A method, apparatus, electronic device, and medium for predicting proximity events.

Abstract

This application discloses a method, apparatus, electronic device, and medium for predicting proximity events. The method includes: constructing a training dataset of satellite proximity events and constructing a satellite knowledge graph; predicting whether a proximity event exists between any two satellite nodes based on the satellite knowledge graph, obtaining a proximity prediction result; inputting the training dataset into a pre-constructed graph neural network for training, obtaining a proximity training result; constructing a loss function using the proximity prediction result and the proximity training result; performing iterative training with the goal of minimizing the loss function, obtaining a target graph neural network; and inputting the feature data of the satellite to be predicted into the target graph neural network to obtain a target proximity prediction result. Thus, by minimizing the loss function constructed from the proximity prediction result and the proximity training result, the training of the graph neural network is completed, resulting in an efficient and accurate target graph neural network that can be directly applied to proximity event prediction.

Description

Technical Field

[0001] This application relates to the field of aerospace engineering technology, and in particular to a method, apparatus, electronic device and medium for predicting proximity events. Background Technology

[0002] With the rapid development of aerospace technology, the number of satellites has exploded, making the space environment in low Earth orbit increasingly complex and crowded. Currently, a large number of satellites and space debris operate within a limited orbital space, resulting in a significant increase in satellite orbital density and increasingly complex relative motion relationships between satellites and between satellites and space debris.

[0003] Consequently, the frequency of satellite proximity events has increased significantly. A proximity event refers to the phenomenon where two or more space objects, during their orbital operation, overlap or approach each other, causing their relative distance to shrink to within the potential collision risk threshold. If timely warnings and avoidance measures are not taken, proximity events may trigger catastrophic on-orbit collisions, not only damaging the relevant satellites but also generating a large amount of new space debris, exacerbating the deterioration of the orbital environment, and threatening the sustainable development of the orbital ecosystem.

[0004] Therefore, how to efficiently and accurately predict satellite proximity events is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, one aspect of this application provides a method for predicting proximity events, the method comprising:

[0006] Construct a training dataset on satellite proximity events and build a satellite knowledge graph;

[0007] Based on the satellite knowledge graph, predict whether there is a proximity event between any two satellite nodes to obtain the proximity prediction result.

[0008] The training dataset is input into a pre-built graph neural network for training to obtain proximity training results;

[0009] A loss function is constructed using the proximity prediction results and the proximity training results;

[0010] With the goal of minimizing the loss function, the step of inputting the training dataset into the pre-built graph neural network to obtain the proximity training result is returned for iterative training to obtain the target graph neural network;

[0011] The feature data of the satellite to be predicted is input into the target graph neural network to obtain the target proximity prediction result of the satellite to be predicted.

[0012] Optionally, the step of predicting whether a proximity event exists between any two satellite nodes based on the satellite knowledge graph, and obtaining a proximity prediction result, includes:

[0013] From the satellite knowledge graph, the node features of each entity node are extracted; the entity nodes include the satellite nodes and orbit nodes.

[0014] The node features are mapped to a unified vector space to obtain feature vectors;

[0015] The feature vector is normalized to obtain the target feature vector;

[0016] The result of element-wise multiplication of any two target feature vectors is input into the multilayer perceptron to obtain the target output result;

[0017] Based on the target output, determine whether there is a proximity event between the two satellite nodes, and obtain the proximity prediction result.

[0018] Optionally, extracting node features of each entity node from the satellite knowledge graph includes:

[0019] A set of vertices in the hypergraph is constructed based on the entity nodes, and a set of hyperedges in the hypergraph is constructed based on whether there is an edge between any two entity nodes.

[0020] A first adjacency matrix is ​​constructed based on the hyperedges between any two satellite nodes, and a second adjacency matrix is ​​constructed based on the hyperedges between the satellite node and the orbit node.

[0021] By concatenating the first adjacency matrix and the second adjacency matrix, the target adjacency matrix is ​​obtained;

[0022] Based on the target adjacency matrix, the features of each entity node are aggregated using a specified graph neural network to obtain the node features.

