A method for state synchronization of low-Earth orbit satellite network twin emulators

Abstract

This invention belongs to the field of network technology, specifically relating to a method for synchronizing the state of a virtual twin emulator in a low-Earth orbit (LEO) satellite network. The method includes: collecting telemetry data from the LEO satellite network; parsing each telemetry data packet and classifying it in real time, including real-time valid, delayed valid, and invalid data; when a telemetry data packet is classified as real-time valid, generating a flow table strategy based on the telemetry data packet; using the flow table strategy to drive the virtual twin network; constructing spatiotemporal graphs of multiple historical moments and inputting them into a spatiotemporal graph neural network prediction model for processing to obtain the predicted network state for future moments; when a telemetry data packet is classified as invalid, using the predicted network state to generate a flow table strategy and drive the virtual twin network; when a telemetry data packet is classified as delayed valid, performing adaptive backtracking calibration based on the predicted network state to obtain a corrected network state; and replacing the predicted network state at the corresponding moment with the corrected network state. This invention achieves high-fidelity state synchronization of the twin and autonomous evolution of the model under real communication constraints.

Description

Technical Field

[0001] This invention belongs to the field of network technology, specifically relating to a method for synchronizing the state of a twin emulator of a low-orbit satellite network. Background Technology

[0002] With the large-scale deployment of low Earth orbit (LEO) satellite constellations, satellite networks are increasingly used in fields such as communication and remote sensing. To effectively manage and control the vast and dynamically changing satellite networks, digital twin technology has been introduced, aiming to achieve real-time mapping and simulation of the physical state and behavior of satellite networks.

[0003] However, the inherent characteristics of LEO satellite networks, such as highly dynamic topology, high latency in satellite-to-ground transmission, and high packet loss rate, often cause delays, interruptions, or noise interference in the telemetry data stream upon which digital twins rely, severely affecting the continuity and accuracy of state updates.

[0004] Therefore, there is an urgent need for a LEO satellite network digital twin state synchronization and prediction mechanism that can still achieve high fidelity and strong robustness under non-ideal telemetry data conditions, in order to support advanced applications such as real-time monitoring, intelligent operation and maintenance, and strategy optimization. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a method for synchronizing the state of a low-Earth orbit satellite network twin, which includes:

[0006] S1: Collect telemetry data from the low-Earth orbit satellite network, parse each telemetry data packet and classify it in real time, including real-time valid, delayed valid and invalid data;

[0007] S2: When telemetry packets are classified as real-time valid, generate flow table policies based on the telemetry packets; use the flow table policies to drive the virtual twin network;

[0008] S3: Construct spatiotemporal graphs for multiple historical moments and input them into a spatiotemporal graph neural network prediction model for processing to obtain the prediction network state for future moments;

[0009] S4: When telemetry packets are classified as invalid, use the predictive network state to generate flow table strategies and drive the virtual twin network;

[0010] S5: When the telemetry data packet is classified as delayed, adaptive backtracking calibration is performed based on the predicted network state to obtain the corrected network state; the corrected network state replaces the predicted network state at the corresponding time.

[0011] Preferably, constructing a spatiotemporal graph across multiple historical moments includes: acquiring the state information of the low-Earth orbit satellite network, constructing a spatiotemporal graph with satellites as nodes and links between satellites as edges, and acquiring the relevant timeframes. The spacetime diagram is represented as ,in, Indicates time The set of satellite nodes Indicates time The set of inter-satellite links, express The node feature matrix at time step.

[0012] Preferably, the process of processing the spatiotemporal graph of multiple historical moments using a spatiotemporal graph neural network prediction model includes:

[0013] Perform graph attention convolution on the spatiotemporal graph at multiple historical moments to obtain the node aggregation matrix at each historical moment;

[0014] The node aggregation matrix of historical moments is processed by causal dilation one-dimensional causal convolution to obtain the temporal feature matrix of historical moments;

[0015] Graph pooling is used to compress the temporal feature matrix of historical moments into a global temporal vector; a scaling dot product attention mechanism is used to process the global temporal vector of historical moments, with temporal attention weights; and the temporal feature matrix of historical moments is weighted and aggregated according to the temporal attention weights to obtain the enhanced temporal features of the current moment.

