Fault repair method and device based on semantic knowledge graph, equipment and medium

Abstract

The application relates to the technical field of data analysis, and discloses a fault repair method and device based on a semantic knowledge graph, equipment and a medium, which comprises the following steps: collecting multi-source heterogeneous data, extracting structured information through business semantic annotation, and constructing a business semantic knowledge graph; when a fault occurs, a double-channel reasoning engine combines the graph to analyze the root cause and generates a candidate repair scheme; the scheme is simulated and executed in digital twinning, risks are evaluated according to the influence of business continuity, and a safe scheme is screened; the scheme is executed according to the risk classification, actual results are fed back to the graph and the engine, and the decision is continuously optimized through reinforcement learning. The application can be applied to business scenarios such as financial technology, multi-source data collection and semantic enhanced knowledge graph construction are combined with double-channel reasoning and digital twinning pre-rehearsal, hierarchical repair and closed-loop feedback are realized, and the efficiency of autonomous fault repair and business continuity are improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a fault repair method, apparatus, equipment and medium based on semantic knowledge graph. Background Technology

[0002] As a core branch of the financial services industry, the insurance industry's IT systems carry critical business processes such as underwriting, claims settlement, risk control, and payment. They naturally possess three significant characteristics: extremely high business continuity requirements, complex system dependencies (including numerous third-party integrations), and stringent compliance audits. However, with the continuous evolution of system architecture towards microservices and cloud-native architectures, cross-service, cross-environment, and cross-team dependencies (covering both business and technical operations) are growing exponentially. Traditional threshold-based monitoring methods are no longer sufficient to capture hidden faults, posing unprecedented challenges to operations and maintenance.

[0003] Currently, troubleshooting in insurance operations relies heavily on human experience. Valuable operational knowledge is scattered across work order systems, manuals, and the minds of experts, leading not only to inefficient troubleshooting but also making high-risk operations such as database switching highly susceptible to triggering secondary failures. Furthermore, any disruption to insurance operations can quickly trigger a chain of risks, including customer complaints and regulatory penalties. Summary of the Invention

[0004] The main objective of this invention is to provide a fault repair method, apparatus, device, and storage medium based on semantic knowledge graphs, aiming to solve the technical problems in existing insurance operation and maintenance, such as the high dependence on human experience in fault repair, difficulty in root cause location, the ease with which high-risk operations can cause secondary faults, and the lack of automated repair capabilities based on business semantic awareness.

[0005] To achieve the above objectives, this invention provides a fault repair method based on semantic knowledge graphs, comprising: Collect multi-source heterogeneous data, and perform business semantic annotation on the multi-source heterogeneous data to extract structured information; Based on the structured information, a business semantic knowledge graph is constructed; When a system failure occurs, the dual-channel inference engine is invoked to perform root cause analysis of the failure in conjunction with the business semantic knowledge graph, and candidate repair solutions are generated based on the analysis results. The candidate repair scheme is simulated and executed in a digital twin pre-simulation environment to obtain pre-simulation evaluation results. The impact value of the candidate repair scheme is evaluated based on the business continuity quantitative index, and the target repair scheme is selected based on the impact value. Based on the pre-evaluation results, a tiered repair operation is performed on the target repair scheme, and the repair results are obtained and fed back to the business semantic knowledge graph and the dual-channel inference engine.

[0006] Furthermore, to achieve the above objectives, the present invention provides a fault repair device based on a semantic knowledge graph, comprising: The data acquisition module is used to collect multi-source heterogeneous data, and to perform business semantic annotation on the multi-source heterogeneous data to extract structured information; The graph construction module is used to construct a business semantic knowledge graph based on the structured information; The candidate repair module is used to call the dual-channel inference engine when the system fails, combine the business semantic knowledge graph to perform root cause analysis of the failure, and generate candidate repair solutions based on the analysis results. The target repair module is used to simulate the execution of the candidate repair scheme in the digital twin pre-simulation environment, obtain the pre-simulation evaluation results, evaluate the impact value of the candidate repair scheme according to the business continuity quantitative index, and select the target repair scheme according to the impact value. The repair execution module is used to perform hierarchical repair operations on the target repair scheme based on the pre-evaluation results, obtain the repair results, and feed them back to the business semantic knowledge graph and the dual-channel inference engine.

[0007] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a fault repair program based on a semantic knowledge graph stored in the memory and executable on the processor, wherein when the fault repair program based on the semantic knowledge graph is executed by the processor, it implements the steps of the fault repair method based on the semantic knowledge graph as described above.

[0008] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a fault repair program based on a semantic knowledge graph, wherein when the fault repair program based on the semantic knowledge graph is executed by a processor, it implements the steps of the fault repair method based on the semantic knowledge graph as described above.

[0009] Beneficial Effects: This invention relates to the field of data analysis technology and discloses a fault repair method, apparatus, device, and medium based on a semantic knowledge graph, comprising: collecting multi-source heterogeneous data and performing business semantic annotation on the multi-source heterogeneous data to extract structured information; constructing a business semantic knowledge graph based on the structured information; when a system fault occurs, calling a dual-channel inference engine to perform fault root cause analysis in conjunction with the business semantic knowledge graph, and generating candidate repair schemes based on the analysis results; simulating the execution of the candidate repair schemes in a digital twin pre-simulation environment to obtain pre-simulation evaluation results, and evaluating the impact value of the candidate repair schemes based on business continuity quantitative indicators, and selecting target repair schemes based on the impact values; performing hierarchical repair operations on the target repair schemes based on the pre-simulation evaluation results, obtaining repair results, and feeding them back to the business semantic knowledge graph and the dual-channel inference engine. This invention can be applied to business scenarios such as fintech. By constructing an insurance business semantic knowledge graph and a dual-channel inference engine, combined with digital twin pre-simulation to assess risks, perform hierarchical repairs, and continuously provide feedback and optimization, it helps operations and maintenance personnel accurately locate the root cause of faults and complete autonomous repairs safely and efficiently, significantly improving fault handling efficiency and system business continuity assurance capabilities. Attached Figure Description

[0010] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for a fault repair method based on semantic knowledge graphs according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the fault repair method based on semantic knowledge graph of the present invention. Figure 3 This is a flowchart illustrating step S100 in one embodiment of the fault repair method based on semantic knowledge graph of the present invention. Figure 4 This is a flowchart illustrating step S200 in one embodiment of the fault repair method based on semantic knowledge graph of the present invention. Figure 5 This is a flowchart illustrating step S300 in one embodiment of the fault repair method based on semantic knowledge graph of the present invention. Figure 6 This is a flowchart illustrating step S400 in one embodiment of the fault repair method based on semantic knowledge graph of the present invention. Figure 7 This is a flowchart illustrating step S500 in one embodiment of the fault repair method based on semantic knowledge graph of the present invention. Figure 8 This is a schematic diagram of the functional modules of a preferred embodiment of the fault repair device based on semantic knowledge graph of the present invention; Figure 9 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 10 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0011] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0012] The fault repair method based on semantic knowledge graph provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can collect multi-source heterogeneous data through the client, perform business semantic annotation on the multi-source heterogeneous data, and extract structured information; based on the structured information, it constructs a business semantic knowledge graph; when a system failure occurs, it calls a dual-channel inference engine, combines the business semantic knowledge graph to perform root cause analysis of the failure, and generates candidate repair solutions based on the analysis results; it simulates the execution of the candidate repair solutions in a digital twin pre-simulation environment, obtains pre-simulation evaluation results, evaluates the impact value of the candidate repair solutions based on business continuity quantitative indicators, and selects the target repair solution based on the impact value; based on the pre-simulation evaluation results, it performs tiered repair operations on the target repair solution, obtains the repair results, and feeds them back to the business semantic knowledge graph and the dual-channel inference engine. This invention can be applied to business scenarios such as fintech. By constructing an insurance business semantic knowledge graph and a dual-channel inference engine, combining digital twin pre-simulation risk assessment, tiered repair execution, and continuous feedback optimization, it helps operation and maintenance personnel accurately locate the root cause of failures and safely and efficiently complete autonomous repairs, significantly improving fault handling efficiency and system business continuity assurance capabilities. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.

