Neural network method for forecasting orbit of spacecraft

Abstract

The invention discloses a neural network method for forecasting a spacecraft orbit, which belongs to the field of spacecraft orbits, and comprises the following steps: S1, predicting a TLE initial value by using a deep learning neural network model and a feedback neural network model; s2, using a deep learning neural network model to improve the precision of the SGP4 model; s3, comparing the predicted value in the S1 with a true value, and evaluating the accuracy of TLE initial value prediction; and S4, comparing the corrected model in the step S2 with a true value, and evaluating the improvement degree of the precision of the SGP4 model. By adopting the neural network method for spacecraft orbit forecasting, the problem of balance between precision and data volume in spacecraft orbit forecasting is solved, the precision of a common SGP4 is improved, and meanwhile, the problem of precision of a TLE initial value for spacecraft orbit indirect forecasting is also solved.

Description

Technical Field

[0001] The invention relates to the field of spacecraft orbits, and in particular to a neural network method for predicting spacecraft orbits. Background Art

[0002] With the increasing frequency of space activities, the number of satellites in orbit is increasing, especially the United States has dozens of reconnaissance satellites in orbit, which can conduct large-scale, all-weather, and high-efficiency reconnaissance and detection of our maritime ship targets and ocean surveillance. In order to effectively avoid their reconnaissance and surveillance or interfere with them, it is necessary to predict their satellite orbits. Orbital satellite missions are booming, among which low-orbit satellites have the characteristics of fast changes in geometric observation configurations and high signal strength. On the one hand, they focus on the initial value of TLE, and on the other hand, they focus on the optimization of each perturbation force model. Although traditional methods are relatively mature, satellite orbit prediction models still have certain limitations and the prediction accuracy needs to be improved. In the existing methods, there is a problem of difficulty in balancing the relationship between accuracy and data set size, and difficulty in balancing the relationship between computational efficiency and accuracy. When the data set is large, the accuracy is high, but the computational efficiency is low. When the data set is small, the computational efficiency is high, but the accuracy is low. At the same time, there is also the problem of low accuracy of orbit prediction error in predicting the initial value of TLE. Summary of the invention

[0003] The purpose of the present invention is to provide a neural network method for spacecraft orbit prediction, which solves the balance problem between accuracy and data volume in spacecraft orbit prediction, improves the accuracy of the commonly used orbit prediction physical model SGP4, and also solves the accuracy problem of the initial value of the spacecraft orbit indirect prediction TLE.

[0004] To achieve the above object, the present invention provides a neural network method for predicting spacecraft orbits, comprising the following steps: S1, use the CNN-SEBlock-LSTM model and the feedback neural network model to predict the initial value of TLE. In this process, the activation function of the CNN-SEBlock-LSTM model is the Relu function and the Sigmoid function, and the activation function of the feedback neural network model is the Tanh function and the Linar function to predict the initial value of TLE; S2, use the CNN-SEBlock-LSTM model to improve the accuracy of the orbit prediction SGP4 model. In this process, the activation functions of the CNN-SEBlock-LSTM model are the LeakyRelu function and the Tanh function; S3, compare the predicted value in S1 with the true value to evaluate the accuracy of the initial TLE value prediction; S4, compare the corrected model in S2 with the true value to evaluate the improvement in the accuracy of the orbit prediction SGP4 model.

[0005] Preferably, in S1, the initial TLE value includes the orbital inclination , right ascension of ascending node , Argument of perigee , eccentricity , average motion , mean anomaly and time , where the CNN-SEBlock-LSTM model is used to predict orbital inclination , eccentricity , average motion and mean anomaly , the feedback neural network model is used to predict the right ascension of the ascending node , Argument of perigee .

[0006] Preferably, in S1, the process of obtaining the predicted value is: S11, input the TLE initial values ​​of a set of historical data into the CNN-SEBlock-LSTM model, and after normalization, each feature is in the range of [0, 1]; S12, the convolutional network and fully connected layer activate Relu and Sigmoid functions to extract important information and weights of the TLE initial value; S13, use the attention mechanism to output data information according to different weights; S14, the long short-term memory layer captures the time relationship in the processed data, and then performs regression prediction on the data to output the orbital inclination , eccentricity , average motion and mean anomaly The predicted value of S15, set the time Input the feedback neural network model, and the feedback neural network model outputs the right ascension of the ascending node , Argument of perigee The predicted value of In S1, the hyperparameters of the CNN-SEBlock-LSTM model are: the number of iterations is 1000, the initial learning rate is 0.01, the learning rate reduction factor is 0.9, the learning rate reduction cycle is 10, the minimum batch is 141 groups, and the data is shuffled in each round of training.

