Conflict detection method and system and electronic equipment

Abstract

The invention discloses a conflict detection method and system and electronic equipment, and relates to the technical field of artificial intelligence. The method comprises the steps of updating a knowledge graph based on an analysis result of a to-be-detected scheme; utilizing the updated knowledge graph to extract entity association features and context constraint features corresponding to the to-be-detected scheme; fusing the entity association feature, the context constraint feature and the entity association link information in the updated knowledge graph to obtain an input feature; inputting the input features into a detection model based on a hybrid expert model MOE framework to obtain a conflict detection result; wherein the detection model comprises a sparse routing network and a plurality of special expert models, the sparse routing network is used for matching the special expert models for the input features, and the special expert models are used for executing special conflict detection according to the input features.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a conflict detection method, system and electronic device. Background Technology

[0002] As telecom operators' networks continue to expand, network fault types are becoming increasingly complex. Network management platforms need to perform conflict detection on massive amounts of fault handling solutions to avoid problems such as fault handling failures and network service interruptions due to conflicting solutions. Currently, the industry generally uses structured data storage (such as relational databases) to record fault information and handling steps, and conducts conflict screening through manually preset rules or simple semantic comparison.

[0003] However, existing solutions can only store isolated data such as faults, operations, and resources, and perform indiscriminate conflict detection using a single model or uniform rules. This results in conflict detection lacking full context support, easily missing hidden conflicts, and involving a large amount of invalid computation and insufficient real-time performance. It is also difficult to accurately identify complex conflicts and adapt to new fault scenarios, making it impossible to meet the actual needs of network operation and maintenance in terms of detection accuracy, efficiency, and adaptability. Summary of the Invention

[0004] In view of this, this application provides a collision detection method, system, and electronic device. It aims to address the technical problems of insufficient detection performance and scenario adaptability in existing collision detection schemes.

[0005] Firstly, this application provides a conflict detection method, including: The knowledge graph is updated based on the analysis results of the scheme to be detected.

[0006] Using the updated knowledge graph, entity association features and context constraint features corresponding to the scheme to be detected are extracted.

[0007] The input features are obtained by integrating entity association features, context constraint features, and entity association link information from the updated knowledge graph.

[0008] Input features are fed into a detection model based on the hybrid expert model (MOE) framework to obtain conflict detection results. The detection model includes a sparse routing network and multiple specialized expert models. The sparse routing network is used to match specialized expert models with the input features, and the specialized expert models are used to perform specialized conflict detection based on the input features.

[0009] Secondly, this application provides a conflict detection system, which includes a knowledge graph module, a feature extraction module, a feature fusion module, and a conflict detection module; wherein: The knowledge graph module is configured to update the knowledge graph based on the parsing results of the scheme to be detected.

[0010] The feature extraction module is configured to use the updated knowledge graph to extract entity association features and context constraint features corresponding to the scheme to be detected.

[0011] The feature fusion module is configured to fuse entity association features, context constraint features, and entity association link information from the updated knowledge graph to obtain input features.

[0012] The conflict detection module is configured to input input features into a detection model based on the hybrid expert model (MOE) framework to obtain conflict detection results. The detection model includes a sparse routing network and multiple specialized expert models. The sparse routing network is used to match specialized expert models with the input features, and the specialized expert models are used to perform specialized conflict detection based on the input features.

[0013] Thirdly, this application provides an electronic device, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the computer program to implement the method of the first aspect.

[0014] In view of the above embodiments, this application provides a conflict detection method, system, and electronic device. This application establishes deep relationships between entities based on knowledge graphs, providing contextual support for conflict detection and effectively reducing the missed detection of implicit conflicts. Furthermore, this application utilizes a detection model constructed through the MOE model framework to achieve intelligent division of labor for conflict detection tasks, activating appropriate specialized expert models on demand to perform detection, thereby reducing unnecessary computation and improving detection efficiency.

[0015] Because specialized expert models are designed for specific conflict types, they can focus on optimizing the core detection logic of the corresponding conflict. Knowledge graphs, on the other hand, contain entity association information and contextual constraint features, which can fully present the related relationships and constraints of conflicts. Furthermore, knowledge graphs support dynamic updates based on the parsing results of the solution to be detected, allowing for the real-time addition of new entities and relationships, and rapid adaptation to new faults and operational scenarios. Therefore, by combining specialized expert models with the dynamic updating capabilities and association constraint features of knowledge graphs for specialized detection, complex conflicts that are difficult to capture using traditional methods can be accurately identified, comprehensively addressing the shortcomings of existing conflict detection solutions in terms of overall performance and scenario adaptability.

[0016] It should be noted that the above content is only a general overview of the technical solution of this application. In order to enable those skilled in the art to clearly understand the core technical means of this application and to accurately implement this solution based on the content disclosed in the specification, the technical details of this application will be described in detail below in conjunction with specific embodiments. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a conflict detection method provided in an embodiment of this application is shown. Figure 2 This paper shows an example diagram of the hybrid expert model (MOE) framework structure applicable to a conflict detection method provided in an embodiment of this application; Figure 3 This paper illustrates an example diagram of a knowledge graph construction process applicable to a conflict detection method provided in an embodiment of this application. Figure 4 This paper illustrates an example of a feature extraction process applicable to a conflict detection method provided in an embodiment of this application. Figure 5 This paper illustrates an example diagram of the feature fusion process applicable to a conflict detection method provided in an embodiment of this application. Figure 6 This illustration shows an example of a conflict detection process applicable to a conflict detection method provided in an embodiment of this application; Figure 7 The illustration shows an example of an implementation of the conflict detection method provided in this application. Figure 8 A schematic diagram of the structure of a collision detection system provided in an embodiment of this application is shown. Detailed Implementation

[0020] To facilitate the explanation of the embodiments of this application, some technical terms and technical means related to the embodiments of this application, as well as the application scenarios of the embodiments of this application, will be introduced first below.

[0021] A knowledge graph (KG) is a structured form of knowledge representation. It models and stores core entities and their relationships from multiple data sources using triples (entity-relationship-entity), providing a clear visual representation of the logical links and constraints between entities and offering comprehensive contextual support for complex reasoning.

[0022] The Mixture of Experts Model (MOE) is a collaborative detection framework that includes a sparse routing network and multiple specialized expert models. The routing network can calculate the matching degree based on the features of the input data and activate the appropriate expert models to perform specialized tasks, achieving precise division of detection tasks and reducing unnecessary computation.

[0023] The embodiments of this application can be applied to conflict detection scenarios in fault handling schemes in multiple fields (such as communications, power grids, and the Internet of Things). That is, in various scenarios where system faults need to be resolved by formulating handling schemes, conflict screening is performed on newly added or pending fault handling schemes to avoid scheme conflicts leading to handling failures, system downtime, or increased losses.

[0024] For example, in the field of communications, telecom operators' network management platforms need to perform conflict detection on solutions for network faults such as port congestion, link interruption, and device offline to ensure the continuous and stable operation of the communication network.

[0025] For example, in the power grid sector, power control systems conduct conflict verification on handling plans for scenarios such as line faults and equipment anomalies to avoid triggering a chain reaction in the power grid during handling operations.

[0026] For example, in the field of the Internet of Things (IoT), IoT platforms conduct conflict screening for handling solutions such as terminal access failures and data transmission anomalies to ensure the coordinated operation of IoT terminal clusters.

[0027] Furthermore, the embodiments of this application can also be applied to scenarios such as industrial control and cloud computing infrastructure that require standardized solutions to handle faults. They are applicable to various systems that require security and timeliness in fault handling, and will not be elaborated further.

[0028] In the conflict detection scenarios of fault handling solutions in the aforementioned fields, existing conflict detection solutions suffer from technical problems such as missing contextual information, low detection efficiency, and insufficient detection accuracy and generalization ability. Specifically: Lacking contextual information, existing conflict detection schemes use structured data storage methods, which can only record isolated faults, operations, or resource data. They cannot establish deep relationships between these three, resulting in a lack of full contextual support during conflict detection, making it easy to miss implicit conflicts such as resource indirect dependency conflicts.

[0029] The detection efficiency is low. Existing conflict detection solutions rely on a single model or manually preset rules to perform indiscriminate detection. When faced with a large number of concurrent fault handling solutions in the scenario, the unified detection process needs to be repeatedly executed, resulting in a large amount of invalid calculations, which makes it difficult to meet the real-time detection requirements of large-scale scenarios.

