METHOD FOR IDENTIFYING THE MANEUVER STATE OF GIANT SATELLITES BASED ON MACHINE LEARNING

Description

2 of the information gathering period gradually becomes a development trend in space situational awareness. In previous research, the method for detecting non-cooperative satellite maneuvers generally relies on adjacent two-line element (TLE) data, with the aim of determining the timing, impulse strength, and direction of the maneuver during the period under consideration. The accuracy of such algorithms is related to the maneuver frequency and the maneuver amplitude. If the satellite's maneuver frequency is higher than the TLE data update frequency, multiple maneuvers can occur between adjacent TLE data, and such an algorithm can only achieve a reliable assessment with difficulty. If the maneuver amplitude is small, the applicability of such algorithms also decreases. For a satellite with continuous small-impulse maneuvers, multiple maneuvers can occur between two adjacent TLE data, leading to...that a conventional method for detecting maneuvers is no longer applicable. At the same time, unavoidable noise and abnormal data in the TLE data impair the evaluation result. Therefore, current algorithms for identifying orbital maneuvers of non-cooperative satellites focus predominantly on detecting single-impulse maneuvers and are hardly able to capture the maneuver state of large constellation satellites with continuous low-thrust maneuvers. There is therefore an urgent need for a method to identify the maneuver state of large constellation satellites with continuous low-thrust maneuvers. CONTENT OF THE PRESENT INVENTION The objective of the present invention is to provide a method for identifying the maneuver state of giant satellites based on machine learning25, which can identify the orbital maneuver state of large constellation satellites with continuous low-thrust maneuvers. To achieve the aforementioned objective,The present invention offers the following solutions: a method for identifying the maneuver state of giant satellites on the basis of machine learning, comprising: obtaining two series of orbital data (Two-LineElements data, TLE data) of a target giant satellite and obtaining time series of the initial major half-axis based on the TLE data; Extracting a low-dimensional nonlinear feature from the time series of the initial major half-axis using an autoencoder neural network to obtain a low-dimensional nonlinear feature vector of the major half-axis, wherein the autoencoder neural network is obtained by training on the basis of originally labeled samples and the two originally labeled samples are time series of a historical major half-axis with labeled maneuver states; and inputting the low-dimensional nonlinear feature vector of the major half-axis into a classification model for orbital maneuver states.to obtain an orbital maneuver state of the target giant satellite, wherein the classification model for orbital maneuver states is constructed on the basis of a decision tree model and is obtained by training on the basis of a training set, and the training set is a low-dimensional nonlinear feature vector of the historical semi-major axis with labeled maneuver states. Optionally, obtaining the originally labeled samples includes: acquiring historical TLE data and acquiring time series of the historical semi-major axis on the basis of the historical TLE data; Subdividing the historical large semi-axis on the basis of the orbital maneuver states, obtaining time series data of the historical large semi-axis of the target giant satellite in the same type of orbital maneuver state, and performing data cleaning and data preprocessing to obtain time series data of orbital elements of equal time intervals, where the orbital maneuver states are controlled lift,Controlled orbit maintenance, controlled descent, and no orbit control are included; and segmenting and normalizing historical semi-major axis time series data within the orbit element time series data using a 25 sliding window algorithm to preserve the originally labeled samples. Optionally, the historical semi-major axis time series data of the target giant satellite in the same type of orbital maneuver state are as follows: ⃗()=[(),(),(),......()],∈{1,2,3,4} where a are semi-major axis data, (() are semi-major axis data at time , a maneuver state is ,The time series data of the historical large semi-axis of the target giant satellite are in the same type of orbital maneuver state. Optionally, performing data cleaning on the time series data of the historical large semi-axis of the target giant satellite in the same type of orbital maneuver state includes the following: BE2025 / 5444 4 Subdividing the adjacent time series data of the historical large semi-axis into time series data of different segments if an update time interval between adjacent time series data of the historical large semi-axis exceeds a preset time; and eliminating abnormal data in the time series data of the historical large semi-axis. Optionally, performing data preprocessing on the time series data of the historical large semi-axis of the target giant satellite in the same type of orbital maneuver state includes the following: performing an orbit prediction on the cleaned data using a 10 SGP4 algorithm,to obtain the time series data of orbital elements of equal time intervals: ⃗()=[(),(),(),……,(),()....