Equipment knowledge graph multi-dimensional feature fusion instance alignment method

Abstract

The invention discloses an equipment knowledge graph multi-dimensional feature fusion instance alignment method, which comprises the following steps of S1, constructing a dual graph, and initializing an instance embedding matrix, a relationship embedding matrix and an attribute embedding matrix; s2, learning relation perception instance embedding by using a graph convolutional neural network, and generating a first layer instance embedding optimization matrix; s3, coding neighbor structures and attribute information of instances through a graph convolutional neural network to obtain a second layer instance embedding optimization matrix; s4, constructing an encoder of a multi-head attention mechanism, and obtaining a final instance embedding matrix and a relation embedding matrix; and S5, determining a final instance alignment result based on an attribute-participated multi-dimensional adaptive convolution decoder. According to the method, the dual graph convolutional neural network, the multi-head attention mechanism and the interactive learning of the dual graph and the original graph are utilized, and the graph convolutional network with the expressway network is combined for coding, so that the accuracy of instance alignment is improved.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graphs, and in particular to a method for aligning multi-dimensional feature fusion instances of equipment knowledge graphs. Background Art

[0002] With the rapid development of big data and artificial intelligence technology, knowledge graphs, as a structured semantic network, have shown great application potential in multiple fields such as intelligent equipment and biomedicine. Knowledge graphs effectively organize and represent complex information in the real world in the form of entity, relationship and attribute triples. However, most current instance alignment research in the field of equipment mainly relies on the structural information of knowledge graphs, while ignoring the deep mining of local semantic information. These methods often only focus on the connection relationship between instance nodes, but lack the effective use of relationship and attribute semantics.

[0003] As a key technology to solve the problem of entity inconsistency between knowledge graphs, instance alignment aims to identify and match instances in different graphs that actually point to the same entity. Traditional instance alignment methods mostly rely on manually defined features or rules, which is not only time-consuming and labor-intensive, but also difficult to cope with large-scale and complex knowledge graphs. In recent years, instance alignment methods based on representation learning have gradually become a research hotspot, especially using methods such as graph neural networks and translation models to achieve instance alignment by mapping instances to low-dimensional vector space and calculating the similarity between vectors, which significantly improves the efficiency and accuracy of alignment.

[0004] However, current instance alignment methods based on representation learning still have some limitations. On the one hand, most methods only use the structural information of knowledge graphs, ignoring the in-depth mining of relational semantics and attribute semantics, resulting in limited alignment effects. On the other hand, structural differences and noisy data between different knowledge graphs also pose challenges to instance alignment. Therefore, how to effectively integrate relational semantics and attribute semantics to improve the accuracy and robustness of instance alignment has become a key issue that needs to be urgently addressed in the current knowledge graph field. Summary of the invention

[0005] The purpose of the present invention is to provide an equipment knowledge graph multi-dimensional feature fusion instance alignment method to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for aligning multi-dimensional feature fusion instances of equipment knowledge graph, comprising the following steps:

[0007] Step S1: construct the dual graph of the equipment knowledge graph and initialize the instance embedding matrix, relationship embedding matrix and attribute embedding matrix;

[0008] Step S2: Use the dual primitive graph convolutional neural network to learn relation-aware instance embedding, optimize the neighbor structure representation of the instance, and generate the first-layer instance embedding optimization matrix;

[0009] Step S3: further encode the neighborhood structure and attribute information of the instance through the graph convolutional neural network to obtain the second-layer instance embedding optimization matrix, and fuse it with the initial instance embedding matrix to obtain the instance embedding fusion matrix;

[0010] Step S4: construct an encoder of the multi-head attention mechanism, train the instance embedding fusion matrix and the relationship embedding matrix, and obtain the final instance embedding matrix and the relationship embedding matrix;

[0011] Step S5: Using a multi-dimensional adaptive convolution decoder based on attribute participation, the attributes of the instance pairs to be aligned are embedded in the feature matrix and input into the decoder. The decoder calculates the score of each pair of feature matrices and determines the final instance alignment result based on the highest score.

