Giant satellite maneuvering state identification method based on machine learning

Abstract

The invention relates to a giant satellite maneuvering state identification method based on machine learning, and the method comprises the steps: obtaining two lines of orbit data TLE data of a target giant satellite, and obtaining an initial semi-major axis time sequence based on the TLE data; a self-encoding neural network is adopted to extract low-dimensional nonlinear features of the initial semi-major axis time sequence, a semi-major axis low-dimensional nonlinear feature vector is obtained, the self-encoding neural network is obtained based on training of an initial labeled sample, and the initial labeled sample is a historical semi-major axis time sequence marking the maneuvering state; the semi-major axis low-dimensional nonlinear feature vectors are input into an orbit maneuver state classification model, the orbit maneuver state of the target giant satellite is obtained, the orbit maneuver state classification model is constructed based on a decision tree model and is obtained based on training of a training set, and the training set is historical semi-major axis low-dimensional nonlinear feature vectors marking the maneuver state. According to the method, the orbit maneuver state of the small-thrust continuous maneuver of the large constellation satellite can be identified.

Description

Technical Field

[0001] The present invention relates to the field of satellite monitoring technology, and in particular to a method for identifying the maneuvering state of a giant satellite based on machine learning. Background Art

[0002] The identification of orbital maneuvering status of large constellation satellites is one of the core technologies for achieving space situational awareness and constellation satellite collision risk assessment. With the rapid deployment of large constellation satellites, the research demand for interpreting and processing massive intelligence, surveillance and reconnaissance data (intelligence, reconnaissance, surveillance) using technical means such as big data and artificial intelligence is becoming increasingly urgent. In order to accurately perceive threats to the space environment, major space powers have successively carried out research on space situational awareness technology. In a broad sense, space situational awareness perceives all events, threats, activities and states occurring in space. It is the basis of all space activities, the premise for ensuring the safety of space assets and the key means to understand the intentions of space activities. The orbital maneuvering of in-orbit satellites is one of the main research contents of space situational awareness. In the next few years, large constellation satellites will occupy a dominant position in space objects and may become the biggest factor affecting the space security situation. At present, there are nearly 5,000 STARLINK satellites in orbit, and it is planned to deploy a constellation system consisting of 42,000 satellites. The total number of satellites it plans to launch far exceeds the total number of satellites launched in the past sixty years of human space activities (15,912 as of August 29, 2023, including 4,851 STARLINK satellites). The recognition of orbital maneuvering intentions of large constellation satellites will become one of the core technologies of space situational awareness. In the face of the "complexity and enlargement" of future space situations, the use of big data, artificial intelligence and other technical means to interpret and process massive intelligence, surveillance and reconnaissance data, thereby shortening the intelligence acquisition cycle, has gradually become a development trend of space situational awareness.

[0003] In existing research, non-cooperative satellite maneuver detection methods are generally based on adjacent TLE data to determine the maneuver time, pulse size, and direction that occur between them. The accuracy of such algorithms is closely related to the maneuver frequency and maneuver amplitude. When the satellite maneuver frequency is greater than the update frequency of TLE data, there may be multiple maneuvers between adjacent TLE data, and such algorithms are difficult to make reliable judgments. When the maneuver amplitude is small, the applicability of such algorithms also decreases. For satellites with small pulse continuous maneuvers, multiple maneuvers may occur between two adjacent TLE data, resulting in conventional maneuver detection methods no longer being applicable. At the same time, the inevitable noise and abnormal data in TLE data will also affect the evaluation results. Therefore, the current non-cooperative satellite orbital maneuver identification algorithms are mostly carried out for single pulse maneuver detection, and it is difficult to identify the maneuver state of large constellation satellites with small thrust continuous maneuvers. The orbital maneuver state identification method for large constellation satellites with small thrust continuous maneuvers needs to be solved urgently. Summary of the invention

[0004] The purpose of the present invention is to provide a giant satellite maneuvering state identification method based on machine learning, which can identify the orbital maneuvering state of a large constellation satellite with low thrust continuous maneuvering.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for identifying the maneuvering state of a giant satellite based on machine learning, comprising:

[0007] Acquire two lines of orbital data TLE data of the target giant satellite, and acquire an initial semi-major axis time series based on the TLE data;

[0008] An autoencoder neural network is used to extract low-dimensional nonlinear features from the initial semi-major axis time series to obtain a semi-major axis low-dimensional nonlinear feature vector, wherein the autoencoder neural network is obtained based on initial labeled samples training, and the initial labeled samples are historical semi-major axis time series of marked maneuvering states;

[0009] The semi-major axis low-dimensional nonlinear feature vector is input into an orbital maneuvering state classification model to obtain the orbital maneuvering state of the target giant satellite, wherein the orbital maneuvering state classification model is constructed based on a decision tree model and is obtained through training based on a training set, wherein the training set is a historical semi-major axis low-dimensional nonlinear feature vector marking the maneuvering state.

