Satellite knowledge graph rapid migration method based on entity similarity

Abstract

The invention discloses a satellite knowledge graph rapid migration method based on entity similarity. The method comprises the following steps: 1, obtaining an attribute graph and an initial entity embedding vector according to a new satellite knowledge graph to be supplemented and an old satellite knowledge graph; 2, training a graph neural network and optimizing network parameters; 3, the trained graph neural network takes the attribute graph structure and the initial entity embedding vector as input to obtain an entity attribute embedding vector containing complete semantic information; 4, obtaining a similarity matrix through an entity attribute body embedding vector, and obtaining entity matching of a new satellite knowledge graph to be supplemented with a correct structure and an old satellite knowledge graph through a graph structure matching assignment matrix; and 5, through entity matching, the entities which exist in the old satellite knowledge graph and do not exist in the new satellite knowledge graph to be supplemented are migrated, and a complete new satellite knowledge graph is output, so that the efficiency of constructing the satellite knowledge graph and the quality of the knowledge graph are improved, and a large amount of time and manpower are saved.

Description

Technical Field

[0001] The present invention relates to a knowledge graph, and in particular to a method for rapidly migrating a satellite knowledge graph based on entity similarity. Background Art

[0002] A knowledge graph is a structured semantic network used to represent the relationships and attributes between entities. Building a knowledge graph usually relies on a large amount of manual annotation and automated extraction technology to extract entities, relationships, and attributes from text, structured data, or semi-structured data to form a graph structure.

[0003] Knowledge graphs can visualize, integrate, analyze, and apply the interconnections between knowledge resources and their carriers, providing deep semantic understanding and efficient retrieval. In the aerospace field, satellite knowledge graphs can integrate all relevant knowledge resources about satellites, playing a vital role in improving the efficiency of satellite management and monitoring. However, the number of satellites is large and diverse, and many satellites of the same family and model used for navigation, communication, and other functions are often highly similar. This makes the traditional method of constructing a knowledge graph for each satellite cumbersome, labor-intensive, and repetitive, wasting a lot of manpower and time. Therefore, new technologies are urgently needed to address these challenges. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a method for rapid migration of satellite knowledge graphs based on entity similarity, which improves the efficiency of constructing satellite knowledge graphs, saves a lot of time and human resources, and improves the accuracy of the new knowledge graph.

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

[0006] A satellite knowledge graph rapid migration method based on entity similarity includes the following steps:

[0007] Step 1: Convert the new satellite knowledge graph to be supplemented and the complete old satellite knowledge graph into an attribute graph. Use a method similar to "one-hot" to encode different attributes of the entity to form an initial entity embedding vector containing semantic information.

[0008] Step 2: Train the graph neural network and optimize the network parameters;

[0009] Step 2.1: Based on the “view horizon” theory, different view horizons are assigned to different types of node entities in the attribute graph, so that entities connect to a wider range of entities as neighbors. The optimized adjacency matrix is obtained as the input of the graph neural network.

[0010] Step 2.2: Randomly select point pairs in the attribute graph to construct samples. In the process of selecting point pairs, combine cross-graph point comparison and weighted node pairs to enable the graph neural network to learn all attribute graphs and focus on learning central range entities.

[0011] Step 2.3: Under unsupervised learning conditions, use the constructed samples to train the graph neural network;

[0012] Step 2.4: Use the mean square error (MSE) as the loss function to optimize the parameters of the graph neural network and obtain the trained graph neural network.

[0013] Step 3: The trained graph neural network takes the attribute graph structure and the initial entity embedding vector as input to obtain the entity attribute embedding vector containing structural and semantic information;

[0014] Step 4: By comparing the entity attribute embedding vectors in the new satellite knowledge graph to be supplemented and the complete old satellite knowledge graph, a similarity matrix is obtained. By matching the assignment matrix through the graph structure, the entity matching of the new satellite knowledge graph to be supplemented and the old satellite knowledge graph with the correct structure is obtained.

[0015] Step 5: Through entity matching, entities that exist in the old satellite knowledge graph but do not exist in the new satellite knowledge graph to be supplemented are regarded as missing entities of the new satellite knowledge graph to be supplemented, and they are migrated to the new knowledge graph to be supplemented, and a complete new satellite knowledge graph is output.

[0016] Furthermore, in step 1, the "one-hot" method is used to encode different attributes of the entity. The process is as follows: the embedding vectors corresponding to all attributes in the attribute graph are horizontally superimposed, and the attribute values that the entity does not have are left blank; the numerical attributes are directly applied with their values as the embedding vectors, and the character attributes are generated by character-by-character hashing. The generated embedding vector is represented as:

[0017] H=concat(σ(H i ))

[0018] In the formula, concat is the connection operation, σ refers to the normalization operation, H i Represents the embedding vector of a single attribute, H does not have the attribute i =0.

