A multi-dimensional logic rule intelligent mining method and system

Abstract

The application belongs to the technical field of artificial intelligence, manufacturing knowledge graph and knowledge reasoning, and discloses a kind of multi-dimensional logic rule intelligent mining method and system.The method obtains and normalizes original manufacturing knowledge graph;Based on five kinds of rules of symmetry, transitivity, hierarchical propagation, horizontal association and parallel association, the candidate rule body is generated directionally;The candidate rule body, field information and relation set are input into the large language model, and the candidate rule head is inferred to obtain the candidate logic rule;The confidence and dispersion are calculated by combining the graph structure support and the multi-round evaluation of the large language model, and the high-confidence logic rule is screened;The high-confidence rule is applied to the graph to generate new triples, and the self-refined graph is obtained after detection, deduplication and merging.The application reduces the reasoning load and illusion risk, and improves the explainability and reliability of the completion.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, manufacturing knowledge graphs, and knowledge reasoning, and specifically relates to a multi-dimensional logical rule intelligent mining method and system for self-refinement of manufacturing knowledge graphs. Background Technology

[0002] Large manufacturing enterprises continuously accumulate vast amounts of domain knowledge across various business processes, including product design, process planning, production, quality control, equipment operation and maintenance, supply chain coordination, and customer service. Manufacturing knowledge graphs can organize this knowledge, scattered across different business systems and knowledge carriers, into entities, relationships, and triples, providing support for knowledge retrieval, reasoning, and business decision-making. However, as manufacturing knowledge continues to update and business scenarios expand, the entities, attributes, and relationships explicitly stored in the original manufacturing knowledge graph are insufficient to fully meet the demands of complex manufacturing tasks for accessing implicit knowledge. Therefore, continuous refinement and completion of the manufacturing knowledge graph are necessary.

[0003] To address the refinement of knowledge graphs, existing research utilizes methods such as symbolic rule mining, association rule analysis, and local path search to summarize interpretable logical rules from the knowledge graph. For example, some methods search for high-frequency co-occurring paths based on the locally closed world assumption or frequent subgraph mining techniques, thereby generating candidate rules; other methods introduce combinations of first-order logic, constant rules, or relational paths into the knowledge graph completion process to improve the interpretability of newly added triples. These methods are effective in scenarios where the rules are clearly defined, but when faced with a large number of entities, complex relationships, and long-tail knowledge in the knowledge graph, the candidate search space often expands rapidly, limiting mining efficiency and rule coverage.

[0004] Some studies employ deep learning and representation learning methods for knowledge graph reasoning completion. These methods typically embed entities and relationships into a low-dimensional vector space, or learn path features and rule patterns through models such as graph neural networks, recurrent neural networks, and reinforcement learning to predict potential relationships. However, manufacturing enterprise knowledge is characterized by its coupling across processes, products, equipment, and business segments, and often lacks a large number of high-quality labeled samples in specific manufacturing domains, affecting the model's generalization ability. Furthermore, the completion results obtained from pure representation learning often fail to directly explain which manufacturing facts and logical rules led to the new knowledge.

[0005] In recent years, large language models have demonstrated strong capabilities in natural language understanding, semantic reasoning, and zero-shot transfer, providing new technical means for discovering high-order rules in manufactured knowledge graphs. Large language models can understand the specialized semantics within manufactured text and infer candidate rule heads based on given graph triples and relation sets. However, directly inputting large-scale manufactured knowledge graphs into large language models for end-to-end rule inference can lead to an excessively large number of candidate rule combinations and high inference costs. Furthermore, the lack of graph structure constraints increases the risk of semantic drift and illusory outputs.

[0006] The aforementioned methods still have shortcomings in handling the self-refinement of manufacturing knowledge graphs: First, the rule types are mostly concentrated on reciprocal relationships or simple transitive relationships, making it difficult to cover composite entities, hierarchical structures, two-hop contexts, and parallel collaborative relationships; second, there is a lack of candidate rule body compression mechanisms for different rule types, making it difficult to control the number of candidates in large-scale graphs; third, the inference results of large language models lack structural support and joint verification of domain semantics, easily leading to low-quality rules; and fourth, newly added manufacturing knowledge triples are difficult to form a traceable closed loop with rule bodies, rule heads, and confidence scores. Therefore, an intelligent mining method is needed that can simultaneously complete multi-dimensional rule definition, candidate rule body generation, large language model rule head inference, mixed confidence score evaluation, and graph updating. Summary of the Invention

[0007] To address the problems in existing technologies such as insufficient coverage of implicit relation types in manufactured knowledge graphs, excessively large candidate rule search space, insufficient reliability of direct reasoning from large language models, and difficulty in continuously refining the graph, this invention provides a multi-dimensional logical rule intelligent mining method and system for the self-refinement of manufactured knowledge graphs.

[0008] A method for intelligent mining of multidimensional logical rules includes the following steps:

[0009] Step S1: Obtain the original manufacturing knowledge graph, and perform structured representation and normalization processing on the manufacturing knowledge entities, manufacturing knowledge relations and manufacturing knowledge triples in the original manufacturing knowledge graph to obtain the entity set, relation set and observed manufacturing knowledge fact triple set;

[0010] Step S2: Based on the preset multi-dimensional logical rule pattern, candidate rule bodies are generated for the original manufacturing knowledge graph, resulting in symmetric rule candidate bodies, transitive rule candidate bodies, hierarchical propagation rule candidate bodies, horizontal association rule candidate bodies, and parallel association rule candidate bodies.

[0011] Step S3: Input the candidate rule body, manufacturing domain information, the relation set and the corresponding logical rule pattern into the large language model, infer the candidate rule head based on the corresponding prompting strategy, and obtain the candidate logical rules;

[0012] Step S4: Perform a mixed confidence evaluation on the candidate logical rules, wherein objective confidence is calculated based on the structural support of the rule body and rule head in the original manufacturing knowledge graph, expert confidence and expert rating dispersion are calculated based on multi-round expert evaluation of the large language model, and a mixed confidence is calculated based on the objective confidence and the expert confidence; the candidate logical rules are screened according to the preset confidence threshold and expert rating dispersion threshold to obtain high-confidence logical rules;

[0013] Step S5: Apply the high-confidence logic rules to the original manufacturing knowledge graph to generate new manufacturing knowledge triples. Then, normalize, detect conflicts, remove duplicates, and merge the new manufacturing knowledge triples to obtain a self-refined manufacturing knowledge graph.

