A Robust Construction Method for Knowledge Graphs in Complex Equipment Supply Chains

By constructing entity authenticity feature vectors and using an adversarial discriminant model to screen credible entities, and combining a large language model and a hybrid Transformer-Mamba model for risk detection, the problem of false entity contamination and poisoning risks is solved, and robust construction of a knowledge graph for complex equipment supply chains is achieved.

CN122840194BActive Publication Date: 2026-11-17DALIAN UNIV OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611343266.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-17
Estimated Expiration
2046-09-01

AI Technical Summary

Technical Problem

In the process of constructing existing knowledge graphs for complex equipment supply chains, false entities or abnormal relationships are easily adopted by large language models, erroneous triples directly pollute the formal graph, cross-layer propagation paths are difficult to trace, poisoning risks are difficult to locate in a timely manner, and polluted subgraphs are difficult to roll back locally.

Method used

By acquiring unstructured text and structured knowledge sources in the complex equipment supply chain domain, entity authenticity feature vectors are constructed. An adversarial entity credibility discrimination model is used to discriminate entity credibility. Candidate triples are generated by combining a large language model. After format consistency verification, they are written into a temporary graph. A hybrid Transformer-Mamba model is used to detect posterior poisoning risks. Finally, the graph is rolled back and regenerated before a clean version is available.

Benefits of technology

It effectively reduces the probability of fake entities entering the generation process, tracks and locates the risk of poisoning, and realizes the construction of a clean version of the graph, ensuring the reliability and traceability of the graph.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122840194B_ABST
    Figure CN122840194B_ABST
Patent Text Reader

Abstract

This invention belongs to the fields of knowledge graph construction, supply chain data governance, artificial intelligence security, and information technology in complex equipment manufacturing. It discloses a robust method for constructing a knowledge graph for a complex equipment supply chain. The method involves acquiring unstructured raw text corpus and structured knowledge sources, extracting the context of target entities, and retrieving candidate triples. For each candidate entity, a authenticity feature vector is constructed, and a credibility score and risk label are obtained using an adversarial entity credibility discrimination model, classifying entities as trustworthy, restricted, and intercepted. A prompt text is constructed, and a large language model is used to generate and validate triples. The validated triples are written into a temporary knowledge graph, recording traceability information and establishing derivation relationships, and implementing posterior poisoning risk detection. If the detection passes, a clean version is submitted; if it fails, the affected subgraph is located, and rollback, constraint regeneration, and re-detection are performed. Based on the clean version, tail entity growth and multi-level iterations are conducted to improve the robustness, traceability, and recovery efficiency of the knowledge graph construction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of knowledge graph construction, supply chain data governance, artificial intelligence security, and information technology in complex equipment manufacturing, and particularly to a robust method for constructing a knowledge graph for a complex equipment supply chain. Background Technology

[0002] Complex equipment typically features deep component hierarchies, multiple suppliers, long process chains, high quality traceability requirements, and strong delivery schedule constraints. Taking aero-engines, rail transit equipment, marine power systems, industrial machine tools, or large energy equipment as examples, their supply chain data includes not only structured information such as suppliers, components, raw materials, batches, qualifications, quality inspection records, and delivery plans, but also a large amount of unstructured text such as bidding documents, procurement contracts, quality reports, delivery instructions, supplier announcements, regulatory notices, maintenance records, and technical documents. To support supply chain risk identification, supplier review, quality traceability, and delivery assurance, it is necessary to extract entities, relationships, and attributes from the above multi-source data to construct a knowledge graph of the complex equipment supply chain.

[0003] Existing knowledge graph construction methods typically rely on named entity recognition, relation extraction, rule matching, and the generation of triples using pre-trained language models or large language models. Large language models possess strong text understanding and relation generation capabilities, enabling the generation of supply chain entity relationships in weakly structured, cross-document, and multi-hop association scenarios. However, the data sources for complex equipment supply chains are diverse, and inconsistencies may exist between publicly available information and internal business information. Some supply chain entities may also have issues such as false qualifications, abnormal batches, identifier confusion, impersonation by entities with the same name, mismatched relationships, or conflicting sources. If triples generated by large language models are directly written into the knowledge graph, erroneous entities or poisoned relationships may continuously propagate along the supplier, component, batch, process, quality inspection, and delivery links, thereby polluting subsequent hierarchical graphs.

[0004] Existing methods have the following shortcomings: First, most methods only perform format or rule checks before writing triples, making it difficult to identify source conflicts, time conflicts, qualification anomalies, and poisoning risks in cross-layer derivation paths; second, the generation process lacks fine-grained source tracing records, making it difficult to reverse-locate the source text, prompt text, and parent header entity of risk nodes, risk relationship edges, or risk paths; third, newly added triples are often directly written into the formal graph, lacking temporary graph storage, clean version snapshots, and local rollback mechanisms, making it difficult to repair at low cost once erroneous relationships are extended to multi-layer links; fourth, traditional risk detection models are insufficient for modeling source tracing records for long sequences and cross-layer derivation paths, making it difficult to simultaneously capture local semantic anomalies and long-range propagation dependencies.

[0005] Therefore, a robust method for constructing knowledge graphs for complex equipment supply chains is needed, which can achieve entity credibility determination, triple verification, temporary storage writing, generation tracing, derivation tracking, posterior poisoning risk detection, affected subgraph location, clean version snapshot rollback, and constraint regeneration during the knowledge graph generation process from a large language model. Summary of the Invention

[0006] This invention aims to address the problems existing in the construction of knowledge graphs for complex equipment supply chains, such as the easy adoption of false entities or abnormal relationships by large language models, the direct contamination of the formal graph by erroneous triples, the difficulty in tracing cross-layer propagation paths, the difficulty in timely locating the risk of poisoning, and the difficulty in locally rolling back contaminated subgraphs. It provides a robust construction method for knowledge graphs of complex equipment supply chains.

