Space target recognition and classification method, device and medium based on non-conservative force characteristics

Abstract

The present invention discloses a method, device and medium for spatial target recognition and classification based on non-conservative force characteristics, which fully considers the effects of non-conservative forces such as solar radiation pressure, earth radiation pressure and atmospheric drag, and obtains characteristics such as attitude, shape, size, surface mass ratio, surface material, reflection coefficient, absorption coefficient, and atmospheric drag coefficient from orbit data with a wide range of sources. Compared with the conservative force model that only considers solar radiation pressure, the non-conservative force characteristics obtained in this way are more reliable and help to improve the accuracy of target recognition. Moreover, the orbit data contains more abundant non-conservative force characteristics compared with photometric curves, angular curves, etc., and can be better applied to the spatial target recognition task. In addition, it is also considered to use a graph neural network that performs well in the recognition and classification task to process the obtained non-conservative force characteristics, significantly improving the accuracy of recognition and classification.

Description

Technical Field

[0001] The present invention belongs to the technical field of identification and classification of low, medium, and high-orbit space targets, and more specifically, relates to a method, device, and medium for identifying and classifying space targets based on non-conservative force characteristics. Background Art

[0002] With the continuous development of space technology, people have increasingly realized the importance of space resources. On the other hand, the available space environment is very limited. In recent years, the activities of various countries in launching artificial space objects and seizing space orbits have shown a rapid growth trend. Many satellites that have reached their designed lifetimes are still in orbit because they cannot be recycled or processed, which further exacerbates the problem of severe orbital congestion.

[0003] So far, ground-based optical and radar systems are still the main means of obtaining space target information, and can complete the detection, identification, and cataloging of large low-orbit targets. The realization of target identification depends on the selection and acquisition of features. The optical system can obtain the radiation information and image information of space targets, and typical representatives are photometric curves, angular curves, etc. Further processing can obtain features such as the attitude, shape, size, and material properties of the target, so as to identify and classify the target, which is mainly applied to the identification of space targets with convex polyhedron shapes. High-resolution radar can directly obtain the attitude, shape, and size information features of space targets, and identify them based on these target features. There are mainly four categories of algorithms for identifying and classifying space targets: distance-based classification, such as the K-nearest neighbor algorithm; statistical-based classification, such as the Bayesian classifier; rule-based classification, such as decision trees; and neural network-based classification algorithms. Classification belongs to a prediction task. In specific applications, which classification algorithm to choose depends on factors such as the characteristics of the data, the complexity of the task, and computing resources. With the development of artificial intelligence technology, neural network-based classification algorithms have been widely applied to the research of space target classification based on photometric data and micro-Doppler radar images, and have achieved fruitful results.

[0004] Currently, in the research of space target identification, the following main deficiencies exist: (1) The fineness of the orbital non-conservative force model needs to be further improved. Currently, when modeling the orbital non-conservative force, atmospheric drag and earth's radiation light pressure are ignored, resulting in the need to improve the accuracy of estimating the characteristic parameters of space targets; (2) The rich non-conservative force characteristics hidden in orbital data have not been used for the identification and classification of space targets; (3) The algorithms for identifying and classifying space targets need to be improved. Using a neural network with an appropriate structure for target identification can greatly improve the identification accuracy. (4) The methods for identifying space targets based on optoelectronic and radar observations are mainly applicable to low-orbit space objects, and are not very applicable to space objects with smaller sizes or higher orbits. Due to the influence of the atmosphere, distance, and its own resolution, optical observation means are difficult to achieve high resolution for them. Summary of the Invention

[0005] The present invention is provided to solve the above problems existing in the prior art. Therefore, there is a need for a method, apparatus, and medium for identifying and classifying space targets based on non-conservative force characteristics. The magnitude and direction of the non-conservative force acting on a space target are affected by the object's own attitude, shape, size, surface mass ratio, surface material, reflection coefficient, absorption coefficient, atmospheric drag coefficient, etc. Therefore, the characteristics such as the target's attitude, shape, and size can be deduced from the non-conservative force. Specifically, fully considering the effects of non-conservative forces such as solar radiation pressure, earth radiation pressure, and atmospheric drag, characteristics such as attitude, shape, size, surface mass ratio, surface material, reflection coefficient, absorption coefficient, and atmospheric drag coefficient are obtained from the widely sourced orbital data. Compared with the conservative force model that only considers solar radiation pressure, the non-conservative force characteristics obtained in this way are more reliable and help to improve the accuracy of target recognition. Moreover, the orbital data contains more abundant non-conservative force characteristics compared to photometric curves, angular curves, etc., and can be better applied to the space target recognition task. In addition, it is also considered to use a graph neural network that performs well in the recognition and classification task to process the obtained non-conservative force characteristics, significantly improving the accuracy of recognition and classification. Most importantly, the accuracy of the non-conservative force characteristics obtained from orbital data is not affected by the target size and orbital altitude, and thus the recognition range can be extended to medium and high-orbit space targets, having wide applicability.

