Equipment manufacturing-oriented multi-data link integration and fusion knowledge graph construction method

Abstract

The application discloses a kind of equipment manufacturing-oriented multi-data chain integration and fusion knowledge graph construction method, first, equipment manufacturing field data is collected;Pretreatment is carried out, and is divided into structured relational data and unstructured text data;Structured data corpus is handled, and is imported into Neo4j graphics database storage;Unstructured text data is extracted using deep learning and relationship extraction;The extraction result is arranged and imported into Neo4j graphics database, and the knowledge graph is preliminarily processed in Neo4j graphics database, and the knowledge graph is exported;The exported knowledge graph data is input into entity alignment model, and entity alignment task training is carried out.The application solves the problem that the knowledge graph data constructed in the prior art is not complete and has a lot of noise.

Description

A Method for Constructing a Multi-Data Chain Integration and Fusion Knowledge Graph for Equipment Manufacturing Technical Field

[0001] This invention belongs to the field of knowledge graph technology for equipment manufacturing, specifically relating to a method for constructing a knowledge graph that integrates and fuses multiple data chains for equipment manufacturing. Background Technology

[0002] The equipment manufacturing industry is the core of the manufacturing sector and a crucial guarantee for the industrial base and national defense. Its development level reflects a nation's comprehensive strength. Equipment manufacturing products, such as basic machinery, electronic components, and major complete sets of technical equipment, often go through multiple stages from research and development to application and subsequent maintenance. These stages include research, design, manufacturing, procurement, component supply, installation, commissioning, delivery, quality control, and after-sales service. This requires collaboration among multiple departments and extensive data interconnection. The provision, acquisition, and processing of information by various stakeholders within the data chain are particularly important for enterprise data management.

[0003] However, enterprises in the equipment manufacturing industry often have multiple data chains, which are interconnected and partially interoperable. This complexity severely impacts data management and business interoperability, hindering the industry's development level and speed. Therefore, achieving the integration and fusion of equipment manufacturing data chains is crucial for improving the data-driven capabilities and collaborative management of equipment manufacturing-related businesses among different enterprises. Knowledge graphs are an efficient way to manage massive amounts of information, offering advantages such as large scale, high quality, and superior structure. By constructing knowledge graphs for the integration and fusion of equipment manufacturing data chains, interconnectivity among enterprises can be achieved, enabling efficient data management.

[0004] Currently, there is a relative lack of technologies for constructing multi-data-chain integration and fusion knowledge graphs for equipment manufacturing. Knowledge graph construction methods from other fields are not very applicable to multi-data-chain integration and fusion in the equipment manufacturing field, and the constructed knowledge graphs suffer from problems such as incomplete data, high noise, and the need for a large amount of manual annotation. Therefore, there is a need for a method for constructing multi-data-chain integration and fusion knowledge graphs for equipment manufacturing. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing, which solves the problems of incomplete and noisy knowledge graph data in the existing technology.

[0006] The technical solution adopted in this invention is a method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing, which is implemented according to the following steps:

[0007] Step 1: Collect data in the equipment manufacturing field;

[0008] Step 2: Preprocess the data from Step 1 and divide it into structured relational data and unstructured text data;

[0009] Step 3: Process the structured data corpus from Step 2 and import it into the Neo4j graph database for storage;

[0010] Step 4: Use deep learning to extract entities and relationships from the unstructured text data in Step 2;

[0011] Step 5: Organize the extraction results from Step 4 and import them into the Neo4j graph database. Perform preliminary processing on the knowledge graph in the Neo4j graph database before exporting the knowledge graph.

[0012] Step 6: Input the knowledge graph data exported in Step 5 into the entity alignment model to train the entity alignment task.

[0013] The invention is further characterized in that,

[0014] Step 1 is implemented in the following steps:

[0015] By using web crawlers to scrape external internet data related to equipment manufacturing from relevant resource websites, we collected texts from the equipment manufacturing process, including production logs, manuals, task sheets, and enterprise information data exported from equipment manufacturing enterprise databases, including manufacturing parts data sheets for extrusion presses and design specification data sheets for continuous casting machines.

[0016] Step 2 is implemented in the following steps:

[0017] Remove useless content from the external internet data obtained in step 1, including page tags, advertising links, and units of measurement. Write corresponding regular expressions and use them to extract useful text. Classify the obtained useful text and the text collected during the equipment manufacturing process in step 1 into unstructured text data. Select and classify the structured relational data containing attributes and relationships in the enterprise information data and save it as structured relational data.

[0018] Step 3 is implemented in the following steps:

[0019] Extract the structured relational data from step 2, select the relevant attribute names and relation names, and set the head and tail pointers of the relations. Generate triples and export them as CSV format. Use Neo4j's APOC plugin to write relevant Cypher statements to import the triples into the graph database.

[0020] Step 4, entity extraction, is implemented according to the following steps:

[0021] Step 4.1: Add a start marker [CLS] to the beginning of each sentence in the unstructured text data obtained in Step 2 and an end marker [SEP] to the end of each sentence. Then, fine-tune the connection parameters of the ALBERT model based on the unstructured text data obtained in Step 2. The internal training parameters of the ALBERT model remain unchanged, and the fine-tuned ALBERT model is obtained.

[0022] Step 4.2: The unstructured text data obtained in Step 2 is annotated with BIO annotation mode. The annotated unstructured text data is divided into training set and validation set, and input into the ALBERT model obtained in Step 4.1 to obtain word vectors.

