Earth-moon L1 point track crossing behavior prediction method and system based on machine learning

By constructing an orbital cluster sample dataset and using the XGBoost model, the problem that traditional orbit prediction methods are unable to accurately determine the behavior of non-cooperative spacecraft was solved, achieving high-precision prediction of the orbital behavior of spacecraft at the Earth-Moon L1 point and improving the reliability of space situational awareness.

CN121766145APending Publication Date: 2026-03-31NORTHWESTERN POLYTECHNICAL UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610176442.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-02-06
Publication Date
2026-03-31

AI Technical Summary

Technical Problem

Traditional orbit prediction methods are unable to accurately determine the motion state and behavioral intentions of non-cooperative spacecraft after crossing the Earth-Moon L1 point, leading to difficulties in space situational awareness and safety management.

Method used

By constructing an orbital cluster sample dataset and employing the XGBoost model, an initial state set is generated by combining the Jacobian constant, velocity direction angle, and y-axis coordinate. Spacecraft behavior categories are defined, and feature preprocessing and adaptive threshold optimization are performed to achieve high-precision orbital behavior prediction.

Benefits of technology

It enables efficient and reliable prediction of the complex and uncertain behavior of spacecraft after crossing the L1 point, thereby enhancing space situational awareness capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121766145A_ABST
    Figure CN121766145A_ABST
Patent Text Reader

Abstract

The invention discloses an earth-moon L1 point track crossing behavior prediction method and system based on machine learning. According to the method, firstly, an initial state set is generated based on an earth-moon plane circular restrictive three-body model, and an orbit sample data set is constructed through numerical integration; secondly, a multi-class behavior classification system is established, and sample data is automatically labeled; on the basis, an XGBoost algorithm is adopted for model training, a category self-adaptive threshold optimization strategy based on a verification set is introduced for the problem of data extreme imbalance, and the overall prediction precision and category fairness of the model are remarkably improved by independently optimizing an optimal decision threshold for each category and applying the optimal decision threshold to normalized probabilistic decision. According to the method, the technical blank of spacecraft behavior prediction in the earth-moon translation point area is filled, and an effective means is provided for space situation awareness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of aerospace technology, specifically to a method and system for predicting the behavior of the Earth-Moon L1 point crossing orbit based on machine learning. Background Technology

[0002] Lagrange of the Earth-Moon system L Point 1, located at a unique position where the Earth and Moon are in gravitational equilibrium, is a crucial node for spacecraft to achieve low-energy transfer between the Earth and the Moon. However, this Lagrange point... L The dynamic environment near point 1 exhibits significant nonlinear and chaotic characteristics. The spacecraft's trajectory is extremely sensitive to the initial state at the time of crossing; even minute differences in initial conditions can fundamentally alter the subsequent evolution path. This dynamic characteristic makes crossing... L Spacecraft at point 1 may exhibit a variety of different behavioral patterns, such as returning directly to Earth's vicinity, being captured by the Moon's gravity, impacting the Moon, or... L Its behavior, such as hovering around 1 o'clock, is highly uncertain and unpredictable. Especially for non-cooperative spacecraft whose mission intentions are unknown, traditional orbit prediction methods struggle to predict their passage. L Reliably judging the movement status and behavioral intentions after point 1 poses a significant challenge to spatial situational awareness and security control. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a machine learning-based Earth-Moon mapping method. A method and system for predicting point crossing trajectory behavior, by constructing a large-scale... By traversing a sample dataset of orbital data points and training the model, a model with high-precision prediction capabilities is ultimately formed, enabling accurate prediction of traversing spacecraft trajectories.

[0004] This invention is achieved through the following technical solution: A machine learning-based method for predicting the Earth-Moon L1 point crossing orbit behavior includes the following steps: Step 1: Based on different energy levels By traversing the initial state set of spacecraft points, an orbital cluster sample dataset is constructed. Define spacecraft orbital behavior categories and classification standards, and establish a spacecraft behavior label classification system; Step 2: Based on the spacecraft behavior label classification system, generate corresponding behavior labels for each sample in the orbital cluster sample dataset to form a labeled orbital sample dataset; Step 3: Divide the labeled track sample dataset into a training set and a validation set. Use the training set to train the XGBoost model to obtain an initial prediction model. Based on the validation set, perform adaptive threshold optimization on each behavior category to determine the optimal probability threshold for each category to form an optimal threshold vector. Use the trained model to obtain the original probability output. Normalize the original probability according to the optimal threshold vector and determine the final behavior prediction category based on the normalized score.

[0005] Preferably, in step 1, the The initial state set of the spacecraft traversed by the point was obtained in the following way: Based on the Earth-Moon planar circular restricted three-body model, the spacecraft traverses the Earth and Moon. Jacobian constant at point Poincaré section Velocity direction angle and y-axis coordinates As three independent variables characterizing its state; The Jacobian constant, velocity direction angle, and y-axis coordinate are discretely sampled within their respective physical ranges to generate the initial state set.

[0006] Preferably, the physical range of the Jacobian constant is determined in the following way: Solve The point Poincaré section contains a Jacobian constant that intersects the zero-velocity curve corresponding to that Jacobian constant. value range .

