Intelligent decision support method and system based on enterprise knowledge graph

Abstract

The invention relates to the technical field of knowledge maps and intelligent decision, and discloses an intelligent decision support method and system based on enterprise knowledge maps, and the intelligent decision support method comprises the steps: collecting original data to obtain a standardized data set; performing knowledge extraction and fusion, and obtaining a domain knowledge graph by adopting an ontology evolution mechanism; symbol rule extraction and vector representation learning are carried out, and a two-way attention fusion mechanism is adopted to obtain an inference result and an explanation link; performing time sequence correlation analysis, and obtaining a decision sequence diagram by adopting a time decay attention mechanism; evaluating the supply chain toughness by adopting a health state prediction method and a network topology vulnerability analysis method to obtain a toughness enhancement scheme; performing multi-objective optimization sorting to obtain a recommended decision scheme; and performing knowledge verification by executing feedback, and obtaining an optimized knowledge graph by adopting an incremental updating method. According to the method, symbol reasoning and vector reasoning can be fused, and accurate and interpretable intelligent decision support is provided.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph and intelligent decision-making technology, and more specifically, to an intelligent decision support method and system based on enterprise knowledge graph. Background Technology

[0002] As the complexity of global supply chain networks continues to increase, enterprises face challenges such as difficulties in integrating multi-source heterogeneous data, fragmented decision-making knowledge, and delayed risk warnings. Traditional decision support systems mainly rely on rule engines or statistical models, which struggle to effectively handle unstructured knowledge and lack a deep understanding of business logic. Knowledge graph technology offers a new solution for the structured representation and reasoning of enterprise knowledge, but existing knowledge graph-based decision support methods still have many shortcomings.

[0003] In existing technologies, while symbolic reasoning-based methods offer strong interpretability, they rely on manually constructed rules and struggle to handle incomplete knowledge and ambiguous scenarios. Vector representation learning-based methods, while exhibiting good generalization capabilities, lack interpretability in their reasoning process, making it difficult to gain the trust of decision-makers. Furthermore, existing methods often employ static knowledge graphs, which cannot adaptively evolve with business development, leading to decreased knowledge timeliness. In supply chain resilience management scenarios, existing methods lack the ability to predict the risk of cascading failures in a network, often relying on reactive responses and failing to achieve proactive prevention.

[0004] Therefore, there is a need for an intelligent decision support method that can integrate the advantages of symbolic reasoning and vector reasoning, support the adaptive evolution of knowledge graphs, and have the ability to predict network-cascaded risks, in order to solve the technical problems faced by enterprises in complex decision-making scenarios, such as difficulties in knowledge integration, insufficient interpretability of reasoning, and delayed risk warning. Summary of the Invention

[0005] This invention provides an intelligent decision support method and system based on enterprise knowledge graphs, which solves the technical problems of knowledge integration difficulties, insufficient interpretability of reasoning, inability of knowledge graphs to adapt and evolve, and delayed risk warning in related technologies.

[0006] This invention provides an intelligent decision support method based on enterprise knowledge graphs, comprising the following steps:

[0007] S100: Collect raw data from the data access channel and use standardization processing to obtain a cleaned standardized data set;

[0008] S200 acquires a standardized dataset for knowledge extraction and fusion, and uses an ontology evolution mechanism to obtain an adaptively updated domain knowledge graph.

[0009] S300 extracts symbolic rules and learns vector representations based on the domain knowledge graph, and uses a bidirectional attention fusion mechanism to obtain inference results and explanation links;

[0010] S400 performs temporal correlation analysis based on the domain knowledge graph and historical decision events, and uses a time decay attention mechanism to obtain a decision time sequence graph.

[0011] S500 acquires domain knowledge graphs and decision sequence diagrams, and uses health status prediction methods and network topology vulnerability analysis methods to assess supply chain resilience and obtain resilience enhancement solutions.

[0012] S600 performs multi-objective optimization based on the reasoning results and resilience enhancement schemes, and ranks them to obtain recommended decision schemes;

[0013] S700 performs knowledge verification through the execution feedback of recommended decision-making schemes and obtains an optimized knowledge graph using an incremental update method.

[0014] In a preferred embodiment, the step of collecting raw data according to the data access channel and obtaining a cleaned standardized data set through standardization processing includes:

[0015] A multi-source data access channel is established based on internal and external data sources. The raw data from each data source is aggregated according to the collection timestamp to form a raw data aggregation pool.

[0016] Based on the heterogeneous data in the original data aggregation pool, a format conversion rule base and a field mapping table are constructed to obtain normalized data;

[0017] A data quality anomaly detection model is constructed based on standardized data, integrating statistical detection and machine learning detection. An active learning mechanism is introduced to incrementally update the parameters of the anomaly detection model.

[0018] In a preferred embodiment, the step of acquiring a standardized data set for knowledge extraction and fusion, and using an ontology evolution mechanism to obtain an adaptively updated domain knowledge graph, includes:

[0019] For structured data, a rule-matching method is used for knowledge extraction, while for unstructured data, a sequence labeling model is used for entity extraction and a relation classification model is used for relation extraction, forming a set of candidate knowledge triples.

[0020] Based on the candidate knowledge triple set, a three-level knowledge fusion method of entity alignment, attribute merging and conflict resolution is adopted to obtain fused knowledge instance data.

[0021] Establish a monitoring mechanism for the triggering conditions of ontology evolution, including new concept emergence detection, relation pattern discovery, concept semantic drift detection, and ontology structure inconsistency detection. When the triggering conditions are met, the ontology evolution rule base is used to evaluate and execute the evolution scheme.

[0022] In a preferred embodiment, the step of extracting symbolic rules and learning vector representations according to the domain knowledge graph, and obtaining the reasoning result and explanation link using a bidirectional attention fusion mechanism, includes:

[0023] Path pattern mining is performed based on domain knowledge graphs. Path patterns that occur more frequently than a preset support threshold are generalized into rules to form a symbol rule base.

[0024] A rule-constrained vector representation learning method is adopted to construct an objective function consisting of a basic score function, a negative sampling loss, and rule constraint terms, thereby obtaining entity vectors and relation vectors;

[0025] The symbolic reasoning engine and the vector reasoning engine are launched in parallel to obtain the symbolic reasoning result set and the vector reasoning result set, respectively.

[0026] A heterogeneous two-layer inference graph structure from symbol to vector is constructed, and the two layers are connected across layers through semantic alignment edges. The structural confidence is calculated in the symbol-to-vector direction, and the fusion weights adopt a meta-learning adaptive adjustment mechanism. Semantic consistency score and context relevance score are calculated in the vector-to-symbol direction. Attention information in both directions is iteratively propagated through the message passing algorithm of graph neural network to output the fused inference result.

[0027] In a preferred embodiment, the step of performing temporal correlation analysis based on the domain knowledge graph and historical decision events, and obtaining the decision time sequence graph using a time decay attention mechanism, includes:

[0028] Based on the enterprise's historical decision-making data, decision timestamps, decision types, measures taken, and evaluation of decision results are extracted to form a set of decision event records;

[0029] Analyze the temporal sequence, causal influence, goal inheritance, and resource competition relationships among decision-making events, and construct a decision-making temporal relationship diagram structure;

[0030] The decision nodes are associated with the domain knowledge graph. The association is divided into three levels: business object association, constraint condition association, and target indicator association.

[0031] A multi-factor time decay model is constructed, with the basic time decay function adopting a two-parameter Weibull distribution. An environmental stability adjustment factor and a decision effect persistence adjustment factor are introduced. For historical decisions with causal relationships, the correlation strength is superimposed to form a historical decision context.

[0032] In a preferred embodiment, the steps of acquiring a domain knowledge graph and a decision sequence graph, and assessing supply chain resilience using health status prediction methods and network topology vulnerability analysis methods to obtain a resilience enhancement scheme include:

[0033] Differentiated health status indicators are designed for each node in the supply chain, and the real-time data of each indicator is collected and normalized.

[0034] A three-layer fusion prediction architecture is constructed. The first layer is a single-indicator time series prediction layer, the second layer is a cross-indicator correlation prediction layer, a Bayesian network is constructed to represent causal dependencies, and the third layer is an external signal enhancement layer.

[0035] The supply chain knowledge graph is abstracted into a weighted directed network structure, and a node capacity model and a dynamic load redistribution model are constructed. If the node load exceeds the preset capacity threshold, cascading propagation is triggered. The cascading scale index, cascading depth index, key trigger node set, and cascading amplification node set are calculated.

[0036] In a preferred embodiment, the step of acquiring a domain knowledge graph and a decision sequence graph, assessing supply chain resilience using health status prediction methods and network topology vulnerability analysis methods, and obtaining a resilience enhancement scheme further includes:

[0037] The generative adversarial network framework for constructing resilient solutions includes a generator network and a discriminator network. The generator network takes the feature vectors and constraint vectors of high-risk nodes as input and outputs candidate resilience enhancement solutions.

[0038] The case-based reasoning method is used to retrieve historical successful cases to verify the feasibility of candidate solutions and optimize parameters.

[0039] Generate differentiated candidate enhancement solutions for different risk types.

[0040] In a preferred embodiment, the step of performing multi-objective optimization based on the reasoning results and resilience enhancement schemes to rank and obtain recommended decision schemes includes:

[0041] Establish a decision-making objective template library and match objective dimension combinations according to the type of decision-making scenario;

[0042] Candidate solution generation is divided into three approaches: historical case retrieval, rule-driven generation, and resilient solution integration, forming a set of candidate decision-making solutions.