()],∈{1,2,3,4} where meinm-ter timepoint and ceinc-ter timepoint is. Optionally, the historical major half-axis time series is segmented within the 15 time series data of orbit elements using a sliding window algorithm as follows: []0111,,..., + + + - = kkkkNXaaa where kX is an k-th sample, a is major half-axis data, and N1 is a sliding window length. 20 Optionally, the extraction of the low-dimensional nonlinear feature from the initial major half-axis time series using an autoencoder neural network includes the following: inputting the initial major half-axis time series as an input vector into an encoder of the autoencoder neural network and outputting a feature of the 25 intermediate layer; and inputting the intermediate layer feature into a decoder for reconstruction. sampling and obtaining a reconstruction vector,dhPreserving low-dimensional nonlinear feature vector of the major semi-axis. Optionally, the constructive classification model for orbital maneuver states based on a decision tree model includes the following: obtaining decision tree models with different layer structures according to a discretization procedure; and training the decision tree models with different layer structures using the training set, filtering the decision tree models with different 35 BE2025 / 5444 5 structural layers and obtaining an optimal decision tree model as a classification model for the orbital maneuver state. The present invention has the following advantageous effects: unlike existing maneuver identification methods, the present invention does not concentrate attention on specific maneuver timings and maneuver impulses, and is intended to divide the orbital maneuver states of the constellation satellites into four types: controlled ascent,Controlled orbit maintenance, controlled descent, and no orbit control, and by training identification models of the four types of orbital maneuver states using historical data, the present invention can identify the orbital maneuver state of large constellation satellites with 10 continuous low-thrust maneuvers. The present invention can be used in the areas of spacecraft monitoring, early warning, and collision avoidance. Based on published TLE data, the orbital state of the non-cooperative large constellation satellite is automatically identified, and it can provide a theoretical basis for early warning of such satellites. 15 DESCRIPTION OF THE DRAWING To illustrate the embodiments of the present invention or the prior art technical solutions more clearly, the attached drawings, which are to be used in the embodiments, are briefly described below. 20 undesistoffe obviously,that the drawings accompanying the following descriptions are merely some embodiments of the present invention, and persons skilled in the art can obtain further drawings according to these drawings without inventive effort. Fig. 1 is a flowchart of a method for identifying the maneuver state 25 of a giant satellite based on machine learning according to an embodiment of the present invention; Fig. 2 is a schematic diagram of an initial sample generated using a sliding window algorithm according to an embodiment of the present invention; and Fig. 30 is a schematic diagram of a training process of an autoencoder neural network according to an embodiment of the present invention. DETAILED DESCRIPTION The technical solutions in the embodiments of the present invention35 are described clearly and completely below with reference to the attached drawings in BE2025 / 5444 6 embodiment of the present invention,It is obvious that the described embodiments represent only a part of the embodiment of the present invention and not all embodiments. All other embodiments that a person skilled in the art can derive from the embodiment of the present invention without creative effort fall within the scope of protection of the present invention. To make the aforementioned subject matter, features, and advantages of the present invention more understandable, the present invention is described in more detail below with reference to the attached drawings and specific embodiments. As shown in Figure 1, the present embodiment provides a method for identifying the maneuver state of giant satellites based on machine learning, which includes the following: obtaining two-line element data,TLE data) of a target giant satellite and obtaining initial major half-axis time series 15 based on the TLE data; extracting a low-dimensional nonlinear feature from the initial major half-axis time series using an autoencoder neural network