Method for constructing aerospace manufacturing knowledge graph based on heterogeneous data joint extraction mode

Abstract

The invention discloses a construction method of an aerospace manufacturing knowledge graph based on a heterogeneous data joint extraction mode. The construction method comprises the following steps: designing and constructing a knowledge model of the knowledge graph by using an automatic modeling tool based on a top-down theory; performing knowledge graph bottom layer storage in a manner of combining a table structure and a graph structure; knowledge extraction and knowledge mining are carried out from structured, semi-structured and unstructured data; carrying out data mode layer fusion and data layer fusion on data from different sources; knowledge discovery and knowledge graph completion are carried out based on rapid knowledge reasoning of a graph neural network; and carrying out intelligent question answering and visual decision support based on semantic search in combination with a specific application form of the knowledge graph and aerospace manufacturing field data and business scenes. According to the method, challenges of various knowledge sources and data formats on knowledge extraction can be handled, so that the aerospace manufacturing knowledge graph can be applied to the ground, and the method has mobility, scientificity and seriousness.

Description

Technical Field

[0001] The present invention belongs to the technical field of knowledge management and mining in the field of aerospace intelligent manufacturing, and specifically relates to a method for constructing an aerospace manufacturing knowledge graph based on a heterogeneous data joint extraction mode. Background Art

[0002] Knowledge graph is a new method of knowledge representation and a new idea of ​​knowledge management, and is no longer limited to general fields such as search and knowledge question answering. A domain knowledge graph construction method focusing on specific industrial scenarios and vertical knowledge can provide in-depth data analysis and decision support for specific industries.

[0003] With the continuous progress and development of knowledge and data science, knowledge graphs in industrial verticals are often used to help companies better understand and optimize their production processes and promote interdisciplinary innovation and decision-making. The aerospace manufacturing field has a strict and rich data model, and is in urgent need of a knowledge management and decision support tool to promote knowledge integration and innovation. However, building a knowledge graph in this field has higher requirements for the richness, depth, and accuracy of aerospace manufacturing knowledge. Traditional knowledge graph construction methods are difficult to use in complex industrial scenarios. Due to the extremely rich and complex sources of knowledge, even if a knowledge graph can be built for a certain business, it is difficult to migrate to other businesses in the field, and existing knowledge graph construction methods often ignore the mining of knowledge decision-making capabilities. Aiming at the difficulties in the field of aerospace manufacturing, the present invention proposes a method for constructing an aerospace manufacturing knowledge graph based on a joint extraction mode of heterogeneous data. The method uses a variety of deep learning algorithms to quickly extract and reason about structured, semi-structured, and unstructured complex data sources, solving the problem of difficult knowledge acquisition in complex industrial scenarios in the field of aerospace manufacturing. Summary of the invention

[0004] In view of the deficiencies in the above-mentioned prior art, the present invention proposes a method for constructing an aerospace manufacturing knowledge graph based on a joint extraction mode of heterogeneous data, and uses a variety of deep learning algorithms to establish a modular knowledge graph life cycle for rapid extraction of complex domain knowledge and construction and application of knowledge graphs in a variety of aerospace manufacturing scenarios. The top-down design model improves the professionalism and seriousness of the knowledge graph to provide reliable decision support to production line engineers and operators.

[0005] In order to achieve the above technical objectives, the present invention provides the following technical solutions:

[0006] A method for constructing an aerospace manufacturing knowledge graph based on a joint extraction mode of heterogeneous data specifically includes the following steps:

[0007] S1. Design and construct the knowledge model of the knowledge graph using automated modeling tools based on top-down theory;

[0008] S2, combining table structure and graph structure to store the underlying knowledge graph;

[0009] S3, knowledge extraction and knowledge mining from structured, semi-structured and unstructured data;

[0010] S4, perform data pattern layer fusion and data instance layer fusion on data from different sources;

[0011] S5. Fast knowledge reasoning based on graph neural networks for knowledge discovery and knowledge graph completion;

[0012] S6. Combine the unique application forms of knowledge graphs with data and business scenarios in the aerospace manufacturing field to provide intelligent question-answering and visual decision support based on semantic search.

