Space debris orbit prediction correction method based on machine learning and triangular feature decomposition

Abstract

The present application relates to the technical field of space debris orbit prediction, in particular to a space debris orbit prediction correction method based on machine learning and triangular feature decomposition; the azimuth angle and the pitch angle are converted into a six-dimensional feature vector through triangular feature decomposition, and a real-time mapping model of orbit prediction deviation is established in combination with a neural network, thereby solving two bottleneck problems of traditional image tracking dependence: 1) in a strong laser interference scene, tracking failure of an optical camera caused by background light saturation is avoided, and deviation prediction driven by historical echo data is used to replace real-time image feedback; 2) when space debris is invisible in the Earth's shadow area, a training set is constructed by using deviation data captured in a successful laser ranging moment, and autonomous correction of prediction error under no-light conditions is realized; the method is free from dependence on optical visibility of space debris, realizes telescope pointing deviation correction in all time and all arc segments, and improves the reliability and continuity of space debris orbit prediction under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of space debris orbit prediction technology, specifically a space debris orbit prediction correction method based on machine learning and triangular feature decomposition. Background Technology

[0002] In recent years, the increasing space activities of major powers have led to a rise in the amount of space debris, seriously threatening the safety of spacecraft in orbit. Space debris monitoring and early warning have become core issues in the international space safety field. The average relative velocity of space debris in low Earth orbit is about 10 km / s, which means that even very small debris can generate huge collision kinetic energy. Millimeter-sized debris impacts can cause satellite malfunctions, while centimeter-sized debris can destroy a satellite. Since its operation, the solar panels of China's space station have been hit by space debris multiple times, resulting in partial loss of power supply. Therefore, avoiding collisions, reducing the number of debris, and clearing existing debris are of great significance for protecting the safe operation of spacecraft. The realization of these technologies all rely on high-precision orbital data. Laser ranging, as the space debris detection technology with the highest single-station measurement accuracy to date, plays an important role in high-precision orbit determination and motion status determination of space debris, and is an important guarantee for the safe operation of spacecraft in orbit.

[0003] The key to successfully achieving laser ranging of space debris lies in accurately illuminating the space debris with a laser and capturing its reflected light. Due to errors in space debris trajectory prediction and the potential for changes in the debris's attitude, in most cases, relying solely on predicted coordinates is insufficient to directly place the debris within the receiving field of view (FOV) of the laser ranging system. Therefore, when performing laser ranging of space debris, it is necessary to continuously and in real-time correct the telescope's pointing direction, i.e., adjust the telescope's azimuth error (Ae) and elevation error (Ee), to guide the space debris into and stabilize it within the effective detection field of view of the laser ranging system.

[0004] When space debris is visible, automatic closed-loop tracking can be achieved through a telescope, which automatically fine-tunes the telescope's pointing based on real-time image feedback. However, this real-time correction method faces two significant challenges: 1. Strong laser interference can easily cause automatic tracking failure: When the ranging laser emits high-power (532nm wavelength) pulses, the resulting strong background light can severely interfere with the optical guide camera, making it difficult for the telescope's automatic closed-loop tracking system to operate normally; 2. Space debris is invisible in the Earth's shadow region, resulting in a lack of deviation information: When space debris enters the Earth's shadow region, due to the lack of sunlight and the fact that the optical guide camera relies on sunlight reflected from the space debris for imaging, the optical guide camera ultimately cannot detect the space debris. At this time, neither the closed-loop tracking system nor the operator has direct visual evidence to judge the angular deviation (i.e., azimuth and elevation angle deviation) between the telescope's pointing and the actual position of the debris. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a space debris orbit prediction correction method based on machine learning and trigonometric feature decomposition. This method solves the problems of automatic tracking failure caused by strong laser interference and loss of deviation information due to the invisibility of space debris in the Earth's shadow region when using real-time image feedback to automatically fine-tune telescope pointing to track predicted space debris orbits.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a space debris orbit prediction correction method based on triangular feature decomposition, which specifically includes the following steps: The telescope ranging and tracking system tracks space debris and simultaneously records the measured azimuth and elevation angles at the time of emission when laser echo signals are present, as well as the telescope azimuth and elevation correction values, to obtain observation data. The telescope azimuth and elevation correction values ​​include azimuth deviation and elevation deviation. The observation data is preprocessed, and the azimuth and elevation angles in the preprocessed observation data are decomposed into trigonometric features to obtain sample data. A neural network model is constructed by dividing the sample data into a training set and a validation set as the input to the neural network model, and using the telescope azimuth and elevation correction values ​​at the corresponding time as the model output to train the neural network model. The azimuth and elevation angles acquired in real time by the telescope ranging and tracking system are decomposed into trigonometric features and then input into the trained neural network model to obtain the telescope azimuth and elevation correction prediction values ​​that are adjusted in real time to adjust the telescope's pointing direction.

