A deep learning-based space debris environment evolution prediction method

Abstract

The application relates to the technical field of aerospace and artificial intelligence, and provides a space debris environment evolution prediction method based on deep learning, which comprises the following steps: S1, constructing an original data set of the state of a space debris environment evolved year by year under different evolution parameter settings; S2, constructing a multi-dimensional representation describing the evolution state of the space debris environment; S3, sequentially performing pairing processing to construct training samples and forming a complete training data set with labels; S4, performing supervised training on a multi-frequency perception fusion neural operator network model to obtain a trained multi-frequency perception fusion neural operator network model; and S5, performing evolution prediction on the space debris environment under different initial states and different evolution parameter conditions based on the trained multi-frequency perception fusion neural operator network model. The application is suitable for fine representation of a space debris environment in a real scene and realizes efficient prediction of a space debris environment at any target moment with parameter robustness.

Description

Technical Field

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

[0002] With the increasing frequency of human space activities, the space debris environment is becoming increasingly crowded and complex. The risk of cascading collisions caused by the Kessler effect continues to rise. Once the debris density reaches a critical threshold, it may lead to an exponential increase in the number of debris, forming an irreversible vicious cycle. This effect has attracted great attention from the international space community and poses a serious challenge to satellite operational safety, space traffic management, and future large-scale constellation deployments. Efficient and accurate characterization of the space debris environment and long-term evolution prediction are of great guiding significance for the formulation of space debris mitigation strategies, the planning of proactive removal missions, and the construction of space situational awareness systems.

[0003] Currently, mainstream methods for analyzing the evolution of space debris environments primarily rely on deterministic physical models. While these methods offer high accuracy in terms of physical realism, they generally suffer from significant computational efficiency bottlenecks and require massive amounts of computational resources, making them unsuitable for practical applications.

[0004] In recent years, deep learning models have demonstrated tremendous potential across various fields, achieving significant improvements in computational efficiency while maintaining high accuracy. However, the application of deep learning methods in space debris environment evolution tasks remains extremely limited. The main reasons include the severe scarcity of datasets and the sharp decline in generalization performance of existing models when evolutionary parameters change. This often necessitates re-collecting data and retraining from scratch for specific parameter configurations, which greatly restricts their applicability in practical engineering applications.

[0005] Currently, several methods exist for improving the efficiency of space debris environment evolution prediction. For example, patents CN113935174A and CN114861570A propose efficient evolution prediction schemes that use orbital altitude and areal density ratio to construct a two-dimensional structure through mesh generation. However, these methods ignore the practical problem that areal density ratio is difficult to obtain accurately in actual observations, thus limiting the applicability of mesh generation. Existing technologies employ deep learning methods to predict the density distribution of space debris environments, typically using sequential input to allow the model to capture potential evolutionary dynamics. However, these methods do not fully consider the scarcity of continuous high-quality data in real-world scenarios, failing to effectively support the need for sequential input and thus limiting their engineering applicability. Summary of the Invention

[0006] This invention primarily addresses the computational efficiency problem of existing technologies for predicting the evolution of space debris environments, as well as the technical problem of insufficient applicability of deep learning methods in practical applications. It proposes a deep learning-based method for predicting the evolution of space debris environments. By correlating orbital height-inclination gridded representations with multi-dimensional features, it achieves a refined representation of space debris environments applicable to real-world scenarios. Evolutionary parameters and time intervals are integrated into the model input features using conditional encoding, driving a multi-frequency sensing fusion neural operator network to achieve parameter-robust and efficient prediction of space debris environments at any target time.

[0007] This invention provides a deep learning-based method for predicting the evolution of spatial fragmentation environments, comprising the following processes:

[0008] Step S1: Construct the original dataset of the space debris environment state evolving year by year under different evolution parameter settings;

[0009] Step S2: Based on the original dataset, construct a multidimensional representation describing the evolutionary state of the space debris environment;

[0010] Step S3: Based on the multidimensional representation and each set of evolution parameters corresponding to the multidimensional representation. Each sample is paired up to construct training samples, forming a complete labeled training dataset.