[0010] Optionally, obtaining the initial labeled sample includes:

[0011] Acquire historical TLE data, and acquire historical semi-major axis time series based on the historical TLE data;

[0012] The historical semi-major axis time series is divided based on the orbital maneuvering state, the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite is obtained, and the data is cleaned and preprocessed to obtain the orbital element time series data with equal time intervals, wherein the orbital maneuvering state includes: controlled climb, controlled maintenance, controlled descent and no orbit control;

[0013] A sliding window algorithm is used to divide and normalize the semi-major axis time series in the orbital element time series data to obtain the initial labeled samples.

[0014] Optionally, the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite is:

[0015]

[0016] Among them, a is the semi-major axis data, a(t n ) is tn The semi-major axis data at time, Y State is the maneuvering state, n is the nth moment, It is the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite.

[0017] Optionally, data cleaning of the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite includes:

[0018] When the update time interval of adjacent historical semi-major axis time series data exceeds a preset time, the adjacent historical semi-major axis time series data are divided into time series data of different segments;

[0019] Abnormal data in the historical semi-major axis time series data is eliminated.

[0020] Optionally, performing data preprocessing on the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite includes:

[0021] The SGP4 algorithm is used to perform orbit prediction on the cleaned data to obtain the orbit element time series data with equal time intervals:

[0022]

[0023] Among them, m is the mth moment, and c is the cth moment.

[0024] Optionally, a sliding window algorithm is used to divide the semi-major axis time series in the orbital element time series data into:

[0025] X k =[a k+0 ,a k+1 ,...,a k+N1-1 ]

[0026] Among them, X k is the kth sample, a is the semi-major axis data, and N1 is the sliding window length.

[0027] Optionally, extracting low-dimensional nonlinear features from the initial semi-major axis time series using an autoencoder neural network includes:

[0028] Inputting the initial semi-major axis time series as an input vector into an encoder of an autoencoder neural network, and outputting intermediate layer features;

[0029] The intermediate layer features are input into the decoder to reconstruct the samples, and the reconstructed vector is obtained, that is, the semi-major axis low-dimensional nonlinear feature vector is obtained.

[0030] Optionally, constructing the orbital maneuvering state classification model based on a decision tree model includes:

[0031] According to the discrete method, the decision tree models with different layer structures are obtained;

[0032] The decision tree models with different number of layers are trained respectively by using the training set, the decision tree models with different number of layers are screened, and the optimal decision tree model is obtained as the orbital maneuvering state classification model.

[0033] The beneficial effects of the present invention are as follows: Unlike the existing maneuver identification methods, the present invention does not focus on the specific maneuvering time and maneuvering pulse, but intends to divide the orbital maneuvering state of the constellation satellite into four categories: "controlled climb, controlled maintenance, controlled descent and no orbital control", and uses historical data to train the four types of orbital maneuvering state identification models, which can identify the orbital maneuvering state of the small-thrust continuous maneuvering of large constellation satellites. The present invention can be applied to the fields of spacecraft monitoring, early warning, and collision avoidance. Based on the publicly released TLE data, the orbital state of non-cooperative large constellation satellites is automatically identified, laying a theoretical foundation for early warning work for such satellites. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0035] Figure 1 A flowchart of a method for identifying the maneuvering state of a giant satellite based on machine learning according to an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of initial samples generated by using a sliding window algorithm according to an embodiment of the present invention;

[0037] Figure 3 Schematic diagram of the training process of the autoencoding neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] like Figure 1 As shown, this embodiment provides a method for identifying the maneuvering state of a giant satellite based on machine learning, including:

[0041] Obtain two lines of orbital data TLE data of the target giant satellite, and obtain the initial semi-major axis time series based on the TLE data;