[0014] Furthermore, step S1 includes the following steps:

[0015] Step S101: Read the node, edge and triple data in the manufacturing knowledge graph, and represent the manufacturing object, material, process, equipment, parameter, performance, organization and product as a unified manufacturing knowledge entity;

[0016] Step S102: The semantics of attributes, composition, use, belonging, influence, cause, generation, requirement and association between entities are uniformly represented as manufacturing knowledge relations;

[0017] Step S103: Represent any manufacturing knowledge fact as a manufacturing knowledge triple consisting of a head entity, a relation, and a tail entity;

[0018] Step S104 involves normalizing synonymous entities, synonymous relations, redundant spaces, invalid symbols, and repeated triples to obtain a raw manufacturing knowledge graph that can be used for logical rule mining.

[0019] Furthermore, the generation of symmetry rule candidate bodies in step S2 includes the following steps:

[0020] Step S201: Extract candidate relations from the relation set, and use a large language model to identify inverse relation pairs and self-symmetric relations in the candidate relations;

[0021] Step S202: Based on the identified reciprocal relationship pairs and self-symmetric relationships, match the corresponding triples in the original manufacturing knowledge graph to form the first round of symmetry rule candidate rule bodies;

[0022] Step S203: Construct a few-sample example based on the identified high-confidence reciprocal relation pairs, and input the unmatched relation set, relation semantics, and the few-sample example into the large language model to identify potential reciprocal relation pairs;

[0023] Step S204: Generate a second round of candidate rules for symmetry rules based on the potential reciprocal relationship, and merge it with the first round of candidate rules for symmetry rules.

[0024] Furthermore, the generation of transitive rule candidate rule bodies in step S2 includes the following steps:

[0025] Step S211: Enumerate consecutive triplet chains in the original manufacturing knowledge graph, wherein the head entity of the next triplet in the consecutive triplet chain is the same as the tail entity of the previous triplet.

[0026] Step S212: Control the length of the candidate triplet chain according to the preset minimum chain length and maximum chain length, and filter out the loop path containing duplicate entities.

[0027] Step S213: Select the triplet chain that satisfies the chain length constraint and the entity non-repetition constraint as the transitive rule candidate rule body;

[0028] Step S214: Input the transitive rule candidate rule body, relation set and manufacturing domain information into the large language model to infer the candidate rule head between the first entity and the last entity of the path.

[0029] Furthermore, the generation of candidate rule bodies for hierarchical propagation rules in step S2 includes the following steps:

[0030] Step S221: Identify candidate composite entities based on the numerical units, connectors, parentheses, component codes, process flow words, and combination structures in the manufacturing entity text;

[0031] Step S222: Use a large language model to split the candidate composite entity into sub-entities with independent manufacturing semantics;

[0032] Step S223: Filter the sub-entities obtained from the splitting based on the sub-entity length, numerical attribute, unit attribute, tolerance attribute, and domain semantic integrity.

[0033] Step S224: Retrieve adjacent triples of the composite entity in the original manufacturing knowledge graph, and construct a candidate rule body for hierarchical propagation rules consisting of the composite entity, effective sub-entities, and external adjacent entities;

[0034] Step S225: Input the candidate rule body of the hierarchical propagation rule into the large language model to infer the candidate rule head between the effective sub-entity and the external adjacent entity.

[0035] Furthermore, the generation of candidate rule bodies for horizontal association rules in step S2 includes the following steps:

[0036] Step S231: Reuse the valid composite entities and valid sub-entities obtained during the hierarchical propagation rule mining process;

[0037] Step S232: Obtain the one-hop neighborhood entity of the effective composite entity, and expand the two-hop neighborhood entity based on the one-hop neighborhood entity to obtain the two-hop neighborhood entity;

[0038] Step S233: Filter the second-hop relationship based on the preset relationship whitelist to obtain a two-hop candidate triple chain that conforms to the semantics of the manufacturing domain;

[0039] Step S234: Construct a candidate rule body for horizontal association rules, consisting of effective sub-entities, composite entities, one-hop neighborhood entities, and two-hop neighborhood entities;

[0040] Step S235: Input the candidate rule body of the horizontal association rule into the large language model to infer the candidate rule head between the effective sub-entity and the two-hop neighbor entity.

[0041] Furthermore, the parallel association rule candidate rule body generation in step S2 includes the following steps:

[0042] Step S241: Enumerate triplet pairs that share the same central entity in the original manufacturing knowledge graph;

[0043] Step S242: Based on the position of the central entity in the two triples, divide the triples with the shared central entity into homogeneous parallel structures and homogeneous parallel structures.

[0044] Step S243: Detect whether there is an explicit relationship between the non-central entities in the two triples;

[0045] Step S244: For parallel structures with explicit relationships, they are used as candidate rules with computable objective confidence; for parallel structures without explicit relationships, they are used as candidate rule bodies to be inferred.

[0046] Step S245: The high-confidence existing parallel rules are used as few-sample examples and input together with the candidate rule body to be inferred into the large language model to infer the candidate rule heads between non-central entities.

[0047] Furthermore, in step S3, the prompting strategy for the large language model to infer candidate rule heads includes the identity positioning of the prompting strategy, the task description of the prompting strategy, the precautions of the prompting strategy, and the output specification of the prompting strategy. Specifically, the identity positioning of the prompting strategy is used to define the large language model as an expert in mining logical rules for creating knowledge graphs; the task description of the prompting strategy is used to define the large language model to infer candidate rule heads based on candidate rule bodies and logical rule patterns; the precautions of the prompting strategy are used to define that the large language model must not modify the input rule body, must not output duplicate rule heads, must preferentially select normative relations from the existing relation set, and determine the direction of the rule head based on entity semantics; the output specification of the prompting strategy is used to define that the large language model outputs the rule type, rule body, rule head, and inference basis in a structured format.

[0048] Furthermore, step S4 includes the following steps:

[0049] Step S401: Count the number of times the rule body of the candidate logical rule appears in the original manufacturing knowledge graph, and the number of times the rule body and rule head are both true, and calculate the objective confidence level.

[0050] Step S402: Using a large language model, candidate logical rules are scored in multiple rounds from four dimensions: logical rationality, domain applicability, semantic consistency, and application value, and expert confidence is calculated.

[0051] Step S403: Calculate the standard deviation of the expert scores over multiple rounds to obtain the dispersion of the expert scores;

[0052] Step S404: Obtain the mixed confidence level based on the weighted result of the objective confidence level and the expert confidence level;

[0053] Step S405: When the mixed confidence level is not lower than a preset confidence threshold and the expert rating dispersion is not higher than a preset dispersion threshold, the candidate logic rule is determined as a high-confidence logic rule.