[0007] Preferably, for each given Jacobian constant The value, and its corresponding y-axis coordinate. Feasible range By solving for the Jacobian constant The intersection of the zero velocity curve at the specified value and the Poincaré section is obtained.

[0008] Preferably, the spacecraft behavior label classification system defined in step 1 includes the following categories: Direct approach to Earth, lunar orbit followed by approach to Earth, direct impact on the Moon, lunar orbit followed by impact on the Moon, short-term lunar capture and L1 binding orbit; The lunar orbiting flight followed by approaching Earth, lunar orbiting flight followed by impacting the Moon, and short-term lunar capture behavior are further subdivided according to the lunar orbiting direction and the number of lunar orbits.

[0009] Preferably, the subdivision based on the direction of lunar orbit and the number of lunar orbits includes: orbiting the moon once clockwise, orbiting the moon once counterclockwise, orbiting the moon multiple times clockwise, orbiting the moon multiple times counterclockwise, and a mixture of lunar orbits.

[0010] Preferably, before dividing the labeled track sample dataset into training and validation sets, a feature preprocessing step is also included: The velocity direction angle Convert to trigonometric function form ; For those containing Jacobi constant y-axis coordinates And perform polynomial interactive feature expansion on the eigenvectors of the trigonometric functions; The expanded features are filtered by variance thresholding to obtain effective feature vectors for model training.

[0011] Preferably, step 3 further includes a category weight calculation step: Calculate the weight of each behavior category based on the number of samples in the training set. ,in It is inversely proportional to the sample size; When training the XGBoost model, the class weights are applied to the loss function to mitigate the class imbalance problem.

[0012] Preferably, the adaptive threshold optimization for each behavior category based on the validation set includes: For each behavior category Iterate through multiple candidate thresholds within a preset probability interval. ; For each candidate threshold Calculate the accuracy of this category. and recall rate And calculate the F1 score based on this. ; Choose to The largest candidate threshold is taken as the optimal probability threshold for that category. .

[0013] A machine learning-based system for predicting the Earth-Moon L1 point crossing orbit behavior includes: The acquisition module is used to collect data based on different energy levels. By traversing the initial state set of spacecraft points, an orbital cluster sample dataset is constructed. The classification criteria module is used to define spacecraft orbital behavior categories and classification criteria, and to establish a spacecraft behavior label classification system. The classification module is used to generate corresponding behavior labels for each sample in the orbital cluster sample dataset according to the spacecraft behavior label classification system, forming a labeled orbital sample dataset; The training module is used to divide the labeled track sample dataset into a training set and a validation set. The XGBoost model is trained using the training set to obtain an initial prediction model. Based on the validation set, adaptive threshold optimization is performed on each behavior category to determine the optimal probability threshold for each category to form an optimal threshold vector. The original probability output is obtained using the trained model. The original probability is normalized according to the optimal threshold vector, and the final behavior prediction category is determined based on the normalized score.

[0014] Compared with the prior art, the present invention has the following beneficial technical effects: This application provides a machine learning-based Earth-Moon... This method for predicting orbital behavior at the Earth-Moon L1 crossing point transforms complex orbital dynamics problems into a standard machine learning classification task. It precisely parameterizes the crossing state at the Earth-Moon L1 point, systematically generating a set of initial conditions covering all physically feasible states using three variables: the Jacobian constant, velocity direction angle, and y-axis coordinate. Subsequently, numerical integration evolves these initial conditions into orbital clusters containing rich dynamic behaviors, and a hierarchical behavior classification system is established based on explicit physical rules, providing high-quality labeled data for machine learning. In the model construction phase, to address the extreme imbalance of the dataset, strategies such as feature enhancement, weight compensation, and adaptive threshold optimization are employed to effectively improve the XGBoost model's ability to identify various orbital behaviors. This method integrates orbital dynamics theory with advanced machine learning techniques, achieving efficient and reliable prediction of the complex and uncertain behaviors of spacecraft after crossing the L1 point.

[0015] This application also proposes a machine learning-based Earth-Moon... A point-crossing orbit behavior prediction system, an electronic device, and a computer storage medium, comprising the aforementioned machine learning-based Earth-Moon... The full advantages of point crossing track behavior prediction methods. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 For the present invention Definition of spacecraft velocity direction angle on the Poincaré section; Figure 2 This is a flowchart of the spacecraft behavior tag classification process of the present invention; Figure 3 ZVC at different energy levels in Embodiment 2 of the present invention and Point Poincaré section view; Figure 4 This is the orbital family sample dataset of Embodiment 2 of the present invention; Figure 5 This is a diagram showing the classification results of spacecraft orbital behavior in Embodiment 2 of the present invention; Figure 6 This is a scatter plot of spacecraft initial state-orbit behavior labels from Embodiment 2 of the present invention. Figure 7 This is a confusion matrix diagram of Embodiment 2 of the present invention.

