Satellite attitude and orbit control hierarchical fusion digital twin modeling method and system

Abstract

This invention discloses a hierarchical fusion digital twin modeling method and system for satellite attitude and orbit control, relating to the field of aerospace control. The method includes: constructing a three-level hierarchical decoupled satellite attitude and orbit control fusion twin basic model; adaptively and dynamically weighting the three-level outputs of the basic model based on the confidence evaluation index of current on-orbit telemetry data to obtain fusion simulation output values; utilizing the residual between the fusion simulation output values ​​and the actual satellite telemetry values, performing closed-loop on-orbit autonomous iterative evolution to obtain a satellite attitude and orbit control fusion twin update model; and deploying the satellite attitude and orbit control fusion twin update model in a lightweight, hierarchical manner at both the satellite and ground levels. This application addresses the technical problem of existing attitude and orbit control system modeling methods where model prediction accuracy gradually degrades with increasing on-orbit time and is difficult to maintain high-precision state reproduction capability throughout the entire lifecycle.

Description

Technical Field

[0001] This application relates to the field of aerospace control, and in particular to a hierarchical fusion digital twin modeling method and system for satellite attitude and orbit control. Background Technology

[0002] The satellite attitude and orbit control system is a core subsystem for spacecraft to complete on-orbit missions, and its control accuracy and stability directly determine the success or failure of the mission. Current mainstream modeling methods for attitude and orbit control systems primarily employ physical modeling based on dynamic equations. This involves establishing satellite orbital motion equations and attitude dynamic equations, combining ground-calibrated actuator parameters and environmental perturbation models, and using numerical integration methods to simulate and predict the satellite's on-orbit state. However, due to uncertainties in the space environment, including atmospheric density fluctuations, changes in solar radiation pressure, non-spherical gravitational perturbations from the Earth, and on-orbit performance degradation of actuators, purely physical models cannot fully and accurately describe the actual dynamic behavior of the satellite on orbit. This leads to deviations between simulation outputs and the actual state, and these deviations gradually increase over time. Simply relying on ground-calibrated parameters and fixed correction coefficients cannot effectively suppress the continuous growth of these deviations.

[0003] At present, the modeling methods for attitude and orbit control systems have technical problems such as the gradual degradation of model prediction accuracy as the on-orbit time increases, and the difficulty in maintaining high-precision state reproduction capability throughout the entire life cycle. Summary of the Invention

[0004] This application provides a hierarchical fusion digital twin modeling method and system for satellite attitude and orbit control. By constructing a three-level hierarchical decoupled fusion twin basic model, adaptively weighting the three-level outputs based on the confidence index of on-orbit telemetry data to obtain fusion simulation values, and using the residual between the fusion simulation values ​​and actual telemetry values ​​to perform closed-loop autonomous iterative evolution to update the model, and using techniques such as lightweight hierarchical deployment at both satellite and ground levels, this application solves the technical problems of existing attitude and orbit control system modeling methods where the model prediction accuracy gradually degrades with the extension of on-orbit time and it is difficult to maintain high-precision state reproduction capability throughout the entire life cycle. This achieves the technical effect of enabling the satellite attitude and orbit control digital twin model to maintain high-precision state reproduction capability throughout the entire life cycle.

[0005] This application provides a hierarchical fusion digital twin modeling method for satellite attitude and orbit control, comprising: constructing a three-level hierarchical decoupled satellite attitude and orbit control fusion twin basic model; based on the confidence evaluation index of the current on-orbit telemetry data, adaptively and dynamically assigning weights to the three-level outputs of the satellite attitude and orbit control fusion twin basic model to obtain fusion simulation output values; using the residual between the fusion simulation output values ​​and the actual satellite telemetry values, performing closed-loop on-orbit autonomous iterative evolution to obtain a satellite attitude and orbit control fusion twin update model; and deploying the satellite attitude and orbit control fusion twin update model in a lightweight, hierarchical manner at both the satellite and ground levels.

[0006] In a possible implementation, a three-level, layered, decoupled satellite attitude and orbit control fusion twin basic model is constructed, and the following processing is performed: a low-frequency error fitting layer is built, which adopts a temporal neural network to output low-frequency compensation; a high-frequency residual compensation layer is built, which adopts a lightweight high-frequency compensation network to output high-frequency compensation; a basic mechanism layer is built, which embeds the satellite attitude and orbit control physical equations to solve the orbital mechanism, and reserves an attitude coupling term correction interface to receive the low-frequency compensation and the high-frequency compensation; based on the basic mechanism layer, the low-frequency error fitting layer, and the high-frequency residual compensation layer, the satellite attitude and orbit control fusion twin basic model is constructed.

[0007] In possible implementations, the following processing is performed: the temporal neural network is an LSTM network or a GRU network; the lightweight high-frequency compensation network is a one-dimensional CNN network or a lightweight Transformer micro-network.

[0008] In a possible implementation, the residual between the fused simulation output value and the actual satellite telemetry value is used to perform closed-loop on-orbit autonomous iterative evolution, and the following processing is performed: based on the fused simulation output value and the actual satellite telemetry value, residual data and residual data change rate are calculated; based on the residual data and the residual data change rate, the low-frequency error fitting layer and the high-frequency residual compensation layer are subjected to closed-loop on-orbit autonomous iterative evolution using the proportional-derivative adjustment formula.

