A digital twin system for low-orbit giant star constellation systems and its construction method

Abstract

The present invention belongs to the field of aerospace technology and relates to a digital twin system for low-orbit giant star constellations and a construction method thereof. The system includes a data management module that comprehensively manages the real-time data, vertical data, horizontal data, simulation data, and fusion data of the giant star constellation system; a model management module that dynamically constructs and comprehensively characterizes the geometric, physical, behavioral, rules, and constraint characteristics of complex physical systems from different dimensions, different spatial scales, and different time scales; an information management module utilizes these multi-scale and multi-level characteristic information to perform local to global network reconstruction of the low-orbit giant star constellation system and restore the complete information network; a collaborative coupling mechanism of data interconnection, information exchange, and model interoperability is adopted to achieve efficient and real-time monitoring and management, providing a solution for digital, visual, and intelligent supervision.

Description

Technical Field

[0001] The present invention relates to the field of low-orbit giant star constellation control technology, and in particular to a digital twin system for a low-orbit giant star constellation system and a construction method thereof. Background Art

[0002] In recent years, low-orbit space has become the focus of global aerospace development. Since low-orbit satellite constellations have obvious advantages in launch costs and global coverage, and have strong anti-destruction capabilities, low transmission delays, and low power links, the low-orbit giant constellation system, characterized by low orbit and large scale, is showing a booming development trend worldwide.

[0003] Unlike traditional single satellites, the Giant Star constellation system has a large number of systems, dynamic and time-varying links and topology, and complex network spatiotemporal behavior. Traditional satellite management models alone cannot meet its operational needs. Therefore, an integrated system comprehensive management model is urgently needed to conduct comprehensive analysis and virtual verification of the problems faced by the Giant Star constellation throughout its life cycle, so as to improve the basic capabilities and management level of the constellation system. Digital twin technology is the key to digital construction in the constellation field.

[0004] As a means of digital development in the constellation sector, digital twins theoretically connect satellite constellations and ground base stations. They provide digital representations of physical objects, interactive connections between physical and virtual interfaces, data integration, fusion, analysis, and mining, and the service-oriented packaging and application of models, data, and functions. Therefore, research into integrating digital twins with the constellation industry will effectively promote digitalization, networking, intelligence, and service-oriented transformation and upgrading, playing a key role in the engineering applications of giant constellations. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a digital twin system for a low-orbit giant star constellation system and a construction method thereof.

[0006] In order to achieve the purpose of the present invention, the present invention will be implemented by adopting the following technical solutions.

[0007] A digital twin system for a low-orbit giant star constellation system includes a data management module, a model management module, and an information management module. The data management module and the model management module interact with each other through data interconnection and model interoperability; the data management module and the information management module interact with each other through data interconnection and information exchange; and the model management module and the information management module interact with each other through model interoperability and information exchange.

[0008] The data management module is a module that comprehensively manages the real-time data, vertical data, horizontal data, simulation data, and fusion data of the low-orbit giant constellation system to achieve data connection and interaction between the physical system and the virtual system; the comprehensive management is implemented through data processing, data classification, data association, and data integration; wherein:

[0009] The data processing is to clean, filter, remove noise and fill missing values ​​of the real-time data, longitudinal data, transverse data, simulation data and fusion data of the low-orbit giant constellation system to ensure the quality and accuracy of the original data;

[0010] The data classification is to decode the data after the data processing and convert it into a format that the system can understand and process. At the same time, the data is organized and labeled and a unified language format is set to ensure that data from different sources can be understood and used within a unified framework;

[0011] The data association is to connect multiple groups of data points after the data classification processing, explore the internal connections and mutual influences, and ensure the compatibility and operability between different data sets;

[0012] The data integration is to integrate data from different sources and different association relationships into a unified data set after the data is associated to support system modeling and analysis, and to perform data conversion and standardization on the integrated data to ensure that the data can work together to support the model management module and the information management module, ensure that the data can be effectively applied and analyzed, and ensure the interoperability of data between different models and modules;

[0013] The real-time data is measured data from on-orbit satellite sensors;

[0014] The longitudinal data covers the historical records, periodic data and performance evolution trends of the LEO giant star constellation system;

[0015] The horizontal data focuses on the relationship between satellites or subsystems, including communication, mutual influence and performance differences;