[0018] Figure 8 This is a flowchart of the method for predicting the Earth-Moon L1 point crossing orbit behavior based on machine learning, as described in this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] See Figure 8 A machine learning-based method for predicting the Earth-Moon L1 point crossing orbit behavior includes the following steps: Step 1: Obtain the set of initial states of the L1 point traversing the spacecraft at different energy levels; specifically including: Based on the Earth-Moon plane circular restricted three-body model, three independent variables characterizing the spacecraft's crossing state at the Poincaré section at point L1 are identified: Jacobian constant. Velocity direction angle and y-axis coordinates ; The Jacobian constant is determined based on the aforementioned circular restricted three-body model of the Earth-Moon plane. The range of values ​​for is determined, and the relationship between and each is solved. The y-axis coordinate corresponding to the value The feasible range; The Jacobian constant Within the range of values, each The y-axis coordinate corresponding to the value Within the feasible range and the velocity direction angle The initial state set is generated by performing grid discrete sampling on a global scale, covering different energy levels and different crossing locations; By constructing a parameterized sampling framework based on a circularly constrained three-body model of the Earth-Moon plane, an initial state set covering different energy levels and spatial positions is generated. This method uses a three-dimensional parameterized representation—characterizing energy level with the Jacobian constant, describing flight direction with the velocity direction angle, and determining spatial position with the y-axis coordinate—to discretize the continuous orbital state space into a tractable parametric grid. The physical feasible region of the parameters is determined using the intersection condition of the zero-velocity curve and the Poincaré section in the three-body model, and complete coverage of the orbital state space is ensured through equally spaced grid sampling.

[0022] Step 2: Based on the initial state set, construct the orbital cluster sample dataset using the numerical integration method; Based on the dynamic equations of the Earth-Moon plane circular restricted three-body model, an adaptive variable-step-size integrator is used to propagate the orbit for each initial state. By setting the integration duration and physical termination conditions, the natural evolution of the spacecraft under non-maneuvering conditions is simulated. Through numerical integration, static initial parameters are mapped to dynamic orbital trajectories, capturing complex nonlinear dynamic characteristics near the L1 point, including chaotic behavior and orbital bifurcation. The generated orbital cluster dataset fully records the spacecraft's evolution from the crossing section to its final state.

[0023] Step 3: Define spacecraft orbital behavior categories and classification criteria, and establish a spacecraft behavior label classification system; The orbital behavior categories include direct approach to Earth, lunar orbit followed by approach to Earth, direct impact to the Moon, lunar orbit followed by impact to the Moon, short-term lunar capture, and L1 tethered orbit. By defining specific behavior determination rules, continuous orbital trajectories are transformed into discrete behavior categories. Based on the orbital endpoint state, geometric features, and relative positional relationships, this system constructs a classification framework encompassing multiple categories, including direct approach to Earth, lunar orbit followed by approach to Earth, direct lunar impact, lunar orbit followed by lunar impact, short-term lunar capture, and L1 bound orbits. The classification logic employs a hierarchical decision-making process, first performing a coarse classification based on whether the orbit approaches Earth or impacts the Moon, and then a finer classification based on lunar orbital characteristics. This transformation of complex orbital dynamics into standardized category labels provides supervisory signals for model learning, enabling the model to learn the mapping relationship from initial state to behavior category.

[0024] Step 4: Based on the spacecraft behavior label classification system, generate corresponding behavior labels for each sample in the orbital cluster sample dataset to form a labeled orbital sample dataset; Unlabeled orbital physical data is transformed into training samples with semantic labels. The labeling process judges each sample step by step according to a defined hierarchical classification standard, ensuring that each sample is accurately assigned to the corresponding behavior category. This gives the original orbital data the input-output correspondence required for supervised learning, constructing a structured dataset containing initial state parameters and final behavior labels.

[0025] Step 5: Using the spacecraft's initial state and corresponding behavior labels as training samples, construct and train a spacecraft behavior prediction model based on the XGBoost algorithm; the training process includes: S5.1: Preprocessing and enhancing the initial state features, including velocity direction angles... The function is converted into trigonometric function form and then subjected to polynomial feature expansion and filtering to obtain effective feature vectors. The labeled track sample dataset is divided into training set, validation set and test set, and class weights are calculated based on the number of samples in each class to alleviate the data imbalance problem. The XGBoost model is fitted and trained using the training set to obtain an initial model for predicting spacecraft orbital behavior; based on the validation set, adaptive threshold optimization is performed on each behavior category to determine the optimal probability threshold for each category and form an optimal threshold vector. During the prediction phase, the original probabilities output by the model are normalized according to the optimal threshold vector, and the final behavior prediction category is determined based on the normalized score.

[0026] Based on the processed orbital dataset, a behavior prediction model based on the XGBoost algorithm was constructed and optimized. By preprocessing and enhancing the initial state features, the model can effectively extract the nonlinear relationships between parameters. To address the extreme class imbalance problem in the dataset, a class weight compensation and adaptive threshold optimization strategy was adopted, significantly improving the model's ability to identify minority class behaviors. During training, a base model was first fitted using the training set, and then the decision thresholds for each class were optimized on the validation set, ultimately forming a prediction model with optimal performance. This training method improves the traditional probability maximization decision-making into a normalized decision-making based on the optimal threshold, enabling the model to achieve a better performance balance among different classes. The trained model can accurately predict the subsequent orbital behavior of a spacecraft based on its initial state when crossing the L1 point.