[0009] In a possible implementation, the following processing is performed: the proportional-derivative adjustment formula is: ; where σ i denoted as , where is the adjustment amount of the hierarchical parameters of the i-th layer data-driven network; e represents the residual data. k represents the rate of change of the residual data. p This is the proportional adjustment coefficient, with a value range of 0.01 to 0.1, used to control the basic correction strength based on the current residual; k d This is the differential adjustment coefficient, with a value range of 0.1 to 1.0. It is used to apply damping correction based on the residual change trend to suppress network iterative oscillations.

[0010] In a possible implementation, based on the confidence index of the current on-orbit telemetry data, adaptive dynamic weight allocation is performed on the three-layer output of the satellite attitude and orbit control fusion twin basic model, and the following processing is performed: when the confidence index is the number of effective telemetry sampling points, the number of current effective telemetry sampling points is counted in real time; when the number of current effective telemetry sampling points is greater than a preset high threshold, the data driving weight of the low-frequency error fitting layer and the high-frequency residual compensation layer is increased; when the number of current effective telemetry sampling points is less than a preset low threshold, the fusion weight of the basic mechanism layer is increased; or, when the confidence index is the residual data, the absolute value or vector magnitude of the residual data is calculated to obtain the residual amplitude; when the residual amplitude is greater than a preset amplitude threshold, the data driving weight of the low-frequency error fitting layer and the high-frequency residual compensation layer is increased; when the residual amplitude is less than or equal to the preset amplitude threshold, the fusion weight of the basic mechanism layer is increased.

[0011] In a possible implementation, the satellite attitude and orbit control fusion twin update model is deployed in a lightweight, two-tiered system on both the satellite and ground, and the following processing is performed: On the satellite, the basic mechanism layer of the satellite attitude and orbit control fusion twin update model is fully preserved, and the low-frequency error fitting layer and the high-frequency residual compensation layer are compressed to generate a lightweight on-board model, which is then mounted on the satellite embedded processor; On the ground, the satellite attitude and orbit control fusion twin update model is fully deployed for global simulation and parameter optimization, and the optimized network parameters are periodically distributed to the lightweight on-board model for version updates.

[0012] This application also provides a hierarchical fusion digital twin modeling system for satellite attitude and orbit control, comprising: a hierarchical decoupled twin basic model construction module for constructing a three-level hierarchical decoupled satellite attitude and orbit control fusion twin basic model; an adaptive weighted fusion module for adaptively and dynamically assigning weights to the three-level outputs of the satellite attitude and orbit control fusion twin basic model based on the confidence evaluation index of the current on-orbit telemetry data to obtain fusion simulation output values; a closed-loop autonomous evolution update module for using the residual between the fusion simulation output values ​​and the actual satellite telemetry values ​​to perform closed-loop on-orbit autonomous iterative evolution to obtain a satellite attitude and orbit control fusion twin update model; and a satellite-ground two-level lightweight hierarchical deployment module for deploying the satellite attitude and orbit control fusion twin update model in a satellite-ground two-level lightweight hierarchical manner.

[0013] This application proposes a hierarchical fusion digital twin modeling method and system for satellite attitude and orbit control. First, a three-level hierarchical decoupled satellite attitude and orbit control fusion twin basic model is constructed. Then, based on the confidence index of current on-orbit telemetry data, adaptive dynamic weight allocation is performed on the three-level outputs of the basic model to obtain fused simulation output values. Next, using the residual between the fused simulation output values ​​and the actual satellite telemetry values, closed-loop on-orbit autonomous iterative evolution is performed to obtain an updated satellite attitude and orbit control fusion twin model. Finally, the updated model is deployed in a lightweight, hierarchical manner at both the satellite and ground levels. Through this process, the proposed method and system achieve the technical effect of enabling the satellite attitude and orbit control digital twin model to maintain high-precision state reproduction capabilities throughout its entire lifecycle. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0015] Figure 1 This is a flowchart illustrating a hierarchical fusion digital twin modeling method for satellite attitude and orbit control provided in an embodiment of this application.

[0016] Figure 2 This is a schematic diagram of the structure of a hierarchical fusion digital twin modeling system for satellite attitude and orbit control provided in an embodiment of this application.

[0017] Figure labeling: Layered decoupled twin basic model construction module 10, adaptive weighted fusion module 20, closed-loop autonomous evolution update module 30, satellite-ground two-level lightweight hierarchical deployment module 40. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] This application provides a hierarchical fusion digital twin modeling method for satellite attitude and orbit control, such as... Figure 1 As shown, the method includes: Step S100: Construct a three-level hierarchical decoupled satellite attitude and orbit control fusion twin basic model.

[0020] Specifically, the satellite attitude and orbit control fusion twin model is divided into three independent modules: a basic mechanism layer, a low-frequency error fitting layer, and a high-frequency residual compensation layer. The basic mechanism layer embeds the satellite attitude and orbit control physical equations to provide basic predicted values ​​for attitude and orbit states. The orbit portion is solved throughout the entire process to avoid over-correction of the data model, which could lead to degradation of orbit determination accuracy. An interface for correction is reserved in the attitude coupling terms for the data-driven layer to access. The low-frequency error fitting layer uses a temporal neural network to fit low-frequency, slowly varying errors such as periodic atmospheric density disturbances and long-term flywheel wear. The high-frequency residual compensation layer uses a lightweight network to fit high-frequency random errors such as gyroscope measurement noise and instantaneous flywheel friction. This step, by decoupling errors by frequency layer, provides a model basis for graded compensation of different error types.