[0016] The simulation data refers to the theoretical analysis data obtained by simulation of the twin system at the digital level;

[0017] The fused data is a complete set of data generated by integrating and reconstructing real-time data, longitudinal data, transverse data and simulation data;

[0018] The model management module uses geometric models, simulation models, mechanism models and data models to dynamically construct and comprehensively characterize the geometry, physics, behavior, rules and constraint characteristics of the physical system from different dimensions, different spatial scales and different time scales;

[0019] The information management module uses the Transformer model as a tool for feature extraction and representation learning in a computing environment provided by a cloud computing platform and distributed training technology, extracts coarse-grained features of twin data information and twin model information at the token level, aggregates all token information at the output of the Encoder by weighted averaging and summarization, and further encodes the input token-level position information into a global representation through position embedding to form a global view; introduces a local window attention branch to learn short-range interactions within the local window, extracts fine-grained feature representations of twin data information and twin model information, and captures and extracts token-level local features; introduces a multi-scale attention mechanism to explore the integration of the two features, refines spatial features in a cross-scale manner through Multi-Head Attention, and captures feature relationships at different scales; uses Reinforcement The learning algorithm optimizes the adaptive fusion process of features and defines state, action, and reward functions, where the state represents the feature representation of the current data, the action represents the method of feature fusion, and the reward function evaluates the quality of the fusion result. Based on the multi-scale attention fusion (MAF) scheme, the interactive connection between global-local token pairs is enhanced, and multiple token-level agents at different stages are trained together in a unified framework. Both local and global features are deployed using linear operations to ensure the consistency of feature distribution. Based on the model's learning process and data feedback, the optimal feature fusion strategy is determined through the reinforcement learning algorithm, and the weights and fusion strategies of global and local features are dynamically adjusted to ensure that the feature fusion process can maximize the performance of the information management module.

[0020] As a preferred embodiment of the present invention, the geometric model is a comprehensive description of the positional relationship of each satellite in the low-orbit giant star constellation system in space by integrating Kepler's laws, quaternions, orbital perturbation theory, spherical harmonics, and orbital element expert knowledge;

[0021] The simulation model simulates the motion trajectory of the satellite constellation in space using dynamics and control models and atmospheric disturbance models;

[0022] The mechanism model is based on expert knowledge of orbital mechanics, communication protocols, laser communication theory, electromagnetic wave propagation theory, and system engineering theory, providing a physical explanation of the behavior of the entire LEO giant star system and a description of the internal logic rules;

[0023] The data model uses statistical methods and machine learning techniques to capture the complex relationships between satellite orbit information from the data management module.

[0024] A method for constructing a digital twin system includes the following steps:

[0025] S1. Build a data management module through data processing, data classification, data association, and data integration to comprehensively manage the real-time data, vertical data, horizontal data, simulation data, and fusion data collected from the low-orbit giant constellation system to achieve data connection and interaction between physical and virtual systems;

[0026] S2. Build a model management module through geometric models, simulation models, mechanism models, and data models to dynamically construct and comprehensively characterize the geometry, physics, behavior, rules, and constraint characteristics of physical systems from different dimensions, spatial scales, and time scales.

[0027] S3. Build an information management module through global feature aggregation, local feature extraction, cross-scale spatial feature refinement, and adaptive feature fusion. Reconstruct the digital twin system from local to global networks from different dimensions, spatial scales, and time scales to restore the complete information network. The specific process of restoring the complete information network is completed in the computing environment provided by the cloud computing platform and distributed training technology, and includes the following steps:

[0028] S31. Using the Transformer model as a tool for feature extraction and representation learning, we extract coarse-grained features of data and model information at the token level. We aggregate all token information at the encoder output through weighted averaging and summarization. We further encode the input token-level position information into a global representation through position embedding to form a global view.

[0029] S32: Introduce a local window attention branch to learn short-range interactions within the local window, extract fine-grained feature representations of data and model information, and capture token-level local features.

[0030] S33. Introducing a multi-scale attention mechanism to explore the integration of two features. Through Multi-Head Attention, we refine spatial features in a cross-scale manner and capture feature relationships at different scales.

