Satellite knowledge graph fast migration method based on entity similarity

Abstract

The application discloses a kind of based on entity similarity's satellite knowledge graph fast migration method, comprising the following steps:1, according to the attribute graph and initial entity embedding vector obtained from old satellite knowledge graph and new satellite knowledge graph to be supplemented;2, train graph neural network and optimize network parameters;3, the graph neural network of training is inputed with attribute graph structure and initial entity embedding vector, and the entity attribute embedding vector containing complete semantic information is obtained;4, the similarity matrix is obtained by entity attribute body embedding vector, and the entity matching of correct structure of old satellite knowledge graph and new satellite knowledge graph to be supplemented is obtained by graph structure matching assignment matrix;5, by entity matching, the entity existing in old satellite knowledge graph and not existing in new satellite knowledge graph to be supplemented is migrated, and the complete new satellite knowledge graph is output, the efficiency of constructing satellite knowledge graph and the quality of knowledge graph are improved, a large amount of time and manpower are saved.

Description

Technical Field

[0001] This invention relates to knowledge graphs, specifically a rapid transfer method for satellite knowledge graphs based on entity similarity. Background Technology

[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 lot of manual annotation and automated extraction techniques 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 interrelationships 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 for satellites, playing a crucial role in improving satellite management and monitoring efficiency. However, the large number and variety of satellites, especially those of the same family and model used for navigation and communication functions, often result in high similarity. This makes traditional methods of constructing knowledge graphs for each satellite labor-intensive, demanding, and repetitive, wasting significant human and time. Therefore, new technologies are urgently needed to address these challenges. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a rapid migration method for satellite knowledge graphs based on entity similarity, which improves the efficiency of constructing satellite knowledge graphs, saves a significant amount of time and human resources, and enhances the accuracy of the new knowledge graph.

[0005] To achieve the above objectives, the present invention employs the following technical solution: A fast transfer method for satellite knowledge graphs based on entity similarity includes the following steps: Step 1: Convert the new satellite knowledge graph to be supplemented and the complete old satellite knowledge graph into attribute graphs. Use the "one-hot" method to encode different attributes of entities 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 field" theory, different view fields are assigned to different types of node entities in the attribute graph, so that the entities can connect to a wider range of entities as neighbors, and 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. During the selection of point pairs, combine cross-graph cross-point comparison and weighted node pairs to enable the graph neural network to learn all attribute graphs and focus on learning entities within the central range. Step 2.3: Under unsupervised learning conditions, train the graph neural network using the constructed samples; Step 2.4: Use the mean squared 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 entity attribute embedding vectors 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. Then, through the graph structure matching assignment matrix, 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 not in the new satellite knowledge graph to be supplemented are identified as missing entities in the new satellite knowledge graph to be supplemented. These entities are then migrated to the new knowledge graph to be supplemented, and a complete new satellite knowledge graph is output.

[0006] Furthermore, in step 1, the "one-hot" method is used to encode different attributes of the entity as follows: the embedding vectors corresponding to all attributes in the attribute graph are horizontally superimposed, and attribute values ​​that the entity does not have are left blank; for numeric attributes, their values ​​are directly used as the embedding vectors, and for character attributes, character-by-character hashing is used to generate the embedding vectors. The generated embedding vectors are represented as follows: In the formula, It's a connection operation. Refers to the normalization operation. This represents the embedding vector of a single attribute, and attributes that are not present. .

[0007] Furthermore, the graph neural network in step 2 is a 4-layer or 6-layer graph convolutional neural network (GCN), and the GCN adopts a residual network structure. Its forward propagation process is represented as follows: In the formula, For the first l Layer entity attribute vector, Represents the ReLU function. Indicates the first i The entity attribute weights of a layer are a characteristic of residual networks. The normalized adjacency matrix, It is a parameter.

[0008] Furthermore, in step 2, the optimizer used to train the graph neural network is the Adam optimizer, and the initial learning rate is set to 0.03.

[0009] Furthermore, the specific process of step 2.1 is as follows: If a neighbor is more important than the entity itself, then only that neighbor is connected; otherwise, the neighbor range is recursively extended to neighbors with lower importance. All entities within the field of view are considered connected to this entity, meaning the corresponding value in the adjacency matrix is ​​not 0. Simultaneously, based on the original distances between entities... Assign different weights , is represented as: In the formula, This is the base value for the view weights, typically set to 2. Entities farther away have lower weights, and the corresponding values ​​in the adjacency matrix are closer to 0.