[0027] Example 1 A machine learning-based method for predicting the Earth-Moon L1 point crossing orbit behavior: Step 1: Obtain energy levels at different energy levels The initial set of states of a point traversing a spacecraft; Specifically, it includes the following processes: S1.1 Based on the Earth-Moon plane circular restricted three-body model, its kinematic equations are as follows:

[0028] In this model, the spacecraft is crossing the Earth and the Moon. Poincaré section of a point ( The traversal state on a given surface can be fully characterized by the following three independent variables: 1. Jacobi constant The integral constant derived from the system's dynamic characteristics has the following mathematical expression:

[0029] in, It is a pseudo-potential function. This represents the speed of the spacecraft. Physically, it represents the total mechanical energy of a spacecraft in a rotating coordinate system, and its value uniquely determines the boundary of the reachable region of the spacecraft's motion (i.e., the zero velocity curve, ZVC).

[0030] Different The values ​​correspond to different energy levels.

[0031] 2. Velocity direction angle Defined as the angle between the spacecraft's velocity vector and the positive x-axis of the rotating coordinate system. Its value range is... This parameter visually reflects the flight direction of the spacecraft when it crosses the cross section.

[0032] 3. y-axis coordinate : The specific position of the spacecraft on the y-axis of the rotating coordinate system as it traverses the Poincaré section. This parameter is related to... Together they determined the geometric location where the crossing occurred.

[0033] Once determined Triples can be passed The magnitude of the velocity can be obtained by inverse solving the definition. and combined The velocity components are calculated to obtain a unique and complete four-dimensional state (position and velocity) of the spacecraft at the moment of transit.

[0034] S1.2 Determine the Jacobian constant Velocity direction angle The range of values ​​for the y-axis coordinate.

[0035] 1. Jacobi constant Determining the global range: its effective interval Two conditions must be met: (a) Point Poincaré section ( It exists on its own.

[0036] (b) For a given Its corresponding ZVC is There is an opening near the point, forming a corridor that spacecraft can pass through.

[0037] By solving the system consisting of the ZVC equations and the cross-section equations, the extent of this energy boundary can be numerically determined (e.g., for the Earth-Moon system, the typical extent is approximately...). ).

[0038] 2. y-axis coordinate Determining the dynamic range: for each given... The value, its corresponding The feasible range of coordinates is not fixed, but is obtained by solving the simultaneous equations of the ZVC equation and the section equation:

[0039] Coordinate range .

[0040] 3. Velocity direction angle Range: The definition of velocity direction angle is as follows Figure 1 As shown, it theoretically covers all possible directions, so a global scope is taken. .

[0041] S1.3, Based on the preset number of discrete points , and The three-dimensional parameter space is discretized by grid sampling to generate spacecraft covering different energy levels and different crossing positions.

[0042] Step 2: Based on the initial state set, construct the orbital cluster sample dataset using numerical integration methods; Specifically, the Earth-Moon Plane Circular Restricted Three-Body Model (PCRTBP) is used as the physical basis for orbital integration. Starting from the initial state of each spacecraft in the initial state set, numerical integration is performed using the ode45 integrator based on the dynamic equations of the Earth-Moon system, and finally, an orbital cluster sample dataset is generated.

[0043] Step 3: Construct behavioral labels and classification standards for spacecraft orbital behavior, and then construct a spacecraft behavior label classification system; S3.1. Behavioral labels for spacecraft orbital behavior include approaching Earth behavior, lunar impact behavior, lunar orbiting behavior, short-term lunar capture behavior, and... For tethered orbits, the classification criteria for each behavior label are determined based on the principles of orbital dynamics as follows: 1) Define approaching Earth behavior: based on the spacecraft's last pass through Earth. The velocity direction angle at the Poincaré section is used as the criterion, i.e., it satisfies... ; 2) Definition of lunar impact behavior: When the distance between the spacecraft's center and the moon's center is less than or equal to the average radius of the moon, i.e. At that point, the integrator stopped calculating and determined that the spacecraft had crashed into the moon; 3) Lunar orbit behavior: If, during the evolution of the orbit, the distance to the lunar perihelion is less than the set radius of the lunar Hill sphere, the spacecraft is determined to have engaged in lunar orbit behavior; In this embodiment, the distance from the lunar center to the spacecraft's perigee is measured in a rotating coordinate system. Less than 0.6 times the radius of the lunar hill sphere ,Right now Based on this, lunar orbits are further divided into 5 categories: one clockwise lunar orbit, one counterclockwise lunar orbit, multiple clockwise lunar orbits, multiple counterclockwise lunar orbits, and mixed lunar orbits (which include both clockwise and counterclockwise lunar orbits). 4) Define short-term lunar capture behavior: This type of orbit does not approach Earth or collide with the Moon, but only exhibits lunar orbit behavior. Based on the relative relationship between the orbital terminus and the lunar hill, it can be further subdivided into Type I and Type II. 5) Definition Constrained orbits, these orbits do not exhibit the aforementioned three behaviors, and are typically in... The point operates in a narrow bottleneck area for a long time.

[0044] S3.2 Classification criteria for spacecraft orbital behavior: Constructing a spacecraft behavior label classification system.

