Digital satellite modeling method, device, equipment, storage medium and product

Abstract

This invention provides a digital satellite modeling method, apparatus, device, storage medium, and product, relating to the field of satellite technology. The method includes: constructing a mechanistic model of the satellite; collecting on-orbit telemetry data measured during the satellite's on-orbit operation, and constructing a data-driven model based on the deviation data between the on-orbit telemetry data and the output results of the mechanistic model; predicting a basic satellite state prediction value at a future time using the mechanistic model, and predicting a deviation compensation value at the future time using the data-driven model; and correcting the basic satellite state prediction value based on the deviation compensation value to obtain a final satellite state prediction value. This invention can improve the prediction accuracy of digital satellite models.

Description

Technical Field

[0001] This invention relates to the field of satellite modeling technology, and in particular to a digital satellite modeling method, apparatus, equipment, storage medium, and product. Background Technology

[0002] With the increasing demands for intelligent operation and maintenance and on-orbit health management of spacecraft, digital twin technology has become a key enabling technology in the aerospace field. As a mirror image of a physical satellite in digital space, the modeling accuracy of a digital satellite directly determines the reliability of state prediction, fault diagnosis, and life assessment. Currently, digital satellite modeling is mainly conducted through pure mechanistic modeling or pure data-driven methods. However, pure mechanistic modeling struggles to accurately describe the uncertainties and time-varying deviations in complex physical processes, resulting in limited prediction accuracy; pure data-driven modeling lacks physical constraints, is sensitive to anomalous data, and lacks interpretability in its prediction results. It is evident that existing single modeling methods all have inherent flaws, making it difficult to balance prediction accuracy, robustness, and interpretability, thus failing to meet the high-reliability engineering application requirements of digital satellite modeling. Summary of the Invention

[0003] This invention provides a digital satellite modeling method, apparatus, device, storage medium, and product to solve the technical problem that the modeling effect of digital satellites cannot meet the requirements of high-reliability engineering applications due to the use of a single modeling method in the prior art.

[0004] This invention provides a digital satellite modeling method, comprising the following steps: Construct a mechanistic model of the satellite; Collect on-orbit telemetry data measured during satellite operation, and construct a data-driven model based on the on-orbit telemetry data. The underlying satellite state prediction value for future times is predicted using the aforementioned mechanism model, and the deviation compensation value for future times is predicted using the aforementioned data-driven model. Based on the deviation compensation value, the basic satellite state prediction value is corrected to obtain the final satellite state prediction value.

[0005] According to a digital satellite modeling method provided by the present invention, the step of constructing a data-driven model based on the on-orbit telemetry data includes: Build the initial data-driven model; The true value of the satellite state at the target time is obtained from the on-orbit telemetry data, and the basic satellite state value at the same target time is determined through the mechanism model. Based on the true satellite state value and the basic satellite state value, determine the deviation vector; Based on the deviation vector, a training sample set is constructed, and the initial data-driven model is trained using the training sample set to obtain a trained data-driven model.

[0006] According to a digital satellite modeling method provided by the present invention, the step of constructing a training sample set based on the deviation vector includes: Obtain the operating condition characteristics corresponding to the target time, wherein the operating condition characteristics include at least the satellite attitude quaternion, solar incidence angle, geomagnetic activity index, and orbital altitude; A training sample set is constructed based on the operating condition characteristics and the deviation vector.

[0007] The digital satellite modeling method provided by the present invention further includes: Acquire real-time on-orbit telemetry data; The status data at multiple time points are obtained from the multiple real-time on-orbit telemetry data, and a sliding time window is constructed; Calculate the real-time deviation between the basic satellite state value output by the mechanism model at each time point within the sliding time window and the corresponding state data; Based on the real-time deviation, the root mean square error is calculated, and the slope of the deviation trend is fitted by linear regression. The data-driven model is updated based on the root mean square error and the slope of the deviation trend.

[0008] According to a digital satellite modeling method provided by the present invention, updating the data-driven model based on the root mean square error and the slope of the deviation trend includes: Based on the root mean square error and the slope of the deviation trend, determine whether the preset triggering condition is met; When the root mean square error is greater than a preset accuracy threshold, or the absolute value of the deviation trend slope is greater than a preset drift threshold, or when an external event flag is received, the model update process is triggered to dynamically update the parameters of the data-driven model.

[0009] According to a digital satellite modeling method provided by the present invention, the dynamic updating of the parameters of the data-driven model includes: The real-time deviations corresponding to each time point within the sliding time window are used as valid samples. Based on the aforementioned valid samples, an incremental training dataset is constructed; The data-driven model is incrementally learned using the incremental training dataset, and the parameters of the data-driven model are dynamically updated.

