A hybrid prediction method for small space target orbits

Abstract

The present invention discloses a hybrid prediction method for the orbit of a small space target, comprising the following steps: Step 1: constructing an orbit prediction error dataset; Step 2: normalizing the raw feature data in the orbit prediction error dataset; Step 3: constructing and training an error prediction model; Step 4: setting the training set to a time series dataset with a step size of 5; Step 5: extracting time series features using a GRU neural network using the time series dataset as input; Step 6: combining the time series features extracted by the GRU neural network with the raw data features, and inputting the combined data into an XGBoost regression model to output the prediction results. This method uses advanced machine learning methods to mine its orbit error patterns to correct the predicted orbit, combining physical interpretability with flexibility to improve generalization ability, reduce manual intervention, and achieve high accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite orbit prediction, and in particular to a hybrid prediction method for micro space target orbits. Background Art

[0002] Leveraging technical support from the U.S. global observing network, a catalog of space objects based on the SGP4 / SDP4 orbit prediction model has been successfully developed and maintained. This database, currently a relatively comprehensive and detailed data source, primarily releases data in a two-line element (TLE) format. The SGP4 / SDP4 model considers only the primary perturbations of Earth's gravity and atmospheric drag on orbits, including a second-order solution for Earth's gravity and a simplified solution for atmospheric drag. It provides analytical state solutions to the equations of motion for arbitrary time periods without increasing computational burden. However, compared to the increasing accuracy demands of orbit predictions, SGP4's modeling of the effects of orbital perturbations is unsatisfactory. For example, the TLE prediction error for the Larets satellite can reach 8 kilometers within 7 days. Collision analysis based on inaccurate orbit predictions can result in false alarms or even missed collision alerts.

[0003] To improve orbit prediction accuracy, researchers at home and abroad have conducted extensive research based on orbital dynamics methods. Levit and Marshall proposed a method for fitting TLE orbits into an orbit estimator based on a high-fidelity mechanical model, while Bennett et al. used estimated biases to correct TLE orbits. Li and Sang employed a least-squares difference correction algorithm to resolve precise orbits. San-Juan et al. developed a hybrid SGP4 (HSGP4) propagator.

[0004] Machine learning (ML) methods offer new insights into orbit prediction. By learning patterns from historical data to predict new data, they have achieved remarkable results in space applications. For example, ML algorithms are being used for automated space object characterization, trajectory planning and optimization, and orbit determination. These applications demonstrate the enormous potential of ML in the field of orbital dynamics.

[0005] However, methods based on orbital dynamics consume enormous computational and time costs, are often computationally inefficient, and are limited in their ability to predict the orbits of cataloged objects, particularly space debris without precise ephemeris. Purely using machine learning for orbit prediction also suffers from shortcomings such as poor interpretability and insufficient generalization. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention proposes a hybrid prediction method for the orbits of small space targets. In this framework, the machine learning model can be regarded as an error corrector for the orbit prediction based on SGP4. Therefore, if the machine learning model captures most of the error patterns, it can achieve accurate prediction of the orbit of space objects over a long period of time.

[0007] In order to solve the above technical problems, the technical solution of the present invention is:

[0008] A hybrid prediction method for small space target orbits includes the following steps:

[0009] Step 1: Construct an orbit prediction error dataset;

[0010] Step 2: Normalize the original feature data in the orbit prediction error dataset, and divide the normalized dataset into a training set and a test set in proportion;

[0011] Step 3: Build and train an error prediction model, which includes a GRU neural network and an XGBoost regression model;

[0012] Step 4: Set the training set to a time series dataset with a step size of 5;

[0013] Step 5: Use the time series dataset as input to extract time series features through the GRU neural network;

[0014] Step 6: Combine the time series features extracted by the GRU neural network with the original data features, and input them into the XGBoost regression model to output the prediction results.

[0015] Preferably, step 1 includes the following sub-steps:

[0016] Step 1.1, acquiring multi-source data, wherein the multi-source data includes high-precision radar data, laser ranging data, optical telescope images, and satellite payload detection data;

[0017] Step 1.2: Multi-source data preprocessing;

[0018] Step 1.3: Use the orbital data obtained by fusion of multi-source data as the state vector The OP error of the TLE data can be calculated by using the approximate value of

[0019] .