[0054] This invention also discloses a multidimensional logical rule intelligent mining system for executing the multidimensional logical rule intelligent mining method, including a manufacturing knowledge graph acquisition and normalization module, a multidimensional rule pattern configuration module, a candidate rule body generation module, a large language model rule head inference module, a mixed confidence evaluation module, a high confidence rule screening module, a manufacturing knowledge triple completion module, and a manufacturing knowledge graph update module.

[0055] The advantages and beneficial effects of this invention are as follows:

[0056] First, this invention defines and mines multidimensional logical rules from five perspectives: relational direction, path structure, entity hierarchy, neighborhood context, and parallel structure. It can cover various implicit knowledge completion scenarios in knowledge graph manufacturing, such as reciprocal semantics, chain transmission, macro and micro levels, two-hop context, and shared center structure.

[0057] Second, this invention does not directly input the complete knowledge graph into the large language model for open reasoning. Instead, it first uses the graph structure to generate constrained candidate rule bodies, and then the large language model performs rule head inference, which can reduce the reasoning load and the risk of hallucination output.

[0058] Third, this invention compresses the candidate rule space through chain length control, sub-entity filtering, two-hop pruning, relationship whitelisting, and shared central entity pairing, enabling multi-dimensional rule mining to adapt to industrial-grade manufacturing knowledge graphs.

[0059] Fourth, this invention simultaneously utilizes graph structure support and expert scoring from large language models for hybrid confidence evaluation, and uses multi-round expert scoring dispersion to control rule stability, thereby improving the reliability of high-confidence logical rule selection.

[0060] Fifth, the newly generated manufacturing knowledge triples can be traced back to the corresponding candidate rule body, rule head, confidence level and inference basis, which is conducive to the evolution of manufacturing knowledge graph from static knowledge storage to a self-refining intelligent knowledge management system. Attached Figure Description

[0061] Figure 1 This is a general flowchart of the method in the embodiments of the present invention.

[0062] Figure 2 This is a diagram of the MKG self-refinement framework driven by multi-dimensional logic rules in an embodiment of the present invention.

[0063] Figure 3 This is a flowchart of the two-stage mining process for symmetry rules in an embodiment of the present invention.

[0064] Figure 4 This is a flowchart of transitivity rule mining in an embodiment of the present invention.

[0065] Figure 5 This is a flowchart of hierarchical propagation and horizontal association rule mining in an embodiment of the present invention.

[0066] Figure 6 This is a flowchart of the parallel association rule mining process in an embodiment of the present invention.

[0067] Figure 7 This is a flowchart of the mixed confidence evaluation and high confidence rule screening in an embodiment of the present invention.

[0068] Figure 8This is a schematic diagram of the system structure in an embodiment of the present invention. Detailed Implementation

[0069] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0070] The present invention provides a multi-dimensional logical rule intelligent mining method for self-refinement of manufacturing knowledge graphs, comprising the following steps:

[0071] Step S1: Obtain the original manufacturing knowledge graph, and perform structured representation and normalization processing on the manufacturing knowledge entities, manufacturing knowledge relations and manufacturing knowledge triples in the original manufacturing knowledge graph to obtain the entity set, relation set and observed manufacturing knowledge fact triple set;

[0072] Step S101: Read the node, edge and triple data in the manufacturing knowledge graph, and represent the manufacturing object, material, process, equipment, parameter, performance, organization and product as a unified manufacturing knowledge entity;

[0073] Step S102: The semantics of attributes, composition, use, belonging, influence, cause, generation, requirement and association between entities are uniformly represented as manufacturing knowledge relations;

[0074] Step S103: Represent any manufacturing knowledge fact as a manufacturing knowledge triple consisting of a head entity, a relation, and a tail entity;

[0075] Step S104 involves normalizing synonymous entities, synonymous relations, redundant spaces, invalid symbols, and repeated triples to obtain a raw manufacturing knowledge graph that can be used for logical rule mining.

[0076] The original manufacturing knowledge graph is represented as a directed multi-relationship graph:

[0077]

[0078] In the formula, Represents a set of knowledge entities for manufacturing. This represents the creation of a set of knowledge relationships. Let represent the set of triples representing observed manufactured knowledge facts. An arbitrary manufactured knowledge triple is represented as... ,in Indicates the head entity. Indicates the tail entity. This represents the manufacturing knowledge relationship between the head entity and the tail entity.

[0079] Step S2: Based on the preset multi-dimensional logical rule pattern, candidate rule bodies are generated for the original manufacturing knowledge graph, resulting in symmetric rule candidate bodies, transitive rule candidate bodies, hierarchical propagation rule candidate bodies, horizontal association rule candidate bodies, and parallel association rule candidate bodies.

[0080] The multidimensional logical rules include symmetry rules, transitivity rules, hierarchical propagation rules, lateral association rules, and parallel association rules. Symmetry rules are used to complete inverse triples based on reciprocal or self-symmetric relationships; transitivity rules are used to infer potential relationships between the first and last entities of a path based on a chain of consecutive triples; hierarchical propagation rules are used to propagate knowledge related to composite entities to sub-entities with independent semantics; lateral association rules are used to discover potential relationships between sub-entities and neighboring entities from the two-hop context of a composite entity; and parallel association rules are used to infer implicit relationships between non-central entities based on parallel structures sharing the same central entity.

[0081] Step S201: Extract candidate relations from the relation set, and use a large language model to identify inverse relation pairs and self-symmetric relations in the candidate relations;

[0082] Step S202: Based on the identified reciprocal relationship pairs and self-symmetric relationships, match the corresponding triples in the original manufacturing knowledge graph to form the first round of symmetry rule candidate rule bodies;

[0083] Step S203: Construct a few-sample example based on the identified high-confidence reciprocal relation pairs, and input the unmatched relation set, relation semantics, and the few-sample example into the large language model to identify potential reciprocal relation pairs;

[0084] Step S204: Generate a second round of candidate rules for symmetry rules based on the potential reciprocal relationship, and merge it with the first round of candidate rules for symmetry rules.

[0085] Candidate rule bodies are not openly enumerated across the entire graph by the large language model, but are generated in a targeted manner based on the graph structure corresponding to different logical rules. Specifically, symmetric rules generate candidate rule bodies based on relational direction matching, transitive rules generate candidate rule bodies based on continuous triple chains, hierarchical propagation rules generate candidate rule bodies based on composite entity splitting and adjacent triples, lateral association rules generate candidate rule bodies based on the two-hop neighborhood of composite entities, and parallel association rules generate candidate rule bodies based on triple pairs of entities sharing a central entity.