[0030] The detection accuracy and generalization ability are insufficient. Existing conflict detection schemes cannot accurately identify complex conflicts such as cross-scheme parameter contradictions and time-series loops through simple semantic comparison or a single rule engine. Moreover, manual rules are difficult to quickly adapt to new faults and operation modes emerging in the scenario, resulting in weak generalization ability.

[0031] This application provides a conflict detection method that uses a knowledge graph to achieve entity association and context modeling, overcoming the shortcomings of traditional storage methods in terms of information loss and solving the problem of missed detections caused by insufficient context support in existing solutions. By relying on the sparse routing of the MOE model architecture and the division of labor among specialized expert models, detection efficiency and accuracy in identifying complex conflicts are improved. Simultaneously, the dynamic updating capability of the knowledge graph is leveraged to adapt to new scenarios, effectively addressing the problems of insufficient generalization ability and weak scenario adaptability in existing solutions.

[0032] The embodiments of this application will now be described in more detail with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0033] like Figure 1 The flowchart shown illustrates a conflict detection method, illustrating the execution process of the conflict detection method of this application. The execution process of this conflict detection method includes the following steps S101 to S104: S101. Update the knowledge graph based on the analysis results of the scheme to be detected.

[0034] The proposed fault handling schemes to be tested include, for example, newly configured fault handling schemes without conflict detection in the conflict detection scenarios of the fault handling schemes in the aforementioned fields, or fault handling schemes that failed conflict detection and were resubmitted for testing after modification. Taking the communications field as an example, these fault handling schemes include core content such as fault type (e.g., port congestion, link interruption, device offline, etc.), corresponding handling operations (e.g., port restart, link switching, parameter adjustment, etc.), network resources involved in the operation (e.g., backup links, core switches, servers, etc.), operation timing requirements (e.g., performing data backup before port restart), and configuration parameter constraints (e.g., bandwidth adjustment limit, device operating threshold, etc.).

[0035] The parsing result is a structured extraction result of the solution to be tested. For example, natural language processing technology is used to identify and extract entity information such as faults, operations, and resources from the text of the solution to be tested, as well as key information such as the relationship between entities (such as the correspondence between faults and handling operations, and the dependency relationship between operations and resources), temporal constraints, and parameter requirements, and convert them into a standardized format that can be recognized by a knowledge graph.

[0036] In some embodiments, the parsing results are connected to the knowledge graph via a graph database interface. Adding, supplementing, or updating operations are performed on the entities, relationships, and attribute information included in the parsing results to obtain the updated knowledge graph.

[0037] S102. Using the updated knowledge graph, extract the entity association features and context constraint features corresponding to the scheme to be detected.

[0038] The entity association features corresponding to the scheme to be detected are extracted from the association links between entities in the knowledge graph, representing the correspondence and dependency relationships between entities related to the scheme (such as faults, operations, resources, etc.). Entity association features can reflect the logical association links between entities. For example, the "fault-corresponding handling operation" association between a fault entity and its corresponding handling operation entity; the "handling operation-involved resource" association between a handling operation entity and a network resource entity; and the "resource-dependent resource" association between network resource entities, clearly presenting the association logic of core entities in the scheme to be detected.

[0039] The contextual constraint features corresponding to the scheme to be detected are feature representations extracted from the knowledge graph, representing constraint information such as timing requirements, parameter limitations, and resource usage conditions related to the scheme to be detected. These contextual constraint features provide a basis for conflict detection rules. Examples include the timing dependency requirement between handling operations such as "performing data backup first, then restarting the port"; the "bandwidth limit" parameter limitation corresponding to bandwidth adjustment operations; and the resource usage condition of "not reserved by other schemes."

[0040] In some embodiments, firstly, the core entity corresponding to the parsed solution to be detected is located in the updated knowledge graph. Then, using the association query capability of the knowledge graph, the direct and indirect association links of the core entity are traversed to extract the association patterns between entities as entity association features. Next, based on the entity attributes, relationship constraints, and other information pre-stored in the knowledge graph, constraints related to the execution of the solution to be detected are selected and integrated to form context constraint features.

[0041] S103. Integrate entity association features, context constraint features, and entity association link information in the updated knowledge graph to obtain input features.

[0042] The entity association link information in the updated knowledge graph refers to the complete multi-level association paths and logical transmission relationships formed between entities related to the solution to be detected in the updated knowledge graph. The entity association link information is a complete logical path without information compression, which can connect multiple types of entities and constraints, restore indirect associations and hierarchical dependencies between entities, and supplement implicit association information not covered by entity association features and context constraint features.

[0043] For example, if the solution to be detected is "port congestion fault - execute port restart operation", the corresponding entity association link information can fully present the multi-dimensional association logic: there is a direct correspondence between port congestion fault and port restart handling operation; the port restart operation relies on the core switch network resource, and the core switch relies on server resources for operational support. Simultaneously, the port restart operation has clear timing dependency requirements; a data backup operation must be performed first, and this data backup operation must meet the configuration parameter constraints of the backup duration threshold. This full-link logic can be represented as: "Port congestion fault - port restart handling operation - core switch (network resource) - server (dependent resource) - data backup handling operation (timing dependency, executed first) - backup duration threshold (configuration parameter requirement)". This link connects multiple entities such as faults, operations, resources, and parameters, and integrates the "correspondence between faults and operations", the "association between resources and dependent resources", the "constraint relationship between operations and timing dependencies", and the "relationship between operations and configuration parameter requirements". It fully restores the multi-dimensional association logic of the port restart operation from fault association to resource dependency, and from timing constraints to parameter limitations.

[0044] In some embodiments, the local association identifier data corresponding to the entity association feature, the rule-based condition data corresponding to the context constraint feature, and the global complete logical path data corresponding to the entity association link information are standardized and then integrated through feature dimension alignment and information complementarity to form a unified input feature that combines local key associations, clear constraint boundaries, and global logical links.

[0045] S104. Input the input features into the detection model based on the hybrid expert model MOE framework to obtain the conflict detection results.

[0046] The detection model includes a sparse routing network and multiple specialized expert models. The sparse routing network is used to match the specialized expert models with the input features, and the specialized expert models are used to perform specialized conflict detection based on the input features.

[0047] like Figure 2 The diagram illustrates an example of a hybrid expert model (MOE) framework. The detection models based on this MOE framework include sparse routing networks and specialized expert models 1, 2, and 3. Each specialized expert model is designed based on the differentiated detection requirements of core conflict types in network fault handling scenarios, and uses a knowledge graph as a unified data foundation. It can call upon entity association data (including entity attributes, direct relationships, multi-level association links, etc.) from the knowledge graph as the reasoning basis for specialized detection.

[0048] After the input features are fed into the detection model, they are sent to the sparse routing network. This network calculates the correlation between the input features and each specialized expert model (e.g., converting it to a matching score in the 0-1 range using the softmax function) to accurately identify the specialized expert model best suited to the conflict type corresponding to the input features. Then, the sparse routing network activates only one or more specialized expert models with the highest matching scores, directing the input features to the corresponding expert model. The specialized expert model then combines the local association identifiers, constraint boundaries, and global logical link information contained in the input features, and calls upon entity association data from the knowledge graph to perform targeted specialized conflict detection, outputting the corresponding specialized detection result.

[0049] The full-process conflict detection mechanism formed by S101~S104 above addresses various potential conflicts in network fault handling schemes, such as resource preemption, configuration parameters, and handling sequence. It effectively solves the core pain points of existing technologies, such as lack of context, low detection efficiency, and insufficient accuracy and generalization ability, and provides an end-to-end reliable solution for conflict detection in network fault handling schemes.

[0050] The conflict detection method provided in the embodiments of this application will now be described in detail.

[0051] This application also provides an embodiment for constructing a knowledge graph to achieve structured modeling and dynamic updating of core entities and their relationships, such as faults, handling operations, network resources, and configuration parameters, providing full contextual support for subsequent conflict detection. Figure 3 The diagram shows an example of a knowledge graph construction process. This embodiment constructs a knowledge graph through the following steps S201 to S205.

[0052] S201. Obtain multi-source heterogeneous data.

[0053] The multi-source heterogeneous data refers to the full-volume associated data related to network fault handling, covering entity information, attribute information, and relationship information required for the entire process of fault occurrence, handling execution, and resource support.