to obtain a low-dimensional nonlinear major half-axis feature vector, wherein the autoencoder neural network is obtained by training on the 20 basis of originally labeled samples and the two originally labeled samples are historical major half-axis time series with labeled maneuver states; and inputting the low-dimensional nonlinear major half-axis feature vector into a classification model for orbital maneuver states to obtain an orbital 25 maneuver state of the target giant satellite,where the classification model for orbital maneuver states is constructed on the basis of a decision tree model and is obtained by training on the basis of a training set, and the training set is a low-dimensional nonlinear feature vector of the historical semi-major axis with labeled maneuver states.30 Furthermore, obtaining the originally labeled samples includes: obtaining historical TLE data and obtaining historical semi-major axis time series on the basis of the historical TLE data; Subdividing the historical large semi-axis on the basis of the orbital maneuver states, preserving time series data of the historical large semi-axis of BE2025 / 5444 7 target giant satellites of the same type of orbital maneuver state, and performing data cleaning and data preprocessing to obtain time series data of orbital elements of equal time intervals, where the orbital maneuver states are controlled lift, controlled orbit maintenance,Controlled descent and no orbit control are included; and 5 Segmenting and normalizing historical semi-major axis time series data within orbit element time series data using a sliding window algorithm to obtain the originally labeled samples. In particular, in this implementation, historical TLE data of the STARLINK satellites are obtained from Space-Track.org, published by the US Space Surveillance Network, and orbital data samples of the constellation satellites with a marker for the orbital maneuver state are generated using a sliding window algorithm, thereby obtaining the originally labeled sample to generate the constellation satellite orbit time series data. In this implementation, the historical TLE data of the STARLINK satellites are extracted from the US Space Surveillance Network.as shown in Table 1: TABLE 1 Orbital Status Sample Controlled Orbit Maintenance The 10 satellites launched on November 11, 2019 (numbers 44713 to 44722) Controlled Ascending 1) The 10 satellites launched on November 11, 2019 (numbers 44713 to 44722) 2) The 50 satellites launched on January 7, 2020 (numbers 44924 to 44973) Controlled Descending All 60 satellites launched on May 24, 2019 No Orbit Control 1) The 5 satellites launched on May 24, 2019 (numbers 44257, 44273, 44279, 44281, 44287) 2) 10 uncontrolled objects on neighboring orbits Orbits (numbers 03127, 03131, 03132, 03137, 03164, 03166, 03177, 03212, 03213, 03214) 20 Furthermore, the time series data of the historical major semi-axis of the target giant satellite are obtained in the same way as the orbital maneuver state as follows: BE2025 / 5444 8 the major semi-axis (designated as sa) is extracted from the TLE data and arranged in chronological order. In this form, the maneuver state of the satellite orbit is divided into four categories: controlled lift (=1),Controlled orbit maintenance (=2), controlled descent (=3), and no orbit control (=4). Based on the four types of satellite data, the major semi-axis time series is further subdivided. Therefore, the major semi-axis time series data of the same satellite and of the same type of orbital maneuver state are plotted as follows: ⃗()=[(),(),(),......()],∈{1,2,3,4}(1) where ai are the major semi-axis data, (()) are the major semi-axis data at time 10, a maneuver state, ni are the time points, and ⃗() are the historical major semi-axis data of the target giant satellite in the same type of orbital maneuver state. Furthermore, performing data cleaning on the historical large-scale half-axis time series data of the target giant satellite in the same type of orbital maneuver state includes the following: subdividing the adjacent historical large-scale half-axis time series data into time series data of different segments,when an update time interval between adjacent time series data of the historical major half-axis exceeds a preset time; and elimination of abnormal data in the time series data of the historical major half-axis. In particular, as can be seen from the Space-Track.org data, the update time of the TLE data is unstable, and the orbital maneuver state of the satellite can change multiple times. To ensure the rationality of the data, for the time series of the major half-axis of the same satellite, if the update interval of the time series of the adjacent major half-axis exceeds one day, the time series data of the adjacent major half-axis are split into two time series data. At the same time, abnormal data whose maneuver amplitude exceeds the theoretical limit are eliminated.eliminated. Furthermore, performing data preprocessing on the time series data of the 30 historical large semi-axis of the target giant satellite in the same type of orbital maneuver state includes, in particular, the following: For the problem of the unstable update frequency of TLE data, the SGP4 algorithm is used to obtain the time series data of orbital elements of equal time intervals: 35 ⃗()=[(),(),(),……,(),()....