[0086] The generation of candidate rule bodies for transitivity rules includes the following steps:

[0087] Step S211: Enumerate consecutive triplet chains in the original manufacturing knowledge graph, wherein the head entity of the next triplet in the consecutive triplet chain is the same as the tail entity of the previous triplet.

[0088] Step S212: Control the length of the candidate triplet chain according to the preset minimum chain length and maximum chain length, and filter out the loop path containing duplicate entities.

[0089] Step S213: Select the triplet chain that satisfies the chain length constraint and the entity non-repetition constraint as the transitive rule candidate rule body;

[0090] Step S214: Input the transitive rule candidate rule body, relation set and manufacturing domain information into the large language model to infer the candidate rule head between the first entity and the last entity of the path.

[0091] The generation of candidate rule bodies for hierarchical propagation rules includes the following steps:

[0092] Step S221: Identify candidate composite entities based on the numerical units, connectors, parentheses, component codes, process flow words, and combination structures in the manufacturing entity text;

[0093] Step S222: Use a large language model to split the candidate composite entity into sub-entities with independent manufacturing semantics;

[0094] Step S223: Filter the sub-entities obtained from the splitting based on the sub-entity length, numerical attribute, unit attribute, tolerance attribute, and domain semantic integrity.

[0095] Step S224: Retrieve adjacent triples of the composite entity in the original manufacturing knowledge graph, and construct a candidate rule body for hierarchical propagation rules consisting of the composite entity, effective sub-entities, and external adjacent entities;

[0096] Step S225: Input the candidate rule body of the hierarchical propagation rule into the large language model to infer the candidate rule head between the effective sub-entity and the external adjacent entity.

[0097] The generation of candidate rule bodies for horizontal association rules includes the following steps:

[0098] Step S231: Reuse the valid composite entities and valid sub-entities obtained during the hierarchical propagation rule mining process;

[0099] Step S232: Obtain the one-hop neighborhood entity of the effective composite entity, and expand the two-hop neighborhood entity based on the one-hop neighborhood entity to obtain the two-hop neighborhood entity;

[0100] Step S233: Filter the second-hop relationship based on the preset relationship whitelist to obtain a two-hop candidate triple chain that conforms to the semantics of the manufacturing domain;

[0101] Step S234: Construct a candidate rule body for horizontal association rules, consisting of effective sub-entities, composite entities, one-hop neighborhood entities, and two-hop neighborhood entities;

[0102] Step S235: Input the candidate rule body of the horizontal association rule into the large language model to infer the candidate rule head between the effective sub-entity and the two-hop neighbor entity.

[0103] The parallel association rule candidate rule body generation includes the following steps:

[0104] Step S241: Enumerate triplet pairs that share the same central entity in the original manufacturing knowledge graph;

[0105] Step S242: Based on the position of the central entity in the two triples, divide the triples with the shared central entity into homogeneous parallel structures and homogeneous parallel structures.

[0106] Step S243: Detect whether there is an explicit relationship between the non-central entities in the two triples;

[0107] Step S244: For parallel structures with explicit relationships, they are used as candidate rules with computable objective confidence; for parallel structures without explicit relationships, they are used as candidate rule bodies to be inferred.

[0108] Step S245: The high-confidence existing parallel rules are used as few-sample examples and input together with the candidate rule body to be inferred into the large language model to infer the candidate rule heads between non-central entities.

[0109] Step S3: Input the candidate rule body, manufacturing domain information, the relation set and the corresponding logical rule pattern into the large language model, infer the candidate rule head based on the corresponding prompting strategy, and obtain the candidate logical rules;

[0110] The prompting strategy for inferring candidate rule heads in a large language model includes the prompting strategy's identity positioning, task description, precautions, and output specifications. Specifically, the prompting strategy's identity positioning defines the large language model as an expert in mining logical rules for knowledge graphs; the task description defines the large language model's ability to infer candidate rule heads based on candidate rule bodies and logical rule patterns; the precautions define the large language model's prohibition on modifying input rule bodies, prohibition on outputting duplicate rule heads, prioritizing the selection of normative relations from existing relation sets, and determining the direction of rule heads based on entity semantics; and the output specifications define the large language model's ability to output rule types, rule bodies, rule heads, and inference basis in a structured format.

[0111] The objective confidence level of the candidate logic rule is calculated using the following formula:

[0112]

[0113] In the formula, Represents candidate logic rules, This indicates the number of times a rule appears in the original manufacturing knowledge graph. This indicates the number of times that both the rule body and the rule header are true simultaneously.

[0114] Step S4: Perform a mixed confidence evaluation on the candidate logical rules, wherein objective confidence is calculated based on the structural support of the rule body and rule head in the original manufacturing knowledge graph, expert confidence and expert rating dispersion are calculated based on multi-round expert evaluation of the large language model, and a mixed confidence is calculated based on the objective confidence and the expert confidence; the candidate logical rules are screened according to the preset confidence threshold and expert rating dispersion threshold to obtain high-confidence logical rules;

[0115] Step S401: Count the number of times the rule body of the candidate logical rule appears in the original manufacturing knowledge graph, and the number of times the rule body and rule head are both true, and calculate the objective confidence level.

[0116] Step S402: Using a large language model, candidate logical rules are scored in multiple rounds from four dimensions: logical rationality, domain applicability, semantic consistency, and application value, and expert confidence is calculated.

[0117] Step S403: Calculate the standard deviation of the expert scores over multiple rounds to obtain the dispersion of the expert scores;

[0118] Step S404: Obtain the mixed confidence level based on the weighted result of the objective confidence level and the expert confidence level;

[0119] Step S405: When the mixed confidence level is not lower than a preset confidence threshold and the expert rating dispersion is not higher than a preset dispersion threshold, the candidate logic rule is determined as a high-confidence logic rule.

[0120] The formula for calculating the expert confidence level of the candidate logic rule is as follows:

[0121]

[0122] In the formula, This indicates the number of expert evaluation rounds for the large language model. Indicates the first Round of expert evaluation of candidate logic rules The rating given.

[0123] The formula for calculating the expert scoring dispersion of the candidate logic rules is as follows:

[0124]

[0125] In the formula, Representation of candidate logic rules The average of multiple rounds of expert scores.

[0126] The formula for calculating the mixed confidence level of the candidate logic rule is as follows:

[0127]

[0128] In the formula, Indicates the objective confidence level weight. This indicates the weight of expert confidence.