[0045] Based on the above criteria, 23 types of spacecraft orbital behaviors were defined, combined with, for example... Figure 2 The classification process shown ultimately constructed a complete spacecraft behavior label classification system, and the correspondence between label numbers and specific behaviors is shown in Table 1.

[0046] Table 1. Spacecraft Orbital Behavior and Labels

[0047] Step 4: Generate a behavior label for each sample in the orbital cluster sample dataset according to the spacecraft behavior label classification system.

[0048] Based on the behavioral label classification system established in step 3, automated classification and labeling are performed on the orbital cluster sample dataset generated in step 2. The specific process is as follows: First, based on the two core terminating events of "whether it approaches Earth" and "whether it impacts the Moon," all samples are roughly divided into three categories. Then, within each coarse category, further distinctions are made based on refined rules such as "whether it orbits the Moon," "orbiting direction and number of orbits," and "positional relationship between the orbital terminus and the lunar Hill sphere." Finally, each sample is assigned to a specific category among the 23 behavioral categories defined in step 3, and a corresponding integer behavioral label is generated for it. Thus, an orbital sample dataset with clear behavioral labels is formed for subsequent machine learning model training.

[0049] In this embodiment, the orbital clusters are first divided into three basic categories: those close to Earth, those that collide with the Moon, and those that do not collide with Earth or the Moon, thus completing the initial screening.

[0050] Then, five types of lunar orbiting flight conditions are superimposed for a second assessment to form detailed labels: For the category of approaching Earth, it is divided into two subcategories: those that approach Earth directly (without lunar orbiting behavior) and five categories that approach Earth after orbiting the Moon, totaling six subcategories. The lunar collision category is further divided into direct lunar collisions (without lunar orbiting) and five categories of lunar orbiting flights followed by lunar collisions, totaling six subcategories.

[0051] For those that do not approach Earth and will not collide with the Moon, the first distinction is made based on their lunar orbit conditions. Traction track type; The remaining samples are further subdivided into 5 types of short-term lunar capture type I and 5 types of short-term lunar capture type II based on the terminal position status (whether it is inside the lunar hill sphere). In summary, the classification system defines a total of 23 types of orbital behavior.

[0052] Step 5: Using the initial state of the spacecraft and the corresponding behavior label results as training samples, train the XGBoost-based multi-classification model to obtain the spacecraft behavior prediction model. The training method is as follows: S5.1. The original three-dimensional features of the spacecraft's initial state are converted into four-dimensional intermediate features, and then subjected to polynomial interactive feature expansion and variance thresholding to obtain effective feature vectors.

[0053] The velocity direction angle in the initial state of the spacecraft By converting the original 3D features into a pair of orthogonal features using trigonometric functions, the original 3D features are transformed into a pair of orthogonal features. Converted into four-dimensional intermediate features, as follows:

[0054] The aforementioned four-dimensional intermediate features are extended using a second-order polynomial to generate an extended feature set that includes the original features, squared terms, and pairwise interaction terms, in order to potentially capture complex synergistic effects between features.

[0055] The variance thresholding method is used to filter the expanded high-dimensional features, removing those with variances below a preset threshold (e.g., ...). The features of the sample are analyzed, retaining only the effective features that contribute to distinguishing the samples, to form the final effective feature vector. ,as follows:

[0056] The above operations transform the original three-dimensional state parameters containing periodic variables into feature vectors that are more suitable for machine learning models and can represent nonlinear relationships.

[0057] S5.2. Based on the behavior labels of each sample in the training set, classify all samples to form samples of multiple track behavior categories. Encode the track behavior category labels with integers, and then calculate the category weight of each track behavior category based on the frequency distribution of each track behavior category. The expression is as follows:

[0058] in, The total number of samples, For the first l Number of samples for each behavior category This represents the total number of behavior categories.

[0059] The category weights are input as sample weights during model training. This allows the model to assign greater importance to categories with fewer samples during loss function optimization, thereby improving the classification accuracy of rare category samples, effectively mitigating the model's bias towards the majority category, and optimizing the overall classification performance of the model.

[0060] In this implementation, the labeled dataset is divided into training, validation, and test sets according to a preset ratio (e.g., 72:8:20). Stratified sampling is used during the partitioning to ensure that the proportion of each category in each subset is consistent with the whole set, preventing rare categories from being missing from the key set.

[0061] S5.3. Input the training set obtained in S5.2 into the XGBoost model for fitting and training to obtain the initial model for predicting spacecraft orbital behavior; The XGBoost (Extreme Gradient Boosting) algorithm is chosen as the base model. Mathematically, XGBoost predicts samples by additively combining multiple regression trees:

[0062] in, This represents the k-th regression tree. Let the function space of all trees be defined. The algorithm is centered around an objective function with regularization:

[0063] Among them, the first item This is the loss function, used to measure the predicted value. With target value The difference between them; the second item This represents the model complexity penalty term, used to control the number of leaf nodes T and the leaf weights. The size of the model should be adjusted to prevent overfitting. and This is the regularization coefficient.

[0064] S5.4 Based on the initial model for spacecraft orbital behavior prediction, predict the validation set to obtain the behavior categories for each sample. l For each category Within a reasonable range (e.g., 0.1 to 0.9), different probability thresholds are iterated. Calculate the category Accuracy at this threshold With recall rate Then calculate the F1 score. According to the score The maximum value determines the optimal threshold for the corresponding category, thus obtaining the final spacecraft orbital behavior prediction model.