[0021] In one possible implementation, a three-level hierarchical decoupled satellite attitude and orbit control fusion twin basic model is constructed. Step S100 further includes step S110, building a low-frequency error fitting layer. The low-frequency error fitting layer adopts a temporal neural network and outputs low-frequency compensation. The temporal neural network is an LSTM network or a GRU network. Specifically, the low-frequency error fitting layer is constructed using a temporal neural network. The temporal neural network uses a long short-term memory network or a gated recurrent unit network. When a long short-term memory network is used, the network structure includes one input layer, two long short-term memory hidden layers, and one fully connected output layer, with 64 memory units in each hidden layer. When a gated recurrent unit network is used, the network structure includes one input layer, two gated recurrent unit hidden layers, and one fully connected output layer, with 64 hidden units in each hidden layer. The input data of the low-frequency error fitting layer are daily telemetry and control low-frequency orbit data and environmental monitoring data. The daily telemetry and control low-frequency orbit data includes the orbit semi-major axis, orbit inclination, eccentricity, right ascension of the ascending node, argument of perigee, and mean perigee angle transmitted by the satellite daily. Environmental monitoring data, including the solar radiation flux index F10.7 and the geomagnetic activity index Kp, are periodically injected into the satellite by ground control stations. The output of the low-frequency error fitting layer is a low-frequency compensation value, which is used to correct low-frequency slowly varying errors. These errors include periodic atmospheric density disturbances, low-frequency solar radiation pressure interference, low-frequency deflection of the solar panels, and torque deviations caused by long-term slow wear of the flywheel. The training data for the low-frequency error fitting layer comes from multi-condition historical data generated by the ground simulation system, including orbital and environmental data under different seasons, different solar activity levels, and different satellite attitude modes.

[0022] Step S120: Construct a high-frequency residual compensation layer. This layer employs a lightweight high-frequency compensation network to output the high-frequency compensation amount. The lightweight high-frequency compensation network is either a one-dimensional CNN network or a lightweight Transformer micro-network. Specifically, the high-frequency residual compensation layer is constructed using a lightweight high-frequency compensation network. The lightweight high-frequency compensation network is either a one-dimensional convolutional neural network or a lightweight Transformer micro-network. When a one-dimensional convolutional neural network is used, the network structure includes one input layer, two convolutional layers, two pooling layers, one flattening layer, and one fully connected output layer. The first convolutional layer has 32 kernels, a kernel size of 3, a stride of 1, and uses a linear rectified function as the activation function. The second convolutional layer has 64 kernels, a kernel size of 3, a stride of 1, and uses a linear rectified function as the activation function. Both the first and second pooling layers use max pooling, with a pooling window size of 2 and a stride of 2. The number of neurons in the fully connected output layer is equal to the dimension of the output vector. When a lightweight Transformer micronetwork is used, the network structure includes an input layer, a position encoding layer, two Transformer encoder layers, and a fully connected output layer. Each Transformer encoder layer contains a multi-head self-attention sublayer and a feedforward network sublayer. The multi-head self-attention sublayer has 4 heads, and the feedforward network has 128 hidden layer dimensions. The input data for the high-frequency residual compensation layer is high-frequency, second-level telemetry data from the star sensor and gyroscope, including the three-axis attitude angles and three-axis attitude angular velocities output by the star sensor and the three-axis angular rates output by the gyroscope. The output of the high-frequency residual compensation layer is a high-frequency compensation value, which is used to correct high-frequency random errors, including gyroscope high-frequency measurement noise, instantaneous disturbances from flywheel dead zone friction, and transient spatial particle interference. The training data for the high-frequency residual compensation layer comes from high-frequency sampling data generated by the ground simulation system, containing star sensor and gyroscope data under different noise levels and disturbance intensities.

[0023] Step S130: Construct a basic mechanism layer. This layer embeds the satellite attitude and orbit control physical equations for solving the orbital mechanism and reserves an interface for attitude coupling term correction to receive the low-frequency and high-frequency compensation quantities. Specifically, the basic mechanism layer embeds the satellite attitude and orbit control physical equations, including the orbital two-body dynamics equations, Earth non-spherical perturbation terms, atmospheric perturbation models, satellite rigid body attitude dynamics equations, thruster standard physical models, and flywheel standard physical models. The Earth non-spherical perturbation terms include J2 and J3 terms, etc. The orbital part of the basic mechanism layer relies entirely on the above physical equations for mechanism solving, i.e., using the orbital two-body dynamics equations to calculate the basic orbital motion of the satellite in the central gravitational field, superimposing Earth non-spherical perturbations and atmospheric perturbations as orbital correction quantities, and outputting the orbital position vector and velocity vector. The attitude part of the basic mechanism layer uses the satellite rigid body attitude dynamics equations to calculate the attitude reference value and reserves correction interfaces at the calculation nodes corresponding to coupling terms such as solar panel deflection and wheel control. All parameters of the basic mechanism layer are fixed based on the satellite's structural and propulsion system parameters at the time of manufacture, including satellite mass, moment of inertia matrix, solar panel area, thruster installation matrix, rated thrust of the thrusters, and rated angular momentum of the flywheel. These parameters are not involved in any network training or parameter iteration updates after the satellite enters orbit. The basic mechanism layer receives the low-frequency compensation output from step S110 and the high-frequency compensation output from step S120 through a reserved attitude coupling term correction interface. It then superimposes the low-frequency and high-frequency compensations onto the calculation results of the attitude coupling term to correct the attitude reference value and output the basic state prediction value.