[0010] Furthermore, 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. Here, a similarity matrix is ​​obtained between each pair of satellite knowledge graphs. The horizontal axis represents each entity of the new satellite knowledge graph, the vertical axis represents each entity of the old satellite knowledge graph, and the matrix value is the cosine similarity of the embedding vectors of the corresponding entities. Step 4.2: Calculate the assignment matrix and the similarity evaluation value between the old satellite knowledge graph and the new satellite knowledge graph using the similarity matrix. Select the old knowledge graph with the highest similarity evaluation value and its correspondence with the new knowledge graph.

[0011] Furthermore, the process in step 4.2 of calculating the assignment matrix and the similarity evaluation values ​​of the old satellite knowledge graph and the new satellite knowledge graph using the similarity matrix is ​​as follows: Step 4.2.1: Match 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. Calculate the local similarity matrix and the assignment relationship between 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 entity does not have any non-corresponding neighbors, the value of its similarity matrix is ​​the cosine similarity of this entity pair. If both entities have non-corresponding neighbors, the value of its similarity matrix is ​​called the matching degree of the entity pair. After obtaining the similarity matrix by recursively entering step 4.2.2, calculate the assignment matrix through the bipartite graph assignment matrix generation step. The matching degree of entity pair a and b is the sum of the matching degrees of all entity pairs pointed to 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 ​​completed. The similarity evaluation value is the sum of the cosine similarity of the main entity and the matching degree of its neighbors.

[0012] Further, the step of generating the assignment matrix through the bipartite graph in step 4.2.2 is as follows: Step 4.2.2.1 Initialization Let the similarity matrix be denoted as , and They are used to record entities in the new satellite knowledge graph and entities in the old satellite knowledge graph, respectively. and Also known as the top index, the array u is initialized with the maximum weight of all i, expressed as: Initialize array v to 0; Step 4.2.2.2: Finding the perfect match The augmented path method is used to find the maximum weight match. An unmatched new satellite knowledge graph entity i is selected. If i has not yet been matched, an augmented path is found 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 already been matched, a better matching scheme is found through the augmented path. Step 4.2.2.3: Adjust the top label: If an augmenting path is found, the values ​​of u and v are adjusted according to the matching situation on the augmenting path in order to find a better match. Specifically, for unmatched new satellite knowledge graph entity i, the value of u[i] is increased, expressed as u[i] = u[i] + a, where a is a positive number and is the smallest non-zero weight in the augmenting path; for old satellite knowledge graph entity j, the value of v[j] is decreased, expressed as v[j] = v[j] - a; for matched edge (i,j), the value of u[i] is increased and the value of v[j] is decreased, so that the value of u[i] + v[j] can reflect the weight of edge (i,j); Step 4.2.2.4: Repeat steps 4.2.2.2 to 4.2.2.3 to search for augmenting paths until no augmenting paths can be found. Step 4.2.2.5, Output Results The resulting u and v arrays contain the top label information of the optimal match. Based on the top label information, the specific set of edges for the maximum weight match is determined.

[0013] Compared with the prior art, the present invention has the following technical effects: On the one hand, the graph neural network of this invention can be implemented using Python scripts, making it easy to deploy in the field and supporting diverse input formats. It uses a "one-hot" approach to handle entity attribute embedding, efficiently utilizing the semantic similarity of entities and improving the accuracy of knowledge graph transfer. On the other hand, to improve the performance of the graph neural network, the following measures were taken during training: First, the intermediate files of the satellite knowledge graph and the graph neural network are stored independently, facilitating high reuse of training results during large-scale / multi-batch use and accelerating network performance. Second, based on the "vision" theory, the adjacency matrix is ​​optimized, assigning different "visions" to different types of node entities, allowing entities to connect to a wider range of entities as neighbors. This overcomes the limitations of slow information transmission and information dilution in traditional adjacency matrices, improving the accuracy and confidence of the new satellite knowledge graph. Third, to address the need for randomly selecting entities from the knowledge graph, the selection weights are modified according to importance, reflected in: cross-graph cross-point comparison to ensure the utilization of... The graph neural network is trained using attribute graphs to prevent overfitting to individual attribute graphs. Furthermore, to address the varying physical importance of different entities, a weighted node selection scheme is employed to improve the accuracy of knowledge graph transfer and the network's training efficiency. This scheme prioritizes the training effect of entities within the central region, preventing dilution by numerous low-weight entities at the edges, thus enhancing training performance and optimizing training speed and efficiency. On another front, the initial entity embedding vectors are optimized using the trained graph neural network to incorporate more structural and semantic information. A similarity matrix is ​​then generated by comparing the initial entity embedding vectors in the new satellite knowledge graph to be supplemented with those in the complete old satellite knowledge graph. This matrix is ​​used to generate an assignment matrix, enabling rapid transfer from the old satellite knowledge graph to the new one, improving the efficiency of satellite knowledge graph construction, saving significant time and manpower, and resulting in a more complete, accurate, and confident structure of the generated new satellite knowledge graph. Attached Figure Description