[0013] Furthermore, step S1 specifically includes:

[0014] S11. Design the aerospace manufacturing knowledge model based on the top-down theory, and define the ontology concept classes, relationships, attributes and instances of the knowledge graph under the constraint of expert knowledge; the ontology concept classes are the basic elements that constitute the knowledge graph structure, which include entities, classes, attributes, relationships, etc., and are used to organize and express the knowledge system in the field of aerospace manufacturing;

[0015] S12. Use automated modeling tools to shield specific ontology description languages ​​and quickly construct ontology models in the aerospace manufacturing field at the conceptual level.

[0016] Furthermore, step S2 specifically includes:

[0017] S21. Use a relational database to store the knowledge graph knowledge model metadata, and store ontology definitions, categories, attributes, relationships, etc. in the form of data tables;

[0018] S22. Use the resource description framework RDF to store triple data of the knowledge graph, i.e., entities, attributes, and relationship instances, and synchronously map the triple data to the graph database;

[0019] S23. Establish relational database, and synchronize data between RDF and graph database.

[0020] Furthermore, step S3 specifically includes:

[0021] S31, for structured aerospace manufacturing data, use R2RML language to directly map relational data into RDF datasets;

[0022] S32. For semi-structured and unstructured aerospace manufacturing data, establish a multi-task learning model that can simultaneously extract entities, relations, and events: accept information in the text through a unified input layer and convert it into word embeddings, use bidirectional long short-term memory networks and transformers in the shared layer to capture word order and context information, apply self-attention or hierarchical attention mechanism after the shared layer to focus on the most task-related parts of the text, design entity extraction layer, relation extraction layer, and event extraction layer after the shared layer and attention mechanism are fused, and define a multi-task joint loss function to optimize the multi-task learning model;

[0023] S33, use the probability model to perform entity disambiguation to solve the entity reference problem in the text, by defining the feature function f(e j ,c j )Entity e j and category c j Mapped to a feature vector, for the Markov model, define the state transition probability P(cj|cj-1), calculate the entity e i Belongs to category c j The probability P(c j |e i ), choose so that P(c j |e i ) The largest category c j As e i The disambiguation results are iteratively optimized using the cross entropy loss function.

[0024] More specifically, in step S32, the multi-task joint loss function optimizes the multi-task learning model as follows:

[0025] Choose θ that minimizes the multi-task joint loss function m is the optimal parameter of the multi-task learning model, and the formula is expressed as:

[0026]

[0027] Among them, θ m * represents the optimal parameters of the multi-task learning model, L total represents a multi-task joint loss function; the definition of the multi-task joint loss function is as follows:

[0028] L total =L entity (Y entity ,Y entity_true )+L relation (Y relation ,Y relation_true )+L event (Y event ,Yevent_true );

[0029] Among them, L entity , L relation and L event Represents independent loss functions for entities, relations, and events respectively; Y entity_true , Y relation_true and Y event_true Represent the true values ​​of entities, relations, and events respectively; Y entity , Y relation and Y event Represents predicted values ​​of entities, relations, and events;

[0030] The predicted value is obtained by extracting the function, and the formula is expressed as:

[0031] Y entity =f entity (H);

[0032] Y relation =f relation (H);

[0033] Y event =f event (H);

[0034] Among them, f entity 、f relation and f event Represent the extraction functions of entities, relations, and events respectively; H is the fusion output of the sharing layer and the attention layer, and its formula is expressed as:

[0035] H=Bi-LSTM(X)+Attention(Bi-LSTM(X));

[0036] Among them, X represents the input information after embedding, Bi-LSTM represents the bidirectional long short-term memory network of the shared layer, and Attention represents the attention function of the attention layer.

[0037] Furthermore, step S4 specifically includes:

[0038] S41. At the data model layer, expert knowledge and high-confidence structured knowledge mapping from aerospace manufacturing are used to merge concepts, merge concept hierarchical relationships, and merge concept attribute definitions.

[0039] S42. At the data instance layer, a deep learning-based method is used to identify and match the same entities in different data sources. The text is converted into feature word vectors through a deep learning model, and instance matching is achieved based on cosine similarity. Optimization and model training are performed through contrast loss. After successful matching, different instances of the same entity are merged into a single entity.