[0042] An autoencoder neural network is used to extract low-dimensional nonlinear features from the initial semi-major axis time series to obtain a low-dimensional nonlinear feature vector of the target semi-major axis, wherein the autoencoder neural network is trained based on initial labeled samples, and the initial labeled samples are historical semi-major axis time series of marked maneuvering states;

[0043] The semi-major axis low-dimensional nonlinear feature vector is input into the orbital maneuvering state classification model to obtain the orbital maneuvering state of the target giant satellite, wherein the orbital maneuvering state classification model is constructed based on a decision tree model and is obtained through training based on a training set, wherein the training set is the historical semi-major axis low-dimensional nonlinear feature vector marking the maneuvering state.

[0044] Furthermore, obtaining the initial labeled samples includes:

[0045] Obtain historical TLE data, and obtain historical semi-major axis time series based on the historical TLE data;

[0046] The historical semi-major axis time series is divided based on the orbital maneuvering state, the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite is obtained, and the data is cleaned and preprocessed to obtain the orbital element time series data with equal time intervals, where the orbital maneuvering state includes: controlled climb, controlled maintenance, controlled descent and no orbit control;

[0047] The sliding window algorithm is used to divide and normalize the semi-major axis time series in the orbital root time series data to obtain the initial labeled samples.

[0048] Specifically, this embodiment obtains the STARLINK satellite historical TLE data publicly released by the US Space Monitoring Network from Space-Track.Org, uses a sliding window algorithm to generate constellation satellite orbit data samples marked with orbital maneuvering states, and obtains initial labeled samples for generating the constellation satellite orbit time series data.

[0049] This embodiment extracts the orbital element data from the STARLINK orbital element data publicly released by the US Space Surveillance Network, as shown in Table 1:

[0050] Table 1

[0051]

[0052]

[0053] Furthermore, obtaining the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite specifically includes:

[0054] The semi-major axis (denoted as a) is extracted from the TLE data and arranged in chronological order. State =1), controlled maintenance (Y State =2), controlled descent (Y State =3), no track control (Y State =4) four categories. Based on the four types of orbital data, the semi-major axis time series is further divided. Therefore, the semi-major axis time series data of the same satellite and the same type of orbital maneuvering state are recorded as:

[0055]

[0056] Among them, a is the semi-major axis data, a(t n ) is t n The semi-major axis data at time, Y State is the maneuvering state, n is the nth moment, It is the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite.

[0057] Furthermore, data cleaning of the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite includes:

[0058] When the update time interval of adjacent historical semi-major axis time series data exceeds the preset time, the adjacent historical semi-major axis time series data are divided into time series data of different segments, and the abnormal data in the historical TLE data are removed.

[0059] Specifically, it can be seen from the Space-Track.Org data that the TLE data update time is unstable and the satellite's orbital maneuvering state will switch multiple times. To ensure the rationality of the data, for the semi-major axis time series of the same satellite, when the update time interval of adjacent semi-major axis time series data exceeds one day, the adjacent semi-major axis time series data will be disconnected into two time series data. At the same time, the abnormal data with maneuvering amplitude exceeding the theoretical limit will be eliminated.

[0060] Furthermore, the data preprocessing of the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite specifically includes:

[0061] To solve the problem of unstable TLE data update frequency, the SGP4 algorithm is used to predict orbits and obtain the time series data of orbit elements with equal time intervals:

[0062]

[0063] t c -t c-1 =Δt,1<c≤m

[0064] Among them, c, m are the c, m-th moments.

[0065] Furthermore, the sliding window algorithm is used to divide the semi-major axis time series in the orbital element time series data into:

[0066] X k =[a k+0 ,a k+1 ,...,a k+N1-1 ] (3)

[0067] Among them, X k is the kth sample, a is the semi-major axis data, and N1 is the sliding window length

[0068] Specifically, the sliding window algorithm is used to generate the initial samples. Let the sliding window step length be Ns, the sliding window length be N1, and each semi-major axis time series be divided to generate a series of semi-major axis time series samples of the same length. Assuming Ns = 1, N1 = 10, the sample division is as follows: Figure 2 As shown. Therefore, the kth sample can be recorded as formula (3).

[0069] However, a reliable classification model should be able to identify the variation pattern of the semi-major axis rather than the specific orbital altitude. Therefore, the semi-major axis sequence in each segment is normalized:

[0070]

[0071] In summary, we get the initial labeled samples of the satellite orbit time series data of this constellation.

