A deep learning-based method for predicting the probability of collision risk of space debris

Abstract

The present application relates to the technical field of aerospace artificial intelligence, and provides a space debris collision risk probability prediction method based on deep learning, which takes ESA Kelvins real rendezvous data as a benchmark, and uses Kessler to generate a synthetic CDM sequence to expand the training sample to improve the coverage of high risk and complex situations; at the same time, a plurality of update records and their associated related events are expressed as a dynamic heterogeneous graph, a dynamic graph neural network is used to learn the law of risk evolution over time and the similarity between events can be optionally used for knowledge transfer, so that the continuous regression prediction result of the final collision risk probability of the rendezvous event can still be stably output under the condition of missing and noise, and the prediction is quickly refreshed when new CDM arrives through the local graph incremental update mechanism, which significantly improves the accuracy, robustness and online availability of risk prediction.

Description

Technical Field

[0001] This invention relates to the field of aerospace artificial intelligence technology, and in particular to a method for predicting the probability of space debris collision risk based on deep learning. Background Technology

[0002] With the continued growth of space activities, the number of space targets in low Earth orbit is rapidly increasing, leading to a rise in space debris density. This has resulted in a significant increase in the frequency of close encounters between spacecraft and space debris, as well as between pieces of space debris themselves. If close encounters are not identified and avoided in a timely manner, they can trigger collisions, disintegration, and a chain reaction of debris, posing a serious threat to on-orbit operational safety, the long-term sustainable operation of constellations, and the reliability of critical payload missions. Therefore, timely and accurate prediction and assessment of the collision risk probability of encounters is a fundamental aspect of space situational awareness and on-orbit operational management.

[0003] In existing engineering processes, rendezvous events are typically detected by ground-based monitoring and orbit determination systems, which generate rendezvous alarm information and provide details such as rendezvous geometry, relative motion, uncertainties, and covariance in the form of a Conjunction Data Message (CDM). Because observational data is continuously updated, the accuracy of orbit determination solutions and uncertainty estimates also change. Multiple update records are often generated for the same rendezvous event before the Time of Closest Approach (TCA). Engineers need to combine records from different update times to repeatedly calculate and screen risks, ultimately deciding whether to trigger evasive maneuvers or take further tracking and observation measures.

[0004] To improve efficiency, the industry has begun to explore machine learning for predicting intersection risks. Kelvins' scoring criteria explicitly state that the task objective is to predict the final risk corresponding to each CDM time series, with predictions of high-risk events being particularly important. However, in existing data-driven solutions, if each CDM is treated as an independent table sample for learning, it is often difficult to fully utilize the evolutionary information updated within the event. If only a general time series model is used to model a single event sequence, the reusable structural information between events is not adequately utilized. In scenarios where high-risk samples are scarce and the distribution has a long tail, problems such as insufficient sensitivity to high-risk events and insufficient prediction stability may still occur.

[0005] Patent application CN116861241A discloses "A method and system for predicting the probability of space debris collisions based on artificial intelligence," which involves using artificial intelligence methods to predict the probability of space debris collisions. It uses an ID3-based decision tree as its core, quantifying and training historical orbits and other data from two rows of root-based TLE data to obtain a collision probability prediction model. Overall, it leans towards static quantification features and a rule-based learning modeling paradigm. However, the aforementioned existing technologies have at least the following shortcomings: First, they are mainly based on static features and rule-based modeling, making it difficult to depict the temporal evolution of multiple CDM records in the same intersection event as observations are updated, and failing to fully utilize the information on the gradual approach and fluctuation of risks within the event; Second, they do not adequately utilize the potential structural relationships between intersection objects and events, making it difficult to adapt to joint risk representation in scenarios with multiple intersections and strong coupling; Third, in real business, the proportion of high-risk samples is extremely low, and the data distribution exhibits a long-tail characteristic, making traditional decision tree methods susceptible to sample imbalance, resulting in insufficient sensitivity to high-risk events; Fourth, when new observation records continuously arrive, existing methods struggle to achieve rapid incremental updates of event risk assessment results, resulting in weak adaptability to online applications. Summary of the Invention