[0011] The input of the training samples Includes: Initial multidimensional characterization of space debris environments Evolution parameters and prediction time interval Output Multidimensional representation of the spatial fragmentation environment at the target time This forms a complete labeled training dataset;

[0012] Step S4: Using the labeled training dataset described in step S3, supervised training is performed on the multi-frequency sensing fusion neural operator network model to obtain the trained multi-frequency sensing fusion neural operator network model.

[0013] The input of the multi-frequency sensing fusion neural operator network model is the initial multidimensional representation of the space debris environment, evolution parameters, and prediction time interval, and the output is the multidimensional representation of the space debris environment at the target time.

[0014] Step S5: Based on the multi-frequency sensing fusion neural operator network model trained in step S4, perform evolution prediction on the space debris environment under different initial states and different evolution parameters, and output a multi-dimensional representation of the space debris environment at the target time.

[0015] Furthermore, step S1 includes the following steps S101 to S102:

[0016] Step S101: Define the initial space debris environment state and evolution parameter space. ;

[0017] The space debris environment state includes at least the number of all space objects in the current space debris environment. and orbital elements for each space target, the orbital elements including orbital inclination. semi-major shaft eccentricity Right ascension of ascending node and perigee argument ;

[0018] The evolution parameter space The evolution parameter space is used to define a reasonable range for multi-scenario evolution scenarios. The evolution parameters include at least the initial target quantity ratio. Emission rate Post-task handling success rate and the failure rate of collision avoidance maneuvers ;

[0019] Step S102: For each set of evolutionary parameters in the evolutionary parameter space Starting from the initial state of the space debris environment, an annual simulation is performed to obtain the state of the space debris environment of all space targets involved in the simulation in each year, forming the original dataset.

[0020] Furthermore, step S2 includes the following steps S201 to S203:

[0021] Step S201: Iterate through each set of evolutionary parameters in the original dataset. and evolution parameters Construct a basic statistical data set based on the corresponding annual spatial debris environment status. Among them, basic statistical data sets At least including: orbital altitude set Orbital Inclination Set Total target quantity Number of controlled satellites Number of failed satellites Number of fragments Number of rocket bodies Eccentricity set set of ascending nodes and right ascension and perigee argument set ;

[0022] Step S202: Establish a track height and orbital inclination The gridded phase space structure is used as the coordinate axis, and the grid cells in the gridded phase space structure are assigned values ​​in the form of target number density distribution;

[0023] The gridded phase space structure consists of multiple grid cells; orbital height exist Within the interval Step size discretization One bin, orbital inclination exist Within the range Step size discretization Each bin constitutes Meshized phase space;

[0024] Step S203: For each grid cell in step S202, calculate and associate the feature vector extracted from the basic statistical feature set to construct a multidimensional representation for describing the space debris environment;

[0025] The mathematical representation of the eigenvector is as follows: Features include the logarithmic form of the target quantity. and dynamic dimension features; the dynamic dimension features include the median eccentricity. Eccentricity interquartile range Right ascension entropy of ascending node and perigee angle entropy ;

[0026] The multidimensional representation is composed of a gridded phase space structure and the feature vectors associated with each grid cell. The gridded phase space structure is constructed in step S202, and the feature vectors are extracted and associated with the corresponding grid cells in step S203.

[0027] Furthermore, the multi-frequency sensing fusion neural operator network model includes a conditional coding module, a frequency domain segmentation module, a cross-frequency interaction module, and a residual hierarchical fusion module connected in sequence.

[0028] The conditional encoding module jointly encodes the input state tensor, scene evolution parameters, and prediction time interval into a high-dimensional hidden representation.