[0024] Step S140: Based on the basic mechanism layer, the low-frequency error fitting layer, and the high-frequency residual compensation layer, construct the satellite attitude and orbit control fusion twin basic model. Specifically, the satellite attitude and orbit control fusion twin basic model is constructed as follows: the basic mechanism layer, the low-frequency error fitting layer, and the high-frequency residual compensation layer are combined in parallel, that is, the inputs of the three layers each independently receive their corresponding data sources, and the outputs of the three layers are all transmitted to the adaptive dynamic weight allocation module, which completes the subsequent weight allocation and fusion output. The low-frequency error fitting layer and the high-frequency residual compensation layer are equipped with enable switches. When telemetry data is missing or the confidence level is too low, the enable switch can shut down the output of the corresponding data-driven layer, at which point only the basic mechanism layer outputs the basic state prediction value.

[0025] Step S200: Based on the confidence evaluation index of the current on-orbit telemetry data, adaptive dynamic weight allocation is performed on the three-layer output of the satellite attitude and orbit control fusion twin basic model to obtain the fusion simulation output value.

[0026] Specifically, the confidence index of the current on-orbit telemetry data is acquired in real time, and the weight coefficients of the three outputs—the basic mechanism layer, the low-frequency error fitting layer, and the high-frequency residual compensation layer—are dynamically adjusted in the final fusion result based on this index. When telemetry data is sufficient or the residuals are large, the weight of the data-driven layer is increased to strengthen the data compensation effect; when telemetry data is missing or the residuals are small, the weight of the basic mechanism layer is increased to suppress the ineffective fitting of the neural network under low-confidence data. This step, through adaptive weight allocation, ensures that the fusion model maintains stable simulation accuracy under different data conditions, avoiding the failure of a single fixed weight under extreme conditions.

[0027] In one possible implementation, based on the confidence index of the current on-orbit telemetry data, adaptive dynamic weight allocation is performed on the three-layer output of the satellite attitude and orbit control fusion twin basic model. Step S200 further includes step S210: when the confidence index is the number of effective telemetry sampling points, the number of current effective telemetry sampling points is counted in real time; when the number of current effective telemetry sampling points is greater than a preset high threshold, the data-driven weights of the low-frequency error fitting layer and the high-frequency residual compensation layer are increased; when the number of current effective telemetry sampling points is less than a preset low threshold, the fusion weight of the basic mechanism layer is increased. Specifically, the number of current effective telemetry sampling points is counted in real time. An effective telemetry sampling point is defined as a telemetry data frame that simultaneously contains low-frequency orbit data, environmental monitoring data, star-sensor data, and gyroscope data at the sampling time, and all of the above data are within the effective measurement range of their respective sensors. When the number of current effective telemetry sampling points is greater than the preset high threshold, it is determined that the telemetry data is sufficient. The preset high threshold is determined based on the satellite type. For example, for large low-Earth orbit satellites, the preset high threshold can be set to 2000 sets per day; for microsatellites, it can be set to 500 sets per day. At this time, the data-driven weights of the low-frequency error fitting layer and the high-frequency residual compensation layer are increased. Specifically, the weight of the low-frequency error fitting layer can be increased from the initial value of 0.30 to 0.40, the weight of the high-frequency residual compensation layer can be increased from the initial value of 0.20 to 0.30, and the weight of the basic mechanism layer can be decreased from the initial value of 0.50 to 0.30. The sum of the weights of the three layers is normalized to 1. When the number of currently valid telemetry sampling points is less than the preset low threshold, telemetry data is considered missing. The preset low threshold is set to 100 sets per day. At this point, the fusion weights of the basic mechanism layer are increased. Specifically, the weight of the basic mechanism layer can be increased from the initial value of 0.50 to 0.80, the weight of the low-frequency error fitting layer can be decreased from 0.30 to 0.12, and the weight of the high-frequency residual compensation layer can be decreased from 0.20 to 0.08. The sum of the weights of the three layers is normalized to 1. When the number of current effective telemetry sampling points is between the preset low threshold and the preset high threshold, the weights of the three layers are kept at their initial values, namely 0.50 for the basic mechanism layer, 0.30 for the low-frequency error fitting layer, and 0.20 for the high-frequency residual compensation layer. The output values ​​of the three layers are multiplied by their corresponding weights and then summed to obtain the fused simulation output value.

[0028] Alternatively, in step S220, when the confidence evaluation index is the residual data, the absolute value or vector magnitude of the residual data is calculated to obtain the residual amplitude. When the residual amplitude is greater than a preset amplitude threshold, the data-driven weights of the low-frequency error fitting layer and the high-frequency residual compensation layer are increased; when the residual amplitude is less than or equal to the preset amplitude threshold, the fusion weights of the basic mechanism layer are increased. Specifically, the absolute value of the residual data is calculated. When the satellite attitude and orbit state is a multi-dimensional vector, the Euclidean magnitude of the vector is calculated to obtain the residual amplitude. When the residual amplitude is greater than the preset amplitude threshold, it is determined that the current model simulation output deviates significantly from the measured value. The preset amplitude threshold is preset according to the satellite control accuracy requirements. For example, for a satellite with an attitude control accuracy requirement better than 0.01°, the preset amplitude threshold can be set to 0.005°; for a satellite with an orbit control accuracy requirement better than 10m, the preset amplitude threshold can be set to 5m. To enhance the compensation effect, the data-driven weights of the low-frequency error fitting layer and the high-frequency residual compensation layer are increased. Specifically, the weight of the low-frequency error fitting layer is increased from the initial value of 0.30 to 0.45, the weight of the high-frequency residual compensation layer is increased from the initial value of 0.20 to 0.35, and the weight of the basic mechanism layer is decreased from the initial value of 0.50 to 0.20. The sum of the weights of the three layers is normalized to 1. When the residual amplitude is less than or equal to the preset amplitude threshold, it is determined that the current model simulation output has approximated the measured value. To maintain the dominance of physical constraints, the fusion weight of the basic mechanism layer is increased. Specifically, the weight of the basic mechanism layer is increased from the initial value of 0.50 to 0.75, the weight of the low-frequency error fitting layer is decreased from 0.30 to 0.15, and the weight of the high-frequency residual compensation layer is decreased from 0.20 to 0.10. The sum of the weights of the three layers is normalized to 1. The fused simulation output value is obtained by multiplying the output value of each of the three layers by its corresponding weight and then summing the results.