[0007] Preferably, the data processing process of the feedback neural network model is as follows: inputting the time T data into the input layer, the input layer transmits the data to the fully connected layer 1 for weighting, and then activating the Tanh function in the fully connected layer 1 to process the data, the processed data is transmitted from the fully connected layer 1 to the fully connected layer 2 for weighting, and then activating the Linar function in the fully connected layer 2 to perform secondary processing on the data, the data after the secondary processing is transmitted from the fully connected layer 2 to the output layer, and the output layer outputs the right ascension of the ascending node , Argument of perigee The predicted value of .

[0008] Preferably, the CNN-SEBlock-LSTM model usage process in S2 is: S21, sorting and normalizing the input values, so that the range of each feature is within the range of [-1, 1]; S22, the convolutional network and fully connected layer activate leakyRelu and Tanh functions to extract local rules and features in the data; S23 uses the attention mechanism to provide the model with a focus on key information and dynamically adjust the impact of different features; In S2, the hyperparameters of the CNN-SEBlock-LSTM model are: the initial learning rate is 0.01, the minimum batch is 200 groups, the maximum number of iterations is 100, the learning rate decreases once every 10 rounds, the decrease factor is 0.8, the sequence is shuffled in each round of training, and the adaptive moment estimation model is used as the optimizer.

[0009] Preferably, the evaluation criteria in S3 are: Evaluation coefficients of the model The calculation is as follows, ; Root mean square error The calculation is as follows, ; Where A represents the orbital inclination , right ascension of ascending node , Argument of perigee , eccentricity , average motion , mean anomaly One of them, N is the total number of samples, True value, is the model's predicted value.

[0010] Preferably, the evaluation criteria in S4 are: Introduce true values ​​for supervision, ; in, represents the true error, for , , , respectively The three components of position; This is The actual measured value at the moment, yes The value predicted by the SGP4 orbit prediction model at the time; Then the evaluation coefficient of the model is and performance indicators The accuracy improvement of the orbit prediction SGP4 model prediction results is evaluated, and the evaluation coefficient The calculation is as follows: ; ; Performance Indicators The calculation is as follows: ; in, represents the residual, Represents the error of the neural network prediction.

[0011] Therefore, the present invention adopts the above-mentioned neural network method for predicting the orbit of a spacecraft, which has the following advantages: (1) The accuracy of the commonly used orbit prediction physical model SGP4 has been improved.

[0012] (2) It balances the relationship between data volume and precision, ensuring accuracy while having faster computing efficiency.

[0013] (3) The initial value of TLE can be predicted with high precision, thus indirectly predicting the orbit more accurately.