[0054] For example, taking the conflict detection scenario in a fault handling scheme in the communication field as an example, the multi-source heterogeneous data includes fault entity data (such as the name, level, and scope of impact of faults like port congestion and link interruption), handling operation data (such as commands and constraints for operations like port restart and bandwidth adjustment), network resource data (such as hardware parameters and associated resources of core switches and servers), configuration parameter data (such as bandwidth limits and device support thresholds), and time-series dependency data (such as operation execution order and dependencies). The data sources for this multi-source heterogeneous data include network management platform database data, historical fault logs, and device configuration manuals. The network management platform database data contains structured fault records, resource occupancy status, and operation execution trajectories. Historical fault logs cover semi-structured / unstructured fault occurrence time, handling process, and result feedback records. Device configuration manuals contain technical document data such as device hardware parameters, supported operation commands, and resource dependencies. Multi-source data collection ensures the comprehensiveness and relevance of the knowledge graph construction.

[0055] In some embodiments, a standardized collection mechanism is used to gather and initially organize fault handling-related data from a preset data source, providing initial support for subsequent knowledge graph construction. For example, for structured data in the network management platform database, queries and batch exports are performed through database interfaces such as JDBC and API calls to directly obtain well-organized fault records, resource usage status, operation execution trajectories, and other data.

[0056] For semi-structured key-value pairs and unstructured natural language descriptions in historical fault logs, log files from distributed storage are captured in real time or offline using log collection tools such as Fluentd and Logstash.

[0057] For technical documents such as equipment configuration manuals, document parsing tools such as PDF parsing components and natural language processing crawlers are used to extract text content such as device hardware parameters, supported operation commands, and resource dependencies.

[0058] Next, the data obtained from multiple sources will be aggregated and preliminarily processed through format verification, duplicate data filtering, and missing value marking to ensure the integrity and usability of the collected data, laying the foundation for subsequent data cleaning and structure transformation.

[0059] S202. Extract entities and their corresponding attribute information from multi-source heterogeneous data.

[0060] In some embodiments, the extraction process includes data preprocessing, entity recognition and classification, attribute association extraction, and preliminary verification of entities and attributes.

[0061] In the data preprocessing step, the multi-source heterogeneous data collected by S201 is cleaned and optimized to remove redundant characters, correct syntax errors, and unify data format standards.

[0062] In the entity recognition and classification step, based on the core associated object types (fault, handling operation, network resources, configuration parameters, etc.) of the network fault handling scenario, entities are identified from semi / unstructured data and mapped from structured data. For example, for semi / unstructured data such as historical fault logs and device configuration manuals, natural language processing models are used to identify entities in the text and label their types. For instance, from the log "Execute backup link switch after link interruption," the fault entity "link interruption" and the handling operation entity "backup link switch" are identified. For structured data such as network management platform database tables, entity types are determined through field mapping; for example, the "fault name" field in the "fault record" table corresponds to the fault entity.

[0063] In the attribute association extraction step, attribute information is extracted based on data type differences. For example, for structured data, it is directly associated through a preset field-attribute mapping relationship. For instance, the "fault level" and "impact range" fields in the "fault record" table of the network management platform database are directly used as the corresponding attributes of the fault entity.

[0064] For example, for semi / unstructured data, semantic dependency analysis and attribute keyword matching (such as parameters, thresholds, constraints, etc.) can be used to mine associations. For instance, from the device configuration manual's statement "Port restart requires bandwidth ≥ 100Mbps", the "bandwidth constraint" attribute and its value for the handling operation entity "Port restart" can be extracted.

[0065] In the initial entity and attribute verification step, the extracted results are verified in conjunction with the business logic constraints of network fault handling, to complete obviously missing core attributes, correct logical errors in associations, and filter out redundant attributes that are irrelevant to fault handling.

[0066] S203. Based on the business logic and entity attribute association features of the fault handling scenario, determine the association relationship between each entity.

[0067] Among them, the business logic of fault handling scenarios refers to the objective operating rules, execution constraints and related logic followed by each core element in the entire process from fault occurrence to resolution. It is used to clarify the role positioning, interaction mode and compliance requirements of different entities in the fault handling process, and is the core basis for determining the relationship between entities.

[0068] For example, taking the conflict detection scenario in a fault handling solution in the field of communications as an example, the business logic includes, but is not limited to, the following a1~a6: a1. Faults are resolved through specific handling operations. This business logic is used to determine the specific operations that need to be performed for a certain type of fault, providing a basis for resource preemption and timing conflict detection.

[0069] a2. The processing operation relies on the corresponding network resources. This business logic is used to locate the resources that the operation depends on, determine whether the resources are occupied or reserved, and detect resource preemption conflicts.

[0070] a3. There are hierarchical dependencies between network resources (e.g., core equipment depends on basic servers to run). This business logic is used to identify indirect dependencies between resources and avoid operation failures due to resource conflicts.

[0071] a4. Multiple processing operations must be executed in a preset order to avoid logical conflicts. This business logic is used to construct the operation sequence chain, detect whether there are timing loops or sequence errors, and ensure that the operations are executed logically.

[0072] a5. The fault affects the operational status of associated network resources. This business logic is used to clarify the scope of the fault's impact on resources and to help determine whether conflict resolution measures will amplify the fault's impact.

[0073] a6. Handling operations must comply with the limitations of the equipment configuration parameters. This business logic is used to provide a basis for configuration parameter conflict detection, verifying whether the parameters are contradictory or exceed the equipment's supported range.

[0074] The association features between business logic and entity attributes refer to common matching clues extracted from the core attributes of the corresponding entities based on the aforementioned business logic, which reflect the potential relationships between entities. Among them: The associated features corresponding to business logic a1 are, for example, semantic adaptation features between fault entities and handling operation entities, representing attribute matching clues that characterize the applicability relationship between the two. For example, the correspondence between the fault type attribute and the operation's adaptation fault type attribute.

[0075] The associated features corresponding to business logic a2 are, for example, functional matching features between the processing operation entity and the network resource entity, representing attribute matching clues that characterize the adaptability of operation resource requirements and resource functions. For example, the fit between the resource requirement attributes of the operation and the functional attributes of the resource.

[0076] The associated features corresponding to business logic a3 are, for example, hierarchical dependency features between network resource entities, representing attribute matching clues that characterize the hierarchical affiliation or dependency relationship between resources. For example, the correspondence between the association identifier attribute of one resource and the identity identifier attribute of another resource.

[0077] The associated features corresponding to business logic a4 are, for example, temporal correlation features between processing operation entities, representing attribute matching clues that characterize the execution order constraints of operations. For example, the correspondence between the pre- / post-identification attributes of one operation and the identity identification attributes of another operation.

[0078] The associated features corresponding to business logic a5 are, for example, impact-related features between faulty entities and network resource entities, representing attribute matching clues that characterize the scope of the fault's effect on the resource. For instance, the correspondence between the fault's impact scope attribute and the resource's attribution attribute.

[0079] The associated features corresponding to business logic a6 are, for example, parameter constraint features between the processing operation entity and the configuration parameter entity, representing attribute matching clues that characterize the operation parameter requirements and parameter adaptability. For instance, the compatibility between the operation's parameter requirement attributes and the parameter's applicable scope attributes.

[0080] In some embodiments, firstly, for the entity combinations corresponding to each business logic, the associated features are compared to preliminarily determine potential associations (e.g., semantic adaptation feature matching preliminarily establishes the correspondence between faults and handling operations). Next, the preliminarily determined potential association results are verified in conjunction with the corresponding business logic to eliminate logical contradictions and invalid associations (e.g., verifying whether there is a circular dependency in time-series associations). Then, the specific types of the verified potential associations are clarified, forming an entity-relationship-entity association chain.

[0081] S204. Based on a predefined construction framework, entities, their corresponding attribute information, and the relationships between entities are converted into triplet data.

[0082] The predefined building framework is used to constrain entity types, attribute specifications, and relationship types. For example, a predefined building framework might be a schema framework. The core constraints of the predefined building framework are as follows: Entity type constraints limit the core entities included in the knowledge graph. For example, the core entities include fault entities, handling operation entities, network resource entities, configuration parameter entities, etc.