[0007] Preferably, the telescope azimuth and elevation correction values ​​include azimuth deviation and elevation deviation.

[0008] Preferably, the sample data includes the original azimuth angle value, the azimuth angle sine value, the azimuth angle cosine value, the original pitch angle value, the pitch angle sine value, and the pitch angle cosine value.

[0009] Preferably, the telescope ranging and tracking system tracks space debris based on two rows of element orbit data.

[0010] Preferably, the structure of the neural network model includes an input layer, a hidden layer, and an output layer; The input layer has a dimension of 6, corresponding to the azimuth angle, pitch angle and their sine and cosine values, and maps the 6-dimensional input features to a 60-dimensional hidden space. The hidden layer contains at least two residual blocks, and each residual block performs the following steps sequentially during each training iteration: Save the 60-dimensional features of the current output of the input layer as the identity residual; A 60→60-dimensional linear transformation is performed through a fully connected layer with shared parameters; The result of the linear transformation is added to the identity residual; Use the ReLU activation function; The output layer is used to output the azimuth prediction error and the pitch prediction error.

[0011] Set the maximum number of iterations and the loss function, use the Adam optimizer to set the initial model parameters, and then train the neural network model; Set the iteration termination condition and initial learning rate, and repeat the forward propagation, loss calculation, back propagation, model parameter update and learning rate adjustment in sequence before the neural network model reaches the iteration termination condition after training times; After every few iterations of training, the obtained model is evaluated using a validation set, and the loss of the validation set is calculated. The iteration termination condition is one of two conditions: the training number of the neural network model reaches the maximum number of iterations and the loss of the validation set does not decrease after several consecutive calculations. After the neural network model reaches the termination condition of the iteration, the current neural network model is output. The trained neural network model is evaluated using a validation set. If the final neural network model passes the evaluation, it is saved and output; otherwise, the training is repeated.

[0012] Preferably, the meaning of "the loss of the validation set has not decreased" is that the difference between the loss of the validation set calculated in several consecutive calculations and the loss of the validation set calculated in the first calculation in those calculations is greater than the loss threshold.

[0013] The technical solution also provides a system for implementing a space debris orbit prediction correction method based on triangular feature decomposition, including a processor and a memory. The memory is used to store a computer program, which, when executed by the processor, implements the space debris orbit prediction correction method.

[0014] Compared with existing technologies, this invention provides a space debris orbit prediction and correction method based on machine learning and triangular feature decomposition, which has the following beneficial effects: 1. This invention transforms azimuth / elevation angles into six-dimensional feature vectors through trigonometric eigenvalue decomposition and establishes a real-time mapping model for orbit prediction deviations using a neural network. This effectively solves two major bottleneck problems inherent in traditional image tracking: 1) In scenarios with strong laser interference, it avoids tracking failure caused by background light saturation of optical cameras, replacing real-time image feedback with deviation prediction driven by historical echo data; 2) When space debris is invisible in Earth's shadow region, it utilizes deviation data successfully captured instantaneously by laser ranging to construct a training set, enabling autonomous correction of prediction errors under no-light conditions. This method eliminates the dependence on the optical visibility of space debris, achieving all-day, all-arc telescope pointing deviation correction, significantly improving the reliability and continuity of space debris orbit predictions under complex conditions.

[0015] 2. This invention fundamentally eliminates the prediction inaccuracy problem caused by the 0°-360° angle boundary jump by decomposing the azimuth and pitch angles into a six-dimensional feature vector of original values, sine values, and cosine values. Traditional models will produce nonlinear errors in the boundary region due to periodic jumps, while the features after triangular decomposition can continuously express the spatial distribution of angles (such as the cosine values ​​of 359° and 1° are close), enabling the neural network to learn more accurately the complex mapping relationship between azimuth and pitch deviations in the horizontal coordinate system, especially improving the deviation prediction accuracy in perturbation-sensitive areas such as perigee and apogee.