[0065] This accuracy With recall rate The calculation method is as follows:

[0066] Based on accuracy With recall rate The F1 score is calculated as follows:

[0067] S5.5, Select the option that makes F1 score Reaching the maximum threshold , as the optimal threshold for this category By iterating through all categories, a set of optimal threshold vectors for each category is obtained.

[0068] The expression for the optimal threshold is as follows: ; The expression for the optimal threshold vector is as follows: .

[0069] S5.6. Normalize the original predicted probability of each category based on the optimal threshold vector, and determine the final behavior prediction category based on the normalized probability score.

[0070] The normalization method is as follows:

[0071] in, This is to prevent the minimum value where the denominator is zero.

[0072] In the final deployment and application phase of the model, the optimal threshold vector learned above is integrated into the decision-making process. For new input samples, the model outputs the original probability vector. Subsequently, instead of directly using the argmax rule, the normalized confidence score for each category is calculated. (in To prevent the possibility of dividing by zero (a very small amount), the final prediction result is: This ensures that predictions for each category are made near their respective optimal decision boundaries.

[0073] Currently, the mainstream metrics used for evaluating the performance of multi-class classification models in machine learning include Precision, Recall, F1-Score, G-Mean (Geometric Mean), MAUC (Multi-class AUC), and MMCC (Multi-class Matthews Correlation Coefficient). In multi-class tasks, the traditional XGBoost model outputs a probability vector for each sample belonging to each class. During prediction, the maximum probability principle is typically used. However, this uniform threshold strategy has significant drawbacks on imbalanced datasets: firstly, the predicted probability distribution of minority class samples is often generally low, and even if correctly classified, their probability value may be lower than that of the majority class; secondly, the confidence distributions of predicted probabilities differ significantly between different classes, and a single threshold cannot simultaneously consider both precision and recall for each class, especially leading to a significant drop in recall for the minority class. This embodiment proposes a class-adaptive threshold optimization strategy based on the validation set. This strategy abandons the uniform decision boundary and instead targets each behavior class... Independently, it seeks an optimal probability threshold on the predicted probability distribution of its validation set. In order to maximize the classification efficiency of this category.

[0074] The proposed category-adaptive threshold optimization strategy achieves targeted control of classification performance by independently and meticulously optimizing the optimal threshold for each behavior category. This strategy effectively mitigates the negative impact of extreme class imbalance by setting a relatively low decision threshold for the minority class, significantly improving its recall rate and thus enhancing the detection capability of rare orbital behaviors. Ultimately, while ensuring stable performance for the majority class, this strategy comprehensively improves the model's performance on macro-level evaluation metrics that emphasize class balance, such as G-Mean and MMCC. This strategy successfully solves the decision bias problem of traditional methods in extremely imbalanced multi-class classification tasks, enabling the final orbital behavior prediction model to maintain high overall accuracy while possessing excellent class fairness and a keen ability to identify key minority behaviors.

[0075] Example 2 Based on the Earth-Moon circular restricted three-body model, a machine learning-based method for predicting the orbital behavior of low-energy spacecraft crossing the Earth-Moon L1 point is proposed. A specific example is introduced below, and the basic parameters used in this example are shown in Figure 2.

[0076] Table 2 Basic parameters of the dynamic model

[0077] Considering Crossing the orbit, therefore the Poincaré section must exist ( And it intersects with ZVC ( ) Therefore, this study sets the range of values ​​for the spacecraft Jacobian constant to be... ZVC within this energy range and its corresponding The cross section of Poincaré is shown in Figure 3.

[0078] The number of discrete values ​​for the three-dimensional input variables is set to Furthermore, since this method only focuses on the spacecraft's movement through... The natural orbital behavior under no-maneuver conditions after the Poincaré section is analyzed. If the integration time is too long, the probability of the spacecraft performing orbital maneuvers will increase significantly, thus reducing the reliability of the prediction results. Therefore, in this example, the integration time is set to... That is, half of the synodic period of the Earth-Moon system. Figure 4 The orbital cluster sample dataset obtained at CJ = 3.03 is presented.

[0079] Taking the Jacobian constant CJ = 3.03 as an example, the orbital behavior classification results and the label scatter plot are as follows: Figure 5 and Figure 6 As shown in these two pictures, it can be seen that Crossing orbits can be well distinguished using the classification strategy proposed in this invention, and the boundaries between different types are relatively clear. By traversing all discrete energy levels, the classification and labeling of all orbital cluster samples are completed. The number and proportion of samples in each category are shown in Table 3, totaling 17 categories of orbital behavior and 18^3 orbital samples.