[0019] Furthermore, the graph neural network in step 2 is a 4-layer or 6-layer graph convolutional neural network GCN, and the graph convolutional neural network GCN adopts a residual network structure, and its forward transmission process is expressed as:

[0020]

[0021] Where H (l)is the entity attribute vector of the lth layer, σ(·) represents the relu function, ω i Represents the entity attribute weight of the i-th layer, which is the characteristic of the residual network. is the normalized adjacency matrix, A (l) is a parameter.

[0022] Furthermore, the optimizer used for training the graph neural network in step 2 is the Adam optimizer, and the initial learning rate is set to 0.03.

[0023] Furthermore, the specific process of step 2.1 is as follows:

[0024] If the neighbor is more important than the entity itself, only this neighbor is connected. Otherwise, the neighbor range is recursively extended to the neighbors with lower importance than itself. All entities within the field of view are considered to be connected to the entity, that is, the corresponding value in the adjacency matrix is not 0. At the same time, different weights w are given according to the original distance D between the entities, expressed as:

[0025]

[0026] Where b is the view weight base, which is generally set to 2. The farther the distance, the lower the weight of the entity, and the closer the corresponding value in the adjacency matrix is to 0.

[0027] Furthermore, the specific process of step 4 is as follows:

[0028] Step 4.1. Calculate the cosine similarity of each pair of embedding vectors containing structural and semantic information obtained in step 3. A similarity matrix is obtained for each pair of satellite knowledge graphs. The horizontal axis represents each entity in the new satellite knowledge graph, and the vertical axis represents each entity in the old satellite knowledge graph. The matrix value is the cosine similarity of the corresponding entity embedding vector.

[0029] Step 4.2: Calculate the assignment matrix and the similarity evaluation values of the old satellite knowledge graph and the new satellite knowledge graph through the similarity matrix, and select the old knowledge graph with the highest similarity evaluation value and its corresponding relationship with the new knowledge graph.

[0030] Furthermore, in step 4.2, the process of calculating the similarity evaluation value of the assignment matrix and the old satellite knowledge graph and the new satellite knowledge graph through the similarity matrix is as follows:

[0031] Step 4.2.1. Align the main entities of the old satellite knowledge graph with the new satellite knowledge graph;

[0032] Step 4.2.2: Take a pair of corresponding entities a and b, and calculate the local similarity matrix and the assignment relationship between neighbors based on the neighbors of a and b. Let the neighbors of a be set A, and the neighbors of b be set B. For a pair of entities in A and B, if one of the entities has no uncorresponding neighbors, the value of its similarity matrix is the cosine similarity of this entity pair; if both entities have uncorresponding neighbors, the corresponding value of their similarity matrix is also called the matching degree of the entity pair. Recursively enter step 4.2.2, and after obtaining the similarity matrix, calculate the assignment matrix through the assignment matrix generation step of the bipartite graph. The matching degree of entity pair a and b is the sum of the matching degrees of all entity pairs indicated by the assignment matrix plus the cosine similarity between a and b.

[0033] Step 4.2.3: When all entities in the new satellite knowledge graph have corresponding entities in the old knowledge graph, the assignment matrix is calculated and the similarity evaluation value is the sum of the cosine similarity of the main entity and the matching degree of its neighbors.

[0034] Furthermore, the steps for generating the assignment matrix through the bipartite graph in step 4.2.2 are as follows:

[0035] Step 4.2.2.1. Initialization

[0036] The similarity matrix is denoted as w, i and j are used to record the new satellite knowledge graph entities and the old satellite knowledge graph entities respectively. i and j are also called top marks. The array u is initialized to the maximum weight of all i, expressed as: u[i] = max(w[i][j]), and the array v is initialized to 0;

[0037] Step 4.2.2.2: Finding a Perfect Match

[0038] Use the augmenting path method to find the maximum weight match. Select an unmatched new satellite knowledge graph entity i. If i has not been matched, find an augmenting path by adjusting the top indices u[i] and v so that the corresponding old satellite knowledge graph entity j can be matched. If j has been matched, find a better matching solution through the augmenting path.

[0039] Step 4.2.2.3, adjust the top mark:

[0040] If an augmenting path is found, adjust the values of u and v according to the matching situation on the augmenting path to find a better match. Specifically, for the unmatched new satellite knowledge graph entity i, increase the value of u[i], expressed as u[i]=u[i]+a, where a is a positive number, usually the smallest non-zero weight in the augmenting path; for the old satellite knowledge graph entity j, reduce the value of v[j], expressed as v[j]=v[j]-a; for the matched edge (i, j), increase the value of u[i] and reduce the value of v[j] so that the value of u[i]+v[j] can reflect the weight of the edge (i, j);

[0041] Step 4.2.2.4: Repeat steps 4.2.2.2 to 4.2.2.3, repeatedly searching for augmenting paths until no augmenting paths can be found.