[0043] Quantitative predictions of candidate solutions are used to construct a solution evaluation matrix;

[0044] A multi-stage ranking optimization framework is constructed. The first stage is to perform non-dominated ranking using Pareto hierarchical structure. The second stage is to calculate the weighted Chebyshev distance using reference point-guided optimization. The third stage is to perform robust adjustment by simulating the ranking variance using Monte Carlo simulation. The fourth stage is to generate a set of recommendation schemes through interactive recommendation.

[0045] In a preferred embodiment, the step of verifying knowledge through the execution feedback of the recommended decision scheme and obtaining the optimized knowledge graph using an incremental update method includes:

[0046] Establish two channels for collecting feedback: explicit feedback and implicit feedback.

[0047] The feedback information is categorized and the problem is identified and analyzed to form a list of knowledge quality issues, and a consistency check is performed to generate a check report;

[0048] To address knowledge quality issues, generate knowledge correction plans, knowledge supplementation plans, knowledge update plans, and conflict resolution plans;

[0049] Knowledge updates are performed incrementally, and a version management mechanism is established to support retrieval and rollback.

[0050] In a preferred embodiment, an intelligent decision support system based on enterprise knowledge graphs is used to perform the steps of the aforementioned intelligent decision support method based on enterprise knowledge graphs, including:

[0051] The data standardization module is used to collect raw data from the data access channel and then perform standardization processing to obtain a cleaned and standardized dataset.

[0052] The knowledge graph construction module is used to acquire standardized data sets for knowledge extraction and fusion, and adopts an ontology evolution mechanism to obtain an adaptively updated domain knowledge graph.

[0053] The neural symbolic reasoning module is used to extract symbolic rules and learn vector representations according to the domain knowledge graph, and uses a bidirectional attention fusion mechanism to obtain reasoning results and interpretation links.

[0054] The temporal analysis module is used to perform temporal correlation analysis based on the domain knowledge graph and historical decision events, and uses a time decay attention mechanism to obtain a decision time sequence graph.

[0055] The resilience assessment module is used to acquire domain knowledge graphs and decision sequence diagrams, and to assess supply chain resilience using health status prediction methods and network topology vulnerability analysis methods, thereby obtaining resilience enhancement solutions.

[0056] The decision optimization module is used to perform multi-objective optimization based on the reasoning results and resilience enhancement schemes, and rank them to obtain recommended decision schemes.

[0057] The knowledge update module is used to verify knowledge through the execution feedback of the recommended decision scheme, and uses an incremental update method to obtain an optimized knowledge graph.

[0058] The beneficial effects of this invention are as follows:

[0059] By constructing a symbol-vector heterogeneous two-layer reasoning graph structure and adopting a bidirectional attention fusion mechanism based on graph neural networks, the deep integration of the logical rigor of symbolic reasoning and the semantic generalization ability of vector reasoning is realized. The fusion weight is dynamically adjusted using a meta-learning adaptive adjustment mechanism, which solves the technical problems in the existing technology that symbolic reasoning has strong interpretability but weak generalization ability, and vector reasoning has strong generalization ability but insufficient interpretability, and provides reasoning results that are both accurate and interpretable.

[0060] By establishing a multi-dimensional monitoring mechanism for ontology evolution triggering conditions, including new concept emergence detection, relation pattern discovery, concept semantic drift detection, and ontology structure inconsistency detection, and combining this with an ontology evolution rule base for evolution scheme evaluation and execution, the adaptive evolution of the knowledge graph with business development is achieved, solving the technical problems of static solidification of knowledge graphs and decreased knowledge timeliness in existing technologies. At the same time, by constructing a three-layer fusion prediction architecture and a dynamic load redistribution model, the prediction of supply chain network cascading failure risks is achieved. Combined with adversarial generative networks and case reasoning methods, preventive resilience enhancement schemes are generated, solving the technical problems of delayed risk warning and lack of proactive prevention capabilities in existing technologies. Attached Figure Description

[0061] Figure 1 This is a flowchart of an intelligent decision support method based on enterprise knowledge graph according to the present invention;

[0062] Figure 2 This is a flowchart of an intelligent decision support method based on enterprise knowledge graphs according to the present invention. Detailed Implementation

[0063] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0064] At least one embodiment of the present invention discloses an intelligent decision support method based on enterprise knowledge graphs, such as... Figures 1 to 2 As shown, it includes the following steps:

[0065] S100: Collect raw data from the data access channel and use standardization processing to obtain a cleaned standardized data set;

[0066] A multi-source data access channel is established based on internal business systems and external data sources. Preprocessing methods such as format conversion, data cleaning, and semantic annotation are used to obtain a standardized dataset for subsequent knowledge graph construction. Specifically, the steps include:

[0067] S101, Establishment of Multi-Source Data Access Channels: Based on the enterprise information system architecture, a unified data interface protocol is used to establish multi-source data access channels, resulting in a raw data aggregation pool. Specifically, for internal data sources, database connectors are used to connect to supplier master data, purchase order data, and inventory data in the Enterprise Resource Planning (ERP) system; customer information and sales order data in the Customer Relationship Management (CRM) system; logistics tracking data and supplier performance data in the Supply Chain Management (SLM) system; and cost data and payment records in the financial system. For external data sources, application programming interfaces (APIs) are used to connect to industry databases to obtain industry dynamics and market trend information; to market research institutions to obtain supplier credit ratings and risk warning information; and to policy and regulatory databases to obtain trade policies and compliance requirements. The raw data from each data source is aggregated according to the collection timestamp to form the raw data aggregation pool.

[0068] S102, Data Format Conversion and Field Mapping: Based on heterogeneous data in the original data aggregation pool, format conversion and field mapping methods are used to obtain standardized data conforming to a unified data schema. Specifically, firstly, a unified data schema specification is defined, including standard attribute sets for supplier entities, product entities, order entities, and logistics entities. Then, a format conversion rule base is constructed to uniformly convert different date formats, numeric formats, and encoding formats into the standard format. Simultaneously, a field mapping table is constructed to map semantically identical but differently named fields from various data sources to unified standard field names. After format conversion and field mapping, heterogeneous data is unified into standardized data conforming to the data schema specification.

[0069] S103, Intelligent Data Cleaning Based on Active Learning: Based on standardized data, an intelligent data cleaning method enhanced by active learning is used to obtain high-quality cleaned data. Specifically, a data quality anomaly detection model is constructed, integrating statistical detection and machine learning detection mechanisms. The statistical detection mechanism uses a quantile-based outlier detection method for numerical fields, marking values ​​that deviate from the median by more than a preset multiple of the interquartile range as statistical anomalies. For categorical fields, a frequency analysis method is used, marking rare values ​​with a frequency below a preset threshold as potential anomalies. The machine learning detection mechanism uses the isolated forest algorithm to construct a multi-dimensional anomaly detection model, calculating anomaly scores for each record through a randomly segmented tree structure.

[0070] Furthermore, an active learning mechanism is introduced to handle detected anomalous data. This mechanism maintains an uncertainty sample pool, adding data records where statistical and machine learning detection results differ. For each record in the pool, the system calculates its uncertainty score, which considers both the confidence difference between the two detection methods and the information gain for model training. The sample pool is then sorted by uncertainty score, and the highest-scoring sample is submitted to domain experts for manual annotation. Expert annotations include whether the record is a true anomaly, the anomaly type, and suggested corrections. The annotated samples are added to the training set, incrementally updating the parameters of the anomaly detection model. Through iterative active learning, the model gradually learns domain-specific data quality rules, continuously improving detection accuracy. For missing values, a knowledge graph-based association inference imputation method is used, leveraging existing entity relationships in the knowledge graph to infer possible values ​​for missing attributes. The cleaning process generates a data cleaning log, recording the cleaning status and confidence score for each data entry.

[0071] S104, Semantic Annotation of Unstructured Data: Based on the unstructured text content in the cleaned data, natural language processing methods are used for semantic annotation to obtain structured annotation results. Specifically, firstly, intelligent word segmentation is performed on the text content, identifying professional terms and entity names based on a supply chain domain dictionary. Then, multi-type named entity recognition is performed to identify entities and their types in the text, such as supplier names, product names, geographical locations, time expressions, and quantity expressions. For the identified entities, entity disambiguation is performed, linking entity mentions in the text to the corresponding standard entities in the knowledge graph. The semantic annotation results are output in the form of annotation sequences, with each annotation containing a text fragment, entity type, and the standard entity identifier of the link.

[0072] S105, Data Quality Assessment and Credibility Labeling: Based on cleaned data and semantic annotation results, a multi-dimensional quality assessment method is employed to obtain a data quality assessment report and credibility labels. Specifically, a data quality assessment indicator system is constructed, including four dimensions: completeness, accuracy, consistency, and timeliness. Based on the scores of the four dimensions, a weighted comprehensive method is used to calculate the overall quality score of each data source. According to the overall quality score, data sources are divided into three levels of credibility: high credibility, medium credibility, and low credibility. The credibility level will affect the weight allocation during subsequent knowledge extraction. The assessment results form a data quality assessment report, which includes the quality score, credibility level, and quality issue statistics for each data source.

[0073] S200 acquires a standardized dataset for knowledge extraction and fusion, and uses an ontology evolution mechanism to obtain an adaptively updated domain knowledge graph.