[0006] According to the first aspect of the present invention, there is provided a method for identifying and classifying space targets based on non-conservative force characteristics, the method comprising:

[0007] Model the motion equation of the space target through the following formula:

[0008]

[0009] where r is the position vector of the space target, μ is the gravitational parameter of the earth, r = ||r||, is the non-spherical perturbation, a sun is the solar gravitational perturbation, a moon is the lunar gravitational perturbation, a SRP is the solar radiation pressure, a ERP is the earth radiation pressure, a drag is the atmospheric drag;

[0010] Compare the result after removing the influence of conservative forces with the original orbital time series to obtain the orbital time series only affected by non-conservative forces;

[0011] After removing the influence of conservative forces, model the non-conservative forces acting on a space target and analyze the influence of these non-conservative forces on the orbital elements, and further perform inversion to obtain the key features of the target. The non-conservative forces acting on the space target include solar radiation pressure, earth radiation pressure, and atmospheric drag received by the space target. The key features of the target include the attitude, shape, size, area-to-mass ratio, surface material, reflection coefficient, absorption coefficient, and atmospheric drag coefficient of the target;

[0012] Utilize the orbital information, the orbital time series under the influence of only non-conservative forces, and the key features of the target to perform feature stitching and fusion to obtain a feature vector, and establish a feature space for space target recognition and classification;

[0013] Establish a model training database based on the feature space;

[0014] Use the model training database to train the model, and realize space target recognition and classification through the trained model.

[0015] It should be noted that in the present invention, TLE in GP category data is adopted as the main data source for non-conservative force feature extraction. Considering that the conservative forces that dominate in orbital mechanical effects are irrelevant to the attitude, shape, size, area-to-mass ratio, surface material, reflection coefficient, absorption coefficient, atmospheric drag coefficient, etc. of the target, while the non-conservative forces are affected by these features. Therefore, it is necessary to remove the influence of conservative forces such as the earth's gravity from the orbital time series, and model the motion equation of the space target through the formulas described above.

[0016] Furthermore, compare the result after removing the influence of conservative forces such as the earth's gravity with the original orbital time series to obtain the orbital time series under the influence of only non-conservative forces. After removing the influence of conservative forces, it is also necessary to model the solar radiation pressure, earth radiation pressure, and atmospheric drag received by the space target and analyze the influence of these non-conservative forces on the orbital elements, and further perform inversion to obtain the key features such as the attitude, shape, size, area-to-mass ratio, surface material, reflection coefficient, absorption coefficient, and atmospheric drag coefficient of the target. For space targets with a relatively high orbital altitude (≥1000 km), solar radiation pressure is the most main non-conservative force acting on them. Here, taking solar radiation pressure as an example, introduce its corresponding calculation model:

[0017]

[0018] In the formula, S F = 1367 W / m 2 , represents the solar radiation constant; c = 299792458 m / s, represents the speed of light in vacuum; d is the distance from the space target to the sun expressed in astronomical units (AU); m objectis the mass of the space target; A(i), ε(i), u n (i) respectively represent the total area, emissivity, and normal vector of the i-th surface element of the space target; R spec (i), R diff (i), R abs (i) respectively represent spectral reflectivity, diffuse reflectivity, and absorption coefficient. Assuming no energy loss during the transmission of electromagnetic waves through the surface element, they have the following relationship:

[0019] R spec (i)+R diff (i)+R abs (i) = 1

[0020] In addition, cos(ψ(i)) ≡ u n (i)·u sun represents the cosine of the angle between the i-th surface element and the sun. The function G[cos(ψ(i))] = max(0, cos(ψ(i))), which means that when the surface element is blocked, its value is 0, and when the surface element is irradiated by the sun, the value is 1.

[0021] Utilize orbital information, the orbital time series under the influence of only non-conservative forces, and non-conservative force characteristics such as the attitude, shape, size, area-to-mass ratio, surface material, reflection coefficient, absorption coefficient, and atmospheric drag coefficient of the target, and perform feature splicing and fusion to obtain a more comprehensive feature vector, and accordingly establish a feature space finally used for space target recognition and classification.

[0022] Furthermore, the data in the model training database includes feature vectors spliced and fused from the orbital characteristics of each space target, the orbital time series under the action of non-conservative forces, and non-conservative force characteristics, as well as the category and number labels of the space target.

