Space target phase combination screening method based on graph neural network

Abstract

The invention relates to the technical field of aviation artificial intelligence, and provides a graph neural network-based space target phase combination screening method, which comprises the following steps of: obtaining a complete catalogue of a space target through an orbit database, and extracting orbit propagation related data; based on all space targets covered in the complete catalogue, establishing a filter framework to screen all target combinations, and constructing a phase combination event set; the combined phase is screened and mapped into a link prediction problem in a graph neural network; a sliding time window method is adopted to construct a graph structure by combining events in a window, and a graph data set is constructed; training the graph neural network model based on the graph data set; by introducing thresholding operation, an edge probability matrix output by the graph neural network model is converted into an adjacent matrix to represent the connection condition between nodes, so that whether a potential phase combination event exists or not is judged. According to the method, the capturing capability of the model on the spatial features can be enhanced, and the combined phase screening efficiency and accuracy are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of aviation artificial intelligence technology, and in particular to a method for selecting phase combinations between space targets based on a graph neural network. Background Art

[0002] With the continuous growth of giant constellations, commercial satellites and other space-based infrastructure, the number of resident space objects, including payloads, satellites and space debris, has increased dramatically. The risk of encounter warning and collision between space objects has increased year by year, which undoubtedly aggravates the Kessler effect and poses a huge challenge to the sustainable development of the space environment. In the increasingly crowded space environment, collision assessment to avoid on-orbit collision events is one of the most important space activities. Compared with collisions between maneuverable space targets, collision events between non-maneuverable space targets occur more frequently. Therefore, collision assessment under full cataloging conditions is more suitable for application in actual scenarios. With the continuous advancement of space observation technology, the number of cataloged space objects will increase significantly, which will put higher computing resource requirements on the current collision assessment tasks and bring huge pressure and challenges to the existing assessment system.

[0003] Conjunction screening plays a vital role in collision assessment. This process propagates the orbits of all space targets and evaluates whether they exceed the preset collision risk threshold at the closest point, thereby determining whether there is a potential conjunction event. This process is usually the most significant part of the entire assessment process in terms of computing time and resource consumption. Traditional conjunction screening methods construct a series of filters to screen target combinations that meet specific rules. Although many researchers have tried to introduce deep learning methods into conjunction screening tasks, most of the current related studies rely on methods based on CDM (Conjunction Data Message) data. This reliance limits its applicability under full cataloging conditions and makes it difficult to meet the needs of comprehensive screening. Summary of the invention

[0004] The present invention mainly solves the problem of computational resource consumption in conjunction screening caused by the increase in the number of space objects, and aims at the technical problem of the limitations of existing deep learning methods based on CDM data in terms of coverage. A conjunction screening method between spatial targets based on graph neural networks is proposed. The graph data construction framework and algorithm application process are adopted, and the graph data generation framework is constructed using TLE data to ensure the wide applicability of the method. By introducing graph neural networks, the model's ability to capture spatial features is significantly enhanced, effectively improving the efficiency and accuracy of conjunction screening.

[0005] The present invention provides a method for selecting spatial objects based on a graph neural network, comprising:

[0006] Step S1: Obtain a complete catalog of space objects through the orbital database and extract orbital propagation related data;

[0007] Step S2: Based on all space targets covered in the complete catalog, a filter framework is established to screen all target combinations and construct a conjunction event set;

[0008] Step S3: Map the conjunction screening to a link prediction problem in graph neural network;

[0009] Step S4: Using the sliding time window method, the conjunction events in the window are constructed into a graph structure to construct a graph data set;

[0010] Step S5: training the graph neural network model based on the graph dataset;

[0011] Step S6: By introducing a thresholding operation, the edge probability matrix output by the graph neural network model is converted into an adjacency matrix to represent the connection between nodes, so as to determine whether there is a potential conjunction event.

[0012] Furthermore, in step S2, the process of establishing the filter framework is as follows:

[0013] Step S201: Establish a time series to clearly define the analysis time range for screening spatial targets and the interval of each time step;

[0014] Step S202: For all space objects in the complete catalog, the orbital initial states are propagated one by one using the orbital propagation tool to generate all orbital states in the time series;

[0015] Step S203: construct a target combination of all space targets in the time series, calculate the orbital distance between each target combination, and compare it with the set minimum collision threshold to determine whether there is a conjunction event;

[0016] Step S204: extract event information and construct a conjunction event set.