[0016] 3. The lightweight residual network design constructed in this invention combines high accuracy and real-time performance: 1) The batch normalization layer solves the training oscillation caused by the difference in the dimensions of astronomical observation data and accelerates model convergence; 2) The residual blocks enhance gradient propagation through skip connections, effectively fitting the strong nonlinear characteristics of orbital deviations (such as non-steady-state errors caused by atmospheric drag perturbations); 3) The overall computational latency is low, meeting the telescope's millisecond-level pointing correction requirements. Compared with the traditional SGP4 physical model, the computational efficiency is significantly improved, and there is no need to rely on complex perturbation modeling. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the space debris orbit prediction correction method based on triangular feature decomposition of the present invention; Figure 2 This is an architecture diagram of the space debris orbit prediction and correction method based on triangular feature decomposition of the present invention; Figure 3 This is a flowchart illustrating the generation of sample data for this invention. Figure 4 This is a structural diagram of the neural network model of the present invention; Figure 5 This is a flowchart of the neural network model training process of the present invention; Figure 6 This is a flowchart of the space debris orbit prediction correction scheme based on triangular feature decomposition of the present invention. Figure 7 The azimuth angle result of the model built based on the observation data of the 02144 space debris in this invention when predicting the observation data; Figure 8 The elevation angle result of the model built based on the observation data of space debris 02144 in this invention when predicting the observation data; Figure 9 The azimuth angle result of the model built based on the observation data of the 02144 space debris in this invention when predicting new observation data of the target; Figure 10 The elevation angle result of the model built based on the observation data of space debris 02144 in this invention when predicting new observation data of the target; Figure 11 This is part of the observation data of space debris 02144 of this invention. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0019] Those skilled in the art will understand that all or part of the steps in the methods of the following embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0020] To address the problems of strong laser interference causing automatic tracking failure and the invisibility of space debris in Earth's shadow causing missing bias information when using real-time image feedback to automatically fine-tune telescope pointing to track and predict space debris orbits, this invention provides a space debris orbit prediction correction method based on trigonometric feature decomposition. During the method development process, we first developed Scheme 1 as follows: Based on measured data, we improved the Two Line Element (TLE) to quickly improve the accuracy of short-term orbit prediction. Based on TLE and the Simplified General Perturbations 4 (SGP4) model, we first generated pseudo-observation data and used a high-precision orbit propagator to precisely determine the orbit of space debris, thereby improving orbit prediction. However, this technical scheme is relatively complex in design and calculation, and not easily applicable to actual observations.

[0021] Subsequently, we developed Scheme 2 as follows: We conducted research on prediction error correction from the non-orbit determination level. We used the distance deviation calculated by the off-target amount at the center of the telescope's receiving field of view of the space debris to improve the distance deviation to within 100 meters within a few minutes, thereby improving the control accuracy of the detector's gate opening time. Alternatively, we can use the apparent position angle deviation of the space debris as an observation and correct the distance prediction deviation in real time to improve the accuracy of distance gate control and the success rate of ranging. Although this scheme achieved good observation results during the dawn and dusk when space debris is visible, it is no longer applicable when the observed space debris camera is "invisible".

[0022] To address this issue, we developed a space debris orbit prediction correction method based on triangular feature decomposition. This method addresses the prediction inaccuracies in the boundary region caused by the periodic jump in azimuth angle from 0° to 360°. It constructs a six-dimensional feature vector composed of the original azimuth angle values ​​(sine, cosine) and the original elevation angle values ​​(sine, cosine). This vector is input into a neural network to establish a mapping model between azimuth and elevation errors. By correcting telescope pointing deviations, this method solves the problem of nonlinear correction of space debris orbit prediction errors in the angular spatial distribution. Figure 1 and Figure 2 As shown, the specific steps include the following: The telescope ranging and tracking system tracks space debris and simultaneously records the measured azimuth and elevation angles at the moment of laser echo signal transmission, as well as the telescope's azimuth and elevation correction values. It captures the deviation between the actual and predicted positions the instant laser ranging is successful, ensuring data validity and thus obtaining observational data. This solves the technical problem of not being able to obtain real-time deviations when space debris is invisible in the Earth's shadow. The telescope's azimuth and elevation correction values ​​generally include azimuth and elevation deviations. The methods used by the telescope ranging and tracking system to track space debris can include two-line element-based orbital data (TLE), precise ephemeris data, orbital state vectors generated by radar or optical measurement fusion, and real-time orbital data provided by commercial orbit determination services. Here, considering data accuracy and the universality of the data source, two-line element-based orbital data is chosen as the data source.