[0023] It should be noted that the data in the training database mainly comes from measured data. The training database stores feature vectors spliced and fused from the orbital characteristics of each space target, the orbital time series under the action of non-conservative forces, non-conservative force characteristics (including attitude, shape, size, area-to-mass ratio, surface material, reflection coefficient, absorption coefficient, atmospheric drag coefficient, etc.), as well as the category and number labels of the space target. The original data mainly comes from the GP category data released by the website https: / / www.space-track.org, including two-line elements (TLE) containing orbital information.

[0024] Furthermore, training the model using the model training database includes:

[0025] Taking each space target as a node of the graph structure and setting edges according to the spatial relationship between each space target;

[0026] Create a node feature vector by using the orbital characteristics of space targets, the time series of non-conservative forces, and the characteristics of non-conservative forces;

[0027] Select a graph convolutional long short-term memory network model for training. The graph convolutional long short-term memory network model includes a graph convolutional neural network suitable for processing the spatial relationship between nodes and a long short-term memory network, which can comprehensively analyze node features, covering orbital characteristics, non-conservative force time series, and non-conservative force characteristics;

[0028] Import the data in the model training database into the graph convolutional long short-term memory network model for model training.

[0029] Further, after importing the data in the model training database into the graph convolutional long short-term memory network model for model training, the method further includes:

[0030] During the training process, use a genetic algorithm to select model parameters for hyperparameters, and perform model training with the selected hyperparameters. The hyperparameters include the number of iterations, learning rate, batch sample size, and optimizer.

[0031] Further, the use of a genetic algorithm to select model parameters for hyperparameters includes:

[0032] Map randomly generated hyperparameters to genotype encoding to create an initial population;

[0033] Calculate the fitness of each individual in the population. The fitness is defined as the target recognition accuracy under the current hyperparameters;

[0034] Select the initial population according to the fitness, and obtain the next generation of individuals through crossover and mutation, and calculate the fitness;

[0035] Select the group of encoding with the strongest fitness, and decode to obtain the corresponding hyperparameter values.

[0036] According to the second technical solution of the present invention, a space target recognition and classification device based on non-conservative force characteristics is provided. The device includes:

[0037] A motion equation modeling module configured to model the motion equation of a space target through the following formula:

[0038]

[0039] where r is the position vector of the space target, μ is the gravitational parameter of the earth, r = ||r||, is the non-spherical perturbation, a sun is the solar gravitational perturbation, a moon is the lunar gravitational perturbation, a SRP is the solar radiation pressure, aERP is the Earth's radiation pressure, a drag is the atmospheric drag;

[0040] A time series comparison module, configured to compare the result after removing the influence of conservative forces with the original orbit time series to obtain the orbit time series under the influence of only non-conservative forces;

[0041] A key feature calculation module, configured to model the non-conservative forces acting on the space target and analyze the influence of the non-conservative forces on the orbital elements after removing the influence of conservative forces, and further invert to obtain the key features of the target. The non-conservative forces acting on the space target include the solar radiation pressure, the Earth's radiation pressure, and the atmospheric drag received by the space target. The key features of the target include the attitude, shape, size, surface-to-mass ratio, surface material, reflectivity, absorptivity, and atmospheric drag coefficient of the target;

[0042] A feature splicing and fusion module, configured to use the orbit information, the orbit time series under the influence of only non-conservative forces, and the key features of the target to perform feature splicing and fusion to obtain a feature vector, and establish a feature space for space target recognition and classification;

[0043] A database establishment module, which establishes a model training database based on the feature space;

[0044] A model training module, configured to train the model using the model training database, and realize space target recognition and classification through the trained model.

[0045] Furthermore, the key feature calculation module is further configured to:

[0046] For space targets with an orbital altitude ≥ 1000 km, taking the solar radiation pressure as the non-conservative force, the key features of the target are obtained through the following calculation model:

[0047]

[0048] In the formula, S F = 1367 W / m 2 , representing the solar radiation constant; c = 299792458 m / s, representing the speed of light propagation in vacuum; d is the distance from the space target to the sun in astronomical units AU; m object is the mass of the space target; A(i), ε(i), u n (i) respectively represent the total area, emissivity, and normal vector of the i-th surface element of the space target; R spec (i), R diff (i), R abs(i) represent the spectral reflectance, diffuse reflectance, and absorption coefficient respectively. In the case where there is no energy loss during the transmission of electromagnetic waves through the surface element, the sum of the spectral reflectance, diffuse reflectance, and absorption coefficient is 1; cos(ψ(i)) ≡ u n (i)·u sun represents the cosine of the angle between the i-th surface element and the sun. The function G[cos(ψ(i))] = max(0, cos(ψ(i))), which means that when the surface element is blocked, its value is 0, and when the surface element is irradiated by the sun, the value is 1.