[0042] Step 4.2.2.5: Output results

[0043] The obtained u and v arrays contain the top label information of the optimal match, and the specific edge set with the maximum weight match is determined through the top label information.

[0044] Compared with the prior art, the present invention has the following technical effects:

[0045] On the one hand, the graph neural network of the present invention can be implemented through Python scripts, is easy to deploy in the field, supports a variety of input formats, uses a solution similar to "single hot" to handle the attribute embedding problem of entities, efficiently utilizes the semantic similarity of entities, and improves the accuracy of knowledge graph migration; on the other hand, in order to improve the performance of the graph neural network, the following measures are taken during the training process: First, the intermediate files of the satellite knowledge graph and the graph neural network are stored independently, which is convenient for high reuse of training results when used on a large scale / multi-batch basis, thereby achieving network performance acceleration; second, based on the "field of view" theory, the adjacency matrix is optimized, and different "fields of view" are assigned to different types of node entities, so that entities are connected to entities in a wider range as neighbors, breaking the limitations of the traditional adjacency matrix that information transmission is slow and information is easily diluted, thereby improving the accuracy and confidence of the new satellite knowledge graph; third, in response to the need to randomly select knowledge graph entities, the selection weights are modified according to importance, which is reflected in: by comparing cross-graph points, ensuring the utilization of all The graph neural network is trained on the attribute graph to prevent the network from overfitting to a single attribute graph. At the same time, in view of the different physical importance of different entities, a weighted node pair selection scheme is adopted to improve the accuracy of network knowledge graph migration and the training efficiency of the network itself. Therefore, according to the characteristics of the satellite knowledge graph, the training effect of the central range entities is emphasized so that they are not diluted by a large number of low-weight entities on the edge, thereby improving the training effect and optimizing the training speed and efficiency. On the other hand, the initial entity embedding vector is first optimized by the trained graph neural network to contain more structural and semantic information. Then, by comparing the initial entity embedding vectors in the new satellite knowledge graph to be supplemented and the complete old satellite knowledge graph, a similarity matrix is obtained, and then an assignment matrix is generated to complete the rapid migration from the old satellite knowledge graph to the new satellite knowledge graph to be supplemented, which improves the efficiency of constructing the satellite knowledge graph, saves a lot of time and human resources, and makes the generated new satellite knowledge graph structure more complete and accurate with higher confidence. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flowchart of the satellite knowledge graph migration process of the present invention;

[0047] Figure 2 This is a schematic diagram of the entity initial embedding vector process of the present invention;

[0048] Figure 3 The satellite knowledge graph migration process and graph convolutional neural network training process of the present invention;

[0049] Figure 4(a) is a schematic diagram of the structure of the graph neural network M-GCN-4 of the present invention;

[0050] Figure 4(b) is a schematic diagram of the structure of the graph neural network M-GCN-6 of the present invention;

[0051] Figure 5 It is the view of the subsystem type entity and parameter type entity of the present invention;

[0052] Figure 6(a) shows the new satellite knowledge graph to be supplemented;

[0053] Figure 6(b) shows the knowledge graph of the old satellite A;

[0054] Figure 6(c) shows the knowledge graph of the old satellite B;

[0055] Figure 6(d) shows the knowledge graph of the old satellite C;

[0056] Figure 7 Schematic diagram of the initial embedding vector of the entity Figure 1 ;

[0057] Figure 8 Schematic diagram of the initial embedding vector of the entity Figure 2 ;

[0058] Figure 9 is the similarity matrix between the new satellite and the old satellite A;

[0059] Figure 10 Schematic diagram of the complete new satellite knowledge graph after migration. DETAILED DESCRIPTION

[0060] The specific contents of the present invention are further explained in detail below with reference to the embodiments.

[0061] The new satellite knowledge graph to be supplemented mentioned in this embodiment only contains a small amount of information describing the main features of the new satellite, while the old satellite knowledge graph contains sufficient and complete feature information of the old satellite. By comparing the similarity between the old satellite knowledge graph and the new satellite knowledge graph to be supplemented, appropriate entities are migrated from the old satellite knowledge graph to the new satellite knowledge graph to be supplemented, and finally a complete new satellite knowledge graph is output for efficient management and monitoring of the new satellite, facilitating maintenance.

[0062] like Figure 1 and 3 As shown in FIG, a satellite knowledge graph rapid migration method based on entity similarity includes the following steps:

[0063] Step 1: Convert the new satellite knowledge graph to be supplemented and the complete old satellite knowledge graph into an attribute graph. Use a method similar to "one-hot" to encode different attributes of the entity to form an initial entity embedding vector containing semantic information.

