DRAM, memory bus direct RAM, direct memory bus dynamic RAM, and memory bus dynamic RAM, etc.

[0043] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0044] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0045] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge graph-based communication network operation and maintenance method, characterized in that, Includes the following steps: Based on a collaborative acquisition architecture deployed in a communication network, data desensitization and feature alignment are performed through a federated learning mechanism, and entities and relations are jointly extracted from unstructured data using a BERT pre-trained model. Based on a dynamic ontology evolution model, this study utilizes concept drift detection to identify new entity types in the network and employs a meta-path reasoning algorithm to mine implicit relationships. For frequently changing business link data, a time-series window-based incremental update strategy is constructed to ensure that the knowledge graph update latency is lower than the target value. This includes: constructing a dynamic time-series window generation model based on the real-time load fluctuation coefficient and historical change patterns of business links, with the window duration dynamically scaling with the link load fluctuation rate; introducing a predictive update triggering mechanism, using an LSTM neural network to predict business link change trends in future periods; if the predicted change amount exceeds a threshold, incremental update preprocessing is initiated; and adopting a hierarchical update architecture, with edge nodes used for incremental updates of basic attributes of business links, and the cloud used for reconstructing implicit relationships across links. Simultaneously, the update operation logs of each node are stored on the blockchain. When update latency anomalies occur, bottleneck nodes are traced and located based on the logs, and computing resources are dynamically allocated, ensuring that both the business link update latency and the overall link average update latency are lower than the target value. The predictive update triggering mechanism, using an LSTM neural network to predict business link changes in future periods... The steps for predicting change trends include: constructing an LSTM input layer that integrates multi-dimensional features, including real-time load data of business links, cross-domain environmental interference data aligned by federated learning, and business scheduling instruction entities extracted by the BERT model, and configuring dynamic weights for different features; introducing an attention mechanism into the forget gate and input gate of the LSTM network, wherein the forget gate dynamically adjusts the forgetting coefficient by calculating the matching degree between historical change data and the current business type, and the input gate filters redundant features based on the implicit association strength obtained by the meta-path inference; establishing a prediction error feedback loop, inputting the deviation value between each prediction result and the actual change data into the network, and adjusting the number of hidden layer neurons of the LSTM through reinforcement learning; and simultaneously correcting the feature weights by combining a dynamic weight optimization mechanism, updating the corrected feature weights to the feature weighting module of the LSTM input layer in real time, and when new real-time data of business links, cross-domain environmental interference data, and business scheduling instruction entities enter the input layer, the features are weighted and fused according to the updated weights, and prediction is performed using the input vector of the generated current network state; In response to fault signals, a causal graph model is used to remove false associations from the knowledge graph, and a graph convolutional network (GCN) is used to extract temporal features from the fault propagation subgraph. A root cause ranking model trained by reinforcement learning is then used to output root cause confidence rankings. The steps for extracting temporal features from the fault propagation subgraph using the GCN include: constructing a hierarchical GCN structure with bidirectional temporal awareness, dividing the fault propagation subgraph into three levels of dynamic subgraphs according to the causal chain triggered by the alarm: the first-level subgraph corresponds to the state features of the fault source candidate nodes and the initial alarm time; the second-level subgraph corresponds to the features of the intermediate nodes affected within t seconds of fault propagation; and the third-level subgraph corresponds to the features of the terminal alarm nodes within 2t seconds of fault propagation. A node importance weight matrix is ​​introduced in the spatial feature extraction layer. This matrix is ​​dynamically generated based on the device association strength obtained from the meta-path inference. A temporal differential convolution module is introduced to fuse the features of the three-level subgraphs in a temporal dimension: the feature change rate of adjacent subgraphs is calculated using a first-order difference operator; the acceleration of feature change is captured using a second-order difference operator; and the differential features and the original spatial features are input into a gated recurrent unit (GRU). The GRU's reset and update gates dynamically filter key temporal nodes. The output enhanced feature vector contains the spatial association weights of the three-level subgraphs, the temporal change rate of adjacent subgraphs, and acceleration features. An attention mechanism is used to dynamically weight each feature dimension. Based on the location results, a dual-objective game model of business SLA and network resources is constructed. The resource scheduling scheme that satisfies the constraints is solved by Nash equilibrium. After verification by digital twin simulation, the policy execution deviation data is fed back to the knowledge graph for model parameter correction.