[0016] A fully connected layer is used to process the enhanced temporal features at the current time to obtain the predicted network state for future time steps;

[0017] The predicted link connectivity probability matrix is ​​calculated based on the enhanced temporal features at the current time. An ensemble mask matrix based on orbital budget is introduced to perform orbital mask filtering on the predicted link connectivity probability matrix to obtain the adjacency matrix at future time.

[0018] Furthermore, the graph attention convolution representation for the spatiotemporal graph at multiple historical moments is as follows:

[0019]

[0020]

[0021] in, Indicates time node In the The output features of the layer Indicates the number of heads of attention. This indicates a splicing operation. This represents the ELU activation function. Indicates the first The attention head is in the first The learnable weight matrix of the layer, Indicates time node In the The output features of the layer Represents a node The set of neighboring nodes, Indicates time No. Each attention node With nodes In the Attention coefficient of layer Represents a learnable attention vector. Indicates time node In the The output features of the layer Indicate Time Neighbor Nodes In the The output features of the layer express Activation function.

[0022] Furthermore, the global temporal vector of historical moments is processed using a scaled dot product attention mechanism and represented as follows:

[0023]

[0024] in, Represents temporal attention weights; , These represent the first and second learnable projection matrices, respectively. Indicates the current time The global time-series vector, Representing historical moments The global time-series vector, Indicates a moment within a historical time window The global time-series vector, Indicates the length of the historical time window. This represents the scaling factor.

[0025] Furthermore, the formula for calculating the predicted link connectivity probability matrix is ​​as follows:

[0026]

[0027] The trajectory mask filtering of the predicted link connectivity probability matrix is ​​expressed as:

[0028]

[0029] in, This represents the predicted link connectivity probability matrix. Represents the adjacency matrix at future time steps. This represents the ELU activation function. Indicates the enhanced temporal features at the current moment. The link prediction weight matrix is ​​a learnable parameter matrix used to map the high-dimensional features of nodes to the probability space of link existence. It represents the Hadamah accumulation. This represents the connectivity probability threshold. Represents a set mask matrix based on orbital budget. Indicates an indicator function.

[0030] Preferably, the adaptive backtracking calibration process includes:

[0031] The delay measurement vector is parsed from the telemetry data packet, and the dynamic prediction covariance is calculated based on the delay measurement vector and the predicted network state.

[0032] Calculate the measurement noise covariance based on the sensor's real-time health status and channel conditions;

[0033] The adaptive Kalman gain is obtained by performing standard Kalman filtering updates based on the dynamic prediction covariance and the measurement noise covariance.

[0034] The delay measurement vector is corrected based on the adaptive Kalman gain and the predicted network state to obtain the corrected network state.

[0035] Furthermore, the formula for correcting the delay measurement vector is as follows:

[0036]

[0037] in, Indicates time node The corrected predicted state Indicates time Adaptive Kalman gain, Indicates time node The delay measurement vector, Indicates time node The predicted state,

[0038] Furthermore, the innovation is defined as the prediction error obtained by subtracting the delayed measurement vector from the prediction network state. The system maintains the innovation sequence. When it is detected that the mean of the innovation sequence is continuously non-zero or the dynamic prediction covariance deviates significantly from the expectation, the parameters in the dynamic prediction covariance calculation formula and the measurement noise covariance calculation formula are adjusted, and the prediction error and the input spatiotemporal graph corresponding to the prediction network state are used as training samples to fine-tune the spatiotemporal graph neural network prediction model online.