[0064] like Figure 2As shown in the figure, a scheme similar to "one-hot" is used to deal with the attribute embedding problem of entities. The process of encoding different attributes of an entity is as follows: the embedding vectors corresponding to all attributes are stacked horizontally, and the attribute values that the entity does not have are left blank; the numerical attributes are directly applied with their values as the embedding vectors, and the character attributes are generated using a character-by-character hashing scheme to generate embedding vectors, thereby efficiently utilizing the semantic similarity of the entities and improving the accuracy. The embedding vector is expressed as:

[0065] H=concat(σ(H i ))

[0066] In the formula, concat is the connection operation, σ refers to the normalization operation, H i Represents the embedding vector of a single attribute, H does not have the attribute i =0, Figure 2 In , the name is embedded as H0, the type is embedded as H1, and the entity weight is embedded as H2;

[0067] Step 2: Train the graph neural network and optimize the network parameters;

[0068] The graph neural network is a 4-layer or 6-layer graph convolutional neural network (GCN) as shown in Figure 4(a) and Figure 4(b), and the graph convolutional neural network (GCN) adopts a residual network structure. The 4-layer or 6-layer graph convolutional neural network (GCN) is denoted as M-GCN-4 and M-GCN-6, respectively, breaking the shackles of the traditional graph convolutional network with shallow layers and small information transmission range. Among them, M-GCN-6 has the best accuracy performance, while M-GCN-4 has higher efficiency while maintaining high accuracy.

[0069] The forward transfer process of M-GCN-4 and M-GCN-6 is expressed as:

[0070]

[0071] Where H (l) is the entity attribute vector of the lth layer, σ(·) represents the relu function, ω i Represents the entity attribute weight of the i-th layer, which is the characteristic of the residual network. is the normalized adjacency matrix, A (l) is a parameter;

[0072] Based on PyTorch, the graph neural network is trained using the Adam optimizer, with the initial learning rate set to 0.03. The training process includes the following steps:

[0073] Step 2.1: Based on the “view horizon” theory, different view horizons are assigned to different types of node entities in the attribute graph, so that entities can connect to a wider range of entities as neighbors, and obtain an optimized adjacency matrix.

[0074] In order to solve the problem of slow node message transmission efficiency in graph neural networks and each layer of the network can only enable nodes to learn information about the neighborhood, the "view" theory is adopted to give different types of node entities different "views", that is, entities connect to entities in a larger range as neighbors, breaking the limitations of the traditional adjacency matrix with slow information transmission speed and easy dilution of information. Figure 5 As shown in Figure 2, different types of entities have different importance, and the process of connecting entities in different ranges and determining the entity's field of view, that is, the new neighbor range, is as follows:

[0075] If the neighbor's importance is higher than the entity itself, only this neighbor is connected. Otherwise, the neighbor range is recursively extended to the neighbors with lower importance than the entity itself. All entities within the field of view are considered to be connected to the entity, that is, the corresponding value in the adjacency matrix is not 0. At the same time, different weights w are given according to the distance D between entities, which can be expressed as:

[0076]

[0077] Where b is the view weight cardinality, which is generally set to 2. The farther the distance, the lower the weight of the entity, and the closer the corresponding value in the adjacency matrix is to 0. The farther the distance, the lower the weight of the entity, and the closer the corresponding value in the adjacency matrix is to 0.

[0078] The optimized adjacency matrix can directly replace the usual adjacency matrix for training graph neural networks. However, normalization is required to ensure the training effect.

[0079] Step 2.2: Use all optimized adjacency matrices as input to the graph neural network. Randomly select point pairs in the attribute graph to construct samples. In the process of selecting point pairs, combine cross-graph point comparison and weighted node pairs to enable the graph neural network to learn all attribute graphs and focus on learning central range entities.

[0080] To improve the module's automation, the training module uses an unsupervised learning approach. Point pairs are randomly selected from the attribute graph to construct samples. The selected nodes in each training session account for approximately two-thirds of the total number of nodes, ensuring sample coverage and training effectiveness. Unlike knowledge graphs in fields like Wikipedia (where adjacent or similar entities have higher similarity), for satellite knowledge graphs, the positive or negative nature of samples is primarily determined by the similarity of their attributes and their neighbors. Therefore, combining knowledge graphs from different fields requires a deep understanding of the relative relationships between entities, which in turn allows for the development of specific learning strategies.