[0129] The set of high-confidence logic rules is represented as follows:

[0130]

[0131] In the formula, Indicates the confidence threshold. This represents the threshold for the dispersion of expert ratings.

[0132] Step S5: Apply the high-confidence logic rules to the original manufacturing knowledge graph to generate new manufacturing knowledge triples. Then, normalize, detect conflicts, remove duplicates, and merge the new manufacturing knowledge triples to obtain a self-refined manufacturing knowledge graph.

[0133] The self-refined manufacturing knowledge graph is represented as follows:

[0134]

[0135] In the formula, This represents a newly added set of manufacturing knowledge entities. This represents a newly added set of manufacturing knowledge relationships. This represents the set of newly added manufacturing knowledge triples generated by high-confidence logic rules and after deduplication and merging.

[0136] This invention also provides a multi-dimensional logical rule intelligent mining system for self-refinement of manufacturing knowledge graphs, including a manufacturing knowledge graph acquisition and normalization module, a multi-dimensional rule pattern configuration module, a candidate rule body generation module, a large language model rule head inference module, a mixed confidence evaluation module, a high-confidence rule screening module, a manufacturing knowledge triple completion module, and a manufacturing knowledge graph update module. The system is used to execute the above methods.

[0137] like Figure 1 As shown, this embodiment provides a multi-dimensional logical rule intelligent mining method for self-refinement of manufacturing knowledge graphs, including steps such as acquiring and structurally representing the original manufacturing knowledge graph, generating multi-dimensional candidate rule bodies, intelligent inference of rule heads from a large language model, mixed confidence evaluation, high-confidence rule screening, and graph completion and self-refinement.

[0138] like Figure 2As shown, this invention takes the original manufacturing knowledge graph as input, mines multi-dimensional logical rules through a large language model semantic inference engine, and then applies the filtered high-confidence logical rules back to the original manufacturing knowledge graph to form a self-refined manufacturing knowledge graph. Different rule types correspond to different graph structures and business semantics. Among them, symmetric rules focus on relationship direction completion, transitive rules focus on path beginning and end completion, hierarchical propagation rules focus on composite entity sinking, lateral association rules focus on two-hop neighborhood connections, and parallel association rules focus on implicit relationship inference under the shared center structure.

[0139] In one specific embodiment, a steel manufacturing knowledge graph is used as the subject of the description. It should be noted that this invention is not limited to the field of steel manufacturing, but can also be applied to equipment manufacturing, process manufacturing, discrete manufacturing, and other industrial scenarios with a manufacturing knowledge graph foundation. The composition of the original manufacturing knowledge graph in this embodiment is shown in Table 1.

[0140] Table 1: Composition of the original manufacturing knowledge graph in the embodiments

[0141]

[0142] Step S1: Acquisition and Structured Representation of the Original Manufacturing Knowledge Graph. Specifically, the original manufacturing knowledge graph is read from a graph database, knowledge file, or structured triple file. Manufacturing objects such as materials, processes, equipment, parameters, performance, defects, organizations, and products are uniformly represented as entities. Semantics such as use, belong, contain, influence, cause, generate, improve, reduce, require, and supply are uniformly represented as relations. Knowledge facts are represented as triples. Synonymous entities, synonymous relations, null relations, redundant symbols, and duplicate triples are normalized.

[0143] In the embodiment, the arbitrary creation knowledge triple is represented as:

[0144]

[0145] In the formula, Indicates the first The head entity of a triplet, Indicates a relationship. This indicates the tail entity.

[0146] Step S2: Generation of Multidimensional Candidate Rule Bodies. Step S2 is used to generate candidate rule bodies from the original manufacturing knowledge graph based on five types of logical rule patterns. The candidate structures and functions of the five types of logical rules are shown in Table 2.

[0147] Table 2: Candidate Structure and Function of Multidimensional Logical Rules

[0148]

[0149] The symmetry rule is expressed as:

[0150]

[0151] in, Denotes a set of candidate symmetric relation pairs, when Time indicates a self-symmetric relationship.

[0152] The transitivity rule is expressed as:

[0153]

[0154] in, Indicates the length of the transmission path. This represents the rule header relationship inferred from the entire transmission path.

[0155] The hierarchical propagation rule is represented as follows:

[0156]

[0157] in, Represents a composite entity. This represents the sub-entities obtained by splitting a composite entity. This represents the external entity adjacent to the composite entity. express yes Sub-entities of.

[0158] Horizontal association rules are represented as follows:

[0159]

[0160] in, This represents a one-hop neighborhood entity of a composite entity. This represents a two-hop neighborhood entity.

[0161] Parallel association rules are represented as follows:

[0162]

[0163] Or it can be expressed as:

[0164]

[0165] in, Represents the shared central entity. and This indicates a non-central entity.

[0166] Step S3: Intelligent Inference of Rule Heads from the Large Language Model. Step S3 is used to transform candidate rule bodies into candidate logical rules. For each type of rule, a corresponding large language model prompting strategy is constructed. The prompting strategy includes identity positioning, task description, precautions, and output specifications.

[0167] Taking the general rule header inference hint strategy as an example, the identity positioning content of the hint strategy includes: You are a domain expert for mining logical rules for creating knowledge graphs, focusing on the {Field_name} creation domain, and familiar with entity semantics, relational semantics and logical rule patterns in creating knowledge graphs.

[0168] The task description of the prompt strategy includes: based on the candidate rule body {Rule_body}, the existing relation set {Relation_set}, and the logical rule pattern {Rule_pattern}, analyze the manufacturing semantics of entities and relations in the rule body, and infer candidate rule heads that conform to the manufacturing domain logic.

[0169] The precautions for the suggestion strategy include: First, the input candidate rule body must not be modified; second, rule header relations should be selected from the existing relation set first. If no suitable relation exists in the existing relation set, a new relation that conforms to the manufacturing domain expression specification should be generated; third, the direction of the rule header should be determined according to the manufacturing semantics; fourth, duplicate rule headers must not be output; and fifth, the output rule headers must be reasonably supported by the candidate rule bodies.

[0170] The output specifications for the suggestion strategy include: returning the data in a structured format, as shown in the template below:

[0171] {

[0172] "rule_type": "rule type",

[0173] "rule_body": ["candidate rule body"],

[0174] "rule_head": {

[0175] "head_entity": "head entity",

[0176] "relation": "relationship"

[0177] "tail_entity": "tail entity"

[0178] },

[0179] "reasoning": "basis for inference"

[0180] }

[0181] Where {Field_name} represents the name of the manufacturing domain, {Rule_body} represents the candidate rule body, {Relation_set} represents the original set of manufacturing knowledge graph relations, and {Rule_pattern} represents the corresponding logical rule pattern.