()],∈{1,2,3,4}(1) BE2025 / 5444 9 1 ,1 cc tttcm   where c and m are the c-th time point and the m-th time point. Furthermore, the time series of the historical major half-axis within the time series data of orbital elements is segmented using a sliding window algorithm as follows: 5 []0111,,..., + + + - = kkkkNXaaa (3) where kX is the kth sample, the data of the major half-axis and N1 is the length of the sliding window. In particular, the initial sample is generated using the sliding window algorithm. The step length of the sliding window is Ns, the length of the sliding window is 10 N1,and each segment of the major semi-axis time series is divided to generate a series of major semi-axis time series samples of equal length. Assuming that Ns=1 and N1=10, the sample distribution is shown in Fig. 2. Therefore, the tth sample can be referred to as formula (3). However, a reliable classification model should be able to identify the major semi-axis change pattern and not the specific orbital height. Therefore, the sequence of the major semi-axes in each segment must be normalized: (4) In summary,that the originally labeled sample 20 of the time series data of the satellite orbit of the constellation is obtained. In this implementation, 11980 samples are created using the sliding window algorithm. Furthermore, the Extracting low-dimensional nonlinear feature from the time series of the initial major half-axis using an autoencoder-25 neural network comprises the following: inputting the time series of the initial major half-axis as an input vector into an encoder of the autoencoder neural network and outputting an intermediate layer feature; and inputting the intermediate layer feature into a decoder for a reconstruction 30 of the samples and obtaining a reconstruction vector, i.e., the obtaining low-dimensional nonlinear major half-axis feature vector.   ˆkk kk XmeanX X meanX   StateYX,ˆ BE2025 / 5444 10 In particular, in this implementation form, the inevitable errors and anomalies in the orbital data of the autoencoders and the corresponding encoder network structure are trained,The low-dimensional nonlinear feature of the samples is extracted using an autoencoder method to achieve dimensionality and noise reduction of the samples, with the 5 samples of the training set being used to train the network parameters. During training, only the data is processed and does not need to be modified, and it is ensured that the low-dimensional vector is also labeled. With the goal that "the output after encoding-decoding can maximally reproduce the initial input," the network parameters are optimized and trained for training, and after training, processing is performed to obtain the labeled low-dimensional vector. In detail, the procedure includes: (1) Using the neural network model, an encoder is generated to achieve the data of the samples. The encoder weight is recorded as (a matrix composed of 15 component vectors),and the deviation is. Thus, the hidden layer can be represented as a nonlinear function (encoder): (5) To reconstruct the samples, a decoder is generated. According to the transposition matrix with the weight of the decoder, the deviation is 20. Thus, the reconstruction vector can be represented as a nonlinear function (encoder): (6) (2) Feature extraction during training of the autoencoder neural network. The main function of the autoencoder neural network is to reproduce the output as closely as possible. The better the network is trained, the more representative the extracted features of the intermediate layer (hidden layer) and the better the effect of feature extraction. In other words, the entire training process of the network aims to obtain the optimal weight, so that the value of the loss function, which represents the difference between the predicted measurement 30 and the actual value, is minimal. [StateYX,ˆ [StateYX,ˆ [StateYx,w mb xm ˆ()xfwXb Tww db X ()T dXgwxb xw BE2025 / 5444 11 To achieve the goal of reproducing the input as closely as possible, the optimization goal function of the algorithm should be recorded as follows: (7) where MSE stands for the mean squared error between the original input5 and the reconstruction vector and represents the minimum value of the variable. Then the loss function of the autoencoder method is expressed as follows: (8) where N represents the number of input samples and is the regularization parameter.10 The training process of the neural network consists of adjusting the weight parameter through continuous iterative optimization to obtain the optimal value of the autoencoder neural network, as i,