[0023] The observation data is then preprocessed. Preprocessing methods such as data cleaning and outlier filtering are used to obtain preprocessed observation data. The azimuth and elevation angles in the preprocessed observation data are then decomposed into trigonometric features to obtain sample data. The sample data includes the original azimuth value, azimuth sine, azimuth cosine, and elevation angle values, as well as the original elevation angle values, elevation sine, and elevation cosine. For each original angle value, it is decomposed into its original value, sine, and cosine values, thereby eliminating boundary jumps between 0° and 360°. Specifically, as shown below... Figure 3 As shown.

[0024] We then constructed a neural network model to establish the mapping between azimuth, elevation, and telescope azimuth and elevation correction values, solving the technical problem of traditional closed-loop tracking failure caused by strong laser interference. First, the sample data was divided into training and validation sets and used as input to the neural network model. Generally, 80% of the sample data was divided into the training set and 20% into the validation set. The telescope azimuth and elevation correction values ​​at the corresponding time were used as the model output to train the neural network model.

[0025] The structure of a neural network model is as follows Figure 4As shown, the neural network model adopts a fully connected feedforward structure, including an input layer, hidden layers, and an output layer. The input layer has a dimension of 6, corresponding to the azimuth angle, pitch angle, and their sine and cosine values; the hidden layers each have 60 neurons, and the ReLU activation function is introduced to enhance nonlinear expression; the output layer has 2 nodes, corresponding to the predicted values ​​of azimuth and pitch deviations. The first input layer receives the input data and can be a fully connected layer with an input dimension of 6. It transforms the 6-dimensional features into several output nodes, for example, setting the output dimension to 60. The activation function can be the ReLU function to enhance non-linear expressiveness, ultimately mapping the 6-dimensional input features to a 60-dimensional hidden space. The second hidden layer includes at least two residual blocks, each containing a linear layer and skip connections. The design of the residual blocks allows the model to better learn complex patterns and avoids gradient vanishing. It also improves the fitting ability of complex patterns to address the strong non-linear characteristics of angular deviations (such as different near-field / far-field error patterns). Each residual block executes the following process in each iteration of training: 1. Save the 60-dimensional features of the current output of the input layer as the identity residual; 2. Perform a 60→60-dimensional linear transformation through a parameter-sharing fully connected layer; 3. Execute the residual connection: add the linear transformation result to the identity residual; 4. Apply the ReLU activation function. The same set of linear layer parameters is used in both loops to achieve efficient parameter utilization. The third output layer can also be a fully connected layer with an input dimension of 60 and an output dimension of 2, which includes both azimuth and pitch prediction errors, and can be linearly output. During the training of this neural network model, the L1 loss function can be used to calculate the average absolute error between the predicted and true values, and the Adam optimization algorithm can be employed.

[0026] The training process of a neural network model is as follows: Figure 5 As shown, it includes: First, you need to configure the parameters required for model training, such as setting the maximum number of iterations, using the L1 loss function, using the ReduceLROnPlateau scheduler based on validation loss, and setting the initial model parameters using the Adam optimizer. The initial learning rate can be set to 10. -4 Then, the neural network model is trained. The model then enters the training loop. First, it is necessary to set the iteration termination condition and the initial learning rate. Before the iteration termination condition is reached after the training number of the neural network model is reached, the forward propagation, loss calculation, backpropagation, model parameter update, and learning rate adjustment are repeatedly executed in sequence. Forward propagation is to use the model to calculate the prediction result from the input data. Loss calculation is to calculate the "loss" based on the difference between the model's prediction result and the true label. This loss value helps us understand the model's performance. Backpropagation is to calculate the impact of each parameter on the loss based on the loss value, thereby determining how to adjust the parameters to improve the model's performance. Parameter update is to update the parameters in the model based on the backpropagation results to make the model more accurate.