[0074] Based on the standardized dataset and initial domain ontology model output in step S100, a domain knowledge graph that can adaptively update with business development is obtained by employing a multi-model integrated knowledge extraction method and a multi-level knowledge fusion method, combined with a dynamic ontology evolution mechanism; specifically, the following steps are included:

[0075] S201, Multi-Model Integration Knowledge Extraction: Based on a standardized dataset and an initial domain ontology model, a multi-model integration knowledge extraction method is used to obtain a set of candidate knowledge triples. Specifically, the initial domain ontology model defines the core concept types and relationship types in the supply chain domain, including entity types such as suppliers, manufacturers, products, components, orders, and logistics nodes, as well as relationship types such as supply relationships, composition relationships, fulfillment relationships, and transportation relationships. For structured data, a rule-matching method is used for knowledge extraction, converting data records into knowledge triples according to the data table structure and field semantics. For unstructured data, a sequence labeling model is used for entity extraction, and a relationship classification model is used for relationship extraction. The rule extraction results and model extraction results are aggregated. For cases where the same knowledge triple is extracted by multiple models, a weighted voting mechanism is used to calculate a comprehensive confidence score, with weights configured based on the performance of each model on the validation set. The extraction results form a set of candidate knowledge triples, with each triple accompanied by a confidence score.

[0076] S202, Multi-level Knowledge Fusion; Based on a set of candidate knowledge triples, a multi-level knowledge fusion method is used to obtain fused knowledge instance data. Specifically, knowledge fusion is divided into three levels: entity alignment, attribute merging, and conflict resolution. At the entity alignment level, for entity mentions from different data sources, a combination of feature similarity calculation and contextual semantic analysis is used to determine whether they point to the same entity. Feature similarity calculation includes dimensions such as name character similarity, attribute value similarity, and overlap of related entities. Contextual semantic analysis compares semantic vectors based on the text context in which the entity appears. For entity mention pairs with similarity higher than a preset alignment threshold, they are determined to be the same entity and are aligned. At the attribute merging level, for aligned entities, attribute values ​​from different data sources are merged. At the conflict resolution level, for cases where the same attribute has different values, a credibility-first strategy is used to resolve conflicts, prioritizing attribute values ​​from high-credibility data sources. For cases with the same credibility, a timeliness-first strategy is used, retaining attribute values ​​with more recent updates. The fusion result forms the fused knowledge instance data.

[0077] S203, Ontology Evolution Trigger Monitoring Based on Semantic Drift Detection; Based on the fused knowledge instance data and instance semantic vector representation, an ontology evolution trigger monitoring mechanism based on semantic drift detection is adopted to obtain ontology evolution trigger signals. Specifically, a monitoring mechanism for multi-dimensional ontology evolution trigger conditions is established.

[0078] The first type of trigger condition is the detection of emerging new concepts. This involves using clustering analysis to semantically cluster entities that cannot be categorized into existing concept types, calculating the cohesion and separation degree of each cluster from existing concepts. Cohesion is measured by the mean cosine similarity of the semantic vectors of instances within the cluster, while separation is measured by the cosine distance between the cluster center vector and the prototype vector of the existing concept. When the cohesion of a cluster exceeds a preset cohesion threshold and the minimum separation degree from all existing concepts exceeds a preset separation threshold, it is determined to be an emerging new concept, triggering evolutionary analysis for adding the new concept type.

[0079] The second type of trigger condition is relation pattern discovery. Frequent subgraph mining methods are used to discover high-frequency entity connection patterns in the knowledge graph that are not covered by existing relation types. For connection patterns with support exceeding a preset relation threshold, their path features and endpoint type constraints are extracted, and they are identified as new relation patterns, triggering an evolutionary analysis that adds new relation types.

[0080] The third type of triggering condition is concept semantic drift detection, which uses a concept prototype vector tracking method to monitor the evolution trend of concept semantics. A concept prototype vector is defined as the centroid of the semantic vectors of all instances under that concept. The movement trajectory of the concept prototype vector is calculated within a sliding time window. When the cumulative movement distance of the prototype vector within a continuous time window exceeds a preset drift threshold, it is determined that concept semantics has drifted. Further analysis of the drift direction reveals that if the prototype vector moves towards a certain subspace, it triggers an evolutionary analysis of concept refinement; if the prototype vectors of multiple concepts approach each other, it triggers an evolutionary analysis of concept generalization or merging.

[0081] The fourth type of trigger condition is ontology structure inconsistency detection. This involves using ontology constraint verification methods to check whether knowledge instances violate the constraint rules defined in the ontology, and statistically analyzing the frequency and distribution of violations of various constraints. When the frequency of violations of a certain type of constraint exceeds a preset tolerance threshold, it is determined that the ontology structure does not match the actual data, triggering an evolutionary analysis of ontology constraint adjustments.

[0082] The monitoring mechanism operates continuously using an incremental calculation method, performing a monitoring cycle each time a new batch of knowledge instances is fused. When any trigger condition is met, an ontology evolution trigger signal is generated, which includes information such as the trigger type, the concepts or relationships involved, the degree of quantification deviation, and relevant instance samples.

[0083] S204, Ontology Evolution Scheme Evaluation: Based on ontology evolution trigger signals, an ontology evolution rule base is used to evaluate evolution schemes and obtain the optimal ontology evolution scheme. Specifically, the ontology evolution rule base contains four types of evolution rules: concept addition rules define the conditions and steps for adding new concept types; concept generalization rules define the conditions and steps for merging multiple similar concepts into a more abstract concept; relation extension rules define the conditions and steps for adding new relation types or extending the domain of existing relations; and hierarchy adjustment rules define the conditions and steps for adjusting the concept inheritance hierarchy. For each trigger signal, applicable evolution rules are matched from the rule base to generate a set of candidate evolution schemes. For each candidate scheme, an impact scope analysis is performed to calculate the scheme's compatibility impact on existing knowledge instances, its effectiveness impact on existing inference rules, and its compatibility impact on existing application programming interfaces. Combining the impact scope analysis results and evolution benefit evaluation, the scheme with controllable impact and the greatest benefit is selected as the optimal ontology evolution scheme.

[0084] S205, Ontology Evolution Execution and Consistency Maintenance: Based on the optimal ontology evolution scheme, a consistency maintenance mechanism is used to execute ontology evolution operations, resulting in an updated domain knowledge graph. Specifically, the ontology evolution execution process includes three stages. In the preparation stage, an ontology evolution transaction is created, and the current ontology version and affected knowledge instances are backed up. In the execution stage, corresponding operations are performed according to the evolution scheme type: for new concepts, a new concept definition and its attribute specification are added to the ontology model, and the entity instances triggering the evolution are migrated to the new concept type; for relation expansion, relation definitions are added or modified in the ontology model, and the domain and value constraints of the relations are updated; for hierarchical adjustments, the inheritance relationships of concepts are reorganized, and the attribute inheritance of affected concepts is updated. In the consistency maintenance stage, all affected knowledge instances are checked to ensure they conform to the new ontology constraints; instances that do not conform are automatically corrected or marked as pending. Simultaneously, inference rules dependent on the ontology structure are updated to ensure consistency with the new ontology. After evolution is complete, the transaction is committed, and the ontology version evolution history is recorded. The evolution result forms the updated domain knowledge graph.

[0085] S300 extracts symbolic rules and learns vector representations based on the domain knowledge graph, and uses a bidirectional attention fusion mechanism to obtain inference results and explanation links;

[0086] Based on the domain knowledge graph and reasoning query request output in step S200, reasoning is performed using symbolic rule extraction and vector representation learning methods respectively. A bidirectional attention fusion mechanism is then used to integrate the two reasoning results, yielding a reasoning result and explanation chain that combines accuracy and interpretability. Specifically, this includes the following steps:

[0087] S301, Symbolic Reasoning Rule Extraction: Based on the structural features and instance data of the domain knowledge graph, inductive logic programming is used to extract symbolic reasoning rules, resulting in a symbolic rule library. Specifically, firstly, path pattern mining is performed on the knowledge graph, statistically analyzing the various types of paths from the source entity to the target entity and their frequency of occurrence. For path patterns whose frequency exceeds a preset support threshold, they are generalized into rule forms. The rule form uses logical implication expressions, where the antecedent of the rule is the sequence of relations traversed in the path and the intermediate entity type constraints, and the consequent is the reasoning target relation. For each rule, its confidence index is calculated, defined as the ratio of the number of positive examples of the rule to the number of matching antecedents. For multiple rules sharing the same consequent, their confidence and coverage are compared, and rules with high confidence and low redundancy are retained. The rule extraction process considers both forward and backward rules. Forward rules are used to deduce new facts from known facts, while backward rules are used to backtrack the required conditions from the target facts. The extraction results form a symbolic rule library, with each rule accompanied by support and confidence evaluation indicators.

[0088] S302, Symbolic Rule-Constrained Vector Representation Learning: Based on a domain knowledge graph and a symbolic rule base, a rule-constrained vector representation learning method is used to obtain entity vectors and relation vectors. Specifically, firstly, low-dimensional dense vector representations of all entities and relations are initialized. Then, an objective function for representation learning is constructed, which consists of three parts: the first part is a basic score function, used to evaluate the reasonableness score of knowledge triples in the vector space, implemented using geometric operations on entity vectors and relation vectors; the second part is a negative sampling loss, which guides the model to learn the ability to distinguish between correct and incorrect knowledge by comparing correct triples with randomly constructed incorrect triples; the third part is the rule constraint term, which encodes symbolic rules as soft constraints in the vector space. The rule constraint term is constructed as follows: for each rule, the entity pair that satisfies the rule's antecedent should have a higher score in the vector space than the entity pair that does not satisfy the rule's antecedent. By jointly optimizing the three parts of the objective function, entity vectors and relation vectors incorporating rule knowledge are obtained. The vector representation results are stored in a vector index, supporting subsequent vector similarity inference.