[0049] Furthermore, the model training module is further configured to:

[0050] Take each space target as a node of the graph structure, and set the edges according to the spatial relationships between the space targets;

[0051] Create node feature vectors using the orbital characteristics, non-conservative force time series, and non-conservative force characteristics of the space targets;

[0052] Select a graph convolutional long short-term memory network model for training. The graph convolutional long short-term memory network model includes a graph convolutional neural network and a long short-term memory network suitable for processing spatial relationships between nodes, and can comprehensively analyze node features, covering orbital characteristics, non-conservative force time series, and non-conservative force characteristics;

[0053] Import the data in the model training database into the graph convolutional long short-term memory network model for model training.

[0054] Furthermore, the model training module is further configured to:

[0055] During the training process, use a genetic algorithm to select hyperparameters for model parameter selection, and perform model training with the selected hyperparameters. The hyperparameters include the number of iterations, learning rate, batch sample size, and optimizer.

[0056] Furthermore, the model training module is further configured to:

[0057] Map the randomly generated hyperparameters to genotype encoding to create an initial population;

[0058] Calculate the fitness of each individual in the population. The fitness is defined as the target recognition accuracy under the current hyperparameters;

[0059] Select the initial population according to the fitness, and obtain the next generation of individuals through crossover and mutation, and calculate the fitness;

[0060] Select the group of encoding with the strongest fitness, and decode to obtain the corresponding hyperparameter values.

[0061] Furthermore, the data in the model training database includes feature vectors obtained by splicing and fusing the orbital features of each space target, the orbital time series under non-conservative forces, and non-conservative force features, as well as the category and number labels of the space targets.

[0062] According to the third technical solution of the present invention, a readable storage medium is provided. The readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the method described above.

[0063] The present invention has at least the following beneficial effects:

[0064] 1. By adopting a refined non-conservative force model, the estimation of non-conservative force characteristic parameters such as attitude, shape, size, surface-to-mass ratio, surface material, reflection coefficient, absorption coefficient, and atmospheric drag coefficient is more accurate.

[0065] 2. Sufficiently excavate the features in the orbital data that can be used for target recognition: the orbital features of each space target, the non-conservative force time series, and the implicit non-conservative force features.

[0066] 3. Represent space targets as nodes with orbital features, non-conservative force time series, and non-conservative force features, consider the relationships between targets, and use the GCN-LSTM network for target recognition and classification, improving the accuracy of target recognition.

[0067] 4. It has strong applicability and can achieve good recognition effects even for small-sized space objects in medium and high orbits. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 Shows a flowchart of a method for identifying and classifying space targets based on non-conservative force features according to an embodiment of the present invention;

[0069] Figure 2 Shows a structural diagram of a device for identifying and classifying space targets based on non-conservative force features according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0070] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be described in detail below in conjunction with the drawings and specific embodiments. The embodiments of the present invention will be further described in detail below in conjunction with the drawings and specific examples, but this is not a limitation to the present invention. For the various steps described herein, if there is no necessity for a sequential relationship between them, the order in which they are described as examples herein should not be regarded as a limitation. Those skilled in the art should know that they can be adjusted in order as long as the logic between them is not destroyed and the entire process cannot be realized.

[0071] Figure 1The flowchart of a spatial target recognition and classification method based on non-conservative force characteristics according to an embodiment of the present invention is shown. As Figure 1 shown, an embodiment of the present invention provides a spatial target recognition and classification method based on non-conservative force characteristics. The method includes steps S1 - S3, which are introduced in detail as follows.

[0072] Step S1, construct a feature space.

[0073] First, directly obtain the target orbit information from the GP category data, such as orbital inclination, eccentricity, right ascension of the ascending node, argument of perigee, mean anomaly, mean motion speed, etc. Then, compare the target orbit time series calculated using the Simplified General Perturbations (SGP4) algorithm with the orbit time series change under the influence of only conservative forces such as the Earth's gravity and the Sun's gravity, so as to strip the action of conservative forces from the orbit time series and obtain the time series of the spatial target under the action of only non-conservative forces. After that, solve the non-conservative force acceleration from the non-conservative force time series and substitute it into the non-conservative force model to realize the inversion of non-conservative force characteristics including attitude, shape, size, surface-to-mass ratio, surface material, reflection coefficient, absorption coefficient, and atmospheric drag coefficient. Finally, use the feature splicing and fusion technology to directly splice different features together to form a more comprehensive feature vector.