[0012] Preferably, in the generation of the dual graph in step S1, each planar graph G has a dual planar graph with the following properties:

[0013] Each point in corresponds to a face in G, and for each edge e in G;

[0014] e belongs to two faces, add edge (,);

[0015] e belongs to only one face f, add the back edge (,) to the dual graph;

[0016] Calculate the maximum flow in the original graph.

[0017] Preferably, the convolutional neural network learning relationship perception in step S2 includes the following steps:

[0018] Step S21: multi-scale similarity measurement, by comparing images at different resolutions and scales;

[0019] Step S22: image pyramid and multi-scale filter are used to create a multi-scale representation of the image, so as to perform cross-scale similarity comparison;

[0020] Step S23: Generate an image similar to the target image by using a generative adversarial network generation model;

[0021] Step S24: comparing the difference between the generated image and the target image, and measuring the image similarity;

[0022] Step S25: using an approximate nearest neighbor neighborhood mapping algorithm to establish a neighbor search graph of the image;

[0023] Step S26: Calculate the similarity between images based on the connectivity in the nearest neighbor search graph.

[0024] Preferably, encoding the neighborhood structure and attribute information of the instance in step S3 comprises the following steps:

[0025] Step S31: converting the image into a compact binary hash value through a perceptual hash algorithm, which represents the visual perception features of the image;

[0026] Step S32: The similarity measure is based on the Hamming distance of the hash value, and a small distance indicates a high similarity;

[0027] Step S33: Divide the image into small blocks and calculate the local binary pattern of the pixel gray value in each block;

[0028] Step S34: The similarity measure is based on the similarity of the LBP histogram, which counts the frequencies of different LBP patterns.

[0029] Preferably, the neighborhood structure measurement method in step S3 includes:

[0030] Step S35: extracting semantic features of the image using a convolutional neural network deep learning model;

[0031] Step S36: comparing the similarity of the semantic features to measure the semantic similarity of the images;

[0032] Step S37: Combining different types of features to perform similarity measurement.

[0033] Preferably, the dual primitive graph convolutional neural network utilizes the neighbor structure information of instance relations and attributes to optimize instance embedding representation, and performs iterative optimization through a multi-layer network structure.

[0034] Preferably, the dual primitive graph convolutional neural network simultaneously learns multiple different feature representations by introducing a multi-head attention mechanism:

[0035] Initialize the instance embedding matrix X0, the relationship embedding matrix R, and the attribute embedding matrix A;

[0036] For each pair of instances ei and ej, the similarity coefficient between them is calculated as follows:

[0037] Among them, the concatenated high-dimensional features are mapped to a scalar, which is implemented by the fully connected layer. and are the node representations in, corresponding to the relation representation in, are the vector concatenation functions, are the weights of each relation in the dual relation graph, and are used to distinguish the semantics of different relations. They are calculated based on the possibility that two different relations in the original graph share similar head instances or tail instances.

[0038] Then introduce the multi-head attention mechanism, calculate the output representation of each attention head, and for the kth attention head, calculate the attention score. The calculation formula is:

[0039] Among them, is the activation function ReLU, is the number of attention heads, is the index set of neighbors, is Softmax, is the similarity coefficient calculated by the th head, and represents the representation of the node in the dual graph.

[0040] Preferably, the multi-head attention is used to integrate the entity representations in the original graph:

[0041] The model first needs to perform a linear transformation on the input node to enable the node to have a more advanced feature expression capability. The similarity coefficient calculated by the model is as follows:

[0042] Among them, the connected high-dimensional features are mapped to a scalar, and the dual representation of the relationship between the instance and is realized by the fully connected layer;

[0043] When learning on the dual graph, the model also uses multi-head attention to integrate the instance representations in the original graph. For the instances in the original graph, the representation calculation formula is as follows:

[0044] Among them, is the activation function, is the number of attention heads, is the index set of the neighbors of the instance in, is Softmax, is the similarity coefficient calculated by the th head, and represents the node representation in the original graph.

[0045] Preferably, the representation learning of the neighbor structure is performed using the dual graph and the original graph, the relational features are integrated into the instance representation, and the neighbor structure information of the knowledge graph is encoded through a graph convolutional neural network with a highway network.

[0046] Preferably, the accuracy of instance alignment is improved by using attribute embedding, and attribute embedding representation is independently learned through a graph convolutional neural network. The convolution calculation formula is as follows:

[0047] Among them, is the combined Laplace, is the layer-specific trainable weight matrix of the th layer in GCN, is the number of features of the th layer, and is the activation function ReLU.