[0089] S303, Parallel Inference Execution: Based on a symbolic rule base, vector representation, and inference query requests, a parallel inference execution method is employed to obtain symbolic inference result sets and vector inference result sets. Specifically, the inference query request is first parsed to extract the type of inference target and constraints. Inference target types include entity prediction, relation prediction, path discovery, etc. Then, the symbolic inference engine and vector inference engine are launched in parallel. The symbolic inference engine selects an applicable subset of rules according to the inference target type, executes a forward chain inference mechanism to apply rules step by step from known facts until the target is derived, and simultaneously executes a backward chain inference mechanism to search backward from the target for inference paths that satisfy the conditions. The symbolic inference result set contains all entities or relations that satisfy the inference target, as well as the inference path and rule sequence applied to derive each result. The vector inference engine performs similarity calculation and link prediction in the vector space according to the inference target. For the entity prediction task, it calculates the vector distance between the candidate entity and the target location; for the relation prediction task, it calculates the score of the candidate relation. The vector inference result set contains candidate results sorted by score and their probability distribution.

[0090] S304, a confidence propagation bidirectional attention fusion mechanism based on graph neural networks; based on symbolic reasoning result sets and vector reasoning result sets, a confidence propagation bidirectional attention fusion mechanism enhanced by graph neural networks is used to obtain fused reasoning results. Specifically, a heterogeneous two-layer reasoning graph structure of symbols and vectors is constructed. The upper layer is a symbolic reasoning graph containing rule nodes and reasoning path nodes, and the lower layer is a vector reasoning graph containing entity vector nodes and relation vector nodes. The two layers are connected across layers through semantic alignment edges. This heterogeneous graph structure supports bidirectional information flow and mutual enhancement between the symbolic space and the vector space.

[0091] In the attention propagation direction from symbols to vectors, the structural confidence of the symbolic reasoning path is first calculated. The calculation of structural confidence considers both path quality and path complexity, and is defined as the product of the geometric mean of the rule confidence at each step of the path and the path length penalty factor. The path length penalty factor uses an exponential decay form to avoid excessively high weights for excessively long paths. Furthermore, a path diversity reward mechanism is introduced. For results that can be derived through multiple independent paths, their structural confidence receives an additional diversity bonus, the magnitude of which is proportional to the logarithm of the number of independent paths. The structural confidence is injected as a prior distribution into the vector reasoning space. For candidate results covered by symbolic reasoning, their vector reasoning scores are weighted and fused with the structural confidence. The fusion weights employ a meta-learning adaptive adjustment mechanism. This mechanism maintains a meta-learner network. The input to the meta-learner is the feature representation of the current reasoning task and the historical performance statistics of symbolic reasoning and vector reasoning, respectively. The output is the optimal fusion weight configuration for the current task. Through training on multiple historical reasoning tasks, the meta-learner learns the relative advantage patterns of symbolic reasoning and vector reasoning under different task features, thereby achieving intelligent dynamic adjustment of the fusion weights.

[0092] In the vector-to-symbol attention propagation direction, semantic consistency score and context relevance score are calculated for vector inference. The semantic consistency score measures the cosine similarity between the vector representation of the candidate result and the inference context vector, while the context relevance score measures the degree of association between the candidate result and the entities involved in the inference query. For results with multiple symbol inference paths, the semantic consistency score and context relevance score are used as soft constraints for path selection, and the overall credibility of each path is calculated. The overall credibility is defined as the weighted harmonic mean of structural confidence, the average semantic consistency score of the entities involved in the path, and the average context relevance score. The weights are adaptively configured based on the variance of each score; scores with smaller variances receive higher weights because they are considered more stable and reliable.

[0093] Attention information in both directions is iteratively propagated through a message passing algorithm based on a graph neural network. Specifically, a graph attention network architecture is adopted. In each iteration, nodes in the upper-layer symbolic graph aggregate information from their neighbors and update their own representations through the attention mechanism. The attention weights are dynamically calculated based on the semantic relevance and structural association between nodes. Nodes in the lower-layer vector graph also aggregate information and update their representations through the graph attention mechanism. Cross-layer information propagation is achieved through edge alignment. The confidence information of symbolic graph nodes propagates to the corresponding vector graph nodes, and the semantic information of vector graph nodes is fed back to the corresponding symbolic graph nodes. The propagation process uses a residual connection mechanism to preserve the original information of nodes in each layer, avoiding excessive smoothing of information in multiple rounds of propagation. At the same time, a gating mechanism is introduced to control the intensity of cross-layer information flow. The gating parameters are adaptively adjusted according to the uncertainty level of the current node, and nodes with higher uncertainty receive more information supplementation from another layer. Iteration is executed for a preset number of rounds or terminated when the representation change of all nodes is lower than the convergence threshold. The fusion inference result and its comprehensive confidence ranking are output. The fusion mechanism achieves a deep integration of the logical rigor of symbolic reasoning and the semantic generalization ability of vector reasoning through the representation learning capability of graph neural networks.

[0094] S305, Reasoning Explanation Link Generation: Based on the fused reasoning results, a reverse tracing algorithm is used to generate reasoning explanation links, resulting in an interpretable description of the reasoning process. Specifically, for each fused reasoning result, it is first determined whether its primary source is symbolic reasoning or vector reasoning. For results primarily from symbolic reasoning, its reasoning path is directly extracted as the skeleton of the explanation link, with each step in the path labeled with the applied rule name, rule confidence, and involved intermediate entities. For results primarily from vector reasoning, attention analysis is used to identify the knowledge triples that contribute most to the result, and these triples are organized into a path-like explanation structure. For results that contribute to both types of reasoning, the symbolic reasoning path and vector reasoning contribution information are integrated to form a hybrid explanation link.

[0095] Each explanatory link is labeled with its step-by-step confidence level, information source, and reasoning type. The explanatory links are output in two forms: structured data and visual graphics. The structured data is used for program calls, and the visual graphics are used for decision-makers' understanding.

[0096] S400 performs temporal correlation analysis based on the domain knowledge graph and historical decision events, and uses a time decay attention mechanism to obtain a decision time sequence graph.

[0097] Based on the domain knowledge graph and enterprise historical decision event records output in step S200, a temporal relationship graph structure between decision events is constructed using temporal association analysis. A time decay attention mechanism is used to extract historical context relevant to the current decision, resulting in a decision time sequence graph to assist in the current decision-making process. Specifically, the process includes the following steps:

[0098] S401, Structured Extraction of Decision Events: Based on historical decision data of the enterprise, a set of decision event records is obtained using structured information extraction methods. Specifically, the sources of historical decision data include management meeting minutes, project decision documents, supply chain adjustment records, risk response reports, etc. For each historical decision record, the following structured information is extracted: decision timestamp, decision type, decision subject, description of the decision problem, list of considerations, measures taken, and evaluation of the decision result. The extraction process comprehensively utilizes template matching and text understanding methods. For unstructured decision documents, paragraph segmentation and semantic role labeling are first performed, and then targeted extraction is carried out based on the feature patterns of each information type. The extraction results form a set of decision event records.

[0099] S402, Decision Sequence Relationship Analysis; Based on a set of decision event records, a multi-dimensional relationship analysis method is used to obtain a decision sequence relationship graph structure. Specifically, four types of sequence relationships between decision events are analyzed. The chronological order relationship is determined by the decision timestamps, establishing chronological relationship edges on the timeline. The causal influence relationship analysis examines the degree of influence of the preceding decision on the subsequent decision. This involves checking whether the problem description of the subsequent decision references the results of the preceding decision, analyzing whether there is a logical connection between the execution effect of the preceding decision and the problem background of the subsequent decision, and assessing whether there are upstream and downstream supply chain relationships between the business objects involved in the preceding and subsequent decisions. The goal inheritance relationship identifies decision sequences with the same or related decision goals. The resource competition relationship identifies parallel decisions competing for the same resources. The results of the four types of relationship analysis are constructed into a graph structure, where nodes represent decision events, edges represent relationships between events, and the type and weight of the edges reflect the category and strength of the relationship.

[0100] S403, Decision-Making and Knowledge Graph Association: Based on the decision-making timeline graph and the domain knowledge graph, entity linking and relation mapping methods are used to obtain the associated decision-making timeline graph. Specifically, each decision node in the decision-making timeline graph is associated with relevant knowledge in the domain knowledge graph. The association is divided into three levels: the first level is business object association, linking business objects such as suppliers, products, orders, and logistics nodes involved in the decision to corresponding entity nodes in the knowledge graph; the second level is constraint association, linking the constraints considered in the decision to relevant attribute values ​​or rule nodes in the knowledge graph; the third level is target indicator association, linking the target indicators pursued by the decision to corresponding metric concepts in the knowledge graph. The association process employs entity disambiguation and link verification mechanisms to ensure the accuracy of the links. The association results enable the understanding and analysis of decision events to be enhanced through the reasoning capabilities of the knowledge graph.