[0039] The beneficial effects of this invention are as follows:

[0040] This invention fully considers the high dynamism, discontinuity, and uncertainty of telemetry data from low-Earth orbit (LEO) satellite networks. By constructing a complete technical closed loop of "real-time perception - intelligent prediction - retrospective calibration - online evolution," it achieves high-fidelity state synchronization of digital twins under real communication constraints, providing reliable state support for intelligent operation and maintenance, dynamic routing, and fault self-healing of LEO constellations. Specifically, the spatiotemporal graph neural network constructed in this invention directly uses satellite node states and inter-satellite link topology as input, and its graph structure dynamically adapts to the time-varying topology of the constellation. By setting delay and noise thresholds, the system can accurately identify and distinguish three typical anomalies: "data delay," "data interruption," and "noisy data," and trigger corresponding prediction and calibration strategies for different anomaly types, realizing a closed-loop response from anomaly detection to state compensation. By integrating the Kalman filtering mechanism based on physical models with graph neural networks based on historical data, while ensuring the accuracy of real-time state mapping, a digital twin time-series consistency system with historical retrospective and state traceability is established. By using the backtracking calibration results as the truth label, the system can continuously collect training samples and perform lightweight incremental updates to the prediction model during operation, enabling the twin system to autonomously evolve in accordance with the actual operating status and service mode changes of the satellite network. This invention achieves high-fidelity state synchronization of the twin and autonomous evolution of the model under real communication constraints. Attached Figure Description

[0041] Figure 1 This is a flowchart of the low-Earth orbit satellite network twin simulation state synchronization method of the present invention;

[0042] Figure 2 This is a framework diagram of the low-Earth orbit satellite network twin simulation state synchronization method in this invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] This invention proposes a method for synchronizing the state of a twin emulator in a low-Earth orbit satellite network, such as... Figure 1 , Figure 2 As shown, the method includes the following:

[0045] S1: Collect telemetry data from the low-Earth orbit satellite network, parse each telemetry data packet and classify it in real time, including real-time valid, delayed valid and invalid data.

[0046] Collect telemetry data from low-Earth orbit satellite networks, and parse each telemetry data packet into a triplet. ,in Indicates the original data collection timestamp (UTC). Represents the original measurement vector. This indicates the timestamp when the data packet arrived in the twin.

[0047] Each telemetry data packet is parsed based on preset thresholds (including the maximum tolerable delay threshold). (and the valid range of physical parameters), classify and label data packets. Real-time valid data (REAL_TIME) simultaneously meets the timeliness requirement. ) and validity (data) (Through physical scope verification and consistency checks). Delayed valid data satisfies the requirements of validity and timeliness; this type of data is crucial for correcting historical states and training models. Invalid data refers to data... The validity check failed.

[0048] S2: When telemetry data packets are classified as real-time valid, generate flow table policies based on the telemetry data packets; use the flow table policies to drive the virtual twin network.

[0049] like This indicates that the data is of high quality and timely; at this point, the telemetry data can be directly used to drive control strategies, generate flow table strategies and apply them to the virtual twin network, thereby driving the virtual twin network to achieve state synchronization between the virtual twin network and the low-Earth orbit satellite network.

[0050] S3: Construct spatiotemporal graphs of multiple historical moments and input them into a spatiotemporal graph neural network prediction model for processing to obtain the prediction network state for future moments.

[0051] Low-Earth orbit (LEO) satellite network telemetry data encompasses satellite orbital parameters, inter-satellite link status, service load, and network topology information. Spatiotemporal maps at multiple historical moments are constructed based on this LEO satellite network telemetry data. Specifically:

[0052] Use a sliding window to capture continuous segments. The network state at each historical moment. The state is represented as a spacetime diagram ,in express A set of satellite nodes, Indicates time The set of inter-satellite links, consisting of an adjacency matrix. The description is defined as an effective communication topology, which depends not only on the ensemble visibility between satellites (determined by orbital mechanics), but also on random blocking factors such as link congestion, equipment failure, or channel interference. This is a node feature matrix, where each row corresponds to a satellite. The values ​​taken on a dimensional feature (position, velocity, load, etc.). Continuous values ​​are considered. A sequence of spatiotemporal graphs of historical moments .