[0027] After every few iterations of training, the obtained model is evaluated using a validation set, and the validation set loss is calculated. For example, every 100 epochs, the model is evaluated and validated. The validation set evaluation uses the previously divided validation set to evaluate the model's performance on unseen data.

[0028] The iteration termination condition is one of two conditions: the neural network model has reached the maximum number of iterations or the loss of the validation set has not decreased after several consecutive calculations. The meaning of "the loss of the validation set has not decreased" is that the difference between the loss of the validation set calculated in several consecutive calculations and the loss of the first validation set calculation in those calculations is greater than a loss threshold. For example, if the loss of the validation set is calculated in 5 consecutive calculations, and the difference between the loss of the subsequent four validation sets and the loss of the first validation set is within the loss threshold (e.g., -0.1), then the condition "the loss of the validation set has not decreased after several consecutive calculations" is met. Therefore, further iterations are unnecessary to reduce computational cost and prevent overfitting.

[0029] After the neural network model reaches the termination condition of iteration, the current neural network model is output.

[0030] The trained neural network model is then evaluated using a validation set. This involves inputting the validation set's sample data into the neural network model and evaluating it based on the actual azimuth and pitch deviations corresponding to the sample data and the azimuth and pitch prediction deviations output by the neural network model. This evaluation can be done, for example, by calculating a loss function. If the final neural network model passes the evaluation, it is saved and output; otherwise, the model is retrained.

[0031] After the neural network model is trained, the azimuth and elevation angles acquired in real time by the telescope ranging and tracking system are decomposed into trigonometric features and input into the trained neural network model. Regardless of whether it is in the Earth's shadow or in a strong laser environment, it can autonomously correct the orbit prediction deviation of space debris all day long. This allows for the real-time adjustment of the telescope's azimuth and elevation correction prediction values, which include both azimuth and elevation prediction deviations. Based on this data, the telescope's pointing can be continuously and in real time corrected, guiding space debris into and stabilizing it within the effective detection field of the laser ranging system. This enables laser ranging and orbit prediction of space debris, improving the accuracy of orbit prediction.

[0032] The technical solution also provides a system for implementing a space debris orbit prediction correction method based on triangular feature decomposition, including a processor and a memory. The memory is used to store computer programs, and when the computer programs are executed by the processor, they implement the space debris orbit prediction correction method.

[0033] In summary, when this invention is used in practice, as... Figure 6 As shown, a large amount of laser ranging data of space debris during the visible time period at dawn and dusk is obtained using a space debris laser ranging system to form a ranging dataset as sample data. Then, the ranging dataset is divided into a training dataset and a test dataset. A model is trained using a machine learning algorithm on the training dataset, and then the trained model is evaluated using the test dataset. Finally, the model is applied to predict the deviation of space debris orbital position and used for laser ranging observations of the ranging system.

[0034] In the dataset, azimuth, pitch, and their sine and cosine values ​​are used as input features, and azimuth deviation and pitch deviation are used as output labels. The neural network is trained using the training dataset. Under the stimulation of external input samples, the network continuously changes the connection weights so that the network output continuously approaches the desired output.

[0035] Specifically, the orbit prediction correction for space debris codenamed 02144 based on triangular feature decomposition was carried out, and the specific experimental process is as follows: The first step is the collection of observational data (sample data). Laser ranging observations of space debris 02144 were conducted using a space debris laser ranging system. Some of the collected data are as follows: Figure 11 As shown, the first column is the space debris code, the second column is the day's accumulated seconds, the third column is the azimuth angle, the fourth column is the elevation angle, the fifth column is the azimuth deviation, and the sixth column is the elevation deviation.

[0036] The second step is data preprocessing to generate feature vectors. That is, according to... Figure 3 The process shown generates the final feature vector.

[0037] The third step is to construct a neural network model according to the neural network model structure and parameters required for model training described above.

[0038] The fourth step involves first reading 653 sample data points from the observation data records of space debris 02144, then dividing the data into a training set of 80% and a test set of 20%, and then using the training set data to train the neural network model. Finally, the trained model can be obtained.