[0031] S34. Utilize the Reinforcement Learning algorithm to optimize the adaptive feature fusion process and define state, action, and reward functions. The state represents the feature representation of the current data, the action represents the method of feature fusion, and the reward function evaluates the quality of the fusion result. Based on the multi-scale attention fusion (MAF) scheme, the interactive connection between global and local token pairs is enhanced. Multiple token-level agents at different stages are trained together in a unified framework. Both local and global features are deployed using linear operations to ensure the consistency of feature distribution.

[0032] S35. Based on the model's learning process and data feedback, the optimal feature fusion strategy is determined through the reinforcement learning algorithm, and the weights and fusion strategies of global features and local features are dynamically adjusted so that the feature fusion process can maximize the performance of the information management module.

[0033] As a preferred embodiment of the present invention, step S1 includes the following specific steps:

[0034] S11. Process the acquired real-time data, longitudinal data, transverse data, simulation data, and fusion data of the LEO giant star constellation system by cleaning, filtering, denoising, and filling in missing values ​​to ensure the quality and accuracy of the original data;

[0035] S12. Decode the data processed in step S11 and convert it into a format that the system can understand and process. At the same time, organize and identify the data and set a unified language format to ensure that data from different sources can be understood and used within a unified framework.

[0036] S13. Connect multiple data points in a unified language format after being decoded, organized, and labeled in step S12 through data association to explore internal connections and mutual influences to ensure compatibility and operability between different data sets;

[0037] S14. For the data associated in step S13, data from different sources and with different association relationships are integrated into a unified data set through data integration to support system modeling and analysis;

[0038] S15. Convert and standardize the data integrated in step S14 to ensure that the data can work together to support model management and information management, ensure that the data can be effectively applied and analyzed, and ensure the interoperability of data between different models and modules.

[0039] As a preferred embodiment of the present invention, step S2 includes the following specific steps:

[0040] S21. Integrate satellite orbit parameters, attitude, reference coordinate system, position and velocity vectors, coordinate system transformation matrix, attitude quaternion, and relative position vector parameter information to form a geometric model to accurately describe the positional relationship of each satellite in the LEO Giant Star constellation system in space;

[0041] S22. Build a simulation model based on a satellite control system designed based on linear and nonlinear control theory to ensure the stability of the attitude and position of each satellite within the constellation.

[0042] S23. Based on expert knowledge in orbital mechanics, communication protocols, laser communication theory, electromagnetic wave propagation theory, and systems engineering theory, construct orbital mechanics models, communication protocol models, and inter-satellite-to-ground communication models. These models are then integrated into a mechanism model to provide a physical explanation of system behavior and a description of the rules governing its internal logic.

[0043] S24. Use statistical methods and machine learning techniques to build data models to accurately capture the complex relationships between satellite orbit information from the data management module;

[0044] S25. Comprehensively manage the functional characteristics of different models to build an accurate and complete digital twin model of the giant star system. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a schematic diagram of the structure of the digital twin system of the present invention;

[0046] Figure 2 This is a structural diagram of the data management module of the present invention;

[0047] Figure 3 This is a flow chart of the model management module of the present invention;

[0048] Figure 4 This is a flow chart of the information management module of the present invention. DETAILED DESCRIPTION

[0049] As an embodiment of the present invention, Figures 1 to 4 As shown, a digital twin system for a low-orbit giant star constellation system includes a data management module, a model management module, and an information management module. The data management module and the model management module interact with each other through data interconnection and model interoperability; the data management module and the information management module interact with each other through data interconnection and information exchange; and the model management module and the information management module interact with each other through model interoperability and information exchange.

[0050] The data management module is a module that comprehensively manages the real-time data, vertical data, horizontal data, simulation data, and fusion data of the low-orbit giant constellation system to achieve data connection and interaction between the physical system and the virtual system; the comprehensive management is implemented through data processing, data classification, data association, and data integration; wherein:

[0051] The data processing is to clean, filter, remove noise and fill missing values ​​of the real-time data, longitudinal data, transverse data, simulation data and fusion data of the low-orbit giant constellation system to ensure the quality and accuracy of the original data;

[0052] The data classification is to decode the data after the data processing and convert it into a format that the system can understand and process. At the same time, the data is organized and labeled and a unified language format is set to ensure that data from different sources can be understood and used within a unified framework;

[0053] The data association is to connect multiple groups of data points after the data classification processing, explore the internal connections and mutual influences, and ensure the compatibility and operability between different data sets;