[0053] This invention employs a pre-trained spatiotemporal graph neural network prediction model (STGNN) to process continuous... A timeline of historical moments, projecting the future. The sequence of node feature matrices at each time step It predicts the network state at future time steps and outputs the adjacency matrix at future time steps.

[0054] The STGNN of this invention adopts a three-level hybrid architecture of Graph Attention Network (GAT) + Causal TCN + Temporal Attention, which sequentially models the spatial dependency of satellite networks, local temporal patterns, and global long-range dependencies. The process of STGNN processing spatiotemporal graphs at multiple historical moments includes:

[0055] Spatial feature extraction layer:

[0056] Graph attention convolution is performed on the spatiotemporal graphs at multiple historical moments. Within a single time slice, each satellite node adaptively and differentially aggregates information about its currently visible neighbors to obtain the node aggregation matrix for each historical moment. The calculation formula is expressed as follows:

[0057]

[0058]

[0059] in, Indicates time node In the The output features of the layer Indicates the number of heads of attention. This indicates a splicing operation. This represents the ELU activation function. Indicates the first The attention head is in the first The learnable weight matrix of the layer, Indicates time node In the The output features of the layer Represents a node The set of neighboring nodes, Indicates time No. Each attention node With nodes In the Attention coefficient of layer Represents a learnable attention vector. Indicates time node In the The output features of the layer express Time Neighbor Nodes In the The output features of the layer Activation function.

[0060] This design allows the model to dynamically focus on the most influential neighboring satellite nodes at any given moment, adapting to rapid topology changes in low-Earth orbit satellite networks.

[0061] Temporal feature extraction layer:

[0062] A one-dimensional causal convolution with causal dilation is used to process the node aggregation matrix at historical time points to capture its short-term local dependency patterns. The aim is to efficiently extract its short-term local dependency patterns and multi-scale temporal features, and output the temporal feature matrix of historical time points.

[0063]

[0064] in, Indicates time node The time-series feature matrix, The kernel size is [size]. It is an expansion factor that grows exponentially with the number of network layers to expand the temporal receptive field; node Input features, Let be the weight matrix of the m-th convolution sum. This is the bias vector. Causal constraints ensure... It satisfies the temporal causal relationship for real-time prediction by relying solely on the input at the current and historical moments. The introduction of dilated convolution effectively expands the temporal receptive field, enabling it to simultaneously capture second-level fluctuations and minute-level state evolution.

[0065] Temporal dependency modeling layer:

[0066] This layer is based on a temporal attention mechanism and receives a temporal feature matrix rich in local patterns output by the TCN. Each element in the sequence Indicates time The feature matrix of all network nodes is composed of the temporal feature vectors of N satellite nodes at that moment. The data is stacked. Then, a temporal self-attention mechanism is used to model global long-range dependencies, adaptively focusing on the historical moments most critical to the current prediction. Specifically:

[0067] Graph pooling is used to extract the temporal feature matrix of historical moments. Compressed into a global time vector , to obtain the sequence A scaled dot product attention mechanism is used to calculate the correlation between the global temporal vectors at historical time points and the global temporal vectors at the current time point, thus obtaining the temporal attention weights:

[0068]

[0069] in, Representing historical moments Features of the current moment Temporal attention weights for prediction results; , These represent the first and second learnable projection matrices, respectively. Indicates time The global time-series vector, derived from time... The temporal feature matrix is ​​obtained by graph pooling compression; Representing historical moments The global time-series vector (as the key vector) is used to match the query vector; Represents any point within a historical time window The global time-series vector, This indicates the length of the historical time window, i.e., the number of frames in the historical spatiotemporal graph sequence of the input model; This represents the scaling factor.

[0070] Based on temporal attention weights The feature matrix of all nodes in the network at historical moments We perform weighted aggregation to obtain the enhanced temporal features at the current time step:

[0071]

[0072] in, The projection matrix is ​​a value. This mechanism enables the model to explicitly capture long-cycle patterns such as orbital cycles and regular business peaks, compensating for the shortcomings of TCN in modeling ultra-long-range dependencies.