[0029] The frequency domain segmentation module applies a two-dimensional Fourier transform to the high-dimensional hidden representation, decomposing the complete spectrum into multiple sub-frequency bands such as low frequency, mid frequency, and high frequency;

[0030] The cross-frequency interaction module performs full-channel spectrum convolution on each sub-frequency band and then models the coupling relationship between different sub-frequency bands through a frequency band mixing matrix.

[0031] The residual hierarchical fusion module uses low-frequency components as a basis, and successively superimposes mid-frequency and high-frequency components. After recovering to the spatial domain through inverse Fourier transform, the prediction result is output in residual form. The prediction result is a multidimensional representation of the spatial debris environment at the target time. .

[0032] This invention proposes a deep learning-based method for predicting the evolution of space debris environments. It constructs a multi-dimensional representation framework for space debris environments and designs an evolution prediction mechanism driven by a deep learning proxy model. Compared to traditional analysis methods that rely on deterministic physical models and suffer from significant bottlenecks in computational efficiency and resource consumption, this invention significantly improves computational efficiency and reduces computational resource requirements while maintaining prediction accuracy.

[0033] To address the problem that existing efficient evolution prediction methods are insufficiently applicable due to the difficulty in obtaining the surface-to-mass ratio parameter, this invention proposes a gridded characterization structure for orbital altitude-inclination. Quantitative and dynamic features are introduced into each grid cell, and while ensuring that the feature parameters are easily obtainable, the grid structure is given a clear physical meaning, thereby achieving a refined characterization of the space debris environment.

[0034] Furthermore, addressing the limitations of existing deep learning methods in predicting the evolution of space debris environments due to the scarcity of continuous high-quality data and the diversity of evolution scenarios, this invention explicitly embeds evolution parameters and time intervals into the model input, constructing a multi-frequency sensing fusion neural operator network model. This enables efficient and accurate long-term evolution prediction of space debris environments under arbitrary evolution scenarios, relying only on the initial state at a single moment. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the implementation of the deep learning-based spatial fragmentation environment evolution prediction method provided by the present invention.

[0036] Figure 2 This is an example diagram of the orbital altitude distribution for the refined characterization of the space debris environment according to the present invention;

[0037] Figure 3 This is an example diagram of the orbital altitude-inclination density distribution for the refined characterization of the space debris environment according to the present invention;

[0038] Figure 4 This is an example diagram of the orbital altitude-inclination-eccentricity distribution for the refined characterization of the space debris environment in this invention;

[0039] Figure 5 This is an example diagram of the orbital altitude-inclination-right ascension entropy distribution for the refined characterization of the space debris environment in this invention;

[0040] Figure 6This is an example diagram of the orbital altitude-inclination-perigee argument entropy distribution for the refined characterization of the space debris environment according to the present invention;

[0041] Figure 7 (a)-(e) are the evaluation results of each performance index of the method of the present invention under different prediction time intervals. Figure 7 (a) shows the performance of MAE. Figure 7 (b) shows the performance of SMAPE. Figure 7 (c) shows the performance of SSIM. Figure 7 (d) shows the performance of Recall. Figure 7 (e) represents the performance of the F1 score. Detailed Implementation

[0042] 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.

[0043] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for predicting the evolution of spatial fragmentation environment based on deep learning, which includes the following process:

[0044] Step S1: Construct the original dataset of the space debris environment state evolving year by year under different evolution parameter settings;

[0045] In this step, raw data is generated to build the training dataset for the neural operator network model. This raw data covers the changes in the space debris environment state of all initial and newly added space targets during long-term evolution, and fully incorporates the dynamic impact of key events on the number and distribution of space targets. The space targets include space debris, rocket bodies, and controlled satellites. The key events include launch missions, reentry events, disintegration events, and active cleanup.

[0046] Step S1 specifically includes the following steps S101 to S102:

[0047] Step S101: Define the initial space debris environment state and evolution parameter space. ;

[0048] The space debris environment state includes at least the number of all space objects in the current space debris environment. and orbital elements for each space target, the orbital elements including orbital inclination. semi-major shaft eccentricity Right ascension of ascending node and perigee argument Among them, the semi-major axis Used to calculate track height .