[0029] Step S300: Using the residual between the fusion simulation output value and the actual satellite telemetry value, a closed-loop on-orbit autonomous iterative evolution is performed to obtain the satellite attitude and orbit control fusion twin update model.

[0030] Specifically, the fused simulation values ​​output in step S200 are compared with the actual telemetry attitude and orbit data on the satellite to obtain residual data and the rate of change of residual data. The residual signal is then converted into correction values ​​for the data-driven layer network parameters using a proportional-derivative adjustment formula. This updates the network weights and biases of the low-frequency error fitting layer and the high-frequency residual compensation layer in real time, while the parameters of the basic mechanism layer remain unchanged. This step enables the model to autonomously iterate and evolve as the satellite's on-orbit environment changes and components age, ensuring that the twin model remains synchronized with the real satellite's state throughout its entire lifecycle.

[0031] In one possible implementation, closed-loop on-orbit autonomous iterative evolution is performed using the residual between the fused simulation output value and the actual satellite telemetry value. Step S300 further includes step S310, calculating the residual data and the residual data change rate based on the fused simulation output value and the actual satellite telemetry value. Specifically, the fused simulation output value and the actual satellite telemetry value are acquired in each control cycle. The fused simulation output value includes simulation results in four dimensions: attitude angle, attitude angular velocity, orbital position, and orbital velocity. The actual satellite telemetry values ​​include the three-axis attitude angle and three-axis attitude angular velocity output from the star sensor, and the orbital position and orbital velocity output from the global navigation satellite system receiver. The fused simulation output values ​​are subtracted from the actual satellite telemetry values ​​at the same time, dimension by dimension. Specifically, for attitude angle, the fused simulation output attitude angle is subtracted from the star sensor's measured attitude angle; for angular velocity, the fused simulation output angular velocity is subtracted from the star sensor's measured angular velocity; for orbital position, the fused simulation output orbital position is subtracted from the GNSS receiver's measured orbital position; and for orbital velocity, the fused simulation output orbital velocity is subtracted from the GNSS receiver's measured orbital velocity. The differences in each dimension are used to form a residual data vector. The residual data vector at the current time is subtracted from the residual data vector at the previous sampling time to obtain a residual data difference vector. This residual data difference vector is then divided by the sampling time interval to obtain the residual data rate of change vector.

[0032] Step S320: Based on the residual data and the rate of change of the residual data, a closed-loop on-orbit autonomous iterative evolution is performed on the low-frequency error fitting layer and the high-frequency residual compensation layer using a proportional-derivative adjustment formula; the proportional-derivative adjustment formula is: ; Where, σ i denoted as , where is the adjustment amount of the hierarchical parameters of the i-th layer data-driven network; e represents the residual data. k represents the rate of change of the residual data. p This is the proportional adjustment coefficient, with a value range of 0.01 to 0.1, used to control the basic correction strength based on the current residual; k d This is the differential adjustment coefficient, with a value range of 0.1 to 1.0. It is used to apply damping correction based on the residual change trend to suppress network iterative oscillations.

[0033] Specifically, in each control cycle, the residual data vector and the residual data change rate vector are substituted into the aforementioned proportional-derivative adjustment formula to calculate the hierarchical parameter correction σ1 for the low-frequency error fitting layer and the hierarchical parameter correction σ2 for the high-frequency residual compensation layer. σ1 is added to all current network parameters of the low-frequency error fitting layer to obtain the updated network parameters; σ2 is added to all current network parameters of the high-frequency residual compensation layer to obtain the updated network parameters. The above parameter updates only apply to the trainable parameters of the low-frequency error fitting layer and the high-frequency residual compensation layer, including the weight matrix and bias vector. All parameters of the basic mechanism layer do not participate in any of the above parameter update operations and remain fixed. After the parameter updates are completed, the low-frequency error fitting layer and the high-frequency residual compensation layer perform forward calculations using the updated network parameters from subsequent control cycles, completing the closed-loop on-orbit autonomous iterative evolution and obtaining the satellite attitude and orbit control fusion twin update model.

[0034] Step S400: Deploy the satellite attitude and orbit control fusion twin update model in a lightweight, tiered manner at both the satellite and ground levels.

[0035] Specifically, the complete model that underwent autonomous on-orbit iterative evolution in step S300 is deployed to both the satellite and the ground. On the satellite, the basic mechanism layer is fully preserved, and the data-driven layer is compressed to generate a lightweight version, which is then mounted on the satellite's embedded processor to achieve millisecond-level real-time state mirroring. On the ground, an uncompressed, high-precision model is deployed, utilizing the ground's high computing power to complete full-condition simulation and global parameter optimization. The optimized network parameters are periodically sent back to the lightweight model on the satellite for version updates. This step, through the division of labor and collaboration between on-board real-time inference and ground-based global optimization, eliminates the asynchrony between virtual and real systems caused by telemetry transmission delays and compensates for the limited computing power on the satellite, which prevents the running of the complete model.