[0073] Value prediction header:

[0074] Direct multi-step prediction is performed using a fully connected layer. This module receives current-time augmented temporal features that incorporate historical context information. Output the future in one go through dimension mapping The predicted network state sequence at each time step.

[0075]

[0076] in, Indicates the future The predicted network state sequence at each time step. This indicates that the current time step enhances the temporal characteristics. This refers to a multilayer perceptron, which learns a nonlinear mapping between input features and future trajectories. dimensional feature space mapped to The output state space of dimension.

[0077] To integrate prior knowledge of orbital mechanics with the stochastic modeling capabilities of deep learning, this invention generates future... The adjacency matrix at each time step At that time, an ensemble mask matrix based on orbital budget was introduced. Specifically:

[0078] Calculate the predicted link connectivity probability matrix based on the enhanced temporal features at the current moment:

[0079] By introducing an ensemble mask matrix based on orbital budget to filter the predicted link connectivity probability matrix using orbital masking, the adjacency matrix at future time steps is obtained:

[0080]

[0081] in, This represents the predicted link connectivity probability matrix. Represents the adjacency matrix at future time steps. This represents the ELU activation function. Indicates the enhanced temporal features at the current moment. This represents the link prediction weight matrix, which is a learnable parameter matrix used to map the high-dimensional features of nodes to the probability space of link existence. It represents the Hadamah accumulation. This represents the connectivity probability threshold. Represents a set mask matrix based on orbital budget. The deterministic geometric visibility matrix (1 for visible, 0 for invisible) is calculated based on satellite ephemeris and used to filter physically unreachable links. Indicates an indicator function.

[0082] This invention ensures that the prediction results strictly adhere to physical constraints (i.e., do not violate the principle of line-of-sight), while accurately capturing non-line-of-sight link interruptions caused by congestion or interference through neural networks.

[0083] S4: When telemetry packets are classified as invalid, use the predictive network state generation flow table strategy and drive the virtual twin network.

[0084] When the system does not receive any data within the expected time, or receives data continuously... When data is missing, a pre-trained Spatiotemporal Graph Neural Network (STGNN) prediction model is immediately employed. Using the spatiotemporal graph constructed from the effective states and topology of historical moments as input, it predicts the current state of the entire network. This current network state is then input into the policy engine to generate and distribute predictive / compensatory flow table policies to maintain the basic operation of the virtual mapping network.

[0085] S5: When the telemetry data packet is classified as delayed, adaptive backtracking calibration is performed based on the predicted network state to obtain the corrected network state; the corrected network state replaces the predicted network state at the corresponding time.

[0086] When telemetry packets marked DELAYED are available at a later time... Upon arrival, the system initiates an adaptive backtracking calibration process to recalibrate historical data. The predicted state is corrected to ensure the eventual consistency of the digital twin's historical state sequence. Specifically:

[0087] The delay measurement vector is parsed from the telemetry data packet. Based on the delay measurement vector and the predicted network state, the dynamic prediction covariance is calculated. Specifically:

[0088] For each satellite node in the low-Earth orbit satellite network At any historical moment Define the predicted state vector This indicates that the STGNN model at time [time value missing] Output node state prediction, This represents the dimensional state, generated by the prediction model when real-time telemetry data is missing or delayed. The prediction state error covariance matrix is... This matrix is ​​a symmetric positive definite matrix. Its diagonal elements represent the prediction error variance of each state component, and its off-diagonal elements represent the error correlation between state components. During system initialization, the STGNN model is evaluated using an independent validation dataset in the offline phase, and the prediction error samples for each state dimension are calculated. and initialize diagonal matrix .

[0089] Delay measurement vector is obtained by parsing telemetry data packets. , indicating at a certain moment Arrived or tagged as The state measurement values ​​parsed from the original telemetry data packet, which carries the original acquisition timestamp. .