[0074] Analyzing the orbit time series requires the orbital parameters at the initial moment first. To achieve this goal, use the orbital data of ultra-short arc segments to estimate and correct the orbital parameters at the initial moment, and then use the corrected initial orbital parameters and combine all conservative force models for orbital integration to obtain the orbit affected only by conservative forces. However, the actual orbit time series is the result of the combined action of conservative forces and non-conservative forces. Therefore, by calculating the difference between the actual orbit and the conservative force orbit and estimating the orbital mechanical parameters accordingly, the orbit time series affected only by non-conservative forces is obtained. More specifically, first, define the orbital parameters at the initial moment: in orbital mechanics, the orbital parameters at the initial moment usually include orbital altitude, orbital velocity, orbital inclination, right ascension of the ascending node, argument of perigee, etc. Then, obtain the orbital data of ultra-short arc segments to accurately estimate and correct the orbital parameters at the initial moment. The orbital data of ultra-short arc segments refers to the high-precision orbital data within a certain time range, which can usually be obtained from various sensors and observation devices including spaceborne radars and optical telescopes. Then, estimate the orbital parameters at the initial moment: by least-squares fitting of the ultra-short arc segment orbital data, the corrected orbital parameters at the initial moment are obtained.

[0075] To obtain the orbital time series under the action of non-conservative forces, the action of conservative forces also needs to be eliminated. Conservative forces generally consider gravitational influences, that is, mainly study the gravitational forces of the Earth, the Sun, and other planets. The magnitude of conservative forces can be accurately quantified based on the celestial body mass and the distance to the target. Starting from the corrected initial orbital parameters and combining with the orbital integration method, the influence of conservative forces on the space target can be quantified, that is, the time series of the target under the action of only conservative forces can be obtained. By comparing with the orbital time series calculated using the SGP4 model, the orbital mechanical parameters are further estimated to remove the influence of conservative forces and calculate the orbital time series under the action of only non-conservative forces. This time series reflects the variation of non-conservative force acceleration and its influence on the space target, and can be used as a key feature for target recognition.

[0076] Features such as attitude, shape, size, and surface mass ratio are of great significance for target recognition and can effectively classify and identify target objects. Through the extraction and analysis of these features, more accurate and efficient target recognition algorithms and applications can be realized. The influence of non-conservative forces on each space target is closely related to the target's own attitude, shape, size, surface mass ratio, surface material, reflection coefficient, absorption coefficient, atmospheric drag coefficient, etc. Therefore, by analyzing the key information such as non-conservative force acceleration hidden in the non-conservative force time series, the corresponding non-conservative force features can be inverted. Combining with the unscented Kalman filter, with attributes such as attitude, size, surface mass ratio, surface material, reflection coefficient, absorption coefficient, and atmospheric drag coefficient as state variables, a system model is established based on the non-conservative force model of the space target, and non-conservative force acceleration and non-conservative force time series are used as observables, so as to achieve the optimal estimation of non-conservative force features.

[0077] Step S1, establish a database.

[0078] Specifically, in the process of creating the training database, one of the cores is to construct the feature vector of the space target. The feature vector of each space target consists of three parts: one part is the features related to the orbit; one part is the time series related to non-conservative force acceleration; and the other part is the features related to the attributes of the space target itself.

[0079] Orbital information such as orbital inclination, eccentricity, right ascension of the ascending node, argument of perigee, mean anomaly, and mean motion speed, the time series of the target under the action of only non-conservative forces, and non-conservative force characteristics such as attitude, shape, size, surface-to-mass ratio, surface material, reflection coefficient, absorption coefficient, and atmospheric drag coefficient are combined with information such as the number, category, and radar cross-section of the space target to improve the space target catalog database, so that the database includes target attributes such as attitude, shape, size, surface-to-mass ratio, surface material, reflection coefficient, absorption coefficient, and atmospheric drag coefficient, as well as related fields such as orbit, usage, and category. To achieve better target recognition results, it is also necessary to perform fusion processing on the existing features. The present invention uses feature splicing to obtain a comprehensive feature vector that is more conducive to recognition, and constructs a training data set based on this.

[0080] Step S3, use the GCN-LSTM neural network to perform recognition and classification of space targets.