[0007] The technical solution of this invention is as follows: A robust construction method for a knowledge graph of a complex equipment supply chain, comprising the following steps: Step S1: Obtain unstructured text corpus and publicly available structured knowledge sources in the field of complex equipment supply chain; extract the target entity context from the unstructured text corpus; and retrieve candidate triples related to the target entity from the structured knowledge sources, including relation information, attribute information and source information. Step S2: Based on the textual semantic information and contextual association information of each candidate entity in the candidate triplet within the target entity context, as well as the relational information, attribute information, and source information of the candidate entity in the candidate triplet, construct the corresponding entity authenticity feature vector, including the textual semantic features of the candidate entity, the contextual semantic features of the candidate entity, the source credibility features, the attribute completeness features, the temporal logic consistency features, the cross-source consistency features, the qualification consistency features, and the compatibility features with the existing knowledge graph topology; input the entity authenticity feature vector into the adversarial entity credibility discrimination model to obtain the entity credibility score and entity risk label of the candidate entity; Step S3: Based on the entity credibility score and entity risk label, classify the candidate entities into trusted entities, restricted entities, and blocked entities; Step S4: Perform domain semantic alignment on the candidate triples to obtain a set of domain-related triples; construct structured prompt text based on the target entity context, the set of domain-related triples, trusted entities, restricted entities, and prohibition on generation constraints. Step S5: Input the structured prompt text into the large language model to generate candidate generated triples for the current level. Based on the predefined rule base, perform format consistency, quantity consistency, logical consistency, and entity usage status checks on the candidate generated triples to obtain triple verification labels. When there are candidate generated triples that fail the checks, write the triple verification labels into the prompt text and prompt the large language model to generate verified triples again. Step S6: First, write the verified triplet into the temporary graph of the current level, and generate a traceability record for each newly added node and each newly added relation edge in the temporary graph; establish a derivation relationship between the newly added node, the newly added relation edge and its parent head entity, which is used to track the intra-layer propagation path of risky entities or risky relation edges in the temporary graph of the current level, as well as the cross-layer propagation path between different level graphs formed in subsequent iterations. Step S7: Based on the generated source traceability records and derivation relationships, construct risk propagation sequences for newly added nodes, newly added relation edges, intra-layer propagation paths, and cross-layer propagation paths in the temporary graph. Input the risk propagation sequences into the hybrid Transformer-Mamba posterior poisoning risk detection model to obtain the risk propagation representation and poisoning risk score of each graph element. When the poisoning risk score of each graph element does not exceed the preset risk threshold, submit the temporary graph as the clean version graph of the current level and establish a clean version snapshot. When the poisoning risk score of at least one graph element exceeds the preset risk threshold, locate the affected subgraph in the temporary graph of the current level and its associated cross-layer derivation paths. Step S8: Perform rollback and regeneration processing on the affected subgraph. Rollback and regeneration processing includes: undoing the write operation corresponding to the affected subgraph, restoring the graph state of the affected subgraph to the most recent clean version snapshot before the affected subgraph was generated; adding risky entities in the affected subgraph to the prohibited entity set or restricted entity set, adding risky relation edges in the affected subgraph to the prohibited relation edge set or prohibited relation type set, and writing the triple check label and poisoning risk score corresponding to the affected subgraph into the regeneration constraints; reconstructing the structured prompt text based on the updated regeneration constraints and calling the large language model to generate alternative triples; re-performing triple check, temporary graph writing, generation of source traceability records, establishment of derivation relations, and posterior poisoning risk detection on the alternative triples. Step S9: Only after the temporary map of the current level is submitted as a clean version map, the tail entity growth decision model is used to determine whether the tail entity of the current level should continue to expand as the head entity of the next level, and the tail entity growth threshold is dynamically adjusted according to the entity credibility score, poisoning risk score and triplet verification result of the current level. Step S10: Repeat steps S1 to S9 until all tail entities are marked as terminated, loops are detected, a preset iteration depth is reached, or the same affected branch reaches a preset maximum number of regenerations, and output a multi-layer complex equipment supply chain knowledge graph.

[0008] The target entity context includes textual semantic information and contextual association information, obtained through the following methods: unstructured raw text corpus is segmented into sentences to obtain a sentence set; pre-trained language models are used to encode sentences and target entities respectively, obtaining sentence vectors and target entity vectors; the semantic similarity between sentence vectors and target entity vectors is calculated, and irrelevant sentences are filtered using keywords from the complex equipment supply chain domain; a ranking score is constructed based on semantic similarity and the frequency of target entity occurrences, and the target entity context is obtained by concatenating these scores from high to low; wherein, the semantic similarity and ranking score are respectively: in, Indicates the first Sentences and target entities Semantic similarity between them Indicates the first One sentence. Represents the target entity. Represents a sentence vector. Represents the target entity vector. Represents the target entity In the The number of times a sentence or its adjacent window appears. This represents the index of each sentence in the sentence set or its corresponding adjacent window. Represents the target entity The maximum number of occurrences in each sentence or corresponding adjacent window. To prevent positive numbers with a denominator of zero, This indicates the frequency of occurrence of the target entity after normalization. This represents the weighting coefficient between semantic similarity and the frequency of occurrence of the target entity. This indicates the sorting score.

[0009] The entity authenticity feature vector is formed by concatenating three or more of the following features: textual semantic features of the candidate entity, contextual semantic features of the candidate entity, source credibility features, attribute completeness features, temporal logical consistency features, cross-source consistency features, qualification consistency features, and compatibility features with the existing knowledge graph topology; when a feature is missing, a preset default value or zero value is used to fill it; the entity authenticity feature vector is represented as: in, Indicates candidate entities The entity authenticity feature vector, The textual semantic features representing candidate entities Represents the contextual semantic features of candidate entities. , , , , , These represent the source credibility feature, attribute completeness feature, temporal logic consistency feature, cross-source consistency feature, qualification consistency feature, and compatibility feature with existing knowledge graph topology, respectively.

[0010] The adversarial entity credibility discrimination model includes a generator and a discriminator. During the training phase, the generator generates adversarial entity feature vectors based on the entity authenticity feature vectors of real supply chain entities, random perturbation vectors, and preset risk types. The discriminator is trained based on the entity authenticity feature vectors of real supply chain entities, entity authenticity feature vectors of fake entity samples, and adversarial entity feature vectors. During the inference phase, the discriminator outputs entity credibility scores and entity risk labels based on the entity authenticity feature vectors of candidate entities. The entity credibility score and entity risk label are represented as follows: in, Indicates candidate entities The entity credibility score, Indicates the entity's risk label, This indicates the confidence level output of the discriminator. Indicates the discriminator on the first Classification output of entity risk labels This indicates the total number of categories of entity risk labels; The adversarial training objective of the adversarial entity credibility discrimination model is: in, Represents a generator. Indicates the discriminator, This represents the adversarial training objective function between the generator and the discriminator. This indicates that the generator parameters are updated with the goal of minimizing the adversarial training objective function; This indicates that the discriminator parameters are updated with the goal of maximizing the adversarial training objective function; The entity authenticity feature vector represents the entity's authenticity within the real supply chain. This indicates that the random perturbation vector is sampled from a preset random perturbation distribution. The distribution of entity authenticity feature vectors representing real supply chain entities; This indicates that the preset risk type is sampled from the preset risk type set. Represents the mathematical expectation; Indicates the loss categorized by entity risk label. This indicates the label loss weight; preset risk types include fake suppliers, fake parts, abnormal qualifications, abnormal batches, time conflicts, relationship mismatches, source conflicts, or abnormal identifiers.

[0011] The entity credibility score is output by the discriminator, and the entity segmentation rule is as follows: in, Indicates candidate entities The entity partitioning results Represents a trusted entity. Represents a restricted entity. Indicates the intercepting entity. This indicates a high confidence threshold. This indicates a low confidence threshold, and , This represents a set of high-risk labels. This represents a set of medium-risk labels; Trusted entities participate in generating subsequent triples, restricted entities only participate in generating subsequent triples when the evidence constraints are met, and intercepting entities do not participate in generating subsequent triples and are written into the prohibition of generation constraints. The evidence constraint is that there is no less than a preset number of mutually independent sources of evidence in the target entity context or domain-related triple set, and there are no source conflicts, time conflicts or entity type conflicts.