[0023] Step 4.3: Use the word vectors obtained in Step 4.2 as input to the Bidirectional Gated Recurrent Unit (BGRU). Use an attention mechanism to obtain the weighted score of each word on different labels. Use a Conditional Random Field (CRF) to constrain the label sequence and reduce erroneous sequences. Use the Vuterbi algorithm to solve for the optimal path and obtain the label with the highest score. The CRF discrimination calculation process is as follows:

[0024]

[0025]

[0026] Where S is the input text sequence, W is the entity label, and score(S, w) is the overall evaluation score, w i Let be the hidden state at time i in a Conditional Random Field (CRF). This indicates that the CRF hidden state is w. i The emission fraction matrix output by the BGRU layer at time. For the hidden state w obtained through CRF learning i to w i+1 The transition matrix, p(w|S) is the probability of the label corresponding to the input sequence, and W is the probability of the label corresponding to the input sequence. s Represents all possible label sequences;

[0027] Step 4.4: Based on the structured relational data obtained in Step 2, construct an entity dictionary for the equipment manufacturing data chain. Using the dictionary and the unstructured text data obtained in Step 2 as input, repeat Steps 4.2 and 4.3 to obtain the entity extraction results. Save the word vector results output by ALBERT during the extraction.

[0028] Step 4, relation extraction, is implemented according to the following steps:

[0029] Step a: Read the text word vectors obtained in step 4.4, and use the Gaussian Graph Generator to construct a potential directed multi-view using the text word vectors as initial features to simulate the relationships between words;

[0030] Step b: Use the Densely Connected Graph Convolutional Network (DCGCN) to extract deep graph features from the directed multi-view network constructed in step a. Then, use Dynamic Temporal Warping Pooling (DTWPool) to sparsify the graph structure generated by DCGCN. The view attention calculation formula of DTWPool is as follows:

[0031]

[0032] in, These are the attention weights calculated before SoftMax activation, where α is the activation function. It is the adjacency matrix between the i-th node and the j-th node in the m-th view, k j Let W be the initial node representation of node j. pool and b pool These are the trainable weights and biases in the pooling operation;

[0033] Step c: Repeat step b three times to obtain the final pooled graph K of the neural network computation graph representation of max pooling. Concatenate the vector representation of the start label [CLS] with the computation result, and input the result into the Softmax classifier to obtain the relation extraction result.

[0034] Step 5 is as follows:

[0035] Step 5.1: Construct triples based on the extraction results of Step 4.3 and Step c, convert the triples to CSV format, and import them using the APOC plugin provided by Neo4j graph database. When importing the knowledge graph, based on the text extraction source of the triples, if the knowledge graph already imported in Step 3 based on the structured relational data has a corresponding knowledge graph label to the source of the extracted and organized triples, then add the triples to the same knowledge graph; otherwise, select to create a new knowledge graph label import to generate a new knowledge graph.

[0036] Step 6.2: In the Neo4j database, use the Cypher statement to delete redundant information such as duplicate relationships between the same pair of nodes and nodes that have no relationship with other nodes, and export the sorted knowledge graph as triples.

[0037] Step 6 is as follows:

[0038] Step 6.1: Select two knowledge graphs E1 and E2 from the knowledge graphs exported in Step 5.2 to be aligned. The knowledge graphs to be aligned should have some knowledge relevance.

[0039] Step 6.2: Input the knowledge graphs E1 and E2 selected in Step 6.1 into the improved Graph Convolutional Network (GCN) model to obtain the entity embedding representations. The improved GCN model consists of three stacked GCN layers, and a Highway Network is used between each GCN layer to reduce accumulated noise. E1 and E2 are used as inputs, and the specific computation method for each layer is as follows:

[0040]

[0041] in, and For the transformation gate T(h) (l) The weight matrix and bias vector, h (l) The input to the (l+1)th layer GCN is used to aggregate high-order domain information through stacked GCN networks. The output of the third layer is processed by the ReLU activation function to obtain the knowledge graph embedding representation.

[0042] Step 6.3: Select candidate entities from the embedding representations of knowledge graphs E1 and E2 obtained in Step 6.2, and in E2, select the entity e from E1. i Sampling is performed on the possible alignment candidate entity set, and then e is calculated. i The probability that an entity in E2 will be sampled as a candidate entity is calculated based on the similarity to the subgraph of entity nodes in the candidate entity set.

[0043]

[0044] Among them, h i and h j For entity e i and entity e j The embedding representation, p(h) j |h i ) is entity e j Sampled as entity e i The probability of the candidate entity. This indicates finding the L1 norm of a vector;

[0045] Step 6.4: Use the subgraph of the node as input to the graph matching network to evaluate entity e in E1. i The neighbor node u and each candidate node c in its candidate entity set i The matching degree of the one-hop neighbor node v is used to perform a weighted summation of the differences between u and v using an attention mechanism, resulting in the matching vector representation z of u. u, z u The calculation method is as follows:

[0046]

[0047]

[0048] Where u and v are entities e i and entities to be aligned One of the neighboring nodes, att uv It is the attention weight, z u It is the matching vector of u, used to measure h. u The difference h between the candidate entity node subgraph and its nearest neighbor node u and h v It is the embedded representation of u and v output by GCN;

[0049] Then, it is concatenated with the output embedding representation of u in GCN to obtain the vector representation of the neighboring nodes;

[0050] Step 6.5: Take the e obtained in step 6.4 i All neighbor representations are aggregated using a weight matrix and compared with e in the GCN. i The entity embedding and splicing results in e i Vector representation that incorporates neighborhood information;

[0051] Step 6.6: Obtain the vector representation of the node fusion neighborhood information for all nodes in E1 and E2 through steps 6.4 and 6.5, and give the entity alignment result by calculating the distance between the pairs of entities to be predicted and sorting them.

[0052] Step 7 is as follows:

[0053] In the triples exported in step 5, extract the relevant knowledge triples of the aligned entity pairs, and remove the duplicate triples in the two sets of triples. Use Neo4j to import the non-duplicate knowledge of entities in knowledge graph E1 into knowledge graph E2, and import the non-duplicate knowledge of entities in knowledge graph E2 into E1. The supplemented and enriched knowledge graph is obtained in Neo4j.