[0090] The system maintains a prediction error sliding window for each node. ,in For a historic moment The prediction error (requires that delayed measurement data has arrived and been corrected at that moment). The dynamic prediction covariance is calculated as follows:

[0091]

[0092] in The time decay weight is calculated as follows:

[0093]

[0094] Among them, parameters The rate of memory decay is typically controlled by setting the window length. of This design allows recent prediction errors to have a greater impact on the current covariance estimate, thus reflecting changes in model prediction performance more promptly.

[0095] The measurement noise covariance is calculated based on the sensor's real-time health status and channel conditions. Specifically:

[0096] Measurement noise covariance matrix This indicates that the sensor is at time [time]. The measurement uncertainty is initially set as the covariance matrix corresponding to the sensor's nominal accuracy. Measure the noise covariance Adaptive and dynamic adjustment based on real-time sensor health status and channel conditions:

[0097]

[0098] in, for The sensor health score at any given time is obtained based on the satellite's self-inspection report, historical fault statistics, and redundant sensor consistency calculations. For a moment The channel quality index is obtained by comprehensively mapping the signal-to-noise ratio (SNR), bit error rate (BER), and jitter of the satellite-to-ground link. The mapping function is: . , For the weighting coefficients, satisfying Default value , This reflects an emphasis on both the intrinsic performance of the sensor and the transmission channel. When the sensor's health deteriorates or the channel quality worsens, As the fluctuation increases, the system reduces the trust weight of that measurement value.

[0099] The adaptive Kalman gain is obtained by performing standard Kalman filtering updates based on the dynamic prediction covariance and the measurement noise covariance. Specifically:

[0100] In obtaining adaptive and Then, the system executes the standard Kalman filter update steps, updating the historical time... The system performs optimal adjustments to the state. It retrieves the corresponding time from the twin's historical state database. Predicted state and its covariance And calculate the adaptive Kalman gain. :

[0101]

[0102] Wherein, the gain matrix Essentially an optimal weight allocator, its value is determined by the prediction uncertainty. With measurement uncertainty The relative size is dynamically determined, accurately representing the quantitative trade-off between "trusting predictions more or trusting delayed measurements more at the current moment".

[0103] The delay measurement vector is corrected based on the adaptive Kalman gain and the predicted network state to obtain the corrected network state, calculated as follows:

[0104]

[0105]

[0106] in, Indicates time node The corrected predicted state This represents the corrected estimated covariance. This covariance matrix will be stored in the database along with the corrected state to characterize the confidence level of the historical data point and provide a basis for subsequent data quality assessment.

[0107] The system will use Overwrite the predicted values ​​originally stored in the historical database This serves as the final authoritative state at that moment, ensuring the consistency of the digital twin's historical timeline. This replacement operation is only used to maintain the consistency of the digital twin's historical timeline, serving as a "truth value" to support subsequent online model fine-tuning and backtracking analysis, and does not re-drive the virtual scene of a past moment.

[0108] In some preferred embodiments of the present invention, innovation is defined as the prediction error obtained by subtracting the delay measurement vector from the prediction network state. .

[0109] System maintenance of new information sequence The statistical properties (mean, covariance, autocorrelation function) of the innovation sequence are considered. If the mean of the innovation sequence is consistently non-zero or the covariance deviates significantly from the expected value, it indicates a systematic bias in the prediction model or an uncalibrated error in the sensor. The system will then trigger an action. and Adaptive fine-tuning of parameters in the estimation formula (specifically including adjusting the measurement noise tradeoff coefficient) , and prediction error decay factor The specific adaptive fine-tuning rules are as follows:

[0110] (1) Prediction error attenuation factor Adaptive adjustment: When the covariance trace of the innovation sequence is detected greater than the set covariance deviation threshold This indicates a significant increase in the model's recent prediction uncertainty. The system reduces this uncertainty through the following formula: To accelerate the forgetting of historical errors and increase the weight of recent measurement data:

[0111]

[0112] in, This is the lower bound of the attenuation factor. To adjust the step size coefficient.