[0048] Technical effects and advantages of the present invention:

[0049] The present invention utilizes a dual graph convolutional neural network and a multi-head attention mechanism to accurately capture and integrate relational features. Through interactive learning of the dual graph and the original graph, the model can more comprehensively understand the relationship between instances, thereby improving the accuracy of alignment. In addition, the model combines a graph convolutional network with a highway network to encode the neighbor structure information of the knowledge graph, further improving the efficiency and effect of feature learning. By processing instance attribute information through an independent GCN module, the model can effectively align instances in the absence of sufficient neighbors. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the framework of the multi-dimensional feature fusion instance alignment method of the equipment knowledge graph of the present invention.

[0051] Figure 2 This is a schematic diagram of the convolutional neural network learning relationship perception process of the present invention.

[0052] Figure 3 It is a schematic diagram of the neighbor structure and attribute information flow of the encoding example of the present invention.

[0053] Figure 4 Schematic diagram of the process of measuring the neighbor structure of the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] The present invention provides Figure 1-Figure 4 The equipment knowledge graph multi-dimensional feature fusion instance alignment method shown includes the following steps:

[0056] Step S1: construct the dual graph of the equipment knowledge graph and initialize the instance embedding matrix, relationship embedding matrix and attribute embedding matrix;

[0057] Step S2: Use the dual primitive graph convolutional neural network to learn relation-aware instance embedding, optimize the neighbor structure representation of the instance, and generate the first-layer instance embedding optimization matrix;

[0058] Step S3: further encode the neighborhood structure and attribute information of the instance through the graph convolutional neural network to obtain the second-layer instance embedding optimization matrix, and fuse it with the initial instance embedding matrix to obtain the instance embedding fusion matrix;

[0059] Step S4: construct an encoder of the multi-head attention mechanism, train the instance embedding fusion matrix and the relationship embedding matrix, and obtain the final instance embedding matrix and the relationship embedding matrix;

[0060] Step S5: Using a multi-dimensional adaptive convolution decoder based on attribute participation, the attributes of the instance pairs to be aligned are embedded in the feature matrix and input into the decoder. The decoder calculates the score of each pair of feature matrices and determines the final instance alignment result based on the highest score.

[0061] Attribute triplets account for a large proportion of knowledge graphs. Therefore, by considering the attribute information of instances, the model can combine multi-faceted information to distinguish candidate instances in the alignment phase, thereby increasing the probability that the correct target instance is discovered. Using the same method to learn instance structure and semantics can improve the performance of models, such as the Inga model and the GCN-Align model.

[0062] The following is a specific example to illustrate the equipment knowledge graph multi-dimensional feature fusion instance alignment method of the present invention. The data of the equipment knowledge graph multi-dimensional feature fusion instance alignment method comes from the public cross-language dataset DBP15K. DBP15K is built based on DBpedia and includes three sub-datasets: DBP15KZH-EN, DBP15KJA-EN, and DBP15KFR-EN. The present invention utilizes the entity, relationship, attribute, relationship triple, and attribute triple information of DBP15K.

[0063] In step S1, the dual graph is generated. Every planar graph G has a dual planar graph with the following properties:

[0064] Each point in corresponds to a face in G, and for each edge e in G;

[0065] e belongs to two faces, add edge (,);

[0066] e belongs to only one face f, add the back edge (,) to the dual graph;

[0067] Calculate the maximum flow in the original graph.

[0068] In step S2, the convolutional neural network learning relationship perception includes the following steps:

[0069] Step S21: multi-scale similarity measurement, by comparing images at different resolutions and scales;

[0070] Step S22: image pyramid and multi-scale filter are used to create a multi-scale representation of the image, so as to perform cross-scale similarity comparison;

[0071] Step S23: Generate an image similar to the target image by using a generative adversarial network generation model;

[0072] Step S24: comparing the difference between the generated image and the target image, and measuring the image similarity;

[0073] Step S25: using an approximate nearest neighbor neighborhood mapping algorithm to establish a neighbor search graph of the image;

[0074] Step S26: Calculate the similarity between images based on the connectivity in the nearest neighbor search graph.