[0006] This invention addresses several technical challenges in existing space debris rendezvous events, including limited data volume, scarcity of high-risk samples, risk fluctuations over time due to multiple CDM updates for the same event, instability in predictions caused by missing records and noise, and difficulty in timely updating assessment results when new records arrive. It proposes a deep learning-based method for predicting the probability of space debris collision risk. Using real ESA Kelvins rendezvous data as a benchmark, the method expands the training samples by generating synthetic CDM sequences using Kessler to improve coverage of high-risk and complex scenarios. Simultaneously, multiple updated records within an event and their relationships are uniformly represented as a dynamic heterogeneous graph. A dynamic graph neural network is used to learn the evolution of risk over time and can optionally utilize the similarity between events for knowledge transfer. This allows for stable continuous regression prediction of the final collision risk probability of the rendezvous event even under conditions of missing and noisy data. Furthermore, a local graph incremental update mechanism rapidly updates the prediction when new CDMs arrive, significantly improving the accuracy, robustness, and online availability of risk prediction.

[0007] This invention provides a method for predicting the probability of space debris collision risk based on deep learning, comprising the following steps:

[0008] Step S1: Obtain real historical CDM data; introduce the Kessler physics simulator and generate synthetic CDM data through the Bayesian probabilistic CDM generation mechanism; and use the synthetic CDM data and real historical CDM data to form a CDM training set.

[0009] Step S2: Perform deep cleaning on the CDM training set and update the CDM training set;

[0010] Step S3: Model the intersection event as a dynamic heterogeneous graph. Construct a dynamic heterogeneous graph topology;

[0011] Step S4: Construct a dynamic graph neural network model, which includes a time encoder and a graph attention mechanism; the dynamic graph neural network model automatically learns the dynamic allocation of observation record weights through message passing on a dynamic heterogeneous graph.

[0012] Step S5: Use the CDM training set to iteratively train the dynamic graph neural network model until the dynamic graph neural network model converges, and obtain the trained dynamic graph neural network model.

[0013] Step S6: Use the dynamic graph neural network model trained in step S5 to predict the probability of collision risk.

[0014] Furthermore, step S1 includes the following steps S11 to S13:

[0015] Step S11: Obtain real historical CDM data;

[0016] Step S12: Using the Kessler physics simulator, synthetic CDM data is generated through the Bayesian probabilistic CDM generation mechanism;

[0017] Step S13: Synthesize CDM data and real historical CDM data to form a CDM training set; perform unified format mapping, field alignment and label standardization on the synthesized CDM data and real historical CDM data, and mix and arrange them according to intersection event level samples to form a CDM training set.

[0018] Furthermore, step S12 includes the following steps S121 to S124:

[0019] Step S121: Using the Kessler physics simulator and the Bayesian probability CDM generation mechanism, set the prior distribution of the positional deviation between the target object and the space debris at the TCA time. ;

[0020] Step S122: Use the SGP4 model to backpropagate the state vector to the time preceding the TCA. Time, to obtain noisy observation points within the corresponding time period;

[0021] Step S123: Based on the obtained noisy observation points, the trajectory is determined using the least squares method, the new state vector and covariance matrix are calculated, and a series of synthetic CDM data are generated.

[0022] Step S124: Calculate the state distance of the synthesized CDM data at the TCA time; if the state distance of the synthesized CDM data at the TCA time is less than the preset safe distance threshold, it is marked as a high-risk collision sample; if the state distance of the synthesized CDM data at the TCA time is greater than or equal to the safe distance threshold, it is marked as a low-risk collision sample.

[0023] Furthermore, step S2 includes the following steps S21 to S22:

[0024] Step S21: For occasional missing covariance off-diagonal terms in CDM data, fill them in using linear interpolation of records from consecutive times within the same event.

[0025] Step S22: Normalize the time offset of the recorded time t relative to the TCA:

[0026] ;

[0027] in, This refers to the generation time of the current CDM record. This is the starting recording time of the CDM sequence corresponding to the intersection event. This is the closest moment of the intersection event. This represents the normalized time offset feature.