[0012] The domain semantic alignment includes: converting candidate triples into triple text descriptions, encoding the triple text descriptions using a sentence vector model to obtain triple semantic vectors; generating domain reference vectors based on a complex equipment supply chain domain terminology library; calculating the domain relevance between the triple semantic vectors and the domain reference vectors, and retaining candidate triples with a domain relevance higher than a semantic threshold and that do not contain intercepting entities; wherein, the domain relevance is: in, Indicates the first Domain relevance of candidate triples Indicates the first Triple text description of each candidate triplet Represents a sentence vector model. This represents a domain reference vector obtained from a terminology database for the complex equipment supply chain domain.

[0013] The generated source traceability record includes node source traceability record and edge source traceability record; the node source traceability record includes node source, entity credibility score, entity risk label, hierarchy, parent head entity, and generation round; the edge source traceability record includes relation source, corresponding structured prompt text, corresponding large language model output, rule detection result, parent head entity, and version identifier; the generated source traceability record for any temporarily stored graph element. Represented as: in, This indicates the graph elements temporarily stored in the current level, which are newly added entity nodes or newly added relation edges; Indicates the source text or source triple. This indicates structured prompt text. Indicates the generation round, Indicates the level to which it belongs. Indicates the parent head entity, Represents the entity's credibility score. This indicates an entity risk label or triplet check label. Indicates version identifier; The structured hint text includes trusted entity sections, restricted entity constraint sections, and blocked entity prohibition sections; Derivation relationship is denoted as It is used to characterize the association between the parent head entity and newly added nodes or new relation edges derived from it.

[0014] The hybrid Transformer-Mamba posterior poisoning risk detection model includes a Transformer self-attention encoding layer and a Mamba selective state space encoding layer. The Transformer self-attention encoding layer is used to capture the semantic associations between risk graph elements and their local context, source text, and adjacent relation edges. The Mamba selective state space encoding layer is used to model long-range dependencies in the generation of source records, derivation paths, and cross-layer propagation sequences. The hybrid Transformer-Mamba posterior poisoning risk detection model obtains the poisoning risk score in the following manner: A risk propagation sequence is constructed for each graph element, including the generation traceability record of the graph element, parent head entity, derivation path, entity credibility score, entity risk label, triple check label, hierarchy, and version identifier; the risk propagation sequence is input into the Transformer self-attention encoding layer to obtain a local association representation, and then input into the Mamba selective state space encoding layer to obtain a long-range propagation representation; the local association representation and the long-range propagation representation are fused to obtain the risk propagation representation; the poisoning risk score is calculated based on the risk propagation representation and risk factors; wherein: in, This indicates the target of posterior poisoning risk detection, which can be an entity node, a relation edge, or a cross-layer path; when When it is an entity node, This represents the entity credibility score of the entity corresponding to the entity node; when When it is a relational edge or a cross-layer path, express The lowest entity credibility score or the weighted average entity credibility score of the associated entity; Representing spectral elements The risk transmission sequence, The parameter is Transformer self-attention encoding layer, The parameter is Mamba Selective State Space Coding Layer Represents the pooling function. This represents a local correlation. This indicates long-range propagation characteristics. This represents the risk propagation characterization obtained through fusion. Representing spectral elements Poisoning risk score, Represents the normalized activation function. , , to Indicates model parameters, , , , , These represent the risks associated with abnormal relationship types, abnormal paths, source conflicts, time conflicts, and abnormal attributes, respectively.

[0015] The affected subgraph is a local graph structure determined from the current-level temporary graph and the cross-layer propagation path associated with the current-level temporary graph. This includes: identifying risk nodes, risk relationship edges, or risk paths whose poisoning risk scores exceed a preset risk threshold; tracing the source text, source triples, structured prompt text, and parent header entities of the risk nodes, risk relationship edges, or risk paths based on generated source tracing records; tracing downstream nodes, downstream relationship edges, and downstream cross-layer paths derived from the risk nodes, risk relationship edges, or risk paths based on derivation relationships; and identifying the risk nodes, risk relationship edges, risk paths, and their downstream nodes, downstream relationship edges, and downstream cross-layer paths as the affected subgraph. Affected subgraph and regenerating constraints They are represented as follows: Among them, when When the risk threshold is exceeded, the target of the subsequent poisoning risk detection is identified as a risk map element; This indicates that the poisoning risk score exceeds the preset risk threshold. The set of risk map elements, Indicates the derivation relationship The downstream nodes and downstream relationship edges are obtained by tracing down from the set of elements in the risk graph. Indicates the affected subgraph; This indicates that constraints will be regenerated. This indicates that the collection of entities is prohibited. Represents a restricted set of entities. This represents the set of prohibited edge pairs or the set of prohibited relation types. Represents the set of risk paths. This represents the set of triple check tags. This represents the set of poisoning risk scores corresponding to each risk map element in the affected subgraph. The regeneration constraints are written into the structured prompt text, which enables the large language model to exclude prohibited entities and prohibited relation edges during regeneration and impose evidence requirements on restricted entities. When the same affected branch is rolled back and regenerated multiple times, the constraint strength of the regeneration constraints is increased until the affected branch reaches the preset regeneration limit and is marked as terminated.

[0016] The tail entity growth decision model outputs the probability that the tail entity will grow as the next-level head entity, and determines whether the tail entity should grow based on a dynamic growth threshold; the dynamic growth threshold is: in, Indicates the first The dynamic growth threshold of the layer Indicates the basic growth threshold. Indicates the first Layer average entity credibility score Indicates the first Average poisoning risk score per layer Indicates the first Layer triplet check tag trigger ratio, Indicates the entity credibility adjustment factor, Indicates the poisoning risk adjustment coefficient, This represents the adjustment factor for the triplet check result. and These represent the lower limit and upper limit of the dynamic growth threshold, respectively; when the growth probability of a tail entity is greater than or equal to the dynamic growth threshold, the corresponding tail entity is marked as growing; when the growth probability of a tail entity is less than the dynamic growth threshold, the corresponding tail entity is marked as terminating.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects.

[0018] First, before generating triples using the large language model, this invention uses entity authenticity feature vectors and an adversarial entity credibility discrimination model to classify candidate entities as trustworthy, restricted, and blocked, and writes high-risk entities into the prohibition generation constraint, which can reduce the probability of fake suppliers, abnormal qualifications, abnormal batches, or entities with conflicting sources entering the generation process.

[0019] Second, instead of directly writing the triples generated by the large language model into the formal graph, this invention first writes them into the temporary graph of the current level, and generates source records and derivation relationships for newly added nodes and newly added relation edge records, so that each newly added graph element can be traced to the source text, candidate triples, prompt text, generation round, parent head entity and version identifier.

[0020] Third, before submitting the temporary graph as a clean version graph, the present invention constructs a risk propagation sequence consisting of generation traceability records, derivation relationship paths, entity credibility scores, entity risk labels, triplet verification labels, belonging level and version identifiers, and calculates the poisoning risk score through a hybrid Transformer-Mamba posterior poisoning risk detection model, which can simultaneously capture local semantic associations and cross-layer long-range propagation dependencies. Attached Figure Description

[0021] Figure 1 A flowchart illustrating a robust method for constructing a knowledge graph for a complex equipment supply chain.