[0083] Attribute specifications define the core attributes of each type of core entity. For example, a fault entity might include attributes such as name, level, and scope of impact. A handling operation entity might include attributes such as command, resource requirements, and constraints. A network resource entity might include attributes such as type, hardware parameters, and associated resources. A configuration parameter entity might include attributes such as name, threshold range, and applicable operations.

[0084] Association type constraints regulate the valid associations between entities. For example, valid associations include, but are not limited to, the correspondence between faults and handling operations, the involvement of handling operations and network resources, the dependency between network resources, the temporal dependency (execution order) between handling operations, the impact of faults on network resources, and the requirement relationship between handling operations and configuration parameters.

[0085] Based on a predefined construction framework, the relationships between entities obtained in S203 are converted into a triplet data format of head entity-relationship-tail entity. Furthermore, attribute information is attached to the head entity and tail entity to obtain triplet data.

[0086] S205. Use a graph database to store triple data to obtain a knowledge graph.

[0087] The graph database can be any commonly used graph database product in this field (such as Neo4j, Nebula Graph, Arango DB, etc.) to adapt to the structured data format of triples, store complex relationships between entities, and support fast relational queries and traversal. This provides efficient data access support for subsequent conflict detection.

[0088] In some embodiments, the triplet data stored in the graph database is shown in the following table:

[0089] By structurally modeling core entities and their relationships, such as faults, handling operations, network resources, and configuration parameters, a dynamically updated knowledge graph is constructed. This provides full contextual support for subsequent conflict detection, effectively avoiding the information fragmentation drawbacks of traditional structured storage. It can also accurately capture implicit relationships such as indirect resource dependencies and cross-scheme parameter associations, forming standardized and traceable entity relationship links, significantly improving the comprehensiveness, accuracy, and efficiency of conflict detection.

[0090] This application also provides embodiments for updating knowledge graphs to achieve dynamic iteration and timeliness maintenance of graph data, ensuring that entities, attributes and relationships are synchronized with actual fault handling scenarios, and continuously improving the accuracy of conflict detection.

[0091] In some embodiments, when a new fault handling plan to be detected is received, or when multi-source heterogeneous data such as network management platform database, historical fault logs, and device configuration manuals are updated, the knowledge graph update process is triggered.

[0092] The following example of receiving a new fault handling solution to be detected illustrates the implementation process of knowledge graph updating. The core logic and operation process of this embodiment are also applicable to knowledge graph updating scenarios when multi-source heterogeneous data is updated, and the implementation process will not be described in detail.

[0093] The knowledge graph update process includes two steps: determining the set of structured information corresponding to the parsing results and performing the update.

[0094] Step 1: Determine the set of structured information corresponding to the parsing results.

[0095] For example, natural language processing (NLP) techniques are used to perform structured parsing of new fault handling solutions to be detected. For instance, at the grammatical level, core elements such as the operator, action, object, time, and parameters are identified. At the semantic level, based on predefined entity types (faults, handling operations, network resources, configuration parameters) and constraints, unstructured / semi-structured solution text is mapped into a knowledge graph-compatible standardized format, forming a structured information set.

[0096] The structured information set includes the target entity, the target entity's target relationships with other entities, and the target entity's target attribute information.

[0097] The target entity is a subset of entities, referring to entities that need to be added or modified in the knowledge graph (such as new faults, new handling operations, etc.) parsed from the new fault handling solutions to be detected. It is the core object of this graph update.

[0098] Target association refers to the association between a target entity and existing entities or other target entities in the knowledge graph (such as the "corresponding handling operation" relationship between a new type of fault and its corresponding handling operation), and its type conforms to the predefined association type constraints.

[0099] Target attribute information refers to attribute data that characterizes the core features of a target entity (such as the level of the fault, the resource requirements for handling operations, etc.), and follows the same standard as the attribute information of existing entities in the knowledge graph.

[0100] Step 2: Perform the update.

[0101] The execution of the update includes performing targeted operations in the following b1~b3 cases: b1. If the target entity does not exist in the knowledge graph, create the target entity and update the knowledge graph based on the target attribute information and target association.

[0102] For example, if the new solution involves a "new type of link congestion fault" (target entity) that is not recorded in the knowledge graph, then the fault entity is created, target attribute information such as "level = high" and "affected scope = backbone network" is entered, and its "corresponding handling operation" relationship with "link expansion operation" is established to complete the knowledge graph supplementation.

[0103] b2. If the target entity already exists in the knowledge graph, but the target relationship is missing, then update the knowledge graph based on the target relationship.

[0104] For example, the knowledge graph already contains "device offline failure" (target entity), but it is not associated with "remote restart operation" in the new solution. In this case, a "corresponding handling operation" relationship is added between the two to complete the graph logic.

[0105] b3. If the target entity already exists in the knowledge graph, and the target attribute information is inconsistent with the attribute information stored in the knowledge graph, then update the knowledge graph based on the target attribute information.

[0106] For example, the original attribute of the "core switch" (target entity) in the knowledge graph is "bandwidth limit = 10G". In the new scheme, its attribute is updated to "bandwidth limit = 20G". At this time, the new attribute value overwrites the original information to ensure data accuracy.

[0107] In some embodiments, after the update is completed, the rationality of the update result is verified by combining the predefined schema framework constraints and the business logic of a1~a6, to check for problems such as entity identifier conflicts, attribute value contradictions, and relational logic conflicts, and to update the standardized feature vector adapted to the MOE model to ensure the accuracy of subsequent conflict detection.

[0108] By constructing a dynamically updated knowledge graph, core entities and their relationships such as faults, handling operations, and network resources are structured and modeled, providing full contextual support, effectively avoiding the information fragmentation problem of traditional structured storage, accurately capturing implicit relationships, and laying a reliable data foundation for conflict detection.

[0109] This application also provides embodiments for extracting entity association features and context constraint features corresponding to the scheme to be detected using an updated knowledge graph. For example... Figure 4 The diagram shows an example of a feature extraction process, which includes the following steps S301 to S304.

[0110] S301. Based on the target entity in the structured information set, determine the triplet association data from the updated knowledge graph.

[0111] Specifically, the triplet association data consists of structured data related to the solution to be detected within the updated knowledge graph. In other words, the triplet association data is the subset of triplet data included in the knowledge graph that is relevant to the solution to be detected.

[0112] In some embodiments, firstly, candidate triples whose head or tail entity types match the target entity type (fault, handling operation, network resource, configuration parameter) are selected from the full set of triple data in the knowledge graph. Then, based on predefined association relationships (such as the corresponding handling operation, involved resources, dependent resources, temporal dependencies, affected resources, and configuration parameter requirements defined above), the data corresponding to the association relationships in the candidate triples that are directly related to the core logic of the solution to be tested are retained, and irrelevant association data is filtered out to obtain the triple association data.

[0113] S302. Based on the link structure and association strength of triplet association data, the neighbor vectors of each entity in the entity set are aggregated through graph convolutional networks to obtain entity association features.

[0114] The entity set includes the target entity and its neighboring entities that are associated with the target entity through triples.

[0115] The link result of triple association data refers to the association path form formed by the target entity and its neighboring entities through triple data based on the predefined association relationship of the knowledge graph.

[0116] Association strength refers to the closeness and importance of the relationship between a target entity and its neighboring entities. In some embodiments, a basic weight is first calculated based on the predefined relation type weights of the knowledge graph (e.g., direct associations such as "corresponding handling operations" and "configuration parameter requirements" have higher weights than indirect associations such as "network resource-dependent resource"). Then, a weighted adjustment is made by combining entity attribute matching (e.g., the degree of fit of core attributes such as entity level, scope of influence, and hardware parameters) to obtain the association strength.

[0117] In some embodiments, the attribute information of each entity (target entity and neighboring entities) in the entity set is first converted into an initial feature vector. Then, an entity association adjacency matrix is ​​constructed based on the link structure of the triple association data.

[0118] Next, the initial feature vectors and the adjacency matrix are input into the graph convolutional network. The network layers of the graph convolutional network slide through the adjacency matrix according to the preset convolutional kernel. Within each convolutional window, the initial feature vectors of all neighboring entities are weighted and summed using the association strength as the weight. Among them, the neighbor vectors with higher association strength have a larger weight in the summation process, and their contribution to the target entity features is higher.

[0119] Subsequently, the weighted summation result is non-linearly transformed using an activation function (such as softmax) to output the target entity feature vector that incorporates neighbor entity association information, i.e., entity association features.