[0039] The fifth step is neural network model validation. After obtaining the trained neural network model, this model is used to predict the azimuth and elevation deviation values ​​of the training and test set sample data used when building the model. The results are as follows: Figure 7 and Figure 8 As shown, in Figure 7 In the meantime, the standard deviation of the azimuth angle is less than 1", and the maximum deviation is 7.5". Figure 8 In the data, the standard deviation of the elevation angle is less than 1", and the maximum deviation is 4.3". The model was used to predict the azimuth and elevation deviations of this target based on the new observation data, and the results are as follows: Figure 9 and Figure 10 As shown, the standard deviation of the azimuth angle is less than 3", and the maximum deviation of the azimuth angle is 8.5". The standard deviation of the pitch angle is less than 3", and the maximum deviation is 12.0".

[0040] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for correcting space debris orbit prediction based on triangular eigenvalue decomposition, characterized in that, The orbit prediction correction method specifically includes the following steps: The telescope ranging and tracking system tracks space debris and simultaneously records the measured azimuth and elevation angles at the moment of emission when laser echo signals are present, as well as the telescope's azimuth and elevation correction values, in order to obtain observation data. The observation data is preprocessed, and the azimuth and elevation angles in the preprocessed observation data are decomposed into trigonometric features to obtain sample data. A neural network model is constructed by dividing the sample data into a training set and a validation set as the input to the neural network model, and using the telescope azimuth and elevation correction values ​​at the corresponding time as the model output to train the neural network model. The azimuth and elevation angles acquired in real time by the telescope ranging and tracking system are decomposed into trigonometric features and then input into the trained neural network model to obtain the telescope azimuth and elevation correction prediction values ​​that are adjusted in real time to adjust the telescope's pointing direction.

2. The space debris orbit prediction correction method according to claim 1, characterized in that, The telescope azimuth and elevation correction values ​​include azimuth deviation and elevation deviation.

3. The space debris orbit prediction correction method according to claim 1, characterized in that, The sample data includes the original azimuth angle value, the azimuth angle sine value, the azimuth angle cosine value, the original elevation angle value, the elevation angle sine value, and the elevation angle cosine value.

4. The space debris orbit prediction correction method according to claim 1, characterized in that, The telescope ranging and tracking system tracks space debris based on two rows of root orbit data.

5. The space debris orbit prediction correction method according to claim 1, characterized in that, The structure of the neural network model includes an input layer, a hidden layer, and an output layer; The input layer has a dimension of 6, corresponding to the azimuth angle, pitch angle and their sine and cosine values, and maps the 6-dimensional input features to a 60-dimensional hidden space. The hidden layer contains at least two residual blocks, and each residual block performs the following steps sequentially during each training iteration: Save the 60-dimensional features of the current output of the input layer as the identity residual; A 60→60-dimensional linear transformation is performed through a fully connected layer with shared parameters; The result of the linear transformation is added to the identity residual; Use the ReLU activation function; The output layer is used to output the azimuth prediction error and the pitch prediction error.

6. The space debris orbit prediction correction method according to claim 1, characterized in that, The training process of the neural network model includes: Set the maximum number of iterations and the loss function, use the Adam optimizer to set the initial model parameters, and then train the neural network model; Set the iteration termination condition and initial learning rate, and repeat the forward propagation, loss calculation, back propagation, model parameter update and learning rate adjustment in sequence before the neural network model reaches the iteration termination condition after training times; After every few iterations of training, the obtained model is evaluated using a validation set, and the loss of the validation set is calculated. The iteration termination condition is one of two conditions: the training number of the neural network model reaches the maximum number of iterations and the loss of the validation set does not decrease after several consecutive calculations. After the neural network model reaches the termination condition of the iteration, the current neural network model is output. The trained neural network model is evaluated using a validation set. If the final neural network model passes the evaluation, it is saved and output; otherwise, the training is repeated.

7. The space debris orbit prediction correction method according to claim 6, characterized in that, The meaning of "the loss of the validation set has not decreased" is that the difference between the loss of the validation set calculated in several consecutive calculations and the loss of the validation set calculated in the first calculation in those calculations is greater than the loss threshold.

8. A system for implementing the space debris orbit prediction correction method according to any one of claims 1-7, characterized in that, It includes a processor and a memory, the memory being used to store a computer program, which, when executed by the processor, implements the space debris orbit prediction correction method as described in any one of claims 1-7.

Images