[0182] Step S4: Symmetry rule mining. For example... Figure 3 As shown, symmetry rule mining employs a two-stage mechanism. The first stage identifies inverse and self-symmetric relation pairs from the original knowledge graph's relation set and performs a first round of completion based on the already matched relation pairs. The second stage uses the high-confidence relation pairs obtained in the first stage as minority sample examples to identify potential symmetric relations in the unmatched relation set, thereby obtaining enhanced completion results.

[0183] If the original manufacturing knowledge graph contains a triple (structural steel, used, galvanized), and the large language model identifies "used" and "applied to" as inverse relationships, then candidate rule heads (galvanized, applied to, structural steel) can be obtained according to the symmetry rule. For the candidate rule heads, a mixed confidence evaluation is performed. Only when the confidence threshold and the expert rating dispersion threshold are met is it written into the self-refined graph as a new manufacturing knowledge triple.

[0184] Step S5: Transitive rule mining. For example... Figure 4 As shown, transitive rule mining first enumerates consecutive triple chains in the original manufacturing knowledge graph. Then, it controls the size of candidate chains based on the minimum and maximum chain lengths and filters out cyclic paths containing duplicate entities. After path filtering, the candidate triple chains are input into the large language model to infer the candidate rule heads between the first entity and the last entity of the path.

[0185] If the original manufacturing knowledge graph contains two triplets (phosphorus, leading to brittleness) and (brittleness, reducing toughness), then according to the transitivity rule, a candidate rule head (phosphorus, reducing toughness) can be deduced. This candidate rule head reflects the indirect influence of alloying elements on material properties, which is helpful in completing the knowledge of material composition, defect mechanisms, and performance effects in the manufacturing knowledge graph.

[0186] Step S6: Hierarchical propagation rule mining. For example... Figure 5 As shown, hierarchical propagation rule mining first identifies composite entities in the manufacturing knowledge graph. Composite entities can be process flows containing multiple steps, material composition descriptions containing multiple elements, technical requirements containing multiple parameter constraints, or entity names formed by combining multiple manufacturing semantic units. After identifying candidate composite entities, a large language model is used to split them into sub-entities with independent manufacturing semantics, and sub-entities lacking independent semantics are removed according to a sub-entity filtering strategy.

[0187] For example, for the "hot-rolled strip steel process flow" and its included process description, the system can identify sub-entities such as "rough rolling" and "finish rolling", and infer fine-grained knowledge related to the sub-entities by combining the adjacent triples of the composite entity.

[0188] Step S7: Horizontal Association Rule Mining. The horizontal association rules reuse the valid composite entities and valid sub-entities obtained from the hierarchical propagation rules. Based on this, the one-hop and two-hop neighbor entities of the composite entities are further obtained. To avoid excessive candidate size due to two-hop expansion, the system sets a relationship whitelist to filter second-hop relationships, retaining only two-hop candidate triple chains that conform to the manufacturing business logic.

[0189] For example, if the original manufacturing knowledge graph contains (40CrNiMo steel, applied to, machinery manufacturing steel) and (machinery manufacturing steel, requires, corrosion resistance), and "Ni" is a valid subentity of 40CrNiMo steel, then the system can infer (Ni, improves, corrosion resistance) based on lateral association rules. This new ternary can make explicit the potential influence relationship between alloying elements and material properties.

[0190] Step S8: Parallel association rule mining. For example... Figure 6 As shown, parallel association rule mining first enumerates triple pairs that share the same central entity, and distinguishes between homologous parallel structures and equivalent parallel structures based on the position of the central entity in the triple. For candidate structures where there is already an explicit relationship between non-central entities, their mixture confidence is calculated; for candidate structures where there is no explicit relationship between non-central entities, high-confidence existing rules are used as few-sample examples to guide the large language model to infer the candidate rule heads between non-central entities.

[0191] For example, if the original manufacturing knowledge graph contains (blast furnace ironmaking, production, molten iron) and (converter steelmaking, use, molten iron), then "molten iron" is a shared central entity, while "blast furnace ironmaking" and "converter steelmaking" are non-central entities. Based on the parallel association rule, (blast furnace ironmaking, supply, converter steelmaking) can be inferred. This new triplet can make the supply collaboration relationships between processes explicit.

[0192] The identity positioning content of the composite entity splitting prompt strategy in the hierarchical propagation rules includes: You are a composite entity splitting expert in the manufacturing field, and you can split composite entities into sub-entities with independent semantics based on manufacturing semantics.

[0193] The task description for the hint strategy includes: analyzing the structure of a given composite entity {Composite_entity} and breaking it down into sub-entities that can independently represent materials, processes, equipment, performance, parameters, or product objects.

[0194] The precautions for the suggestion strategy include: first, retaining sub-entities with independent manufacturing semantics; second, not outputting pure numerical values, pure units, single punctuation marks, and fragments without entity meaning; third, the splitting result must not be separated from the input composite entity; and fourth, the splitting result should conform to the expression habits of the manufacturing domain.

[0195] The output specifications for the prompting strategy include:

[0196] {

[0197] "composite_entity": "composite entity",

[0198] "sub_entities": ["sub-entity1", "sub-entity2"],

[0199] "reasoning": "basis for splitting"

[0200] }

[0201] Step S9: Combined confidence level evaluation and high-confidence rule screening. For example... Figure 7 As shown, after candidate logical rules are generated, the system evaluates them using a mixed confidence score. For candidate rules whose co-occurrence of rule bodies and rule heads can be statistically analyzed in the original manufacturing knowledge graph, the objective confidence score is calculated. For new relations that lack sufficient structural support but have potential manufacturing semantics, multiple rounds of expert evaluation are performed using a large language model to obtain expert confidence scores and expert score dispersion. Then, the objective confidence score and expert confidence score are weighted to obtain the mixed confidence score.

[0202] The identity positioning content of the expert evaluation prompt strategy for large language models includes: You are an expert in evaluating logical rules of a knowledge graph in the manufacturing domain, and you can evaluate candidate logical rules from the aspects of logical rationality, domain applicability, semantic consistency and application value.

[0203] The task description of the prompt strategy includes: based on the candidate rule body and candidate rule head {Rule_body_and_head}, determine whether it constitutes a reliable logical rule for manufacturing a knowledge graph, and give a score of 0 to 5.