[0080] Table 3 Classification results of all orbital cluster behavior labels

[0081] Table 3 shows that the number and proportion of labels for different orbital behaviors vary significantly in the dataset. To avoid interference from an extremely small number of labels (label 15, accounting for only 0.0005%) on model training, and to ensure the generality of the research, samples corresponding to this label were first removed. Even so, the remaining samples still have a high imbalance ratio (IR). The dataset still belongs to the extremely imbalanced dataset (EHIRD). The core parameter settings of the XGBoost algorithm of this invention are shown in Table 4. The simulation was completed on a computer equipped with an AMD EPYC 7B13 CPU, an RTX 4060 Ti GPU, and a main frequency of 2.25GHz. The generation of the orbital sample dataset and behavior annotation took about 26.4 minutes, and the training of the XGBoost model took about 53.8 minutes.

[0082] Figure 7 Table 5 presents the confusion matrix and performance metrics of the trained model on the test set. The data shows that the model performs exceptionally well across all multi-class evaluation metrics, indicating that the classifier possesses good discriminative ability and generalization performance for various orbital behaviors of spacecraft. However, traditional multi-class models often default to using the "maximum probability" as the classification decision criterion. This can easily lead to an imbalance in precision or recall for specific classes when there is uneven class distribution or asymmetric decision boundaries between different classes. To further improve the model's discrimination performance in multi-class scenarios, this study introduces a class threshold optimization method based on validation set performance metrics. As can be seen from the comparison data in Table 5, the model performance shows a significant improvement trend after threshold optimization: although Recall decreased slightly by 0.0021 (from 0.9414 to 0.9393), the decrease is negligible, and the model's ability to capture positive class samples remains basically stable; while key metrics such as Precision, F1-Score, G-Mean, and MMCC all improved, with Precision increasing by 0.0226 and F1-Score increasing by 0.0111, fully demonstrating the optimization effect.

[0083] In summary, the category-adaptive threshold optimization strategy effectively improves the model's overall performance in imbalanced multi-class classification tasks. It ensures a high ability to capture key orbital behaviors while reducing the risk of misclassification, making the prediction results more consistent with the Earth-Moon relationship. Research needs for predicting point orbit behavior.

[0084] Table 4 XGBoost Algorithm Parameter Settings

[0085] Table 5. Model performance index results (comparison before and after threshold optimization)

[0086] Example 2 Correspondingly, this application also provides a machine learning-based system for predicting the Earth-Moon L1 point crossing orbit behavior, including: The acquisition module is used to collect data based on different energy levels. By traversing the initial state set of spacecraft points, an orbital cluster sample dataset is constructed. The classification criteria module is used to define spacecraft orbital behavior categories and classification criteria, and to establish a spacecraft behavior label classification system. The classification module is used to generate corresponding behavior labels for each sample in the orbital cluster sample dataset according to the spacecraft behavior label classification system, forming a labeled orbital sample dataset; The training module is used to divide the labeled track sample dataset into a training set and a validation set. The XGBoost model is trained using the training set to obtain an initial prediction model. Based on the validation set, adaptive threshold optimization is performed on each behavior category to determine the optimal probability threshold for each category to form an optimal threshold vector. The original probability output is obtained using the trained model. The original probability is normalized according to the optimal threshold vector, and the final behavior prediction category is determined based on the normalized score.

[0087] This invention proposes a method for predicting spacecraft orbital behavior in low-energy crossing scenarios at the Earth-Moon L1 point. The method first establishes a classification system covering typical spacecraft orbital behaviors, capable of accurately distinguishing between various behaviors such as approaching Earth, lunar orbit, lunar impact, and bounded orbits. Based on this, a machine learning-based prediction model for Earth-Moon L1 point crossing orbital behavior is developed, achieving high-precision prediction of orbital behavior. This research fills a technological gap in the field of spacecraft behavior prediction in the Earth-Moon translational point region and provides a new technical means to enhance space situational awareness capabilities.

[0088] It should be noted that, in the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another device, or some features may be ignored or not executed. The modules described as separate components may or may not be physically separated. The components shown as modules may be one or more physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs.

[0089] Furthermore, in the various embodiments of the present invention, the modules can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0090] An electronic device provided in this application includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the machine learning-based method for predicting the Earth-Moon L1 point crossing orbit behavior as described in any of the above embodiments.

[0091] Another electronic device provided in this application embodiment may further include: an input port connected to a processor for transmitting multimodal data collected by an external acquisition device to the processor; a display unit connected to the processor for displaying the processor's processing results to the outside world; and a communication module connected to the processor for enabling communication between the electronic device and the outside world. The display unit may be a display panel, a laser scanning display, etc.; the communication method adopted by the communication module includes, but is not limited to, Mobile High Definition Link (HML), Universal Serial Bus (USB), High Definition Multimedia Interface (HDMI), and wireless connection (including Wi-Fi, Bluetooth, Bluetooth Low Energy, and IEEE 802.11s-based communication technology).

[0092] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the machine learning-based method for predicting the Earth-Moon L1 point crossing orbit behavior as described in any of the above embodiments.