[0049] Specifically, the evolution parameter space The evolution parameter space is used to define a reasonable range for multi-scenario evolution scenarios. The evolution parameters include at least the initial target quantity ratio. Emission rate Post-task handling success rate and the failure rate of collision avoidance maneuvers Evolutionary parameter space Defined as: , , , .

[0050] Step S102: For each set of evolutionary parameters in the evolutionary parameter space Starting from the initial state of the space debris environment, an annual simulation is performed to obtain the state of the space debris environment of all space targets involved in the simulation in each year, forming the original dataset.

[0051] Specifically, each set of evolutionary parameters From the evolution parameter space The parameters are obtained by random sampling to ensure the diversity of evolution parameter distribution in the simulation scenario. For each set of randomly sampled evolution parameters, orbit propagation is performed on all current space targets, orbit propagation calculation is performed on all current space targets, and the total number of space targets and the individual orbital elements of each space target are updated in real time to obtain the space debris environment status for each year.

[0052] The orbital propagation calculation fully considers the dynamic source-sink effects, including the increase in the number of space targets due to repeated launches and disintegration events, and the decrease in the number of space targets due to reentry events and active removal.

[0053] Specifically, for each set of evolutionary parameter configurations, the long-term evolution simulation of the space debris environment state is advanced year by year, and the state data at the end of each year is independently stored as a timestamp snapshot, while the specific values ​​of the set of evolutionary parameters are recorded in association.

[0054] Step S2: Based on the original dataset, construct a multidimensional representation describing the evolutionary state of the space debris environment;

[0055] In this step, a multidimensional dataset is constructed based on the original dataset generated in step S1. This multidimensional dataset can simultaneously meet the requirements for refined representation of the spatial fragmentation environment and the supervised training requirements for the neural operator network model used for environmental evolution. Step S2 specifically includes the following steps S201 to S203:

[0056] Step S201: Iterate through each set of evolutionary parameters in the original dataset. and evolution parameters Construct a basic statistical data set based on the corresponding annual spatial debris environment status. ;

[0057] Specifically, basic statistical data sets At least including: orbital altitude set Orbital Inclination Set Total target quantity Number of controlled satellites Number of failed satellites Number of fragments Number of rocket bodies Eccentricity set set of ascending nodes and right ascension and perigee argument set Among them, orbital altitude and orbital inclination are mainly used to characterize the density distribution of space debris in phase space, reflecting the macroscopic structural characteristics of the space dimension; the total number of targets, the number of controlled satellites, the number of failed satellites, the number of debris, and the number of rocket bodies are used to depict the overall evolution of the space debris environment, reflecting the dynamic characteristics of the macroscopic quantitative scale; eccentricity, right ascension of the ascending node, and argument of perigee are mainly used for the orbital geometry characterization in the dynamic dimension.

[0058] Step S202: Establish a track height and orbital inclination The gridded phase space structure is used as the coordinate axis, and the grid cells in the gridded phase space structure are assigned values ​​in the form of target number density distribution;

[0059] Specifically, the gridded phase space structure includes multiple grid cells; orbital height exist Within the interval Step size discretization One bin, orbital inclination exist Within the range Step size discretization Each bin constitutes Meshable phase space. For example, orbital altitude. exist The interval is discretized into 180 bins with a step size of 10km, and the orbital inclination is... exist Within the range The step size is discretized into 36 bins, forming a 180×36 gridded phase space. Then, for each grid cell, from the basic statistical data set... Extract and correlate the number of spatial fragments in the corresponding region.

[0060] Step S203: For each grid cell in step S202, calculate and associate the feature vector extracted from the basic statistical feature set to construct a multidimensional representation for describing the space debris environment.