[0040] More specifically, the specific process of optimizing and training the model by contrast loss in step S42 is as follows:

[0041] Choose θ that minimizes the contrast loss c To compare the optimal parameters of the learning model, the formula is expressed as:

[0042]

[0043] Among them, θ c * represents the optimal parameters of the contrastive learning model, L is the contrastive loss, and its calculation formula is:

[0044]

[0045] Among them, y i is a weight term to ensure that only negative samples contribute to the loss function, and are the positive and negative samples of the i-th instance pair in the batch, respectively, and f is a deep learning model used to convert entities into embedding vectors.

[0046] Furthermore, step S5 specifically includes:

[0047] S51. Train the graph neural network model, represent the entities and relationships in the knowledge graph as a graph structure, where entities are nodes and relationships are edges. Update the state of each node by aggregating the information of neighboring nodes to complete information transfer. Use the embedded representation of entities and relationships learned by GNN to perform knowledge reasoning, and use the loss function to supervise the training of the model.

[0048] S52. Use the trained graph neural network model to perform knowledge reasoning and knowledge graph completion when building and maintaining the knowledge graph under the supervision of expert knowledge.

[0049] More specifically, the loss function used in step S51 to train the supervised model is:

[0050] Select θ with the minimum loss as the optimal parameter of the supervised model, and the formula is expressed as:

[0051] θ * =argmin θ L;

[0052] Among them, L is the loss function, and its calculation formula is:

[0053]

[0054] Among them, v i ,v j is an entity, Yes i ,v j The true label of the relationship between Is entity v i ,v j The predicted value of the relationship between , where the predicted value reasoning method is as follows:

[0055]

[0056] where f readout is the readout function, and the readout function is calculated as follows:

[0057]

[0058] Where N(v) is the set of neighbor nodes of v, c uv is the weight or coefficient of the edge between nodes u and v, W (t) is the weight matrix at time step t, σ is the nonlinear activation function, Represents the feature representation of node u at layer t.

[0059] Further, step S6 specifically includes:

[0060] S61. Use intelligent question-answering technology based on semantic search to understand user intent in aerospace manufacturing scenarios and perform contextual analysis, provide personalized search results and recommendations based on user preferences and historical behaviors, and provide explainable search results to help users understand the reasoning process of the search algorithm;

[0061] S62. Promote knowledge discovery through complex query and decision understanding, use visualization, reasoning and retrieval techniques in aerospace manufacturing scenarios, enhance user interactive experience, and help users quickly understand business models.

[0062] Based on the above technical solution, the present invention has the following beneficial effects:

[0063] The present invention provides a complete set of knowledge graph construction paradigms, which solves the problems of difficult knowledge integration and low utilization in aerospace manufacturing scenarios, making it possible to build knowledge graphs in the aerospace manufacturing field.

[0064] Based on the joint extraction of structured, semi-structured and unstructured complex heterogeneous data, the richness of knowledge sources and the transferability of construction methods are guaranteed;

[0065] Through a variety of deep learning and machine learning methods, it provides a full-process knowledge extraction, knowledge fusion and knowledge discovery model, which can cope with the challenges brought by various knowledge sources and data formats to knowledge extraction, and solve the difficulties in the application of knowledge graphs in the aerospace manufacturing field and the poor task transferability of existing construction methods; based on top-down logic, it ensures the scientificity and seriousness of building knowledge graphs in this field;

[0066] At the same time, the intelligent question-answering and visual decision-making support functions of the knowledge graph are used to provide users with a complete set of intelligent and fast aerospace manufacturing knowledge understanding and knowledge mining solutions, and the knowledge intelligent decision-making function of the knowledge graph is used to further assist the transformation of the aerospace manufacturing business. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is an overall flow chart of the aerospace manufacturing knowledge graph construction method based on the heterogeneous data joint extraction mode proposed in the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0069] Although the steps in the present invention are arranged with numbers, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used in this article involves and covers any and all possible combinations of one or more of the associated listed items.

[0070] like Figure 1 As shown, a method for constructing an aerospace manufacturing knowledge graph based on a heterogeneous data joint extraction mode proposed by the present invention is given, which specifically includes the following steps:

[0071] S1. Design and construct the knowledge model of the knowledge graph using automated modeling tools based on top-down theory;

[0072] As a preferred implementation of step S1, step S1 specifically includes:

[0073] S11. Design aerospace manufacturing knowledge model based on top-down theory and define the ontology concepts of knowledge graph under the constraint of expert knowledge, including classes, relationships, attributes and instances;

[0074] S12. Use automated modeling tools to shield specific ontology description languages ​​and quickly construct ontology models in the aerospace manufacturing field at the conceptual level.