[0010] The present invention also provides a digital satellite modeling device, comprising the following modules: Building blocks are used to construct mechanistic models of satellites; The acquisition module is used to acquire on-orbit telemetry data measured during the satellite's on-orbit operation, and to construct a data-driven model based on the on-orbit telemetry data. The prediction module is used to predict the basic satellite state prediction value at a future time through the mechanistic model, and to predict the deviation compensation value at the future time through the data-driven model. The correction module corrects the basic satellite state prediction value based on the deviation compensation value to obtain the final satellite state prediction value.

[0011] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the digital satellite modeling method as described above.

[0012] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the digital satellite modeling method as described above.

[0013] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the digital satellite modeling method as described above.

[0014] This invention provides a digital satellite modeling method, apparatus, device, storage medium, and product, which constructs a mechanistic model of a satellite; collects on-orbit telemetry data measured during the satellite's on-orbit operation, and constructs a data-driven model based on the on-orbit telemetry data; predicts the basic satellite state prediction value at future times using the mechanistic model, and predicts the deviation compensation value at the future times using the data-driven model; and corrects the basic satellite state prediction value based on the deviation compensation value to obtain the final satellite state prediction value. This invention solves the technical problem that the modeling effect of digital satellites cannot meet the high reliability requirements of engineering applications due to the use of a single modeling method. Compared with existing technologies, this invention provides a physically meaningful basic prediction through a mechanistic model, ensuring the physical interpretability and reliability of the prediction results. Simultaneously, it uses a data-driven model constructed based on real on-orbit telemetry data to predict and specifically correct the output deviation of the mechanistic model, effectively improving the prediction accuracy of the satellite state and the model's adaptability to time-varying uncertainties. This achieves organic synergy between physical laws and data-driven approaches, significantly enhancing the robustness, long-term operational stability, and engineering practical value of the digital satellite model. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is one of the flowcharts of the digital satellite modeling method provided by the present invention.

[0017] Figure 2 This is an example diagram of the data-driven model of the digital satellite modeling method provided by the present invention.

[0018] Figure 3 This is a detailed flowchart illustrating the digital satellite modeling method provided by the present invention.

[0019] Figure 4 This is the second flowchart of the digital satellite modeling method provided by the present invention.

[0020] Figure 5 This is a schematic diagram of the structure of the digital satellite modeling device provided by the present invention.

[0021] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] The following is combined Figure 1 and Figure 4 The present invention describes a digital satellite modeling method applicable to any digital satellite modeling. The subject executing this method can be an electronic device or a digital satellite modeling device installed in the electronic device. The digital satellite modeling device can be implemented by software, hardware, or a combination of both.

[0024] Figure 1 This is one of the flowcharts illustrating the digital satellite modeling method provided by this invention, such as... Figure 1 As shown, the method includes the following: Step 101: Construct a mechanistic model of the satellite; It should be noted that the satellite's mechanistic model refers to a mathematical model built based on the physical laws and first principles of satellite design. It is used to output physically meaningful basic predictions of the satellite's state, providing fundamental physical constraints and initial prediction basis for digital satellites. A satellite consists of multiple subsystems, including the orbital subsystem, energy subsystem, thermal control subsystem, attitude and orbit control subsystem, propulsion subsystem, telemetry, tracking and command (TT&C) and data transmission subsystem, structural and mechanical subsystem, integrated electronic subsystem, and power supply and distribution subsystem. Each subsystem corresponds to an independent mechanistic model, and these subsystem mechanistic models together constitute the overall satellite mechanistic model system.

[0025] Understandably, by constructing a mechanistic model, the basic simulation capability of digital satellites has been established. This model can calculate the theoretical state of the satellite at any future moment based on the satellite's initial state (such as position, velocity, attitude, temperature of various components, battery charge, etc.) and external input conditions (such as solar radiation, atmospheric drag, Earth's gravitational field, etc.).