[0101] S404, Context-Aware Adaptive Time Decay Attention Mechanism Design: Based on the timestamp information and decision context features of the decision time series graph, a context-aware adaptive time decay attention mechanism design method is adopted to obtain a dynamically adjusted time decay attention function. Specifically, a multi-factor time decay model is constructed, which not only considers the time interval but also comprehensively considers the stability of the decision environment and the persistence of the decision effect. First, a basic time decay function is defined, adopting a two-parameter Weibull distribution. The shape parameter controls the concavity and convexity of the decay curve, and the scale parameter controls the decay rate. Compared with traditional exponential decay, it can more flexibly characterize the decay characteristics of different stages. Then, an environmental stability adjustment factor is introduced. The environmental stability is quantified by analyzing the rate of change of environmental feature vectors within adjacent time windows in the decision time series graph. The environmental feature vectors include dimensions such as market volatility indicators, supply chain disruption frequency, and policy change frequency. When the environmental stability is high, the reference value of historical decisions decays slowly, and the adjustment factor takes a larger value; when the environmental volatility is severe, the timeliness of historical decisions decreases, and the adjustment factor takes a smaller value. Furthermore, a decision-effect persistence adjustment factor is introduced. Based on historical decision-effect tracking data recorded in the knowledge graph, the average effective duration of various decision measures is calculated. Decision types with longer effective durations receive larger persistence adjustment factors. The comprehensive time decay function is defined as the product of the base decay value and the environmental stability adjustment factor and the effect persistence adjustment factor. In addition, a decision association strength enhancement mechanism is introduced. For historical decisions that have a causal relationship or goal inheritance relationship with the current decision, an association strength weighting is added on top of the time decay. The association strength is quantified by the edge weights in the decision time sequence graph. The resulting adaptive time decay attention function can dynamically adjust the decay mode according to the specific decision scenario and historical decision characteristics, exhibiting stronger scenario adaptability compared to a decay function with fixed parameters.

[0102] S405, Historical Decision Context Extraction: Based on the decision time sequence graph, time decay attention function, and current decision scenario features, a weighted historical decision context is obtained using relevance calculation and context aggregation methods. Specifically, firstly, the feature vector of the current decision scenario is parsed, with feature dimensions including decision type, type of business object involved, type of problem faced, and type of constraint. Then, historical decision nodes related to the current decision features are retrieved from the decision time sequence graph. The relevance calculation comprehensively considers three aspects: similarity of decision type, overlap of involved entities, and similarity of problem description. The relevance is multiplied by the time decay weight to obtain a comprehensive weight, and historical decisions are sorted according to the comprehensive weight. Historical decisions with a comprehensive weight higher than a preset relevance threshold are selected, and their measures taken and result evaluation information are extracted and weighted to form the historical decision context. The context information includes a list of measures related to the historical decisions, effect evaluation statistics, summary of successful experiences, and warnings of lessons learned from failures.

[0103] S500 acquires domain knowledge graphs and decision sequence diagrams, and uses health status prediction methods and network topology vulnerability analysis methods to assess supply chain resilience and obtain resilience enhancement solutions.

[0104] Based on the domain knowledge graph output in step S200, the decision sequence diagram output in step S400, and the real-time status data of supply chain nodes, the resilience of the supply chain is assessed using health status prediction methods and network topology vulnerability analysis methods. Preventive resilience enhancement schemes are generated using case-based reasoning methods, resulting in a resilience assessment report and a set of schemes for decision support. Specifically, the process includes the following steps:

[0105] S501, Construction of a Supply Chain Node Health Status Indicator System: Based on the definition of supply chain entities and attributes in the domain knowledge graph, a multi-dimensional indicator design method is adopted to obtain a supply chain node health status indicator system. Specifically, differentiated health status indicators are designed for different types of nodes in the supply chain, such as supplier nodes, logistics nodes, and warehousing nodes. The health indicator dimensions for supplier nodes include: financial health, capacity utilization, on-time delivery rate, quality stability, and geographical risk exposure. The health indicator dimensions for logistics nodes include capacity adequacy, timeliness stability, and route coverage. Real-time data for each dimension is collected and normalized, converting it into a standardized score in the range of zero to one.

[0106] S502, a node health status prediction method based on multi-source heterogeneous signal fusion, uses historical time-series data of health status indicators, knowledge graph association information, and external early warning signals to obtain the future health status evolution trajectory of nodes and its uncertainty quantification. Specifically, a three-layer fusion prediction architecture is constructed.

[0107] The first layer is a single-indicator time-series prediction layer, which constructs adaptive prediction models for each node and each dimension of health indicators. The prediction model adopts an ensemble learning framework, integrating three types of basic predictors: a trend predictor uses a locally weighted regression method to capture the long-term trend changes of the indicators; a cycle predictor uses Fourier transform and autocorrelation analysis to identify the periodic patterns of the indicators and perform cycle extrapolation; and a mutation predictor uses a change point detection method to identify abnormal jump patterns of the indicators and assess the risk of mutation. The outputs of the three types of predictors are dynamically weighted and integrated, with the weights adaptively adjusted according to the prediction errors of each predictor within the recent sliding window.

[0108] The second layer is a cross-indicator correlation prediction layer, which uses causal relationship knowledge from the knowledge graph to establish a dependency model between health indicators. For example, a deterioration in a supplier's financial health may precede a decline in on-time delivery rate, and an increase in geographical risk exposure may lead to fluctuations in capacity utilization. A Bayesian network structure is constructed to represent the causal dependencies between indicators. The network structure is automatically discovered from historical data through a structure learning algorithm, and the network parameters are learned using the maximum likelihood estimation method. During prediction, the single-indicator prediction results from the first layer are used as observational evidence and input into the Bayesian network. The joint prediction distribution of all indicators is updated through probabilistic inference propagation, achieving mutual correction between indicators and uncertainty propagation.

[0109] The third layer is the external signal enhancement layer, integrating early warning signals from external data sources of the knowledge graph, including industry risk warnings, geopolitical events, natural disaster forecasts, and policy change notifications. For each type of external signal, an impact mapping model from the signal to health indicators is established. This model is learned from historical event cases and describes the direction, magnitude, and time lag of the impact of specific types of external events on each health indicator. When a relevant external signal is detected, the predicted distribution of the second layer's output is adjusted according to the impact mapping model, and the expected impact of the external shock is superimposed. The final output of the three-layer fusion prediction is the predicted trajectory and confidence interval of each node's health indicators for each dimension within the future time window. The confidence interval is estimated from the predicted distribution using the Monte Carlo sampling method. The prediction results of each dimension are synthesized into the evolution trajectory of the node's comprehensive health status score using a hierarchical weighting method. The weight configuration considers the contribution of each dimension to comprehensive health and the focus of the current decision-making scenario. For nodes whose overall health status score shows a downward trend and whose predicted trajectory lower bound is below the preset health threshold, an early warning mechanism is triggered. The early warning level is determined based on the predicted time to reach the threshold, the rate of decline, and the prediction confidence level. It is divided into three levels: Level 1, Level 2, and Level 3. Level 1 early warning indicates a high confidence level of near-term risk, requiring immediate countermeasures.

[0110] S503, Supply Chain Network Cascade Failure Prediction Based on Dynamic Load Redistribution; Based on the supply chain network structure and node capacity constraints in the domain knowledge graph, a dynamic load redistribution model is used to predict cascade failures, resulting in network topology vulnerability analysis. Specifically, the supply chain knowledge graph is abstracted into a weighted directed network structure, where nodes represent supply chain entities with capacity and load attributes, and edges represent supply relationships or logistics connections with flow and capacity attributes.

[0111] A node capacity model is constructed, where node capacity is defined as the maximum workload a node can handle under normal operating conditions, including the maximum capacity of suppliers, the maximum throughput of logistics nodes, and the maximum storage capacity of warehousing nodes. The initial load of a node is defined as the actual workload it currently undertakes, obtained through real-time operational data from the knowledge graph. Node capacity margin is defined as the difference between capacity and load, reflecting the node's resilience to shocks.

[0112] A dynamic load redistribution model is then constructed to simulate the cascading propagation process after a node failure. When a node fails due to deteriorating health, the load it originally handled needs to be redistributed to other nodes in the network. Load redistribution follows the shortest path first principle and capacity constraint principle. The upstream node of the failed node reroutes its output traffic to an available alternative path, and nodes on the alternative path take on the additional load. For nodes taking on additional load, their load is checked to see if it exceeds a capacity threshold, which is set as a preset proportion of the node's capacity, typically between 0.8 and 0.9. If the node's load exceeds the capacity threshold, the node enters an overload state. Overloaded nodes have a higher probability of failure, and the failure probability is positively correlated with the degree of overload. Overloaded nodes may fail in the next time step, triggering a new round of load redistribution, forming cascading propagation. The cascading propagation process is simulated iteratively using discrete time steps. At each time step, the load and failure states of all nodes are updated until the network reaches a new stable state or a large-scale collapse occurs. Based on cascading failure simulation, network vulnerability indices are calculated: the cascading scale index measures the total number of nodes that ultimately fail due to the failure of a single node; the cascading depth index measures the maximum number of time steps in cascading propagation; the set of key triggering nodes identifies the initial failure node most likely to trigger large-scale cascading failures; and the set of cascading amplification nodes identifies intermediate nodes that play a key role in the transmission process of cascading propagation.

[0113] Furthermore, the robustness metrics of the computational network were evaluated using two attack modes: random attack and deliberate attack. Random attack simulated scenarios where nodes failed randomly, while deliberate attack simulated scenarios where nodes failed sequentially according to their importance. The degradation curves of network performance under the two attack modes were compared. The vulnerability analysis results were visualized as cascading failure propagation diagrams and vulnerability heatmaps, providing accurate target identification for the design of resilience enhancement schemes.

[0114] S504, Comprehensive Supply Chain Resilience Assessment: Based on node health status predictions and network topology vulnerability analysis, a multi-factor comprehensive assessment method is employed to obtain the comprehensive supply chain resilience assessment results. Specifically, a comprehensive supply chain resilience assessment model is constructed, which comprehensively considers three types of factors: node's own health status, node topological importance, and historical associated risks. These three factors are combined using configurable weights to calculate the comprehensive vulnerability score for each node; a higher score indicates a greater potential risk for that node. All nodes are ranked according to their comprehensive vulnerability scores, identifying a set of high-risk nodes requiring priority attention. The assessment results are visualized as a risk heatmap, colored according to vulnerability scores on the supply chain network topology map, allowing decision-makers to intuitively identify areas of concentrated risk.