[0020] Preferably, in step 1, the SP3 data of the global navigation satellite system is obtained, and the position information of the SP3 is used as the state vector The OP error of the TLE data can be calculated by using the approximate value of

[0021] ;

[0022] in, The true error, is the real TLE-based OP error; It's a real track.

[0023] Preferably, the multi-source data preprocessing method is:

[0024] For high-precision radar data, effective echo signals are extracted through denoising filtering;

[0025] For laser ranging data, outliers are removed by calibrating time and position;

[0026] For optical telescope images, image processing is performed through denoising, enhancement and feature extraction;

[0027] For satellite payload data, multi-source sensor data is collated and preliminary analysis is performed.

[0028] Preferably, the multi-source data preprocessing method further includes data calibration, and the data calibration includes:

[0029] Time calibration, calibrate the timestamps of all data;

[0030] Spatial calibration, converting all data to the ECEF coordinate system;

[0031] Sensor calibration, calibrating the measurement errors and biases of multi-source sensors.

[0032] Preferably, in step 1.4, data fusion is performed through Kalman filtering.

[0033] Preferably, the data fusion method is:

[0034] First, define the state vector , contains the orbital parameters of the fragment:

[0035] ;

[0036] Define the observation vector , containing the sensor's measurements:

[0037] ;

[0038] Assume that the system moves with constant acceleration in the time interval Δt, describe the system from time At the time The state transition matrix of the state change , the expression is as follows:

[0039] ;

[0040] Through the state vector and observation vector to the observation matrix , the expression is as follows:

[0041] ;

[0042] By process noise covariance matrix It reflects the uncertainty or process noise of the satellite navigation system. The commonly used expression method is:

[0043] ;

[0044] By observing the noise covariance matrix It reflects the noise characteristics in the observation data and is set according to the accuracy and noise characteristics of the sensor:

[0045] ;

[0046] During initialization, the initial state estimate x 0|0 and the initial error covariance P 0|0 .

[0047] At the prediction step the predicted state is:

[0048] ;

[0049] Forecast error covariance:

[0050] ;

[0051] The Kalman gain is calculated during the update step:

[0052] ;

[0053] Update the state estimate:

[0054] ;

[0055] Update error covariance:

[0056] .

[0057] Preferably, the fused multi-source data includes orbital propagation time, x, y, z coordinate prediction values ​​in the ECEF coordinate system, velocity prediction value and position error.

[0058] Preferably, the GRU neural network includes two GRU layers, two Dropout layers and a final fully connected layer, wherein the fully connected layer uses a linear activation function, and L2 regularization is introduced into the GRU neural network.

[0059] Preferably, the training method of the error prediction model is:

[0060] The GRU neural network is compiled using the adam optimizer and the mean square error loss function;

[0061] The XGBoost regression model randomly selects parameter combinations from a specified parameter distribution through random search, performs a specified number of iterations, uses a different parameter combination to train the model in each iteration, and selects the best parameters based on the results of cross-validation. The parameters include maximum depth, learning rate, subsample ratio, and L1 and L2 regularization terms.

[0062] Preferably, the evaluation method of the error prediction model is:

[0063] The x-coordinate error, y-coordinate error, and z-coordinate error in the test set are trained separately, and the prediction is performed after the optimal parameters are determined. The fitting degree R is used. 2 The mean absolute percentage error (MAPE) is used as the evaluation index to evaluate the prediction results of the model. 2 The closer the value is to 1, the smaller the MAPE value is, which means the prediction result is more accurate.

[0064] The present invention has the following characteristics and beneficial effects:

[0065] Unlike traditional methods based purely on orbital dynamics or machine learning, this approach integrates advanced machine learning into the standard TLE / SGP4 system, mining its orbital error patterns to correct the predicted orbit. Combining physical interpretability with flexibility, this approach improves generalization, reduces manual intervention, and achieves high accuracy.

[0066] Existing machine learning methods often use CPF ephemeris to determine the actual orbital position, with an accuracy of up to the meter level. This solution uses SP3 precise ephemeris to determine the actual orbital position of navigation satellites, with an accuracy of up to the centimeter level.