[0061] Specifically, the mathematical representation of the feature vector in step S203 is as follows: The features include the logarithmic form of the target quantity. and dynamic dimension features; the dynamic dimension features include the median eccentricity. Eccentricity interquartile range Right ascension entropy of ascending node and perigee angle entropy .

[0062] Specifically, the multidimensional representation described in step S203 is composed of a gridded phase space structure and the feature vectors associated with each grid cell. The gridded phase space structure is constructed in step S202, and the feature vectors are extracted and associated with the corresponding grid cells in step S203.

[0063] The multidimensional representation is represented as a numerical feature tensor extracted and organized from the original dataset of the space debris environment. This numerical feature tensor includes, but is not limited to, orbital altitude-inclination gridded density statistics and the feature vector obtained in step S203. These multidimensional representations serve as direct inputs to the neural operator network model, driving its learning and prediction processes. Figures 2 to 6 All images are the result of visualizing the multidimensional representations of the model input. These visualizations are only used to assist in understanding the state of the space debris environment and to demonstrate prediction results; they do not constitute any input data during model training or inference. Figure 2 Using orbital altitude as the horizontal axis and the number of objects as the vertical axis, the data is overlaid and displayed according to controlled satellites, failed satellites, debris, and rocket bodies, revealing the distribution characteristics of space targets at different orbital altitudes. Figure 3 Using orbital altitude and orbital inclination as coordinate axes, the banded aggregation and sparse distribution characteristics of space debris environment in orbital altitude-inclination phase space are characterized; Figure 4 By superimposing eccentricity statistical features in the orbital height-inclination phase space, the orbital shape dispersion of debris groups in different grids is characterized, reflecting the dynamic evolution state of each orbital region. Figure 5 The RAAN (right ascension of ascending node) entropy features are superimposed in the orbital altitude-inclination phase space to quantitatively describe the distribution uniformity of RAAN of the debris population in each grid. Figure 6 By superimposing perigee argument entropy in the orbital altitude-inclination phase space, the uniformity of perigee argument distribution of debris groups within each grid is quantitatively described.

[0064] Step S3: Based on the multidimensional representation and each set of evolution parameters corresponding to the multidimensional representation. Each sample is paired up to construct training samples, forming a complete labeled training dataset.

[0065] Specifically, the input of the training samples Includes: Initial multidimensional characterization of space debris environments Evolution parameters and prediction time interval Output Multidimensional representation of the spatial fragmentation environment at the target time This forms a complete labeled training dataset. The output... As input The truth labels, which correspond one-to-one, together constitute supervised training sample pairs.

[0066] The number of training samples is increased by a one-to-one pairing process. Specifically, any two states at any time within the same evolution parameter scenario are paired one-to-one. The spatial debris environment state, evolution parameters, and time interval at each time are used as inputs, and the spatial debris environment state at the corresponding target time is used as the output. Multiple sets of training samples are extracted, thereby effectively expanding the size of the training dataset without increasing the number of simulation runs.

[0067] Step S4: Using the labeled training dataset described in step S3, supervised training is performed on the multi-frequency perception fusion neural operator network model to obtain the trained multi-frequency perception fusion neural operator network model (network proxy model).

[0068] Specifically, the multi-frequency sensing fusion neural operator network model takes as input an initial multidimensional representation of the space debris environment, evolution parameters, and a prediction time interval, and outputs a multidimensional representation of the space debris environment at the target time. The multi-frequency sensing fusion neural operator network model includes a conditional coding module, a frequency domain segmentation module, a cross-frequency interaction module, and a residual hierarchical fusion module connected in sequence. Compared to traditional neural operators that primarily rely on low-frequency components for function mapping learning, the multi-frequency sensing fusion neural operator network model achieves multi-frequency band perception of the evolutionary patterns of the space debris environment by explicitly segmenting and interactively fusing features from multiple frequency bands.

[0069] The conditional encoding module jointly encodes the input state tensor, scene evolution parameters, and prediction time interval into a high-dimensional hidden representation.