[0014] Figure 1 This is a flowchart of the satellite knowledge graph migration process of the present invention; Figure 2 This is a schematic diagram of the entity initial embedding vector process of the present invention; Figure 3 This invention provides a satellite knowledge graph transfer process and a graph convolutional neural network training process. Figure 4(a) is a schematic diagram of the structure of the graph neural network M-GCN-4 of the present invention; Figure 4(b) is a schematic diagram of the structure of the graph neural network M-GCN-6 of the present invention; Figure 5 This is the view of the subsystem type entity and parameter type entity of the present invention; Figure 6(a) shows the knowledge graph of the new satellites to be supplemented; Figure 6(b) shows the knowledge graph of the old satellite A; Figure 6(c) shows the knowledge graph of the old satellite B; Figure 6(d) shows the knowledge graph of the old satellite C; Figure 7 A diagram of the initial embedding vector of an entity. Figure 1 ; Figure 8 A diagram of the initial embedding vector of an entity. Figure 2 ; Figure 9 This is the similarity matrix between the new satellite and the old satellite A; Figure 10 This is a schematic diagram of the complete new satellite knowledge graph after migration. Detailed Implementation

[0015] The specific content of the present invention will be further explained in detail below with reference to the embodiments.

[0016] The new satellite knowledge graph mentioned in this embodiment contains only 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 and for easy maintenance.