[0081] The data set consists of the fused feature vectors and label vectors such as target number, category, and radar cross-section, and is divided into a training set, a validation set, and a test set according to the ratio of 8:1:1. An appropriate graph structure can greatly improve the target recognition performance of the graph neural network, so the design of the graph structure is very important. Adjust the data to a specific graph structure, that is, use each space target as a node of the graph structure, set the edges according to the spatial relationship between the targets, and use the spliced and fused features of the space targets as the node feature vectors. Input the data with the graph structure into the GCN-LSTM model and train it until the model converges to obtain a preliminary space target recognition result. Finally, use the genetic algorithm to adjust and optimize the hyperparameters to find a set of hyperparameters with the best target recognition performance of the model. Under the values of this set of hyperparameters, test the data in the test set to evaluate the recognition and classification accuracy of the trained GCN-LSTM model for space targets. Specifically, map the randomly generated hyperparameters to genotype encoding to create an initial population, and then calculate the fitness of each individual in the population. The fitness is defined as the target recognition accuracy under the current hyperparameters. Select the initial population according to the fitness, and obtain the next-generation individuals through crossover and mutation, and calculate the fitness. Repeat the above process multiple times, select a set of codes with the strongest fitness, and decode to obtain the corresponding hyperparameter values, that is, parameters such as the number of iterations, learning rate, batch sample size, and optimizer.

[0082] Figure 2 The structural diagram of a space target recognition and classification device based on non-conservative force characteristics according to an embodiment of the present invention is shown. An embodiment of the present invention provides a space target recognition and classification device based on non-conservative force characteristics, as Figure 2 shown, the device 200 includes:

[0083] The motion equation modeling module 201 is configured to model the motion equation of a space target through the following formula:

[0084]

[0085] where r is the position vector of the space target, μ is the gravitational parameter of the Earth, r = ||r||, is the non-spherical perturbation, a sun is the solar gravitational perturbation, a moon is the lunar gravitational perturbation, a SRP is the solar radiation pressure, a ERP is the Earth radiation pressure, a drag is the atmospheric drag;

[0086] The time series comparison module 202 is configured to compare the result after removing the influence of conservative forces with the original orbit time series to obtain the orbit time series only affected by non-conservative forces;

[0087] The key feature calculation module 203 is configured to model the non-conservative forces acting on the space target after removing the influence of conservative forces and analyze the influence of the non-conservative forces on the orbital elements, and further perform inversion to obtain the key features of the target. The non-conservative forces acting on the space target include the solar radiation pressure, the Earth radiation pressure, and the atmospheric drag acting on the space target. The key features of the target include the attitude, shape, size, surface-to-mass ratio, surface material, reflection coefficient, absorption coefficient, and atmospheric drag coefficient of the target;

[0088] The feature stitching and fusion module 204 is configured to use the orbit information, the orbit time series only affected by non-conservative forces, and the key features of the target to perform feature stitching and fusion to obtain a feature vector and establish a feature space for space target recognition and classification;

[0089] The database establishment module 205 establishes a model training database based on the feature space;

[0090] The model training module 206 is configured to train the model using the model training database and realize space target recognition and classification through the trained model.

[0091] In some embodiments, the key feature calculation module is further configured to:

[0092] For space targets with an orbital altitude ≥ 1000 km, taking the solar radiation pressure as the non-conservative force, the key features of the target are obtained through the following calculation model:

[0093]

[0094] where S F= 1367 W / m 2 , representing the solar radiation constant; c = 299792458 m / s, representing the speed of light propagation in vacuum; d is the distance from the space target to the sun in astronomical units AU; m object is the mass of the space target; A(i), ε(i), u n (i) respectively represent the total area, emissivity, and normal vector of the i-th surface element of the space target; R spec (i), R diff (i), R abs (i) respectively represent spectral reflectivity, diffuse reflectivity, and absorption coefficient. In the case of no energy loss during the transmission of electromagnetic waves through the surface element, the sum of spectral reflectivity, diffuse reflectivity, and absorption coefficient is 1; cos(ψ(i)) ≡ u n (i)·u sun represents the cosine of the angle between the i-th surface element and the sun. The function G[cos(ψ(i))] = max(0, cos(ψ(i))), which means that when the surface element is blocked, its value is 0, and when the surface element is irradiated by the sun, the value is 1.

[0095] In some embodiments, the model training module is further configured to:

[0096] Take each space target as a node of the graph structure and set the edges according to the spatial relationship between each space target;

[0097] Create node feature vectors using the orbital characteristics, non-conservative force time series, and non-conservative force characteristics of the space target;

[0098] Select the graph convolutional long short-term memory network model for training. The graph convolutional long short-term memory network model includes a graph convolutional neural network and a long short-term memory network suitable for processing the spatial relationship between nodes, and can comprehensively analyze node features, covering orbital characteristics, non-conservative force time series, and non-conservative force characteristics;

[0099] Import the data in the model training database into the graph convolutional long short-term memory network model for model training.

[0100] In some embodiments, the model training module is further configured to:

[0101] During the training process, use the genetic algorithm to select model parameters for hyperparameters, and perform model training with the selected hyperparameters. The hyperparameters include the number of iterations, learning rate, batch sample quantity, and optimizer.