[0075] S2, combining table structure and graph structure to store the underlying knowledge graph;

[0076] As a preferred implementation of step S2, step S2 specifically includes:

[0077] S21. Use relational database to store knowledge graph knowledge model metadata, including ontology definition, categories, attributes, relationships and other information;

[0078] S22. Use RDF to store entities, attributes, and relationship instances of the knowledge graph, and synchronously map triple data to the graph database;

[0079] S23. Establish relational database, and synchronize data between RDF and graph database.

[0080] In this embodiment, the mapping relationship between RDF and graph database is constructed to facilitate various data queries in industrial scenarios and provide quick visualization support; and the data synchronization between relational database, RDF and graph database is established to ensure that the update of knowledge model and specific knowledge instances can be carried out without conflict.

[0081] S3, knowledge extraction and knowledge mining from structured, semi-structured and unstructured data;

[0082] As a preferred implementation of step S3, step S3 specifically includes:

[0083] S31, for structured aerospace manufacturing data, use R2RML language to directly map relational data into RDF datasets;

[0084] S32. For semi-structured and unstructured aerospace manufacturing data, establish a multi-task learning model that can simultaneously extract entities, relations, and events: accept information in the text through a unified input layer and convert it into word embeddings, use bidirectional long short-term memory networks and transformers in the shared layer to capture word order and context information, apply self-attention or hierarchical attention mechanism after the shared layer to focus on the most task-related parts of the text, design entity extraction layer, relation extraction layer, and event extraction layer after the shared layer and attention mechanism are fused, and define a multi-task joint loss function to optimize the multi-task learning model;

[0085] More specifically, the optimization process is:

[0086] Select θ that minimizes the multi-task joint loss function as the optimal parameter of the multi-task learning model, and the formula is expressed as:

[0087] θ * =argmin θ Ltotal ;

[0088] Among them, θ * represents the optimal parameters of the multi-task learning model, L total represents a multi-task joint loss function; the definition of the multi-task joint loss function is as follows:

[0089] L total =L entity (Y entity ,Y entity_true )+L relation (Y relation ,Y relation_true )+L event (Y event ,Y event_true );

[0090] Among them, L entity , L relation and L event Represents independent loss functions for entities, relations, and events respectively; Y entity_true , Y relation_true and Y event_true Represent the true values ​​of entities, relations, and events respectively; Y entity , Y relation and Y event Represents predicted values ​​of entities, relations, and events;

[0091] The predicted value is obtained by extracting the function, and the formula is expressed as:

[0092] Y entity =f entity (H);

[0093] Y relation =f relation (H);

[0094] Y event =f event (H);

[0095] Among them, f entity 、f relation and f event Represent the extraction functions of entities, relations, and events respectively; H is the fusion output of the sharing layer and the attention layer, and its formula is expressed as:

[0096] H=Bi-LSTM(X)+Attention(Bi-LSTM(X));

[0097] Among them, Bi-LSTM represents the bidirectional long short-term memory network of the shared layer, and Attention represents the attention function of the attention layer.

[0098] S33, use the probability model to perform entity disambiguation to solve the entity reference problem in the text, by defining the feature function f(e j ,c j )Entity e j and category c j Mapped to a feature vector, for the Markov model, define the state transition probability P(cj|cj-1), calculate the entity e i Belongs to category c j The probability P(c j |e i ), choose so that P(c j |e i ) The largest category c j As e i The disambiguation results are iteratively optimized using the cross entropy loss function.

[0099] S4, perform data pattern layer fusion and data instance layer fusion on data from different sources;

[0100] As a preferred implementation of step S4, step S4 specifically includes:

[0101] S41. At the data model layer, expert knowledge and high-confidence structured knowledge mapping from aerospace manufacturing are used to merge concepts, merge concept hierarchical relationships, and merge concept attribute definitions.