[0026] In practical implementation, the satellite object to be modeled can be determined first, such as a communication satellite or a remote sensing satellite. Then, for the satellite's subsystems such as orbit, energy, thermal control, attitude and orbit control, propulsion, telemetry and data transmission, structure and mechanism, integrated electronics, and power supply and distribution, the corresponding differential equations or algebraic equations (i.e., the mechanism models corresponding to each subsystem of the satellite) can be established respectively. For example, for the orbital subsystem, a mechanism model can be constructed using universal gravitation, celestial perturbation theory, and numerical integration methods; for the energy subsystem, a mechanism model can be constructed using photovoltaic array output models, battery charging and discharging models, and energy balance equations; for the thermal control subsystem, a mechanism model can be constructed using the first law of thermodynamics and heat transfer theory; for the attitude and orbit control subsystem, a mechanism model can be constructed using rigid body dynamics and the law of angular momentum; for the propulsion subsystem, a mechanism model can be constructed using fluid dynamics, thermodynamics, and Newton's third law; for the telemetry, tracking, and command (TT&C) and data transmission subsystem, a mechanism model can be constructed using electromagnetics and radio frequency circuit theory; for the structural and mechanism subsystem, a mechanism model can be constructed using mechanics of materials and structural dynamics; for the integrated electronic subsystem, a mechanism model can be constructed using computer architecture and real-time control theory; and for the power supply and distribution subsystem, a mechanism model can be constructed using circuit theory.

[0027] Step 102: Collect on-orbit telemetry data measured during the satellite's on-orbit operation, and construct a data-driven model based on the on-orbit telemetry data. It should be noted that on-orbit telemetry data refers to the real-time sensor measurements transmitted from a satellite to the ground via its telemetry system during its actual operation in space, such as temperature, voltage, current, attitude angles, and orbital elements. Data-driven models are statistical learning models that make predictions by learning mapping relationships from historical data, such as neural networks and support vector regression. Their role is to capture complex dynamics and biases that mechanistic models cannot fully describe.

[0028] Understandably, collecting real-world on-orbit telemetry data provides training material for data-driven models. This data contains real dynamic characteristics that mechanistic models cannot accurately predict due to factors such as space environment disturbances, component wear, and modeling errors. Building data-driven models based on this data aims to enable machines to automatically learn the complex mapping relationship between the "mechanistic model output" and the "real state."

[0029] In the implementation, ground receiving stations receive telemetry data streams from satellites in real time, and then decode, verify, time-align, and store them to form a historical telemetry database. Training data is then extracted from this database. When building the data-driven model, a suitable machine learning model architecture for the task is first selected, such as a multi-layer feedforward neural network or a recurrent neural network. The model's input is designed to be operational features related to prediction bias, and its output is designed to be a bias compensation value of the same dimension as the satellite's state. The model parameters are initially set randomly and are subsequently optimized using training samples.

[0030] Step 103: Predict the basic satellite state prediction value at future time using the mechanism model, and predict the deviation compensation value at the future time using the data-driven model; It should be noted that the basic satellite state prediction value is the theoretical state value at a future moment, calculated by the mechanistic model based on the current input state and external force conditions through numerical integration or algebraic operations. The deviation compensation value is an estimate of the possible difference between the mechanistic model prediction value and the actual state, calculated by the data-driven model based on the current operating condition characteristics or historical state sequences.

[0031] Understandably, the mechanistic model provides a baseline prediction that conforms to physical laws, while the data-driven model specifically targets the systematic errors of this baseline prediction under the current specific conditions. Working together, they output the baseline values ​​and correction values ​​required for the final prediction, respectively, thus preparing for subsequent high-precision fusion predictions.

[0032] In specific implementations, such as Figure 2The diagram shows an example of a data-driven model. This model employs a time-series neural network architecture (such as LSTM or GRU) based on a gating mechanism to learn the dynamic mapping relationship between the operating condition feature sequence and the deviation compensation value. The diagram illustrates the connection structure between adjacent time steps t-1 and t: x t-1 x t h is the input working condition feature vector. t-1 This represents the hidden state from the previous time step. σ represents the Sigmoid gating function, which controls the selective memorization and forgetting of information; tanh is the hyperbolic tangent function, used to generate candidate states. Through a combination of multiplication and addition operations, the hidden state is updated and propagated, enabling the network to capture the temporal evolution of operational characteristics and provide accurate bias compensation for the mechanistic model.

[0033] Step 104: Based on the deviation compensation value, correct the basic satellite state prediction value to obtain the final satellite state prediction value.

[0034] It should be noted that the final satellite state prediction value refers to the result obtained by fusing the basic satellite state prediction value output by the mechanistic model with the deviation compensation value output by the data-driven model. Its function is to serve as the final state estimate output by the digital satellite model, which retains the physical constraints of the mechanistic model and incorporates the self-learning and correction capabilities of the data-driven model.