[0036] In one possible implementation, the satellite attitude and orbit control fusion twin update model is deployed in a lightweight, hierarchical manner at both the satellite and ground levels. Step S400 further includes step S410: on the satellite, the basic mechanism layer of the satellite attitude and orbit control fusion twin update model is fully preserved, and the low-frequency error fitting layer and the high-frequency residual compensation layer are compressed to generate a lightweight on-board model, which is then mounted on the onboard embedded processor. Specifically, the basic mechanism layer of the satellite attitude and orbit control fusion twin update model is fully preserved, that is, all physical equations and factory-fixed parameters are retained without any pruning or compression. The low-frequency error fitting layer and the high-frequency residual compensation layer are compressed to generate a lightweight on-board model. The specific implementation of model compression is as follows: structured pruning is performed on the low-frequency error fitting layer and the high-frequency residual compensation layer. Specifically, the layers are sorted according to the sum of the absolute values ​​of the weights of each channel, and the channels with the lowest sum of weights are removed. The pruning ratio is set to 30%, that is, the 30% of channels with the lowest contribution and their corresponding connection weights are removed. After pruning, the pruned network is compressed using 8-bit quantization, which converts the floating-point representation of each layer's network parameters into an 8-bit integer representation. The quantization range is the dynamic range determined by the maximum absolute value of each layer's parameters. After model compression, the basic mechanism layer, the compressed low-frequency error fitting layer, and the compressed high-frequency residual compensation layer are combined in parallel with the same structure as in step S140 to generate a lightweight on-board model. This lightweight on-board model is then mounted on an onboard embedded processor, which is an FPGA or ARM architecture chip. An alternative implementation for on-board model compression is knowledge distillation. This involves using a fully uncompressed, complete, three-level fusion high-precision twin model deployed on the ground as the teacher model, and a small on-board network to be deployed as the student network. The student network's size is 20% to 30% of the teacher model's parameters. The teacher model's output is used as a soft label for distillation training of the student network. The distillation temperature parameter is set to 4, the distillation loss weight is set to 0.7, and the student network's cross-entropy loss weight is set to 0.3. After training, the student network is deployed as the lightweight on-board model.

[0037] Step S420: On the ground, the satellite attitude and orbit control fusion twin update model is fully deployed for global simulation and parameter optimization, and the optimized network parameters are periodically distributed to the on-board lightweight model for version updates. Specifically, the uncompressed satellite attitude and orbit control fusion twin update model is fully deployed on the ground simulation server, that is, all physical equations and parameters of the basic mechanism layer, all long short-term memory network or gated recurrent unit network parameters of the low-frequency error fitting layer, and all one-dimensional convolutional neural network or lightweight Transformer micro-network parameters of the high-frequency residual compensation layer are fully preserved without any pruning or quantization compression. The ground-based high-computing-power server is used to complete full-condition simulation, fault simulation, and global parameter optimization. The full-condition simulation includes satellite attitude and orbit state simulation under different orbital altitudes, seasons, and solar activity levels; the fault simulation includes the injection and simulation of typical fault modes such as flywheel jamming, gyroscope drift, and thruster failure; global parameter optimization adopts a Bayesian optimization algorithm, with the root mean square error between the model simulation output and historical telemetry data as the objective function, and the optimization variables including all network parameters of the low-frequency error fitting layer and the high-frequency residual compensation layer, with the number of iterations set to 500 rounds. After the global parameter optimization is completed, the optimized network parameters of the low-frequency error fitting layer and the high-frequency residual compensation layer are uploaded to the satellite via the satellite-to-ground communication link. The satellite receives the data and replaces the network parameters of the corresponding layer in the lightweight model on the satellite, completing the version update. The version update time cycle is set according to the satellite orbital revolution. For example, for low-Earth orbit satellites, it can be set to update once every one orbital revolution; for geostationary orbit satellites, it can be set to update once every 24 hours.

[0038] This application's embodiments address the technical problem of existing attitude and orbit control system modeling methods where model prediction accuracy gradually degrades with increasing on-orbit time and is difficult to maintain high-precision state reproduction capability throughout the entire lifecycle. This is achieved by constructing a three-level hierarchical decoupled fusion twin basic model, adaptively weighting the three-level outputs based on the confidence index of on-orbit telemetry data to obtain fusion simulation values, using the residual between the fusion simulation values ​​and actual telemetry values ​​for closed-loop autonomous iterative evolution to update the model, and deploying the updated model in a lightweight, tiered manner at both the satellite and ground levels.

[0039] In the above text, refer to Figure 1 A hierarchical fusion digital twin modeling method for satellite attitude and orbit control according to an embodiment of the present invention is described in detail. Next, reference will be made to... Figure 2 This invention describes a hierarchical fusion digital twin modeling system for satellite attitude and orbit control according to an embodiment of the present invention.

[0040] A satellite attitude and orbit control hierarchical fusion digital twin modeling system according to an embodiment of the present invention addresses the technical problem that existing attitude and orbit control system modeling methods suffer from a gradual degradation in model prediction accuracy with increasing on-orbit time and difficulty in maintaining high-precision state reproduction capability throughout the entire lifecycle. The system achieves the technical effect of enabling the satellite attitude and orbit control digital twin model to maintain high-precision state reproduction capability throughout its entire lifecycle. The satellite attitude and orbit control hierarchical fusion digital twin modeling system includes: a hierarchical decoupled twin basic model construction module 10, an adaptive weighted fusion module 20, a closed-loop autonomous evolution update module 30, and a satellite-ground two-level lightweight hierarchical deployment module 40.