[0023] Optionally, the loss function is a binary classification cross-entropy loss function.

[0024] Optionally, the training dataset further includes a test dataset, and the method includes:

[0025] The test dataset is input into the target graph neural network to obtain the target proximity prediction result;

[0026] Based on the target proximity prediction results, a quality score is determined to evaluate the confidence of the target graph neural network proximity event prediction.

[0027] When the quality score is not greater than the threshold, return to the step of constructing the training dataset about satellite proximity events and perform subsequent steps.

[0028] Optionally, construct a satellite knowledge graph, including:

[0029] Obtain the target dataset regarding satellite proximity events;

[0030] The satellites and orbits in the target dataset are used as entity nodes in the satellite knowledge graph;

[0031] Based on proximity events between satellite nodes and the operational relationships between the satellite nodes and orbital nodes, edges between the entity nodes are generated;

[0032] Satellite parameters are injected into the corresponding satellite nodes, and orbital parameters are injected into the corresponding orbital nodes to construct the satellite knowledge graph.

[0033] Optionally, the satellite parameters include at least the satellite number, payload, and physical parameters, and the orbital parameters include at least the orbital semi-major axis, orbital inclination, and eccentricity.

[0034] Another aspect of this application provides a proximity event prediction device, the device comprising:

[0035] The target building module is used to build a training dataset on satellite proximity events and construct a satellite knowledge graph;

[0036] The prediction module is used to predict whether there is a proximity event between any two satellite nodes based on the satellite knowledge graph, and obtain the proximity prediction result.

[0037] The network training module is used to input the training dataset into a pre-built graph neural network for training to obtain proximity training results; construct a loss function based on the proximity prediction results and the proximity training results; and perform iterative training by returning to the step of inputting the training dataset into the pre-built graph neural network to obtain proximity training results with the goal of minimizing the loss function, to obtain the target graph neural network.

[0038] The input module is used to input the feature data of the satellite to be predicted into the target graph neural network to obtain the target proximity prediction result of the satellite to be predicted.

[0039] Another aspect of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor executes the program to implement the steps of the proximity event prediction method.

[0040] Another aspect of this application provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the proximity event prediction method.

[0041] The method, apparatus, electronic device, and medium for predicting proximity events provided in this application have the following beneficial effects: By transforming satellites and orbits into a graph structure of knowledge graphs, the existence of proximity events between satellite nodes can be quickly predicted based on the characteristics of each node in the knowledge graph. Simultaneously, by training a graph neural network using a training dataset, the graph neural network can handle complex nonlinear relationships to learn about proximity events between satellites. Furthermore, by minimizing the loss function constructed from the proximity prediction results and proximity training results, the graph neural network is trained, resulting in an efficient and accurate target graph neural network that can be directly applied to proximity event prediction, thus avoiding catastrophic on-orbit collisions caused by proximity events and achieving sustainable development of the orbital environment. Attached Figure Description

[0042] Figure 1 A flowchart illustrating a proximity event prediction method provided in an embodiment of this application;

[0043] Figure 2 A schematic diagram illustrating the principle of a proximity event prediction method provided in an embodiment of this application;

[0044] Figure 3 This is a schematic diagram illustrating the principle of node feature extraction provided in an embodiment of this application.

[0045] Figure 4 A schematic diagram of the structure of a proximity event prediction device provided in an embodiment of this application;

[0046] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0047] The attached diagram is labeled as follows: 40 is the target construction module, 41 is the prediction module, 42 is the network training module, 43 is the input module, 50 is the memory, 51 is the processor, 52 is the display screen, 53 is the input / output interface, 54 is the communication interface, 55 is the power supply, 56 is the communication bus, 501 is the computer program, 502 is the operating system, and 503 is the data. Detailed Implementation

[0048] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0049] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0050] Figure 1 This is a flowchart illustrating a proximity event prediction method provided in an embodiment of this application, as shown below. Figure 1 As shown, the method includes:

[0051] S10: Construct a training dataset on satellite proximity events and build a satellite knowledge graph;

[0052] S11: Based on the satellite knowledge graph, predict whether there is a proximity event between any two satellite nodes and obtain the proximity prediction result;

[0053] S12: Input the training dataset into the pre-built graph neural network for training to obtain the proximity training results;

[0054] Figure 2 This is a schematic diagram illustrating the principle of a proximity event prediction method provided in an embodiment of this application. In a specific embodiment, such as... Figure 1 and Figure 2 As shown, in order to obtain a high-precision target graph neural network that can be directly applied to satellite proximity event prediction, a training dataset needs to be constructed first. This training dataset can be selected and obtained from open-source satellite databases.