[0017] Furthermore, in step S3, the method of mapping the phase screening to the link prediction problem in the graph neural network is:

[0018] All space objects are represented as nodes and identified by their NORAD IDs; the characteristics of each node include the time t when the conjunction event occurs. i And its corresponding orbital state E ti ; In addition, the existence or absence of edges between nodes is used to indicate whether a conjunction event occurs between different space targets.

[0019] Furthermore, in step S4, the sliding time window method is used to construct the conjunction events in the window into a graph structure as follows:

[0020] Step S401: define a sliding time window;

[0021] Step S402: applying a sliding time window method to the conjunction event set constructed in step S2, extracting conjunction event information within each time window, and storing them separately;

[0022] Step S403: For the conjunction events in each time window, a corresponding graph structure is constructed and stored as a graph dataset for training the graph neural network.

[0023] Furthermore, in step S402, the sliding time window operation method is:

[0024] The conjunction event set is divided into training set, validation set and test set, and the overlapping window method is used to select samples in the three data sets.

[0025] Furthermore, in step S403, the graph structure construction method is:

[0026] The conjunction events within the sliding time window are counted, all relevant space targets are regarded as nodes, and edge sets are established based on the target pairs involved in the events within the window. The event time and the corresponding orbital state information are matched as node features to construct the graph structure.

[0027] Furthermore, in step S5, the training method of the graph neural network is:

[0028] In each training cycle, negative samples are generated using a global random sampling strategy, which randomly extracts them from the unconnected part except the positive samples. In addition, the sampling rate parameter is set to accurately control the ratio of positive and negative samples.

[0029] The present invention provides a method for screening spatial objects based on graph neural network, which takes spatial objects as nodes and the relationships between spatial objects as edges. It uses the classification idea of ​​machine learning to view the screening task and maps it to the link prediction problem in the graph neural network to capture the spatial dependency between large-scale objects. The present invention discloses a data set construction framework for generating the original phase data and the corresponding graph structure required for the algorithm to run. The graph data construction framework and the algorithm application process are adopted, and the graph data generation framework is constructed using TLE data. Compared with the CDM data relied on by most studies, the present invention has the characteristics of high update frequency and wide coverage, thereby ensuring the wide applicability of the method; by introducing the graph neural network, the model's ability to capture spatial features is significantly enhanced, and the efficiency and accuracy of phase screening are effectively improved. The present invention uses TLE with higher update frequency and wider coverage as raw data, builds a filter framework to generate graph structure data, and is used to train the graph neural network model. The method provided by the present invention provides a new means for screening spatial objects, which is a new attempt of artificial intelligence methods in the field of aerospace and has good practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of the method for selecting spatial targets based on graph neural network in the present invention;

[0031] Figure 2 It is a filter framework logic diagram for constructing a conjunction event set in the present invention;

[0032] Figure 3 It is a sliding time window logic diagram used to construct a graph data set in the present invention;

[0033] Figure 4 This is a diagram of the training method of the neural network in the present invention;

[0034] Figure 5 It is a performance curve diagram of the accuracy index of the present invention. DETAILED DESCRIPTION

[0035] In order to make the technical problems solved by the present invention, the technical solutions adopted and the technical effects achieved clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention are shown in the accompanying drawings, rather than all the contents.

[0036] like Figure 1 As shown, an embodiment of the present invention provides a method for selecting spatial target phase combinations based on a graph neural network, including:

[0037] Step S1: Obtain a complete catalog of space objects through the orbital database and extract orbital propagation related data.

[0038] Specifically, the complete catalog contains all cataloged space objects and is given in TLE (Two-line Element) format; at the same time, the extracted orbital propagation related data include Julian epochs and orbital elements.

[0039] Step S2: Based on all space targets covered in the complete catalog, a filter framework is established to screen all target combinations and construct a conjunction event set.