[0204] The key considerations for the strategy include: first, evaluating whether there is clear logical support between the rule body and the rule header; second, evaluating whether the rule conforms to common sense and professional semantics in the manufacturing field; third, evaluating whether the relationship between the rule headers is accurately expressed; fourth, evaluating whether the rule has application value for map completion; and fifth, scoring should be cautious, and excessively high scores should not be given due to similar text expression.

[0205] The output specifications for the prompting strategy include:

[0206] {

[0207] "score": "0-5.0",

[0208] "evaluation": "Reasons for the evaluation",

[0209] "decision": "accept / reject"

[0210] }

[0211] In this embodiment, when the mixed confidence level of a candidate logical rule is not lower than a preset threshold, and the standard deviation of multiple rounds of expert scoring is not higher than a preset threshold, the candidate logical rule is determined as a high-confidence logical rule. This mechanism can avoid the influence of single-round large language model evaluation bias on the rule selection results.

[0212] Step S10: Manufacturing Knowledge Graph Completion and Self-Refining. After completing the high-confidence logical rule screening, the system transforms the rule header of each high-confidence logical rule into a newly added manufacturing knowledge triple. For each newly added manufacturing knowledge triple, it first checks whether it already exists in the original manufacturing knowledge graph; if it already exists, it is not written again; if it does not exist, further relation name normalization, entity name normalization, and semantic conflict detection are performed. The newly added triples after normalization, deduplication, and conflict detection are merged with the original triples to form the self-refined manufacturing knowledge graph.

[0213] Representative mining examples of the five types of logical rules are shown in Table 3.

[0214] Table 3: Examples of Mining Five Types of Logical Rules

[0215]

[0216] In this embodiment, five types of logical rule mining and completion are performed based on a single large language model, and the results are shown in Table 4. It should be noted that Table 4 is only used to illustrate that the method of the present invention can perform rule mining and graph completion, and is not intended to limit the present invention to a specific large language model or a specific data scale.

[0217] Table 4: Single-model rule completion results in the examples

[0218]

[0219] As shown in Table 4, with the original manufacturing knowledge graph containing 912 triples, this embodiment generates 874 candidate new triples through five types of logical rules. After deduplication, 866 new manufacturing knowledge triples are obtained. After merging, the self-refined manufacturing knowledge graph contains 1778 triples. Each new triple can be traced back to its corresponding candidate rule body, candidate rule head, confidence level, and inference basis.

[0220] System architecture. For example... Figure 8 As shown, the present invention also provides a multi-dimensional logical rule intelligent mining system for self-refinement of manufacturing knowledge graphs, including a manufacturing knowledge graph acquisition and standardization module, a multi-dimensional rule pattern configuration module, a candidate rule body generation module, a large language model rule head inference module, a mixed confidence evaluation module, a high confidence rule screening module, a manufacturing knowledge triple completion module, and a manufacturing knowledge graph update module.

[0221] The Manufacturing Knowledge Graph Acquisition and Normalization module acquires the original manufacturing knowledge graph and performs normalization processing on entities, relations, and triples. The Multidimensional Rule Pattern Configuration module configures rule patterns and related parameters for symmetric rules, transitive rules, hierarchical propagation rules, lateral association rules, and parallel association rules. The Candidate Rule Body Generation module generates candidate rule bodies from the original manufacturing knowledge graph based on the rule patterns. The Large Language Model Rule Head Inference module calls the large language model and infers candidate rule heads based on a prompting strategy. The Mixed Confidence Evaluation module calculates objective confidence, expert confidence, and mixed confidence. The High-Confidence Rule Filtering module filters high-confidence logical rules based on threshold conditions. The Manufacturing Knowledge Triple Completion module transforms the rule heads of high-confidence logical rules into new manufacturing knowledge triples. The Manufacturing Knowledge Graph Update module normalizes, detects conflicts, removes duplicates, and merges the new triples to obtain a self-refined manufacturing knowledge graph.

[0222] The system of this invention can be deployed on a manufacturing enterprise knowledge management platform, enterprise resource planning system, manufacturing execution system, or a standalone server for manufacturing knowledge services. The system can read the original manufacturing knowledge graph from a graph database via an interface, or it can read CSV, JSON, RDF, or other structured triple files. The self-refined manufacturing knowledge graph output by the system can be further used for knowledge retrieval, process analysis, quality traceability, customer demand response, and business decision support.

[0223] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent mining of multi-dimensional logical rules, characterized in that, Includes the following steps: Step S1: Obtain the original manufacturing knowledge graph, and perform structured representation and normalization processing on the manufacturing knowledge entities, manufacturing knowledge relations and manufacturing knowledge triples in the original manufacturing knowledge graph to obtain the entity set, relation set and observed manufacturing knowledge fact triple set; Step S2: Based on the preset multi-dimensional logical rule pattern, candidate rule bodies are generated for the original manufacturing knowledge graph, resulting in symmetric rule candidate bodies, transitive rule candidate bodies, hierarchical propagation rule candidate bodies, horizontal association rule candidate bodies, and parallel association rule candidate bodies. Step S3: Input the candidate rule body, manufacturing domain information, the relation set and the corresponding logical rule pattern into the large language model, infer the candidate rule head based on the corresponding prompting strategy, and obtain the candidate logical rules; Step S4: Perform a mixed confidence evaluation on the candidate logical rules, wherein objective confidence is calculated based on the structural support of the rule body and rule head in the original manufacturing knowledge graph, expert confidence and expert rating dispersion are calculated based on multi-round expert evaluation of the large language model, and a mixed confidence is calculated based on the objective confidence and the expert confidence; the candidate logical rules are screened according to the preset confidence threshold and expert rating dispersion threshold to obtain high-confidence logical rules; Step S5: Apply the high-confidence logic rules to the original manufacturing knowledge graph to generate new manufacturing knowledge triples. Then, normalize, detect conflicts, remove duplicates, and merge the new manufacturing knowledge triples to obtain a self-refined manufacturing knowledge graph.

2. The multidimensional logical rule intelligent mining method according to claim 1, characterized in that: Step S1 includes the following steps: Step S101: Read the node, edge and triple data in the manufacturing knowledge graph, and represent the manufacturing object, material, process, equipment, parameter, performance, organization and product as a unified manufacturing knowledge entity; Step S102: The semantics of attributes, composition, use, belonging, influence, cause, generation, requirement and association between entities are uniformly represented as manufacturing knowledge relations; Step S103: Represent any manufacturing knowledge fact as a manufacturing knowledge triple consisting of a head entity, a relation, and a tail entity; Step S104 involves normalizing synonymous entities, synonymous relations, redundant spaces, invalid symbols, and repeated triples to obtain a raw manufacturing knowledge graph that can be used for logical rule mining.