[0054] The data integration is to integrate data from different sources and different association relationships into a unified data set after the data is associated to support system modeling and analysis, and to perform data conversion and standardization on the integrated data to ensure that the data can work together to support the model management module and the information management module, ensure that the data can be effectively applied and analyzed, and ensure the interoperability of data between different models and modules;

[0055] The real-time data is derived from actual measurements by on-orbit satellite sensors, providing key real-time insights for real-time monitoring and decision-making;

[0056] The longitudinal data covers the historical records, periodic data and performance evolution trends of the LEO giant star system, which helps to understand the performance changes, problem trends and long-term evolution patterns of the LEO giant star system in different time periods;

[0057] The horizontal data focuses on the interrelationships between satellites or subsystems, including communications, mutual influences, and performance differences. Data records of similar or similar constellations in a certain LEO giant star constellation are used as a reference and comparison to optimize the coordinated operation of the entire LEO giant star constellation system and improve overall performance.

[0058] The simulation data refers to theoretical analysis data obtained through simulation of the digital twin system at the digital level. It supports virtual testing of the digital twin system's response, helps to identify potential problems before actual operation, and improves system design and performance.

[0059] The fusion data is generated by integrating and reconstructing real-time data, vertical data, horizontal data and simulation data into a more comprehensive and accurate data set that runs through theory, simulation and reality;

[0060] The model management module uses geometric models, simulation models, mechanism models and data models to dynamically construct and comprehensively characterize the geometry, physics, behavior, rules and constraint characteristics of the physical system from different dimensions, different spatial scales and different time scales;

[0061] The information management module uses the Transformer model as a tool for feature extraction and representation learning in a computing environment provided by a cloud computing platform and distributed training technology, extracts coarse-grained features of twin data information and twin model information at the token level, aggregates all token information at the output of the Encoder by weighted averaging and summarization, and further encodes the input token-level position information into a global representation through position embedding to form a global view; introduces a local window attention branch to learn short-range interactions within the local window, extracts fine-grained feature representations of twin data information and twin model information, and captures and extracts token-level local features; introduces a multi-scale attention mechanism to explore the integration of the two features, refines spatial features in a cross-scale manner through Multi-Head Attention, and captures feature relationships at different scales; uses Reinforcement The learning algorithm optimizes the adaptive fusion process of features and defines state, action, and reward functions, where the state represents the feature representation of the current data, the action represents the method of feature fusion, and the reward function evaluates the quality of the fusion result. Based on the multi-scale attention fusion (MAF) scheme, the interactive connection between global-local token pairs is enhanced, and multiple token-level agents at different stages are trained together in a unified framework. Both local and global features are deployed using linear operations to ensure the consistency of feature distribution. Based on the model's learning process and data feedback, the optimal feature fusion strategy is determined through the reinforcement learning algorithm, and the weights and fusion strategies of global and local features are dynamically adjusted to ensure that the feature fusion process can maximize the performance of the information management module.

[0062] The connection between physical and virtual systems emphasizes data interconnection, information exchange, and model interoperability. Therefore, a digital twin system for the LEO system is designed with three modules: data management, model management, and information management. As the driving force of the digital twin system, data is characterized by massive volume, strong correlation, and high coupling. Due to the limited transit time of satellites, data transmission has a certain timeliness requirement. Furthermore, LEO data is fragmented and discrete. Therefore, a data management module is designed to comprehensively manage the real-time, vertical, horizontal, simulation, and fused data of the physical system. This includes data processing, data classification, data association, and data integration, enabling data connectivity and interaction between the physical and virtual systems. Models, as the carriers of the digital twin system, are used to simulate and emulate the operational processes of the physical world. Therefore, a model management module, comprising geometric, simulation, mechanism, and data models, is designed to dynamically construct and comprehensively characterize the geometric, physical, behavioral, rules, and constraints characteristics of complex physical systems from different dimensions, spatial, and temporal scales. Relying on twin data and twin models, the information management module uses these multi-scale and multi-level feature information to reconstruct the network of the digital twin system from local to global. Through global feature aggregation, local feature extraction, cross-scale refinement of spatial features and adaptive feature fusion, it restores the complete information network and builds a digital twin system.

[0063] As an embodiment of the present invention, Figure 3 As shown, the geometric model is a comprehensive description of the positional relationship of each satellite in the low-orbit giant star system in space by integrating Kepler's law, quaternions, orbital perturbation theory, spherical harmonics, and orbital roots.