[0014] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a schematic diagram of the structure of a CNN-SEBlock-LSTM model in a neural network method S1 for predicting spacecraft orbits in the present invention; Figure 2 A schematic diagram of the structure of a CNN-SEBlock-LSTM model in a neural network method S2 for predicting a spacecraft orbit according to the present invention; Figure 3 A schematic diagram of the structure of a feedback neural network of a neural network method for predicting a spacecraft orbit according to the present invention; Figure 4A neural network method for predicting the orbit of a spacecraft according to the present invention Schematic diagram of the comparison between the predicted results and the true values; Figure 5 A neural network method for predicting the orbit of a spacecraft according to the present invention Schematic diagram of the difference between the predicted result and the true value; Figure 6 The eccentricity of a neural network method for predicting spacecraft orbits according to the present invention Schematic diagram of the comparison between the predicted results and the true values; Figure 7 The eccentricity of a neural network method for predicting a spacecraft orbit according to the present invention Schematic diagram of the difference between the predicted result and the true value; Figure 8 The present invention is a spacecraft orbit prediction neural network method average motion Schematic diagram of the comparison between the predicted results and the true values; Fig. 9 The present invention is a spacecraft orbit prediction neural network method average motion Schematic diagram of the difference between the predicted result and the true value; Fig.10 A neural network method for predicting the orbit of a spacecraft using the present invention Schematic diagram of the comparison between the predicted results and the true values; Fig.11 A neural network method for predicting the orbit of a spacecraft using the present invention Schematic diagram of the difference between the predicted result and the true value; Fig.12 The invention provides a neural network method for predicting the orbit of a spacecraft. Schematic diagram of the comparison between the predicted results and the true values; Fig.13 The invention provides a neural network method for predicting the orbit of a spacecraft. Schematic diagram of the difference between the predicted result and the true value; Fig.14 A neural network method for predicting the orbit of a spacecraft using the perigee angle Schematic diagram of the comparison between the predicted results and the true values; Fig.15 A neural network method for predicting the orbit of a spacecraft using the perigee angle Schematic diagram of the difference between the predicted result and the true value; Fig.16 It is a schematic diagram of the X-direction correction effect of the neural network method for predicting the orbit of a spacecraft according to the present invention on the SGP4 after correction; Fig.17 A residual histogram of the X-direction correction effect of the neural network method for predicting the orbit of a spacecraft according to the present invention on the SGP4 correction; Fig.18 It is a schematic diagram of the Y-direction correction effect of the neural network method for predicting the orbit of a spacecraft according to the present invention on the SGP4 after correction; Fig.19 A residual histogram of the Y-direction correction effect of the neural network method for predicting the orbit of a spacecraft according to the present invention on the SGP4 after correction; Fig. 20 It is a schematic diagram of the Z-direction correction effect of the neural network method for predicting the orbit of a spacecraft according to the present invention on the SGP4 after correction; Fig.21 The residual histogram of the Z-direction correction effect of the neural network method for spacecraft orbit prediction of the present invention on SGP4 after correction. DETAILED DESCRIPTION

[0016] Example In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0017] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "setting", "installation" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0018] Some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0019] In the description of the present invention, it should be noted that the directions or positional relationships indicated by terms such as “upper”, “lower”, “left”, “right”, “inside” and “outside” are based on the directions or positional relationships shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description. They do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they cannot be understood as limitations on the present invention.

[0020] The specific model specifications need to be selected and determined based on the actual specifications of the device, and the specific selection calculation method adopts the existing technology in this field, so it will not be described in detail.

[0021] The present invention provides a neural network method for predicting spacecraft orbits, comprising the following steps: S1, use the CNN-SEBlock-LSTM model and the feedback neural network model to predict the initial value of TLE. In this process, the activation function of the CNN-SEBlock-LSTM model is the Relu function and the Sigmoid function, and the activation function of the feedback neural network model is the Tanh function and the Linar function to predict the initial value of TLE; The activation function is mainly a function that runs on neurons. Its function is to scale and transform the output of neurons to make them have nonlinear or linear characteristics, helping the network learn and represent complex functional relationships.

[0022] The initial TLE value includes the orbital inclination , right ascension of ascending node , Argument of perigee , eccentricity , average motion 、Mean Anomaly and time , where the CNN-SEBlock-LSTM model is used to predict the orbital inclination , eccentricity , average motion and mean anomaly , the feedback neural network model is used to predict the right ascension of the ascending node , Argument of perigee .

[0023] The process of getting the predicted value is: S11, input the TLE initial values ​​of a set of historical data into the CNN-SEBlock-LSTM model, and after normalization, each feature is in the range of [0, 1]; In this embodiment, the historical data used are the six orbital elements in the two-line element number (TLE) released by the North American Defense Command, which are used to predict the six orbital elements at a certain moment in the future. The six elements are converted into position and speed through analytical formulas to achieve the predicted orbit. The input set is: orbital inclination , right ascension of ascending node , Argument of perigee , eccentricity , average motion 、Mean Anomaly and time (TLE era Julian day, such as 24061, which is the 61st day of 2024, 2024.03.01).

[0024] The purpose of normalization is to avoid the influence of the data size of different features on the weight of the result. Because they are all non-negative numbers, the formula is: ; in Represents feature data, Indicates the minimum value of characteristic data, Indicates the maximum value of characteristic data, K j represents the normalized eigenvalue.