[0055] Simultaneously, to improve the accuracy of the target graph neural network, in addition to training the initially constructed graph neural network with a training dataset to obtain proximity training results, a satellite knowledge graph is also constructed concurrently. It is understood that a satellite knowledge graph can more accurately describe the relationships between satellites and between satellites and their orbits. Therefore, in one optional embodiment, proximity event prediction can be performed based on the constructed satellite knowledge graph to obtain proximity prediction results.

[0056] It should be noted that the pre-built graph neural network (GNN) can be a GCN (Graph Convolutional Network), a GAT (Graph Attention Network), or a GIN (Graph Isomorphism Network), and this application does not limit it.

[0057] Furthermore, it should be noted that when predicting whether there are proximity events between satellite nodes based on satellite knowledge graphs, inference and prediction can be performed using a large language model in the satellite domain, or a pre-trained prediction model can be used. This application does not limit the scope of the prediction.

[0058] S13: Construct a loss function using proximity prediction results and proximity training results;

[0059] S14: With the goal of minimizing the loss function, return to step S12 for iterative training to obtain the target graph neural network;

[0060] S15: Input the feature data of the satellite to be predicted into the target graph neural network to obtain the target proximity prediction result of the satellite to be predicted.

[0061] like Figure 2 As shown, in a specific embodiment, the graph neural network is trained simultaneously to obtain proximity training results, and proximity event prediction is performed based on the satellite knowledge graph to obtain proximity prediction results. For two identical satellite nodes, if the proximity training results and proximity prediction results are consistent, it indicates that the current graph neural network has learned the proximity event facts between these two satellite nodes. If the proximity training results and proximity prediction results are inconsistent, the proximity knowledge learned by the graph neural network may be incorrect, which will lead to lower accuracy in proximity event prediction for the final target graph neural network.

[0062] Therefore, in an alternative embodiment, such as Figure 2 As shown, the near training results and near prediction results are combined to construct a loss function, which is used to characterize the error between the near training results and the near prediction results.

[0063] In specific embodiments, it can be understood that when the loss function is minimized, the error between the representation proximity training result and the proximity prediction result is minimized, and at this point, the prediction result of the graph neural network is closer to the true result. Therefore, as Figure 2 As shown, after obtaining the loss function, it is determined whether the current loss function is at its minimum. If the current loss function has not reached its minimum, it indicates that the error between the proximity training result and the proximity prediction result is not within an acceptable range, and the graph neural network needs to be trained further. At this time, the process returns to step S12, that is, the training dataset is input again to continuously iterate and train the graph neural network. Of course, if it is at its minimum, it indicates that the graph neural network obtained by the current training meets the prediction accuracy requirements, and the graph neural network obtained by the last iteration of training can be directly used as the target graph neural network.

[0064] The target graph neural network trained using the above methods can be directly applied to the prediction of satellite proximity events. Specifically, by inputting the relevant information of the satellite to be predicted into the target graph neural network, the prediction results of the proximity events of the satellite to be predicted can be obtained directly. This allows for intervention in satellites that may experience proximity events in the future, thus preventing on-orbit collisions.

[0065] Therefore, the proximity event prediction method provided in this application transforms satellites and orbits into a graph structure of knowledge graphs, enabling rapid prediction of the existence of proximity events between satellite nodes based on the characteristics of each node in the knowledge graph. Simultaneously, the graph neural network is trained using a training dataset, allowing it to handle complex nonlinear relationships and learn about proximity events between satellites. Furthermore, the graph neural network is trained by minimizing the loss function constructed from the proximity prediction results and the proximity training results, resulting in an efficient and accurate target graph neural network that can be directly applied to proximity event prediction. This avoids catastrophic on-orbit collisions caused by proximity events and achieves sustainable development of the orbital environment.