[0075] By introducing the methods of relationship perception and attribute participation, the accuracy of instance alignment is improved. First, the relationship information is integrated into the instance representation through the dual original graph convolutional neural network, and the interactive learning of the dual graph and the original graph is used to capture the relationship characteristics. Specifically, the dual graph is built on the basis of the original graph, and the nodes and edges are swapped to form a new graph structure. The complex relationship characteristics are accurately captured through the improved attention mechanism and convolution operation. In addition, through the multi-head attention mechanism, multiple different feature representations can be learned at the same time, thereby improving the ability to model relationships, and integrating the learning process of the dual graph and the original graph to obtain a more accurate instance representation.

[0076] In order to further improve the alignment accuracy, the method also introduces the encoding of the knowledge graph neighbor structure information, and uses a graph convolutional network with a highway network to process the neighbor information. This process extracts the spatial features between nodes by calculating the Laplacian matrix of the graph, thereby improving the computational efficiency and feature learning capabilities. At the same time, in order to deal with instances that lack sufficient neighbors, the attribute information of the instance is learned through an independent GCN module, and the structure and attribute features are combined to improve the instance alignment effect. Through this comprehensive method, the complex relationship between instances can be better captured, and more accurate instance alignment can be achieved.

[0077] The encoding of the neighborhood structure and attribute information of the instance in step S3 includes the following steps:

[0078] Step S31: converting the image into a compact binary hash value through a perceptual hash algorithm, which represents the visual perception features of the image;

[0079] Step S32: The similarity measure is based on the Hamming distance of the hash value, and a small distance indicates a high similarity;

[0080] Step S33: Divide the image into small blocks and calculate the local binary pattern of the pixel gray value in each block;

[0081] Step S34: The similarity measure is based on the similarity of the LBP histogram, which counts the frequencies of different LBP patterns.

[0082] The measurement method of the neighbor structure in step S3 includes:

[0083] Step S35: extracting semantic features of the image using a convolutional neural network deep learning model;

[0084] Step S36: comparing the similarity of the semantic features to measure the semantic similarity of the images;

[0085] Step S37: Combining different types of features to perform similarity measurement.

[0086] The present invention also uses 30% of the seed set as training data and the rest of the seed set as test data. The stochastic gradient descent method is used to optimize the model in the attribute embedding part. The present invention uses Tensorflow, an open source machine learning platform, which provides a wide range of tools, libraries, and community resources. After multiple experimental verifications, the final hyperparameter settings of the model are shown in Table 1.

[0087] Table 1 Model hyperparameter settings

[0088]

[0089] The present invention compares the performance of each model based on the value of the evaluation index. The evaluation indexes used are Hits@k (k=1, 10) and MRR. The higher their values ​​are, the better the model performance is.

[0090] The dual primitive graph convolutional neural network utilizes the neighbor structure information of instance relations and attributes to optimize the instance embedding representation and performs iterative optimization through a multi-layer network structure.

[0091] To evaluate the proposed model, we will compare it with some competitive entity alignment methods, which mainly include translation-based models and graph neural network-based models.

[0092] Translation-based models: JAPE, BootEA, NAEA.

[0093] Models based on graph neural networks: GCN-Align, HGCN, NMN, RAGA.

[0094] The experimental results show that the performance of the proposed model on the three datasets is better than almost all the comparison methods. The Hits@1 index reflects the accuracy of entity alignment. The experimental results are shown in Table 2.

[0095] Table 2 Comparison of model instance alignment performance

[0096]

[0097] It can be seen that the performance of the model of the present invention is better than other models because it makes comprehensive use of the dual graph convolutional neural network and the multi-head attention mechanism to accurately capture and incorporate relational features. Through interactive learning of the dual graph and the original graph, the model can more comprehensively understand the relationship between instances, thereby improving the accuracy of alignment. In addition, the model combines the encoding of the knowledge graph neighbor structure information with the graph convolutional network with a highway network, further improving the efficiency and effectiveness of feature learning. By processing instance attribute information through an independent GCN module, the model can still effectively align instances in the absence of sufficient neighbors.