[0013] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the fault repair method based on semantic knowledge graph provided by the present invention. It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0014] like Figure 2 As shown, the fault repair method based on semantic knowledge graph proposed in this invention includes the following steps: S100. Collect multi-source heterogeneous data, and perform business semantic annotation on the multi-source heterogeneous data to extract structured information; S200. Based on the structured information, construct a business semantic knowledge graph; S300. When a system failure occurs, the dual-channel inference engine is invoked to perform root cause analysis of the failure in conjunction with the business semantic knowledge graph, and candidate repair solutions are generated based on the analysis results. S400. Simulate the execution of the candidate repair scheme in the digital twin pre-simulation environment, obtain the pre-simulation evaluation results, evaluate the impact value of the candidate repair scheme according to the business continuity quantitative index, and select the target repair scheme according to the impact value. S500. Based on the pre-evaluation results, perform a graded repair operation on the target repair scheme, obtain the repair results, and feed them back to the business semantic knowledge graph and the dual-channel inference engine.

[0015] In this embodiment, the system first collects multi-source heterogeneous data and performs deep business semantic annotation. The collected data covers four main categories: transaction logs and API call logs reflecting business processes; host and microservice operation and maintenance indicators reflecting infrastructure health; dynamic events recording version changes and emergency responses; and business rules containing insurance terms and regulatory red lines. For the unstructured text in this data, the system introduces Natural Language Processing (NLP) technology, relying on a specialized dictionary for the insurance field, and employs a sequence annotation model combining Bidirectional Long Short-Term Memory (BiLSTM) networks and Conditional Random Fields (CRF) to accurately extract core business entities and their attribute information, such as policy numbers, insurance types, and coverage amounts, from the original logs and documents. Subsequently, a graph-based dependency parsing model is used to specifically handle complex long-distance dependencies and nested sentence structures in Chinese text, identifying the semantic links between business actions and associated objects, and generating standard triples in the form of "business action-object-constraint". Meanwhile, for system anomaly logs, the system uses a semantic classification model to label them with specific business impact tags. For example, "claims service calls risk control engine timed out" is labeled as "affects claims timeliness". Finally, all processing results are integrated into structured information that can be directly used by downstream models.

[0016] After acquiring structured information, a semantically enhanced knowledge graph specific to the insurance business domain is constructed. This graph is built using a hybrid "top-down" and "bottom-up" strategy: First, based on the insurance business architecture document, core entity types are predefined, encompassing business services, business objects, technical components, operational events, and business rules. The semantic relationship skeleton between entities, such as business dependencies, change impacts, historical root cause associations, and compliance constraints, is then clearly defined, forming an ontology model. Building upon this, the structured instance data generated in the previous steps is categorized according to entity type and populated into the established relationship skeleton, thus generating the initial graph structure data. To address the dynamic changes brought about by the rapid iteration of insurance business, the system introduces a dynamic evolution mechanism: when a new business line or entirely new entity is detected, the system first determines whether it belongs to an existing type. If not, a new entity type is automatically created, and a knowledge graph completion algorithm (such as TransE) is invoked to infer the potential relationships between the new entity and existing entities. This completes the incremental update of the knowledge graph. All graph data is ultimately stored in a graph database to support efficient relational queries and subsequent graph computations.

[0017] When a business failure or performance anomaly is detected, the built-in dual-channel inference engine is quickly triggered, closely integrating with the aforementioned knowledge graph to perform in-depth root cause analysis. In the symbolic inference channel, the system pre-configures an "IF-THEN" production rule base, derived from insurance business rules, compliance red lines, and historical failure cases. Rules are assigned differentiated priority weights based on business importance; for example, rules involving core links such as claims payments receive higher trigger priority, thus quickly matching known failure patterns and outputting the first type of candidate repair solutions. Simultaneously, the neural network inference channel is launched. This channel employs a graph neural network (GNN) model with an incorporating attention mechanism, whose input is the relational subgraph surrounding the failure node in the knowledge graph. This model not only integrates technical indicators such as CPU utilization into node embedding but also innovatively embeds business attributes such as policy type risk weights and claims processing timeliness requirements. It automatically identifies the neighboring nodes that contribute the most to the current anomaly through a multi-head attention mechanism, ultimately outputting the probability distribution of the anomaly's root cause. For latent fault modes with probability values ​​exceeding a preset threshold, the system uses their corresponding repair strategies as second-class candidate solutions, thus complementing the symbol channel and balancing the rapid response of known rules with the accurate discovery of unknown anomalies.

[0018] To fundamentally eliminate the risk of secondary failures caused by remediation operations, all candidate remediation solutions must undergo rigorous simulation and verification in a digital twin pre-operation environment before actual implementation. The system constructs a digital twin of the core insurance business based on lightweight containerization technologies (such as Docker and Kubernetes). This includes simplified images of core microservices, anonymized real business sample data, and simulators for simulating external dependencies such as reinsurance systems and payment gateways, enabling a high degree of real-world business state flow and data interaction. During the pre-operation, the system converts candidate remediation solutions into specific operational instructions and injects them into the twin environment. After simulation execution, the system quantitatively evaluates the solutions based on a preset Business Continuity Impact Score (BCIS). This impact score is a comprehensive function whose calculation factors include business interruption duration, the number of affected policies, and whether compliance violations are triggered. Compliance violations are checked in real-time by a built-in rule engine. If the calculated impact value exceeds a safety threshold, the solution is marked as high-risk and automatically triggers a manual review process; if the impact value is within an acceptable range, it is considered a safe and feasible target remediation solution. In addition, the execution trajectory and result data of the entire pre-training process will be recorded synchronously and fed back to the knowledge graph for subsequent model optimization training.

[0019] Ultimately, the system executes tiered repair operations based on the pre-simulation assessment results and establishes a comprehensive feedback and learning loop. The specific tiered strategy is as follows: for low-risk operations such as cleaning log files and refreshing caches, the system directly invokes the operations and maintenance robot for fully automated execution without manual intervention; for high-risk operations such as database master-slave switching and core service restarts, the system automatically generates a repair work order containing a detailed pre-simulation report, step-by-step operation instructions, and a standardized rollback plan, and pushes it to the operations and maintenance personnel's terminals. Execution is only permitted after manual confirmation, thus organically combining automated efficiency with manual safety safeguards. After the actual repair operation is completed, the system will collect the actual execution results (including whether it was successful or not, the time taken, and whether it caused other service anomalies or performance degradation, etc.) and compare and analyze them with historical cases stored in the knowledge graph. Then, it will use reinforcement learning (RL) algorithms to dynamically adjust the reward function parameters inside the dual-channel inference engine—for example, giving positive incentive scores to successful repairs and imposing strong negative penalties on operations that cause secondary failures—in order to continuously optimize the policy trigger threshold of the symbolic inference channel and the model weight of the neural network inference channel, driving the entire system to iterate and evolve in continuous operation and maintenance practices, and finally forming a complete closed-loop capability from data perception, knowledge accumulation, intelligent decision-making, security verification to autonomous execution.

[0020] This embodiment constructs an insurance business knowledge graph by collecting multi-source heterogeneous data and labeling it with business semantics. In case of a failure, a dual-channel inference engine is invoked, combining symbolic rules and neural networks to perform root cause analysis and generate repair solutions. After pre-simulating and assessing the impact on business continuity in a digital twin environment, target solutions are selected, repairs are executed in stages, and results are fed back to achieve closed-loop optimization. This significantly shortens fault recovery time, reduces the probability of customer complaints and compliance risks caused by interruptions, and significantly improves the autonomous repair capability and operational security of the core insurance business system.