[0054] The beneficial effects of this invention are that the method for constructing a multi-data-chain integration and fusion knowledge graph for equipment manufacturing constructs a knowledge graph by extracting entities and relationships from unstructured text data of the equipment manufacturing industry, and combines it with a knowledge graph constructed from structured relational data of the equipment manufacturing industry. The knowledge graph is further enriched by entity alignment between the knowledge graphs, thus realizing the construction of a multi-data-chain integration and fusion knowledge graph for equipment manufacturing. Attached Figure Description

[0055] Figure 1 is a flowchart of the method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing according to the present invention.

[0056] Figure 2 is a model flowchart of the entity alignment process in a multi-data chain integration and fusion knowledge graph construction method for equipment manufacturing in this invention. Detailed Implementation

[0057] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0058] This invention presents a method for constructing a multi-data-chain integration and fusion knowledge graph for equipment manufacturing. The flowchart is shown in Figure 1. Based on collected data, a knowledge graph is constructed by extracting entities and relationships from unstructured text data of the equipment manufacturing industry. This knowledge graph is then combined with a knowledge graph constructed from structured relational data of the equipment manufacturing industry. The knowledge graph is further enriched by entity alignment between the knowledge graphs. The Neo4j graph database is used for storage, thus realizing the construction of a multi-data-chain integration and fusion knowledge graph for equipment manufacturing.

[0059] This invention provides a method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing, which is implemented according to the following steps:

[0060] Step 1: Collect data in the equipment manufacturing field;

[0061] Step 1 is implemented in the following steps:

[0062] By using web crawlers, external internet data related to equipment manufacturing was collected from relevant resource websites, including "Tianxia Machinery Network" and "China Machine Tool Network". Texts from the equipment manufacturing process were collected, including production logs, manuals, and task sheets, as well as enterprise information data exported from equipment manufacturing enterprise databases, including manufacturing parts data sheets for extrusion presses and design specification data sheets for continuous casting machines.

[0063] Step 2 is implemented in the following steps:

[0064] Remove useless content from the external internet data obtained in step 1, including page tags, advertising links, and units of measurement. Write corresponding regular expressions and use them to extract useful text. Classify the obtained useful text and the text collected during the equipment manufacturing process in step 1 into unstructured text data. Select and classify the structured relational data containing attributes and relationships in the enterprise information data and save it as structured relational data.

[0065] Step 2: Preprocess the data obtained in Step 1 and divide it into structured relational data and unstructured text data;

[0066] Step 3: Process the structured data corpus from Step 2 and import it into the Neo4j graph database for storage;

[0067] Step 3 is implemented in the following steps:

[0068] Extract the structured relational data from step 2, select the relevant attribute names and relation names, and set the head and tail pointers of the relations. Generate triples and export them as CSV format. Use Neo4j's APOC plugin to write relevant Cypher statements to import the triples into the graph database.

[0069] Referring to Figure 2, step 4 involves using deep learning to extract entities and relationships from the unstructured text data in step 2.

[0070] Step 4, entity extraction, is implemented according to the following steps:

[0071] Step 4.1: Add a start marker [CLS] to the beginning of each sentence in the unstructured text data obtained in Step 2 and an end marker [SEP] to the end of each sentence. Then, fine-tune the connection parameters of the ALBERT model based on the unstructured text data obtained in Step 2. The internal training parameters of the ALBERT model remain unchanged, and the fine-tuned ALBERT model is obtained.

[0072] Step 4.2: The unstructured text data obtained in Step 2 is annotated with BIO annotation mode. The annotated unstructured text data is divided into training set and validation set, and input into the ALBERT model obtained in Step 4.1 to obtain word vectors.

[0073] Step 4.3: Use the word vectors obtained in Step 4.2 as input to the Bidirectional Gated Recurrent Unit (BGRU). Use an attention mechanism to obtain the weighted score of each word on different labels. Use a Conditional Random Field (CRF) to constrain the label sequence and reduce erroneous sequences. Use the Vuterbi algorithm to solve for the optimal path and obtain the label with the highest score. The CRF discrimination calculation process is as follows:

[0074]

[0075]

[0076] Where S is the input text sequence, W is the entity label, and score(S, w) is the overall evaluation score, w i Let be the hidden state at time i in a Conditional Random Field (CRF). This indicates that the CRF hidden state is w. i The emission fraction matrix output by the BGRU layer at time. For the hidden state w obtained through CRF learning i to w i+1 The transition matrix, p(w|S) is the probability of the label corresponding to the input sequence, and W is the probability of the label corresponding to the input sequence. s Represents all possible label sequences;

[0077] Step 4.4: Based on the structured relational data obtained in Step 2, construct an entity dictionary for the equipment manufacturing data chain. Using the dictionary and the unstructured text data obtained in Step 2 as input, repeat Steps 4.2 and 4.3 to obtain the entity extraction results. Save the word vector results output by ALBERT during the extraction.

[0078] Step 4, relation extraction, is implemented according to the following steps:

[0079] Step a: Read the text word vectors obtained in step 4.4, and use the Gaussian Graph Generator to construct a potential directed multi-view using the text word vectors as initial features to simulate the relationships between words;

[0080] Step b: Use the Densely Connected Graph Convolutional Network (DCGCN) to extract deep graph features from the directed multi-view network constructed in step a. Then, use Dynamic Temporal Warping Pooling (DTWPool) to sparsify the graph structure generated by DCGCN. The view attention calculation formula of DTWPool is as follows:

[0081]

[0082] in, These are the attention weights calculated before SoftMax activation, where α is the activation function. It is the adjacency matrix between the i-th node and the j-th node in the m-th view, k j Let W be the initial node representation of node j. pool and b pool These are the trainable weights and biases in the pooling operation;

[0083] Step c: Repeat step b three times to obtain the final pooled graph K of the neural network computation graph representation of max pooling. Concatenate the vector representation of the start label [CLS] with the computation result, and input the result into the Softmax classifier to obtain the relation extraction result.