[0115] S505 is an intelligent generation of resilience enhancement schemes based on adversarial generative networks. Based on the high-risk node identification results and historical enhancement cases in the knowledge graph, an innovative resilience enhancement scheme is obtained by combining adversarial generative networks with case reasoning.

[0116] Specifically, a resilient generative adversarial network framework is constructed, which includes two core components: a generator network and a discriminator network. The generator network takes as input the feature vectors and constraint vectors of high-risk nodes. The feature vectors encode information such as the node's type, health status indicators, and topological location, while the constraint vectors encode restrictions such as available resource budgets, implementation time windows, and business continuity requirements.

[0117] The generator network learns the mapping relationship from input features to combinations of solutions through a multi-layer neural network structure, outputting a structured representation of candidate resilience-enhancing solutions, including a sequence of solution types, parameter configurations, and implementation timing. The discriminator network evaluates the feasibility and effectiveness of the generated solutions. Its training data comes from historical successful solutions recorded in knowledge graphs and decision-making time-series graphs as positive samples, and randomly constructed unreasonable solutions as negative samples. The discriminator network learns feature patterns that distinguish between high-quality and low-quality solutions and scores the candidate solutions output by the generator. The generator and discriminator are jointly optimized through an adversarial training mechanism. The generator aims to generate high-quality solutions that can deceive the discriminator, while the discriminator aims to accurately identify solution quality. The adversarial training process allows the generator to gradually learn deep patterns from historical successful solutions, while also developing the ability to generate innovative solutions that surpass historical examples.

[0118] After generating initial candidate solutions using the Generative Adversarial Network (GAN), a case-based reasoning approach is employed to refine and validate these solutions. Successful historical cases addressing similar risk nodes are retrieved from the knowledge graph and decision sequence graph, with search criteria including node type matching, similar risk type, and similar business background. For each retrieved historical case, key information such as the enhancement measures taken, implementation costs, effective time, and final results are extracted to construct a case feature library. The candidate solutions output by the GAN are then matched against the case feature library to identify the set of historical cases most similar to the candidate solutions. Based on the implementation experience of similar historical cases, the feasibility of the candidate solutions is verified and parameters are optimized. Feasibility verification checks whether the candidate solutions violate business rules, such as exceeding resource budget limits or conflicting with existing contract terms. Parameter optimization fine-tunes the parameters of the candidate solutions based on implementation data from historical cases; for example, strategic inventory levels are adjusted based on historical inventory turnover data, and the startup time of alternative suppliers is adjusted based on historical supplier switching cycles.

[0119] For each high-risk node, multiple candidate enhancement solutions are generated, with the solution types differentiated based on the risk type and node characteristics. For supplier financial risk, alternative supplier reserve solutions, supplier support plans, or supplier equity investment solutions are generated; for capacity bottleneck risk, capacity expansion support solutions, demand diversification solutions, or process optimization solutions are generated; for logistics route risk, alternative route contingency plans, multimodal transport solutions, or strategic inventory pre-positioning solutions are generated; for geographical concentration risk, supplier geographical dispersion solutions or regional supply chain localization solutions are generated. The generated candidate solution set includes both conservative solutions based on historical experience and innovative solutions based on the generated network, providing decision-makers with diverse choices.

[0120] S506, Solution Simulation and Evaluation: Based on a set of candidate resilience enhancement solutions, simulation and multi-dimensional evaluation methods are used to obtain solution evaluation results and priority ranking. Specifically, simulations are performed on each candidate solution to evaluate its performance under assumed risk scenarios. The simulation inputs include a snapshot of the current supply chain state, assumed node failure scenarios, and the specific measures of the candidate solutions. The simulation process simulates changes in the supply chain state after solution implementation and calculates the effectiveness indicators of the solutions in mitigating risks. Based on the simulations, a multi-dimensional solution evaluation is conducted, including: implementation cost, implementation cycle, resource requirements, operational impact, and risk mitigation effect. Based on the multi-dimensional evaluation results, a weighted comprehensive method is used to calculate the overall priority score of the solutions, and the solutions are ranked according to priority.

[0121] S600 performs multi-objective optimization based on the reasoning results and resilience enhancement schemes, and ranks them to obtain recommended decision schemes;

[0122] Based on the resilience enhancement scheme set and decision objective weight configuration output in step S500, the inference results and explanation chain output in step S300, and the decision sequence diagram output in step S400, a multi-objective decision scheme generation method and a comprehensive evaluation method are used to obtain an optimized and ranked set of recommended decision schemes for decision support; specifically, the following steps are included:

[0123] S601, Determining the Multidimensional Decision-Making Objective System: Based on the characteristics of the current decision-making scenario and business needs, a multidimensional objective system for the current decision is obtained using objective matching and weight configuration methods. Specifically, a decision-making objective template library is established, pre-defining a set of objective dimensions applicable to various decision-making scenarios. Objective dimension types include: cost-related objectives, risk-related objectives, benefit-related objectives, and time-sensitive objectives. Based on the type of the current decision-making scenario, the corresponding combination of objective dimensions is matched from the template library. Based on the matched objective dimension combinations, decision-makers can make adjustments according to the actual situation, including adding or removing objective dimensions and adjusting the relative importance weights of each dimension. Weight configuration supports both direct numerical input and pairwise comparison. The determined multidimensional objective system serves as the evaluation standard for subsequent scheme evaluation.

[0124] S602, Candidate Decision Solution Generation: Based on the domain knowledge graph, decision sequence diagram, and resilience enhancement solution set, a combination of case retrieval and rule generation is used to obtain a candidate decision solution set. Specifically, candidate solution generation is divided into three approaches. The first approach is historical case retrieval, which directly reuses decision solutions that have achieved good results in similar scenarios in the past based on the historical decision context extracted in step S405, and makes appropriate adjustments according to the differences in the current scenario when necessary. The second approach is rule-driven generation, which automatically derives decision solutions that meet the rules based on the decision rules and best practice knowledge stored in the knowledge graph and the constraints of the current decision scenario. The third approach is resilience solution integration, which incorporates the resilience enhancement solution set output in step S506 into the candidate solutions. The candidate solutions generated by the three approaches are aggregated and deduplicated to form a candidate decision solution set. Each solution includes information such as a description of specific measures, expected resource requirements, and expected implementation cycle.

[0125] S603, a multi-objective quantitative evaluation model, is based on a set of candidate decision-making schemes and a multi-dimensional objective system. It employs a predictive model to perform multi-objective quantitative evaluation, resulting in an evaluation matrix. Specifically, each candidate scheme undergoes quantitative prediction across all objective dimensions. This prediction is based on causal relationship knowledge in a knowledge graph and historical data statistics. For cost-related objectives, the historical cost distribution of similar measures in the knowledge graph is queried based on the types of measures involved in the scheme, combined with current market environment factors for cost prediction. For risk-related objectives, the risk level is predicted based on the scheme's coverage of risk factors and the risk mitigation effects of historical cases. For benefit-related objectives, benefit prediction is based on the causal correlation strength between the scheme's measures and business returns, and historical benefit realization. For time-sensitive objectives, time-sensitive prediction is based on the complexity of the measures involved in the scheme and historical implementation cycle statistics. Each predicted value is accompanied by a confidence interval, reflecting the degree of uncertainty in the prediction. The predicted values ​​of all schemes across all objective dimensions constitute the evaluation matrix.

[0126] S604, Reference Point-Guided Adaptive Pareto Ranking and Recommendation: Based on the alternative evaluation matrix, target weight configuration, and decision-maker preferred reference points, an adaptive Pareto ranking method guided by reference points is adopted to obtain personalized alternative priority rankings. Specifically, a multi-stage ranking optimization framework is constructed. The first stage is Pareto stratification, which performs non-dominated ranking on the alternative evaluation matrix, dividing all alternatives into multiple Pareto levels. The first level is the Pareto front, containing all non-dominated alternatives; the second level contains alternatives dominated by the first level but non-dominated among the remaining alternatives; and so on until all alternatives are assigned to a certain level. Pareto stratification can identify the absolute superiority-inferiority relationship of alternatives, but alternatives at the front cannot be directly compared. The second stage is reference point-guided optimization, introducing a decision-maker preferred reference point mechanism. The reference point is the ideal combination of target values ​​specified by the decision-maker in the target space, representing the target state that the decision-maker most wants to achieve.

[0127] For each scheme on the Pareto front, its distance to the reference point is calculated using a weighted Chebyshev distance metric, where the weights reflect the relative importance of each objective dimension. The weighted Chebyshev distance is defined as the weighted maximum of the normalized differences between the scheme's value and the reference point's value for each objective dimension. This distance metric balances the optimization level of each objective dimension, preventing excessive sacrifice of any one objective. Schemes on the Pareto front are sorted from smallest to largest distance to the reference point, with the scheme having the smallest distance closest to the decision-maker's ideal preference. The third stage is robustness adjustment, considering the impact of uncertainty in scheme evaluation on the ranking results. Since scheme evaluation is based on a prediction model, the predicted values ​​for each objective dimension have confidence intervals. A Monte Carlo simulation method is used to randomly sample within the confidence intervals of each scheme and each objective, generating multiple possible evaluation matrices. For each evaluation matrix, the Pareto ranking and reference point distance calculation are repeatedly performed, and the average ranking and ranking variance of each scheme across multiple simulations are calculated. Schemes with smaller ranking variances exhibit higher robustness, and their evaluation is less sensitive to evaluation uncertainty.