[0102] S42. At the data instance layer, a deep learning-based method is used to identify and match the same entities in different data sources. The text is converted into feature word vectors through a deep learning model, and instance matching is achieved based on cosine similarity. Optimization and model training are performed through contrast loss. After successful matching, different instances of the same entity are merged into a single entity.

[0103] More specifically, the specific process of optimization and model training through contrast loss is:

[0104] Select θ with the minimum contrast loss as the optimal parameter of the deep learning model, and the formula is expressed as:

[0105] θ * =argmin θ L

[0106] Among them, L is the contrast loss, and its calculation formula is:

[0107]

[0108] Among them, y i is a label, with a value of 1 or 0. and are the positive and negative samples of the i-th instance pair in the batch, and f is the deep learning model used to convert entities into embedding vectors.

[0109] In this embodiment, positive samples refer to sample pairs that are similar or related in some sense, and negative samples refer to sample pairs that are dissimilar or unrelated in some sense. This loss function encourages the model to bring the feature representations of positive samples closer and push the feature representations of negative samples further away.

[0110] S5. Fast knowledge reasoning based on graph neural networks for knowledge discovery and knowledge graph completion;

[0111] As a preferred implementation of step S5, step S5 specifically includes:

[0112] S51. Train the graph neural network model, represent the entities and relationships in the knowledge graph as a graph structure, where entities are nodes and relationships are edges. Update the state of each node by aggregating the information of neighboring nodes to complete information transfer. Use the embedded representation of entities and relationships learned by GNN to perform knowledge reasoning, and use the loss function to supervise the training of the model.

[0113] More specifically, the loss function is used to train the supervised model as follows:

[0114] Choose θ with the minimum loss g is the optimal parameter of the graph neural network model, and the formula is expressed as:

[0115] θ g * =argmin θ L;

[0116] Among them, θ g * represents the optimal parameters of the graph neural network model, L is the loss function, and its calculation formula is:

[0117]

[0118] Among them, v i ,v j is an entity, Yes i ,v j The true label of the relationship between Is entity v i ,v j The predicted value of the relationship between , where the predicted value reasoning method is as follows:

[0119]

[0120] in, and For entity node v i With v j The feature representation, f readout is the readout function, and the readout function is calculated as follows:

[0121]

[0122] Where N(v) is the set of neighbor nodes of v, c uv is the normalization coefficient, W (t) is the weight matrix at the time step, and σ is the nonlinear activation function. Since the function is updated in real time, is the update rule of the function;

[0123] S52. Use the trained graph neural network model to perform knowledge reasoning and knowledge graph completion when building and maintaining the knowledge graph under the supervision of expert knowledge.

[0124] S6. Combine the unique application forms of knowledge graphs with data and business scenarios in the aerospace manufacturing field to conduct intelligent question-answering and visual decision support based on semantic search;

[0125] As a preferred implementation of step S6, step S6 specifically includes:

[0126] S61. Use intelligent question-answering technology based on semantic search to understand user intent in aerospace manufacturing scenarios and perform contextual analysis, provide personalized search results and recommendations based on user preferences and historical behaviors, and provide explainable search results to help users understand the reasoning process of the search algorithm;

[0127] S62. Promote knowledge discovery through complex query and decision understanding, use visualization, reasoning and retrieval techniques in aerospace manufacturing scenarios, enhance user interactive experience, and help users quickly understand business models.

[0128] So far, this embodiment has realized the construction of aerospace manufacturing knowledge graph based on the joint extraction mode of heterogeneous data, and completed the vertical implementation of knowledge graph in the complex field of aerospace manufacturing. The method proposed in the present invention is based on the knowledge graph construction paradigm of multiple deep learning methods, and jointly extracts and applies knowledge to structured, semi-structured and unstructured knowledge sources to achieve the rapid construction of knowledge graphs in various aerospace manufacturing businesses. Through the top-down business logic, the construction of the knowledge graph is carried out under the guidance of expert knowledge, which will not cause confusion in knowledge structure and logic. At the same time, through intelligent question and answer and visual decision support, the knowledge graph is helped to quickly extract knowledge, understand knowledge and make knowledge decisions in the aerospace manufacturing scenario.