[0081] On the one hand, this embodiment adopts a full-range cross-point comparison scheme, thereby utilizing all attribute graphs to train the graph neural network, preventing the network from overfitting to a single attribute graph;

[0082] On the other hand, in view of the different physical importance of different entities, in order to improve the overall accuracy and training efficiency, a weighted node pair selection scheme is adopted. In the satellite knowledge graph, each knowledge graph has a main entity, and the remaining entities are radially spread out. The closer the entity is to the main entity, the higher the importance, and the importance of the edge entities decreases in turn. They have higher weights in similarity evaluation. Therefore, when training graph neural networks, the training effect of the entities in the center range should be considered. Therefore, when randomly selecting point pairs, this part of entities should also be selected to avoid diluting the training effect by a large number of low-importance entities on the edge, thereby wasting training time. Specifically, according to the shortest path length D between the entity and the main entity, the ratio q of the probability of selecting this entity to the probability of selecting the main entity is expressed as:

[0083]

[0084] In the formula, a is the cardinality of the point selection weight, which is 8;

[0085] Step 2.3: Under unsupervised learning conditions, use the constructed samples to train the graph neural network;

[0086] Step 2.4: Use the mean square error (MSE) as the loss function to optimize the parameters of the graph neural network and obtain the trained graph neural network.

[0087] Step 3: The trained graph neural network takes the attribute graph structure and the initial entity embedding vector as input and outputs the entity attribute embedding vector that contains the most structural and semantic information.

[0088] Step 4: By comparing the initial entity embedding vectors in the new satellite knowledge graph to be supplemented and the complete old satellite knowledge graph, a similarity matrix is obtained. By matching the assignment matrix through the graph structure, the entity matching of the new satellite knowledge graph to be supplemented and the old satellite knowledge graph with the correct structure is obtained. The specific process is as follows:

[0089] Step 4.1. Calculate the cosine similarity of each pair of embedding vectors containing structural and semantic information obtained in step 3. A similarity matrix is obtained for each pair of satellite knowledge graphs. The horizontal axis represents each entity in the new satellite knowledge graph, and the vertical axis represents each entity in the old satellite knowledge graph. The matrix value is the cosine similarity of the corresponding entity embedding vector.

[0090] Step 4.2: Calculate the assignment matrix and the similarity evaluation values of the old satellite knowledge graph and the new satellite knowledge graph through the similarity matrix, and select the old knowledge graph with the highest similarity evaluation value and its corresponding relationship with the new knowledge graph;

[0091] The process of calculating the assignment matrix and the similarity evaluation value of the old satellite knowledge graph and the new satellite knowledge graph through the similarity matrix is as follows:

[0092] Step 4.2.1. Align the main entities of the old satellite knowledge graph with the new satellite knowledge graph;

[0093] Step 4.2.2: Take a pair of corresponding entities a and b, and calculate the local similarity matrix and the assignment relationship between neighbors based on the neighbors of a and b. Let the neighbors of a be set A, and the neighbors of b be set B. For a pair of entities in A and B, if one of the entities has no uncorresponding neighbors, the value of its similarity matrix is the cosine similarity of this entity pair; if both entities have uncorresponding neighbors, the corresponding value of their similarity matrix is also called the matching degree of the entity pair, which is obtained by recursively entering step 4.2.2. After obtaining the similarity matrix, the assignment matrix is generated through the assignment matrix of the bipartite graph. The matching degree of entity pair a and b is the sum of the matching degrees of all entity pairs indicated by the assignment matrix plus the cosine similarity between a and b.

[0094] The steps to generate the assignment matrix through the bipartite graph are:

[0095] Step 4.2.2.1. Initialization

[0096] The similarity matrix is denoted as w, i and j are used to record the new satellite knowledge graph entities and the old satellite knowledge graph entities respectively. i and j are also called top marks. The array u is initialized to the maximum weight of all i, expressed as: u[i] = max(w[i][j]), and the array v is initialized to 0;

[0097] Step 4.2.2.2: Finding a Perfect Match

[0098] Use the augmenting path method to find the maximum weight match. Select an unmatched new satellite knowledge graph entity i. If i has not been matched, find an augmenting path by adjusting the top indices u[i] and v so that the corresponding old satellite knowledge graph entity j can be matched. If j has been matched, find a better matching solution through the augmenting path.

[0099] Step 4.2.2.3, adjust the top mark:

[0100] If an augmenting path is found, adjust the values of u and v according to the matching situation on the augmenting path to find a better match. Specifically, for the unmatched new satellite knowledge graph entity i, increase the value of u[i], expressed as u[i]=u[i]+a, where a is a positive number, usually the smallest non-zero weight in the augmenting path; for the old satellite knowledge graph entity j, reduce the value of v[j], expressed as v[j]=v[j]-a; for the matched edge (i, j), increase the value of u[i] and reduce the value of v[j] so that the value of u[i]+v[j] can reflect the weight of the edge (i, j);

[0101] Step 4.2.2.4: Repeat steps 4.2.2.2 to 4.2.2.3, repeatedly searching for augmenting paths until no augmenting paths can be found.