[0040] Specifically, the combination of all targets is composed of all the space targets in the catalog in pairs, and does not include pairs consisting of the same targets; the conjunction event set is composed of events that meet the preset minimum collision threshold rule, and the event information includes the time of event occurrence, the relevant space targets, and the orbital state vector of the relevant space targets at that time.

[0041] Specifically, if Figure 2 The process of establishing the filter framework is shown as:

[0042] Step S201: Establish a time series to clearly define the analysis time range for screening spatial targets and the interval between each time step.

[0043] The specific process of this step: Establishing a time series t i is the time point, where i = [0, 1, 2, …, fs, f], t0 represents the initial moment, f represents the time range, s represents the time step interval, and l = f / s represents the number of time points, which aims to clarify the analysis time range of the screening spatial target and the interval of each time step to ensure that the screening process is carried out within a specific time period;

[0044] Step S202: For all space objects in the complete catalog, use the orbit propagation tool to propagate their initial orbital states one by one to generate all orbital states in a time series.

[0045] The specific process of this step: for all spatial targets in the catalog Use the orbit propagation tool to determine the initial state of its orbit Propagate one by one to generate all orbital states under time series T Where n represents the number of space targets, and v represents the dimension of the orbital state vector;

[0046] Step S203: construct a target combination of all space targets in the time series, calculate the orbital distance between each target combination, and compare it with the set minimum collision threshold to determine whether there is a conjunction event.

[0047] The specific process of this step is to construct the target combination of all space targets in the time series and calculate the orbital distance between each target combination. And with the set minimum collision threshold d min Compare to determine whether there is a conjunction event. Take the target combination (a, b) as an example, t i The orbital distance between the targets at the time The calculation method is:

[0048]

[0049] Step S204: extract event information and construct a conjunction event set.

[0050] The event information includes the time t when the event occurred i , the NORAD ID (North American Aerospace Defense Command catalog number) of the associated space target and the orbital status of the associated space target at that moment

[0051] Step S3: Map the conjunction screening to a link prediction problem in graph neural network;

[0052] Specifically, the method of mapping conjunction screening to the link prediction problem in graph neural network is:

[0053] All space targets are represented as nodes and identified by their NORAD ID (North American Aerospace Defense Command directory number); the characteristics of each node include the time t when the conjunction event occurs i And its corresponding orbital state E ti ; In addition, the existence or absence of edges between nodes is used to indicate whether a conjunction event occurs between different space targets.

[0054] Step S4: Using the sliding time window method, the conjunction events in the window are constructed into a graph structure to construct a graph data set;

[0055] Specifically, if Figure 3 As shown in the figure, the process of constructing the conjunction events in the window into a graph structure using the sliding time window method is as follows:

[0056] Step S401: define a sliding time window.

[0057] Set the window length to w l , and set the window sliding step size to w s ;

[0058] Step S402: applying a sliding time window method to the conjunction event set constructed in step S2, extracting conjunction event information within each time window, and storing them separately;

[0059] Specifically, the operation method of the sliding time window is to split the conjunction event set into a training set, a validation set, and a test set, and use overlapping windows to select samples in the three data sets respectively.

[0060] Step S403: For the conjunction events in each time window, a corresponding graph structure is constructed and stored as a graph dataset for training the graph neural network.

[0061] Specifically, the method of constructing a graph structure using information in a time window is as follows: counting conjunction events within a sliding time window, treating all relevant spatial targets as nodes, and establishing an edge set based on the target pairs involved in the events within the window, matching event times and corresponding orbital state information as node features to construct a graph structure.

[0062] Step S5: Train the graph neural network model based on the graph dataset.

[0063] Among them, the input of the graph neural network model includes the graph topology structure and its node characteristics, and the output of the graph neural network model is a matrix representation of the existence probability of each edge in the graph.

[0064] Specifically, if Figure 4 As shown in the figure, the training method of the graph neural network model is as follows: in each training cycle, negative samples are generated by using a global random sampling method, and are randomly extracted from the unconnected part except for positive samples. This method is designed to ensure that the model can be exposed to as many negative samples as possible in a limited number of training rounds, thereby improving the generalization ability of the model. In addition, the sampling rate parameter is set to control the ratio of positive and negative samples and balance the training process of the model.