[0067] A comprehensive multi-source data fusion framework was proposed to integrate high-precision radar data, laser ranging data, optical telescope imagery, and satellite payload detection data. This framework solves the information silo problem in traditional data fusion methods and improves the accuracy and reliability of orbit determination for small space debris.

[0068] Improving orbit prediction accuracy: By introducing the GRU-XGBoost hybrid model to mine the error patterns of the TLE / SGP4 system, the accuracy of orbit prediction is significantly improved, solving the problem of insufficient accuracy of the traditional SGP4 model in long-term predictions.

[0069] Improve response speed: The hybrid model has relatively high computational efficiency and can achieve real-time early warning and avoidance functions to meet the needs of rapid response. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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 any creative work.

[0071] Figure 1 2 is a structural diagram of a hybrid model in an embodiment of the present invention.

[0072] Figure 2 Schematic diagram of the x-coordinate error prediction effect of satellite USA132 after applying an embodiment of the present invention.

[0073] Figure 3 Schematic diagram of the y-coordinate error prediction effect of satellite USA132 after applying an embodiment of the present invention.

[0074] Figure 4 Schematic diagram of the z-coordinate error prediction effect of satellite USA132 after applying an embodiment of the present invention.

[0075] Figure 5 Schematic diagram of the x-coordinate error prediction effect of satellite USA293 after applying an embodiment of the present invention.

[0076] Figure 6 The figure is a schematic diagram of the y-coordinate error prediction effect of the satellite USA293 after applying the embodiment of the present invention.

[0077] Figure 7 Schematic diagram of the z-coordinate error prediction effect of satellite USA293 after applying an embodiment of the present invention.

[0078] Figure 8 Schematic diagram of the x-coordinate error prediction effect of satellite USA343 after applying an embodiment of the present invention.

[0079] Figure 9 The figure is a schematic diagram of the y-coordinate error prediction effect of the satellite USA343 after applying the embodiment of the present invention.

[0080] Figure 10Schematic diagram of the z-coordinate error prediction effect of satellite USA343 after applying an embodiment of the present invention.

[0081] Figure 11 Schematic diagram of the final orbit prediction error of satellite USA132 after applying an embodiment of the present invention.

[0082] Figure 12 Schematic diagram of the final orbit prediction error of satellite USA293 after applying an embodiment of the present invention.

[0083] Figure 13 Schematic diagram of the final orbit prediction error of satellite USA343 after applying an embodiment of the present invention.

[0084] Figure 14 Schematic diagram comparing the x-coordinate error residual rates of three models in an embodiment of the present invention.

[0085] Figure 15 Schematic diagram comparing the y-coordinate error residual rates of three models in an embodiment of the present invention.

[0086] Figure 16 Schematic diagram comparing the z-coordinate error residual rates of three models in an embodiment of the present invention. DETAILED DESCRIPTION

[0087] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0088] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0089] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0090] Example 1

[0091] The present invention provides a hybrid prediction method for small space object orbits. This method, in this embodiment, aims to integrate advanced machine learning methods into a standard TLE / SGP4 system. The machine learning method uses a GRU-XGBoost hybrid model to learn orbit prediction error patterns from historical TLE data. This model is used to generate orbit prediction errors for orbits predicted based on the latest TLE data, improving their accuracy through error compensation or correction. In this framework, the machine learning model can be viewed as an error corrector for SGP4-based orbit predictions. Therefore, if the machine learning model captures most of the error patterns, it can accurately predict the orbits of space objects over long periods of time.

[0092] The specific steps are as follows:

[0093] Step 1: Construction of orbit prediction error dataset

[0094] It should be noted that since the disturbance force considered in the SGP4 propagator is relatively simplified, Usually contains large errors. Therefore, The true error can be measured by the following formula:

[0095] (1)

[0096] in, is the real TLE-based OP error; Is the true or exact track.

[0097] In fact, the state vector of a space object The true value of can never be known exactly and can only be estimated through precise orbital determination.