[0102] Step 4.2.2.5: Output results

[0103] The obtained u and v arrays contain the top label information of the best match. Through the top label information, the specific edge set with the maximum weight match is determined;

[0104] Step 4.2.3: When all entities in the new satellite knowledge graph have corresponding entities in the old knowledge graph, the assignment matrix is calculated and the similarity evaluation value is the sum of the cosine similarity of the main entity and its neighbors.

[0105] Step 5: Through entity matching, entities that exist in the old satellite knowledge graph but do not exist in the new satellite knowledge graph to be supplemented are regarded as missing entities of the new satellite knowledge graph to be supplemented. They are migrated to the new knowledge graph to be supplemented, and a complete new satellite knowledge graph is output. The specific process is as follows:

[0106] By comparing the similarity evaluation values of all old satellite knowledge graphs and the new satellite knowledge graphs to be supplemented, the old satellite knowledge graph with the highest similarity is selected as the migration target of the new knowledge graph to be supplemented. When performing the migration operation, the existing assignment relationship between the old and new satellite knowledge graphs is used to obtain the missing parts of the new satellite knowledge graph compared with the old satellite knowledge graph. These parts are used as the missing parts of the new satellite knowledge graph and migrated in a grafting manner to obtain a complete new satellite knowledge graph. During the migration process, the confidence of the migrated entity is obtained by the similarity of the entities corresponding to the old satellites for manual review and improvement;

[0107] In order to verify the effectiveness of the satellite knowledge graph rapid migration method based on entity similarity proposed in this embodiment, the new satellite knowledge graph and Figure 6(b) to Figure 6(d) The knowledge graph of old satellite A, old satellite B and old satellite C is verified, including the following steps:

[0108] S1. According to Figure 6(a) to Figure 6(d) knowledge graph, generate attribute graph and initial entity embedding vector;

[0109] Figure 6(a) to Figure 6(d) Each circle in the knowledge graph represents an entity, which has different attributes (not marked in the figure). Assuming that all entities include the four attributes of name, type, quality and manufacturer, the initial entity embedding vector of the entity "new satellite" with four attributes in the new satellite knowledge graph to be supplemented is as follows Figure 7 As shown, the entity "high orbit" in the new satellite knowledge graph to be supplemented does not have quality and manufacturer attributes, and its initial entity embedding vector is as follows Figure 8 As shown;

[0110] Figure 7 and Figure 8 In which, each H_subpart indicates that the sub-embedding vector contains the semantic information of this part. For character data, the vector is scaled and normalized to express it as: For numerical data, a single vector dimension is occupied and the value is normalized by the inverse tangent function, expressed as:

[0111] S2. The attribute graph (also known as the graph structure of the knowledge graph or the adjacency matrix optimized by the "view horizon" theory) and the initial entity embedding vector are used as the input of the trained graph neural network. The graph neural network outputs the entity attribute embedding vector containing complete semantic information.

[0112] S3. Calculate the cosine similarity and assignment matrix by embedding the entity attribute body vector. Take the new satellite knowledge graph and the old satellite A knowledge graph as examples to show the steps of calculating the assignment matrix. The similarity, matching degree and other figures involved are only for demonstration.

[0113] 1) Make the main entities of "new satellite" and "old satellite A" correspond;

[0114] 2) Take a pair of corresponding entities, namely "new satellite" and "old satellite A". Among the neighbors of "new satellite" there is a neighbor entity "power subsystem", and among the neighbors of "old satellite A" there are neighbor entities "attitude and orbit control subsystem" and "power subsystem". For these parts in the similarity matrix, recursively calculate the matching degree, and for the other parts, directly calculate the cosine similarity of the corresponding entities;

[0115] 3) Compare the "power subsystem" in the new satellite with the "attitude and orbit control subsystem" in the old satellite A. The cosine similarity between the two is 0.7. The neighbors of the two can be directly compared. The cosine similarity value of the entity embedding of "Y model battery" in the new satellite and "attitude control jet" in the old satellite A is 0.6. Therefore, the value of the similarity matrix between the two in step 2) is 0.7 + 0.6 = 1.3;

[0116] 4) Compare the "power subsystem" in the new satellite with the "power subsystem" in the old satellite A. The cosine similarity between the two is 0.9. The neighbors of the two can be directly compared. The cosine similarity value of the entity embedding of "Y model battery" in the new satellite and "Y model battery" in the old satellite A is 0.9. Therefore, the value of the similarity matrix between the two in step 2) is 0.9 + 0.9 = 1.8.