[0017] like Figure 1 and 3 As shown, a fast transfer method for satellite knowledge graphs based on entity similarity includes the following steps: Step 1: Convert the new satellite knowledge graph to be supplemented and the complete old satellite knowledge graph into attribute graphs. Use a method similar to "one-hot" to encode different attributes of entities to form an initial entity embedding vector containing semantic information. like Figure 2 As shown, a "one-hot" approach is used to handle the attribute embedding problem of entities. The process of encoding different entity attributes is as follows: the embedding vectors corresponding to all attributes are horizontally superimposed, and attribute values ​​that the entity does not have are left blank; numeric attributes directly use their numerical values ​​as embedding vectors, and character attributes use a character-by-character hashing scheme to generate embedding vectors, thereby efficiently utilizing the semantic similarity of entities and improving the accuracy. The embedding vector is represented as follows: In the formula, It's a connection operation. 'Refers to the normalization operation,' This represents the embedding vector of a single attribute, and attributes that are not present. , Figure 2 In the middle, the name is embedded as Type embedding is The physical weight is embedded as ; Step 2: Train the graph neural network and optimize the network parameters; The graph neural network is a 4-layer or 6-layer graph convolutional neural network (GCN) as shown in Figures 4(a) and 4(b). The GCN adopts a residual network structure. The 4-layer or 6-layer GCN is denoted as M-GCN-4 and M-GCN-6, respectively. It breaks the constraints of the traditional graph convolutional network's shallow layers and small information transmission range. Among them, M-GCN-6 has the best accuracy, while M-GCN-4 has higher efficiency while maintaining high accuracy. The forward propagation process of M-GCN-4 and M-GCN-6 is represented as follows: In the formula, For the first l Layer entity attribute vector, Represents the ReLU function. Indicates the first i The entity attribute weights of a layer are a characteristic of residual networks. The normalized adjacency matrix, It is a parameter; Based on PyTorch, a graph neural network is trained for the Adam optimizer using an optimizer, with the initial learning rate set to 0.03. The training process includes the following steps: Step 2.1: Based on the "view field" theory, different view fields are assigned to different types of node entities in the attribute graph, so that the entities can connect to a wider range of entities as neighbors, and thus obtain the optimized adjacency matrix. To address the issues of slow message passing efficiency and the limitation of each layer only enabling nodes to learn information from their neighborhood in graph neural networks, the "vision" theory is employed. Different types of nodes are assigned different "visions," meaning that entities connect to a wider range of other entities as neighbors. This overcomes the limitations of traditional adjacency matrices, such as slow information transmission speed and information dilution. Figure 5 As shown, different types of entities have different levels of importance. The process of connecting entities within different ranges to determine the entity's view range, i.e., the new neighbor range, is as follows: If a neighbor is more important than itself, only that neighbor is connected; otherwise, the neighbor range is recursively extended to neighbors with lower importance. All entities within the viewport are considered connected to this entity, meaning the corresponding value in the adjacency matrix is ​​not 0. Furthermore, the connection is based on the distance between entities. Different weights are assigned to distance. , represented as: In the formula, The value is the base of the view weight, which is generally 2. The farther away an entity is, the lower its weight, and the closer the corresponding value in the adjacency matrix is ​​to 0. The optimized adjacency matrix can directly replace the usual adjacency matrix for training graph neural networks. However, normalization is necessary to ensure the training effect. 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. During the selection of point pairs, combine cross-graph cross-point comparison and weighted node pairs to enable the graph neural network to learn all attribute graphs and focus on learning entities within the central range. To improve the automation level of the module, the training module adopts an unsupervised learning method. Point pairs are randomly selected in the attribute graph to construct samples. Each training session selects about 2 / 3 of the total number of nodes, which can ensure the sample coverage area and training effect. Unlike knowledge graphs in fields such as Wikipedia (adjacent or similar entities have higher similarity), for satellite knowledge graphs, the positive or negative of a sample is mainly determined by the similarity of its attributes and the similarity of its neighbors. Therefore, for the combined work of knowledge graphs from different fields, it is necessary to have a deep understanding of the relative relationships of the entities and then specify the learning strategy. On the one hand, this embodiment adopts a full-range cross-point comparison scheme, thereby using all attribute graphs to train the graph neural network and preventing the network from overfitting to a single attribute graph; On the other hand, to address the varying degrees of physical importance among different entities, a weighted node pair selection scheme is adopted to improve overall accuracy and training efficiency. In the satellite knowledge graph, each knowledge graph has a main entity, with the remaining entities arranged radially. Entities closer to the main entity have higher importance, while the importance of edge entities decreases sequentially, giving them higher weight in similarity evaluation. Therefore, when training the graph neural network, the training effect of entities within the central region should be emphasized. Consequently, when randomly selecting point pairs, this part of the entities should also be prioritized to avoid diluting the training effect with a large number of low-importance entities at the edges, thus saving training time. Specifically, the selection is based on the shortest path length between the entity and the main entity. The ratio of the probability of selecting this entity to the probability of selecting the main entity. Represented as: In the formula, This is the base value for the point selection weight, and its value is 8. Step 2.3: Under unsupervised learning conditions, train the graph neural network using the constructed samples; Step 2.4: Use the mean squared 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 and outputs entity attribute embedding vectors that contain structural and semantic information to the greatest extent. 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. Then, through graph structure matching and assignment matrices, 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: Step 4.1: Calculate the cosine similarity of each pair of embedding vectors containing structural and semantic information obtained in Step 3. Here, a similarity matrix is ​​obtained between each pair of satellite knowledge graphs. The horizontal axis represents each entity of the new satellite knowledge graph, the vertical axis represents each entity of the old satellite knowledge graph, and the matrix value is the cosine similarity of the embedding vectors of the corresponding entities. Step 4.2: Calculate the assignment matrix and the similarity evaluation value between the old satellite knowledge graph and the new satellite knowledge graph using the similarity matrix. Select the old knowledge graph with the highest similarity evaluation value and its correspondence with the new knowledge graph. The process of calculating the assignment matrix and the similarity evaluation values ​​between the old and new satellite knowledge graphs using the similarity matrix is ​​as follows: Step 4.2.1: Match 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 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 value of its similarity matrix is ​​called the matching degree of the entity pair, which is obtained by recursively entering step 4.2.2. After obtaining the similarity matrix, generate the assignment matrix 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 pointed to by the assignment matrix plus the cosine similarity between a and b. The steps for generating the assignment matrix from a bipartite graph are as follows: Step 4.2.2.1 Initialization Let the similarity matrix be denoted as , and They are used to record entities in the new satellite knowledge graph and entities in the old satellite knowledge graph, respectively. and Also known as the top index, it initializes the array. For all The maximum weight is expressed as: Initialize array v to 0; Step 4.2.2.2: Finding the perfect match The augmented path method is used to find the maximum weight match. An unmatched new satellite knowledge graph entity i is selected. If i has not yet been matched, an augmented path is found 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 already been matched, a better matching scheme is found through the augmented path. Step 4.2.2.3: Adjust the top label: If an augmenting path is found, the values ​​of u and v are adjusted according to the matching situation on the augmenting path in order to find a better match. Specifically, for unmatched new satellite knowledge graph entity i, the value of u[i] is increased, expressed as u[i] = u[i] + a, where a is a positive number and is the smallest non-zero weight in the augmenting path; for old satellite knowledge graph entity j, the value of v[j] is decreased, expressed as v[j] = v[j] - a; for matched edge (i,j), the value of u[i] is increased and the value of v[j] is decreased, so that the value of u[i] + v[j] can reflect the weight of edge (i,j); Step 4.2.2.4: Repeat steps 4.2.2.2 to 4.2.2.3 to search 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. Based on the top label information, the specific set of edges for the maximum weight match is determined. 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. Step 5: Through entity matching, entities that exist in the old satellite knowledge graph but not in the new satellite knowledge graph to be supplemented are identified as missing entities in the new satellite knowledge graph to be supplemented. These entities are then migrated to the new knowledge graph to be supplemented, and a complete new satellite knowledge graph is output. The specific process is as follows: By comparing the similarity evaluation values ​​of all old satellite knowledge graphs and the new satellite knowledge graph 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. During the migration operation, the existing assignment relationship between the old and new satellite knowledge graphs is used to determine the missing parts of the new satellite knowledge graph compared to the old satellite knowledge graph. These missing parts are then migrated in a grafting manner to obtain a complete new satellite knowledge graph. During the migration process, the confidence level of the migrated entities is obtained through the similarity of the entities associated with the old satellite knowledge graph, which is then used for manual review and improvement. To verify the effectiveness of the satellite knowledge graph fast transfer method based on entity similarity proposed in this embodiment, the verification was performed using the new satellite knowledge graph shown in Figure 6(a) and the knowledge graphs of old satellite A, old satellite B, and old satellite C shown in Figures 6(b) to 6(d), including the following steps: S1. Based on the knowledge graphs in Figures 6(a) to 6(d), generate attribute graphs and initial entity embedding vectors; In Figures 6(a) to 6(d), each circle in the knowledge graph represents an entity with different attributes (not shown in the figures). Assuming all entities include four attributes: name, type, quality, and manufacturer, the initial entity embedding vectors of the "new satellite" entities with these four attributes in the supplementary new satellite knowledge graph are as follows: Figure 7 As shown, the entity "High Orbit" in the new satellite knowledge graph to be supplemented does not have mass and manufacturer attributes, and its initial entity embedding vector is as follows: Figure 8 As shown; Figure 7 and Figure 8 In this context, each H_ sub-part represents the semantic information contained in the sub-embedding vector. For character data, the vector is scaled and normalized, and represented as follows: For numerical data, a single vector dimension is used, and the values ​​are normalized using the arctangent function, represented as: ; S2. The attribute graph (also known as the graph structure of a knowledge graph or the adjacency matrix optimized by the "view" theory) and the initial entity embedding vector are used as inputs to the trained graph neural network. The graph neural network outputs entity attribute embedding vectors containing complete semantic information. S3. By embedding vectors into entity attribute bodies, calculate cosine similarity and assignment matrix. Using the knowledge graph of the new satellite to be supplemented and the knowledge graph of the old satellite A as examples, the steps of assignment matrix calculation are shown. The numbers such as similarity and matching degree involved are for display purposes only. 1) Match the main entities of the "new satellite" with those of the "old satellite A"; 2) Take a pair of corresponding entities, namely "new satellite" and "old satellite A". Among the neighbors of "new satellite", there is the entity "power subsystem" that has a neighbor. Among the neighbors of "old satellite A", there are the entities "attitude and orbit control subsystem" and "power subsystem" that have a neighbor. For these parts in the similarity matrix, the matching degree is calculated recursively. For other parts, the cosine similarity of the corresponding entities can be calculated directly. 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. Their neighbors can be directly compared. The entity embedding cosine similarity value between the "Y-type battery" in the new satellite and the "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. 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. Their neighbors can be directly compared. The entity embedding cosine similarity value of "Y-type battery" in the new satellite and "Y-type battery" in the old satellite A is 0.9. Then the value of the similarity matrix between the two in step 2) is 0.9 + 0.9 = 1.8.