[0041] The hierarchical decoupled twin model construction module 10 is used to construct a three-level hierarchical decoupled satellite attitude and orbit control fusion twin basic model; the adaptive weighted fusion module 20 is used to adaptively and dynamically allocate weights to the three-level outputs of the satellite attitude and orbit control fusion twin basic model based on the confidence evaluation index of the current on-orbit telemetry data, so as to obtain the fusion simulation output value; the closed-loop autonomous evolution update module 30 is used to use the residual between the fusion simulation output value and the actual satellite telemetry value to perform closed-loop on-orbit autonomous iterative evolution, so as to obtain the satellite attitude and orbit control fusion twin update model; the space-ground two-level lightweight hierarchical deployment module 40 is used to deploy the satellite attitude and orbit control fusion twin update model in a space-ground two-level lightweight hierarchical manner.

[0042] The detailed configuration of the layered decoupled twin model construction module 10 is explained as follows: As mentioned above, to construct a three-level layered decoupled satellite attitude and orbit control fusion twin basic model, the layered decoupled twin model construction module 10 may further include: a low-frequency error timing fitting unit for building a low-frequency error fitting layer, which uses a timing neural network to output low-frequency compensation; a high-frequency residual lightweight compensation unit for building a high-frequency residual compensation layer, which uses a lightweight high-frequency compensation network to output high-frequency compensation; a basic physical mechanism fixed model unit for building a basic mechanism layer, which embeds the satellite attitude and orbit control physical equations to solve the orbital mechanism and reserves an attitude coupling term correction interface to receive the low-frequency compensation and the high-frequency compensation; and a three-level decoupled twin basic model integration unit for constructing the satellite attitude and orbit control fusion twin basic model based on the basic mechanism layer, the low-frequency error fitting layer, and the high-frequency residual compensation layer.

[0043] The hierarchical decoupling twin model construction module 10 may further include: the temporal neural network is an LSTM network or a GRU network; the lightweight high-frequency compensation network is a one-dimensional CNN network or a lightweight Transformer micro-network.

[0044] The detailed description of the specific configuration of the closed-loop autonomous evolution update module 30 is explained as follows: As mentioned above, closed-loop on-orbit autonomous iterative evolution is performed using the residual between the fused simulation output value and the actual satellite telemetry value. The closed-loop autonomous evolution update module 30 may further include: a residual and rate of change calculation unit for calculating residual data and residual data change rate based on the fused simulation output value and the actual satellite telemetry value; and a proportional-derivative closed-loop evolution control unit for performing closed-loop on-orbit autonomous iterative evolution of the low-frequency error fitting layer and the high-frequency residual compensation layer using a proportional-derivative adjustment formula based on the residual data and the residual data change rate.

[0045] The proportional-derivative closed-loop evolution control unit may further include: the proportional-derivative adjustment formula is: ; where σ i denoted as , where is the adjustment amount of the hierarchical parameters of the i-th layer data-driven network; e represents the residual data. k represents the rate of change of the residual data. p This is the proportional adjustment coefficient, with a value range of 0.01 to 0.1, used to control the basic correction strength based on the current residual; k d This is the differential adjustment coefficient, with a value range of 0.1 to 1.0. It is used to apply damping correction based on the residual change trend to suppress network iterative oscillations.

[0046] The specific configuration of the adaptive weighted fusion module 20 is described in detail below: As mentioned above, based on the current on-orbit telemetry data confidence evaluation index, adaptive dynamic weight allocation is performed on the three-layer output of the satellite attitude and orbit control fusion twin basic model. The adaptive weighted fusion module 20 may further include: a sampling point density-driven weight allocation unit used to count the current effective telemetry sampling points in real time when the confidence evaluation index is the number of effective telemetry sampling points, and to increase the weight of the low-frequency error fitting layer and the weight of the sampled data when the current effective telemetry sampling points are greater than a preset high threshold. The data-driven weights of the high-frequency residual compensation layer are described above; when the number of currently valid telemetry sampling points is less than a preset low threshold, the fusion weights of the basic mechanism layer are increased; the residual amplitude-triggered weight allocation unit is used to calculate the absolute value or vector magnitude of the residual data to obtain the residual amplitude when the confidence evaluation index is the residual data; when the residual amplitude is greater than a preset amplitude threshold, the data-driven weights of the low-frequency error fitting layer and the high-frequency residual compensation layer are increased; when the residual amplitude is less than or equal to the preset amplitude threshold, the fusion weights of the basic mechanism layer are increased.

[0047] The detailed description of the specific configuration of the satellite-ground two-level lightweight hierarchical deployment module 40 is explained as follows: As mentioned above, the satellite attitude and orbit control fusion twin update model is deployed in a satellite-ground two-level lightweight hierarchical manner. The satellite-ground two-level lightweight hierarchical deployment module 40 may further include: a satellite-borne lightweight compression deployment unit, which is used on the satellite to fully retain the basic mechanism layer of the satellite attitude and orbit control fusion twin update model, and to compress the low-frequency error fitting layer and the high-frequency residual compensation layer to generate a satellite-borne lightweight model, which is carried on the satellite-borne embedded processor; and a ground-based global optimization and parameter distribution unit, which is used on the ground to fully deploy the satellite attitude and orbit control fusion twin update model for global simulation and parameter optimization, and to periodically distribute the optimized network parameters to the satellite-borne lightweight model for version updates.