[0022] Figure 2This is a schematic diagram for judging the credibility of candidate entities and classifying entities. Detailed Implementation

[0023] In this embodiment, the robust construction method for complex equipment supply chain knowledge graphs is applied to scenarios involving the construction of supply chain knowledge graphs for aero-engines, rail transit equipment, marine power systems, industrial machine tools, or large energy equipment. The system uses key components, suppliers, batches, qualifications, quality inspection reports, process nodes, assembly units, or delivery plans of a complex piece of equipment as target entities, constructing a multi-layered knowledge graph from multi-source supply chain data. The overall execution process of the system corresponds to... Figure 1 This includes acquiring unstructured raw text corpora and structured knowledge sources, entity credibility discrimination and segmentation, domain semantic alignment, structured hints and triplet verification, writing to the current level temporary graph, posterior poisoning risk detection, rollback and regeneration processing, establishing clean version snapshots, and tail entity growth decisions and multi-level expansion.

[0024] The target entity in this embodiment can be a supplier, manufacturer, component, raw material, batch, contract, qualification certificate, quality inspection report, process node, assembly unit, delivery plan, or other entities related to the complex equipment supply chain. Candidate triples are represented in the form of head entity, relation, and tail entity, such as "a supplier – supplies – a certain type of bearing", "a certain batch – corresponding quality inspection report – a certain report number", or "a certain component – ​​used for – a certain assembly unit".

[0025] (a) Obtaining the target entity context and candidate triples.

[0026] This involves acquiring unstructured raw text corpora and publicly available structured knowledge sources within the complex equipment supply chain domain. Unstructured raw text corpora include one or more of the following: procurement announcements, supplier qualification documents, contract texts, delivery records, quality inspection reports, maintenance records, regulatory notices, news announcements, and technical documents. Structured knowledge sources include one or more of the following: enterprise master data, supplier directories, bills of materials, quality management systems, batch traceability systems, publicly available business registration data, qualification certificate databases, and industry knowledge bases.

[0027] This method extracts the context of target entities from unstructured raw text corpora. Specifically, the unstructured raw text corpora are segmented into sentences to obtain a set of sentences; pre-trained language models are used to encode sentences and target entities respectively, resulting in sentence vectors and target entity vectors; the semantic similarity between sentence vectors and target entity vectors is calculated, and irrelevant sentences are filtered using keywords from the complex equipment supply chain domain. The semantic similarity is represented as: in, Indicates the first Sentences and target entities Semantic similarity between them Indicates the first One sentence. Represents the target entity. Represents a sentence vector. Represents the target entity vector.

[0028] In order to ensure that the occurrence count of target entities can be used in conjunction with semantic similarity for ranking, the occurrence count of target entities is normalized: in, Represents the target entity In the The number of times a sentence or its adjacent window appears. This represents the index of each sentence in the sentence set or its corresponding adjacent window. Represents the target entity The maximum number of occurrences in each sentence or corresponding adjacent window. To prevent positive numbers with a denominator of zero, This represents the frequency of occurrence of the target entity after normalization.

[0029] The ranking score is expressed as: in, This represents the weighting coefficient between semantic similarity and the frequency of occurrence of the target entity. This represents the ranking score. The system sorts the filtered sentences from highest to lowest ranking score and concatenates them to obtain the target entity context. Simultaneously, the system retrieves candidate triples related to the target entity from structured knowledge sources.

[0030] (II) Entity authenticity feature vector and adversarial entity credibility judgment.

[0031] Construct entity authenticity feature vectors for the target entity context and candidate entities involved in the candidate triples. The entity authenticity feature vector is formed by concatenating three or more of the following features: textual semantic features of the candidate entity, contextual semantic features of the candidate entity, source credibility features, attribute completeness features, temporal logical consistency features, cross-source consistency features, qualification consistency features, and compatibility features with the existing knowledge graph topology. When a feature is missing, a preset default value or zero value is used to fill in the missing value. The entity authenticity feature vector is represented as: in, Indicates candidate entities The entity authenticity feature vector, Represents the semantic features of candidate entities. Represents the contextual semantic features of candidate entities. , , , , and These represent the source credibility feature, attribute completeness feature, temporal logic consistency feature, cross-source consistency feature, qualification consistency feature, and compatibility feature with existing knowledge graph topology, respectively.

[0032] Source credibility features are determined based on data source type, data source authority level, source update time, and historical accuracy. Attribute completeness features are determined based on the completeness of entity name, unified social credit code, qualification number, batch number, material code, time field, and associated subject field. Time logic consistency features are determined based on whether there are contradictions between qualification validity period, supply time, delivery time, batch production time, and quality inspection time. Cross-source consistency features are determined based on whether entity name, identifier, address, contact person, qualification level, or component model are consistent across multiple independent sources. Qualification consistency features are determined based on whether the entity's declared supply scope is consistent with its qualification certificate, industry access scope, or quality certification scope. Compatibility features with existing knowledge graph topology are determined based on whether the candidate entity is compatible with the adjacency relationships, relationship types, and path patterns in the existing supply chain knowledge graph.

[0033] The entity authenticity feature vector is input into the adversarial entity credibility discrimination model. The adversarial entity credibility discrimination model includes a generator and a discriminator. During the training phase, the generator generates adversarial entity feature vectors based on the entity authenticity feature vectors of real supply chain entities, random perturbation vectors, and preset risk types; the discriminator is trained based on the entity authenticity feature vectors of real supply chain entities, as well as the entity authenticity feature vectors and adversarial entity feature vectors of fake entity samples. During the inference phase, the discriminator outputs an entity credibility score and an entity risk label based on the entity authenticity feature vectors of candidate entities. The entity credibility score and entity risk label are expressed as follows: in, Indicates candidate entities The entity credibility score, Indicates candidate entities Entity risk label, This indicates the confidence level output of the discriminator. Indicates the discriminator on the first The classification output of entity risk labels, This indicates the total number of categories of entity risk labels.

[0034] The adversarial training objective of the adversarial entity credibility discrimination model is represented as: in, Represents a generator. Indicates the discriminator, This represents the adversarial training objective function between the generator and the discriminator. This indicates that the generator parameters are updated with the goal of minimizing the adversarial training objective function; This indicates that the discriminator parameters are updated with the goal of maximizing the adversarial training objective function; The entity authenticity feature vector represents the entity's authenticity within the real supply chain. This indicates that the random perturbation vector is sampled from a preset random perturbation distribution. The distribution of entity authenticity feature vectors representing real supply chain entities; This indicates that the preset risk type is sampled from the preset risk type set. Represents the mathematical expectation; Indicates the loss categorized by entity risk label. This indicates the label loss weight; preset risk types include fake suppliers, fake parts, abnormal qualifications, abnormal batches, time conflicts, relationship mismatches, source conflicts, or abnormal identifiers.