[0018] For entity pairs where one of the other two has no other neighbors, the cosine similarity can be used to represent the degree of matching. The resulting similarity matrix between "new satellite" and "old satellite A" is as follows: Figure 9 As shown, using this similarity matrix and the assignment matrix calculation method for bipartite graphs, the correspondence between entities appearing in this similarity matrix can be calculated. Steps 1) to 4) are repeated for these entities with determined correspondence until there are no entities without corresponding entities in the new satellite, thus completing the determination of the assignment relationship between the new satellite and the old satellite A. 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 matching degree of its neighbor pairs with determined correspondence. S4. Compare the similarity assessment values ​​of the new satellite with all the old satellites. From a semantic perspective, the new satellite's orbital type and subsystems differ from those of the old satellite A knowledge graph, resulting in a lower similarity score. Furthermore, the battery type differs from the old satellite B knowledge graph's power subsystem. The old satellite C knowledge graph contains entities with high similarity to all entities in the old satellite C knowledge graph. Therefore, the old satellite C knowledge graph has the highest similarity assessment with the new satellite knowledge graph to be supplemented. Consequently, the old satellite C knowledge graph is selected as the satellite knowledge graph most similar to the new satellite knowledge graph. S5. Define the portion of the old satellite C knowledge graph that is more abundant than the new satellite knowledge graph as the missing portion to be supplemented in the new satellite knowledge graph, perform migration, and obtain, as shown below. Figure 10The diagram shows the complete new satellite knowledge graph and the confidence levels of the migrated entities. In the diagram, the confidence level of "satellite frame" is the cosine similarity of the embedding vectors of the structural subsystems of the new satellite knowledge graph and the old satellite C knowledge graph; the confidence levels of "X-type solar panel" and "unable to supply power" are the cosine similarity of the embedding vectors of the power supply subsystems of the new satellite knowledge graph and the old satellite C knowledge graph; and the confidence level of "communication subsystem" is the cosine similarity of the embedding vectors of the master nodes of the new satellite knowledge graph and the old satellite C knowledge graph.