[0128] In the final ranking, considering both the distance to the reference point and robustness indicators, a multi-attribute decision-making method is used to calculate the comprehensive priority score of each solution. The priority score is defined as the weighted sum of the inverse of the distance to the reference point and the robustness weight. The fourth stage is interactive recommendation. Instead of outputting a single optimal solution, the system generates a diverse set of recommended solutions. This set includes three categories: the optimal balance solution (the one with the highest comprehensive priority score); the extreme optimization solution (the one that performs best on a single objective dimension, used to demonstrate the optimization limits of each objective); and the compromise alternative solution (representative solutions evenly distributed on the Pareto front, used to demonstrate the effects of different trade-off strategies). The recommendation results are presented in an interactive visualization interface. Decision-makers can adjust the reference point position and objective weights, and the system updates the recommendation ranking in real time, supporting decision-makers in exploring optimal solutions under different preference settings. Furthermore, the system provides a solution comparison analysis function. Decision-makers can select multiple solutions for side-by-side comparison. The comparison view displays the performance differences of each solution on each objective dimension in the form of radar charts and parallel coordinate graphs, helping decision-makers understand the trade-offs between solutions.

[0129] S605, Causal Interpretability Analysis and Presentation: Based on the recommended solutions and their evaluation results, causal tracing and visualization methods are used to present the causal interpretation of the solutions. Specifically, causal interpretability analysis is performed on each recommended solution, and the analysis includes three levels. The first level is measure causal analysis, explaining how each measure in the solution acts on the target dimension, tracing the impact transmission path from measures to the target based on the causal relationship links in the knowledge graph. The second level is hypothesis analysis, explaining the key assumptions in the solution evaluation, and explaining which changes in external factors may affect the solution's effectiveness. The third level is sensitivity analysis, analyzing the sensitivity of the solution evaluation results to each input parameter, and identifying the key variables that have the greatest impact on the solution's effectiveness. The causal analysis results are visualized in the form of a causal graph, where nodes represent measures, intermediate variables, and target results, and edges represent causal relationships and their strength.

[0130] S700 performs knowledge verification through the execution feedback of recommended decision-making schemes and obtains the optimized knowledge graph using an incremental update method;

[0131] Based on the recommended decision-making scheme execution record and user feedback information output in step S600, the domain knowledge graph is continuously verified and updated using knowledge quality detection methods and feedback-driven optimization methods to obtain an optimized knowledge graph for supporting subsequent decision-making cycles; specifically, the following steps are included:

[0132] S701, User Feedback Collection Mechanism: Based on the decision-making process, a multi-channel feedback collection method is used to obtain a set of user feedback information. Specifically, two collection channels are established: explicit feedback and implicit feedback. The explicit feedback collection channel sets feedback entry points at key nodes in decision-making execution, allowing decision-makers to proactively submit evaluations of the accuracy of recommended solutions, the comprehensibility of reasoning explanations, and corrections regarding the correctness of related knowledge. Feedback forms support both structured ratings and free text descriptions. The implicit feedback collection channel infers the degree of acceptance of the system output by analyzing the decision-maker's operational behavior. Behavioral indicators include solution adoption rate, explanation expansion rate, solution adjustment rate, and dwell time. Explicit and implicit feedback converge to form a set of user feedback information.

[0133] S702, Knowledge Quality Issue Localization; Based on the user feedback information set, an issue tracing analysis method is used to obtain the knowledge quality issue localization results. Specifically, the collected feedback information is classified and analyzed for issue localization. For knowledge errors pointed out in explicit feedback, the specific knowledge point identifiers involved in the feedback are extracted, and the corresponding entities, attributes, or relationships are located in the knowledge graph and marked as pending verification. For improvement suggestions in explicit feedback, the knowledge domains involved in the suggestions are analyzed to identify potentially omitted or outdated knowledge areas. For cases of abnormal implicit feedback indicators, potential knowledge quality issues are traced back. The issue localization results form a knowledge quality issue list, where each record includes the issue type, the scope of knowledge involved, the severity assessment of the issue, and information on the source of the issue.

[0134] S703, Multi-level Knowledge Consistency Verification; Based on a knowledge quality issue list and domain knowledge graph, a multi-level consistency verification method is employed to obtain a consistency verification report. Specifically, three levels of consistency verification are performed. Entity attribute consistency verification checks whether the attribute values ​​of the same entity satisfy predefined constraints; detected violations are marked as attribute conflicts. Cross-source data consistency verification checks whether the attribute values ​​of the same entity from different data sources are consistent; for inconsistencies, the credibility and timeliness of each data source are evaluated to determine the correct value. Knowledge timeliness verification checks whether the knowledge update timestamp exceeds a preset validity period threshold; knowledge exceeding the validity period is marked as pending update. The results of each level of verification are summarized to form a consistency verification report, which details the various consistency issues found and their locations.

[0135] S704, Knowledge Correction and Supplementation Scheme Generation: Based on the knowledge quality issue list and consistency verification report, a scheme planning method is used to obtain knowledge correction and supplementation schemes. Specifically, differentiated handling schemes are generated for different types of knowledge quality issues. For confirmed knowledge errors, a knowledge correction scheme is generated, including the location of the erroneous knowledge, the corrected value, and the source of the correction basis. For identified knowledge omissions, a knowledge supplementation scheme is generated, including the type of missing knowledge, the suggested supplementary content, and the method of obtaining the supplementary information. For discovered outdated knowledge, a knowledge update scheme is generated, including the identification of outdated knowledge, the suggested update process, and the update cycle configuration. For detected knowledge conflicts, a conflict resolution scheme is generated, including the identification of the conflicting parties, the selection of resolution strategies, and the retained value after resolution. All schemes are prioritized according to their scope of impact and urgency, forming a knowledge correction and supplementation scheme queue.

[0136] S705, Incremental Knowledge Update Execution: Based on a knowledge correction and supplementation scheme queue, an incremental update method is used to update the knowledge graph, resulting in an updated knowledge graph. Specifically, knowledge updates are executed incrementally to minimize the impact on online services. First, a knowledge update transaction is created, recording the scope and content of this update. For knowledge correction operations, the target knowledge is located in the knowledge graph and its value is updated, along with the version number and modification timestamp. For knowledge supplementation operations, new entities, attributes, or relationships are added to the knowledge graph, establishing associations with existing knowledge. For knowledge deletion operations, outdated or erroneous knowledge is marked as obsolete rather than physically deleted, preserving historical traceability. After the knowledge instance update is completed, incremental updates of the relevant vector representations are triggered. An incremental learning method is used to update the vector representations of affected entities and relationships, avoiding the overhead of full retraining. Simultaneously, it is checked whether any inference rules are affected by the knowledge update, and the affected rules are re-evaluated for effectiveness and adjusted as necessary. After the update transaction is completed, the changes are committed, and a detailed update log is recorded.

[0137] S706, Knowledge Evolution History Management; Based on knowledge update logs, a version management approach is adopted to obtain a knowledge evolution history record. Specifically, a version management mechanism for the knowledge graph is established to support the tracing and rollback of knowledge changes. Each knowledge update operation generates a change record, which includes information such as change timestamp, change type, changed object, previous and new values, change reason, and change operator. Change records are organized in a timeline format and support retrieval by time range, change type, affected objects, and other conditions. For scenarios requiring rollback, the system supports selecting a historical version for restoration, and the restoration process generates a reverse change record to maintain audit integrity. The knowledge evolution history record serves as the data foundation for continuous monitoring of knowledge quality and is also used to analyze knowledge evolution trends and identify systemic quality issues. The optimized knowledge graph will serve as input for the next round of decision support processes, forming a closed loop of continuous optimization.

[0138] In one embodiment of the present invention, an application example is provided, focusing on the field of supply chain resilience management for multinational manufacturing enterprises, using a real-world application of a global automotive parts manufacturing group as an example. This group has established a complex global supply chain network, encompassing suppliers in multiple countries and regions, multiple logistics and transportation routes, and production bases distributed across different regions. In its daily operations, the group needs to cope with multiple factors such as fluctuations in raw material prices, changes in supplier fulfillment capabilities, geopolitical risks, and the impact of natural disasters, placing high demands on supply chain resilience management.

[0139] The intelligent decision support system of this invention, in its application within the group, achieves comprehensive digital modeling of the supply chain network by constructing an enterprise knowledge graph encompassing multi-dimensional knowledge such as supplier information, product information, logistics information, and historical decision-making information. The system continuously monitors the health status of each supply chain node, analyzes the vulnerability characteristics of the network topology, predicts potential cascading disruption risks, and proactively generates preventative resilience enhancement schemes when the accumulated risks reach a warning threshold, providing interpretable intelligent decision support for decision-makers.

[0140] Table 1 shows an example of the supply chain supplier node health status data collected by the system:

[0141] Table 1. Supplier Node Health Status Data

[0142]

[0143] In Table 1, each health status indicator was collected and standardized based on the multidimensional health status indicator system constructed in step S501. The comprehensive health score was calculated using the weighted comprehensive method in step S502, and the warning level was determined by comparing the comprehensive health score with a preset threshold. The data in the table represents an overview of the health status of different supplier nodes, providing a data foundation for subsequent vulnerability analysis and resilience enhancement scheme generation.

[0144] Based on the comprehensive assessment results of supply chain resilience, the system-generated evaluation data for preventative resilience enhancement schemes is shown in Table 2:

[0145] Table 2. Evaluation data of toughness enhancement schemes

[0146]

[0147] Table 2 shows the evaluation results based on the simulation and multi-dimensional evaluation methods in step S506. Each scheme designed differentiated resilience enhancement measures for different high-risk nodes, and the evaluation dimensions covered implementation costs, implementation cycle, risk mitigation effects, and robustness. The overall priority was calculated based on the Pareto ranking and weighted comprehensive scoring method in step S604, providing decision-makers with a reference for the priority order of scheme selection.