[0035] Based on entity trust scores and entity risk labels, the system classifies candidate entities into trusted entities, restricted entities, and blocked entities. The entity classification rules are expressed as follows: in, Indicates candidate entities The entity partitioning results Represents a trusted entity. Represents a restricted entity. Indicates the intercepting entity. This indicates a high confidence threshold. This indicates a low confidence threshold, and , This represents a set of high-risk labels. This represents a set of medium-risk labels; Trusted entities participate in generating subsequent triples; restricted entities participate in generating subsequent triples only when the evidentiary constraints are met; intercepted entities do not participate in generating subsequent triples and are written into the prohibition on generation constraints. Evidence constraints require that there be no less than a preset number of mutually independent sources of evidence in the target entity's context or domain-related triple set, and that there are no source conflicts, time conflicts, or entity type conflicts.

[0036] (iii) Domain semantic alignment, structured hints and triple validation.

[0037] Candidate triples are converted into triple text descriptions, and the text descriptions are encoded using a sentence vector model to obtain triple semantic vectors. Domain reference vectors are generated based on a terminology database for the complex equipment supply chain domain. The domain relevance between the triple semantic vectors and the domain reference vectors is calculated, and candidate triples with a domain relevance higher than the semantic threshold and that do not contain intercepting entities are retained. Domain relevance is expressed as: in, Indicates the first Domain relevance of candidate triples Indicates the first Triple text description of each candidate triplet Represents a sentence vector model. This represents a domain reference vector obtained from a terminology database for the complex equipment supply chain domain.

[0038] A structured hint text is constructed based on the target entity context, the domain-related triple set, trusted entities, restricted entities, and prohibition constraints on generation. The structured hint text includes a target entity context segment, a domain-related triple set segment, a trusted entity segment, a restricted entity constraint segment, an interception entity prohibition segment, a relation type constraint segment, an output format constraint segment, and a validation feedback segment. The trusted entity segment indicates that the large language model should prioritize using trusted entities; the restricted entity constraint segment indicates that the large language model can only use restricted entities when the evidence constraints are satisfied; and the interception entity prohibition segment explicitly prohibits the large language model from generating triples containing intercepted entities.

[0039] The large language model generates candidate generated triples for the current level based on structured prompt text. Based on a predefined rule base, the candidate generated triples are validated for format consistency, quantity consistency, logical consistency, and entity usage status, resulting in triple validation labels. The predefined rule base includes format rules, quantity rules, logical consistency rules, entity usage status rules, relation type rules, time order rules, attribute constraint rules, and duplicate relation rules. Triple validation labels include one or more of the following: format error, quantity error, entity type conflict, time conflict, source conflict, qualification mismatch, relation direction error, use of intercepted entities, insufficient evidence for restricted entities, and duplicate generation. When a candidate generated triple fails validation, the system writes the triple validation label into the prompt text, prompting the large language model to generate a validated triple again.

[0040] (iv) Temporarily store the map, generate source records and derivation relationships.

[0041] Instead of directly writing the validated triples into the final graph, they are first written into a temporary graph at the current level. This temporary graph is used to hold new nodes and edges that have not yet passed the posterior poisoning risk detection. Only after the temporary graph passes the posterior poisoning risk detection is it submitted as the clean version of the graph at the current level.

[0042] The generated source traceability records include node source traceability records and edge source traceability records. Node source traceability records include node origin, entity credibility score, entity risk label, hierarchy, parent head entity, and generation round. Edge source traceability records include relationship origin, corresponding structured hint text, corresponding large language model output, rule detection results, parent head entity, and version identifier. Any temporary graph is written to an object. The generation traceability record is represented as: in, This indicates the graph elements temporarily stored in the current level, which are newly added entity nodes or newly added relation edges; Indicates the source text or source triple. This indicates structured prompt text. Indicates the generation round, Indicates the level to which it belongs. Indicates the parent head entity, Represents the entity's credibility score. This indicates an entity risk label or triplet check label. Indicates version identifier; Derivation relationship is denoted as It is used to characterize the association between the parent head entity and newly added nodes or new relation edges derived from it.

[0043] Through derivation relationships, the system tracks the intra-layer propagation path of risky entities or risky relationship edges in the current-level temporary graph and the cross-layer propagation path in subsequent-level temporary graphs.

[0044] (v) Post-transformer-Mamba posterior risk detection of poisoning.

[0045] Based on the generated source tracing records and derivation relationships, risk propagation sequences are constructed for newly added nodes, newly added relation edges, intra-layer propagation paths, and cross-layer propagation paths in the temporary graph. The posterior poisoning risk detection object is denoted as... The post-hoc poisoning risk detection targets are entity nodes, relation edges, intra-layer propagation paths, or cross-layer propagation paths. The risk propagation sequence includes the generation traceability record of graph elements, parent header entity, derived relation path, entity credibility score, entity risk label, triple check label, belonging level, and version identifier. The risk propagation sequence is represented as follows: in, This indicates the target of posterior poisoning risk detection, which can be an entity node, a relation edge, or a cross-layer path; when When it is an entity node, This represents the entity credibility score of the entity corresponding to the entity node; when When it is a relational edge or a cross-layer path, express The lowest entity credibility score or the weighted average entity credibility score of the associated entity; Representing spectral elements The risk transmission sequence, The parameter is Transformer self-attention encoding layer, The parameter is Mamba Selective State Space Coding Layer Represents the pooling function. This represents a local correlation. This indicates long-range propagation characteristics. This represents the risk propagation characterization obtained through fusion. Representing spectral elements Poisoning risk score, Represents the normalized activation function. , , to Indicates model parameters, , , , , These represent the risks associated with abnormal relationship types, abnormal paths, source conflicts, time conflicts, and abnormal attributes, respectively.

[0046] A hybrid Transformer-Mamba posterior poisoning risk detection model is proposed, comprising a Transformer self-attention encoding layer and a Mamba selective state space encoding layer. The Transformer self-attention encoding layer is used to capture the semantic associations between risk graph elements and their local context, source text, and adjacent relation edges; the Mamba selective state space encoding layer is used to model long-range dependencies in the generation source record, derivation path, and cross-layer propagation sequence.

[0047] When the poisoning risk scores of all posterior poisoning risk detection objects do not exceed the preset risk threshold, the system submits the temporary map as a clean version map for the current level and establishes a clean version snapshot. When the poisoning risk score of at least one posterior poisoning risk detection object exceeds the preset risk threshold, the system enters the affected submap location process.

[0048] (vi) Location of affected subgraphs, snapshot rollback and constraint regeneration.

[0049] The specific process for locating the affected subgraph, rolling back the snapshot, and regenerating constraints is as follows: First, the system identifies risk nodes, risk relationship edges, or risk paths whose poisoning risk scores exceed a preset risk threshold; based on the generated source tracing records, it traces the source text, source triples, structured prompt text, and parent header entities of the risk nodes, risk relationship edges, or risk paths in reverse; based on the derivation relationship, it traces the downstream nodes, downstream relationship edges, and downstream cross-layer paths derived from the risk nodes, risk relationship edges, or risk paths downwards; and it identifies the risk nodes, risk relationship edges, risk paths, and their downstream nodes, downstream relationship edges, and downstream cross-layer paths as the affected subgraphs.