Claims

1. A fast transfer method for satellite knowledge graphs based on entity similarity, characterized in that, Includes the following steps: Step 1: Convert the new satellite knowledge graph to be supplemented and the complete old satellite knowledge graph into attribute graphs. Use the "one-hot" method to encode different attributes of entities to form an initial entity embedding vector containing semantic information. The "one-hot" method is used to encode different attributes of an entity as follows: The embedding vectors corresponding to all attributes in the attribute graph are horizontally superimposed, leaving attribute values ​​that the entity does not possess blank; numeric attributes directly use their numerical values ​​as embedding vectors, while character attributes are generated using character-by-character hashing. The generated embedding vectors are represented as follows: In the formula, It's a connection operation. 'Refers to the normalization operation,' An embedding vector representing a single attribute; Step 2: Train the graph neural network and optimize the network parameters; Step 2.1: Based on the "viewfield" theory, different viewfields are assigned to different types of node entities in the attribute graph, allowing entities to connect to a wider range of entities as neighbors, resulting in an optimized adjacency matrix, which serves as the input to the graph neural network. The specific process is as follows: If a neighbor is more important than the entity itself, then only that neighbor is connected; otherwise, the neighbor range is recursively extended to neighbors with lower importance. All entities within the field of view are considered connected to this entity, meaning the corresponding value in the adjacency matrix is ​​not 0. Simultaneously, based on the original distances between entities... Assign different weights , is represented as: In the formula, This is the base value for the view weights, typically set to 2. The farther away an entity is, the lower its weight, and the closer the corresponding value in the adjacency matrix is ​​to 0. Step 2.2: Randomly select point pairs in the attribute graph to construct samples. During the selection of point pairs, combine cross-graph cross-point comparison and weighted node pairs to enable the graph neural network to learn all attribute graphs and focus on learning entities within the central range. Step 2.3: Under unsupervised learning conditions, train the graph neural network using the constructed samples; Step 2.4: Use the mean squared 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 entity attribute embedding vectors 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. An assignment matrix is ​​generated by matching the similarity matrix and graph structure to obtain the entity matching of the new satellite knowledge graph to be supplemented and the old satellite knowledge graph with the correct structure. Step 5: Through entity matching, entities that exist in the old satellite knowledge graph but not in the new satellite knowledge graph to be supplemented are identified as missing entities in the new satellite knowledge graph to be supplemented. These entities are then migrated to the new knowledge graph to be supplemented, and a complete new satellite knowledge graph is output.