[0120] The above fusion process combines the relationship logic and importance differences between entities, so that the final features accurately reflect the core relationships and dependency levels between entities, providing more targeted feature support for conflict detection in subsequent MOE models.

[0121] S303. Based on the target entities and target relationships in the structured information set, extract the constraint information of the scheme to be detected from the updated knowledge graph.

[0122] Among them, constraint information is the original data basis of context constraint features, and context constraint features are the standardized feature forms of constraint information after numerical encoding.

[0123] In some embodiments, the extraction process is triggered in a targeted manner based on predefined associations in the knowledge graph.

[0124] For example, for the fault-corresponding handling operation association (such as serial number 1 in the table above), the matching requirements of fault and handling operation in the solution to be tested are extracted. Port congestion fault requires port restart operation, which is a typical scenario, and operation matching constraints are formed.

[0125] For disposal operations involving resource relationships (such as item 2 in the table above), extract the types and quantities of resources directly occupied by the disposal operation. Link switching operations require the use of backup links, which is a typical scenario, forming direct resource constraints.

[0126] Based on the network resource-dependent resource relationship (as shown in item 3 in the table above), extract the indirect dependent resource requirements required for the normal operation of network resources. The core switch's dependence on the server is a typical scenario, forming indirect resource constraints.

[0127] For the processing operation-time sequence dependency relationship (such as serial number 4 in the table above), the execution order requirements of the processing operations in the scheme to be tested are extracted. The data backup operation must precede the port restart operation, which is a typical scenario, thus forming a time sequence constraint.

[0128] For the fault-affected resource correlation (such as serial number 5 in the table above), extract the resource range requirements that will be affected after the fault occurs. Device offline faults will affect routers, which is a typical scenario, thus forming resource impact constraints.

[0129] Based on the relationship between handling operations and configuration parameter requirements (as shown in item 6 in the table above), extract the parameter thresholds and rules that the handling operations must follow. Bandwidth adjustment operations must meet the bandwidth upper limit requirements, which is a typical scenario, thus forming configuration parameter constraints.

[0130] S304. Based on the rule attributes and numerical boundaries of the constraint information, the constraint information is numerically encoded to obtain context constraint features.

[0131] In some embodiments, the coding process is designed to adapt to the core features of various constraints.

[0132] For example, for operation matching constraints, the matching relationship between fault types and handling operations is mapped to a binary feature vector. For instance, a matching state is encoded as 1, and a non-matching state is encoded as 0, forming an operation matching feature vector.

[0133] To address direct resource constraints, resource types are first mapped to basic values ​​according to their importance level, and then a multi-dimensional resource feature vector is constructed by combining the resource usage quantity and usage duration.

[0134] For indirect resource constraints, the resource dependency level is used as the weight, and the type and status information of the dependent resources are encoded into a dependency feature vector.

[0135] For timing constraints, the execution order of operations is mapped to ordered integer pairs, with the first operation encoded as 1 and the second operation encoded as 2, and then combined with the execution interval time to quantize into a timing feature vector.

[0136] To address resource impact constraints, the impact of faults on resources is categorized into levels and mapped to corresponding values. An impact feature vector is then constructed by combining the quantity and type of affected resources.

[0137] For configuration parameter constraints, numerical information such as parameter threshold range and default value is directly converted into a standardized vector.

[0138] By accurately extracting triplet association data and constraint information from knowledge graphs, and combining the association feature aggregation and adaptive numerical encoding of graph convolutional networks, we can achieve accurate representation of entity association logic, dependency hierarchy and constraint rules, provide high-quality input features for MOE models, and thus improve the efficiency, accuracy and generalization ability of conflict detection, effectively avoiding the problem of missed detection of implicit conflicts.

[0139] This application also provides an embodiment for obtaining input features by fusing the entity association features, the context constraint features, and the entity association link information in the updated knowledge graph. For example... Figure 5 The diagram shown is an example of a feature fusion process. This embodiment includes the following steps S401 to S404: S401. Standardize the entity association features and context constraint features to obtain the entity association feature vector and context constraint feature vector.

[0140] For example, standardization processing may include the Z-score standardization method. First, the mean and standard deviation of the entity association features and context constraint features are calculated. Then, the values ​​of the two types of features are mapped to the same scale interval using the formula (feature value - mean) / standard deviation, eliminating the dimensional differences between different feature dimensions. Subsequently, the standardized features are dimensionally aligned to ensure that the entity association feature vector and the context constraint feature vector have the same number of dimensions, meeting the format requirements for subsequent feature fusion.

[0141] S402. Construct an entity relationship adjacency matrix based on entity association link information.

[0142] The entity association link information includes the direct association path, indirect dependency path, and link level between the target entity and its neighboring entities. A direct association path refers to the association path established directly between the target entity and its neighboring entities through a single triple, corresponding to the direct mapping of head entity-relationship-tail entity in the knowledge graph. For example, a disposal operation involves resources that are then linked to network resources.

[0143] Indirect dependency paths refer to the association paths established between a target entity and other entities through at least one intermediate entity, corresponding to the links formed by connecting multi-level triples in a knowledge graph. For example, a disposal operation involves resources, a core switch, dependent resources, and a server.

[0144] Link level refers to the number of hops from the target entity to other entities in the associated path. The level of a direct associated path is 1, and the level of an indirect dependent path increases with the number of intermediate entities.

[0145] The entity relationship adjacency matrix is ​​structured data that quantifies various relationship paths and hierarchical relationships between a target entity and its related entities in matrix form, used to fully present the global relationship logic between entities. Compared to the adjacency matrix in step S302, the core difference lies in the fact that the adjacency matrix in S302 is based on the link structure of triplet relationship data, focusing on whether there is a direct relationship between entities, with matrix elements weighted by the strength of the relationship. In contrast, the entity relationship adjacency matrix covers all links of direct relationships and indirect dependencies, and incorporates link hierarchy weights, enabling a more comprehensive representation of the deep relationship logic between entities. The entity relationship link information includes the direct relationship path, indirect dependency path, and link hierarchy between the target entity and its neighboring entities.

[0146] For example, first, the target entity and all associated entities (including directly / indirectly associated entities) are identified and arranged in a unique identifier order as rows and columns of a matrix. The matrix element values ​​consist of "association existence identifier + link hierarchy weight".

[0147] Next, if there is a direct path between two entities, the base value is set to 1; if there is an indirect path, the base value is set to 0.5 (indirect associations have lower weights than direct associations); and if there is no association, it is set to 0. The base value is then multiplied by the reciprocal of the link level (higher levels have lower weights to avoid interference from long-distance associations) to obtain the final matrix element values, forming a complete entity relationship adjacency matrix.

[0148] S403. Concatenate the entity association feature vector, the context constraint feature vector, and the flattened vector of the entity relationship adjacency matrix to obtain the initial fusion vector.

[0149] The flattened vector of the entity relation adjacency matrix refers to converting a two-dimensional adjacency matrix into a one-dimensional continuous vector in row-major or column-major order, thereby realizing the serialized representation of matrix data.

[0150] In some embodiments, first obtain entity association feature vectors (e.g., dimension d1) and context constraint feature vectors (e.g., dimension d1) with consistent dimensions. Then, flatten the entity relationship adjacency matrix (e.g., size n×n) into a matrix of dimension n. 2 A one-dimensional vector.

[0151] Subsequently, following the order of entity association feature vector - context constraint feature vector - adjacency matrix flattened vector, the three types of vectors are concatenated end to end to form a vector with dimension d1+d1+n. 2 The one-dimensional vector, i.e. the initial fusion vector.

[0152] S404. Perform regularization on the initial fusion vector to obtain the input features.

[0153] In some embodiments, L2 regularization is used for regularization. For example, firstly, the L2 norm (the square root of the sum of the squares of the elements in the vector) of the initial fused vector is calculated, and then each element in the vector is divided by this L2 norm to make the norm of the processed vector equal to 1. Next, redundant elements with values ​​close to zero in the vector are filtered out to further optimize the vector sparsity and obtain the input features.

[0154] Through the collaborative processing of standardized alignment, full-link adjacency matrix construction, multi-dimensional feature splicing, and regularization optimization, the deep integration of entity association, constraint rules, and global link information is achieved, generating high-quality input features adapted to the MOE model, and providing comprehensive and standardized data support for accurate conflict detection.