[0117] For other entity pairs where one of the two has no other neighbors, the cosine similarity can be used to represent the degree of matching. Finally, the similarity matrix between "new satellite" and "old satellite A" is obtained as follows Figure 9 As shown, through this similarity matrix and the assignment matrix calculation method for the bipartite graph, the corresponding relationship between the entities appearing in the similarity matrix can be calculated, and steps 1) to 4) are repeated for these entities with determined corresponding relationships until there are no entities without corresponding entities in the new satellite, and the assignment relationship between the new satellite and the old satellite A can be determined; the similarity evaluation value between the new satellite and the old satellite A is the sum of the cosine similarity of the main entity pair and the neighbor pairs with which the corresponding relationship is determined;

[0118] S4. Compare the similarity evaluation values of the new satellite with all old satellites

[0119] From a semantic perspective, the new satellite has different orbit types and subsystems from the old satellite A knowledge graph, resulting in a low similarity score. The new satellite has a different battery type in the power subsystem from the old satellite B knowledge graph. All entities in the old satellite C knowledge graph have high similarity, and the old satellite C knowledge graph has the highest similarity assessment with the new satellite knowledge graph to be supplemented. Therefore, the old satellite C knowledge graph is selected as the satellite knowledge graph most similar to the new satellite knowledge graph.

[0120] S5. Define the part of the old satellite C knowledge graph that is larger than the new satellite as the missing part of the new satellite knowledge graph to be supplemented, and migrate it to obtain the following: Figure 10 The confidence of the complete new satellite knowledge graph and the migrated entities is shown in the figure. In the figure, the confidence of "satellite framework" is the cosine similarity of the embedding vector of the structural subsystem of the new satellite knowledge graph and the old satellite C knowledge graph; the confidence of "X model solar panel" and "unable to power" is the cosine similarity of the embedding vector of the power subsystem of the new satellite knowledge graph and the old satellite C knowledge graph; the confidence of "communication subsystem" is the cosine similarity of the embedding vector of the main node of the new satellite knowledge graph and the old satellite C knowledge graph.

Claims

1. A satellite knowledge graph rapid migration method based on entity similarity, characterized in that: The steps include: Step 1: Convert the new satellite knowledge graph to be supplemented and the complete old satellite knowledge graph into an attribute graph. Use a method similar to "one-hot" to encode different attributes of the entity to form an initial entity embedding vector containing semantic information. Step 2: Train the graph neural network and optimize the network parameters; Step 2.1: Based on the "view horizon" theory, different view horizons are assigned to different types of node entities in the attribute graph, so that entities connect to a wider range of entities as neighbors. The optimized adjacency matrix is obtained as the input of the graph neural network. Step 2.2: Randomly select point pairs in the attribute graph to construct samples. In the process of selecting point pairs, combine cross-graph point comparison and weighted node pairs to enable the graph neural network to learn all attribute graphs and focus on learning central range entities. Step 2.3: Under unsupervised learning conditions, use the constructed samples to train the graph neural network; Step 2.4: Use the mean square error (MSE) as the loss function to optimize the parameters of the graph neural network and obtain the trained graph neural network. Step 3: The trained graph neural network takes the attribute graph structure and the initial entity embedding vector as input to obtain the entity attribute embedding vector containing structural and semantic information; Step 4: By comparing the entity attribute embedding vectors in the new satellite knowledge graph to be supplemented and the complete old satellite knowledge graph, a similarity matrix is obtained. By matching the assignment matrix through the graph structure, the entity matching of the new satellite knowledge graph to be supplemented and the old satellite knowledge graph with the correct structure is obtained. Step 5: Through entity matching, entities that exist in the old satellite knowledge graph but do not exist in the new satellite knowledge graph to be supplemented are regarded as missing entities of the new satellite knowledge graph to be supplemented, and they are migrated to the new knowledge graph to be supplemented, and a complete new satellite knowledge graph is output.

2. The satellite knowledge graph rapid migration method based on entity similarity according to claim 1 is characterized in that: In step 1, the "one-hot" method is used to encode different attributes of the entity. The process is as follows: the embedding vectors corresponding to all attributes in the attribute graph are stacked horizontally, and the attribute values that the entity does not have are left blank. Numeric attributes directly use their values as embedding vectors, and character attributes use character-by-character hashing to generate embedding vectors. The generated embedding vectors are expressed as: H=concat(σ(H i )) In the formula, concat is the connection operation, σ refers to the normalization operation, H i Represents the embedding vector of a single attribute, H does not have the attribute i =0.

3. The satellite knowledge graph rapid migration method based on entity similarity according to claim 1 is characterized in that: The graph neural network in step 2 is a 4-layer or 6-layer graph convolutional neural network GCN, and the graph convolutional neural network GCN adopts a residual network structure. Its forward transmission process is expressed as: Where H (l) is the entity attribute vector of the lth layer, σ(·) represents the relu function, ω i Represents the entity attribute weight of the i-th layer, which is the characteristic of the residual network. is the normalized adjacency matrix, A (l) is a parameter.