[0102] In some embodiments, the model training module is further configured to:

[0103] Map the randomly generated hyperparameters to genotype coding to create an initial population;

[0104] Calculate the fitness of each individual in the population, where the fitness is defined as the target recognition accuracy under the current hyperparameters;

[0105] Select the initial population according to the fitness, and obtain the next-generation individuals through crossover and mutation, and calculate the fitness;

[0106] Select the group of codes with the strongest fitness and decode to obtain the corresponding hyperparameter values.

[0107] In some embodiments, the data in the model training database includes feature vectors spliced and fused from the orbital characteristics of each space target, the orbital time series under the action of non-conservative forces, and non-conservative force characteristics, as well as the category and number labels of the space targets.

[0108] It should be noted that the device described in this embodiment belongs to the same technical concept as the method described above and can achieve the same technical effects, which will not be elaborated here.

[0109] The embodiment of the present invention provides a readable storage medium, which stores one or more programs, and the one or more programs can be executed by one or more processors to implement the methods described in the above embodiments.

[0110] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more of their solutions) can be used in combination with each other. For example, those of ordinary skill in the art can use other embodiments when reading the above description. Additionally, in the above specific embodiments, various features can be grouped together to simplify the present invention. This should not be construed as an intention that the features of an unclaimed invention are necessary for any claim. On the contrary, the subject matter of the present invention may be less than all the features of a particular embodiment of the invention. Thus, the following claims are incorporated herein as examples or embodiments into the specific embodiments, where each claim stands alone as a separate embodiment, and it is contemplated that these embodiments can be combined with each other in various combinations or permutations. The scope of the present invention should be determined with reference to the appended claims and the full scope of the equivalents to which these claims are entitled.

Claims

1. A space target recognition and classification method based on non-conservative force features, characterized in that: The method comprises: The motion equation of the space target is modeled using the following formula: Among them, r is the position vector of the space target, μ is the gravitational parameter of the earth, r = ||r||, is a non-spherical perturbation, a sun is the solar gravitational perturbation, a moon is the lunar gravitational perturbation, a SRP is the solar radiation pressure, a ERP is the Earth's radiation pressure, a drag is the atmospheric drag; Compare the result after removing the conservative force with the original orbit time series to obtain the orbit time series affected only by the non-conservative force; After removing the influence of conservative forces, the non-conservative forces on the space target are modeled and the influence of the non-conservative forces on the orbital elements is analyzed, and further inversion is performed to obtain the key characteristics of the target. The non-conservative forces on the space target include the solar radiation pressure, the earth radiation pressure and the atmospheric drag on the space target. The key characteristics of the target include the target's attitude, shape, size, surface-to-mass ratio, surface material, reflection coefficient, absorption coefficient and atmospheric drag coefficient. By using orbital information, orbital time series affected only by non-conservative forces, and key features of the target, feature splicing and fusion are performed to obtain feature vectors, and a feature space for space target recognition and classification is established. Establishing a model training database based on the feature space; The model is trained using the model training database, and spatial target recognition and classification are achieved through the trained model.

2. The method according to claim 1, characterized in that After removing the conservative force, the non-conservative force on the space target is modeled and the impact of the non-conservative force on the orbital elements is analyzed. The key features of the target are obtained by further inversion, including: For space targets with an orbital altitude of ≥1000km, the solar radiation pressure is used as a non-conservative force and the key characteristics of the target are obtained through the following calculation model: In the formula, S F =1367W / m 2 , represents the solar radiation constant; c = 299792458m / s, represents the propagation speed of light in vacuum; d is the distance from the space target to the sun expressed in astronomical units AU; m object is the mass of the space target; A(i), ε(i), u n (i) represents the total area, emissivity and normal vector of the i-th facet of the space target; R spec (i) R diff (i) R abs (i) represents the spectral reflectivity, diffuse reflectivity and absorption coefficient respectively. When there is no energy loss in the transmission of electromagnetic waves through the surface element, the sum of the spectral reflectivity, diffuse reflectivity and absorption coefficient is 1; cos(ψ(i))≡u n (i) u sun Represents the cosine of the angle between the i-th face element and the sun. The function G[cos(ψ(i))]=max(0,cos(ψ(i))) means that when the face element is blocked, its value is 0, and when the face element is exposed to solar radiation, its value is 1.

3. The method according to claim 1, characterized in that The data in the model training database include orbital characteristics of each space target, feature vectors spliced ​​and fused by orbital time series under the action of non-conservative forces and non-conservative force characteristics, and category and number labels of space targets.