[0098] The dual original graph convolutional neural network introduces a multi-head attention mechanism to simultaneously learn multiple different feature representations:

[0099] Initialize the instance embedding matrix X0, the relationship embedding matrix R, and the attribute embedding matrix A;

[0100] For each pair of instances ei and ej, the similarity coefficient between them is calculated as follows:

[0101] Among them, the concatenated high-dimensional features are mapped to a scalar, which is implemented by the fully connected layer. and are the node representations in, corresponding to the relation representation in, are the vector concatenation functions, are the weights of each relation in the dual relation graph, and are used to distinguish the semantics of different relations. They are calculated based on the possibility that two different relations in the original graph share similar head instances or tail instances.

[0102] Then introduce the multi-head attention mechanism, calculate the output representation of each attention head, and for the kth attention head, calculate the attention score. The calculation formula is:

[0103] Among them, is the activation function ReLU, is the number of attention heads, is the index set of neighbors, is Softmax, is the similarity coefficient calculated by the th head, and represents the representation of the node in the dual graph.

[0104] Use multi-head attention to integrate entity representations in the original graph:

[0105] The model first needs to perform a linear transformation on the input node to enable the node to have a more advanced feature expression capability. The similarity coefficient calculated by the model is as follows:

[0106] Among them, the connected high-dimensional features are mapped to a scalar, and the dual representation of the relationship between the instance and is realized by the fully connected layer;

[0107] When learning on the dual graph, the model also uses multi-head attention to integrate the instance representations in the original graph. For the instances in the original graph, the representation calculation formula is as follows:

[0108] Among them, is the activation function, is the number of attention heads, is the index set of the neighbors of the instance in, is Softmax, is the similarity coefficient calculated by the th head, and represents the node representation in the original graph.

[0109] The dual graph and the original graph are used to learn the representation of the neighborhood structure, the relational features are integrated into the instance representation, and the neighborhood structure information of the knowledge graph is encoded through a graph convolutional neural network with a highway network.

[0110] Attribute embedding is used to improve the accuracy of instance alignment. The attribute embedding representation is independently learned through the graph convolutional neural network. The convolution calculation formula is as follows:

[0111] Among them, is the combined Laplace, is the layer-specific trainable weight matrix of the th layer in GCN, is the number of features of the th layer, and is the activation function ReLU.

[0112] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for aligning instances of multi-dimensional feature fusion of equipment knowledge graph, characterized in that: The following steps are involved: Step S1: construct the dual graph of the equipment knowledge graph and initialize the instance embedding matrix, relationship embedding matrix and attribute embedding matrix; Step S2: Use the dual primitive graph convolutional neural network to learn relation-aware instance embedding, optimize the neighbor structure representation of the instance, and generate the first-layer instance embedding optimization matrix; Step S3: further encode the neighborhood structure and attribute information of the instance through the graph convolutional neural network to obtain the second-layer instance embedding optimization matrix, and fuse it with the initial instance embedding matrix to obtain the instance embedding fusion matrix; Step S4: construct an encoder of the multi-head attention mechanism, train the instance embedding fusion matrix and the relationship embedding matrix, and obtain the final instance embedding matrix and the relationship embedding matrix; Step S5: Using a multi-dimensional adaptive convolution decoder based on attribute participation, the attributes of the instance pairs to be aligned are embedded in the feature matrix and input into the decoder. The decoder calculates the score of each pair of feature matrices and determines the final instance alignment result based on the highest score.

2. According to claim 1, a method for aligning instances of multi-dimensional feature fusion of equipment knowledge graph is characterized in that: In the generation of the dual graph in step S1, each planar graph G has a dual planar graph with the following properties: Each point in corresponds to a face in G, and for each edge e in G; e belongs to two faces, add edge (,); e belongs to only one face f, add the back edge (,) to the dual graph; Calculate the maximum flow in the original graph.

3. According to claim 1, a method for aligning instances of multi-dimensional feature fusion of equipment knowledge graph is characterized in that: The convolutional neural network learning relationship perception in step S2 includes the following steps: Step S21: multi-scale similarity measurement, by comparing images at different resolutions and scales; Step S22: image pyramid and multi-scale filter are used to create a multi-scale representation of the image, so as to perform cross-scale similarity comparison; Step S23: Generate an image similar to the target image by using a generative adversarial network generation model; Step S24: comparing the difference between the generated image and the target image, and measuring the image similarity; Step S25: using an approximate nearest neighbor neighborhood mapping algorithm to establish a neighbor search graph of the image; Step S26: Calculate the similarity between images based on the connectivity in the nearest neighbor search graph.