[0025] Each feature is converted to the range of [0,1] to avoid the data size of different features affecting the output results. The total data set of the simulation is 175 groups of six numbers for 175 days, of which 141 groups of data are used to train the model, and the other 34 groups are used to test the effect of the neural network and verify its generalization.

[0026] S12, the convolutional network and fully connected layer activate Relu and Sigmoid functions to extract important information and weights of the TLE initial value; S13, use the attention mechanism to output data information according to different weights; S14, the long short-term memory layer captures the time relationship in the processed data, and then performs regression prediction on the data to output the orbital inclination , eccentricity , average motion and mean anomaly The predicted value of S15, set the time Input the feedback neural network model, and the feedback neural network model outputs the right ascension of the ascending node , Argument of perigee The predicted value of The CNN-SEBlock-LSTM model has 13 layers, and the activation functions are Relu, Sigmoid function, and Relu function: ; Sigmoid function: ; like Figure 1 As shown, the orbital inclination , right ascension of ascending node , Argument of perigee , eccentricity , average motion , mean anomaly Six items of data are input into the input layer, and the input layer transmits the six items of data to the sequence folding layer. The sequence folding layer processes the data to obtain the original data, and one copy of the original data is transmitted to the sequence expansion layer for processing. One copy of the original data is transmitted to the convolution layer 1. After the convolution layer 1 preliminarily extracts features from the original data, one copy of the original data with the preliminarily extracted features is transmitted to the global average pooling layer, and then the activation function 1 Relu processes another copy of the original data with the preliminarily extracted features. After the processing is completed, the data is transmitted to the convolution layer 2 activation function 2 Relu for feature extraction again, and then the local features are obtained, and the local features are transmitted to the multiplication layer; The global average pooling layer transmits the raw data of the initial feature extraction to the fully connected layer 1. The fully connected layer 1 performs weighted activation function 3 Relu on the data, and then transmits the processed data to the fully connected layer 2 for a second weighted processing of the data. The weighted activation function 4 Sigmoid processes the second weighted data, and then transmits the processed data to the multiplication layer. The local features and the processed data are multiplied in the multiplication layer, and then enter the sequence expansion layer together with the original data to expand. After expansion, the data sequence is obtained, and the data sequence is sent to the flattening layer for weighting. The weighted data sequence is sent to the long short-term memory layer for prediction. The long short-term memory layer processes the data sequence and sends it to the fully connected layer 3. The fully connected layer 3 organizes the data and extracts the data points. Then the data points are sent to the regression output layer for regression prediction and output of the orbital inclination. , eccentricity , average motion and mean anomaly The predicted value of The hyperparameters of the CNN-SEBlock-LSTM model are: number of iterations: 1000, initial learning rate: 0.01, learning rate reduction factor: 0.9, learning rate reduction cycle: 10, minimum batch size: 141 groups, and after each round of training, the data is shuffled for training.

[0027] In S1, the CNN-SEBlock-LSTM model takes a TLE data and time of the spacecraft history as input, and outputs the orbital inclination at a certain moment in the future. , eccentricity , average motion and mean anomaly , where the prediction time interval can be customized.

[0028] The input of the feedback neural network model is time T, and the output is the right ascension of the ascending node and the argument of perigee. The reason for choosing this model is that it is found that the right ascension of the ascending node and the argument of perigee show an approximately linear relationship with time.