[0050] Affected subgraph and regenerating constraints They are represented as follows: Among them, when When the risk threshold is exceeded, the target of the subsequent poisoning risk detection is identified as a risk map element; This indicates that the poisoning risk score exceeds the preset risk threshold. The set of risk map elements, Indicates the derivation relationship The downstream nodes and downstream relationship edges are obtained by tracing down from the set of elements in the risk graph. Indicates the affected subgraph; This indicates that constraints will be regenerated. This indicates that the collection of entities is prohibited. Represents a restricted set of entities. This represents the set of prohibited edge pairs or the set of prohibited relation types. Represents the set of risk paths. This represents the set of triple check tags. This represents the set of poisoning risk scores corresponding to each risk map element in the affected subgraph. The affected subgraph undergoes a rollback and regeneration process, including: undoing the write operations corresponding to the affected subgraph and restoring the graph state of the affected subgraph to the most recent clean version snapshot before the affected subgraph was generated; adding risky entities in the affected subgraph to the prohibited entity set or restricted entity set; adding risky relation edges in the affected subgraph to the prohibited relation edge set or prohibited relation type set; writing the triple check label and poisoning risk score corresponding to the affected subgraph into the regeneration constraints; reconstructing the structured prompt text based on the updated regeneration constraints and calling the large language model to generate alternative triples; and re-performing triple check, temporary graph writing, source tracing record generation, derivation relation establishment, and posterior poisoning risk detection on the alternative triples.

[0051] When the same affected branch is rolled back and regenerated multiple times, the constraint strength of the regeneration constraint is increased, such as increasing the requirement for restricted entity evidence, expanding the set of prohibited relation types, reducing the probability of tail entity growth, or marking the corresponding branch as requiring manual review; when the same affected branch reaches the preset upper limit of the number of regenerations, the affected branch is marked as terminated.

[0052] (vii) Tail entity growth and exemplary execution results.

[0053] The tail entity growth decision and multi-level graph iterative construction process specifically involves using a tail entity growth decision model to determine whether a tail entity at the current level should continue to expand as a next-level head entity only after the current-level temporary graph has been submitted as a clean version. The tail entity growth decision model outputs the growth probability of a tail entity becoming a next-level head entity based on the tail entity context, entity credibility score, poisoning risk score, triplet verification results, current level, parent path, and historical expansion status. The dynamic growth threshold is expressed as: in, Indicates the first The dynamic growth threshold of the layer Indicates the basic growth threshold. Indicates the first Layer average entity credibility score Indicates the first Average poisoning risk score per layer Indicates the first Layer triplet check tag trigger ratio, , , Indicates the adjustment factor. and These represent the lower and upper limits of the dynamic growth threshold, respectively. When the growth probability of a tail entity is greater than or equal to the dynamic growth threshold, the corresponding tail entity is marked as growing and continues the above process as the head entity of the next level; when the growth probability of a tail entity is less than the dynamic growth threshold, the corresponding tail entity is marked as terminated.

[0054] When targeting "a certain type of aircraft engine bearing" as the target entity, the system extracts the target entity context from procurement announcements, supplier qualification documents, quality inspection reports, and material master data, and retrieves candidate triples from structured knowledge sources. Candidate triples include relationships such as supplier, batch, model, material, quality inspection report, delivery record, and assembly unit. For candidate supplier entities, the system constructs an entity authenticity feature vector and inputs it into an adversarial entity credibility discrimination model. If a supplier's name in publicly available business registration data differs from its name in qualification documents, and the scope of its quality certification does not fully match the category of supplied parts, the supplier can be marked as a restricted entity. If the supplier has consistent qualification numbers, validity periods, and supply scopes from multiple independent sources, it meets the evidentiary constraints and can participate in subsequent triple generation. If a candidate entity has a forged qualification number, time conflict, or source conflict, it is marked as an intercepted entity and a prohibition on generation constraint is added.

[0055] After the large language model generates triples based on the structured prompt text, the system first writes the verified triples into the current level's temporary graph and generates traceability records for newly added supplier nodes, batch nodes, quality report nodes, and supply relationship edge records. The system constructs intra-layer and cross-layer propagation paths based on derivation relationships, such as "aircraft engine bearing – supplier – batch – quality inspection report – assembly unit". A hybrid Transformer-Mamba posterior poisoning risk detection model detects this path; the Transformer self-attention encoding layer captures the local semantic associations between supplier name, qualification number, quality report, and adjacent relationships; the Mamba selective state space encoding layer models the long-range propagation dependencies from the target entity to the supplier, batch, inspection report, and assembly unit. If the hybrid Transformer-Mamba posterior poisoning risk detection model detects source and temporal conflicts in this path, and the poisoning risk score exceeds a preset risk threshold, the system locates the affected subgraph containing the supplier node, related batch nodes, quality report nodes, and their downstream relationship edges.

[0056] The write operation corresponding to the affected subgraph is revoked, and the graph is restored to the most recent clean version snapshot before the affected subgraph was generated. The supplier or related relationship type is added to the prohibited entity set, restricted entity set, or prohibited relationship edge set. Subsequently, the triple check label and poisoning risk score are written into the regeneration constraint, the structured prompt text is reconstructed, and a large language model is called to generate alternative triples. Only after the alternative triples pass rule validation, temporary writing, source tracing, derivation relationship establishment, and posterior poisoning risk detection are they allowed to be submitted as a clean version graph.

[0057] In this embodiment, the method can be collaboratively implemented by a data access module, a context extraction module, a candidate triplet retrieval module, an entity credibility judgment module, an entity classification module, a domain semantic alignment module, a structured prompt construction module, a large language model generation module, a triplet verification module, a temporary graph management module, a source tracing record generation module, a derivation relationship management module, a hybrid Transformer-Mamba posterior poisoning risk detection module, an affected subgraph location module, a clean version snapshot management module, a rollback and regeneration module, and a tail entity growth decision module. The temporary graph and clean version graph can be implemented using graph databases, relational databases, document databases, or in-memory graph structures. The source tracing record generation and clean version snapshot generation can be indexed using version number, timestamp, parent header entity identifier, and generation round. The hybrid Transformer-Mamba posterior poisoning risk detection model can be deployed as an independent model service or as a risk detection submodule in a graph construction pipeline.