[0093] For descriptions of relevant parts of the machine learning-based Earth-Moon L1 point crossing orbit behavior prediction system, electronic device, and computer-readable storage medium provided in this application's embodiments, please refer to the detailed descriptions of the corresponding parts in the machine learning-based Earth-Moon L1 point crossing orbit behavior prediction method provided in this application's embodiments; they will not be repeated here. Furthermore, parts of the technical solutions provided in this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0094] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for predicting the Earth-Moon L1 point crossing orbit behavior based on machine learning, characterized in that, Includes the following steps: Step 1: Based on different energy levels By traversing the initial state set of spacecraft points, an orbital cluster sample dataset is constructed. Define spacecraft orbital behavior categories and classification standards, and establish a spacecraft behavior labeling and classification system; Step 2: Based on the spacecraft behavior label classification system, generate corresponding behavior labels for each sample in the orbital cluster sample dataset to form a labeled orbital sample dataset; Step 3: Divide the labeled track sample dataset into a training set and a validation set. Use the training set to train the XGBoost model to obtain an initial prediction model. Based on the validation set, perform adaptive threshold optimization on each behavior category to determine the optimal probability threshold for each category to form an optimal threshold vector. Use the trained model to obtain the original probability output. Normalize the original probability according to the optimal threshold vector and determine the final behavior prediction category based on the normalized score.

2. The method for predicting the Earth-Moon L1 point crossing orbit behavior based on machine learning according to claim 1, characterized in that, In step 1, the The initial state set of the spacecraft traversed by the point was obtained in the following way: Based on the Earth-Moon planar circular restricted three-body model, the spacecraft traverses the Earth and Moon. Jacobian constant at point Poincaré section Velocity direction angle and y-axis coordinates As three independent variables characterizing its state; The Jacobian constant, velocity direction angle, and y-axis coordinate are discretely sampled within their respective physical ranges to generate the initial state set.

3. The method for predicting the Earth-Moon L1 point crossing orbit behavior based on machine learning according to claim 1, characterized in that, The physical range of the Jacobian constant is determined in the following way: Solve The point Poincaré section contains a Jacobian constant that intersects the zero-velocity curve corresponding to that Jacobian constant. value range .

4. A method for predicting the Earth-Moon L1 point crossing orbit behavior based on machine learning, as described in claim 2 or 3, characterized in that, For each given Jacobian constant The value, and its corresponding y-axis coordinate. Feasible range By solving for the Jacobian constant The intersection of the zero velocity curve at the specified value and the Poincaré section is obtained.

5. The method for predicting the Earth-Moon L1 point crossing orbit behavior based on machine learning according to claim 1, characterized in that, The spacecraft behavior label classification system defined in step 1 includes the following categories: Direct approach to Earth, lunar orbit followed by approach to Earth, direct impact on the Moon, lunar orbit followed by impact on the Moon, short-term lunar capture and L1 binding orbit; The lunar orbiting flight followed by approaching Earth, lunar orbiting flight followed by impacting the Moon, and short-term lunar capture behavior are further subdivided according to the lunar orbiting direction and the number of lunar orbits.

6. The method for predicting the Earth-Moon L1 point crossing orbit behavior based on machine learning according to claim 5, characterized in that, The subdivision based on the direction and number of revolutions around the moon includes: one clockwise revolution around the moon, one counterclockwise revolution around the moon, multiple clockwise revolutions around the moon, multiple counterclockwise revolutions around the moon, and mixed revolutions around the moon.

7. The method for predicting the Earth-Moon L1 point crossing orbit behavior based on machine learning according to claim 2, characterized in that, Before dividing the labeled orbital sample dataset into training and validation sets, a feature preprocessing step is also included: The velocity direction angle Convert to trigonometric function form ; For those containing Jacobi constant y-axis coordinates And perform polynomial interactive feature expansion on the eigenvectors of the trigonometric functions; The expanded features are filtered by variance thresholding to obtain effective feature vectors for model training.

8. The method for predicting the Earth-Moon L1 point crossing orbit behavior based on machine learning according to claim 1, characterized in that, Step 3 also includes a category weight calculation step: Calculate the weight of each behavior category based on the number of samples in the training set. ,in It is inversely proportional to the sample size; When training the XGBoost model, the class weights are applied to the loss function to mitigate the class imbalance problem.

9. The method for predicting the Earth-Moon L1 point crossing orbit behavior based on machine learning according to claim 1, characterized in that, The adaptive threshold optimization for each behavior category based on the validation set includes: For each behavior category Iterate through multiple candidate thresholds within a preset probability interval. ; For each candidate threshold Calculate the accuracy of this category. and recall rate And calculate the F1 score based on this. ; Choose to The largest candidate threshold is taken as the optimal probability threshold for that category. .

10. A machine learning-based system for predicting the behavior of Earth-Moon L1 point crossing orbits, characterized in that, include: The acquisition module is used to collect data based on different energy levels. By traversing the initial state set of spacecraft points, an orbital cluster sample dataset is constructed. The classification criteria module is used to define spacecraft orbital behavior categories and classification criteria, and to establish a spacecraft behavior label classification system. The classification module is used to generate corresponding behavior labels for each sample in the orbital cluster sample dataset according to the spacecraft behavior label classification system, forming a labeled orbital sample dataset; The training module is used to divide the labeled track sample dataset into a training set and a validation set. The XGBoost model is trained using the training set to obtain an initial prediction model. Based on the validation set, adaptive threshold optimization is performed on each behavior category to determine the optimal probability threshold for each category to form an optimal threshold vector. The original probability output is obtained using the trained model. The original probability is normalized according to the optimal threshold vector, and the final behavior prediction category is determined based on the normalized score.