[0098] Specifically, the following sub-steps are included:

[0099] Step 1.1, acquiring multi-source data, wherein the multi-source data includes high-precision radar data, laser ranging data, optical telescope images, and satellite payload detection data;

[0100] Step 1.2: Multi-source data preprocessing;

[0101] Step 1.3: Use the orbital data obtained by fusion of multi-source data as the state vector The OP error of the TLE data can be calculated by using the approximate value of

[0102] ;

[0103] in, The true error, is the actual OP error based on TLE.

[0104] It is understandable that for satellites and space debris that lack precise ephemeris, especially tiny space debris, precise orbit determination can be achieved through multi-source data fusion.

[0105] In this embodiment, it is proposed to use high-precision radar data, laser ranging data, optical telescope images and satellite payload detection data to achieve this goal.

[0106] Specifically, we first collect relevant data from various sensors and platforms:

[0107] Use high-precision radar to collect information such as distance, speed, and angle.

[0108] Apply laser ranging to obtain precise distance and time-stamp data.

[0109] Optical telescopes are used to obtain image data and analyze the shape and position of the debris.

[0110] Satellite payload detection is used to collect multi-band information, orbital data and other relevant measurements.

[0111] The collected multi-source data is then preprocessed for subsequent processing and fusion:

[0112] For high-precision radar data, effective echo signals are extracted through denoising filtering;

[0113] For laser ranging data, outliers are removed by calibrating time and position;

[0114] For optical telescope images, image processing is performed through denoising, enhancement and feature extraction;

[0115] For satellite payload data, multi-source sensor data is collated and preliminary analysis is performed.

[0116] It can be understood that multi-source sensor data refers to data obtained by different sensors on different satellites that can assist in locating a satellite or space debris, such as infrared imaging data and lidar data. Preliminary analysis also refers to data preprocessing such as screening out outliers.

[0117] Then calibrate and register the preprocessed multi-source data:

[0118] Time alignment ensures that all data has consistent timestamps.

[0119] The data were converted to the ECEF coordinate system through spatial calibration.

[0120] Through sensor calibration, the measurement error and deviation of each sensor are calibrated.

[0121] Finally, data fusion is performed:

[0122] Specifically, in this embodiment, Kalman filtering is used for data fusion to integrate multi-source data and obtain the most accurate debris orbit information. The following are detailed application steps:

[0123] Define the state vector , contains the orbital parameters of the fragment:

[0124] (2)

[0125] in, Refers to the x, y, z coordinates in the ecef coordinate system, Refers to x,y,z velocity;

[0126] Define the observation vector , containing the sensor's measurements:

[0127] (3)

[0128] in, Refers to the radar range value, Refers to the angle of the target relative to the sensor; Refers to the vertical angle of the target relative to the sensor (the sensor is radar or laser); Refers to the laser ranging value.

[0129] State transition matrix Describe the system from time At the time The state change of , assuming that the system moves with constant acceleration in the time interval Δt:

[0130] (4)

[0131] Observation matrix Describe the relationship between the state vector and the observation vector:

[0132] (5)

[0133] Process noise covariance matrix It reflects the uncertainty or process noise in the system model. The commonly used expression method is:

[0134] (6)

[0135] in, Is the noise input matrix, which defines the degree of influence of process noise on the system state. If a state variable is very sensitive to process noise, then The corresponding element value in may be larger. It is at the moment The process noise covariance matrix, is a fixed process noise covariance matrix, which is used to describe the statistical characteristics of the process noise itself.

[0136] Observation noise covariance matrix It reflects the noise characteristics in the observation data and is set according to the accuracy and noise characteristics of the sensor:

[0137] (7)

[0138] Initial state estimate x at initialization 0|0 and the initial error covariance P 0|0 ,

[0139] It's important to note that these two estimates are self-determined. The initial state estimate is a preliminary estimate of the system state at time t = 0, typically based on the measured data at that initial time (since the exact data is unknown). The initial error covariance is typically set as a diagonal matrix, with the values ​​on the main diagonal reflecting the initial uncertainty of each state variable. For example, the uncertainty in position can be small, while the uncertainty in velocity can be large. This provides the starting point for the Kalman filtering process. The filter's state updates and predictions are based on this initial estimate, which directly affects the filter's convergence speed and the accuracy of the final estimate.