[0028] Furthermore, step S3 includes the following steps S31 to S34:

[0029] Step S31: Model the i-th intersection event as the i-th dynamic graph. , where the set of nodes Includes event node sequence and time-series record node sequence edge set Includes time edges connecting adjacent record nodes And the bidirectional home edge connecting record nodes and event nodes. ;

[0030] Step S32: Define the attribute of the time edge as the time interval between adjacent records. It is used to handle the irregularity of CDM data arrival time;

[0031] Step S33: Organize the event nodes, time sequence record nodes, time edges and bidirectional belonging edges according to the unified node type and edge type rules, and associate the CDM numerical features, time features and edge attributes corresponding to each node to form a dynamic heterogeneous graph topology of a single intersection event.

[0032] Step S34: Following the methods in steps S31 to S33, model all intersection events as dynamic heterogeneous graphs. This forms a dynamic heterogeneous graph topology for all intersection events.

[0033] Furthermore, step S4 includes steps S41 to S44:

[0034] Step S41: Construct a dynamic graph neural network model; the dynamic graph neural network model includes at least an input feature mapping module, a temporal evolution module, a graph attention aggregation module, and a regression output module;

[0035] Step S42: Use a fully connected layer to map the node features of the dynamic heterogeneous graph to a unified hidden layer dimension:

[0036] ;

[0037] in, It is a non-linear activation function; Indicates the first The first of the exchange events The input feature vector corresponding to each record, This represents the weight matrix of the input mapping layer. This represents the bias vector of the input mapping layer; This represents the initial hidden layer obtained after mapping through a fully connected layer.

[0038] Step S43: A gated recurrent unit is used on the time edge to recursively update the state within the intersection event, and the time interval encoding vector constructed from the time interval between adjacent records is used as an auxiliary input to the hidden state update process. This allows the model to utilize the current observation and historical state while further perceiving the irregularity of the record arrival time, so as to characterize the nonlinear temporal evolution of the CDM sequence.

[0039] ;

[0040] in, The time interval encoding vector corresponding to the t-th record is used to represent the time interval information between the current record and the previous record; This is the time-coding mapping matrix, used to map the time-interval coding vector to a feature space that matches the hidden state of the node; Let be the initial hidden layer representation of node v at time t; This represents the historical hidden state of node v at time t-1. Let v be the hidden state of node v after being updated by the gated loop unit at time t; T is the total number of records corresponding to this intersection event; t is the time step index of the record within the event.

[0041] Step S43: On the bidirectional home edges, using the graph attention network mechanism, model the association strength between record nodes and intersection event nodes through message passing on the dynamic graph, and dynamically calculate the contribution weight of each record to the risk of the intersection event according to the following formula. This allows for the automatic learning of the dynamic allocation of observation record weights.

[0042] ;

[0043] ;

[0044] in, This can be represented as the transpose of the learnable parameter vector in the graph attention mechanism. denoted as the unnormalized attention score corresponding to the t-th record, which is used to characterize the importance of the record in representing the risk of intersection events under the current graph structure and context; represents the normalized attention weight, reflecting the contribution of the t-th record to the risk prediction of its respective intersection event; T represents the total number of records corresponding to the intersection event; k represents the record index participating in the normalization calculation;

[0045] Using normalized weights The information from all record nodes is weighted and summed to update the global representation of the event node:

[0046] ;

[0047] in, Represents the global representation vector of event node u; This represents the hidden layer representation of the t-th record node v associated with event node u; This represents the normalized attention weight corresponding to the t-th record node; This represents the weight matrix used to linearly transform the features of the record nodes; T represents the total number of records corresponding to this intersection event. This represents the result obtained by weighted aggregation of all recorded node information according to attention weights; Represents a non-linear activation function; the updated This represents the global representation of the event node after integrating information from all record nodes;

[0048] Step S44: Update the event vector Input to a multilayer perceptron, output predicted value .