[0148] This application example demonstrates the application of an intelligent decision support method based on enterprise knowledge graphs in a real-world supply chain resilience management scenario. By continuously monitoring and analyzing the health status of supply chain nodes and combining it with network topology vulnerability assessment, the system can proactively identify potential risks and generate preventative resilience enhancement solutions. This achieves a transformation from a traditional reactive response to a proactive prevention decision-making model, effectively improving the resilience management level of the enterprise's supply chain.

[0149] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An intelligent decision support method based on enterprise knowledge graphs, characterized in that, Includes the following steps: S100: Collect raw data from the data access channel and use standardization processing to obtain a cleaned standardized data set; S200 acquires a standardized dataset for knowledge extraction and fusion, and uses an ontology evolution mechanism to obtain an adaptively updated domain knowledge graph. S300 extracts symbolic rules and learns vector representations based on the domain knowledge graph, and uses a bidirectional attention fusion mechanism to obtain inference results and explanation links; S400 performs temporal correlation analysis based on the domain knowledge graph and historical decision events, and uses a time decay attention mechanism to obtain a decision time sequence graph. S500 acquires domain knowledge graphs and decision sequence diagrams, and uses health status prediction methods and network topology vulnerability analysis methods to assess supply chain resilience and obtain resilience enhancement solutions. S600 performs multi-objective optimization based on the reasoning results and resilience enhancement schemes, and ranks them to obtain recommended decision schemes; S700 performs knowledge verification through the execution feedback of recommended decision-making schemes and obtains an optimized knowledge graph using an incremental update method.

2. The intelligent decision support method based on enterprise knowledge graph according to claim 1, characterized in that, The process of collecting raw data through the data access channel and then performing standardized processing to obtain a cleaned, standardized dataset includes: A multi-source data access channel is established based on internal and external data sources. The raw data from each data source is aggregated according to the collection timestamp to form a raw data aggregation pool. Based on the heterogeneous data in the original data aggregation pool, a format conversion rule base and a field mapping table are constructed to obtain normalized data; A data quality anomaly detection model is constructed based on standardized data, integrating statistical detection and machine learning detection. An active learning mechanism is introduced to incrementally update the parameters of the anomaly detection model.

3. The intelligent decision support method based on enterprise knowledge graph according to claim 1, characterized in that, The process of acquiring a standardized data set for knowledge extraction and fusion, and using an ontology evolution mechanism to obtain an adaptively updated domain knowledge graph, includes: For structured data, a rule-matching method is used for knowledge extraction, while for unstructured data, a sequence labeling model is used for entity extraction and a relation classification model is used for relation extraction, forming a set of candidate knowledge triples. Based on the candidate knowledge triple set, a three-level knowledge fusion method of entity alignment, attribute merging and conflict resolution is adopted to obtain fused knowledge instance data. Establish a monitoring mechanism for the triggering conditions of ontology evolution, including new concept emergence detection, relation pattern discovery, concept semantic drift detection, and ontology structure inconsistency detection. When the triggering conditions are met, the ontology evolution rule base is used to evaluate and execute the evolution scheme.

4. The intelligent decision support method based on enterprise knowledge graph according to claim 1, characterized in that, The process of extracting symbolic rules and learning vector representations based on the domain knowledge graph, and using a bidirectional attention fusion mechanism to obtain the reasoning results and explanation links, includes: Path pattern mining is performed based on domain knowledge graphs. Path patterns that occur more frequently than a preset support threshold are generalized into rules to form a symbol rule base. A rule-constrained vector representation learning method is adopted to construct an objective function consisting of a basic score function, a negative sampling loss, and rule constraint terms, thereby obtaining entity vectors and relation vectors; The symbolic reasoning engine and the vector reasoning engine are launched in parallel to obtain the symbolic reasoning result set and the vector reasoning result set, respectively. A heterogeneous two-layer inference graph structure from symbol to vector is constructed, and the two layers are connected across layers through semantic alignment edges. The structural confidence is calculated in the symbol-to-vector direction, and the fusion weights adopt a meta-learning adaptive adjustment mechanism. Semantic consistency score and context relevance score are calculated in the vector-to-symbol direction. Attention information in both directions is iteratively propagated through the message passing algorithm of graph neural network to output the fused inference result.

5. The intelligent decision support method based on enterprise knowledge graph according to claim 1, characterized in that, The step involves performing temporal correlation analysis based on the domain knowledge graph and historical decision events, and using a time decay attention mechanism to obtain a decision time sequence graph, including: Based on the enterprise's historical decision-making data, decision timestamps, decision types, measures taken, and evaluation of decision results are extracted to form a set of decision event records; Analyze the temporal sequence, causal influence, goal inheritance, and resource competition relationships among decision-making events, and construct a decision-making temporal relationship diagram structure; The decision nodes are associated with the domain knowledge graph. The association is divided into three levels: business object association, constraint condition association, and target indicator association. A multi-factor time decay model is constructed, with the basic time decay function adopting a two-parameter Weibull distribution. An environmental stability adjustment factor and a decision effect persistence adjustment factor are introduced. For historical decisions with causal relationships, the correlation strength is superimposed to form a historical decision context.

6. The intelligent decision support method based on enterprise knowledge graph according to claim 1, characterized in that, The process involves acquiring domain knowledge graphs and decision sequence diagrams, employing health status prediction methods and network topology vulnerability analysis methods to assess supply chain resilience, and obtaining resilience enhancement schemes, including: Differentiated health status indicators are designed for each node in the supply chain, and the real-time data of each indicator is collected and normalized. A three-layer fusion prediction architecture is constructed. The first layer is a single-indicator time series prediction layer, the second layer is a cross-indicator correlation prediction layer, a Bayesian network is constructed to represent causal dependencies, and the third layer is an external signal enhancement layer. The supply chain knowledge graph is abstracted into a weighted directed network structure, and a node capacity model and a dynamic load redistribution model are constructed. If the node load exceeds the preset capacity threshold, cascading propagation is triggered. The cascading scale index, cascading depth index, key trigger node set, and cascading amplification node set are calculated.

7. The intelligent decision support method based on enterprise knowledge graph according to claim 1, characterized in that, The process of acquiring domain knowledge graphs and decision sequence diagrams, assessing supply chain resilience using health status prediction methods and network topology vulnerability analysis methods, and obtaining resilience enhancement solutions also includes: The generative adversarial network framework for constructing resilient solutions includes a generator network and a discriminator network. The generator network takes the feature vectors and constraint vectors of high-risk nodes as input and outputs candidate resilience enhancement solutions. The case-based reasoning method is used to retrieve historical successful cases to verify the feasibility of candidate solutions and optimize parameters. Generate differentiated candidate enhancement solutions for different risk types.

8. The intelligent decision support method based on enterprise knowledge graph according to claim 1, characterized in that, The step of performing multi-objective optimization based on the reasoning results and resilience enhancement schemes to rank and obtain recommended decision schemes includes: Establish a decision-making objective template library and match objective dimension combinations according to the type of decision-making scenario; Candidate solution generation is divided into three approaches: historical case retrieval, rule-driven generation, and resilient solution integration, forming a set of candidate decision-making solutions. Quantitative predictions of candidate solutions are used to construct a solution evaluation matrix; A multi-stage ranking optimization framework is constructed. The first stage is to perform non-dominated ranking using Pareto hierarchical structure. The second stage is to calculate the weighted Chebyshev distance using reference point-guided optimization. The third stage is to perform robust adjustment by simulating the ranking variance using Monte Carlo simulation. The fourth stage is to generate a set of recommendation schemes through interactive recommendation.

9. The intelligent decision support method based on enterprise knowledge graph according to claim 1, characterized in that, The process of verifying knowledge through the execution feedback of the recommended decision-making scheme and obtaining the optimized knowledge graph using an incremental update method includes: Establish two channels for collecting feedback: explicit feedback and implicit feedback. The feedback information is categorized and the problem is identified and analyzed to form a list of knowledge quality issues, and a consistency check is performed to generate a check report; To address knowledge quality issues, generate knowledge correction plans, knowledge supplementation plans, knowledge update plans, and conflict resolution plans; Knowledge updates are performed incrementally, and a version management mechanism is established to support retrieval and rollback.

10. An intelligent decision support system based on enterprise knowledge graphs, used to execute the steps of an intelligent decision support method based on enterprise knowledge graphs as described in any one of claims 1-9, characterized in that, include: The data standardization module is used to collect raw data from the data access channel and then perform standardization processing to obtain a cleaned and standardized dataset. The knowledge graph construction module is used to acquire standardized data sets for knowledge extraction and fusion, and adopts an ontology evolution mechanism to obtain an adaptively updated domain knowledge graph. The neural symbolic reasoning module is used to extract symbolic rules and learn vector representations according to the domain knowledge graph, and uses a bidirectional attention fusion mechanism to obtain reasoning results and interpretation links. The temporal analysis module is used to perform temporal correlation analysis based on the domain knowledge graph and historical decision events, and uses a time decay attention mechanism to obtain a decision time sequence graph. The resilience assessment module is used to acquire domain knowledge graphs and decision sequence diagrams, and to assess supply chain resilience using health status prediction methods and network topology vulnerability analysis methods, thereby obtaining resilience enhancement solutions. The decision optimization module is used to perform multi-objective optimization based on the reasoning results and resilience enhancement schemes, and rank them to obtain recommended decision schemes. The knowledge update module is used to verify knowledge through the execution feedback of the recommended decision scheme, and uses an incremental update method to obtain an optimized knowledge graph.

Images