[0021] The core application in the fintech business lies in building an intelligent operation and maintenance (O&M) system that integrates business and technology. Addressing the common challenges faced by financial institutions such as banks, securities firms, and insurance companies—including system complexity, stringent regulations, and high business continuity requirements—it deeply connects core business objects such as transactions, payments, and risk control with underlying technical components through a business semantic knowledge graph, enabling the O&M system to understand the boundaries of business impact. A dual-channel inference engine ensures rapid response to known compliance faults and timely discovery of unknown hidden risks. A digital twin pre-simulation mechanism simulates the impact and conducts compliance pre-checks for high-risk changes involving fund flows and customer data operations, preventing transaction interruptions or regulatory violations caused by configuration changes or fault switching. A tiered repair strategy ensures that low-frequency, high-impact core financial operations require pre-simulation and manual confirmation, while high-frequency, low-impact routine O&M tasks can be executed fully automatically. This guarantees transaction consistency, data security, and regulatory compliance in financial businesses, while continuously learning from feedback to gradually accumulate institution-specific O&M knowledge assets, reducing reliance on expert experience and providing a reliable O&M technology foundation for the digital transformation of the financial industry.

[0022] In one embodiment, S100 includes: S101. Collect business transaction data, operation and maintenance status data, change and event data, and business strategy data, which together constitute multi-source heterogeneous data. S102. Based on an insurance domain dictionary, a sequence labeling model based on bidirectional long short-term memory network and conditional random field is used to extract business entities and their attribute information from the unstructured text of the multi-source heterogeneous data. S103. Based on the business entities, process the long-distance dependencies and complex sentence structures in the unstructured text to identify the business actions and related objects between the business entities. S104. Generate a triplet structure based on the business actions and associated objects between the business entities; S105. Perform semantic classification on the abnormal logs and label the business impact tag corresponding to each abnormal log. S106. Generate structured information based on the business entity and its attribute information, the triplet structure, and the labeled business impact tags.

[0023] In this embodiment, as Figure 3 As shown, extensive data collection was conducted targeting the characteristics of the insurance business system, acquiring four core data sources in parallel: business transaction data reflecting transaction details and interface call status of application systems across various business lines; operational status data covering the availability of hosts, microservices, databases, and third-party systems; change and event data recording version releases, emergency plan execution, and compliance checks; and business strategy data containing strategic constraints such as insurance clause exception handling logic and regulatory red lines. These four data types encompass structured logs, semi-structured logs, and unstructured text, collectively forming a typical multi-source heterogeneous dataset, providing raw materials for subsequent semantic analysis.

[0024] Building upon this foundation, the system employs natural language processing (NLP) technology for deep analysis of unstructured text. On one hand, relying on lexical features and boundary cues provided by a professional insurance dictionary, a bidirectional long short-term memory (BiLSTM) network is used to encode the text sequence in both forward and backward directions, fully capturing the long-distance semantic dependencies of each word within its context. Simultaneously, a conditional random field (CRF) model is introduced into the sequence labeling layer, utilizing its transition probability matrix to constrain the global legality of the label sequence, avoiding unreasonable results such as "insurance-entity" label jumps. This allows for the accurate extraction of business entity information, including policy number, insurance type, sum insured, and policyholder, along with their associated attributes, from unstructured text such as original logs, policy descriptions, and operation records.

[0025] For the extracted business entities, the system further addresses the complex sentence structures and long-distance dependencies unique to the Chinese language. In Chinese sentences, modifying or constraining components are often far from the core predicate, and nested structures are frequent, making it difficult for traditional local window analysis methods to accurately capture semantic meaning. Therefore, the system employs a graph-based dependency parsing model, constructing a directed graph by treating the words of the entire sentence as nodes and the syntactic dependency relationships between words as edges. By iteratively calculating the dependency paths between nodes in the graph, it effectively traverses long-distance sentence structures such as "if...then..." and "for...as far as...", accurately identifying the specific business actions and associated objects between business entities. For example, it clarifies that the action of "refusing to pay" is applied to the "policyholder" and that the triggering condition is "failure to disclose truthfully."

[0026] After identifying business actions and related objects, the system reorganizes these semantic elements into standardized triple structures, uniformly expressed as "business action-object-constraint" or "entity-relationship-entity," transforming unstructured natural language knowledge into graph relational data that computers can directly process. Simultaneously, for abnormal logs generated during system operation, the system does not treat them as simple technical errors, but rather categorizes them by business dimensions using a pre-trained semantic classification model, labeling each abnormal log with a clear business impact tag. For example, "timeout when calling the risk control engine" is categorized as "affecting claims processing time," and "database connection pool exhaustion" is categorized as "potentially causing underwriting interruption," thus establishing a direct mapping between underlying technical anomalies and upper-level business impacts. Finally, the system integrates the extracted business entities and their attribute information, the triple structures representing semantic relationships, and the labeled business impact tags, uniformly encapsulating them into standardized structured information as precise input for subsequent knowledge graph construction.

[0027] This embodiment employs a dual-track strategy of "basic corpus accumulation + automatic new word discovery" to construct an insurance domain dictionary. First, the system collects raw corpora including insurance business terms and conditions, regulatory compliance documents, historical maintenance work order descriptions, and product manuals. Combined with a publicly available financial and insurance terminology database, and after manual verification by domain experts and regular expression matching, a basic dictionary is formed covering insurance product names, policy fields, business actions, and common error code aliases. Building upon this foundation, for new abbreviations, new service names, and special expressions in the context of maintenance and operation arising from rapid iterations in insurance business, the system introduces an unsupervised new word discovery algorithm based on mutual information and left / right entropy. This algorithm performs frequent pattern mining on unstructured text, calculates the internal cohesion and boundary freedom of candidate word strings, automatically filters high-confidence new words, and incrementally incorporates them into the dictionary after expert sampling and verification, thereby ensuring the dictionary's continuous and complete coverage of insurance business and maintenance terminology.

[0028] To label the business impact of system anomaly logs, this embodiment adopts a fine-tuning classification scheme based on a pre-trained language model. Specifically, a BERT model pre-trained on Chinese financial corpus is selected as the text encoding backbone, with a fully connected classification layer and a Softmax output layer stacked on top. Multiple categories of business impact labels are predefined, including "affecting claims processing timeliness," "affecting underwriting continuity," "affecting payment settlement," "affecting risk control judgment," and "no direct impact." During offline training, anomaly log samples annotated by experts are extracted from the historical operations database to form a training set. The model parameters are fine-tuned in a supervised manner, enabling the model to learn the ability to map technical log text to the probability distribution of business impact labels. During online inference, the model outputs the confidence score for each category in real time. The system selects the label with the highest confidence score exceeding a preset threshold (e.g., 0.75) as the semantic classification result for that log. For ambiguous samples with confidence scores below the threshold, they are automatically transferred for manual review and included in subsequent incremental training sets, driving continuous iterative optimization of the classification model.

[0029] This embodiment first collects insurance business transaction data, operation and maintenance status data, change and event data, and business strategy data to form a multi-source heterogeneous data foundation. Then, based on an insurance-specific dictionary, it employs a Bidirectional Long Short-Term Memory (BiLSTM) network combined with a Conditional Random Field (CRF) sequence labeling model to accurately extract business entities and their attribute information from unstructured text. Furthermore, it processes complex long-distance dependencies and nested sentence structures in Chinese text, identifies business actions and related objects between entities, and generates standard triple structures. Simultaneously, it uses semantic classification to label abnormal logs with business impact tags. This technology unifies the transformation of unstructured text, semi-structured logs, and structured indicators in the insurance field into standardized structured information, providing high-quality semantic data input for downstream knowledge graph construction.

[0030] In the fintech business sector, such as in core banking systems, the BiLSTM+CRF model extracts key entities like transaction serial numbers, amounts, and channels from failed interbank transfer logs. Combined with dependency parsing, it identifies causal relationships such as "timeout-payment gateway-retry," generating triples to quickly pinpoint the root cause and label it with business tags like "affects settlement timeliness." In the securities sector, it can be used to extract entity relationships such as shareholder changes and pledge ratios from listed company announcements, assisting risk control systems in timely identifying credit risk signals. This technology enables financial IT systems to understand unstructured data at the business semantic level, significantly improving fault location accuracy and business impact assessment efficiency, providing intelligent data support for financial business continuity assurance and compliance review.