[0049] Furthermore, step S5 includes the following steps S51 to S52:

[0050] Step S51: Construct the piecewise weighted mean square error loss function :

[0051] ;

[0052] in, This represents the piecewise weighted mean squared error loss function used in model training. Let N represent the set of parameters to be optimized in the dynamic graph neural network model, where N represents the total number of training samples and i represents the index of the i-th training sample. This represents the true risk label corresponding to the i-th intersection event sample. This represents the predicted risk value output by the model for the i-th intersection event sample. Indicates based on real risk labels A defined weighting coefficient function; where:

[0053] ;

[0054] Step S52: Use the CDM training set to train the dynamic graph neural network model and obtain the trained dynamic graph neural network model.

[0055] The present invention provides a method for predicting the probability of space debris collision risk based on deep learning, which has the following beneficial effects:

[0056] 1. This invention organizes multiple CDM update records of the same intersection event into an event sequence in chronological order, and uses a dynamic graph neural network to propagate and aggregate information along the time edges, enabling the model to learn evolutionary patterns rather than relying solely on isolated features at a single moment. This effectively suppresses prediction jumps caused by factors such as single observation errors, covariance estimation jitter, and outliers, improving the consistency of the fit and trend judgment ability of the event's final collision risk probability. Consequently, it reduces unnecessary false alarms in alarm screening and threshold decision-making, and lowers the probability of missing truly high-risk events.

[0057] 2. Due to the low proportion and scarcity of high-risk events in real intersection data, traditional training is easily overwhelmed by a large number of low-risk samples, resulting in the model's insensitivity to high-risk patterns. This invention uses real ESA Kelvins data as a benchmark, introduces Kessler-generated synthetic CDM sequences to expand the training set, and improves the coverage of high-risk intervals and boundary conditions through distribution alignment and targeted sampling. Simultaneously, a weight that monotonically increases with the real risk probability is introduced into the training loss, allowing the model to learn more fully about high-risk events. These mechanisms significantly enhance the model's generalization ability under long-tailed distributions, enabling it to maintain high prediction quality and usability for rare but crucial high-risk intersection scenarios, and improving the efficiency of selecting key events in actual operation.

[0058] 3. This invention introduces a dynamic graph neural network, using spacecraft and potential threat objects as nodes, and intersection relationships within the same event or time window as edges. It encodes relative motion, covariance, uncertainty, minimum distance, and other information from the CDM (Content Management Model) into nodes and features. Subsequently, the dynamic graph neural network is used to jointly model the temporal evolution of the graph structure and features, directly outputting continuous risk values ​​oriented towards the event. This representation more closely reflects the evolution of the situation as observations are updated, which is beneficial for forming stable risk regression predictions in multi-intersection and strongly coupled scenarios, and is more adaptable and scalable in high-risk long-tail scenarios and scenarios with continuously updated data. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the implementation of the deep learning-based spatial debris collision risk probability prediction method provided by this invention. Detailed Implementation

[0060] To make the technical problems solved by this invention, the technical solutions adopted, and the technical effects achieved clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings, not all of them.

[0061] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for predicting the probability of space debris collision risk based on deep learning, which includes the following process:

[0062] Step S1: Obtain real historical CDM data; introduce the Kessler physics simulator and generate synthetic CDM data through the Bayesian probabilistic CDM generation mechanism; and use the synthetic CDM data and real historical CDM data to form a CDM training set; Step S1 includes the following steps S11 to S13:

[0063] Step S11: Obtain real historical CDM data;

[0064] This embodiment can obtain real historical CDM data from the ESA Kelvins competition platform.

[0065] Step S12: Using the Kessler physics simulator, generate synthetic CDM data through the Bayesian probabilistic CDM generation mechanism; Step S12 includes the following steps S121 to S124:

[0066] Step S121: Using the Kessler physics simulator and the Bayesian probability CDM generation mechanism, set the prior distribution of the positional deviation between the target object and the space debris at the TCA time. .

[0067] The prior distribution Using a Gaussian distribution with a mean of 0 and a very small variance ensures that the generated samples have a high probability of close-range intersection. The variance is set to 1% of the actual positional deviation between the target and the debris. This variance setting ensures that the generated samples are more concentrated in the close-range region at the TCA time, thereby increasing the probability of generating high-risk intersection samples.

[0068] Step S122: Use the SGP4 model to backpropagate the state vector to the time preceding the TCA. Time, to obtain noisy observation points within the corresponding time period;

[0069] in, This indicates the time elapsed since the TCA was last encountered, and is typically set to 7 days.