[0072] In this embodiment, 11,980 samples are constructed according to the sliding window algorithm.

[0073] Furthermore, the autoencoder neural network is used to extract low-dimensional nonlinear features from the initial semi-major axis time series, including:

[0074] The initial semi-major axis time series is input as an input vector into the encoder of the autoencoder neural network, and the intermediate layer features are output;

[0075] The intermediate layer features are input into the decoder to reconstruct the samples and obtain the reconstructed vector.

[0076] Specifically, this embodiment trains the AutoEncoder and the corresponding Encoder network structure to solve the inevitable errors and anomalies in the track data, and uses the AutoEncoder method to extract the low-dimensional nonlinear features of the samples to achieve dimensionality reduction and noise reduction of the samples. Train the network parameters. Processing, Y state No changes are required to ensure that the low-dimensional vector is also labeled. With the goal of "the output after encoding-decoding can reproduce the initial input to the greatest extent", the network parameters are optimized and trained. After training, Processing is performed to obtain a low-dimensional vector [x,Y State ]. Specifically include:

[0077] ①Use the neural network model to generate an encoder to achieve sample dimensionality reduction. The encoder weight is w (a matrix composed of component vectors) and the deviation is b m . Thus, the hidden layer x can be expressed as a nonlinear function (encoder):

[0078]

[0079] Generate a decoder to reconstruct the sample. According to the transposed matrix w of the decoder weight w T , the deviation is b d . So, the reconstruction vector It can be expressed as a nonlinear function (encoder):

[0080]

[0081] ② Feature extraction autoencoder neural network training. The main function of the autoencoder neural network is to restore the output as much as possible to the input. The better the network is trained, the more representative the extracted intermediate layer features (hidden layer x) are, and the better the feature extraction effect is. In other words, the entire network training process is to obtain the optimal weight w so that the loss function representing the difference between the predicted value and the actual value is minimized.

[0082] In order to achieve the goal of restoring the input as much as possible, the optimization objective function of the algorithm should be recorded as:

[0083]

[0084] Where MSE represents the original input and the reconstruction vector The mean square error between , Min represents the minimum value of the variable. Therefore, the loss function of the autoencoder method is expressed as:

[0085]

[0086] Among them, N represents the number of input samples and λ is the regularization parameter.

[0087] The training process of the neural network is to adjust the weight w parameter through continuous iterative optimization, and then obtain the optimal value of the autoencoder neural network, see Figure 3 .

[0088] Furthermore, the orbital maneuvering state classification model based on the decision tree model includes:

[0089] According to the discrete method, the decision tree models with different layer structures are obtained;

[0090] The decision tree models with different number of layers are trained using the training sets, and the decision tree models with different number of layers are screened to obtain the optimal decision tree model as the orbital maneuvering state classification model.

[0091] Specifically, the orbital maneuvering state recognition based on the decision tree includes:

[0092] ① Training set and test set division. All labeled feature samples, i.e., the historical semi-major axis low-dimensional nonlinear feature vectors [x,Y State ] is divided into two parts: training set and test set. Both the training set and the test set must contain four different orbital maneuvering states.

[0093] ② Training and screening of decision tree models. According to the discrete method, multiple decision tree structures (number of decision tree layers) are set. For each decision tree structure, the training set is used to train it separately, and the network classification results (model calculation results) of the training set are compared with the sample labels (real results). The "model accuracy" is used as the judgment standard for the optimal nodes and decision rules to obtain the optimal nodes and decision rules. Different decision tree structures are compared horizontally to screen the optimal decision tree model, that is, the orbital maneuvering state classification model.

[0094] ③ Quantitative evaluation of orbital maneuvering state classification model. The model is evaluated using the test set and orbital maneuvering state classification model. For the multi-classification model, the confusion matrix of a certain type k is shown in Table 2:

[0095] Table 2

[0096]

[0097] The general evaluation parameter is the model precision (the proportion of correct model judgments among all results where the model prediction is Positive). For a certain type k, the judgment result is:

[0098]

[0099] For the multi-classification model of this embodiment, the macro average evaluation index can be expressed as:

[0100]

[0101] In this embodiment, the track state classification is obtained by combining the maneuvering state recognition algorithm as shown in Table 3:

[0102] Table 3

[0103]

[0104] According to formula (10), for the four orbit states of controlled orbit maintenance (CM), controlled orbit climb (CR), controlled orbit descent (CL), and no maneuvering (NM), the macro precision of the algorithm of this patent in the above example is 93.8%.