[0031] In one embodiment, S200 includes: S201. Predefine an ontology model that includes core entity types and business semantic relationships; S202. Based on the core entity types and business semantic relationships in the ontology model, and in conjunction with the insurance business architecture document, determine the entity types and their corresponding relationship skeletons. S203. Automatically extract instance data from the structured information, classify the instance data according to entity type, and fill it into the relation skeleton to generate a business semantic knowledge graph. S204. When a new insurance business or a new entity is detected, determine whether the new insurance business or the new entity belongs to an existing entity type based on the core entity type in the ontology model. S205. If it belongs to an existing entity type, the new insurance business or new entity shall be classified into the relation skeleton corresponding to the existing entity type. S206. If it does not belong to an existing entity type, then create an entity type for the new insurance business or new entity and populate it into the corresponding relation skeleton. S207. Call the knowledge graph completion algorithm to infer the association between the new insurance business or new entity and the existing entity type, and update the business semantic knowledge graph according to the association.

[0032] In this embodiment, as Figure 4 As shown, the construction of the knowledge graph begins with the ontology design at the schema layer. Based on the inherent logic of the insurance business domain, core entity types (such as technical components and business objects like underwriting services, claims, and database clusters) and the business semantic relationships between these entities (such as service dependencies, change impacts, historical root cause associations, and compliance constraints) are clearly defined, thus forming a dedicated ontology model. Building upon this, and in conjunction with the insurance business architecture document, each entity type and its corresponding relational skeleton are specifically defined within this ontology framework, clarifying the structural blueprint and association paths of knowledge storage. This is the "top-down" abstract modeling process, providing precise semantic anchors for subsequent data instantiation. After completing the schema definition, the system automatically extracts specific instance data using the structured information obtained from the previous processing. These instances are categorized according to preset entity types and accurately populated into the corresponding relational skeletons, thereby completing the transformation from abstract schemas to concrete instances and generating an initial business semantic knowledge graph, laying the data foundation for subsequent relational queries and inference calculations.

[0033] To address the dynamic changes brought about by the rapid iteration of insurance business and the continuous launch of new product lines, the system incorporates a dynamic evolution mechanism for its knowledge graph. When a new insurance business or entirely new entity is detected (e.g., the addition of a "new energy vehicle insurance" product line and its associated assessment services), the system first determines whether the new entity belongs to an existing category based on the core entity types defined in the ontology model. If it belongs to an existing category (e.g., the service still belongs to the "risk control engine" category), it is directly assigned to the corresponding relational skeleton and instantiated. If the new entity does not belong to any existing category, an entity type is dynamically added at the schema layer for the new insurance business or entity, and a corresponding relational skeleton structure is established. Subsequently, the system calls a knowledge graph completion algorithm (such as TransE, or Translation Embedding Model, which maps entities and relations to a low-dimensional continuous vector space and uses vector translation operations to efficiently calculate semantic similarity and potential associations between entities) to predict links. It automatically infers the possible dependencies, influences, or constraints between the new entity and existing entity types, and performs incremental updates to the knowledge graph based on the inferred association results. This ensures that the entire knowledge system can remain synchronized with the evolution of insurance business, providing comprehensive and timely knowledge support for intelligent reasoning in new models.

[0034] This embodiment predefines an insurance ontology model and relational skeleton, automatically extracts instances from structured information to fill in and generate a knowledge graph. When new business or entities appear, it automatically determines their affiliation, adds new types, and calls a completion algorithm to infer associations to achieve dynamic updates. The result is the construction of an evolvable insurance business semantic knowledge graph, breaking down the disconnect between operations and business, and providing accurate and dynamic knowledge support for intelligent reasoning.

[0035] In the fintech sector, for example, when banks add "digital RMB wallet" services, the system automatically identifies the type of entity to which it belongs or adds the service, completing the connections with the "settlement system" and "anti-money laundering rules" to ensure the knowledge base is synchronized with the business. In the securities sector, new regulatory rules can be dynamically integrated, automatically inferring their impact on trading and risk control services. This mechanism ensures that the knowledge system is continuously updated with business innovation, supporting accurate root cause analysis and compliance reasoning, reducing manual maintenance costs, and enhancing the self-adaptive capabilities of the financial system.

[0036] In one embodiment, S300 includes: S301. When a system failure occurs, the symbolic reasoning channel of the dual-channel inference engine performs inference based on the preset strategy library and outputs the first candidate repair scheme corresponding to the known failure mode. S302. The neural network reasoning channel of the dual-channel reasoning engine uses a graph neural network model with an attention fusion mechanism for reasoning. The subgraph related to the fault node in the business semantic knowledge graph is used as the input of the graph neural network model, and the root cause probability distribution is output. S303. The repair strategy corresponding to the root cause whose root cause probability distribution exceeds the preset threshold is taken as the second candidate repair scheme. S304. A candidate repair scheme is formed by the first candidate repair scheme and the second candidate repair scheme.

[0037] In this embodiment, as Figure 5 As shown, when a fault or performance anomaly is detected, the dual-channel inference engine immediately starts, with its internal symbolic inference channel taking over first. The core of this channel relies on a pre-built strategy library, which was constructed offline by deeply abstracting and formally encoding insurance business rules, regulatory compliance red lines, and numerous historical fault cases. The rules use an "IF-THEN" production rule representation; for example, the operational experience of "if the claims service calls the payment gateway and the payment gateway is in normal status, then trigger the retry mechanism" is solidified into a computer-executable logical decision unit. To ensure that critical business operations receive priority resource protection, the system also introduces an insurance business priority weighting mechanism—rules involving core transaction links such as underwriting, claims, and payment are given higher triggering priority, effectively preventing edge faults from preempting inference resources. When real-time fault characteristics are input into this channel, the system quickly matches rules in the strategy library that meet the conditions through forward chain inference, directly outputting the first candidate repair solution corresponding to the known fault mode.

[0038] Simultaneously, the neural network inference channel is launched in parallel. This channel employs a graph neural network model with an incorporating attention mechanism, specifically designed to handle complex, latent faults lacking clear historical precedents. Its input data is not the entire knowledge graph, but rather a subgraph extracted from the faulty node within a multi-hop range. This subgraph contains all connected business services, technical components, historical events, and constraint rule nodes surrounding the faulty entity. The core innovation of this model lies in its enhanced node embedding strategy: each node's feature vector not only integrates conventional technical indicators such as CPU utilization and memory usage, but also innovatively embeds insurance business attributes—such as the policy's risk weight coefficient, the urgency of the current claims stage, and the real-time transaction volume carried by the service—giving the model a native ability to perceive "business impact." The model completes the transfer and aggregation of neighborhood information through stacked multi-layer graph convolution operations and introduces a multi-head attention mechanism to automatically calculate and focus on the neighboring nodes that contribute the most to the current anomaly (such as a third-party payment gateway with a long-standing interface fluctuation record), ultimately mapping the output layer to the probability distribution of each candidate root cause.

[0039] This probability distribution represents the quantified likelihood of different potential root causes triggering the current failure. The system sets a strict empirical probability threshold; only when the probability distribution of a certain root cause exceeds this threshold is the system identified as a high-confidence anomalous pattern, and its corresponding remediation strategy is included in the second set of candidate remediation solutions. Thus, the symbolic reasoning channel provides rapid response capabilities for deterministic scenarios through "rule matching," while the neural network channel uncovers unknown hidden dangers in complex relationships through "deep learning." The candidate solutions output by the two channels are combined to form a complete set of candidate remediation solutions covering both known and unknown aspects, balancing efficiency and depth, providing ample and diverse input options for subsequent digital twin pre-simulation verification.

[0040] In this embodiment, when a system failure occurs, the symbolic inference channel quickly matches known faults based on a pre-set strategy library and outputs a first solution; the neural network channel takes the fault node subgraph as input, uses a graph neural network with an attention fusion mechanism to output a root cause probability distribution, and those exceeding a threshold are used as a second solution, which are then combined to form candidate repair solutions. This approach balances rapid response to known rules with accurate detection of unknown anomalies, improving the coverage and accuracy of inference.