[0070] Step S123: Based on the obtained noisy observation points, the trajectory is determined using the least squares method, the new state vector and covariance matrix are calculated, and a series of synthetic CDM data are generated.

[0071] Step S124: Calculate the state distance of the synthesized CDM data at the TCA time; if the state distance of the synthesized CDM data at the TCA time is less than the preset safe distance threshold, it is marked as a high-risk collision sample; if the state distance of the synthesized CDM data at the TCA time is greater than or equal to the safe distance threshold, it is marked as a low-risk collision sample.

[0072] Because high-risk collision samples account for a small percentage (less than 1%) and are scarce in real historical CDM data, traditional training is easily overwhelmed by a large number of low-risk collision samples, making the model insensitive to high-risk collision patterns. This invention uses real ESA Kelvins data as a benchmark and introduces Kessler to generate synthetic CDM sequences to expand the training set, which can directionally generate a large number of high-risk collision samples (20-50%).

[0073] Step S13: Synthesize CDM data and real historical CDM data to form a CDM training set; perform unified format mapping, field alignment and label standardization on the synthesized CDM data and real historical CDM data, and mix and arrange them according to intersection event level samples to form a CDM training set.

[0074] Specifically, after converting the synthetic CDM sequence and the real historical CDM sequence into the same field structure, time index form and risk label form, they are mixed and sampled according to a preset ratio. Among them, the real historical CDM data is mainly used to maintain the real distribution characteristics, and the synthetic CDM data is mainly used to supplement high-risk collision samples and boundary condition samples, thereby forming a CDM training set that combines authenticity, coverage and category balance.

[0075] The CDM training set formed by this invention has the following characteristics: First, it simultaneously includes real historical intersection samples and directed high-risk intersection samples, balancing authenticity and high-risk coverage; second, the in-event update records are complete, reflecting the dynamic process of risk evolution over time; third, the proportion of high-risk samples is significantly increased compared to the original real data, thereby alleviating the problem of extreme class imbalance; and fourth, the field structure is unified and the label expression is consistent, which facilitates subsequent dynamic graph modeling and neural network training.

[0076] Step S2: Perform deep cleaning on the CDM training set and update the CDM training set.

[0077] To adapt to neural network training, this step performs deep cleaning on the CDM training set, which includes interpolation to complete missing terms in the covariance matrix. Step S2 includes the following steps S21 to S22:

[0078] Step S21: For occasional missing covariance off-diagonal terms in CDM data, fill them in using linear interpolation of records from previous and subsequent times within the same event.

[0079] Step S22: Normalize the time offset of the recorded time t relative to the TCA:

[0080]

[0081] in, This refers to the generation time of the current CDM record. This is the starting recording time of the CDM sequence corresponding to the intersection event. This is the closest moment of the intersection event. The normalized time offset features are represented by the normalization formula. Using this formula, CDM records at different absolute time positions in different intersection events are mapped to a unified relative time interval, thereby improving the comparability of time features between different events and enhancing the model's ability to represent the evolution of intersection risk.

[0082] Step S3: Model the intersection event as a dynamic heterogeneous graph. Construct a dynamic heterogeneous graph topology; step S3 includes the following steps S31 to S34:

[0083] Step S31: Model the i-th intersection event as the i-th dynamic graph. , Represents a set of nodes. Let the set of edges be represented; and the set of nodes be represented. Includes event node sequence and time-series record node sequence edge set Includes time edges connecting adjacent record nodes And the bidirectional home edge connecting record nodes and event nodes. ;

[0084] The time edge is used to transmit messages between local observations and the global state, realizing the hierarchical aggregation of information; the bidirectional attribution edge is used to establish the hierarchical association between the record node and its corresponding intersection event node, so that the local observation information of a single CDM record can converge to the event node to form a global risk representation.

[0085] Step S32: Define the attribute of the time edge as the time interval between adjacent records. It is used to handle the irregularity of CDM data arrival time.