[0084] Step 5: Organize the extraction results from Step 4 and import them into the Neo4j graph database. Perform preliminary processing on the knowledge graph in the Neo4j graph database before exporting the knowledge graph.

[0085] Step 5 is as follows:

[0086] Step 5.1: Construct triples based on the extraction results of Step 4.3 and Step c, convert the triples to CSV format, and import them using the APOC plugin provided by Neo4j graph database. When importing the knowledge graph, based on the text extraction source of the triples, if the knowledge graph already imported in Step 3 based on the structured relational data has a corresponding knowledge graph label to the source of the extracted and organized triples, then add the triples to the same knowledge graph; otherwise, select to create a new knowledge graph label import to generate a new knowledge graph.

[0087] Step 6.2: In the Neo4j database, use the Cypher statement to delete redundant information such as duplicate relationships between the same pair of nodes and nodes that have no relationship with other nodes, and export the sorted knowledge graph as triples.

[0088] Step 6: Input the knowledge graph data exported in Step 5 into the entity alignment model to train the entity alignment task.

[0089] Step 6 is as follows:

[0090] Step 6.1: Select two knowledge graphs E1 and E2 from the knowledge graphs exported in Step 5.2 to be aligned. The knowledge graphs to be aligned should have some knowledge relevance.

[0091] Step 6.2: Input the knowledge graphs E1 and E2 selected in Step 6.1 into the improved Graph Convolutional Network (GCN) model to obtain the entity embedding representations. The improved GCN model consists of three stacked GCN layers, and a Highway Network is used between each GCN layer to reduce accumulated noise. E1 and E2 are used as inputs, and the specific computation method for each layer is as follows:

[0092]

[0093] in, and For the transformation gate T(h) (l) The weight matrix and bias vector, h (l) The input to the (l+1)th layer GCN is used to aggregate high-order domain information through stacked GCN networks. The output of the third layer is processed by the ReLU activation function to obtain the knowledge graph embedding representation.

[0094] Step 6.3: Select candidate entities from the embedding representations of knowledge graphs E1 and E2 obtained in Step 6.2, and in E2, select the entity e from E1. i Sampling is performed on the possible alignment candidate entity set, and then e is calculated. iThe probability that an entity in E2 will be sampled as a candidate entity is calculated based on the similarity to the subgraph of entity nodes in the candidate entity set.

[0095]

[0096] Among them, h i and h j For entity e i and entity e j The embedding representation, p(h) j |h i ) is entity e j Sampled as entity e i The probability of the candidate entity. This indicates finding the L1 norm of a vector;

[0097] Step 6.4: Use the subgraph of the node as input to the graph matching network to evaluate entity e in E1. i The neighbor node u and each candidate node c in its candidate entity set i The matching degree of the one-hop neighbor node v is used to perform a weighted summation of the differences between u and v using an attention mechanism, resulting in the matching vector representation z of u. u , z u The calculation method is as follows:

[0098]

[0099]

[0100] Where u and v are entities e i and entities to be aligned One of the neighboring nodes, att uv It is the attention weight, z u It is the matching vector of u, used to measure h. u The difference h between the candidate entity node subgraph and its nearest neighbor node u and h v It is the embedded representation of u and v output by GCN;

[0101] Then, it is concatenated with the output embedding representation of u in GCN to obtain the vector representation of the neighboring nodes;

[0102] Step 6.5: Take the e obtained in step 6.4 i All neighbor representations are aggregated using a weight matrix and compared with e in the GCN. i The entity embedding and splicing results in e i Vector representation that incorporates neighborhood information;

[0103] Step 6.6: Obtain the vector representation of the node fusion neighborhood information for all nodes in E1 and E2 through steps 6.4 and 6.5, and give the entity alignment result by calculating the distance between the pairs of entities to be predicted and sorting them.

[0104] Step 7: Using the entity alignment results from Step 6, the knowledge graph is enriched through mutual information supplementation between different knowledge graphs, and the newly added knowledge is imported into the Neo4j graph database.

[0105] Step 7 is as follows:

[0106] In the triples exported in step 5, extract the relevant knowledge triples of the aligned entity pairs, and remove the duplicate triples in the two sets of triples. Use Neo4j to import the non-duplicate knowledge of entities in knowledge graph E1 into knowledge graph E2, and import the non-duplicate knowledge of entities in knowledge graph E2 into E1. The supplemented and enriched knowledge graph is obtained in Neo4j.

[0107] Example 1

[0108] This invention provides a method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing, which is implemented according to the following steps:

[0109] Step 1: Collect data in the equipment manufacturing field;

[0110] Step 1 is implemented in the following steps:

[0111] By using web crawlers, external internet data related to equipment manufacturing was collected from relevant resource websites, including "Tianxia Machinery Network" and "China Machine Tool Network". Texts from the equipment manufacturing process were collected, including production logs, manuals, and task sheets, as well as enterprise information data exported from equipment manufacturing enterprise databases, including manufacturing parts data sheets for extrusion presses and design specification data sheets for continuous casting machines.

[0112] Step 2 is implemented in the following steps:

[0113] Remove useless content from the external internet data obtained in step 1, including page tags, advertising links, and units of measurement. Write corresponding regular expressions and use them to extract useful text. Classify the obtained useful text and the text collected during the equipment manufacturing process in step 1 into unstructured text data. Select and classify the structured relational data containing attributes and relationships in the enterprise information data and save it as structured relational data.