[0113] (2) Measure the noise tradeoff factor and Adaptive adjustment: and These correspond to penalty weights for sensor health status and channel conditions, respectively. When the mean of the innovation sequence... When the value remains non-zero, the system calculates the sensor health score for a specific time period (e.g., within a sliding window). variance With channel quality index variance Dynamically redistribute weights based on volatility:

[0114]

[0115]

[0116] If the sensor's state fluctuates more drastically, that is Then adaptively increase This enables the system to calculate the measurement noise covariance. Higher penalties are applied to sensor noise, and vice versa.

[0117] Retrieve the historical spatiotemporal graph sequence context of the input STGNN that triggered this prediction from the cache. .in This represents the input feature vector corresponding to the prediction time, containing continuous... The node feature matrix and adjacency matrix at each historical moment. (The tuples are then used to...) The samples are stored in the incremental learning sample pool for subsequent online fine-tuning of the STGNN model.

[0118] In summary, this invention uses a twin emulator to continuously receive real-time telemetry data streams from the physical network and performs timestamp alignment and validity verification on the data. When data delays, interruptions, or excessive noise are detected, the system employs a pre-trained spatiotemporal graph neural network (STGNN) to generate a predicted value of the entire network node state at the current missing moment, using the state sequence and dynamic topology of each node in the network at the previous moment as input. This predicted value is then used as the instantaneous state input of the twin emulator to maintain the continuity and real-time performance of the simulation. When the delayed actual measurement data arrives at a later moment, a backtracking calibration process is initiated. This process fuses the delayed data with the neural network prediction results from the corresponding historical moment to achieve backtracking correction of the historical state. The system calculates the error between the predicted state and the corrected state, and uses this error, along with the historical network context that triggered the prediction, to form a training sample, which is then provided to the STGNN for online, lightweight incremental learning. By constructing a complete technical closed loop of "real-time perception - intelligent prediction - backtracking calibration - online evolution," this invention achieves high-fidelity state synchronization of the twin emulator under real communication constraints, providing reliable state support for intelligent operation and maintenance and application feasibility analysis of low-Earth orbit constellations.

[0119] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for synchronizing the state of a twin emulator in a low-Earth orbit satellite network, characterized in that, include: S1: Collect telemetry data from the low-Earth orbit satellite network, parse each telemetry data packet and classify it in real time, including real-time valid, delayed valid and invalid data; S2: When the telemetry data packet is classified as real-time valid, a flow table strategy is generated based on the telemetry data packet; Virtual twin networks are driven using flow table strategies; S3: Construct spatiotemporal graphs for multiple historical moments and input them into a spatiotemporal graph neural network prediction model for processing to obtain the prediction network state for future moments; S4: When telemetry packets are classified as invalid, use the predictive network state to generate flow table strategies and drive the virtual twin network; S5: When the telemetry data packet is classified as delayed, adaptive backtracking calibration is performed based on the predicted network state to obtain the corrected network state; the corrected network state replaces the predicted network state at the corresponding time.

2. The method for synchronizing the state of a low-Earth orbit satellite network twin as described in claim 1, characterized in that, Constructing a spatiotemporal graph across multiple historical moments involves: acquiring the state information of the low-Earth orbit satellite network, treating satellites as nodes and inter-satellite links as edges to construct the spatiotemporal graph, and acquiring the relevant data. The spacetime diagram is represented as ,in, Indicates time The set of satellite nodes Indicates time The set of inter-satellite links, express The node feature matrix at time step.

3. The method for synchronizing the state of a low-Earth orbit satellite network twin simulation as described in claim 1, characterized in that, The process of using a spatiotemporal graph neural network prediction model to process spatiotemporal graphs at multiple historical moments includes: Perform graph attention convolution on the spatiotemporal graph at multiple historical moments to obtain the node aggregation matrix at each historical moment; The node aggregation matrix of historical moments is processed by causal dilation one-dimensional causal convolution to obtain the temporal feature matrix of historical moments; Graph pooling is used to compress the temporal feature matrix of historical moments into a global temporal vector; a scaling dot product attention mechanism is used to process the global temporal vector of historical moments, with temporal attention weights; and the temporal feature matrix of historical moments is weighted and aggregated according to the temporal attention weights to obtain the enhanced temporal features of the current moment. A fully connected layer is used to process the enhanced temporal features at the current time to obtain the predicted network state for future time steps; The predicted link connectivity probability matrix is ​​calculated based on the enhanced temporal features at the current time. An ensemble mask matrix based on orbital budget is introduced to perform orbital mask filtering on the predicted link connectivity probability matrix to obtain the adjacency matrix at future time.