[0086] Step S33: Organize the event nodes, time sequence record nodes, time edges, and bidirectional belonging edges according to unified node type and edge type rules, and associate the CDM numerical features, time features, and edge attributes corresponding to each node to form a dynamic heterogeneous graph topology for a single intersection event.

[0087] Step S34: Following the methods in steps S31 to S33, model all intersection events as dynamic heterogeneous graphs. This forms a dynamic heterogeneous graph topology for all intersection events.

[0088] Step S4: Construct a dynamic graph neural network model, which includes a time encoder and a graph attention mechanism; the dynamic graph neural network model automatically learns the dynamic allocation of observation record weights through message passing on a dynamic heterogeneous graph. Step S4 includes steps S41 to S44:

[0089] Step S41: Construct a dynamic graph neural network model.

[0090] The dynamic graph neural network model includes at least an input feature mapping module, a temporal evolution module, a graph attention aggregation module, and a regression output module.

[0091] Step S42: Use a fully connected layer (MLP) to map the node features of the dynamic heterogeneous graph to a unified hidden layer dimension:

[0092] ;

[0093] in, It is a non-linear activation function. Indicates the first The first of the exchange events The input feature vector corresponding to each record, This represents the weight matrix of the input mapping layer. This represents the bias vector of the input mapping layer; This represents the initial hidden layer obtained after mapping through a fully connected layer, and its dimension is the same as the hidden layer dimension.

[0094] Step S43: A gated recurrent unit (GRU) is used at the time edge to recursively update the state within the intersection event, and the time interval encoding vector constructed from the time interval between adjacent records is introduced as an auxiliary input into the hidden state update process. This allows the model to utilize the current observation and historical state while further perceiving the irregularity of the record arrival time, so as to characterize the nonlinear temporal evolution of the CDM sequence.

[0095] ;

[0096] in, The time interval encoding vector corresponding to the t-th record is used to represent the time interval information between the current record and the previous record; This is the time-coding mapping matrix, used to map the time-interval coding vector to a feature space that matches the hidden state of the node; Let be the initial hidden layer representation of node v at time t; This represents the historical hidden state of node v at time t-1. Let v be the hidden state of node v after being updated by the gated loop unit at time t; T is the total number of records corresponding to this intersection event; and t is the time step index of the record within the event.

[0097] Step S43: On the bidirectional home edges, using the Graph Attention Network (GAT) mechanism, the association strength between record nodes and intersection event nodes is modeled through message passing on the dynamic graph. The contribution weight of each record to the risk of the intersection event is dynamically calculated according to the following formula. This allows for the automatic learning of the dynamic allocation of observation record weights.

[0098] ;

[0099] ;

[0100] in, This can be represented as the transpose of the learnable parameter vector in the graph attention mechanism. denoted as the unnormalized attention score corresponding to the t-th record, which is used to characterize the importance of the record in representing the risk of intersection events under the current graph structure and context; represents the normalized attention weight, reflecting the contribution of the t-th record to the risk prediction of its respective intersection event; T represents the total number of records corresponding to the intersection event; k represents the record index participating in the normalization calculation.

[0101] Using normalized weights The information from all record nodes is weighted and summed to update the global representation of the event node:

[0102] ;

[0103] in, Represents the global representation vector of event node u; This represents the hidden layer representation of the t-th record node v associated with event node u; This represents the normalized attention weight corresponding to the t-th record node; This represents the weight matrix used to linearly transform the features of the record nodes; T represents the total number of records corresponding to this intersection event. This represents the result obtained by weighted aggregation of all recorded node information according to attention weights; Represents a non-linear activation function; the updated This represents the global representation of the event node after integrating the information from each record node.

[0104] Step S44: Update the event vector Input to a multilayer perceptron (Regression Head), output predicted value .

[0105] Output predicted value This represents the probability of collision risk.

[0106] Step S5: Using the CDM training set, iteratively train the dynamic graph neural network model until it converges, obtaining the trained dynamic graph neural network model. Step S5 includes the following steps S51 to S52:

[0107] Step S51: Construct the piecewise weighted mean square error loss function :

[0108] ;

[0109] in, This represents the piecewise weighted mean squared error loss function used in model training. Let N represent the set of parameters to be optimized in the dynamic graph neural network model, where N represents the total number of training samples and i represents the index of the i-th training sample. This represents the true risk label corresponding to the i-th intersection event sample. This represents the predicted risk value output by the model for the i-th intersection event sample. Indicates based on real risk labels A defined weighting coefficient function. Where:

[0110] ;

[0111] This allows samples with higher risk labels to receive greater attention and weight during model training.