[0029] like Figure 3 The data processing process of the feedback neural network model shown is as follows: the time T data is input into the input layer, the input layer transmits the data to the fully connected layer 1 for weighting, and then the Tanh function is activated in the fully connected layer 1 to process the data. After the processing, the data is transmitted from the fully connected layer 1 to the fully connected layer 2 for weighting, and then the Linar function is activated in the fully connected layer 2 for secondary processing of the data. After the secondary processing, the data is transmitted from the fully connected layer 2 to the output layer, and the output layer outputs the right ascension of the ascending node. , Argument of perigee The predicted value of S2, use the CNN-SEBlock-LSTM model to improve the accuracy of the orbit prediction SGP4 model. In this process, the activation functions of the CNN-SEBlock-LSTM model are the LeakyRelu function and the Tanh function; leakyRelu function: ; Tanh function: ; The usage process of the CNN-SEBlock-LSTM model in S2 is as follows: S21, sorting and normalizing the input values, so that the range of each feature is within the range of [-1, 1]; The input value in S21 is a data set, with a total of 11 quantities, including the position R predicted by SGP4 in three directions, R X , R Y , R Z ; The predicted speed V in three directions is V X 、V Y 、V Z As well as the ascending node right ascension, declination, azimuth, flight path angle and the time difference with the reference point as the zero point, the daily measurement values ​​of the International Laser Ranging Service (ILRS) are recorded by the European data center for real values ​​to supervise as the target set of the neural network. The selected time period of the data set is 2024.06.03-2024.06.10, with the first point of the selected time as the time reference point, with 1 minute as a step length, and the position and speed of the spacecraft with different step lengths within a week are predicted by the physical model SGP4, and then the ascending node right ascension, declination, azimuth, flight path angle and the time difference with the reference point as the zero point are calculated to form the learning variable set of the neural network: ① , For the current moment, is the future time; unit: minute; ② Ascending node right ascension : The angle between the ascending node and the vernal equinox, in degrees; ③ Declination ; The equator is 0 degrees, north is positive, and it will change over time; ④ Azimuth ; The angle from the North Pole to the target in a clockwise direction; ⑤ Flight path angle : The angle between the velocity vector direction and the vertical position vector direction; ⑥ The position predicted by SGP4 and speed ; During normalization, since the values ​​of different features are not in the same order of magnitude, each feature needs to be normalized. Generally, normalization changes the data to [0,1]. Since this data set contains negative values, the data is changed to [-1,1] here, which allows the neural network to better map the relationship between input and output. The reason for choosing the two activation functions, LeakyRelu and Tanh, is that the normalization range is [-1,1], which is consistent with the value range of the function.

[0030] There are a total of 10081 groups in the dataset, with the training set time period being 2024.06.03-2024.06.08 and the test set time period being (2024.06.08-2024.06.10).

[0031] S22, the convolutional network and fully connected layer activate leakyRelu and Tanh functions to extract local rules and features in the data; S23 uses the attention mechanism to provide the model with a focus on key information and dynamically adjust the impact of different features; like Figure 2 As shown, the above eleven data are input into the input layer, and the input layer transmits the eleven data to the sequence folding layer. The sequence folding layer processes the data to obtain the original data, and one copy of the original data is transmitted to the sequence expansion layer for processing, and one copy of the original data is transmitted to the convolution layer 1. After the convolution layer 1 preliminarily extracts features from the original data, one copy of the original data with the preliminarily extracted features is transmitted to the global average pooling layer, and then the activation function 1 LeakyRelu processes another copy of the original data with the preliminarily extracted features. After the processing is completed, the data is transmitted to the convolution layer 2 activation function 2 LeakyRelu for feature extraction again, and then the local features are obtained, and the local features are transmitted to the multiplication layer; The global average pooling layer transmits the original data of the preliminary extracted features to the fully connected layer 1. The fully connected layer 1 performs weighted activation function 3 LeakyRelu on the data, and then transmits the processed data to the fully connected layer 2 for a second weighted processing of the data. The weighted activation function 4 tanhLayer processes the second weighted data, and then transmits the processed data to the multiplication layer; The local features and the processed data are multiplied in the multiplication layer, and then enter the sequence expansion layer together with the original data to expand. After expansion, the data sequence is obtained, and the data sequence is sent to the flattening layer for weighting. The weighted data sequence is sent to the long short-term memory layer for prediction. The long short-term memory layer processes the data sequence and then sends it to the fully connected layer 3. The fully connected layer 3 organizes the data to obtain the error value, and then sends the error value to the regression output layer for integration processing, and outputs the adjustment value of the SGP4 model; The hyperparameters of the CNN-SEBlock-LSTM model are: using the adaptive moment estimation model as the optimizer, the initial learning rate is 0.01, the minimum batch is 200 groups, the maximum number of iterations is 100, the learning rate decreases every 10 rounds, the decrease factor is 0.8, and the sequence is shuffled in each round of training.

[0032] S3, compare the predicted value in S1 with the true value to evaluate the accuracy of the initial TLE value prediction; The evaluation criteria in S3 are: Evaluation coefficients of the model The calculation is as follows, ; Root mean square error The calculation is as follows, ; Where A represents the orbital inclination , right ascension of ascending node , Argument of perigee , eccentricity , average motion , mean anomaly One of them, N is the total number of samples, True value, is the model's predicted value.