[0155] This application also provides an embodiment that takes input features as input, inputs the input features into a detection model based on the Hybrid Expert Model (MOE) framework, and obtains conflict detection results. For example... Figure 6 The diagram shows an example of a conflict detection process, which includes the following steps S501 to S503.

[0156] S501. The correlation between the input features and the conflict type adaptation features of each specialized expert model is calculated through a sparse routing network to determine the target conflict type corresponding to the input features.

[0157] In some embodiments, a conflict type adaptation feature template is pre-configured for each specialized expert model. For example, the first expert model has a feature template for resource preemption conflicts, including key dimensions such as resource type, occupation duration, and dependency level. The second expert model has a feature template for configuration parameter conflicts, including dimensions such as parameter thresholds and device limitations. The third expert model has a feature template for handling timing conflicts, including dimensions such as execution order and interval time.

[0158] In some embodiments, the sparse routing network dimensionally aligns the input features with each template, calculates the correlation score (range 0-1) using a cosine similarity algorithm, and selects the adaptive feature template with the highest correlation score exceeding a preset threshold (e.g., 0.6). The corresponding conflict type is the target conflict type of the input feature. The preset threshold can be determined based on business fault tolerance requirements, historical conflict data distribution, and model detection accuracy targets. For example, a lower threshold (0.4-0.5) is used in high-reliability scenarios (e.g., power grids) to reduce missed detections, while a higher threshold (0.6-0.7) is used in efficiency-priority scenarios to reduce invalid computations.

[0159] In some embodiments, if all relevance scores fail to reach the threshold, the target conflict type is determined to be a composite conflict, which covers a single conflict type corresponding to multiple adaptive feature templates.

[0160] S502. Based on the target conflict type, activate the specialized expert model in the MOE framework that matches the target conflict type.

[0161] In some embodiments, if the target conflict type is a resource preemption conflict, the first expert model is activated. If it is a configuration parameter conflict, the second expert model is activated. If it is a timing conflict, the third expert model is activated. If it is a complex conflict, multiple specialized expert models with the highest suitability ranking are activated.

[0162] During the activation process, the model scheduling module allocates dedicated computing resources (such as computing power quota and memory space) to the target expert model and loads entity association link data strongly related to this conflict type from the knowledge graph. Furthermore, unmatched expert models are kept in a dormant state to avoid unnecessary resource consumption and ensure detection efficiency.

[0163] S503. Using the input features as input, perform conflict detection through a specialized expert model to obtain the conflict detection results.

[0164] In some embodiments, each specialized expert model uses the updated knowledge graph as a unified data foundation, combining entity associations, constraint rules, and global link information contained in the input features to execute specialized detection logic for the conflict type it is responsible for. For example, by leveraging the knowledge graph's association query and link traversal capabilities, it obtains core data such as entity attributes, associations, and constraints required for conflict detection. Then, combined with preset detection rules and inference algorithms, it accurately identifies the corresponding type of conflict and outputs conflict detection results including whether the conflict exists, core conflict information, and root cause link information.

[0165] This application also provides an embodiment that uses input features as input, performs conflict detection through a specialized expert model, and obtains conflict detection results. This embodiment includes: If the specialized expert model is the first expert model, then the resource occupation demand and resource occupation status are judged by the first expert model to determine whether there is a resource preemption conflict, and the first conflict detection result is output.

[0166] In some embodiments, firstly, the resource occupancy requirements (including resource type, quantity, and time period) of the scheme to be detected are extracted from the input features using a first expert model. Based on the knowledge graph's "disposal operation - involved resources" and "network resources - dependent resources" association relationships, the direct resources and indirect dependent resource links involved in the scheme to be detected are traversed.

[0167] Next, the resource occupancy status (occupied, idle, reserved) and time period information of other scheduled schemes stored in the knowledge graph are retrieved. By comparing whether the resource occupancy time periods of the scheme to be tested overlap with those of other schemes, and whether there are occupancy conflicts of indirectly dependent resources, it is determined whether there is a resource preemption conflict.

[0168] Subsequently, the first conflict detection result is output. The first detection result includes the conflict identifier (present / absent), the conflict resource identifier (direct resource + indirect dependent resource), the conflict period, and the conflict-related link. The conflict-related link is, for example, the scheme to be detected - involved resources - core switch - dependent resources - server - already occupied scheme.

[0169] If the specialized expert model is a second expert model, then the second expert model is used to judge the constraints of configuration parameters and equipment parameters to determine whether there are configuration parameter conflicts, and output the second conflict detection result.

[0170] In some embodiments, firstly, the configuration parameter information (parameter name, setting value, threshold range) in the input features is parsed by a second expert model. Based on the knowledge graph "disposal operation-configuration parameter requirements" association, the device configuration parameter constraints (such as the upper / lower limit of parameters supported by the device, default configuration value) corresponding to the disposal operation are retrieved.

[0171] Next, by using the "Disposal Operation - Involved Resources" association link, we query the configuration parameter settings of the same device or the same resource in other disposal schemes, verify whether the configuration parameters of the scheme under test exceed the device constraint range, whether there is a contradiction with the cross-scheme parameters, and determine whether there is a configuration parameter conflict.

[0172] Subsequently, the second conflict detection result is output. This result includes the conflict identifier (present / absent), conflict parameter items, parameter settings, compliance scope, associated device identifier, and conflict-related links. The conflict-related links, for example, are: the detected solution - configuration parameters - bandwidth limit - involved resources - core switch - other solution parameters.

[0173] If the specialized expert model is a third expert model, then the third expert model is used to judge the step sequence and timing constraint rules to determine whether there is a timing conflict in the handling, and output the third conflict detection result.

[0174] In some embodiments, firstly, based on the temporal feature vector in the input features, a third expert model is used to extract the step sequence and temporal constraint rules of the scheme to be detected (such as operation A must be executed before operation B) by combining the "disposal operation-temporal dependency" association relationship of the knowledge graph.

[0175] Next, an operation time-series dependency graph is constructed using a knowledge graph. The associated links in the graph are traversed, and a loop detection algorithm is used to detect whether there are time-series cycles (such as A must precede B, B must precede C, and C must precede A) or violations of preconditions (such as starting a port restart without performing a data backup operation, which violates the association rule of "data backup-time-series dependency-port restart"), and to determine whether there are any time-series conflicts to be handled.

[0176] Subsequently, the third conflict detection result is output. The third conflict detection result includes the conflict identifier (existence / non-existence), conflict steps, correct execution order, temporal dependency chain, and conflict root cause (such as "temporal loop", "violation of preceding operation constraints").

[0177] In some embodiments, the conflict detection result includes at least one of the first conflict detection result, the second conflict detection result, and the third conflict detection result.

[0178] If the target conflict type is a composite conflict and multiple specialized expert models are activated simultaneously, each expert model executes the above detection process and outputs the corresponding results. Subsequently, the multiple detection results are integrated and deduplicated, and the conflict type combinations (such as "resource preemption conflict + configuration parameter conflict") and cross-influence relationships (such as "resource preemption causes the tested solution to be unable to obtain the target resource, which in turn causes the configuration parameters to be unable to be set according to the constraints") are labeled, forming a complete conflict detection result that covers multiple types of conflicts and is logically coherent, ensuring that no composite conflicts are missed.

[0179] By leveraging the comprehensive entity association links and multi-dimensional constraint data provided by the knowledge graph, combined with the targeted detection logic of various specialized expert models, this system accurately identifies single-type conflicts such as resource preemption, configuration parameters, and handling sequence. Furthermore, it integrates and analyzes complex conflicts to cover cross-type related conflicts, achieving comprehensive and accurate conflict detection. In addition, the knowledge graph's association links allow for the traceability of conflict root causes, providing a clear basis for optimizing subsequent fault handling solutions.

[0180] After determining the conflict detection results, this application also provides an embodiment for displaying the conflict detection results, which transforms the abstract conflict detection results into intuitive and interpretable visual information, helping operation and maintenance personnel to quickly locate the root cause of the conflict, efficiently formulate optimization solutions, and shorten the iteration cycle of fault handling solutions. This embodiment includes: if the conflict detection result indicates the existence of a conflict and indicates the conflict type, then using a knowledge graph visualization method to annotate the entity association links corresponding to the conflict type to present the root cause of the conflict; based on the conflict type and the root cause, generating a detection report containing a description of the conflict type and visual root cause annotations.