[0070] The frequency domain segmentation module applies a two-dimensional Fourier transform to the high-dimensional hidden representation, decomposing the complete spectrum into multiple sub-frequency bands such as low frequency, mid frequency, and high frequency;

[0071] The cross-frequency interaction module performs full-channel spectrum convolution on each sub-frequency band and then models the coupling relationship between different sub-frequency bands through a frequency band mixing matrix.

[0072] The residual hierarchical fusion module uses low-frequency components as a basis, and successively superimposes mid-frequency and high-frequency components. After recovering to the spatial domain through inverse Fourier transform, the prediction result is output in residual form. The prediction result is a multidimensional representation of the spatial debris environment at the target time. .

[0073] Specifically, in the supervised training process described in step S4, the training dataset is adjusted according to the evolution parameters. The model was divided into training, validation, and testing subsets. The training loss function used was MSE (mean squared error). Model performance was quantitatively evaluated, including target quantity prediction performance, overall distribution similarity evaluation, and high-density region classification performance. Specifically, target quantity prediction performance was evaluated using MAE (mean absolute error) and SMAPE (symmetric mean absolute percentage error), measuring the absolute and relative errors in spatial target quantity prediction, respectively. Overall distribution similarity evaluation used SSIM (structural similarity index), measuring the structural similarity between the predicted and true distributions. High-density region classification performance was evaluated using Recall and F1 score, measuring the accuracy and overall performance of the multi-frequency sensing fusion neural operator network model in identifying high-density regions of interest.

[0074] Step S5: Based on the multi-frequency sensing fusion neural operator network model trained in step S4, perform evolution prediction on the space debris environment under different initial states and different evolution parameters, and output a multi-dimensional representation of the space debris environment at the target time.

[0075] Specifically, as shown in Table 1, the method proposed in this invention has achieved good performance on the test set in terms of MAE, SMAPE, SSIM, Recall and F1 score, which verifies the effectiveness of the model in terms of target quantity prediction accuracy, overall distribution structure similarity and high-density region identification ability.

[0076] Table 1 shows the comprehensive performance evaluation results of the method of the present invention on the test set.

[0077]

[0078] like Figure 7 As shown in (a)-(e), Figure 7 (a)–(e) show the performance of MAE, SMAPE, SSIM, Recall, and F1 score, respectively. The horizontal axis represents the prediction time interval, and the vertical axis represents the numerical performance of each evaluation index. As the prediction time interval increases, although the proposed method shows a decrease in performance across all indices, it still maintains high prediction performance, indicating that the model has good long-term evolution prediction capabilities.

[0079] This invention provides a deep learning-based method for predicting the evolution of space debris environments, enabling refined characterization and efficient evolution prediction of these environments. Accurate and efficient space debris environment evolution prediction methods have significant guiding implications for formulating space debris mitigation strategies, planning proactive removal tasks, and constructing space situational awareness systems. Considering the scarcity of measured data on space debris environments, the core objective of this invention is to achieve effective characterization of the space debris environment based on readily available features, and to achieve efficient and accurate long-term evolution prediction under arbitrary evolution parameters, relying only on a single-time initial state. This provides a feasible solution for deep learning methods in space debris environment evolution tasks.