[0064] The simulation model simulates the motion trajectory of the satellite constellation in space using dynamics and control models and atmospheric disturbance models;

[0065] The mechanism model is based on expert knowledge of orbital mechanics, communication protocols, laser communication theory, electromagnetic wave propagation theory, and system engineering theory, providing a physical explanation of the behavior of the entire LEO giant star system and a description of the internal logic rules;

[0066] The data model uses statistical methods and machine learning techniques to capture the complex relationships between satellite orbit information from the data management module.

[0067] As an embodiment of the present invention, Figure 1 As shown, a method for constructing a digital twin system includes the following steps:

[0068] S1. Build a data management module through data processing, data classification, data association, and data integration to comprehensively manage the real-time data, vertical data, horizontal data, simulation data, and fusion data collected from the low-orbit giant constellation system to achieve data connection and interaction between physical and virtual systems;

[0069] S2. Build a model management module through geometric models, simulation models, mechanism models, and data models to dynamically construct and comprehensively characterize the geometry, physics, behavior, rules, and constraint characteristics of the physical system from different dimensions, spatial scales, and time scales;

[0070] S3. Build an information management module through global feature aggregation, local feature extraction, cross-scale refinement of spatial features, and adaptive fusion of features. Reconstruct the network of the digital twin system from local to global from different dimensions, different spatial scales, and different time scales to restore the complete information network.

[0071] As an embodiment of the present invention, Figure 2 As shown, step S1 includes the following steps:

[0072] S11. Process the acquired physical system data by cleaning, filtering, denoising, and filling in missing values ​​to ensure the quality and accuracy of the original data;

[0073] S12. Decode the data processed in step S11 and convert it into a format that the system can understand and process. At the same time, organize and identify the data and set a unified language format to ensure that data from different sources can be understood and used within a unified framework.

[0074] S13. Connect multiple data points in a unified language format after being decoded, organized, and labeled in step S12 through data association to explore internal connections and mutual influences to ensure compatibility and operability between different data sets;

[0075] S14. For the data associated in step S13, data from different sources and with different association relationships are integrated into a unified data set through data integration to support system modeling and analysis;

[0076] S15. For the data integrated in step S14, ensure that the data can work together through data conversion and standardization to support model management and information management, ensure that the data can be effectively applied and analyzed, and ensure the interoperability of data between different models and modules.

[0077] As an embodiment of the present invention, Figure 3 As shown, step S2 includes the following steps:

[0078] S21. Integrate satellite orbit parameters, attitude, reference coordinate system, position and velocity vectors, coordinate system transformation matrix, attitude quaternion, and relative position vector parameter information to form a geometric model to accurately describe the positional relationship of each satellite in the LEO Giant Star constellation system in space;

[0079] S22. Build a simulation model based on a satellite control system designed based on linear and nonlinear control theory to ensure the stability of the attitude and position of each satellite within the constellation.

[0080] S23. Based on expert knowledge of orbital mechanics, communication protocols, laser communication theory, electromagnetic wave propagation theory, and systems engineering theory, construct orbital mechanics models, communication protocol models, and inter-satellite-to-ground communication models. These models are then integrated into a mechanism model to provide a physical explanation of system behavior and a description of the rules governing its internal logic.

[0081] S24. Use statistical methods and machine learning techniques to build data models to accurately capture the complex relationships between satellite orbit information from the data management module;

[0082] S25. Comprehensively manage the functional characteristics of different models to build an accurate and complete digital twin model of the giant star system.

[0083] As an embodiment of the present invention, Figure 4 As shown, step S3 includes the following steps:

[0084] S31. Using the Transformer model as a tool for feature extraction and representation learning, we extract coarse-grained features of data and model information at the token level. We aggregate all token information at the encoder output through weighted averaging and summarization. We further encode the input token-level position information into a global representation through position embedding to form a global view.

[0085] S32: Introduce a local window attention branch to learn short-range interactions within the local window, extract fine-grained feature representations of data and model information, and capture token-level local features.

[0086] S33. Introducing a multi-scale attention mechanism to explore the integration of two features. Through Multi-Head Attention, we refine spatial features in a cross-scale manner and capture feature relationships at different scales.