[0033] S4, compare the corrected model in S2 with the true value to evaluate the improvement in the accuracy of the orbit prediction SGP4 model.

[0034] The evaluation criteria in S4 are: Introduce true values ​​for supervision, ; in, represents the true error, for , , , respectively The three components of position; This is The actual measured value at the moment, yes The value predicted by the SGP4 orbit prediction model at the time; Then the evaluation coefficient of the model is and performance indicators The accuracy improvement of the orbit prediction SGP4 model prediction results is evaluated, and the evaluation coefficient The calculation is as follows, ; ; Performance Indicators The calculation is as follows, ; in, represents the residual, Represents the error of the neural network prediction.

[0035] For the evaluation criteria, when the evaluation coefficient The larger it is, the better the prediction effect of the neural network model is, and the smaller the gap between the predicted value and the true value is. Conversely, the prediction effect is not good. When the RMSE is smaller, the better the prediction result is, and the distribution of the predicted value and the true value is more similar, which indicates the concentration degree of the data on the best fit line.

[0036] like Figure 4 Figure 5 As shown, the inclination angle predicted by the CNN-SEBlock-LSTM model The prediction results are basically consistent with the true values, and the evaluation coefficient of the model The root mean square error is 0.99689. It is 0.00052307.

[0037] like Figure 6 Figure 7 As shown, the eccentricity predicted by the CNN-SEBlock-LSTM model The prediction results are basically consistent with the true values, and the evaluation coefficient of the model The root mean square error is 0.99196. is 1.1941e-06.

[0038] like Figure 8 Fig. 9 As shown, the average motion predicted by the CNN-SEBlock-LSTM model The prediction results are basically consistent with the true values, and the evaluation coefficient of the model The root mean square error is 0.97952. is 1.6315e-07.

[0039] like Fig.10 Fig.11 As shown, the mean anomaly angle predicted by the CNN-SEBlock-LSTM model The prediction results are basically consistent with the true values, and the evaluation coefficient of the model The root mean square error is 0.99878. is 3.3263.

[0040] like Fig.12 Fig.13 As shown in Figure 2, the right ascension of the ascending node predicted by the feedback neural network model The prediction results are basically consistent with the true values, and the evaluation coefficient of the model The root mean square error is 0.9989. for .

[0041] like Fig.14 Fig.15 As shown in Figure 2, the perigee angle predicted by the feedback neural network model The prediction results are basically consistent with the true values, and the evaluation coefficient of the model for , root mean square error for .

[0042] The correction effect of CNN-SEBlock-LSTM model on SGP4 model is: like Fig.16 Fig.17 As shown, the position is in the X direction: Training set: , , The bigger, The smaller it is, the better the model effect is. It can be seen that the error on the training set is reduced by 92.2%.

[0043] like Fig.18 Fig.19 As shown, the position error prediction results in the Y direction: Training set: , , the original error is reduced by 91.34%.

[0044] like Fig. 20 Fig.21 As shown, the position error prediction in the Z direction: Training set: , , the original error is reduced by 92.94%.

[0045] In all the above directions, quantitative indicators They are all less than 100%, from which it can be concluded that the designed neural network algorithm can improve the orbit prediction accuracy of the SGP4 model.

[0046] Therefore, the present invention adopts the above-mentioned neural network method for spacecraft orbit prediction, which solves the balance problem between accuracy and data volume in spacecraft orbit prediction, improves the accuracy of the commonly used orbit prediction physical model SGP4, and also solves the accuracy problem of the initial value of the spacecraft orbit indirect prediction TLE.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A neural network method for predicting spacecraft orbits, characterized in that: The following steps are included: S1, use the CNN-SEBlock-LSTM model and the feedback neural network model to predict the initial value of TLE. In this process, the activation function of the CNN-SEBlock-LSTM model is the Relu function and the Sigmoid function, and the activation function of the feedback neural network model is the Tanh function and the Linar function to predict the initial value of TLE; S2, use the CNN-SEBlock-LSTM model to improve the accuracy of the orbit prediction SGP4 model. In this process, the activation functions of the CNN-SEBlock-LSTM model are the LeakyRelu function and the Tanh function; S3, compare the predicted value in S1 with the true value to evaluate the accuracy of the initial TLE value prediction; S4, compare the corrected model in S2 with the true value to evaluate the improvement in the accuracy of the orbit prediction SGP4 model.