[0066] In one optional embodiment, based on a satellite knowledge graph, the presence of a proximity event between any two satellite nodes is predicted to obtain a proximity prediction result, including:

[0067] Extract node features of each entity node from the satellite knowledge graph; entity nodes include satellite nodes and orbital nodes.

[0068] The node features are mapped to a unified vector space to obtain feature vectors;

[0069] The feature vectors are normalized to obtain the target feature vectors;

[0070] The result of element-wise multiplication of any two target feature vectors is input into the multilayer perceptron to obtain the target output result;

[0071] Based on the target output, determine whether there is a proximity event between the two satellite nodes and obtain the proximity prediction result.

[0072] Figure 3 This is a schematic diagram illustrating the principle of node feature extraction provided in an embodiment of this application. In a specific embodiment, it can be understood that each entity node in a satellite knowledge graph has its own features, but these features may not be sufficient to express the semantic information of the node. For example, a satellite node may only have the satellite name as a feature. Therefore, extracting as many features as possible from entity nodes can improve the accuracy of proximity event prediction. In an optional embodiment, such as... Figure 3As shown, a graph neural network (GNN) can be used to learn the position of each entity node in the satellite knowledge graph and its relationship with other entity nodes, thereby obtaining rich feature representations of each entity node. That is, satellite node features and orbit node features can be extracted through GNN.

[0073] In another alternative embodiment, the satellite knowledge graph can be reasoned and analyzed using a large language model in the satellite domain to extract node feature representations for each entity node. This application does not limit the methods used to improve node feature representations.

[0074] It should be noted that, in specific embodiments, the entity nodes in the satellite knowledge graph include satellite nodes and orbital nodes. When extracting node features, since the two types of nodes are closely related, in order to obtain richer node features, it is necessary to extract the node features of both types of nodes.

[0075] Furthermore, in order to facilitate subsequent prediction of nearby events based on node features, the node features need to be processed. First, the node features can be filtered and screened to remove outliers. Then, the node features are mapped to a unified vector space to obtain feature vectors, thereby ensuring that the node features are transformed into learnable quantifiable data.

[0076] In one optional embodiment, to ensure data stability and further improve the prediction accuracy of proximity prediction results, the feature vectors are normalized to obtain target feature vectors. Then, any two target feature vectors are multiplied element-wise, and the multiplication result is input into the multilayer perceptron to obtain the final target output result.

[0077] Specifically, in one optional embodiment, the calculation formula is shown in formula (1):

[0078] (1)

[0079] in, For proximity prediction results, the value range is: . For activation function, As a multi-layer sensing mechanism, For element-wise multiplication, and They represent the first The satellite node and the first The target feature vector of a satellite node, that is, the feature vector obtained from the node features of two different satellite nodes.

[0080] The entire process of determining the target output can be understood as first performing local cross-multiplication, and then global fusion. That is, first, the two target feature vectors are multiplied element-wise, and then a multilayer perception mechanism is incorporated. Learn non-linear interaction.

[0081] In one optional embodiment, based on the target output, it can be determined whether a proximity event exists between the two satellite nodes. Specifically, when the proximity prediction result... When, it indicates that there is a proximity event between two satellite nodes, and when the proximity prediction result At this time, it indicates that there is no proximity event between the two satellite nodes.

[0082] In one optional embodiment, node features of each entity node are extracted from the satellite knowledge graph, including:

[0083] Construct a set of vertices in the hypergraph based on entity nodes, and construct a set of hyperedges in the hypergraph based on whether there is an edge between any two entity nodes;

[0084] Construct a first adjacency matrix based on the hyperedges between any two satellite nodes, and construct a second adjacency matrix based on the hyperedges between a satellite node and an orbital node.

[0085] By concatenating the first and second adjacency matrices, the target adjacency matrix is ​​obtained.