[0181] In some embodiments, core conflict entities, relationships, and key conflict attributes (such as resource occupation periods, parameter settings, and temporal dependency rules) are first extracted from the conflict detection results to construct a core conflict dataset associated with the knowledge graph. Then, a hierarchical entity relationship graph is generated using a graph visualization tool. Conflict association links for resource preemption, parameter configuration, and temporal conflict handling are marked with red, orange, and blue colors plus differentiated line types (solid lines, dashed lines, and solid lines with arrows), respectively. Core attributes are labeled next to entity nodes, and dynamic flashing markers are added to conflict nodes. Interactive link drill-down is also supported to help operations and maintenance personnel intuitively trace the conflict propagation path and core root causes.

[0182] In some embodiments, before generating a detection report, the validity of the conflict detection results is first verified. If the detection results contain duplicate records or essentially identical conflict information caused by the same entity's associated links, deduplication processing is performed first. Deduplication processing compares the core entities and relationships of the conflict-related links, eliminates completely overlapping or highly similar redundant records, and merges conflict information with different descriptions but the same root cause, ensuring that the detection results are concise and accurate, and avoiding redundant content from interfering with the judgment of operation and maintenance personnel.

[0183] In some embodiments, after deduplication, conflicts are prioritized according to their impact on the network. For example, resource preemption conflicts are listed as the highest priority because they directly lead to operation failure and network service interruption. Configuration parameter conflicts are listed as medium priority because they may cause abnormal device operation. Conflicts related to handling timing are listed as low priority because they are mostly logical sequence issues with a relatively limited scope of impact. After prioritization, the conflicts are presented in the detection report, allowing maintenance personnel to focus on high-impact conflicts and improve problem-solving efficiency.

[0184] like Figure 7 The diagram illustrates an implementation example of a conflict detection method. Leveraging the full contextual support of a knowledge graph and the specialized capabilities of the MOE (Hybrid Expert Model) framework, it achieves end-to-end conflict detection from feature extraction to result output in fault handling solutions. This addresses the problems of missed implicit conflicts, excessive invalid computation, and low efficiency and accuracy in traditional detection methods, making it suitable for fault handling scenarios in fields such as communications, power grids, and the Internet of Things.

[0185] First, input the fault handling solution to be tested (such as a port congestion handling solution in the communications field or a line fault handling solution in the power grid field), and update the knowledge graph based on the solution parsing results. This knowledge graph contains three core entities: faults, operations, and resources, as well as relationships between entities such as "fault-corresponding handling operation", "operation-involved resources", and "resource-dependent resources". It can supplement new entities (such as new faults and new operations) and related logic in the solution in real time.

[0186] Next, based on the triple association data of the knowledge graph, the association information of the target entity and its neighboring entities is aggregated through a graph convolutional network to obtain entity association features; the temporal constraints, parameter restrictions and other constraint information in the knowledge graph are extracted and numerically encoded to obtain context constraint features.

[0187] Next, the entity association features and context constraint features are standardized and transformed into a unified dimension vector. Then, the entity association adjacency matrix is ​​constructed by combining the entity association link information (including multi-level dependency paths and link levels) from the knowledge graph. The two types of feature vectors are flattened and concatenated with the adjacency matrix vectors, and then regularization is performed to form a unified input feature adapted to the MOE framework.

[0188] Subsequently, the input features are fed into a sparse routing network, and the cosine similarity algorithm is used to calculate the correlation between the input features and the corresponding conflict type adaptation features of each specialized expert model (resource preemption detection, configuration parameter detection, and disposal timing detection). The adaptation features with the highest correlation and meeting the preset threshold are selected to determine the target conflict type.

[0189] Finally, based on the target conflict type, the matching specialized expert model in the MOE framework is activated. The model combines the input features with the knowledge graph entity association data to perform specialized detection and obtain the conflict detection results.

[0190] The above process, through the contextual support of knowledge graphs and the specialized division of labor of MOE models, not only solves the problem of traditional detection easily missing implicit conflicts, but also reduces invalid computation through sparse activation, thus balancing detection accuracy and efficiency.

[0191] Furthermore, this embodiment provides a collision detection system, such as... Figure 8 The diagram shows the structure of a conflict detection system, which includes: a knowledge graph module 810, a feature extraction module 820, a feature fusion module 830, and a conflict detection module 840; wherein: The knowledge graph module 810 is configured to update the knowledge graph based on the parsing results of the scheme to be detected.

[0192] The feature extraction module 820 is configured to use the updated knowledge graph to extract entity association features and context constraint features corresponding to the scheme to be detected.

[0193] The feature fusion module 830 is configured to fuse entity association features, context constraint features, and entity association link information in the updated knowledge graph to obtain input features.

[0194] The conflict detection module 840 is configured to input input features into a detection model based on the hybrid expert model MOE framework to obtain conflict detection results. The detection model includes a sparse routing network and multiple specialized expert models. The sparse routing network is used to match specialized expert models to the input features, and the specialized expert models are used to perform specialized conflict detection based on the input features.

[0195] In some embodiments, the knowledge graph module 810 is configured to determine the structured information set corresponding to the parsing result; the structured information set includes the target entity, the target association relationship between the target entity and other entities, and the target attribute information of the target entity.

[0196] If the target entity does not exist in the knowledge graph, the target entity is created, and the knowledge graph is updated based on the target attribute information and target relationships.

[0197] If the target entity already exists in the knowledge graph but the target relationship is missing, then the knowledge graph is updated based on the target relationship.

[0198] If the target entity already exists in the knowledge graph, and the target attribute information is inconsistent with the attribute information stored in the knowledge graph, then the knowledge graph is updated based on the target attribute information.

[0199] In some embodiments, the feature extraction module 820 is configured to determine triple association data from the updated knowledge graph based on the target entity in the structured information set; the triple association data is the structured data in the updated knowledge graph related to the scheme to be detected.

[0200] Based on the link structure and association strength of triple-linked data, the neighbor vectors of each entity in the entity set are aggregated through a graph convolutional network to obtain entity association features; wherein, the entity set includes the target entity and the neighbor entities associated with the target entity through triples.

[0201] Based on the target entities and their relationships in the structured information set, constraint information of the scheme to be detected is extracted from the updated knowledge graph.

[0202] Based on the rule attributes and numerical boundaries of constraint information, the constraint information is numerically encoded to obtain context constraint features.

[0203] In some embodiments, the feature fusion module 830 is configured to perform standardization processing on entity association features and context constraint features to obtain entity association feature vectors and context constraint feature vectors.

[0204] An entity relationship adjacency matrix is ​​constructed based on entity association link information; the entity association link information includes the direct association path, indirect dependency path and link level between the target entity and its neighboring entities.

[0205] The initial fusion vector is obtained by concatenating the entity association feature vector, the context constraint feature vector, and the flattened vector of the entity relationship adjacency matrix.

[0206] Regularization is performed on the initial fusion vector to obtain the input features.

[0207] In some embodiments, the conflict detection module 840 is configured to perform correlation calculations between the input features and the conflict type adaptation features of each specialized expert model through a sparse routing network to determine the target conflict type corresponding to the input features.

[0208] Based on the target conflict type, activate the specialized expert model in the MOE framework that matches the target conflict type.

[0209] Using the input features as input, conflict detection is performed through a specialized expert model to obtain the conflict detection results.

[0210] In some embodiments, the conflict detection module 840 is configured to, if the specialized expert model is the first expert model, judge the matching between resource occupancy demand and resource occupancy status through the first expert model, determine whether there is a resource preemption conflict, and output the first conflict detection result.

[0211] If the specialized expert model is a second expert model, then the consistency between the configuration parameters and the equipment parameter constraints is judged by the second expert model to determine whether there is a configuration parameter conflict, and the second conflict detection result is output.

[0212] If the specialized expert model is a third expert model, then the conformity between the step sequence and the timing constraint rules is judged by the third expert model to determine whether there is a timing conflict in the handling, and the third conflict detection result is output.

[0213] The conflict detection results include at least one of the first conflict detection results, the second conflict detection results, and the third conflict detection results.

[0214] In some embodiments, the apparatus further includes a display module, which is configured to, if the conflict detection result indicates that a conflict exists and indicates the type of conflict, mark the entity association links corresponding to the conflict type in a knowledge graph visualization manner to present the root cause of the conflict.

[0215] Based on the conflict type and root cause, a detection report is generated that includes a description of the conflict type and a visual annotation of the root cause.