[0140] At the prediction step the predicted state is:

[0141] (8)

[0142] Forecast error covariance:

[0143] (9)

[0144] The Kalman gain is calculated during the update step:

[0145] (10)

[0146] Update the state estimate:

[0147] (11)

[0148] Update error covariance:

[0149] (12)

[0150] It should be noted that the error covariance is just a parameter in the Kalman filter method, which provides new data for the next time step. (Kalman filtering is a progressive process, and the data of the next time step is updated based on the previous time step.)

[0151] It can be understood that orbit determination and prediction are performed based on the fused data. Through these steps, precise orbit determination of space targets can be achieved.

[0152] Step 2: Data preprocessing

[0153] It should be noted that in order to eliminate the influence of different dimensions and feature scales on model training, the fused data needs to be normalized. The normalization process uses the standardization method, and the formula is:

[0154] (13)

[0155] in, is the original feature data, is the mean of the data, is the standard deviation of the data.

[0156] In this embodiment, the data set is divided into a training set and a test set in a ratio of 4:1. The data in the training set is used to train and learn the prediction model, and the data in the test set is then put into the trained model for testing.

[0157] Step 3: Error prediction model construction

[0158] Specifically, such as Figure 1 As shown, in this embodiment, a hybrid model of a GRU neural network and an XGBoost regression model is used to predict the orbit propagation error, taking the orbit propagation time, the predicted x, y, and z coordinates in the ECEF coordinate system, and the predicted velocity as network inputs. This hybrid model fully utilizes the advantages of the GRU neural network in processing time series data and the efficiency of the XGBoost regression model in handling nonlinear problems.

[0159] Furthermore, the construction process of the hybrid model is introduced in detail.

[0160] Time series construction

[0161] It should be noted that, considering that the orbit error prediction is closely related to the past orbit state, the input data is constructed as a time series dataset. In this example, the time series step size is set to 5, that is, each sample contains data from the past 5 time steps, which is used to capture the temporal dependencies in the historical data.

[0162] Feature extraction

[0163] The normalized time series is fed into a GRU model to extract time series features. This GRU model consists of two GRU layers and two Dropout layers to prevent overfitting. The final layer is a fully connected layer using a linear activation function, and L2 regularization is introduced to further enhance the model's generalization capabilities. The model is compiled using the Adam optimizer and the mean squared error loss function.

[0164] Position error prediction

[0165] The features extracted by the GRU are combined with the original features and fed into the XGBoost regression model for training. During the regression model training process, a random parameter combination is randomly selected from the specified parameter distribution through random search. A specified number of iterations are performed, and a different parameter combination is used to train the model in each iteration. The optimal parameters such as maximum depth, learning rate, subsample ratio, and L1 and L2 regularization terms are selected based on the results of cross-validation (CV).

[0166] Output

[0167] The predicted data is denormalized and saved with the corresponding feature output.

[0168] In order to ensure the accuracy of the prediction results output by the above technical solution, this embodiment evaluates the constructed error prediction model.

[0169] Use the test set to train the model for the x-coordinate error, y-coordinate error, and z-coordinate error, and then make predictions after determining the optimal parameters. 2 The mean absolute percentage error (MAPE) is used as the evaluation index to evaluate the prediction results of the model. 2 The closer the value is to 1, the smaller the MAPE value is, indicating that the prediction result is more accurate. The expressions of the evaluation indicators are:

[0170] (14)

[0171] (15)

[0172] Where n is the total number of predictions, and are the true value and predicted value of the position error of the i-th sample point respectively; Represents the average value of the true value of the position error.

[0173] Example 2

[0174] The difference between this embodiment and embodiment 1 is that in step 1, SP3 is available data for precise orbit determination of satellites in global navigation satellite systems such as GPS, GLONASS, Galileo, BeiDou, etc.