[0041] In the fintech field, the symbolic channel provides millisecond-level rule matching and handling suggestions for high-frequency known faults such as payment timeouts and interface timeouts; the neural network channel performs graph structure modeling for complex anomalies involving cross-services and cross-systems (such as clearing link jitter and risk control engine response anomalies), and combines multi-head attention mechanisms to identify hidden risk nodes in fund flows. Both output candidate solutions in parallel, ensuring rapid repair of common faults while also uncovering potential related faults, effectively reducing false positives and false negatives, and providing efficient and reliable decision support for financial business continuity.

[0042] In one embodiment, S400 includes: S401, Pre-build a digital twin rehearsal environment; S402. Obtain the business continuity impact score from the preset business continuity quantitative indicators; S403. Based on the calculation factors of the business continuity impact score and combined with the state transition logic of the core insurance business process, configure a business state tracking strategy in the digital twin pre-simulation environment. S404. Convert the candidate repair scheme into operation instructions, and input the operation instructions into the configured digital twin pre-simulation environment; S405. Simulate the repair operation corresponding to the operation instruction in the configured digital twin pre-simulation environment to obtain the pre-simulation evaluation result; S406. Calculate the impact value of performing the repair operation based on the pre-evaluation results; S407. If the impact value exceeds the preset safety threshold, the candidate repair solution is marked as a high-risk solution, and the review process is triggered. S408. If the impact value does not exceed the preset safety threshold, the candidate repair scheme is marked as a safe and feasible scheme, and the safe and feasible scheme is used as the target repair scheme.

[0043] In this embodiment, as Figure 6 As shown, before performing any remediation operations, the system needs to pre-build a high-fidelity digital twin pre-simulation environment. This environment is built on lightweight containerization technologies (such as Docker, an operating system-level virtualization technology that packages applications and their dependencies into standardized units; and Kubernetes, a container orchestration platform used for automated deployment, scaling, and management of containerized applications) to fully simulate the critical links of the core insurance business. This twin consists of three core components: first, simplified images of microservices such as underwriting, claims, and payment, which remove non-core auxiliary functions but retain the main business logic and state transition capabilities; second, a data sandbox, which is filled with anonymized real policy and claims sample data and can simulate reasonable business traffic to restore real load pressure; and third, a third-party system simulator, which uses mock server technology to simulate the response behavior of external dependencies such as reinsurance systems and payment gateways under different operating conditions, supports configurable switching between normal and abnormal states, and provides a complete external interactive environment for the pre-simulation.

[0044] Within the completed twin environment, the system further introduces a business continuity quantitative assessment mechanism, with its core indicator being the Business Continuity Impact Score (BCIS). This score is a comprehensive function whose calculation factors include the duration of business interruption, the number of affected policies, and the existence of any compliance or non-compliance items. To ensure that this indicator accurately reflects the characteristics of the insurance business, the system combines the standard status transition logic of core business processes such as underwriting, claims, and payment (e.g., the complete lifecycle of a claim from "under review" to "payment" and then to "case closed") and configures a detailed business status tracking strategy within the twin environment. This ensures that every status change of every business object during the simulation is accurately captured and recorded, providing fine-grained process data for subsequent impact score calculations.

[0045] Once candidate remediation plans are generated by the dual-channel inference engine, the system first converts these high-level remediation strategies (such as "restarting claims service" or "switching payment gateway routing") into specific operational instructions that can be directly executed in the digital twin environment. These instructions are then injected into the pre-configured digital twin pre-simulation environment. Within a secure, isolated sandbox, the system simulates the remediation actions corresponding to these instructions, driving business components, data flows, and third-party simulators within the digital twin to operate according to real logic, fully replicating the entire process from operation triggering to business process response. After the simulation, the system uses the execution trajectory data collected by the pre-configured state tracking strategy, substitutes it into the business continuity impact function, and calculates the comprehensive impact value of the remediation plan in terms of interruption duration, number of affected policies, and compliance dimensions.

[0046] The system then compares the impact value with a preset safety threshold. If the calculated impact value exceeds the threshold (e.g., an impact score greater than 0.6), the candidate remediation plan is marked as a high-risk plan, and the system automatically triggers a manual review process, submitting the detailed report, operation steps, and potential risk warnings generated during the simulation to the operations and maintenance experts for manual decision-making. Conversely, if the impact value is within the safety threshold range, the plan is deemed safe and feasible, and it is selected as the target remediation plan for the next step of actual execution. The entire execution trajectory of the simulation process, the impact value calculation results, and the record of whether the final review is passed are persistently stored and fed back into the knowledge graph as historical cases for subsequent optimization of the reasoning model and supplementation of the rule base, forming a complete governance closed loop from simulation to verification, from risk assessment to knowledge accumulation.

[0047] This embodiment pre-constructs a lightweight digital twin environment, introduces a Business Continuity Impact Score (BCIS), and configures a state tracking strategy. Candidate solutions are converted into instructions and simulated for execution within the twin environment. Impact values ​​are calculated and compared to safety thresholds. If the threshold is exceeded, manual review is triggered; otherwise, the solution is marked as the target solution. Virtual verification mitigates risks upfront, ensuring high-risk operations are safe and controllable, and preventing secondary failures.

[0048] In the fintech field, for example, before a bank plans to switch database master-slave nodes, it injects the operation into a twin environment to simulate fund flows and account status changes, calculates the interruption duration, the number of affected transactions, and compliance issues (such as customer fund segregation requirements), and comprehensively quantifies the impact. If the threshold is exceeded, a work order containing a pre-simulation report and rollback plan is generated and pushed to the operations and maintenance department for confirmation. This mechanism effectively reduces the risks of transaction interruptions, fund errors, or regulatory violations caused by operational changes, ensuring the safe and continuous operation of core financial businesses.

[0049] In one embodiment, S500 includes: S501. When the pre-assessment result is a low-risk repair operation, the repair operation of the target repair plan is executed directly. S502. When the pre-simulation assessment result is a high-risk or operation impact range exceeds the preset repair threshold, a repair work order containing a pre-simulation report, detailed operation steps and rollback plan is generated, and the repair work order is pushed to the operation and maintenance terminal. When a confirmation repair instruction is received, the repair operation is executed. S503. Collect the repair results after the actual repair operation is performed; S504. Compare the repair results with the historical case repair results in the business semantic knowledge graph to obtain the comparison results; S505. Update the reward function parameters of the dual-channel inference engine according to the comparison results, and adjust the output confidence and matching weight of the inference channel through the updated reward function parameters.

[0050] In this embodiment, as Figure 7 As shown, the remediation plan is moved from decision-making to actual execution, and differentiated operational strategies are implemented based on risk levels. First, the assessment results output from the digital twin pre-simulation stage are used to determine risk. When the pre-simulation assessment results indicate that the business continuity impact of a remediation operation is low and does not violate any compliance red lines (e.g., routine maintenance actions such as cleaning expired log files or refreshing distributed caches), the system classifies the plan as a low-risk operation and directly calls the maintenance robot (RPA, i.e., Robotic Process Automation, which uses scripts to simulate manual operations to execute repetitive instructions) to automatically execute the target remediation plan without human intervention, thereby minimizing fault recovery time.

[0051] Conversely, when the pre-simulation assessment results indicate that the repair operation carries a high risk of business interruption, or its impact extends to core transaction links, involves sensitive data operations, or exceeds the preset repair operation impact threshold (e.g., performing high-risk actions such as database master-slave switching, core microservice version rollback, or payment gateway route redirection), the system will not blindly automate the operation. Instead, it will automatically trigger a semi-automatic execution process. In this process, the system integrates and encapsulates the detailed assessment report generated during the pre-simulation phase, the precise operation steps verified in a twin environment, and a complete standardized rollback scheme (including rollback triggering conditions and specific execution instructions) into a structured "repair work order," which is then pushed to the operations personnel's terminals via the operations work order system or instant messaging. This work order provides sufficient contextual information and risk assessment basis for human decision-making. Only when the operations personnel review and confirm the work order and issue a clear confirmation repair instruction will the system execute the repair operation in the real production environment, thereby achieving an organic combination of automated efficiency and human safety safeguards.