[0216] In some embodiments, the knowledge graph module 810 is configured to acquire multi-source heterogeneous data.

[0217] Extract entities and their corresponding attribute information from multi-source heterogeneous data; Based on the business logic and entity attribute association characteristics of the fault handling scenario, the relationship between each entity is determined.

[0218] Based on a predefined construction framework, entities, their corresponding attribute information, and the relationships between entities are converted into triplet data; the predefined construction framework is used to constrain entity types, attribute specifications, and relationship types.

[0219] A knowledge graph is obtained by storing triple data using a graph database.

[0220] It should be noted that other corresponding descriptions of the functional units (modules or components) involved in the conflict detection system provided in this embodiment can be found in the description of the conflict detection method in the above embodiments, and will not be repeated here.

[0221] Based on the conflict detection method shown in the above embodiments, and Figure 8 To achieve the above objectives, the present application also provides an electronic device, as illustrated in the conflict detection system embodiment. This electronic device includes a storage medium and a processor; the storage medium stores a computer program; the processor executes the computer program to implement the method shown in the above embodiment.

[0222] Optionally, the aforementioned electronic device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optionally, a USB interface, a card reader interface, etc. Optionally, the network interface may include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0223] Based on the conflict detection method shown in the above embodiments, this embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method shown in the above embodiments.

[0224] Based on the methods shown in the above embodiments, this embodiment also provides a computer program product on which a computer program is stored, and when the computer program product is executed by a processor, it implements the methods shown in the above embodiments.

[0225] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.). The storage medium includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0226] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0227] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0228] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented in hardware. Compared with current related technologies, this application can establish deep relationships between entities based on knowledge graphs, providing contextual support for conflict detection and effectively reducing the missed detection of implicit conflicts. Furthermore, the detection model constructed by this application through the MOE model framework enables intelligent division of labor for conflict detection tasks, activating adapted specialized expert models on demand to perform detection, thereby reducing invalid computation and improving detection efficiency.

[0229] Because specialized expert models are designed for specific conflict types, they can focus on optimizing the core detection logic of the corresponding conflict. Knowledge graphs, on the other hand, contain entity association information and contextual constraint features, which can fully present the related relationships and constraints of conflicts. Furthermore, knowledge graphs support dynamic updates based on the parsing results of the solution to be detected, allowing for the real-time addition of new entities and relationships, and rapid adaptation to new faults and operational scenarios. Therefore, by combining specialized expert models with the dynamic updating capabilities and association constraint features of knowledge graphs for specialized detection, complex conflicts that are difficult to capture using traditional methods can be accurately identified, comprehensively addressing the shortcomings of existing conflict detection solutions in terms of overall performance and scenario adaptability.

[0230] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0231] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A collision detection method, characterized in that, include: The knowledge graph is updated based on the analysis results of the scheme to be detected; Using the updated knowledge graph, entity association features and context constraint features corresponding to the scheme to be detected are extracted; The input features are obtained by fusing the entity association features, the context constraint features, and the entity association link information in the updated knowledge graph. The input features are input into a detection model based on the hybrid expert model (MOE) framework to obtain conflict detection results; wherein, the detection model includes a sparse routing network and multiple specialized expert models, the sparse routing network is used to match specialized expert models to the input features, and the specialized expert models are used to perform specialized conflict detection based on the input features.

2. The method according to claim 1, characterized in that, The updating of the knowledge graph based on the parsing results of the scheme to be detected includes: Determine the structured information set corresponding to the parsing result; the structured information set includes the target entity, the target association relationship between the target entity and other entities, and the target attribute information of the target entity; If the target entity does not exist in the knowledge graph, the target entity is created, and the knowledge graph is updated based on the target attribute information and the target association. If the target entity already exists in the knowledge graph, but the target association is missing, then the knowledge graph is updated based on the target association. If the target entity already exists in the knowledge graph, and the target attribute information is inconsistent with the attribute information stored in the knowledge graph, then the knowledge graph is updated based on the target attribute information.

3. The method according to claim 2, characterized in that, The step of extracting entity association features and context constraint features corresponding to the scheme to be detected using the updated knowledge graph includes: Based on the target entity in the structured information set, triplet association data is determined from the updated knowledge graph; the triplet association data is the structured data in the updated knowledge graph related to the scheme to be detected. Based on the link structure and association strength of the triple association data, the neighbor vectors of each entity in the entity set are aggregated through a graph convolutional network to obtain the entity association features; wherein, the entity set includes the target entity and the neighbor entities associated with the target entity through triples; Based on the target entities and their relationships in the structured information set, the constraint information of the scheme to be detected is extracted from the updated knowledge graph; Based on the rule attributes and numerical boundaries of the constraint information, the constraint information is numerically encoded to obtain the context constraint features.

4. The method according to any one of claims 1-3, characterized in that, The input features are obtained by fusing the entity association features, the context constraint features, and the entity association link information in the updated knowledge graph, including: The entity association features and the context constraint features are standardized to obtain entity association feature vectors and context constraint feature vectors. An entity relationship adjacency matrix is ​​constructed based on the entity association link information; the entity association link information includes the direct association path, indirect dependency path, and link level between the target entity and its neighboring entities; The entity association feature vector, the context constraint feature vector, and the flattened vector of the entity relationship adjacency matrix are concatenated to obtain the initial fusion vector; The initial fusion vector is regularized to obtain the input features.

5. The method according to claim 4, characterized in that, The input features are fed into a detection model based on the Hybrid Expert Model (MOE) framework to obtain conflict detection results, including: The sparse routing network is used to calculate the correlation between the input features and the conflict type adaptation features of each specialized expert model to determine the target conflict type corresponding to the input features. Based on the target conflict type, activate the specialized expert model in the MOE framework that matches the target conflict type; Using the input features as input, conflict detection is performed through the specialized expert model to obtain the conflict detection result.

6. The method according to claim 5, characterized in that, The process of using the input features as input, performing conflict detection through the specialized expert model, and obtaining the conflict detection result includes: If the specialized expert model is the first expert model, then the resource occupation demand and resource occupation status are judged by the first expert model to determine whether there is a resource preemption conflict, and the first conflict detection result is output. If the specialized expert model is a second expert model, then the second expert model is used to judge the constraints of configuration parameters and equipment parameters to determine whether there is a configuration parameter conflict, and outputs the second conflict detection result; If the specialized expert model is a third expert model, then the third expert model is used to judge the step sequence and timing constraint rules to determine whether there is a timing conflict in the handling, and outputs the third conflict detection result; The conflict detection result includes at least one of the first conflict detection result, the second conflict detection result, and the third conflict detection result.

7. The method according to claim 6, characterized in that, The method further includes: If the conflict detection result indicates that a conflict exists and indicates the conflict type, the entity association links corresponding to the conflict type are labeled through the visualization method of the knowledge graph to present the root cause of the conflict. Based on the conflict type and the conflict root cause, a detection report is generated that includes a description of the conflict type and a visual annotation of the root cause.

8. The method according to any one of claims 1-3, characterized in that, The method further includes: Acquire multi-source heterogeneous data; Extract entities and their corresponding attribute information from the multi-source heterogeneous data; Based on the business logic and entity attribute association features of the fault handling scenario, the association relationship between the entities is determined. Based on a predefined construction framework, the entities, their corresponding attribute information, and the relationships between the entities are converted into triplet data; wherein, the predefined construction framework is used to constrain entity types, attribute specifications, and relationship types; The knowledge graph is obtained by storing the triple data in a graph database.

9. A collision detection system, characterized in that, The conflict detection system includes a knowledge graph module, a feature extraction module, a feature fusion module, and a conflict detection module; wherein: The knowledge graph module is configured to update the knowledge graph based on the parsing results of the scheme to be detected; The feature extraction module is configured to use the updated knowledge graph to extract entity association features and context constraint features corresponding to the scheme to be detected. The feature fusion module is configured to fuse the entity association features, the context constraint features, and the entity association link information in the updated knowledge graph to obtain input features; The conflict detection module is configured to input the input features into a detection model based on the hybrid expert model (MOE) framework to obtain conflict detection results; wherein, the detection model includes a sparse routing network and multiple specialized expert models, the sparse routing network is used to match specialized expert models to the input features, and the specialized expert models are used to perform specialized conflict detection based on the input features.

10. An electronic device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 8.

Images