[0112] Step S52: Use the CDM training set to train the dynamic graph neural network model and obtain the trained dynamic graph neural network model.

[0113] When the piecewise weighted mean square error loss function Training ends when convergence reaches the threshold.

[0114] Step S6: Use the dynamic graph neural network model trained in step S5 to predict the probability of collision risk.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions for some or all of the technical features, do 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 predicting the probability of space debris collision risk based on deep learning, characterized in that, The process includes the following: Step S1: Obtain real historical CDM data; introduce the Kessler physics simulator and generate synthetic CDM data through the Bayesian probabilistic CDM generation mechanism; and use the synthetic CDM data and real historical CDM data to form a CDM training set. Step S2: Perform deep cleaning on the CDM training set and update the CDM training set; Step S3: Model the intersection event as a dynamic heterogeneous graph. Construct a dynamic heterogeneous graph topology; Step S4: Construct a dynamic graph neural network model, which includes a time encoder and a graph attention mechanism; the dynamic graph neural network model automatically learns the dynamic allocation of observation record weights through message passing on a dynamic heterogeneous graph. Step S4 includes steps S41 to S45: Step S41: Construct a dynamic graph neural network model; the dynamic graph neural network model includes at least an input feature mapping module, a temporal evolution module, a graph attention aggregation module, and a regression output module; Step S42: Use a fully connected layer to map the node features of the dynamic heterogeneous graph to a unified hidden layer dimension: ； in, It is a non-linear activation function; Let represent the input feature vector corresponding to the nth record in the nth intersection event. This represents the weight matrix of the input mapping layer. This represents the bias vector of the input mapping layer; This represents the initial hidden layer obtained after mapping through a fully connected layer. Step S43: A gated recurrent unit is used on the time edge to recursively update the state within the intersection event, and the time interval encoding vector constructed from the time interval between adjacent records is used as an auxiliary input to the hidden state update process. This allows the model to utilize the current observation and historical state while further perceiving the irregularity of the record arrival time, so as to characterize the nonlinear temporal evolution of the CDM sequence. ； in, For the first t The time interval encoding vector corresponding to each record is used to represent the time interval information between the record and its predecessor; This is the time-coding mapping matrix, used to map the time-interval coding vector to a feature space that matches the hidden state of the node; For nodes v In the t The initial hidden layer representation corresponding to each time step; For nodes v In the t -1 time step of historical hidden state; For nodes v In the t The hidden state after each time step is updated by the gated loop unit; T This represents the total number of records corresponding to this intersection event; t This is to record the time step index within the intersection event; Step S44: On the bidirectional home edges, using the graph attention network mechanism, model the association strength between record nodes and intersection event nodes through message passing on the dynamic graph, and dynamically calculate the contribution weight of each record to the risk of the intersection event according to the following formula. This allows for the automatic learning of the dynamic allocation of observation record weights. ； ； in, This can be represented as the transpose of the learnable parameter vector in the graph attention mechanism. Indicates the first t The unnormalized attention score corresponding to each record is used to characterize the importance of that record to the risk representation of the intersection event under the current graph structure and context. This represents the normalized attention weights, reflecting the first... t The contribution of each record to the risk prediction of its respective intersection event; T This indicates the total number of records corresponding to this intersection event; k Indicates the record index that participated in the normalization calculation; Using normalized weights The information from all record nodes is weighted and summed to update the global representation of the event node: ； in, Represents event nodes u The global representation vector; Indicates the relationship with event nodes u The associated first t Record nodes v Hidden layer representation; Indicates the first t Normalized attention weights corresponding to each record node; This represents the weight matrix used to perform a linear transformation on the features of the record nodes; T This indicates the total number of records corresponding to this intersection event; This represents the result obtained by weighted aggregation of all recorded node information according to attention weights; Represents a non-linear activation function; the updated This represents the global representation of the event node after integrating information from all record nodes; Step S45: Update the event vector Input to a multilayer perceptron, output predicted value ; Step S5: Use the CDM training set to iteratively train the dynamic graph neural network model until the dynamic graph neural network model converges, and obtain the trained dynamic graph neural network model. Step S6: Use the dynamic graph neural network model trained in step S5 to predict the probability of collision risk.