[0052] Regardless of whether a fully automatic or semi-automatic execution path is used, the system initiates a result collection process upon completion. The collected repair results cover multi-dimensional information, including whether the operation was successfully completed, the total execution time, and whether any unexpected side effects were caused (such as increased latency in surrounding services or degradation of some non-core functions). This fine-grained execution data constitutes the key input for the system's self-optimization closed loop.

[0053] After data collection, the system performs a structured comparison between the current repair results and the massive historical case repair results stored in the business semantic knowledge graph. The comparison process not only focuses on the binary judgment of success and failure but also delves into the repair time efficiency, the similarity of side effect patterns, and the degree of matching between operational conditions and environmental context. Based on this comprehensive comparison result, the system invokes a reinforcement learning (RL) algorithm to iteratively update the reward function parameters within the dual-channel inference engine. In its implementation, the reward function's assignment logic follows a clear reward and punishment principle—significant positive reward scores are given to successful repairs without any side effects, while high-intensity negative penalties are imposed on results that cause secondary failures or expand the scope of impact. This reward signal propagates back to the symbolic inference channel, dynamically adjusting the trigger confidence threshold and priority ranking weight of each rule; it is also transmitted to the neural network inference channel, serving as an important supervisory signal for the gradient update of the graph neural network model parameters. This drives the model to gradually increase the output confidence of high-yield repair strategies and suppress the matching weight of high-risk strategies in subsequent inference, thereby enabling the entire system to continuously accumulate experience and optimize decision quality in ongoing operation and maintenance practices, forming a spiral upward evolution closed loop of "execution-feedback-learning-optimization".

[0054] In this embodiment, tiered repairs are implemented based on the pre-simulated risk level. Low-risk operations are executed fully automatically by the robot, while high-risk operations generate work orders with rollback plans, which are then pushed to manual confirmation before execution. Repair results are collected and compared with historical cases. The reward function is dynamically updated through reinforcement learning, continuously optimizing the confidence and matching weights of the inference engine, thereby achieving a dual improvement in repair efficiency and operational safety.

[0055] In the fintech field, taking insurance claims as an example, when the system detects an anomaly in the payment gateway, low-risk cache refresh operations are automatically executed for rapid recovery. High-risk operations, such as database master-slave switching, are simulated using a virtual twin to generate detailed work orders, which are then pushed to operations experts for confirmation and execution, preventing accidental fund miscalculations or data inconsistencies. This closed-loop mechanism can also be used for scenarios such as handling timeouts in bank transfers and backlogged securities trading orders. Through continuous learning and accumulation of domain-specific operations knowledge, it reduces reliance on expert experience and provides a solid guarantee for the continuity and compliance of financial services.

[0056] In one embodiment, after performing a tiered repair operation on the target repair scheme based on the pre-evaluation results, obtaining the repair results, and feeding them back to the business semantic knowledge graph and the dual-channel inference engine, the method further includes: Obtain the pre-simulation evaluation results and repair results; The pre-evaluation results and the repair results are used together as incremental feedback data and synchronously input into the business semantic knowledge graph to incrementally update the business semantic knowledge graph; The incremental feedback data is simultaneously transmitted to the dual-channel inference engine, and the policy triggering threshold in the symbolic inference channel and the model parameters in the neural network inference channel are adjusted according to the incremental feedback data.

[0057] In this embodiment, after a candidate remediation plan undergoes digital twin simulation and enters the actual execution phase, the system acquires feedback data in parallel from two paths: one from the digital twin simulation environment, including the business interruption duration recorded during the simulation, the number of affected policies, compliance violation detection results, and whether the final security threshold assessment is passed; the other from the actual execution process in the real production environment, including whether the remediation operation was successfully executed, whether the actual time consumption was within the expected range, whether the operation caused unexpected log errors or business interruptions, and whether the business ultimately returned to normal. These two types of data characterize the effectiveness and risk features of the same remediation plan from the two dimensions of "simulation simulation" and "real-world verification," respectively, complementing each other.

[0058] These two types of feedback data are integrated into incremental feedback data and synchronously input into the construction layer of the business semantic knowledge graph, driving the dynamic evolution of the graph. Specifically, the execution trajectory, impact score calculation details, and records of whether manual review was triggered in the pre-performance evaluation results are transformed into new operation and maintenance event nodes (such as "Pre-performance report of a certain repair solution"), and relationships such as "pre-performance completed" or "scope of impact" are established with related business services and technical components. Successful cases in the repair results are added as "historical successful repair cases" nodes, and relationships such as "resolved" or "root cause cleared" are established with the original fault work order nodes and business component nodes. Failed cases or operations that cause secondary faults are marked as "negative cases" and new relationship edges such as "caused fault" or "execution failed" are added. For new abnormal patterns or fault propagation paths discovered during pre-performance and actual execution that have not yet been recorded in the graph, the system will also supplement them into the graph through automated entity and relationship extraction, thereby realizing the continuous enrichment and self-correction of the knowledge graph in the operation and maintenance process.

[0059] Meanwhile, incremental feedback data is also synchronously transmitted to the dual-channel inference engine, serving as a key signal to drive iterative optimization of inference capabilities. For the symbolic inference channel, the system dynamically adjusts the trigger threshold and priority weight of each rule in the policy library based on the feedback results: for example, if a repair solution corresponding to a rule is proven to be safe, efficient, and have extremely low impact in multiple pre-runs or actual executions, the confidence threshold in its triggering conditions is appropriately lowered, making it easier to be matched and activated in similar fault scenarios; conversely, if a solution corresponding to a rule frequently causes side effects or triggers compliance alarms in actual execution, its trigger threshold is increased, making it more cautiously selected during inference. For the neural network inference channel, the system utilizes reinforcement learning (RL, a machine learning paradigm that learns optimal strategies by interacting with the environment and based on reward signals) to dynamically adjust the specific parameters of the reward function based on the repair results—for example, operations that successfully repair and whose business impact value is below a set threshold are given a higher positive reward (e.g., +10 points), while operations that cause secondary faults, business interruptions, or regulatory violations are given a strong negative penalty (e.g., -50 points). The updated reward function is further backpropagated to the weight update process of the graph neural network model through methods such as policy gradient, adjusting the attention intensity of the attention mechanism on different neighbor nodes and the decision boundary of the root cause classification layer. This makes the subsequent inference output confidence and matching weights more inclined to those repair strategies that have been verified to be safe and effective in practice. Through this continuous closed-loop process of "simulation-actual execution-feedback collection-knowledge accumulation-parameter tuning", the system can continuously accumulate experience in actual operation and maintenance scenarios, avoid repeating mistakes, and gradually approach the optimal autonomous repair decision-making capability.

[0060] This embodiment uses both the digital twin pre-simulation evaluation results and the actual production environment repair results as incremental feedback data, which are synchronously input into the knowledge graph for dynamic updates and transmitted to the dual-channel inference engine to adjust the rule trigger threshold of the symbolic channel and the model parameters of the neural network channel. Through closed-loop feedback, the inference decision-making capability is continuously optimized, forming a positive cycle of knowledge accumulation and model evolution.

[0061] In the fintech field, taking the core insurance system as an example, during each fault handling process, the differences between the twin simulation results and the actual repair effects (such as boundary conditions not exposed during the simulation or sudden dependency timeouts in actual production) are incrementally recorded in the knowledge graph, continuously supplementing the fault case library and repair strategy graph, making the symbol rule matching for subsequent similar faults more accurate. Simultaneously, feedback data continuously fine-tunes the distribution of neural network attention weights, enabling the model to have a more acute perception of the hidden risks unique to peak financial transaction periods (such as queue backlogs caused by a surge in claims at the end of the month). After long-term operation of this mechanism, the operation and maintenance system can gradually break free from path dependence on senior experts, achieving self-evolving intelligent operation and maintenance capabilities, providing lasting and continuously optimized technical support for the stable operation of financial businesses.