[0105] The method of this embodiment can better realize the automatic identification of the orbital status of large-scale constellation satellites, and can lay a theoretical foundation for the monitoring and early warning of the earth's orbital space environment.

[0106] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for identifying the maneuvering state of a giant satellite based on machine learning, characterized in that: include: Acquire two lines of orbital data TLE data of the target giant satellite, and acquire an initial semi-major axis time series based on the TLE data; An autoencoder neural network is used to extract low-dimensional nonlinear features from the initial semi-major axis time series to obtain a semi-major axis low-dimensional nonlinear feature vector, wherein the autoencoder neural network is obtained based on initial labeled samples training, and the initial labeled samples are historical semi-major axis time series of marked maneuvering states; The semi-major axis low-dimensional nonlinear feature vector is input into an orbital maneuvering state classification model to obtain the orbital maneuvering state of the target giant satellite, wherein the orbital maneuvering state classification model is constructed based on a decision tree model and is obtained through training based on a training set, wherein the training set is a historical semi-major axis low-dimensional nonlinear feature vector marking the maneuvering state.

2. The method for identifying the maneuvering state of a giant satellite based on machine learning according to claim 1, characterized in that: Obtaining the initial labeled sample includes: Acquire historical TLE data, and acquire historical semi-major axis time series based on the historical TLE data; The historical semi-major axis time series is divided based on the orbital maneuvering state, the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite is obtained, and the data is cleaned and preprocessed to obtain the orbital element time series data with equal time intervals, wherein the orbital maneuvering state includes: controlled climb, controlled maintenance, controlled descent and no orbit control; A sliding window algorithm is used to divide and normalize the semi-major axis time series in the orbital element time series data to obtain the initial labeled samples.

3. The method for identifying the maneuvering state of a giant satellite based on machine learning according to claim 2, characterized in that: The historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite is: Among them, a is the semi-major axis data, a(t n ) is t n The semi-major axis data at time, Y State is the maneuvering state, n is the nth moment, It is the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite.

4. The method for identifying the maneuvering state of a giant satellite based on machine learning according to claim 2, characterized in that: Data cleaning of the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite includes: When the update time interval of adjacent historical semi-major axis time series data exceeds a preset time, the adjacent historical semi-major axis time series data are divided into time series data of different segments; Abnormal data in the historical semi-major axis time series data is eliminated.

5. The method for identifying the maneuvering state of a giant satellite based on machine learning according to claim 3, characterized in that: The data preprocessing of the historical semi-major axis time series data of the same type of orbital maneuvering state of the target giant satellite includes: The SGP4 algorithm is used to perform orbit prediction on the cleaned data to obtain the orbit element time series data with equal time intervals: Among them, m is the mth moment, and c is the cth moment.

6. The method for identifying the maneuvering state of a giant satellite based on machine learning according to claim 2, characterized in that: The sliding window algorithm is used to divide the semi-major axis time series in the orbital element time series data into: X k =[a k+0 ,a k+1 ,...,a k+N1-1 ] Among them, X k is the kth sample, a is the semi-major axis data, and N1 is the sliding window length.

7. The method for identifying the maneuvering state of a giant satellite based on machine learning according to claim 2, characterized in that: The method of extracting low-dimensional nonlinear features from the initial semi-major axis time series using an autoencoder neural network includes: Inputting the initial semi-major axis time series as an input vector into an encoder of an autoencoder neural network, and outputting intermediate layer features; The intermediate layer features are input into the decoder to reconstruct the samples, and the reconstructed vector is obtained, that is, the semi-major axis low-dimensional nonlinear feature vector is obtained.

8. The method for identifying the maneuvering state of a giant satellite based on machine learning according to claim 1, characterized in that: Constructing the orbital maneuvering state classification model based on the decision tree model includes: According to the discrete method, the decision tree models with different layer structures are obtained; The decision tree models with different number of layers are trained respectively by using the training set, the decision tree models with different number of layers are screened, and the optimal decision tree model is obtained as the orbital maneuvering state classification model.

Images