[0129] In summary, the present invention proposes a knowledge graph construction method based on a heterogeneous data joint extraction mode, which can adapt to various complex knowledge sources and data modes, and solve the problems of difficult implementation, poor migration and poor interpretability of knowledge graphs in the aerospace manufacturing field. This knowledge graph construction method can be applied to a variety of aerospace manufacturing task scenarios and provide efficient knowledge decision support.

[0130] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0131] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for constructing an aerospace manufacturing knowledge graph based on a joint extraction mode of heterogeneous data, characterized in that: The following steps are involved: S1. Design and construct the knowledge model of the knowledge graph using automated modeling tools based on top-down theory; S2, combining table structure and graph structure to store the underlying knowledge graph; S3, knowledge extraction and knowledge mining from structured, semi-structured and unstructured data; S4, perform data pattern layer fusion and data instance layer fusion on data from different sources; S5. Fast knowledge reasoning based on graph neural networks for knowledge discovery and knowledge graph completion; S6. Combine the unique application forms of knowledge graphs with data and business scenarios in the aerospace manufacturing field to provide intelligent question-answering and visual decision support based on semantic search.

2. The method for constructing an aerospace manufacturing knowledge graph based on a heterogeneous data joint extraction mode according to claim 1 is characterized in that: Step S1 specifically includes: S11. Design aerospace manufacturing knowledge model based on top-down theory and define the ontology concepts of knowledge graph under the constraint of expert knowledge, including classes, relationships, attributes and instances; S12. Use automated modeling tools to shield specific ontology description languages ​​and quickly construct ontology models in the aerospace manufacturing field at the conceptual level.

3. The method for constructing an aerospace manufacturing knowledge graph based on a heterogeneous data joint extraction mode according to claim 1 is characterized in that: Step S2 specifically includes: S21. Use a relational database to store the knowledge graph knowledge model metadata, and store ontology definitions, categories, attributes, relationships, etc. in the form of data tables; S22. Use the resource description framework RDF to store triple data of the knowledge graph, i.e., entities, attributes, and relationship instances, and synchronously map the triple data to the graph database; S23. Establish relational database, and synchronize data between RDF and graph database.

4. The method for constructing an aerospace manufacturing knowledge graph based on a joint extraction mode of heterogeneous data according to claim 1 is characterized in that: Step S3 specifically includes: S31, for structured aerospace manufacturing data, use R2RML language to directly map relational data into RDF datasets; S32. For semi-structured and unstructured aerospace manufacturing data, establish a multi-task learning model that can simultaneously extract entities, relations, and events: accept information in the text through a unified input layer and convert it into word embeddings, use bidirectional long short-term memory networks and transformers in the shared layer to capture word order and context information, apply self-attention or hierarchical attention mechanism after the shared layer to focus on the most task-related parts of the text, design entity extraction layer, relation extraction layer, and event extraction layer after the shared layer and attention mechanism are fused, and define a multi-task joint loss function to optimize the multi-task learning model; S33, use the probability model to perform entity disambiguation to solve the entity reference problem in the text, by defining the feature function f(e j , c j )Entity e j and category c j Mapped to a feature vector, for the Markov model, define the state transition probability P(cj|cj-1), calculate the entity e i Belongs to category c j The probability P(c j |e i ), choose so that P(c j |e i ) The largest category c j As e i The disambiguation results are iteratively optimized using the cross entropy loss function.

5. The method for constructing an aerospace manufacturing knowledge graph based on a heterogeneous data joint extraction mode according to claim 4 is characterized in that: The optimization process in step S32 is specifically as follows: Choose θ that minimizes the multi-task joint loss function m is the optimal parameter of the multi-task learning model, and the formula is expressed as: Among them, θ m * represents the optimal parameters of the multi-task learning model, L total represents a multi-task joint loss function; the definition of the multi-task joint loss function is as follows: L total =L entity (AND entity ,AND entity_true )+L relation (AND relation ,AND relation_true )+L event (AND event ,AND event_true ); Among them, L entity , L relation and L event Represents independent loss functions for entities, relations, and events respectively; Y entity_true , Y relation_true and Y event_true Represent the true values ​​of entities, relations, and events respectively; Y entity , Y relation and Y event Represents predicted values ​​of entities, relations, and events; The predicted value is obtained by extracting the function, and the formula is expressed as: Y entity =f entity (H); Y relation =f relation (H); Y event =f event (H); Among them, f entity 、f relation and f event Represent the extraction functions of entities, relations, and events respectively; H is the fusion output of the sharing layer and the attention layer, and its formula is expressed as: H=Bi-LSTM(X)+Attention(Bi-LSTM(X)); Among them, X represents the input information after embedding, Bi-LSTM represents the bidirectional long short-term memory network of the shared layer, and Attention represents the attention function of the attention layer.