[0086] Based on the target adjacency matrix, the features of each entity node are aggregated through a specified graph neural network to obtain node features.

[0087] In a specific embodiment, a hypergraph is first constructed. SuperMap Includes vertex sets Hyper-edge set and weight matrix Among them, the weight matrix It is a diagonal matrix, and each hyperedge There is a corresponding weight .

[0088] Vertex set Constructed based on entity nodes in satellite knowledge graphs, and with hyperedge sets The construction is based on whether there is an edge between any two entity nodes. Specifically, this includes whether there is an edge between satellite nodes and whether there is an edge between a satellite node and an orbital node.

[0089] Through vertex set and hyper-edge set The first adjacency matrix between satellite nodes can be constructed. and the second adjacency matrix between satellite nodes and orbital nodes. Specifically, the first adjacency matrix can be constructed based on the hyperedges between any two satellite nodes. Meanwhile, a second adjacency matrix can be constructed based on the hyperedges between satellite nodes and orbital nodes. .

[0090] Furthermore, the first adjacency matrix Second Adjacency Matrix By concatenating the matrices, we can obtain the target adjacency matrix. Specifically, the splicing formula is formula (2):

[0091] (2)

[0092] in, This indicates a feature splicing operation.

[0093] It is understandable that the target adjacency matrix yes The matrix, the target adjacency matrix elements in Represents vertices Does (entity node) belong to a hyperedge? In one alternative embodiment, 1 can be used to indicate belonging and 0 to indicate not belonging.

[0094] In an optional embodiment, node features can be aggregated using a specified graph neural network to obtain rich node feature representations for each entity node. Specifically, see formula (3):

[0095] (3)

[0096] in, For the first Node characteristics of an entity node For learnable linear transformations, For aggregation functions (e.g., attention mechanisms, etc.). Let be the vertex degree matrix. Given the target adjacency matrix, This is the weight matrix. The inverse of the edge degree matrix. For the first The adjacent nodes of the nth entity node, that is, the nth The neighbors of each vertex. Each vertex (entity node) (belongs to vertex set) ,Right now, Each vertex Corresponding to a target feature vector ,in, This constitutes the input feature matrix. .

[0097] In specific embodiments, this application does not limit the type of loss function constructed. In one optional embodiment, in order to ensure clearer gradients, more efficient training, and avoid gradient saturation, the binary cross-entropy loss function (BCE Loss) can be selected.

[0098] Specifically, the formula for calculating BCE Loss is formula (4):

[0099] (4)

[0100] in, The value of the loss function. For the training dataset, For the first The satellite node and the first The results of proximity training for each satellite node. This is the result of a near-term prediction.

[0101] In an optional embodiment, the training dataset further includes a test dataset, and the proximity event prediction method includes:

[0102] Input the test dataset into the target graph neural network to obtain the target proximity prediction results;

[0103] Based on the target proximity prediction results, a quality score is determined to evaluate the confidence of the target graph neural network's proximity event prediction.

[0104] If the quality score is not greater than the threshold, return to the step of building a training dataset about satellite proximity events and perform subsequent steps.

[0105] In one alternative embodiment, such as Figure 2 As shown, to further improve the accuracy of proximity event prediction, after obtaining the target graph neural network, it can be tested using a test dataset. Specifically, the test dataset is input into the target graph neural network, and a quality score is calculated based on the target proximity prediction results.

[0106] The quality score is used to evaluate the confidence level of the target graph neural network's proximity event predictions; in other words, it assesses whether the accuracy of the target graph neural network meets the requirements. When the quality score is greater than a threshold, it indicates that the accuracy of the target graph neural network has reached the expected level, and the target graph neural network can be used as the final graph neural network, i.e., as the final network used for satellite proximity event prediction. When the quality score is not greater than the threshold, such as... Figure 2 As shown, the training dataset is continued to be input into the graph neural network for iterative training until a network that meets the conditions is obtained.