2. The neural network method for predicting spacecraft orbit according to claim 1, characterized in that: In S1, the initial TLE value includes the orbital inclination , right ascension of ascending node , Argument of perigee , eccentricity , average motion , mean anomaly and time , where the CNN-SEBlock-LSTM model is used to predict orbital inclination , eccentricity , average motion and mean anomaly , the feedback neural network model is used to predict the right ascension of the ascending node , Argument of perigee .

3. The neural network method for predicting spacecraft orbit according to claim 2, characterized in that: In S1, the process of obtaining the predicted value is: S11, input the TLE initial values ​​of a set of historical data into the CNN-SEBlock-LSTM model, and after normalization, each feature is in the range of [0, 1]; S12, the convolutional network and fully connected layer activate Relu and Sigmoid functions to extract important information and weights of the TLE initial value; S13, use the attention mechanism to output data information according to different weights; S14, the long short-term memory layer captures the time relationship in the processed data, and then performs regression prediction on the data to output the orbital inclination , eccentricity , average motion and mean anomaly The predicted value of S15, set the time Input the feedback neural network model, and the feedback neural network model outputs the right ascension of the ascending node , Argument of perigee The predicted value of In S1, the hyperparameters of the CNN-SEBlock-LSTM model are: the number of iterations is 1000, the initial learning rate is 0.01, the learning rate reduction factor is 0.9, the learning rate reduction cycle is 10, the minimum batch is 141 groups, and the data is shuffled in each round of training.

4. A neural network method for spacecraft orbit prediction according to claim 3, characterized in that: The data processing process of the feedback neural network model is as follows: the time T data is input into the input layer, the input layer sends the data to the fully connected layer 1 for weighting, and then the Tanh function is activated in the fully connected layer 1 to process the data. After the processing, the data is sent from the fully connected layer 1 to the fully connected layer 2 for weighting, and then the Linar function is activated in the fully connected layer 2 for secondary processing of the data. After the secondary processing, the data is sent from the fully connected layer 2 to the output layer, and the output layer outputs the right ascension of the ascending node. , Argument of perigee The predicted value of .

5. The neural network method for predicting spacecraft orbit according to claim 1, characterized in that: The usage process of the CNN-SEBlock-LSTM model in S2 is: S21, sorting and normalizing the input values, so that the range of each feature is within the range of [-1, 1]; S22, the convolutional network and fully connected layer activate leakyRelu and Tanh functions to extract local rules and features in the data; S23 uses the attention mechanism to provide the model with a focus on key information and dynamically adjust the impact of different features; In S2, the hyperparameters of the CNN-SEBlock-LSTM model are: the initial learning rate is 0.01, the minimum batch is 200 groups, the maximum number of iterations is 100, the learning rate decreases once every 10 rounds, the decrease factor is 0.8, the sequence is shuffled in each round of training, and the adaptive moment estimation model is used as the optimizer.

6. The neural network method for predicting spacecraft orbit according to claim 1, characterized in that: The evaluation criteria in S3 are: Evaluation coefficients of the model The calculation is as follows, ； Root mean square error The calculation is as follows, ； Where A represents the orbital inclination , right ascension of ascending node , Argument of perigee , eccentricity , average motion , mean anomaly One of them, N is the total number of samples, True value, is the model's predicted value.

7. The neural network method for predicting spacecraft orbit according to claim 1, characterized in that: The evaluation criteria in S4 are: Introduce true values ​​for supervision, ； in, represents the true error, for , , , respectively The three components of position; This is The actual measured value at the moment, yes The value predicted by the SGP4 orbit prediction model at the moment; Then the evaluation coefficient of the model is and performance indicators The accuracy improvement of the orbit prediction SGP4 model prediction results is evaluated, and the evaluation coefficient The calculation is as follows: ； ； Performance Indicators The calculation is as follows: ； in, represents the residual, Represents the error of the neural network prediction.

Images