2. The knowledge graph-based communication network operation and maintenance method according to claim 1, characterized in that, The steps for jointly extracting entities and relations from unstructured data using a BERT pre-trained model include: Unstructured data from communication networks is preprocessed and categorized into equipment maintenance documents, fault alarm description texts, and business interaction logs based on data type. Fine-tuned corpora containing a domain terminology dictionary are generated for each type of data. The fine-tuned corpus is input into the BERT model pre-trained based on corpus from the communication domain. By setting a dual task head in the output layer, which includes an entity recognition head and a relation classification head, and adding association labels to the entity recognition head, entity boundary labels and relation labels between entities are obtained. The output results are cross-validated. If the confidence of the joint extraction of entities and relations is lower than the preset threshold, the existing entity-relation pairs in the knowledge graph are called to correct the results, and the corrected samples are returned to the BERT model for incremental training.

3. The knowledge graph-based communication network operation and maintenance method according to claim 2, characterized in that, The steps for mining implicit associations using the meta-path reasoning algorithm include: Based on the cross-domain data features aligned by federated learning, an initial meta-path candidate set containing device-fault-business class meta-paths is generated; Apply network topology constraints and business rule constraints to the initial meta-path candidate set, and eliminate paths that do not conform to the physical connection logic; Introduce entity link state constraints: when the link packet loss rate exceeds the packet loss threshold, reduce the priority of the relevant meta-paths of the current link. We employ a weighted meta-path reasoning algorithm to incorporate implicit associations obtained through reasoning into the knowledge graph after passing a temporal correlation test.

4. The knowledge graph-based communication network operation and maintenance method according to claim 3, characterized in that, In the steps of the weighted meta-path reasoning algorithm, the weights are jointly determined by the entity relationship confidence level extracted from the BERT model and the frequency of historical fault associations, specifically including: The initial weights are calculated by weighting the entity relationship confidence scores extracted from the BERT model and the frequency of historical fault associations, wherein the frequency of historical fault associations is assigned differentiated weights according to the fault level; the initial weights are expressed as follows: In the formula, Indicates the confidence level of the entity relationship. This represents the standardized value of the frequency of historical fault associations. Indicates the fault level coefficient. Indicates the initial weight allocation coefficients; Real-time monitoring of network load rate, a network status parameter, and dynamic adjustment of weight allocation are represented as: In the formula, Indicates network load rate. This represents the dynamic adjustment coefficient for the weights. Indicates the initial weight allocation coefficients. This represents the dynamically adjusted weighting coefficient; When the network load rate exceeds a preset threshold, the weighting of historical fault association frequency is increased, while the weighting of entity relationship confidence is decreased. When new business types are added to the network, the weighting of entity relationship confidence is increased, while the weighting of historical fault association frequency is decreased. At each preset interval, the basic weight ratio is adjusted based on the accuracy of implicit association reasoning within that interval, forming a dynamic weight optimization mechanism.

5. The knowledge graph-based communication network operation and maintenance method according to claim 4, characterized in that, In the step of adjusting the number of hidden layer neurons in the LSTM through reinforcement learning, the number of hidden layer neurons is expressed as: In the formula, Indicates the prediction error. Indicates the error threshold. Indicates the adjustment factor. This indicates the adjusted number of neurons. This indicates the number of neurons before adjustment.

6. The knowledge graph-based communication network operation and maintenance method according to claim 5, characterized in that, In the step of dynamically generating the node importance weight matrix based on the device association strength obtained from the meta-path inference, the spatial weight of the i-th node is... Represented as: In the formula, This represents the basic importance of node i. This represents the strength of the meta-path association between node i and its neighboring node j. Let n represent the set of adjacent nodes of node i, and n represent the total number of nodes in the subgraph. The rate of change is expressed as: , The subgraph eigenvalues ​​representing time t; Acceleration is expressed as: T represents the time interval; In the step of dynamically weighting each feature dimension using an attention mechanism, the feature dimension weights... Represented as: In the formula, Represents the spatial correlation weight at time t. Represents the rate of change of the time series at time t. This represents the spatial association weight of the i-th level subgraph. This represents the temporal rate of change of the i-th level subgraph.

7. The knowledge graph-based communication network operation and maintenance method according to claim 6, characterized in that, The steps for constructing a bi-objective game model of service SLA and network resources based on location results include: Convert the root cause confidence ranking in the location results into game priority coefficients; A dynamic compliance threshold is introduced into the business SLA objective function. The dynamic compliance threshold is determined based on the business type entity extracted by the BERT model and the fault impact duration feature extracted by the GCN. The network resource objective function incorporates the resource recovery cost coefficient. At the same time, combined with the real-time resource utilization rate updated by the time series window, a dynamic strategy space for the game participants is constructed. The fault propagation subgraph in the location result is used as the boundary. Nodes inside the subgraph are set as players, and nodes outside the subgraph are set as auxiliary players. The strategy execution authority of the players increases linearly with the root cause confidence ranking. A dual-objective balancing mechanism is set up. When the difference between the business SLA compliance rate and the resource utilization rate exceeds a preset threshold, the dynamic optimization mechanism of the weight is invoked to adjust the weight of the objective function, giving priority to the objective corresponding to a positive difference, and the adjustment result is synchronized to the strategy rule layer of the knowledge graph.