[0062] In one embodiment, a fault repair device based on a semantic knowledge graph is provided, which corresponds one-to-one with the fault repair method based on a semantic knowledge graph in the above embodiments. (Refer to...) Figure 8 , Figure 8 This is a schematic diagram of the functional modules of a preferred embodiment of the fault repair device based on semantic knowledge graph of the present invention. The module includes a data acquisition module 10, a knowledge graph construction module 20, a candidate repair module 30, a target repair module 40, and a repair execution module 50. Detailed descriptions of each functional module are as follows: The data acquisition module 10 is used to acquire multi-source heterogeneous data, and to perform business semantic annotation on the multi-source heterogeneous data to extract structured information. The graph construction module 20 is used to construct a business semantic knowledge graph based on the structured information; The candidate repair module 30 is used to call the dual-channel inference engine when the system fails, combine the business semantic knowledge graph to perform root cause analysis of the failure, and generate candidate repair solutions based on the analysis results. The target repair module 40 is used to simulate the execution of the candidate repair scheme in the digital twin pre-simulation environment, obtain the pre-simulation evaluation results, evaluate the impact value of the candidate repair scheme according to the business continuity quantitative index, and select the target repair scheme according to the impact value. The repair execution module 50 is used to perform hierarchical repair operations on the target repair scheme based on the pre-evaluation results, obtain the repair results, and feed them back to the business semantic knowledge graph and the dual-channel inference engine.

[0063] In one embodiment, the data acquisition module 10 includes: Collect business transaction data, operation and maintenance status data, change and event data, and business strategy data, which together constitute multi-source heterogeneous data; Based on an insurance domain dictionary, a sequence labeling model based on bidirectional long short-term memory networks and conditional random fields is used to extract business entities and their attribute information from the unstructured text of the multi-source heterogeneous data. Based on the business entities, the long-distance dependencies and complex sentence structures in the unstructured text are processed to identify the business actions and related objects between the business entities. Generate a triplet structure based on the business actions and associated objects between the business entities; Perform semantic classification on the abnormal logs and label each abnormal log with a business impact tag. Structured information is generated based on the business entity and its attribute information, the triplet structure, and the labeled business impact tags.

[0064] In one embodiment, the map construction module 20 includes: Predefine an ontology model that includes core entity types and business semantic relationships; Based on the core entity types and business semantic relationships in the ontology model, and in conjunction with the insurance business architecture document, determine the entity types and their corresponding relationship skeletons; Instance data is automatically extracted from the structured information, the instance data is categorized according to entity type, and populated into the relation skeleton to generate a business semantic knowledge graph; When a new insurance business or a new entity is detected, it is determined whether the new insurance business or the new entity belongs to an existing entity type based on the core entity type in the ontology model. If it belongs to an existing entity type, the new insurance business or new entity will be classified into the relation skeleton corresponding to the existing entity type. If it does not belong to an existing entity type, then create an entity type for the new insurance business or new entity and populate it into the corresponding relationship skeleton; The knowledge graph completion algorithm is invoked to infer the association between the new insurance business or new entity and existing entity types, and the business semantic knowledge graph is updated based on the association.

[0065] In one embodiment, the candidate repair module 30 includes: When a system failure occurs, the symbolic reasoning channel of the dual-channel inference engine performs inference based on a pre-set strategy library and outputs the first candidate repair scheme corresponding to the known failure mode. The neural network inference channel of the dual-channel inference engine uses a graph neural network model with an attention fusion mechanism for inference. The subgraph related to the fault node in the business semantic knowledge graph is used as the input of the graph neural network model, and the root cause probability distribution is output. The repair strategy corresponding to the root cause whose root cause probability distribution exceeds the preset threshold is taken as the second candidate repair scheme. The first and second candidate repair schemes are used to form a candidate repair scheme.

[0066] In one embodiment, the target repair module 40 includes: Pre-build a digital twin rehearsal environment; Obtain the business continuity impact score from the preset business continuity quantitative indicators; Based on the calculation factors of the business continuity impact score, and combined with the state transition logic of the core insurance business process, a business state tracking strategy is configured in the digital twin pre-simulation environment. The candidate repair schemes are converted into operation instructions, and the operation instructions are input into the configured digital twin pre-simulation environment; The repair operation corresponding to the operation instruction is simulated and executed in the configured digital twin pre-simulation environment to obtain the pre-simulation evaluation result; Based on the results of the pre-simulation assessment, calculate the impact value of performing the repair operation; If the impact value exceeds a preset safety threshold, the candidate repair solution will be marked as a high-risk solution, and the review process will be triggered. If the impact value does not exceed the preset safety threshold, the candidate repair scheme is marked as a safe and feasible scheme, and the safe and feasible scheme is used as the target repair scheme.

[0067] In one embodiment, the repair execution module 50 includes: When the pre-analysis assessment result is a low-risk repair operation, the repair operation of the target repair plan is executed directly. When the pre-simulation assessment result is a high-risk or the scope of operation impact exceeds the preset repair threshold, a repair work order containing a pre-simulation report, detailed operation steps and rollback plan is generated and pushed to the operation and maintenance terminal. When a confirmation repair instruction is received, the repair operation is executed. Collect the repair results after the actual repair operation is performed; The repair results are compared with the repair results of historical cases in the business semantic knowledge graph to obtain the comparison results; The reward function parameters of the dual-channel inference engine are updated based on the comparison results, and the output confidence and matching weight of the inference channel are adjusted by the updated reward function parameters.

[0068] In one embodiment, the map and inference update module further includes: Obtain the pre-simulation evaluation results and repair results; The pre-evaluation results and the repair results are used together as incremental feedback data and synchronously input into the business semantic knowledge graph to incrementally update the business semantic knowledge graph; The incremental feedback data is simultaneously transmitted to the dual-channel inference engine, and the policy triggering threshold in the symbolic inference channel and the model parameters in the neural network inference channel are adjusted according to the incremental feedback data.

[0069] Specific limitations regarding the fault repair device based on semantic knowledge graphs can be found in the aforementioned limitations on fault repair methods based on semantic knowledge graphs, and will not be repeated here. Each module in the aforementioned fault repair device based on semantic knowledge graphs can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0070] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side fault repair method based on semantic knowledge graphs.

[0071] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps of a fault repair method based on semantic knowledge graphs.

[0072] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Collect multi-source heterogeneous data, and perform business semantic annotation on the multi-source heterogeneous data to extract structured information; Based on the structured information, a business semantic knowledge graph is constructed; When a system failure occurs, the dual-channel inference engine is invoked to perform root cause analysis of the failure in conjunction with the business semantic knowledge graph, and candidate repair solutions are generated based on the analysis results. The candidate repair scheme is simulated and executed in a digital twin pre-simulation environment to obtain pre-simulation evaluation results. The impact value of the candidate repair scheme is evaluated based on the business continuity quantitative index, and the target repair scheme is selected based on the impact value. Based on the pre-evaluation results, a tiered repair operation is performed on the target repair scheme, and the repair results are obtained and fed back to the business semantic knowledge graph and the dual-channel inference engine.

[0073] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, and a computer program is stored thereon. When the computer program is executed by a processor, it performs the following steps: Collect multi-source heterogeneous data, and perform business semantic annotation on the multi-source heterogeneous data to extract structured information; Based on the structured information, a business semantic knowledge graph is constructed; When a system failure occurs, the dual-channel inference engine is invoked to perform root cause analysis of the failure in conjunction with the business semantic knowledge graph, and candidate repair solutions are generated based on the analysis results. The candidate repair scheme is simulated and executed in a digital twin pre-simulation environment to obtain pre-simulation evaluation results. The impact value of the candidate repair scheme is evaluated based on the business continuity quantitative index, and the target repair scheme is selected based on the impact value. Based on the pre-evaluation results, a tiered repair operation is performed on the target repair scheme, and the repair results are obtained and fed back to the business semantic knowledge graph and the dual-channel inference engine.