[0080] 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 evolution of spatial fragmentation environments based on deep learning, characterized in that, The process includes the following: Step S1: Construct the original dataset of the space debris environment state evolving year by year under different evolution parameter settings; Step S2: Based on the original dataset, construct a multidimensional representation describing the evolutionary state of the space debris environment; Step S2 includes the following steps S201 to S203: Step S201: Iterate through each set of evolutionary parameters in the original dataset. and evolution parameters Construct a basic statistical data set based on the corresponding annual spatial debris environment status. Among them, basic statistical data sets At least including: orbital altitude set Orbital Inclination Set Total target quantity Number of controlled satellites Number of failed satellites Number of fragments Number of rocket bodies Eccentricity set set of ascending nodes and right ascension and perigee argument set ; Step S202: Establish a track height and orbital inclination The gridded phase space structure is used as the coordinate axis, and the grid cells in the gridded phase space structure are assigned values ​​in the form of target number density distribution; The gridded phase space structure consists of multiple grid cells; orbital height exist Within the interval Step size discretization One bin, orbital inclination exist Within the range Step size discretization Each bin constitutes Meshized phase space; Step S203: For each grid cell in step S202, calculate and associate the feature vector extracted from the basic statistical feature set to construct a multidimensional representation for describing the space debris environment; The mathematical representation of the feature vector is as follows: Features include the logarithmic form of the target quantity. and dynamic dimension features; the dynamic dimension features include the median eccentricity. Eccentricity interquartile range Right ascension entropy of ascending node and perigee angle entropy ; The multidimensional representation is composed of a gridded phase space structure and the feature vectors associated with each grid cell. The gridded phase space structure is constructed by step S202, and the feature vectors are extracted by step S203 and associated with the corresponding grid cells. Step S3: Based on the multidimensional representation and each set of evolution parameters corresponding to the multidimensional representation. Each sample is paired up to construct training samples, forming a complete labeled training dataset. The input of the training samples Includes: Initial multidimensional characterization of space debris environments Evolution parameters and prediction time interval Output Multidimensional representation of the spatial fragmentation environment at the target time This forms a complete labeled training dataset; Step S4: Using the labeled training dataset described in step S3, supervised training is performed on the multi-frequency sensing fusion neural operator network model to obtain the trained multi-frequency sensing fusion neural operator network model. The input of the multi-frequency sensing fusion neural operator network model is the initial multidimensional representation of the space debris environment, evolution parameters, and prediction time interval, and the output is the multidimensional representation of the space debris environment at the target time. Step S5: Based on the multi-frequency sensing fusion neural operator network model trained in step S4, perform evolution prediction on the space debris environment under different initial states and different evolution parameters, and output a multi-dimensional representation of the space debris environment at the target time.

2. The method for predicting the evolution of spatial fragmentation environment based on deep learning according to claim 1, characterized in that, Step S1 includes the following steps S101 to S102: Step S101: Define the initial space debris environment state and evolution parameter space. The space debris environment state includes at least the number of all space targets in the current space debris environment. and orbital elements for each space target, the orbital elements including orbital inclination. semi-major shaft eccentricity Right ascension of ascending node and perigee argument ; The evolution parameter space The evolution parameter space is used to define a reasonable range for multi-scenario evolution scenarios. The evolution parameters include at least the initial target quantity ratio. Emission rate Post-task handling success rate and the failure rate of collision avoidance maneuvers ; Step S102: For each set of evolutionary parameters in the evolutionary parameter space Starting from the initial state of the space debris environment, an annual simulation is performed to obtain the state of the space debris environment of all space targets involved in the simulation in each year, forming the original dataset.

3. The spatial fragmentation environment evolution prediction method based on deep learning according to claim 2, characterized in that, The multi-frequency sensing fusion neural operator network model includes a conditional coding module, a frequency domain segmentation module, a cross-frequency interaction module, and a residual hierarchical fusion module connected in sequence. The conditional encoding module jointly encodes the input state tensor, scene evolution parameters, and prediction time interval into a high-dimensional hidden representation. The frequency domain segmentation module applies a two-dimensional Fourier transform to the high-dimensional hidden representation, decomposing the complete spectrum into multiple sub-frequency bands of low frequency, mid frequency, and high frequency. The cross-frequency interaction module performs full-channel spectrum convolution on each sub-frequency band and then models the coupling relationship between different sub-frequency bands through a frequency band mixing matrix. The residual hierarchical fusion module uses low-frequency components as a basis, and successively superimposes mid-frequency and high-frequency components. After recovering to the spatial domain through inverse Fourier transform, the prediction result is output in residual form. The prediction result is a multidimensional representation of the spatial debris environment at the target time. .

Images