[0065] Step S6: By introducing a thresholding operation, the edge probability matrix output by the graph neural network model is converted into an adjacency matrix to represent the connection between nodes, so as to determine whether there is a potential conjunction event.

[0066] Specifically, the thresholding operation method is: by comparing the predicted probability p of the edge with the preset threshold θ, a clear binary classification result is obtained; when p ≥ θ, the output classification result is 1; when p < θ, the output classification result is 0, which indicates whether there is a conjunction event between the nodes. The preset threshold can be adjusted according to the specific event requirements to improve the accuracy of the prediction.

[0067] Under the action of the spatial target phase screening method based on graph neural network, such as Figure 5As shown, the accuracy index shows good convergence regardless of whether the negative sampling rate is 1 or 2. In addition, when random negative sampling is used during training, the curve does not show large oscillation, which shows that the method provided by the present invention is not only suitable for the phase screening task, but also has good robust performance and excellent generalization ability.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that modifying the technical solutions described in the aforementioned embodiments, or replacing part or all of the technical features therein by equivalents, does not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for selecting spatial objects based on graph neural network, characterized in that: include: Step S1: Obtain a complete catalog of space objects through the orbital database and extract orbital propagation related data; Step S2: Based on all space targets covered in the complete catalog, a filter framework is established to screen all target combinations and construct a conjunction event set; Step S3: Map the conjunction screening to a link prediction problem in graph neural network; Step S4: Using the sliding time window method, the conjunction events in the window are constructed into a graph structure to construct a graph data set; Step S5: training the graph neural network model based on the graph dataset; Step S6: By introducing a thresholding operation, the edge probability matrix output by the graph neural network model is converted into an adjacency matrix to represent the connection between nodes, so as to determine whether there is a potential conjunction event.

2. The method for selecting spatial objects based on graph neural network according to claim 1 is characterized in that: In step S2, the process of establishing the filter framework is: Step S201: Establish a time series to clearly define the analysis time range for screening spatial targets and the interval of each time step; Step S202: For all space objects in the complete catalog, the orbital initial states are propagated one by one using the orbital propagation tool to generate all orbital states in the time series; Step S203: construct a target combination of all space targets in the time series, calculate the orbital distance between each target combination, and compare it with the set minimum collision threshold to determine whether there is a conjunction event; Step S204: extract event information and construct a conjunction event set.

3. The method for selecting spatial objects based on graph neural network according to claim 2 is characterized in that: In step S3, the method of mapping the phase screening to the link prediction problem in the graph neural network is: All space objects are represented as nodes and identified by their NORAD IDs; the characteristics of each node include the time t when the conjunction event occurs. i And its corresponding orbital state E ti ; In addition, the existence or absence of edges between nodes is used to indicate whether a conjunction event occurs between different space targets.

4. The method for selecting spatial objects based on graph neural network according to claim 3 is characterized in that: In step S4, the sliding time window method is used to construct the conjunction events in the window into a graph structure as follows: Step S401: define a sliding time window; Step S402: applying a sliding time window method to the conjunction event set constructed in step S2, extracting conjunction event information within each time window, and storing them separately; Step S403: For the conjunction events in each time window, a corresponding graph structure is constructed and stored as a graph dataset for training the graph neural network.

5. The method for selecting spatial objects based on graph neural network according to claim 4 is characterized in that: In step S402, the sliding time window operation method is: The conjunction event set is divided into training set, validation set and test set, and the overlapping window method is used to select samples in the three data sets respectively.

6. The method for selecting spatial objects based on graph neural network according to claim 5, characterized in that: In step S403, the graph structure construction method is: The conjunction events within the sliding time window are counted, all relevant space targets are regarded as nodes, and edge sets are established based on the target pairs involved in the events within the window. The event time and the corresponding orbital state information are matched as node features to construct the graph structure.

7. The method for selecting spatial objects based on graph neural network according to claim 1, characterized in that: In step S5, the training method of the graph neural network is: In each training cycle, negative samples are generated using a global random sampling strategy, which randomly extracts them from the unconnected part except the positive samples. In addition, the sampling rate parameter is set to accurately control the ratio of positive and negative samples.

Images