[0114] Step 2: Preprocess the data obtained in Step 1 and divide it into structured relational data and unstructured text data;

[0115] Step 3: Process the structured data corpus from Step 2 and import it into the Neo4j graph database for storage;

[0116] Step 4: Use deep learning to extract entities and relationships from the unstructured text data in Step 2;

[0117] Step 5: Organize the extraction results from Step 4 and import them into the Neo4j graph database. Perform preliminary processing on the knowledge graph in the Neo4j graph database before exporting the knowledge graph.

[0118] Step 6: Input the knowledge graph data exported in Step 5 into the entity alignment model to train the entity alignment task.

[0119] Step 7: Using the entity alignment results from Step 6, the knowledge graph is enriched through mutual information supplementation between different knowledge graphs, and the newly added knowledge is imported into the Neo4j graph database.

[0120] Example 2

[0121] This invention provides a method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing, which is implemented according to the following steps:

[0122] Step 1: Collect data in the equipment manufacturing field;

[0123] Step 1 is implemented in the following steps:

[0124] By using web crawlers, external internet data related to equipment manufacturing was collected from relevant resource websites, including "Tianxia Machinery Network" and "China Machine Tool Network". Texts from the equipment manufacturing process were collected, including production logs, manuals, and task sheets, as well as enterprise information data exported from equipment manufacturing enterprise databases, including manufacturing parts data sheets for extrusion presses and design specification data sheets for continuous casting machines.

[0125] Step 2 is implemented in the following steps:

[0126] Remove useless content from the external internet data obtained in step 1, including page tags, advertising links, and units of measurement. Write corresponding regular expressions and use them to extract useful text. Classify the obtained useful text and the text collected during the equipment manufacturing process in step 1 into unstructured text data. Select and classify the structured relational data containing attributes and relationships in the enterprise information data and save it as structured relational data.

[0127] Step 2: Preprocess the data obtained in Step 1 and divide it into structured relational data and unstructured text data;

[0128] Step 3: Process the structured data corpus from Step 2 and import it into the Neo4j graph database for storage;

[0129] Step 3 is implemented in the following steps:

[0130] Extract the structured relational data from step 2, select the relevant attribute names and relation names, and set the head and tail pointers of the relations. Generate triples and export them as CSV format. Use Neo4j's APOC plugin to write relevant Cypher statements to import the triples into the graph database.

[0131] Step 4: Use deep learning to extract entities and relationships from the unstructured text data in Step 2;

[0132] Step 4, entity extraction, is implemented according to the following steps:

[0133] Step 4.1: Add a start marker [CLS] to the beginning of each sentence in the unstructured text data obtained in Step 2 and an end marker [SEP] to the end of each sentence. Then, fine-tune the connection parameters of the ALBERT model based on the unstructured text data obtained in Step 2. The internal training parameters of the ALBERT model remain unchanged, and the fine-tuned ALBERT model is obtained.

[0134] Step 4.2: The unstructured text data obtained in Step 2 is annotated with BIO annotation mode. The annotated unstructured text data is divided into training set and validation set, and input into the ALBERT model obtained in Step 4.1 to obtain word vectors.

[0135] Step 4.3: Use the word vectors obtained in Step 4.2 as input to the Bidirectional Gated Recurrent Unit (BGRU). Use an attention mechanism to obtain the weighted score of each word on different labels. Use a Conditional Random Field (CRF) to constrain the label sequence and reduce erroneous sequences. Use the Vuterbi algorithm to solve for the optimal path and obtain the label with the highest score. The CRF discrimination calculation process is as follows:

[0136]

[0137]

[0138] Where S is the input text sequence, W is the entity label, and score(S, w) is the overall evaluation score, w i Let be the hidden state at time i in a Conditional Random Field (CRF). This indicates that the CRF hidden state is w. i The emission fraction matrix output by the BGRU layer at time. For the hidden state w obtained through CRF learning i to w i+1 The transition matrix, p(w|S) is the probability of the label corresponding to the input sequence, and W is the probability of the label corresponding to the input sequence. s Represents all possible label sequences;

[0139] Step 4.4: Based on the structured relational data obtained in Step 2, construct an entity dictionary for the equipment manufacturing data chain. Using the dictionary and the unstructured text data obtained in Step 2 as input, repeat Steps 4.2 and 4.3 to obtain the entity extraction results. Save the word vector results output by ALBERT during the extraction.

[0140] Step 4, relation extraction, is implemented according to the following steps:

[0141] Step a: Read the text word vectors obtained in step 4.4, and use the Gaussian Graph Generator to construct a potential directed multi-view using the text word vectors as initial features to simulate the relationships between words;

[0142] Step b: Use the Densely Connected Graph Convolutional Network (DCGCN) to extract deep graph features from the directed multi-view network constructed in step a. Then, use Dynamic Temporal Warping Pooling (DTWPool) to sparsify the graph structure generated by DCGCN. The view attention calculation formula of DTWPool is as follows:

[0143]

[0144] in, These are the attention weights calculated before SoftMax activation, where α is the activation function. It is the adjacency matrix between the i-th node and the j-th node in the m-th view, k j Let W be the initial node representation of node j. pool and b pool These are the trainable weights and biases in the pooling operation;

[0145] Step c: Repeat step b three times to obtain the final pooled graph K of the neural network computation graph representation of max pooling. Concatenate the vector representation of the start label [CLS] with the computation result, and input the result into the Softmax classifier to obtain the relation extraction result.

[0146] Step 5: Organize the extraction results from Step 4 and import them into the Neo4j graph database. Perform preliminary processing on the knowledge graph in the Neo4j graph database before exporting the knowledge graph.

[0147] Step 6: Input the knowledge graph data exported in Step 5 into the entity alignment model to train the entity alignment task.