[0087] S34. Utilize the Reinforcement Learning algorithm to optimize the adaptive feature fusion process and define state, action, and reward functions. The state represents the feature representation of the current data, the action represents the method of feature fusion, and the reward function evaluates the quality of the fusion result. Based on the Multi-Scale Attention Fusion (MAF) scheme, the interactive connection between global and local token pairs is enhanced. Multiple token-level agents at different stages are trained together in a unified framework. Both local and global features are deployed using linear operations to ensure consistent feature distribution.

[0088] S35. Based on the model's learning process and data feedback, the optimal feature fusion strategy is determined through a reinforcement learning algorithm, and the weights and fusion strategies of global and local features are dynamically adjusted so that the feature fusion process can maximize the performance of the information management module.

[0089] S36. Use cloud computing platform and distributed training technology to provide an efficient, flexible and scalable computing environment.

[0090] The technical solution of the present invention is described in detail above in conjunction with the embodiments / drawings, but the present invention is not limited to the above technical solution. For ordinary technicians in this technical field, after knowing the contents recorded in the present invention, they can make several equivalent transformations and substitutions without departing from the principles of the present invention. These equivalent transformations and substitutions should also be regarded as falling within the scope of protection of the present invention.

Claims

1. A digital twin system for a low-Earth orbit giant system, characterized by: The digital twin system includes a data management module, a model management module and an information management module. The data management module and the model management module interact with each other through data interconnection and model interoperability; the data management module and the information management module interact with each other through data interconnection and information exchange; the model management module and the information management module interact with each other through model interoperability and information exchange; wherein: The data management module is a module that comprehensively manages the real-time data, vertical data, horizontal data, simulation data, and fusion data of the low-orbit giant constellation system; the comprehensive management includes data processing, data classification, data association, and data integration; wherein: The data processing is to clean, filter, remove noise and fill missing values ​​of the real-time data, longitudinal data, transverse data, simulation data and fusion data of the low-orbit giant star constellation system obtained; The data classification is to decode the data after the data processing and convert it into a format that the system can understand and process. At the same time, the data is organized and marked and a unified language format is set; The data association is to connect multiple groups of data points after the data classification processing; The data integration is to integrate the data from different sources and different association relationships into a unified data set after the data association, and perform data conversion and standardization on the data after data integration; The real-time data is measured data from on-orbit satellite sensors; The longitudinal data covers the historical records, periodic data and performance evolution trends of the LEO giant star constellation system; The horizontal data focuses on the relationship between satellites or subsystems, including communication, mutual influence and performance differences; The simulation data refers to the theoretical analysis data obtained by simulation of the digital twin system at the digital level; The fused data is a complete set of data generated by integrating and reconstructing real-time data, longitudinal data, transverse data and simulation data; The model management module uses geometric models, simulation models, mechanism models and data models to dynamically construct and comprehensively characterize the geometry, physics, behavior, rules and constraint characteristics of the physical system from different dimensions, different spatial scales and different time scales; The information management module uses the Transformer model as a tool for feature extraction and representation learning in a computing environment provided by a cloud computing platform and distributed training technology, extracts coarse-grained features of twin data information and twin model information at the token level, aggregates all token information on the output of the Encoder by weighted averaging and summarization, and encodes the input token-level position information into a global representation through position embedding to form a global view; by introducing a local window attention branch, short-range interactions within the local window are learned, and fine-grained feature representations of twin data information and twin model information are extracted to capture and extract token-level local features; a multi-scale attention mechanism is introduced to explore the integration of the two features, and spatial features are refined in a cross-scale manner through Multi-Head Attention to capture feature relationships at different scales; and reinforcement learning is used to learn short-range interactions within the local window. The learning algorithm optimizes the adaptive fusion process of features and defines state, action, and reward functions. The state represents the feature representation of the current data, the action represents the method of feature fusion, and the reward function evaluates the quality of the fusion result. Based on the multi-scale attention fusion (MAF) scheme, the interactive connection between global and local token pairs is enhanced, and multiple token-level agents at different stages are jointly trained in a unified framework. Both local and global features are deployed using linear operations. Based on the model's learning process and data feedback, the feature fusion strategy is determined through the reinforcement learning algorithm, and the weights and fusion strategies of global and local features are dynamically adjusted.