[0107] In one alternative embodiment, constructing a satellite knowledge graph includes:

[0108] Obtain the target dataset regarding satellite proximity events;

[0109] The satellites and orbits in the target dataset are used as entity nodes in the satellite knowledge graph;

[0110] Based on proximity events between satellite nodes and the operational relationships between satellite nodes and orbital nodes, generate edges between entity nodes;

[0111] Satellite parameters are injected into the corresponding satellite nodes, and orbital parameters are injected into the corresponding orbital nodes to construct a satellite knowledge graph.

[0112] In a specific embodiment, a target dataset of satellite proximity events is obtained from an open-source satellite database. Furthermore, the satellites and orbits in the target dataset are treated as entity nodes, such as... Figure 3 As shown, the satellite knowledge graph includes two types of nodes: satellite nodes and orbital nodes. In one optional embodiment, the satellite knowledge graph may include 27,433 satellite nodes and 27,433 orbital nodes.

[0113] The satellite knowledge graph includes two types of edges, such as... Figure 3 As shown, one type of edge is the edge between satellite nodes, used to characterize whether a proximity event occurs between the satellites corresponding to two satellite nodes at a certain time. In addition, edges also include those between satellite nodes and orbital nodes, used to characterize the satellite corresponding to a satellite node operating in the orbit corresponding to the orbital node. Therefore, in a specific embodiment, edges between entity nodes can be generated based on proximity events between satellite nodes and the operational relationships between satellite nodes and orbital nodes. In an optional embodiment, the satellite knowledge graph may include 71,572 edges representing satellite proximity relationships and 27,433 edges representing satellite operational and orbital relationships.

[0114] The target dataset also includes a large amount of information about each satellite node and orbital node. To ensure the accuracy of subsequent proximity event predictions, satellite parameters are injected into the corresponding satellite nodes, and orbital parameters are injected into the corresponding orbital nodes. The satellite parameters include at least the satellite number, payload, and physical parameters, while the orbital parameters include at least the orbital semi-major axis, orbital inclination, and eccentricity.

[0115] In the above embodiments, the method for predicting proximity events has been described in detail. This application also provides an embodiment of a device for predicting proximity events.

[0116] Figure 4 This is a schematic diagram of the structure of a proximity event prediction device provided in an embodiment of this application, as shown below. Figure 4 As shown, the device includes:

[0117] The target building module 40 is used to build a training dataset on satellite proximity events and to construct a satellite knowledge graph;

[0118] Prediction module 41 is used to predict whether there is a proximity event between any two satellite nodes based on the satellite knowledge graph, and obtain the proximity prediction result.

[0119] The network training module 42 is used to input the training dataset into the pre-built graph neural network for training and obtain the proximity training result; construct the loss function through the proximity prediction result and the proximity training result; and return to the prediction module with the goal of minimizing the loss function to obtain the target graph neural network.

[0120] Input module 43 is used to input the feature data of the satellite to be predicted into the target graph neural network to obtain the target proximity prediction result of the satellite to be predicted.

[0121] Furthermore, the proximity event prediction device provided in this application embodiment also includes:

[0122] The node feature extraction module is used to extract the node features of each entity node from the satellite knowledge graph; entity nodes include satellite nodes and orbital nodes.

[0123] The mapping module is used to map node features to a unified vector space to obtain feature vectors;

[0124] The normalization module is used to normalize the feature vectors to obtain the target feature vectors;

[0125] The multilayer sensing mechanism module is used to input the element-wise multiplication result of any two target feature vectors into the multilayer sensing mechanism to obtain the target output result;

[0126] The proximity event determination module is used to determine whether a proximity event exists between two satellite nodes based on the target output results, and to obtain the proximity prediction results.

[0127] The hypergraph construction module is used to construct the vertex set in the hypergraph based on entity nodes, and to construct the hyperedge set in the hypergraph based on whether there is an edge between any two entity nodes;

[0128] The adjacency matrix construction module is used to construct a first adjacency matrix based on the corresponding hyperedge between any two satellite nodes, and to construct a second adjacency matrix based on the corresponding hyperedge between a satellite node and an orbital node.