[0148] Step 7: Using the entity alignment results from Step 6, the knowledge graph is enriched through mutual information supplementation between different knowledge graphs, and the newly added knowledge is imported into the Neo4j graph database.

[0149] Example 3

[0150] This invention provides a method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing, which is implemented according to the following steps:

[0151] Step 1: Collect data in the equipment manufacturing field;

[0152] Step 1 is implemented in the following steps:

[0153] By using web crawlers, external internet data related to equipment manufacturing was collected from relevant resource websites, including "Tianxia Machinery Network" and "China Machine Tool Network". Texts from the equipment manufacturing process were collected, including production logs, manuals, and task sheets, as well as enterprise information data exported from equipment manufacturing enterprise databases, including manufacturing parts data sheets for extrusion presses and design specification data sheets for continuous casting machines.

[0154] Step 2 is implemented in the following steps:

[0155] Remove useless content from the external internet data obtained in step 1, including page tags, advertising links, and units of measurement. Write corresponding regular expressions and use them to extract useful text. Classify the obtained useful text and the text collected during the equipment manufacturing process in step 1 into unstructured text data. Select and classify the structured relational data containing attributes and relationships in the enterprise information data and save it as structured relational data.

[0156] Step 2: Preprocess the data obtained in Step 1 and divide it into structured relational data and unstructured text data;

[0157] Step 3: Process the structured data corpus from Step 2 and import it into the Neo4j graph database for storage;

[0158] Step 3 is implemented in the following steps:

[0159] Extract the structured relational data from step 2, select the relevant attribute names and relation names, and set the head and tail pointers of the relations. Generate triples and export them as CSV format. Use Neo4j's APOC plugin to write relevant Cypher statements to import the triples into the graph database.

[0160] Step 4: Use deep learning to extract entities and relationships from the unstructured text data in Step 2;

[0161] Step 4, entity extraction, is implemented according to the following steps:

[0162] Step 4.1: Add a start marker [CLS] to the beginning of each sentence in the unstructured text data obtained in Step 2 and an end marker [SEP] to the end of each sentence. Then, fine-tune the connection parameters of the ALBERT model based on the unstructured text data obtained in Step 2. The internal training parameters of the ALBERT model remain unchanged, and the fine-tuned ALBERT model is obtained.

[0163] Step 4.2: The unstructured text data obtained in Step 2 is annotated with BIO annotation mode. The annotated unstructured text data is divided into training set and validation set, and input into the ALBERT model obtained in Step 4.1 to obtain word vectors.

[0164] Step 4.3: Use the word vectors obtained in Step 4.2 as input to the Bidirectional Gated Recurrent Unit (BGRU). Use an attention mechanism to obtain the weighted score of each word on different labels. Use a Conditional Random Field (CRF) to constrain the label sequence and reduce erroneous sequences. Use the Vuterbi algorithm to solve for the optimal path and obtain the label with the highest score. The CRF discrimination calculation process is as follows:

[0165]

[0166]

[0167] Where S is the input text sequence, W is the entity label, and score(S, w) is the overall evaluation score, w i Let be the hidden state at time i in a Conditional Random Field (CRF). This indicates that the CRF hidden state is w. i The emission fraction matrix output by the BGRU layer at time. For the hidden state w obtained through CRF learning i to w i+1 The transition matrix, p(w|S) is the probability of the label corresponding to the input sequence, and W is the probability of the label corresponding to the input sequence. s Represents all possible label sequences;

[0168] Step 4.4: Based on the structured relational data obtained in Step 2, construct an entity dictionary for the equipment manufacturing data chain. Using the dictionary and the unstructured text data obtained in Step 2 as input, repeat Steps 4.2 and 4.3 to obtain the entity extraction results. Save the word vector results output by ALBERT during the extraction.

[0169] Step 4, relation extraction, is implemented according to the following steps:

[0170] Step a: Read the text word vectors obtained in step 4.4, and use the Gaussian Graph Generator to construct a potential directed multi-view using the text word vectors as initial features to simulate the relationships between words;

[0171] Step b: Use the Densely Connected Graph Convolutional Network (DCGCN) to extract deep graph features from the directed multi-view network constructed in step a. Then, use Dynamic Temporal Warping Pooling (DTWPool) to sparsify the graph structure generated by DCGCN. The view attention calculation formula of DTWPool is as follows:

[0172]

[0173] in, These are the attention weights calculated before SoftMax activation, where α is the activation function. It is the adjacency matrix between the i-th node and the j-th node in the m-th view, k j Let W be the initial node representation of node j. pool and b pool These are the trainable weights and biases in the pooling operation;

[0174] Step c: Repeat step b three times to obtain the final pooled graph K of the neural network computation graph representation of max pooling. Concatenate the vector representation of the start label [CLS] with the computation result, and input the result into the Softmax classifier to obtain the relation extraction result.

[0175] Step 5: Organize the extraction results from Step 4 and import them into the Neo4j graph database. Perform preliminary processing on the knowledge graph in the Neo4j graph database before exporting the knowledge graph.

[0176] Step 5 is as follows:

[0177] Step 5.1: Construct triples based on the extraction results of Step 4.3 and Step c, convert the triples to CSV format, and import them using the APOC plugin provided by Neo4j graph database. When importing the knowledge graph, based on the text extraction source of the triples, if the knowledge graph already imported in Step 3 based on the structured relational data has a corresponding knowledge graph label to the source of the extracted and organized triples, then add the triples to the same knowledge graph; otherwise, select to create a new knowledge graph label import to generate a new knowledge graph.

[0178] Step 6.2: In the Neo4j database, use the Cypher statement to delete redundant information such as duplicate relationships between the same pair of nodes and nodes that have no relationship with other nodes, and export the sorted knowledge graph as triples.

[0179] Step 6: Input the knowledge graph data exported in Step 5 into the entity alignment model to train the entity alignment task.