2. The digital twin system for the LEO giant star system according to claim 1, characterized in that: The geometric model is a combination of Kepler's laws, quaternions, orbital perturbation theory, spherical harmonics, and orbital element expert knowledge to describe the positional relationship of each satellite in the low-orbit giant star system in space; The simulation model simulates the motion trajectory of the satellite constellation in space using dynamics and control models and atmospheric disturbance models; The mechanism model is based on expert knowledge of orbital mechanics, communication protocols, laser communication theory, electromagnetic wave propagation theory, and system engineering theory, providing a physical explanation of the behavior of the entire LEO giant star system and a description of the internal logic rules; The data model uses statistical methods and machine learning techniques to capture the complex relationships between satellite orbit information from the data management module.

3. A method for constructing a digital twin system, characterized by: The following steps are involved: S1. Build a data management module through data processing, data classification, data association, and data integration to comprehensively manage the real-time data, vertical data, horizontal data, simulation data, and fusion data collected from the low-orbit giant constellation system; S2. Build a model management module through geometric models, simulation models, mechanism models, and data models to dynamically construct and comprehensively characterize the geometry, physics, behavior, rules, and constraint characteristics of physical systems from different dimensions, spatial scales, and time scales. S3. Build an information management module through global feature aggregation, local feature extraction, cross-scale spatial feature refinement, and adaptive feature fusion. Reconstruct the digital twin system from local to global networks from different dimensions, spatial scales, and time scales to restore the complete information network. The specific process of restoring the complete information network is completed in the computing environment provided by the cloud computing platform and distributed training technology, and includes the following steps: S31. Using the Transformer model as a tool for feature extraction and representation learning, we extract coarse-grained features of data and model information at the token level. We aggregate all token information at the encoder output through weighted averaging and summarization. We further encode the input token-level position information into a global representation through position embedding to form a global view. S32: Introduce a local window attention branch to learn short-range interactions within the local window, extract fine-grained feature representations of data and model information, and capture token-level local features. S33. Introducing a multi-scale attention mechanism to explore the integration of two features. Through Multi-Head Attention, we refine spatial features in a cross-scale manner and capture feature relationships at different scales. S34. Utilize the Reinforcement Learning algorithm to optimize the adaptive feature fusion process and define state, action, and reward functions. The state represents the feature representation of the current data, the action represents the method of feature fusion, and the reward function evaluates the quality of the fusion result. Based on the multi-scale attention fusion (MAF) scheme, the interactive connection between global and local token pairs is enhanced. Multiple token-level agents at different stages are trained together in a unified framework. Both local and global features are deployed using linear operations. S35. Based on the model's learning process and data feedback, the feature fusion strategy is determined through the reinforcement learning algorithm, and the weights and fusion strategies of global features and local features are dynamically adjusted.

4. The method for constructing a digital twin system according to claim 3, characterized in that: The step S1 includes the following specific steps: S11. Process the acquired real-time data, longitudinal data, transverse data, simulation data, and fusion data of the LEO giant star constellation system by cleaning, filtering, denoising, and filling in missing values; S12, decoding the data processed in step S11 and converting it into a format that the system can understand and process. At the same time, organizing and identifying the data and setting a unified language format; S13, connecting multiple groups of data points through data association for the data in a unified language format that has been decoded, organized, and labeled in step S12; S14, integrating the data associated in step S13 from different sources and with different association relationships into a unified data set through data integration; S15. Perform data conversion and standardization on the data integrated in step S14.

5. The method for constructing a digital twin system according to claim 3, wherein: The step S2 includes the following specific steps: S21, integrating satellite orbit parameters, attitude, reference coordinate system, position and velocity vectors, coordinate system transformation matrix, attitude quaternion, and relative position vector parameter information to form a geometric model; S22. Satellite control systems designed based on linear and nonlinear control theory are used to establish simulation models; S23. Based on the expert knowledge of orbital mechanics, communication protocols, laser communication theory, electromagnetic wave propagation theory, and system engineering theory, construct an orbital mechanics model, a communication protocol model, and an inter-satellite-to-ground communication model, and integrate the orbital mechanics model, communication protocol model, and inter-satellite-to-ground communication model into a mechanism model. S24. Use statistical methods and machine learning techniques to build data models; S25. Build a digital twin model of the giant star system by comprehensively managing the functional characteristics of different models.

Images