[0129] The matrix concatenation module is used to concatenate the first adjacency matrix and the second adjacency matrix to obtain the target adjacency matrix;

[0130] The feature aggregation module is used to aggregate the features of each entity node based on the target adjacency matrix and through a specified graph neural network to obtain node features.

[0131] The testing module is used to input the test dataset into the target graph neural network and obtain the target proximity prediction results;

[0132] The quality score determination module is used to determine the quality score for evaluating the confidence of the target graph neural network proximity event prediction based on the target proximity prediction results; when the quality score is not greater than the threshold, it returns to call the target construction module.

[0133] The target dataset acquisition module is used to acquire target datasets related to satellite proximity events;

[0134] The entity node determination module is used to identify satellites and orbits in the target dataset as entity nodes in the satellite knowledge graph.

[0135] The edge generation module is used to generate edges between entity nodes based on proximity events between satellite nodes and the operational relationship between satellite nodes and orbital nodes.

[0136] The injection module is used to inject satellite parameters into the corresponding satellite nodes and orbital parameters into the corresponding orbital nodes to construct a satellite knowledge graph.

[0137] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 5 As shown, the electronic device includes: a memory 50 for storing computer programs;

[0138] The processor 51 is configured to implement the steps of the proximity event prediction method as described in the above embodiments when executing a computer program.

[0139] The electronic devices provided in this embodiment may include, but are not limited to, laptops or desktop computers.

[0140] The processor 51 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 51 may be implemented using at least one of the following hardware forms: Digital Signal Processor (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 51 may also include a main processor and a coprocessor. The main processor, also known as the Central Processing Unit (CPU), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 51 may integrate a Graphics Processing Unit (GPU), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 51 may also include an Artificial Intelligence (AI) processor, which is used to handle computational operations related to machine learning.

[0141] The memory 50 may include one or more computer-readable storage media, which may be non-transitory. The memory 50 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In this embodiment, the memory 50 is used to store at least the following computer program 501, which, after being loaded and executed by the processor 51, is capable of implementing the relevant steps of the proximity event prediction method disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 50 may also include an operating system 502 and data 503, etc., and the storage method may be temporary storage or permanent storage. The operating system 502 may include Windows, Unix, Linux, etc. The data 503 may include, but is not limited to, relevant data involved in the proximity event prediction method.

[0142] In some embodiments, the electronic device may further include a display screen 52, an input / output interface 53, a communication interface 54, a power supply 55, and a communication bus 56.

[0143] Those skilled in the art will understand that Figure 5 The structures shown do not constitute a limitation on electronic devices and may include more or fewer components than those shown.

[0144] The electronic device provided in this application includes a memory and a processor. When the processor executes a program stored in the memory, it can implement the proximity event prediction method described in the above embodiments.

[0145] It should be noted that although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

Claims

1. A method of predicting a near-miss event, the method comprising: The method comprises: constructing a training data set about satellite close approach events, and constructing a satellite knowledge graph; based on the satellite knowledge graph, predicting whether there is a close approach event between any two satellite nodes to obtain a close approach prediction result; inputting the training data set into a pre-constructed graph neural network for training to obtain a close approach training result; constructing a loss function through the close approach prediction result and the close approach training result; iteratively training the step of inputting the training data set into the pre-constructed graph neural network to obtain the close approach training result with the goal of minimizing the loss function to obtain a target graph neural network; inputting feature data of a satellite to be predicted into the target graph neural network to obtain a target close approach prediction result of the satellite to be predicted; the step of predicting whether there is a close approach event between any two satellite nodes based on the satellite knowledge graph to obtain a close approach prediction result comprises: learning the positions of each entity node in the satellite knowledge graph and the relationships between the entity nodes and other entity nodes through a graph neural network to extract node features of each entity node; the entity nodes include the satellite nodes and orbit nodes; mapping the node features to a unified vector space to obtain feature vectors; normalizing the feature vectors to obtain target feature vectors; inputting an element-by-element multiplication result of any two target feature vectors into a multilayer perception mechanism to obtain a target output result; determining whether there is a close approach event between two satellite nodes according to the target output result to obtain the close approach prediction result; when the close approach prediction result is 1, it indicates that there is a close approach event between the two satellite nodes; when the close approach prediction result is 0, it indicates that there is no close approach event between the two satellite nodes.