[0175] SP3 data is a precise orbit and clock data format provided by the International GNSS Service of the International Earth Rotation and Reference System Service and other organizations. The data file in SP3 format contains the position, velocity and related metadata of the satellite at a specific point in time, such as time information, orbit type, coordinate reference system, etc. The data of the SP3 file is recorded at a certain time interval, generally 15 minutes or 5 minutes, and its data accuracy is at the centimeter level, while the accuracy of the SGP4 model is generally considered to be between hundreds of meters and kilometers. There is an order of magnitude difference in accuracy between the two. Therefore, in this embodiment, GPS satellites are selected as the research object, and the position information provided by SP3 data is used as an approximation of its state vector. According to Formula 1, the OP error of TLE data can be approximately obtained by the following formula:

[0176] Therefore, using TLE data at a specific moment in time with the SGP4 model for backward propagation, we obtain target coordinate and velocity predictions for subsequent moments. These coordinate results are then compared with the SP3 ephemeris at the corresponding moment to obtain the coordinate error after orbit propagation, thereby constructing the dataset required for model training. Because the reference frames of TLE / SGP4 and SP3 differ in actual calculations, in this embodiment, the position information obtained from both is uniformly converted to the ECEF reference frame for calculation.

[0177] Example 3

[0178] In this embodiment, the last TLE data of three GPS satellites on March 1, 2024 is propagated backward for 30 days, with one data recorded every 5 minutes. The position error data of the satellites from 00:00:00 on March 1, 2024 UTC to 24:00:00 on March 30, 2024 UTC for 30 consecutive days are used to predict the position error using the error prediction model pre-trained in Example 1. The results are as follows: Figure 2-Figure 10 As shown in Figure 2, it can be seen that the predicted values ​​of the x, y, and z coordinate errors of the three satellites have the same overall change trend compared with the actual values.

[0179] Based on the orbit predictions of the TLE / SGP4 system, the orbits were corrected using the predictions of the pre-trained error prediction model and compared with the precise orbit data (SP3 data). Finally, the orbit prediction errors of the three satellites within 30 days were as follows: Figure 11-13 As shown, the errors of the x, y, and z coordinates do not exceed 200 meters.

[0180] Comparative Example

[0181] To further verify the scientific nature and superiority of the hybrid model proposed in this article in terms of overall prediction accuracy, the prediction results of the prediction regression model (GRU-XGBoost regression model) in Example 1 are compared with the prediction results of the GRU model and XGBoost regression model after parameter tuning as follows:

[0182] Table 1 Comparison of prediction results of various models

[0183]

[0184] From Table 1, it can be seen that the trajectory error prediction accuracy of the hybrid model used in Example 1 is higher. For the error prediction of x, y, and z coordinates, R 2 The values ​​are closer to 1 than those of the other two methods, indicating a better fit; the MAPE is also significantly lower than that of the other two methods. The hybrid model significantly improves on both accuracy indicators compared to the single model, indicating that the hybrid model has good accuracy.

[0185] pass Figure 14-16 Taking the USA 132 data as an example, the residual rates of the three models on each sample are shown, which can intuitively compare the stability of the three models in error prediction. As can be seen from the figure, for each sample, the residual rate of the hybrid model is more concentrated near 0 compared with the other two models, indicating that the stability of the prediction error of this model is also more advantageous than that of the other two models.

[0186] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It will be apparent to those skilled in the art that various changes, modifications, substitutions, and variations of these embodiments, including components, without departing from the principles and spirit of the present invention are still within the scope of protection of the present invention.