6. The method for constructing an aerospace manufacturing knowledge graph based on a heterogeneous data joint extraction mode according to claim 1 is characterized in that: Step S4 specifically includes: S41. At the data model layer, expert knowledge and high-confidence structured knowledge mapping from aerospace manufacturing are used to merge concepts, merge concept hierarchical relationships, and merge concept attribute definitions. S42. At the data instance layer, a deep learning-based method is used to identify and match the same entities in different data sources. The text is converted into feature word vectors through a deep learning model, and instance matching is achieved based on cosine similarity. Optimization and model training are performed through contrast loss. After successful matching, different instances of the same entity are merged into a single entity.

7. The method for constructing an aerospace manufacturing knowledge graph based on a heterogeneous data joint extraction mode according to claim 6 is characterized in that: The specific process of optimization and model training by contrast loss in step S42 is as follows: Choose θ that minimizes the contrast loss c To compare the optimal parameters of the learning model, the formula is expressed as: Among them, θ c * represents the optimal parameters of the contrastive learning model, L is the contrastive loss, and its calculation formula is: Among them, y i is a weight term to ensure that only negative samples contribute to the loss function, and are the positive and negative samples of the i-th instance pair in the batch, respectively, and f is a deep learning model used to convert entities into embedding vectors.

8. The method for constructing an aerospace manufacturing knowledge graph based on a heterogeneous data joint extraction mode according to claim 1 is characterized in that: Step S5 specifically includes: S51. Train the graph neural network model, represent the entities and relationships in the knowledge graph as a graph structure, where entities are nodes and relationships are edges. Update the state of each node by aggregating the information of neighboring nodes to complete information transfer. Use the embedded representation of entities and relationships learned by GNN to perform knowledge reasoning, and use the loss function to supervise the training of the model. S52. Use the trained graph neural network model to perform knowledge reasoning and knowledge graph completion when building and maintaining the knowledge graph under the supervision of expert knowledge.

9. The method for constructing an aerospace manufacturing knowledge graph based on a heterogeneous data joint extraction mode according to claim 8 is characterized in that: In step S51, the training of the supervised model using the graph neural network loss function is specifically as follows: Choose θ with the minimum loss g is the optimal parameter of the graph neural network model, and the formula is expressed as: <h2 style=";text-align:left;direction:ltr">θ<h2 style=";text-align:left;direction:ltr"> g <h2 style=";text-align:left;direction:ltr"> * <h2 style=";text-align:left;direction:ltr"> =argmin<h2 style=";text-align:left;direction:ltr"> θ <h2 style=";text-align:left;direction:ltr"> L; Among them, θ g * represents the optimal parameters of the graph neural network model, L is the loss function, and its calculation formula is: Among them, v i , v j is an entity, Yes i , v j The true label of the relationship between Is entity v i , v j The predicted value of the relationship between , where the predicted value reasoning method is as follows: in, and For entity node v i With v j The feature representation, f readout is the readout function, and the readout function is calculated as follows: Where N(v) is the set of neighbor nodes of v, c uv is the weight or coefficient of the edge between nodes u and v, W (t) is the weight matrix at time step t, σ is the nonlinear activation function, Represents the feature representation of node u at layer t.

10. The method for constructing an aerospace manufacturing knowledge graph based on a heterogeneous data joint extraction mode according to claim 1 is characterized in that: Step S6 specifically includes: S61. Use semantic search-based intelligent question-answering technology to understand user intent in aerospace manufacturing scenarios and conduct contextual analysis, provide personalized search results and recommendations based on user preferences and historical behaviors, and explain search results to help users understand the reasoning process of search algorithms; S62. Promote knowledge discovery through complex query and decision understanding, use visualization, reasoning and retrieval techniques in aerospace manufacturing scenarios, enhance user interactive experience, and help users quickly understand business models.

Images