8. An operation and maintenance apparatus for implementing the knowledge graph-based communication network operation and maintenance method as described in any one of claims 1 to 7, characterized in that, The maintenance equipment includes the following modules: The data processing module is used to perform data desensitization and feature alignment through a federated learning mechanism based on a collaborative acquisition architecture deployed on a communication network, and to jointly extract entities and relations from unstructured data using a BERT pre-trained model. The knowledge graph dynamic construction module is used to identify new entity types in the network based on a dynamic ontology evolution model, utilizing concept drift detection, and mining implicit relationships using a meta-path reasoning algorithm. It constructs an incremental update strategy based on time-series windows for frequently changing business link data, ensuring that the knowledge graph update latency is lower than the target value. This includes: constructing a dynamic time-series window generation model based on the real-time load fluctuation coefficient and historical change patterns of business links, with the window duration dynamically scaling with the link load fluctuation rate; introducing a predictive update triggering mechanism, using an LSTM neural network to predict business link change trends in future periods; if the predicted change amount exceeds a threshold, incremental update preprocessing is initiated; adopting a hierarchical update architecture, with edge nodes used for incremental updates of basic attributes of business links, and the cloud used for reconstructing implicit relationships across links; simultaneously, the update operation logs of each node are stored on the blockchain; when update latency anomalies occur, bottleneck nodes are traced and dynamically allocated computing resources based on the logs, ensuring that both business link update latency and the average update latency across the entire link are lower than the target value; and introducing a predictive update triggering mechanism, using an LSTM neural network to predict future periods... The steps for predicting changes in business links within the network include: constructing an LSTM input layer that integrates multi-dimensional features, including real-time load data of business links, cross-domain environmental interference data aligned by federated learning, and business scheduling instruction entities extracted by the BERT model, and configuring dynamic weights for different features; introducing an attention mechanism into the forget gate and input gate of the LSTM network, wherein the forget gate dynamically adjusts the forgetting coefficient by calculating the matching degree between historical change data and the current business type, and the input gate filters redundant features based on the implicit association strength obtained by the meta-path inference; establishing a prediction error feedback loop, inputting the deviation value between each prediction result and the actual change data into the network, and adjusting the number of hidden layer neurons of the LSTM through reinforcement learning; simultaneously correcting the feature weights by combining a dynamic weight optimization mechanism, updating the corrected feature weights to the feature weighting module of the LSTM input layer in real time, and when new real-time data of business links, cross-domain environmental interference data, and business scheduling instruction entities enter the input layer, the features are weighted and fused according to the updated weights, and prediction is performed using the input vector of the generated current network state; The fault localization module, in response to fault signals, uses a causal graph model to remove false associations from the knowledge graph and employs a graph convolutional network (GCN) to extract temporal features from the fault propagation subgraph. It then combines this with a root cause ranking model trained through reinforcement learning to output root cause confidence rankings. The steps for extracting temporal features from the fault propagation subgraph using the GCN include: constructing a hierarchical GCN structure with bidirectional temporal awareness; dividing the fault propagation subgraph into three levels of dynamic subgraphs based on the causal chain triggered by the alarm: the first-level subgraph corresponds to the state features of the fault source candidate nodes and the initial alarm time; the second-level subgraph corresponds to the features of the intermediate nodes affected within t seconds of fault propagation; and the third-level subgraph corresponds to the terminal alarms within 2t seconds of fault propagation. The system identifies key nodes and introduces a node importance weight matrix in the spatial feature extraction layer. This matrix is ​​dynamically generated based on the device association strength obtained from the meta-path inference. A temporal differential convolution module is introduced to fuse the features of the three-level subgraphs temporally: a first-order difference operator calculates the feature change rate of adjacent subgraphs; a second-order difference operator captures the acceleration of feature changes; and the differential features and original spatial features are input into a gated recurrent unit (GRU), where the reset and update gates dynamically filter key temporal nodes. The output enhanced feature vector includes the spatial association weights of the three-level subgraphs, the temporal change rate of adjacent subgraphs, and acceleration features, and dynamically weights each feature dimension through an attention mechanism. The resource scheduling module is used to construct a dual-objective game model of service SLA and network resources based on the location results. It solves the resource scheduling scheme that satisfies the constraints through Nash equilibrium. After verification by digital twin simulation, the strategy execution deviation data is fed back to the knowledge graph for model parameter correction.

9. A readable storage medium, characterized in that, The system includes a stored program, wherein, when the program is executed, it controls the device containing the readable storage medium to perform the knowledge graph-based communication network operation and maintenance method as described in any one of claims 1 to 7.

Images