[0180] Step 6 is as follows:

[0181] Step 6.1: Select two knowledge graphs E1 and E2 from the knowledge graphs exported in Step 5.2 to be aligned. The knowledge graphs to be aligned should have some knowledge relevance.

[0182] Step 6.2: Input the knowledge graphs E1 and E2 selected in Step 6.1 into the improved Graph Convolutional Network (GCN) model to obtain the entity embedding representations. The improved GCN model consists of three stacked GCN layers, and a Highway Network is used between each GCN layer to reduce accumulated noise. E1 and E2 are used as inputs, and the specific computation method for each layer is as follows:

[0183]

[0184] in, and For the transformation gate T(h) (l) The weight matrix and bias vector, h (l) The input to the (l+1)th layer GCN is used to aggregate high-order domain information through stacked GCN networks. The output of the third layer is processed by the ReLU activation function to obtain the knowledge graph embedding representation.

[0185] Step 6.3: Select candidate entities from the embedding representations of knowledge graphs E1 and E2 obtained in Step 6.2, and in E2, select the entity e from E1. i Sampling is performed on the possible alignment candidate entity set, and then e is calculated. iThe probability that an entity in E2 will be sampled as a candidate entity is calculated based on the similarity to the subgraph of entity nodes in the candidate entity set.

[0186]

[0187] Among them, h i and h j For entity e i and entity e j The embedding representation, p(h) j |h i ) is entity e j Sampled as entity e i The probability of the candidate entity. This indicates finding the L1 norm of a vector;

[0188] Step 6.4: Use the subgraph of the node as input to the graph matching network to evaluate entity e in E1. i The neighbor node u and each candidate node c in its candidate entity set i The matching degree of the one-hop neighbor node v is used to perform a weighted summation of the differences between u and v using an attention mechanism, resulting in the matching vector representation z of u. u , z u The calculation method is as follows:

[0189]

[0190]

[0191] Where u and v are entities e i and entities to be aligned One of the neighboring nodes, att uv It is the attention weight, z u It is the matching vector of u, used to measure h. u The difference h between the candidate entity node subgraph and its nearest neighbor node u and h v It is the embedded representation of u and v output by GCN;

[0192] Then, it is concatenated with the output embedding representation of u in GCN to obtain the vector representation of the neighboring nodes;

[0193] Step 6.5: Take the e obtained in step 6.4 i All neighbor representations are aggregated using a weight matrix and compared with e in the GCN. i The entity embedding and splicing results in e i Vector representation that incorporates neighborhood information;

[0194] Step 6.6: Obtain the vector representation of the node fusion neighborhood information for all nodes in E1 and E2 through steps 6.4 and 6.5, and give the entity alignment result by calculating the distance between the pairs of entities to be predicted and sorting them.

[0195] Step 7: Using the entity alignment results from Step 6, the knowledge graph is enriched through mutual information supplementation between different knowledge graphs, and the newly added knowledge is imported into the Neo4j graph database.

[0196] Step 7 is as follows:

[0197] In the triples exported in step 5, extract the relevant knowledge triples of the aligned entity pairs, and remove the duplicate triples in the two sets of triples. Use Neo4j to import the non-duplicate knowledge of entities in knowledge graph E1 into knowledge graph E2, and import the non-duplicate knowledge of entities in knowledge graph E2 into E1. The supplemented and enriched knowledge graph is obtained in Neo4j.

Claims

1. A method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing, characterized in that, The specific implementation follows these steps: Step 1: Collect data from the equipment manufacturing field; Step 2: After preprocessing, divide the data into structured relational data and unstructured text data; Step 3: Process the structured data corpus and import it into the Neo4j graph database for storage; Step 4: Use deep learning to extract entities and relations from the unstructured text data; Entity extraction in Step 4 is specifically implemented as follows: Step 4.1: Add a start marker [CLS] to the beginning of each sentence in the unstructured text data obtained in Step 2 and an end marker [SEP] to the end of each sentence, and fine-tune the connection parameters of the ALBERT model based on the unstructured text data obtained in Step 2. The training parameters remain unchanged to obtain a fine-tuned ALBERT model. Step 4.2: The unstructured text data obtained in Step 2 is annotated using the BIO annotation mode. The annotated unstructured text data is then divided into training and validation sets, which are input into the ALBERT model obtained in Step 4.1 to obtain word vectors. Step 4.3: The word vectors obtained in Step 4.2 are used as input to a bidirectional gated recurrent unit (BGRU). An attention mechanism is used to obtain the weighted score of each word on different labels. A Conditional Random Field (CRF) is used to constrain the label sequence, reducing erroneous sequences. The Vuterbi algorithm is used to solve for the optimal path to obtain the label with the highest score. The CRF discrimination calculation process is as follows: （1） (2) Among them, Given a sequence of text as input, For the corresponding tags of the entities, To comprehensively evaluate the scores, For Conditional Random Fields (CRF) The hidden state at all times Indicates that the CRF hidden state is The emission fraction matrix output by the BGRU layer at time. To learn from the hidden state through CRF arrive The transition matrix, This represents the probability of a label corresponding to an input sequence. Representing all possible label sequences; Step 4.4: Based on the structured relational data obtained in Step 2, construct an entity dictionary for the equipment manufacturing data chain, and use the dictionary and the unstructured text data obtained in Step 2 as input, repeating Steps 4.2 and 4.3 to obtain entity extraction results, and save the word vector results output by ALBERT during extraction; The relation extraction in Step 4 is specifically implemented according to the following steps: Step a: Read the text word vectors obtained in Step 4.4, and use the Gaussian Graph Generator to construct a potential directed multi-view using the text word vectors as initial features to simulate inter-word relationships; Step b: Use the Densely Connected Graph Convolutional Network (DCGCN) to extract deep graph features from the directed multi-view constructed in Step a, and then use Dynamic Time Warping Pooling (DTWPool) to sparsify the graph structure generated by DCGCN. The view attention calculation formula of DTWPool is as follows: (3) Among them, These are the attention weights calculated before SoftMax activation. For activation function, It is the adjacency matrix between the i-th node and the j-th node in the m-th view. Let j be the initial node representation. and This refers to the trainable weights and biases in the pooling operation; Step c: Repeat step b three times to obtain the final pooled graph K, which is the neural network computation graph representation of max pooling. The vector representation of the starting identifier [CLS] is concatenated with the computation result, and the result is input into the Softmax classifier to obtain the relation extraction result; Step 5: Organize the extraction result and import it into the Neo4j graph database; Step 5 is specifically as follows: Step 5.1: Construct triples based on the extraction results of steps 4.3 and c, convert the triples to CSV format, and import them using the APOC plugin provided by the Neo4j graph database. When importing the knowledge graph, based on the text extraction source of the triples, if the structured relation extraction in step 3 is used... If the imported knowledge graph contains knowledge graph tags corresponding to the extracted and organized triples, then the triples are added to the same knowledge graph; otherwise, a new knowledge graph tag is created and imported to generate a new knowledge graph. Step 5.2: In the Neo4j database, the Cypher statement is used to delete duplicate relationships between the same pair of nodes and redundant information about nodes that have no relationship with other nodes, and the organized knowledge graph is exported as triples. Step 6: The exported knowledge graph data is input into the entity alignment model for entity alignment training. Step 6 is specifically as follows: Step 6.1: Select two knowledge graphs to be aligned from the knowledge graph exported in Step 5.