[0058] The above embodiments are merely examples illustrating the technical solutions of the present invention. Any equivalent substitutions or modifications made based on the technical concept of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for robust construction of a complex equipment supply chain knowledge graph, characterized in that, Includes the following steps: Step S1: Obtain unstructured text corpus and publicly available structured knowledge sources in the field of complex equipment supply chain; extract the target entity context from the unstructured text corpus; and retrieve candidate triples related to the target entity from the structured knowledge sources, including relation information, attribute information and source information. Step S2: Based on the textual semantic information and contextual association information of each candidate entity in the candidate triplet within the target entity context, as well as the relational information, attribute information, and source information of the candidate entity in the candidate triplet, construct the corresponding entity authenticity feature vector, including the textual semantic features of the candidate entity, the contextual semantic features of the candidate entity, the source credibility features, the attribute completeness features, the temporal logic consistency features, the cross-source consistency features, the qualification consistency features, and the compatibility features with the existing knowledge graph topology; input the entity authenticity feature vector into the adversarial entity credibility discrimination model to obtain the entity credibility score and entity risk label of the candidate entity; Step S3: Based on the entity credibility score and entity risk label, divide the candidate entities into trusted entities, restricted entities, and blocked entities; Step S4: Perform domain semantic alignment on the candidate triples to obtain a set of domain-related triples; Structured prompt text is constructed based on the target entity context, the domain-related triple set, trusted entities, restricted entities, and prohibition on generation constraints; Step S5: Input the structured prompt text into the large language model to generate candidate generated triples for the current level. Based on the predefined rule base, perform format consistency, quantity consistency, logical consistency and entity usage status verification on the candidate generated triples to obtain triple verification labels. When there are candidate generated triples that fail the validation, the triple validation label is written into the prompt text, and the large language model is prompted to generate validated triples again. Step S6: First, write the verified triplet into the temporary graph of the current level, and generate a traceability record for each newly added node and each newly added relation edge in the temporary graph; establish a derivation relationship between the newly added node, the newly added relation edge and its parent head entity, which is used to track the intra-layer propagation path of risky entities or risky relation edges in the temporary graph of the current level, as well as the cross-layer propagation path between different level graphs formed in subsequent iterations. Step S7: Based on the generated source traceability records and derivation relationships, construct risk propagation sequences for newly added nodes, newly added relation edges, intra-layer propagation paths, and cross-layer propagation paths in the temporary graph. Input the risk propagation sequences into the hybrid Transformer-Mamba posterior poisoning risk detection model to obtain the risk propagation representation and poisoning risk score of each graph element. When the poisoning risk score of each graph element does not exceed the preset risk threshold, the temporary graph is submitted as the clean version of the graph at the current level, and a clean version snapshot is created; when the poisoning risk score of at least one graph element exceeds the preset risk threshold, the affected subgraph is located in the temporary graph at the current level and its associated cross-level derivation path. Step S8: Perform rollback and regeneration processing on the affected subgraphs; The rollback and regeneration process includes: undoing the write operations corresponding to the affected subgraph, restoring the graph state of the affected subgraph to the most recent clean version snapshot before the affected subgraph was generated; adding risky entities in the affected subgraph to the prohibited entity set or restricted entity set, adding risky relation edges in the affected subgraph to the prohibited relation edge set or prohibited relation type set, and writing the triple check label and poisoning risk score corresponding to the affected subgraph into the regeneration constraints; reconstructing the structured prompt text based on the updated regeneration constraints and calling the large language model to generate alternative triples; and re-performing triple check, temporary graph writing, generation of source traceability records, establishment of derivation relations, and posterior poisoning risk detection on the alternative triples. Step S9: Only after the temporary map of the current level is submitted as a clean version map, the tail entity growth decision model is used to determine whether the tail entity of the current level should continue to expand as the head entity of the next level, and the tail entity growth threshold is dynamically adjusted according to the entity credibility score, poisoning risk score and triplet verification result of the current level. Step S10: Repeat steps S1 to S9 until all tail entities are marked as terminated, loops are detected, a preset iteration depth is reached, or the same affected branch reaches a preset maximum number of regenerations, and output a multi-layer complex equipment supply chain knowledge graph. 2.The method of claim 1, wherein, The target entity context includes textual semantic information and contextual association information, obtained through the following methods: unstructured raw text corpus is segmented into sentences to obtain a sentence set; pre-trained language models are used to encode sentences and target entities respectively, obtaining sentence vectors and target entity vectors; the semantic similarity between sentence vectors and target entity vectors is calculated, and irrelevant sentences are filtered using keywords from the complex equipment supply chain domain; a ranking score is constructed based on semantic similarity and the frequency of target entity occurrences, and the target entity context is obtained by concatenating these scores from high to low; wherein, the semantic similarity and ranking score are respectively: in, Indicates the first Sentences and target entities Semantic similarity between them Indicates the first One sentence. Represents the target entity. Represents a sentence vector. Represents the target entity vector. Represents the target entity In the The number of times a sentence or its adjacent window appears. This represents the index of each sentence in the sentence set or its corresponding adjacent window. Represents the target entity The maximum number of occurrences in each sentence or corresponding adjacent window. To prevent positive numbers with a denominator of zero, This indicates the frequency of occurrence of the target entity after normalization. This represents the weighting coefficient between semantic similarity and the frequency of occurrence of the target entity. This indicates the sorting score.

3. The robust construction method for a knowledge graph of a complex equipment supply chain according to claim 1, characterized in that, The entity authenticity feature vector is formed by concatenating three or more of the following features: textual semantic features of the candidate entity, contextual semantic features of the candidate entity, source credibility features, attribute completeness features, temporal logical consistency features, cross-source consistency features, qualification consistency features, and compatibility features with the existing knowledge graph topology; when a feature is missing, a preset default value or zero value is used to fill it; the entity authenticity feature vector is represented as: in, Indicates candidate entities The entity authenticity feature vector, The textual semantic features representing candidate entities Represents the contextual semantic features of candidate entities. , , , , , These represent the source credibility feature, attribute completeness feature, temporal logic consistency feature, cross-source consistency feature, qualification consistency feature, and compatibility feature with existing knowledge graph topology, respectively.

4. The robust construction method for a knowledge graph of a complex equipment supply chain according to claim 1, characterized in that, The adversarial entity credibility discrimination model includes a generator and a discriminator. During the training phase, the generator generates adversarial entity feature vectors based on the entity authenticity feature vectors of real supply chain entities, random perturbation vectors, and preset risk types. The discriminator is trained based on the entity authenticity feature vectors of real supply chain entities, entity authenticity feature vectors of fake entity samples, and adversarial entity feature vectors. During the inference phase, the discriminator outputs entity credibility scores and entity risk labels based on the entity authenticity feature vectors of candidate entities. The entity credibility score and entity risk label are represented as follows: in, Indicates candidate entities The entity credibility score, Indicates the entity's risk label, This indicates the confidence level output of the discriminator. Indicates the discriminator on the first Classification output of entity risk labels This indicates the total number of categories of entity risk labels; The adversarial training objective of the adversarial entity credibility discrimination model is: in, Represents a generator. Indicates the discriminator, This represents the adversarial training objective function between the generator and the discriminator. This indicates that the generator parameters are updated with the goal of minimizing the adversarial training objective function; This indicates that the discriminator parameters are updated with the goal of maximizing the adversarial training objective function; The entity authenticity feature vector represents the entity's authenticity within the real supply chain. This indicates that the random perturbation vector is sampled from a preset random perturbation distribution. The distribution of entity authenticity feature vectors representing real supply chain entities; This indicates that the preset risk type is sampled from the preset risk type set. Represents the mathematical expectation; Indicates the loss categorized by entity risk label. This indicates the label loss weight; preset risk types include fake suppliers, fake parts, abnormal qualifications, abnormal batches, time conflicts, relationship mismatches, source conflicts, or abnormal identifiers.