4. According to claim 1, a method for aligning instances of multi-dimensional feature fusion of equipment knowledge graphs is characterized in that: The encoding of the neighbor structure and attribute information of the instance in step S3 comprises the following steps: Step S31: converting the image into a compact binary hash value through a perceptual hash algorithm, which represents the visual perception features of the image; Step S32: The similarity measure is based on the Hamming distance of the hash value, and a small distance indicates a high similarity; Step S33: Divide the image into small blocks and calculate the local binary pattern of the pixel gray value in each block; Step S34: The similarity measure is based on the similarity of the LBP histogram, which counts the frequencies of different LBP patterns.

5. According to claim 1, a method for aligning instances of multi-dimensional feature fusion of equipment knowledge graph is characterized in that: The method for measuring the neighbor structure in step S3 includes: Step S35: extracting semantic features of the image using a convolutional neural network deep learning model; Step S36: comparing the similarity of the semantic features to measure the semantic similarity of the images; Step S37: Combining different types of features to perform similarity measurement.

6. The method for aligning instances of multi-dimensional feature fusion of equipment knowledge graph according to claim 1 is characterized in that: The dual primitive graph convolutional neural network utilizes the neighbor structure information of instance relations and attributes to optimize instance embedding representation, and performs iterative optimization through a multi-layer network structure.

7. The method for aligning instances of multi-dimensional feature fusion of equipment knowledge graph according to claim 1 is characterized in that: The dual primitive graph convolutional neural network introduces a multi-head attention mechanism to simultaneously learn multiple different feature representations: Initialize the instance embedding matrix X0, the relationship embedding matrix R, and the attribute embedding matrix A; For each pair of instances ei and ej, the similarity coefficient between them is calculated as follows: Among them, the concatenated high-dimensional features are mapped to a scalar, which is implemented by the fully connected layer. and are the node representations in, corresponding to the relation representation in, are the vector concatenation functions, are the weights of each relation in the dual relation graph, and are used to distinguish the semantics of different relations. They are calculated based on the possibility that two different relations in the original graph share similar head instances and tail instances. Then introduce the multi-head attention mechanism, calculate the output representation of each attention head, and for the kth attention head, calculate the attention score. The calculation formula is: Among them, is the activation function ReLU, is the number of attention heads, is the index set of neighbors, is Softmax, is the similarity coefficient calculated by the th head, and represents the representation of the node in the dual graph.

8. The method for aligning instances of multi-dimensional feature fusion of equipment knowledge graph according to claim 1 is characterized in that: The multi-head attention is used to integrate the entity representation in the original graph: The model first needs to perform a linear transformation on the input node to enable the node to have a more advanced feature expression capability. The similarity coefficient calculated by the model is as follows: Among them, the connected high-dimensional features are mapped to a scalar, and the dual representation of the relationship between the instance and is realized by the fully connected layer; When learning on the dual graph, the model also uses multi-head attention to integrate the instance representations in the original graph. For the instances in the original graph, the representation calculation formula is as follows: Among them, is the activation function, is the number of attention heads, is the index set of the neighbors of the instance in, is Softmax, is the similarity coefficient calculated by the th head, and represents the node representation in the original graph.

9. The method for aligning instances of multi-dimensional feature fusion of equipment knowledge graph according to claim 1 is characterized in that: The dual graph and the original graph are used to learn the representation of the neighbor structure, the relational features are integrated into the instance representation, and the neighbor structure information of the knowledge graph is encoded through a graph convolutional neural network with a highway network.

10. The method for aligning instances of multi-dimensional feature fusion of equipment knowledge graph according to claim 1, characterized in that: The accuracy of instance alignment is improved by using attribute embedding. The attribute embedding representation is independently learned through the graph convolutional neural network. The convolution calculation formula is as follows: Among them, is the combined Laplace, is the layer-specific trainable weight matrix of the th layer in GCN, is the number of features of the th layer, and is the activation function ReLU.

Images