[0048] The satellite attitude and orbit control hierarchical fusion digital twin modeling system provided in this embodiment of the invention can execute the satellite attitude and orbit control hierarchical fusion digital twin modeling method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0049] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A hierarchical fusion digital twin modeling method for satellite attitude and orbit control, characterized in that, The method includes: Construct a three-level, layered, decoupled satellite attitude and orbit control fusion twin basic model; Based on the current confidence evaluation index of on-orbit telemetry data, the three-layer output of the satellite attitude and orbit control fusion twin basic model is adaptively and dynamically weighted to obtain the fusion simulation output value. By utilizing the residual between the fusion simulation output value and the actual satellite telemetry value, a closed-loop on-orbit autonomous iterative evolution is performed to obtain the satellite attitude and orbit control fusion twin update model; The satellite attitude and orbit control fusion twin update model is deployed in a lightweight, tiered manner at both the satellite and ground levels.

2. The satellite attitude and orbit control hierarchical fusion digital twin modeling method as described in claim 1, characterized in that, Construct a three-level, layered, decoupled satellite attitude and orbit control fusion twin basic model, including: A low-frequency error fitting layer is constructed, which adopts a temporal neural network to output low-frequency compensation. A high-frequency residual compensation layer is constructed, which adopts a lightweight high-frequency compensation network and outputs high-frequency compensation amount; A basic mechanism layer is constructed, in which the physical equations of satellite attitude and orbit control are fixedly embedded to solve the mechanism of orbit, and an attitude coupling term correction interface is reserved to receive the low-frequency compensation amount and the high-frequency compensation amount; Based on the aforementioned basic mechanism layer, the low-frequency error fitting layer, and the high-frequency residual compensation layer, the satellite attitude and orbit control fusion twin basic model is constructed.

3. The satellite attitude and orbit control hierarchical fusion digital twin modeling method as described in claim 2, characterized in that, The temporal neural network is an LSTM network or a GRU network; the lightweight high-frequency compensation network is a one-dimensional CNN network or a lightweight Transformer micro-network.

4. The satellite attitude and orbit control hierarchical fusion digital twin modeling method as described in claim 2, characterized in that, Using the residual between the fused simulation output value and the actual satellite telemetry value, a closed-loop on-orbit autonomous iterative evolution is performed, including: Based on the fusion simulation output value and the actual satellite telemetry value, the residual data and the residual data change rate are calculated; Based on the residual data and the rate of change of the residual data, the low-frequency error fitting layer and the high-frequency residual compensation layer are subjected to closed-loop on-orbit autonomous iterative evolution using the proportional-derivative adjustment formula.

5. The satellite attitude and orbit control hierarchical fusion digital twin modeling method as described in claim 4, characterized in that, The proportional-derivative adjustment formula is as follows: ； Where, σ i Let be the hierarchical parameter correction amount of the i-th layer data-driven network; e is the residual data; is the rate of change of the residual data; k p This is the proportional adjustment coefficient, with a value range of 0.01 to 0.1, used to control the basic correction strength based on the current residual; k d This is the differential adjustment coefficient, with a value range of 0.1 to 1.

0. It is used to apply damping correction based on the residual change trend to suppress network iterative oscillations.

6. The satellite attitude and orbit control hierarchical fusion digital twin modeling method as described in claim 4, characterized in that, Based on the current confidence evaluation index of on-orbit telemetry data, an adaptive dynamic weight allocation is performed on the three-layer output of the satellite attitude and orbit control fusion twin basic model, including: When the confidence evaluation index is the number of valid telemetry sampling points, the number of current valid telemetry sampling points is counted in real time. When the number of current valid telemetry sampling points is greater than a preset high threshold, the data driving weights of the low-frequency error fitting layer and the high-frequency residual compensation layer are increased. When the number of current valid telemetry sampling points is less than a preset low threshold, the fusion weights of the basic mechanism layer are increased. Alternatively, when the confidence evaluation index is the residual data, the absolute value or vector magnitude of the residual data is calculated to obtain the residual amplitude. When the residual amplitude is greater than a preset amplitude threshold, the data driving weights of the low-frequency error fitting layer and the high-frequency residual compensation layer are increased; when the residual amplitude is less than or equal to the preset amplitude threshold, the fusion weights of the basic mechanism layer are increased.

7. The satellite attitude and orbit control hierarchical fusion digital twin modeling method as described in claim 2, characterized in that, The satellite attitude and orbit control fusion twin update model is deployed in a lightweight, tiered manner at both the space and ground levels, including: On the satellite, the basic mechanism layer of the satellite attitude and orbit control fusion twin update model is fully preserved, and the low-frequency error fitting layer and the high-frequency residual compensation layer are compressed to generate a lightweight on-board model, which is carried on the satellite embedded processor. On the ground, the satellite attitude and orbit control fusion twin update model is fully deployed for global simulation and parameter optimization, and the optimized network parameters are periodically sent to the on-board lightweight model for version updates.

8. A hierarchical fusion digital twin modeling system for satellite attitude and orbit control, characterized in that, The system is used to implement the satellite attitude and orbit control hierarchical fusion digital twin modeling method according to any one of claims 1-7, the system comprising: A hierarchical decoupled twin model construction module is used to construct a three-level hierarchical decoupled satellite attitude and orbit control fusion twin basic model; The adaptive weighted fusion module is used to adaptively and dynamically allocate weights to the three-layer output of the satellite attitude and orbit control fusion twin basic model based on the confidence evaluation index of the current on-orbit telemetry data, so as to obtain the fusion simulation output value. The closed-loop autonomous evolution update module is used to perform closed-loop on-orbit autonomous iterative evolution by utilizing the residual between the fusion simulation output value and the actual telemetry value of the satellite to obtain the satellite attitude and orbit control fusion twin update model. The satellite-ground two-level lightweight hierarchical deployment module is used to deploy the satellite attitude and orbit control fusion twin update model in a satellite-ground two-level lightweight hierarchical manner.

Images