5. The robust construction method for a knowledge graph of a complex equipment supply chain according to claim 4, characterized in that, The entity credibility score is output by the discriminator, and the entity segmentation rule is as follows: in, Indicates candidate entities The entity partitioning results Represents a trusted entity. Represents a restricted entity. Indicates the intercepting entity. This indicates a high confidence threshold. This indicates a low confidence threshold, and , This represents a set of high-risk labels. This represents a set of medium-risk labels; Trusted entities participate in generating subsequent triples, restricted entities only participate in generating subsequent triples when the evidence constraints are met, and intercepting entities do not participate in generating subsequent triples and are written into the prohibition of generation constraints. The evidence constraint is that there is no less than a preset number of mutually independent sources of evidence in the target entity context or domain-related triple set, and there are no source conflicts, time conflicts or entity type conflicts.

6. The robust construction method for a knowledge graph of a complex equipment supply chain according to claim 1, characterized in that, The domain semantic alignment includes: converting candidate triples into triple text descriptions, encoding the triple text descriptions using a sentence vector model to obtain triple semantic vectors; generating domain reference vectors based on a complex equipment supply chain domain terminology library; calculating the domain relevance between the triple semantic vectors and the domain reference vectors, and retaining candidate triples with a domain relevance higher than a semantic threshold and that do not contain intercepting entities; wherein, the domain relevance is: in, Indicates the first Domain relevance of candidate triples Indicates the first Triple text description of each candidate triplet Represents a sentence vector model. This represents a domain reference vector obtained from a terminology database for the complex equipment supply chain domain.

7. The robust construction method for a knowledge graph of a complex equipment supply chain according to claim 1, characterized in that, The generated source traceability record includes node source traceability record and edge source traceability record; the node source traceability record includes node source, entity credibility score, entity risk label, hierarchy, parent head entity, and generation round; the edge source traceability record includes relation source, corresponding structured prompt text, corresponding large language model output, rule detection result, parent head entity, and version identifier; the generated source traceability record for any temporarily stored graph element. Represented as: in, This indicates the graph elements temporarily stored in the current level, which are newly added entity nodes or newly added relation edges; Indicates the source text or source triple. This indicates structured prompt text. Indicates the generation round, Indicates the level to which it belongs. Indicates the parent head entity, Represents the entity's credibility score. This indicates an entity risk label or triplet check label. Indicates version identifier; The structured hint text includes trusted entity sections, restricted entity constraint sections, and blocked entity prohibition sections; Derivation relationship is denoted as It is used to characterize the association between the parent head entity and newly added nodes or new relation edges derived from it.

8. The robust construction method for a knowledge graph of a complex equipment supply chain according to claim 1, characterized in that, The hybrid Transformer-Mamba posterior poisoning risk detection model includes a Transformer self-attention encoding layer and a Mamba selective state space encoding layer. The Transformer self-attention encoding layer is used to capture the semantic associations between risk graph elements and their local context, source text, and adjacent relation edges. The Mamba selective state space encoding layer is used to model long-range dependencies in the generation of source records, derivation paths, and cross-layer propagation sequences. The hybrid Transformer-Mamba posterior poisoning risk detection model obtains the poisoning risk score in the following manner: A risk propagation sequence is constructed for each graph element, including the generation traceability record of the graph element, parent head entity, derivation path, entity credibility score, entity risk label, triple check label, hierarchy, and version identifier; the risk propagation sequence is input into the Transformer self-attention encoding layer to obtain a local association representation, and then input into the Mamba selective state space encoding layer to obtain a long-range propagation representation; the local association representation and the long-range propagation representation are fused to obtain the risk propagation representation; the poisoning risk score is calculated based on the risk propagation representation and risk factors; wherein: in, This indicates the target of posterior poisoning risk detection, which can be an entity node, a relation edge, or a cross-layer path; when When it is an entity node, This represents the entity credibility score of the entity corresponding to the entity node; when When it is a relational edge or a cross-layer path, express The lowest entity credibility score or the weighted average entity credibility score of the associated entity; Representing spectral elements The risk transmission sequence, The parameter is Transformer self-attention encoding layer, The parameter is Mamba Selective State Space Coding Layer Represents the pooling function. This represents a local correlation. This indicates long-range propagation characteristics. This represents the risk propagation characterization obtained through fusion. Representing spectral elements Poisoning risk score, Represents the normalized activation function. , , to Indicates model parameters, , , , , These represent the risks associated with abnormal relationship types, abnormal paths, source conflicts, time conflicts, and abnormal attributes, respectively.

9. The robust construction method for a knowledge graph of a complex equipment supply chain according to claim 1, characterized in that, The affected subgraph is a local graph structure determined from the current-level temporary graph and the cross-layer propagation path associated with the current-level temporary graph. This includes: identifying risk nodes, risk relationship edges, or risk paths whose poisoning risk scores exceed a preset risk threshold; tracing the source text, source triples, structured prompt text, and parent header entities of the risk nodes, risk relationship edges, or risk paths based on generated source tracing records; tracing downstream nodes, downstream relationship edges, and downstream cross-layer paths derived from the risk nodes, risk relationship edges, or risk paths based on derivation relationships; and identifying the risk nodes, risk relationship edges, risk paths, and their downstream nodes, downstream relationship edges, and downstream cross-layer paths as the affected subgraph. Affected subgraph and regenerating constraints They are represented as follows: Among them, when When the risk threshold is exceeded, the target of the subsequent poisoning risk detection is identified as a risk map element; This indicates that the poisoning risk score exceeds the preset risk threshold. A set of risk map elements, Indicates the derivation relationship The downstream nodes and downstream relationship edges are obtained by tracing down from the set of elements in the risk graph. Indicates the affected subgraph; This indicates that constraints will be regenerated. This indicates that the collection of entities is prohibited. Represents a restricted set of entities. This represents the set of prohibited edge pairs or the set of prohibited relation types. Represents the set of risk paths. This represents the set of triple check tags. This represents the set of poisoning risk scores corresponding to each risk map element in the affected subgraph. The regeneration constraints are written into the structured prompt text, which enables the large language model to exclude prohibited entities and prohibited relation edges during regeneration and impose evidence requirements on restricted entities. When the same affected branch is rolled back and regenerated multiple times, the constraint strength of the regeneration constraints is increased until the affected branch reaches the preset regeneration limit and is marked as terminated.

10. The robust construction method for a knowledge graph of a complex equipment supply chain according to claim 1, characterized in that, The tail entity growth decision model outputs the probability that the tail entity will grow as the next-level head entity, and determines whether the tail entity should grow based on a dynamic growth threshold; the dynamic growth threshold is: in, Indicates the first The dynamic growth threshold of the layer Indicates the basic growth threshold. Indicates the first Layer average entity credibility score Indicates the first Average poisoning risk score per layer Indicates the first Layer triplet check tag trigger ratio, Indicates the entity credibility adjustment factor, Indicates the poisoning risk adjustment coefficient, This represents the adjustment factor for the triplet check result. and These represent the lower limit and upper limit of the dynamic growth threshold, respectively; when the growth probability of a tail entity is greater than or equal to the dynamic growth threshold, the corresponding tail entity is marked as growing; when the growth probability of a tail entity is less than the dynamic growth threshold, the corresponding tail entity is marked as terminating.

Citation Information

Patent Citations

  • Bidding supplier risk intelligent detection method and device based on knowledge graph

    CN121481275A

  • KR1018206350000B1