4. The method for synchronizing the state of a low-Earth orbit satellite network twin simulation as described in claim 3, characterized in that, The graph attention convolution representation for performing a spatiotemporal graph over multiple historical moments is as follows: ； ； in, Indicates time node In the The output features of the layer Indicates the number of heads of attention. This indicates a splicing operation. This represents the ELU activation function. Indicates the first The attention head is in the first The learnable weight matrix of the layer, Indicates time node In the The output features of the layer Represents a node The set of neighboring nodes, Indicates time No. Each attention node With nodes In the Attention coefficient of layer Represents a learnable attention vector. Indicates time node In the The output features of the layer Indicate Time Neighbor Nodes In the The output features of the layer express Activation function.

5. The method for synchronizing the state of a low-Earth orbit satellite network twin as described in claim 3, characterized in that, The global time-series vector of historical moments is processed using a scaled dot product attention mechanism and represented as follows: ； in, Represents temporal attention weights; , These represent the first and second learnable projection matrices, respectively. Indicates the current time The global time-series vector, Representing historical moments The global time-series vector, Indicates a moment within a historical time window The global time-series vector, Indicates the length of the historical time window. This represents the scaling factor.

6. The method for synchronizing the state of a low-Earth orbit satellite network twin as described in claim 3, characterized in that, The formula for calculating the predicted link connectivity probability matrix is: ； The trajectory mask filtering of the predicted link connectivity probability matrix is ​​expressed as: ； in, This represents the predicted link connectivity probability matrix. Represents the adjacency matrix at future time steps. This represents the ELU activation function. Indicates the enhanced temporal features at the current moment. The link prediction weight matrix is ​​a learnable parameter matrix used to map the high-dimensional features of nodes to the probability space of link existence. It represents the Hadamah accumulation. This represents the connectivity probability threshold. Represents a set mask matrix based on orbital budget. This indicates an indicator function.

7. The method for synchronizing the state of a low-Earth orbit satellite network twin simulation as described in claim 1, characterized in that, The process of adaptive backtracking calibration includes: The delay measurement vector is parsed from the telemetry data packet, and the dynamic prediction covariance is calculated based on the delay measurement vector and the predicted network state. Calculate the measurement noise covariance based on the sensor's real-time health status and channel conditions; The adaptive Kalman gain is obtained by performing standard Kalman filtering updates based on the dynamic prediction covariance and the measurement noise covariance. The delay measurement vector is corrected based on the adaptive Kalman gain and the predicted network state to obtain the corrected network state.

8. The method for synchronizing the state of a low-Earth orbit satellite network twin simulation according to claim 7, characterized in that, The formula for correcting the delay measurement vector is: ； in, Indicates time node The corrected predicted state, Indicates time Adaptive Kalman gain, Indicates time node The delay measurement vector, Indicates time node The predicted state.

9. The method for synchronizing the state of a low-Earth orbit satellite network twin simulation according to claim 7, characterized in that, Also includes: The innovation is defined as the prediction error obtained by subtracting the delayed measurement vector from the prediction network state. The system maintains the innovation sequence. When the mean of the innovation sequence is detected to be continuously non-zero or the dynamic prediction covariance deviates significantly from the expectation, the parameters in the dynamic prediction covariance calculation formula and the measurement noise covariance calculation formula are adjusted, and the prediction error and the input spatiotemporal graph corresponding to the prediction network state are used as training samples to fine-tune the spatiotemporal graph neural network prediction model online.

Images