[0074] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0075] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0076] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0077] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0078] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A fault repair method based on semantic knowledge graph, characterized in that, Includes the following steps: Collect multi-source heterogeneous data, and perform business semantic annotation on the multi-source heterogeneous data to extract structured information; Based on the structured information, a business semantic knowledge graph is constructed; When a system failure occurs, the dual-channel inference engine is invoked to perform root cause analysis of the failure in conjunction with the business semantic knowledge graph, and candidate repair solutions are generated based on the analysis results. The candidate repair scheme is simulated and executed in a digital twin pre-simulation environment to obtain pre-simulation evaluation results. The impact value of the candidate repair scheme is evaluated based on the business continuity quantitative index, and the target repair scheme is selected based on the impact value. Based on the pre-evaluation results, a tiered repair operation is performed on the target repair scheme, and the repair results are obtained and fed back to the business semantic knowledge graph and the dual-channel inference engine.

2. The fault repair method based on semantic knowledge graph as described in claim 1, characterized in that, Collect multi-source heterogeneous data, and perform business semantic annotation on the multi-source heterogeneous data to extract structured information, including: Collect business transaction data, operation and maintenance status data, change and event data, and business strategy data, which together constitute multi-source heterogeneous data; Based on an insurance domain dictionary, a sequence labeling model based on bidirectional long short-term memory networks and conditional random fields is used to extract business entities and their attribute information from the unstructured text of the multi-source heterogeneous data. Based on the business entities, the long-distance dependencies and complex sentence structures in the unstructured text are processed to identify the business actions and related objects between the business entities. Generate a triplet structure based on the business actions and associated objects between the business entities; Perform semantic classification on the abnormal logs and label each abnormal log with a business impact tag. Structured information is generated based on the business entity and its attribute information, the triplet structure, and the labeled business impact tags.

3. The fault repair method based on semantic knowledge graph as described in claim 1, characterized in that, Based on the structured information, a business semantic knowledge graph is constructed, including: Predefine an ontology model that includes core entity types and business semantic relationships; Based on the core entity types and business semantic relationships in the ontology model, and in conjunction with the insurance business architecture document, determine the entity types and their corresponding relationship skeletons; Instance data is automatically extracted from the structured information, the instance data is categorized according to entity type, and populated into the relation skeleton to generate a business semantic knowledge graph; When a new insurance business or a new entity is detected, it is determined whether the new insurance business or the new entity belongs to an existing entity type based on the core entity type in the ontology model. If it belongs to an existing entity type, the new insurance business or new entity will be classified into the relation skeleton corresponding to the existing entity type. If it does not belong to an existing entity type, then create an entity type for the new insurance business or new entity and populate it into the corresponding relationship skeleton; The knowledge graph completion algorithm is invoked to infer the association between the new insurance business or new entity and existing entity types, and the business semantic knowledge graph is updated based on the association.

4. The fault repair method based on semantic knowledge graph as described in claim 1, characterized in that, When a system failure occurs, the dual-channel inference engine is invoked to perform root cause analysis in conjunction with the business semantic knowledge graph. Based on the analysis results, candidate repair solutions are generated, including: When a system failure occurs, the symbolic reasoning channel of the dual-channel inference engine performs inference based on a pre-set strategy library and outputs the first candidate repair scheme corresponding to the known failure mode. The neural network inference channel of the dual-channel inference engine uses a graph neural network model with an attention fusion mechanism for inference. The subgraph related to the fault node in the business semantic knowledge graph is used as the input of the graph neural network model, and the root cause probability distribution is output. The repair strategy corresponding to the root cause whose root cause probability distribution exceeds the preset threshold is taken as the second candidate repair scheme. The first and second candidate repair schemes are used to form a candidate repair scheme.

5. The fault repair method based on semantic knowledge graph as described in claim 1, characterized in that, The candidate repair schemes are simulated and executed in a digital twin pre-simulation environment to obtain pre-simulation evaluation results. The impact values ​​of the candidate repair schemes are then evaluated based on business continuity quantification indicators. Target repair schemes are selected based on these impact values, including: Pre-build a digital twin rehearsal environment; Obtain the business continuity impact score from the preset business continuity quantitative indicators; Based on the calculation factors of the business continuity impact score, and combined with the state transition logic of the core insurance business process, a business state tracking strategy is configured in the digital twin pre-simulation environment. The candidate repair schemes are converted into operation instructions, and the operation instructions are input into the configured digital twin pre-simulation environment; The repair operation corresponding to the operation instruction is simulated and executed in the configured digital twin pre-simulation environment to obtain the pre-simulation evaluation result; Based on the results of the pre-simulation assessment, calculate the impact value of performing the repair operation; If the impact value exceeds a preset safety threshold, the candidate repair solution will be marked as a high-risk solution, and the review process will be triggered. If the impact value does not exceed the preset safety threshold, the candidate repair scheme is marked as a safe and feasible scheme, and the safe and feasible scheme is used as the target repair scheme.

6. The fault repair method based on semantic knowledge graph as described in claim 1, characterized in that, Based on the pre-evaluation results, a tiered repair operation is performed on the target repair plan, and the repair results are obtained and fed back to the business semantic knowledge graph and the dual-channel inference engine, including: When the pre-analysis assessment result is a low-risk repair operation, the repair operation of the target repair plan is executed directly. When the pre-simulation assessment result is a high-risk or the scope of operation impact exceeds the preset repair threshold, a repair work order containing a pre-simulation report, detailed operation steps and rollback plan is generated and pushed to the operation and maintenance terminal. When a confirmation repair instruction is received, the repair operation is executed. Collect the repair results after the actual repair operation is performed; The repair results are compared with the repair results of historical cases in the business semantic knowledge graph to obtain the comparison results; The reward function parameters of the dual-channel inference engine are updated based on the comparison results, and the output confidence and matching weight of the inference channel are adjusted by the updated reward function parameters.

7. The fault repair method based on semantic knowledge graph as described in claim 4, characterized in that, Based on the pre-simulation evaluation results, after performing tiered repair operations on the target repair plan, obtaining the repair results, and feeding them back to the business semantic knowledge graph and the dual-channel inference engine, the process further includes: Obtain the pre-simulation evaluation results and repair results; The pre-evaluation results and the repair results are used together as incremental feedback data and synchronously input into the business semantic knowledge graph to incrementally update the business semantic knowledge graph; The incremental feedback data is simultaneously transmitted to the dual-channel inference engine, and the policy triggering threshold in the symbolic inference channel and the model parameters in the neural network inference channel are adjusted according to the incremental feedback data.

8. A fault repair device based on semantic knowledge graph, characterized in that, The fault repair device based on semantic knowledge graph includes: The data acquisition module is used to collect multi-source heterogeneous data, and to perform business semantic annotation on the multi-source heterogeneous data to extract structured information; The graph construction module is used to construct a business semantic knowledge graph based on the structured information; The candidate repair module is used to call the dual-channel inference engine when the system fails, combine the business semantic knowledge graph to perform root cause analysis of the failure, and generate candidate repair solutions based on the analysis results. The target repair module is used to simulate the execution of the candidate repair scheme in the digital twin pre-simulation environment, obtain the pre-simulation evaluation results, evaluate the impact value of the candidate repair scheme according to the business continuity quantitative index, and select the target repair scheme according to the impact value. The repair execution module is used to perform hierarchical repair operations on the target repair scheme based on the pre-evaluation results, obtain the repair results, and feed them back to the business semantic knowledge graph and the dual-channel inference engine.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a semantic knowledge graph-based fault repair program stored in the memory and executable on the processor. When executed by the processor, the semantic knowledge graph-based fault repair program implements the steps of the semantic knowledge graph-based fault repair method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a fault repair program based on a semantic knowledge graph, which, when executed by a processor, implements the steps of the fault repair method based on a semantic knowledge graph as described in any one of claims 1-7.

Images