4. The satellite knowledge graph rapid migration method based on entity similarity according to claim 1 is characterized in that: The optimizer used for training the graph neural network in step 2 is the Adam optimizer, and the initial learning rate is set to 0.

03.

5. The satellite knowledge graph rapid migration method based on entity similarity according to claim 1 is characterized in that: The specific process of step 2.1 is as follows: If the neighbor is more important than the entity itself, only this neighbor is connected. Otherwise, the neighbor range is recursively extended to the neighbors with lower importance than itself. All entities within the field of view are considered to be connected to the entity, that is, the corresponding value in the adjacency matrix is not 0. At the same time, different weights w are given according to the original distance D between the entities, expressed as: Where b is the view weight base, which is generally set to 2. The farther the distance, the lower the weight of the entity, and the closer the corresponding value in the adjacency matrix is to 0.

6. The satellite knowledge graph rapid migration method based on entity similarity according to claim 1 is characterized in that: The specific process of step 4 is as follows: Step 4.

1. Calculate the cosine similarity of each pair of embedding vectors containing structural and semantic information obtained in step 3. A similarity matrix is obtained for each pair of satellite knowledge graphs. The horizontal axis represents each entity in the new satellite knowledge graph, and the vertical axis represents each entity in the old satellite knowledge graph. The matrix value is the cosine similarity of the corresponding entity embedding vector. Step 4.2: Calculate the assignment matrix and the similarity evaluation values of the old satellite knowledge graph and the new satellite knowledge graph through the similarity matrix, and select the old knowledge graph with the highest similarity evaluation value and its corresponding relationship with the new knowledge graph.

7. The satellite knowledge graph rapid migration method based on entity similarity according to claim 6 is characterized in that: The process of calculating the similarity evaluation values of the assignment matrix and the old satellite knowledge graph and the new satellite knowledge graph through the similarity matrix in step 4.2 is as follows: Step 4.2.

1. Align the main entities of the old satellite knowledge graph with the new satellite knowledge graph; Step 4.2.2: Take a pair of corresponding entities a and b, and calculate the local similarity matrix and the assignment relationship between neighbors based on the neighbors of a and b. Let the neighbors of a be set A, and the neighbors of b be set B. For a pair of entities in A and B, if one of the entities has no uncorresponding neighbors, the value of its similarity matrix is the cosine similarity of this entity pair; if both entities have uncorresponding neighbors, the corresponding value of their similarity matrix is also called the matching degree of the entity pair. Recursively enter step 4.2.2, and after obtaining the similarity matrix, calculate the assignment matrix through the assignment matrix generation step of the bipartite graph. The matching degree of entity pair a and b is the sum of the matching degrees of all entity pairs indicated by the assignment matrix plus the cosine similarity between a and b. Step 4.2.3: When all entities in the new satellite knowledge graph have corresponding entities in the old knowledge graph, the assignment matrix is calculated and the similarity evaluation value is the sum of the cosine similarity of the main entity and the matching degree of its neighbors.

8. The satellite knowledge graph rapid migration method based on entity similarity according to claim 7 is characterized in that: The steps for generating the assignment matrix through the bipartite graph in step 4.2.2 are: Step 4.2.2.

1. Initialization The similarity matrix is denoted as w, i and j are used to record the new satellite knowledge graph entities and the old satellite knowledge graph entities respectively. i and j are also called top marks. The array u is initialized to the maximum weight of all i, expressed as: u[i] = max(w[i][j]), and the array v is initialized to 0; Step 4.2.2.2: Finding a Perfect Match Use the augmenting path method to find the maximum weight match. Select an unmatched new satellite knowledge graph entity i. If i has not been matched, find an augmenting path by adjusting the top indices u[i] and v so that the corresponding old satellite knowledge graph entity j can be matched. If j has been matched, find a better matching solution through the augmenting path. Step 4.2.2.3, adjust the top mark: If an augmenting path is found, adjust the values of u and v according to the matching situation on the augmenting path to find a better match. Specifically, for the unmatched new satellite knowledge graph entity i, increase the value of u[i], expressed as u[i]=u[i]+a, where a is a positive number, usually the smallest non-zero weight in the augmenting path; for the old satellite knowledge graph entity j, reduce the value of v[j], expressed as v[j]=v[j]-a; for the matched edge (i, j), increase the value of u[i] and reduce the value of v[j] so that the value of u[i]+v[j] can reflect the weight of the edge (i, j); Step 4.2.2.4: Repeat steps 4.2.2.2 to 4.2.2.3, repeatedly searching for augmenting paths until no augmenting paths can be found. Step 4.2.2.5: Output results The obtained u and v arrays contain the top label information of the optimal match, and the specific edge set with the maximum weight match is determined through the top label information.

Images