Claims

1. A hybrid prediction method for small space target orbits, characterized by: The steps include: Step 1: Construct an orbit prediction error dataset; The following sub-steps are included: Step 1.1, acquiring multi-source data, wherein the multi-source data includes high-precision radar data, laser ranging data, optical telescope images, and satellite payload detection data; Step 1.2: Multi-source data preprocessing; Step 1.3: Use the orbital data obtained by fusion of multi-source data as the state vector The OP error of the TLE data can be calculated by using the approximate value of Among them, X TLE The true error of (t; t0), is a true TLE-based OP error; the multi-source data fusion method is to perform data fusion through Kalman filtering; the fused multi-source data includes the orbit propagation time, the x, y, z coordinate prediction values ​​in the ECEF coordinate system, the velocity prediction value and the position error; Step 2: Normalize the original feature data in the orbit prediction error dataset, and divide the normalized dataset into a training set and a test set in proportion; Step 3: Build and train an error prediction model, which includes a GRU neural network and an XGBoost regression model; the GRU neural network includes two GRU layers, two Dropout layers, and a fully connected layer at the end, the fully connected layer uses a linear activation function, and L2 regularization is introduced in the GRU neural network; The training method of the error prediction model is: The GRU neural network is compiled using the adam optimizer and the mean square error loss function; The XGBoost regression model randomly selects parameter combinations from a specified parameter distribution through random search, performs a specified number of iterations, uses a different parameter combination to train the model in each iteration, and selects the best parameters based on the results of cross-validation, which include maximum depth, learning rate, subsample ratio, and L1 regularization term and L2 regularization term; Step 4: Set the training set to a time series dataset with a step size of 5; Step 5: Use the time series dataset as input to extract time series features through the GRU neural network; Step 6: Combine the time series features extracted by the GRU neural network with the original data features, and input them into the XGBoost regression model to output the prediction results.

2. A hybrid prediction method for micro space target orbits according to claim 1, characterized in that: In step 1, the SP3 data of the global navigation satellite system is obtained, and the position information of the SP3 is used as the state vector The OP error of the TLE data can be calculated by using the approximate value of Among them, X TLE The true error of (t; t0), is the real TLE-based OP error; It's a real track.

3. The hybrid prediction method for micro space target orbits according to claim 1, characterized in that: The preprocessing method of the multi-source data is: For high-precision radar data, effective echo signals are extracted through denoising filtering; For laser ranging data, outliers are removed by calibrating time and position; For optical telescope images, image processing is performed through denoising, enhancement and feature extraction; For satellite payload data, multi-source sensor data is collated and preliminary analysis is performed.

4. A hybrid prediction method for micro space target orbits according to claim 3, characterized in that: The multi-source data preprocessing method further includes data calibration, which includes: Time calibration, calibrate the timestamps of all data; Spatial calibration, converting all data to the ECEF coordinate system; Sensor calibration, calibrating the measurement errors and biases of multi-source sensors.

5. A hybrid prediction method for micro space target orbits according to claim 4, characterized in that: The data fusion method is: First, define the state vector x k , contains the orbital parameters of the fragment: Among them, x k ,y k ,z k Refers to the x, y, z coordinates in the ecef coordinate system, Refers to x, y, z speed; Define the observation vector z k , containing the sensor's measurements: Among them, R k Refers to the radar range value, θ k Refers to the angle of the target relative to the sensor; φ k Refers to the vertical angle of the target relative to the sensor; L k Refers to the laser ranging value. Assuming that the system moves with constant acceleration within the time interval Δt, the state transfer matrix F describing the state change of the system from time k-1 to time k is k , the expression is as follows: Through the state vector and the observation vector to the observation matrix H k , the expression is as follows: Through the process noise covariance matrix Q k It reflects the uncertainty or process noise of the satellite navigation system. The commonly used expression method is: Among them, G k is the noise input matrix, Q v is the covariance matrix of the process noise, Q k is the process noise covariance matrix at time k; By observing the noise covariance matrix R k It reflects the noise characteristics in the observation data and is set according to the accuracy and noise characteristics of the sensor: During initialization, the initial state estimate x 0|0 and the initial error covariance P 0|0 , At the prediction step the predicted state is: x k|k-1 =F k-1 x k-1|k-1 +B k-1 u k-1 ; Forecast error covariance: The Kalman gain is calculated during the update step: Update the state estimate: x k|k =x k|k-1 +K k (z k -H k x k|k-1 ); Update error covariance: P k|k =(I-K k H k )P k|k-1 。 6. A hybrid prediction method for micro space target orbits according to any one of claims 1 to 5, characterized in that: The evaluation method of the error prediction model is: The x-coordinate error, y-coordinate error, and z-coordinate error in the test set are trained separately, and the prediction is performed after the optimal parameters are determined. The fitting degree R is used. 2 The mean absolute percentage error (MAPE) is used as the evaluation index to evaluate the prediction results of the model. 2 The closer the value is to 1, the smaller the MAPE value is, which means the prediction result is more accurate.

Images