2. The method for predicting the probability of space debris collision risk based on deep learning according to claim 1, characterized in that, Step S1 includes the following steps S11 to S13: Step S11: Obtain real historical CDM data; Step S12: Using the Kessler physics simulator, synthetic CDM data is generated through the Bayesian probabilistic CDM generation mechanism; Step S13: Synthesize CDM data and real historical CDM data to form a CDM training set; perform unified format mapping, field alignment and label standardization on the synthesized CDM data and real historical CDM data, and mix and arrange them according to intersection event level samples to form a CDM training set.

3. The method for predicting the probability of space debris collision risk based on deep learning according to claim 2, characterized in that, Step S12 includes the following steps S121 to S124: Step S121: Using the Kessler physics simulator and the Bayesian probability CDM generation mechanism, set the prior distribution of the positional deviation between the target object and the space debris at the TCA time. ; Step S122: Use the SGP4 model to backpropagate the state vector to the time preceding the TCA. Time, to obtain noisy observation points within the corresponding time period; Step S123: Based on the obtained noisy observation points, the trajectory is determined using the least squares method, the new state vector and covariance matrix are calculated, and a series of synthetic CDM data are generated. Step S124: Calculate the state distance of the synthesized CDM data at the TCA time; if the state distance of the synthesized CDM data at the TCA time is less than the preset safe distance threshold, it is marked as a high-risk collision sample; if the state distance of the synthesized CDM data at the TCA time is greater than or equal to the safe distance threshold, it is marked as a low-risk collision sample.

4. The method for predicting the probability of space debris collision risk based on deep learning according to claim 1, characterized in that, Step S2 includes the following steps S21 to S22: Step S21: For occasional missing covariance off-diagonal terms in CDM data, fill them in using linear interpolation of records from consecutive times within the same event. Step S22: Record the time of generation. Normalize the time offset relative to TCA: ； in, This refers to the generation time of the current CDM record. This is the starting recording time of the CDM sequence corresponding to the intersection event. This is the closest moment of the intersection event. This represents the normalized time offset feature.

5. The method for predicting the probability of space debris collision risk based on deep learning according to claim 4, characterized in that, Step S3 includes the following steps S31 to S34: Step S31, the first i The first intersection event is modeled as the... i Animated GIF , where the set of nodes Includes event node sequence and time-series record node sequence edge set Includes time edges connecting adjacent record nodes And the bidirectional home edge connecting record nodes and event nodes. ; Step S32: Define the attribute of the time edge as the time interval between adjacent records. It is used to handle the irregularity of CDM data arrival time; Step S33: Organize the event nodes, time sequence record nodes, time edges and bidirectional belonging edges according to the unified node type and edge type rules, and associate the CDM numerical features, time features and edge attributes corresponding to each node to form a dynamic heterogeneous graph topology of a single intersection event. Step S34: Following the methods in steps S31 to S33, model all intersection events as dynamic heterogeneous graphs. This forms a dynamic heterogeneous graph topology for all intersection events.

6. The method for predicting the probability of space debris collision risk based on deep learning according to claim 5, characterized in that, Step S5 includes the following steps S51 to S52: Step S51: Construct the piecewise weighted mean square error loss function : ； in, This represents the piecewise weighted mean squared error loss function used in model training. This represents the set of parameters to be optimized in a dynamic graph neural network model. N This represents the total number of training samples. i Indicates the first i The index of each training sample. Indicates the first i The actual risk label corresponding to each sample of intersection events The model represents the first i The predicted risk value output from a sample of intersection events. Indicates based on real risk labels A defined weighting coefficient function; where: ； Step S52: Use the CDM training set to train the dynamic graph neural network model and obtain the trained dynamic graph neural network model.

Images