2. The rapid transfer method for satellite knowledge graphs based on entity similarity according to claim 1, characterized in that, The graph neural network in step 2 is a 4-layer or 6-layer graph convolutional neural network (GCN), and the GCN adopts a residual network structure. Its forward propagation process is represented as follows: In the formula, For the first l Layer entity attribute vector, Represents the ReLU function. Indicates the first i The entity attribute weights of a layer are a characteristic of residual networks. The normalized adjacency matrix, It is a parameter.

3. The rapid transfer method for satellite knowledge graphs based on entity similarity according to claim 1, characterized in that, In step 2, the optimizer used to train the graph neural network is the Adam optimizer, and the initial learning rate is set to 0.

03.

4. The rapid transfer method for satellite knowledge graphs based on entity similarity according to claim 1, 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. Here, a similarity matrix is ​​obtained between each pair of satellite knowledge graphs. The horizontal axis represents each entity of the new satellite knowledge graph, the vertical axis represents each entity of the old satellite knowledge graph, and the matrix value is the cosine similarity of the embedding vectors of the corresponding entities. Step 4.2: Calculate the assignment matrix and the similarity evaluation value between the old satellite knowledge graph and the new satellite knowledge graph using the similarity matrix. Select the old knowledge graph with the highest similarity evaluation value and its correspondence with the new knowledge graph.

5. The rapid transfer method for satellite knowledge graphs based on entity similarity according to claim 4, characterized in that, The process in step 4.2 of calculating the assignment matrix and the similarity evaluation value between the old satellite knowledge graph and the new satellite knowledge graph using the similarity matrix is ​​as follows: Step 4.2.1: Match 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. Calculate the local similarity matrix and the assignment relationship between 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 entity does not have any non-corresponding neighbors, the value of its similarity matrix is ​​the cosine similarity of this entity pair. If both entities have non-corresponding neighbors, the value of its similarity matrix is ​​called the matching degree of the entity pair. After obtaining the similarity matrix by recursively entering step 4.2.2, calculate the assignment matrix through the bipartite graph assignment matrix generation step. The matching degree of entity pair a and b is the sum of the matching degrees of all entity pairs pointed to 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 ​​completed. The similarity evaluation value is the sum of the cosine similarity of the main entity and the matching degree of its neighbors.

6. The method for rapid transfer of satellite knowledge graphs based on entity similarity according to claim 5, characterized in that, The step 4.2.2 in which the assignment matrix is ​​generated using the bipartite graph is as follows: Step 4.2.2.1 Initialization Let the similarity matrix be denoted as , and They are used to record entities in the new satellite knowledge graph and entities in the old satellite knowledge graph, respectively. and Also known as the top index, it initializes the array. For all The maximum weight is expressed as: Initialize array v to 0; Step 4.2.2.2: Finding the perfect match The augmented path method is used to find the maximum weight match. An unmatched new satellite knowledge graph entity i is selected. If i has not yet been matched, an augmented path is found 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 already been matched, a better matching scheme is found through the augmented path. Step 4.2.2.3: Adjust the top label: If an augmenting path is found, the values ​​of u and v are adjusted according to the matching situation on the augmenting path in order to find a better match. Specifically, for unmatched new satellite knowledge graph entity i, the value of u[i] is increased, expressed as u[i] = u[i] + a, where a is a positive number and is the smallest non-zero weight in the augmenting path; for old satellite knowledge graph entity j, the value of v[j] is decreased, expressed as v[j] = v[j] - a; for matched edge (i,j), the value of u[i] is increased and the value of v[j] is decreased, so that the value of u[i] + v[j] can reflect the weight of edge (i,j); Step 4.2.2.4: Repeat steps 4.2.2.2 to 4.2.2.3 to search for augmenting paths until no augmenting paths can be found. Step 4.2.2.5, Output Results The resulting u and v arrays contain the top label information of the optimal match. Based on the top label information, the specific set of edges for the maximum weight match is determined.

Images