4. The method according to claim 1, characterized in that The model is trained using the model training database, including: Each spatial target is regarded as a node of the graph structure, and edges are set according to the spatial relationship between each spatial target; Create node feature vectors using orbital features, non-conservative force time series, and non-conservative force features of space targets; A graph convolutional long short-term memory network model is selected for training, wherein the graph convolutional long short-term memory network model includes a graph convolutional neural network and a long short-term memory network suitable for processing spatial relationships between nodes, and can comprehensively analyze node characteristics, including orbital characteristics, non-conservative force time series, and non-conservative force characteristics; The data in the model training database is imported into the graph convolutional long short-term memory network model for model training.

5. The method according to claim 4, characterized in that After importing the data in the model training database into the graph convolutional long short-term memory network model for model training, the method further includes: During the training process, the genetic algorithm model parameters are used to select hyperparameters, and the model is trained with the selected hyperparameters, which include the number of iterations, the learning rate, the number of batch samples, and the optimizer.

6. The method according to claim 5, characterized in that The method of selecting hyper parameters by using genetic algorithm model parameters includes: Map randomly generated hyperparameters to genotype codes to create an initial population; Calculate the fitness of each individual in the population. Fitness is defined as the target recognition accuracy under the current hyperparameters. The initial population is selected according to fitness, and the next generation of individuals is obtained through crossover and mutation, and the fitness is calculated; Select the set of codes with the strongest fitness and decode them to get the corresponding hyperparameter values.

7. A space target recognition and classification device based on non-conservative force features, characterized in that: The device comprises: The motion equation modeling module is configured to perform motion equation modeling of a space target using the following formula: Among them, r is the position vector of the space target, μ is the gravitational parameter of the earth, r = ||r||, is a non-spherical perturbation, a sun is the solar gravitational perturbation, a moon is the lunar gravitational perturbation, a SRP is the solar radiation pressure, a ERP is the Earth's radiation pressure, a drag is the atmospheric drag; A time series comparison module is configured to compare the result after removing the influence of conservative forces with the original orbit time series to obtain the orbit time series under the influence of non-conservative forces only; A key feature calculation module is configured to model the non-conservative forces on the space target after removing the influence of the conservative forces, analyze the influence of the non-conservative forces on the orbital elements, and further invert to obtain the key features of the target. The non-conservative forces on the space target include the solar radiation pressure, the earth radiation pressure and the atmospheric drag on the space target. The key features of the target include the target's attitude, shape, size, surface mass ratio, surface material, reflection coefficient, absorption coefficient and atmospheric drag coefficient. A feature splicing and fusion module is configured to utilize orbit information, orbit time series only affected by non-conservative forces, and key features of the target to perform feature splicing and fusion to obtain feature vectors and establish a feature space for space target recognition and classification; A database establishment module, establishing a model training database based on the feature space; The model training module is configured to train the model using the model training database, and realize space target recognition and classification through the trained model.

8. The device according to claim 7, characterized in that The key feature calculation module is further configured as follows: For space targets with an orbital altitude of ≥1000km, the solar radiation pressure is used as a non-conservative force and the key characteristics of the target are obtained through the following calculation model: In the formula, S F =1367W / m 2 , represents the solar radiation constant; c = 299792458m / s, represents the propagation speed of light in vacuum; d is the distance from the space target to the sun expressed in astronomical units AU; m object is the mass of the space target; A(i), ε(i), u n (i) represents the total area, emissivity and normal vector of the i-th facet of the space target; R spec (i) R diff (i) R abs (i) represents the spectral reflectivity, diffuse reflectivity and absorption coefficient respectively. When there is no energy loss in the transmission of electromagnetic waves through the surface element, the sum of the spectral reflectivity, diffuse reflectivity and absorption coefficient is 1; cos(ψ(i))≡u n (i) u sun Represents the cosine of the angle between the i-th face element and the sun. The function G[cos(ψ(i))]=max(0,cos(ψ(i))) means that when the face element is blocked, its value is 0, and when the face element is exposed to solar radiation, its value is 1.

9. The device according to claim 7, characterized in that The model training module is further configured as follows: Each spatial target is regarded as a node of the graph structure, and edges are set according to the spatial relationship between each spatial target; Create node feature vectors using orbital features, non-conservative force time series, and non-conservative force features of space targets; A graph convolutional long short-term memory network model is selected for training, wherein the graph convolutional long short-term memory network model includes a graph convolutional neural network and a long short-term memory network suitable for processing spatial relationships between nodes, and can comprehensively analyze node characteristics, including orbital characteristics, non-conservative force time series, and non-conservative force characteristics; The data in the model training database is imported into the graph convolutional long short-term memory network model for model training.

10. A readable storage medium, characterized in that: The readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the method according to any one of claims 1 to 7.

Images