3. The intelligent mining method for multidimensional logical rules according to claim 1, characterized in that: The generation of symmetry rule candidate rule bodies in step S2 includes the following steps: Step S201: Extract candidate relations from the relation set, and use a large language model to identify inverse relation pairs and self-symmetric relations in the candidate relations; Step S202: Based on the identified reciprocal relationship pairs and self-symmetric relationships, match the corresponding triples in the original manufacturing knowledge graph to form the first round of symmetry rule candidate rule bodies; Step S203: Construct a few-sample example based on the identified high-confidence reciprocal relation pairs, and input the unmatched relation set, relation semantics, and the few-sample example into the large language model to identify potential reciprocal relation pairs; Step S204: Generate a second round of candidate rules for symmetry rules based on the potential reciprocal relationship, and merge it with the first round of candidate rules for symmetry rules.

4. The intelligent mining method for multidimensional logical rules according to claim 1, characterized in that: The generation of transitive rule candidate rule bodies in step S2 includes the following steps: Step S211: Enumerate consecutive triplet chains in the original manufacturing knowledge graph, wherein the head entity of the next triplet in the consecutive triplet chain is the same as the tail entity of the previous triplet. Step S212: Control the length of the candidate triplet chain according to the preset minimum chain length and maximum chain length, and filter out the loop path containing duplicate entities. Step S213: Select the triplet chain that satisfies the chain length constraint and the entity non-repetition constraint as the transitive rule candidate rule body; Step S214: Input the transitive rule candidate rule body, relation set and manufacturing domain information into the large language model to infer the candidate rule head between the first entity and the last entity of the path.

5. The intelligent mining method for multidimensional logical rules according to claim 1, characterized in that: The generation of candidate rule bodies for hierarchical propagation rules in step S2 includes the following steps: Step S221: Identify candidate composite entities based on the numerical units, connectors, parentheses, component codes, process flow words, and combination structures in the manufacturing entity text; Step S222: Use a large language model to split the candidate composite entity into sub-entities with independent manufacturing semantics; Step S223: Filter the sub-entities obtained from the splitting based on the sub-entity length, numerical attribute, unit attribute, tolerance attribute, and domain semantic integrity. Step S224: Retrieve adjacent triples of the composite entity in the original manufacturing knowledge graph, and construct a candidate rule body for hierarchical propagation rules consisting of the composite entity, effective sub-entities, and external adjacent entities; Step S225: Input the candidate rule body of the hierarchical propagation rule into the large language model to infer the candidate rule head between the effective sub-entity and the external adjacent entity.

6. The intelligent mining method for multidimensional logical rules according to claim 1, characterized in that: The generation of candidate rule bodies for horizontal association rules in step S2 includes the following steps: Step S231: Reuse the valid composite entities and valid sub-entities obtained during the hierarchical propagation rule mining process; Step S232: Obtain the one-hop neighborhood entity of the effective composite entity, and expand the two-hop neighborhood entity based on the one-hop neighborhood entity to obtain the two-hop neighborhood entity; Step S233: Filter the second-hop relationship based on the preset relationship whitelist to obtain a two-hop candidate triple chain that conforms to the semantics of the manufacturing domain; Step S234: Construct a candidate rule body for horizontal association rules, consisting of effective sub-entities, composite entities, one-hop neighborhood entities, and two-hop neighborhood entities; Step S235: Input the candidate rule body of the horizontal association rule into the large language model to infer the candidate rule head between the effective sub-entity and the two-hop neighbor entity.

7. The intelligent mining method for multidimensional logical rules according to claim 1, characterized in that: The parallel association rule candidate rule body generation in step S2 includes the following steps: Step S241: Enumerate triplet pairs that share the same central entity in the original manufacturing knowledge graph; Step S242: Based on the position of the central entity in the two triples, divide the triples with the shared central entity into homogeneous parallel structures and homogeneous parallel structures. Step S243: Detect whether there is an explicit relationship between the non-central entities in the two triples; Step S244: For parallel structures with explicit relationships, they are used as candidate rules with computable objective confidence; for parallel structures without explicit relationships, they are used as candidate rule bodies to be inferred. Step S245: The high-confidence existing parallel rules are used as few-sample examples and input together with the candidate rule body to be inferred into the large language model to infer the candidate rule heads between non-central entities.

8. The intelligent mining method for multidimensional logical rules according to claim 1, characterized in that: In step S3, the prompting strategy for the large language model to infer candidate rule heads includes the identity positioning of the prompting strategy, the task description of the prompting strategy, the precautions of the prompting strategy, and the output specification of the prompting strategy. Specifically, the identity positioning of the prompting strategy is used to define the large language model as an expert in mining logical rules for creating knowledge graphs; the task description of the prompting strategy is used to define the large language model to infer candidate rule heads based on candidate rule bodies and logical rule patterns; the precautions of the prompting strategy are used to define that the large language model must not modify the input rule body, must not output duplicate rule heads, must preferentially select normative relations from the existing relation set, and determine the direction of the rule head based on entity semantics; the output specification of the prompting strategy is used to define that the large language model outputs the rule type, rule body, rule head, and inference basis in a structured format.

9. The intelligent mining method for multidimensional logical rules according to claim 1, characterized in that: Step S4 includes the following steps: Step S401: Count the number of times the rule body of the candidate logical rule appears in the original manufacturing knowledge graph, and the number of times the rule body and rule head are both true, and calculate the objective confidence level. Step S402: Using a large language model, candidate logical rules are scored in multiple rounds from four dimensions: logical rationality, domain applicability, semantic consistency, and application value, and expert confidence is calculated. Step S403: Calculate the standard deviation of the expert scores over multiple rounds to obtain the dispersion of the expert scores; Step S404: Obtain the mixed confidence level based on the weighted result of the objective confidence level and the expert confidence level; Step S405: When the mixed confidence level is not lower than a preset confidence threshold and the expert rating dispersion is not higher than a preset dispersion threshold, the candidate logic rule is determined as a high-confidence logic rule.

10. A multi-dimensional logical rule intelligent mining system, used to execute the multi-dimensional logical rule intelligent mining method as described in any one of claims 1-9, characterized in that, It includes modules for manufacturing knowledge graph acquisition and normalization, multi-dimensional rule pattern configuration, candidate rule body generation, large language model rule head inference, mixed confidence evaluation, high confidence rule screening, manufacturing knowledge triple completion, and manufacturing knowledge graph update.

Images