2. The method of claim 1, wherein the near miss event is predicted based on the at least one of the plurality of parameters. learning the positions of each entity node in the satellite knowledge graph and the relationships between the entity nodes and other entity nodes through a graph neural network to extract node features of each entity node; the entity nodes include the satellite nodes and orbit nodes; constructing a vertex set in a hypergraph based on the entity nodes, and constructing a hyperedge set in the hypergraph based on whether there is an edge between any two entity nodes; constructing a first adjacency matrix according to the corresponding hyperedges between any two satellite nodes, and constructing a second adjacency matrix according to the corresponding hyperedges between the satellite nodes and the orbit nodes; concatenating the first adjacency matrix and the second adjacency matrix to obtain a target adjacency matrix; based on the target adjacency matrix, aggregating the features of each entity node through a specified graph neural network to obtain the node features.

3. The method of claim 1, wherein the step of predicting the near miss event is performed by a computer system. The loss function is a binary classification cross-entropy loss function.

4. The method of claim 1, wherein the step of predicting the near miss event is performed by a computer system. The training data set further comprises a test data set, and the method comprises: inputting the test data set into the target graph neural network to obtain a target close approach prediction result; determining a quality score for evaluating the close approach event prediction confidence of the target graph neural network according to the target close approach prediction result; when the quality score is not greater than a threshold value, returning to the step of constructing a training data set about satellite close approach events and performing subsequent steps.

5. The method of claim 1, wherein the step of predicting the near miss event is performed by a computer system. constructing a satellite knowledge graph comprises: obtaining a target data set about satellite close approach events; taking satellites and orbits in the target dataset as entity nodes in the satellite knowledge graph; generating edges between the entity nodes according to proximity events between satellite nodes and operational relationships between the satellite nodes and orbit nodes; injecting satellite parameters into corresponding satellite nodes and injecting orbit parameters into corresponding orbit nodes to construct the satellite knowledge graph.

6. The method of claim 5, wherein the step of predicting the near miss event is performed by a neural network. The satellite parameters at least include satellite number, payload and physical parameters, and the orbit parameters at least include orbit semi-major axis, orbit inclination and eccentricity.

7. A prediction device of a near contact event, characterized by, The device comprises: a target construction module configured to construct a training dataset about satellite proximity events and construct a satellite knowledge graph; a prediction module configured to predict whether there is a proximity event between any two satellite nodes based on the satellite knowledge graph to obtain a proximity prediction result; a network training module configured to input the training dataset into a pre-constructed graph neural network for training to obtain a proximity training result, construct a loss function through the proximity prediction result and the proximity training result, and iteratively train the step of inputting the training dataset into the pre-constructed graph neural network to obtain the proximity training result to obtain a target graph neural network; an input module configured to input feature data of a satellite to be predicted into the target graph neural network to obtain a target proximity prediction result of the satellite to be predicted; a node feature extraction module configured to learn positions of each entity node in the satellite knowledge graph and relationships with other entity nodes through graph neural network learning to extract node features of each entity node; the entity nodes include the satellite nodes and orbit nodes; a mapping module configured to map the node features to a unified vector space to obtain feature vectors; a normalization module configured to normalize the feature vectors to obtain target feature vectors; a multi-layer perception mechanism module configured to input an element-by-element multiplication result of any two target feature vectors into a multi-layer perception mechanism to obtain a target output result; a proximity event determination module configured to determine whether there is a proximity event between two satellite nodes according to the target output result to obtain the proximity prediction result; when the proximity prediction result is 1, it indicates that there is a proximity event between the two satellite nodes; when the proximity prediction result is 0, it indicates that there is no proximity event between the two satellite nodes.

8. An electronic device comprising a memory and a processor, said memory having stored thereon a computer program operable to run on the processor, characterized in that, The processor executes the program to realize the steps of the proximity event prediction method in any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the steps of the proximity event prediction method in any one of claims 1 to 6.

Images