2. 、 The knowledge graphs that need to be aligned should have some knowledge relevance; Step 6.2: align the knowledge graphs selected in Step 6.

1. 、 The entity embedding representation is obtained by feeding the input into an improved Graph Convolutional Network (GCN) model. The improved GCN model consists of three stacked GCN layers, and a Highway Network is used between the GCN layers to reduce accumulated noise. 、 As input, the specific calculation method for each layer is as follows: (4) Among them, and Transform gate The weight matrix and bias vector, For the first The input to the layer GCN is aggregated with high-order domain information through a stacked GCN network. The output of the third layer is then processed by the ReLU activation function to obtain the knowledge graph embedding representation. Step 6.3 involves processing the knowledge graph obtained in step 6.

2. 、 The embedding represents the selection of candidate entities, in China medium entity Sampling of possible alignment candidate entity sets, and then calculation. The similarity between the entity node subgraph and the candidate entity set is used to... The probability that a certain entity is sampled as a candidate entity is calculated as follows: (5) Among them, and For entities and entity Embedded representation, For entities Sampled as entities The probability of the candidate entity. This indicates calculating the L1 norm of a vector; Step 6.4: Use the subgraph of the node as input to the graph matching network and evaluate... medium entity neighboring nodes Each candidate node in its candidate entity set One-hop neighbor node The degree of matching, for and The difference is weighted and summed using an attention mechanism to obtain... Matching vector representation ， The calculation method is as follows: （6） (7) Among them, and They are entities and entities to be aligned One of the neighboring nodes, It is attention weight. yes The matching vector is used for measurement. The difference between the candidate entity node and its nearest neighbor node in the subgraph. and It is output by GCN. and The embedded representation; then it is combined with the GCN The output embedding representations are concatenated to obtain the vector representations of neighboring nodes; Step 6.5: The vector representations obtained in Step 6.4 are concatenated. All neighbor representations are aggregated using a weight matrix and compared with those in the GCN. The entity is embedded and spliced ​​to obtain Vector representation incorporating neighborhood information; Step 6.6, for 、 All nodes in the process obtain vector representations of node fusion neighborhood information through steps 6.4 and 6.

5. The entity alignment results are given by calculating the distance between the pairs of entities to be predicted and sorting them. In step 7, the knowledge graph is enriched by supplementing the mutual information between different knowledge graphs using the entity alignment results from step 6, and the newly added knowledge is imported into the Neo4j graph database.

2. The method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing according to claim 1, characterized in that, Step 1 is implemented in the following steps: using web crawler tools to crawl external Internet data on equipment manufacturing from relevant resource websites, collecting texts from the equipment manufacturing process, including production logs, manuals, task sheets, and enterprise information data exported from the equipment manufacturing enterprise database, including manufacturing parts data tables for extrusion presses and design specification data tables for continuous casting machines.

3. The method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing according to claim 2, characterized in that, Step 2 is implemented in the following steps: Delete useless content from the external Internet data obtained in Step 1, including page tags, advertising links, and units of measurement, and write corresponding regular expressions to extract useful text. Classify the obtained useful text and the text collected in the equipment manufacturing process in Step 1 into unstructured text data. Select and classify the structured relational data containing attributes and relationships in the enterprise information data and save it as structured relational data.

4. The method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing according to claim 3, characterized in that, Step 3 is implemented in the following steps: extract the structured relational data from step 2, select the relevant attribute names and relation names, set the head and tail pointers of the relations, generate triples and export them as CSV format, and use the Neo4j APOC plugin to write relevant Cypher statements to import the triples into the graph database.

5. The method for constructing a multi-data chain integration and fusion knowledge graph for equipment manufacturing according to claim 4, characterized in that, Step 7 is as follows: From the triples exported in step 5, extract the relevant knowledge triples for aligned entity pairs, remove duplicate triples from both sets of triples, and use Neo4j to process the knowledge graph. Importing non-repeating knowledge of entities into knowledge graphs In the middle, knowledge graphs Import of non-repeating knowledge of medium entities In Neo4j, the knowledge graph is supplemented and enriched.

Images