[0035] Understandably, by superimposing the deviation compensation value onto the basic satellite state prediction value, collaborative prediction of the dual-drive model is achieved. Specifically, for any prediction time... First, the basic satellite state prediction values ​​are obtained from the mechanism model. Simultaneously, bias compensation values ​​are obtained from the trained data-driven model. Then, the final satellite state prediction is calculated using an additive compensation method: This correction process can be applied to each component of the state vector (such as position and velocity components). For the orbital subsystem, this means that the position and velocity in the x, y, and z directions are compensated component by component. Through this correction method, the final predicted value will not deviate from the physical laws (because the basic predicted value is derived from rigorous mechanical equations, providing reasonable physical boundaries) and can effectively approximate the actual state (because the deviation compensation value continuously learns from telemetry data and corrects the difference between the model and reality). This achieves a balance between physical interpretability and data adaptability, solving the technical problems of insufficient accuracy of pure mechanistic models and lack of physical constraints in pure data-driven models, and significantly improving the prediction accuracy and robustness of digital satellite models.

[0036] It should be noted that the final satellite state prediction is the core output of the digital satellite model, used to characterize the physical satellite at the predicted time t. k The estimated state. This predicted value can be directly input into the satellite ground operation and maintenance system or the onboard autonomous management module, serving the following specific application scenarios: In status monitoring, it serves as the basis for judging the current health status of the satellite, identifying potential faults such as orbital deviation, attitude anomalies, or insufficient energy by comparing it with preset safety thresholds; In fault diagnosis, it serves as the input to fault detection and isolation algorithms, locating the specific subsystem or component where the anomaly occurred through residual analysis with real telemetry data; In life assessment, it serves as the basis for long-term trend analysis, assessing the degree of propellant consumption, battery degradation, or optical device performance degradation through the time series changes of accumulated predicted values; In mission planning, it serves as the basis for orbit prediction and attitude pointing prediction, assisting in calculating the Earth observation window, satellite-to-ground communication link budget, and orbital maneuvering timing.

[0037] In specific implementations, such as Figure 3 The flowchart shown details the complete execution flow of the digital satellite modeling method. After the process begins, the system executes two paths in parallel: on the one hand, it acquires the true orbital state value in real time; on the other hand, it constructs a satellite orbit extrapolation mechanism model and a data-driven model. Subsequently, based on the true orbital state value and the output of the mechanism model, it constructs a deviation training sample set and trains the data-driven model. After training, the mechanism model and the data-driven model are fused to obtain the final predicted orbital state value. Based on this, the system calculates the root mean square error and trend slope of the final predicted orbital state value and the telemetry deviation, and determines whether to trigger the model update process. Triggering conditions include accuracy degradation, trend drift, and significant event input. If no update is triggered, the process ends. If an update is triggered, it enters the incremental model training phase. First, it constructs an incremental deviation sample set, then performs incremental training to update the data-driven model parameters. After the update is completed, it loops back to continue acquiring the real-time orbital state value, forming a continuous adaptive closed-loop optimization process.

[0038] This invention constructs a mechanistic model of a satellite; collects on-orbit telemetry data measured during the satellite's operation, and builds a data-driven model based on this data; predicts the basic satellite state at future times using the mechanistic model, and predicts the deviation compensation value at those future times using the data-driven model; and corrects the basic satellite state prediction value based on the deviation compensation value to obtain the final satellite state prediction value. This solves the technical problem that using a single modeling method results in digital satellite modeling that cannot meet the high-reliability engineering application requirements. Compared to existing technologies, this invention provides physically meaningful basic predictions through a mechanistic model, ensuring the physical interpretability and reliability of the prediction results. Simultaneously, it uses a data-driven model built based on real on-orbit telemetry data to predict and specifically correct the output deviation of the mechanistic model, effectively improving the prediction accuracy of the satellite state and the model's adaptability to time-varying uncertainties. This achieves organic synergy between physical laws and data-driven approaches, significantly enhancing the robustness, long-term operational stability, and engineering practical value of the digital satellite model.

[0039] Based on any of the above embodiments, the step of constructing a data-driven model based on the on-orbit telemetry data includes: Build the initial data-driven model; The true value of the satellite state at the target time is obtained from the on-orbit telemetry data, and the basic satellite state value at the same target time is determined through the mechanism model. Based on the true satellite state value and the basic satellite state value, determine the deviation vector; Based on the deviation vector, a training sample set is constructed, and the initial data-driven model is trained using the training sample set to obtain a trained data-driven model.

[0040] It should be noted that the initial data-driven model refers to a machine learning model with random or pre-set initialization parameters that has not yet been trained, such as a deep neural network with randomly initialized weights. Its role is to provide a learnable functional framework, waiting for subsequent learning processes to fit specific bias mapping relationships. The ground truth satellite state refers to the actual measured values ​​representing the satellite's true physical state, directly obtained from on-orbit telemetry data, such as the position and velocity output by a GPS receiver. Its role is to serve as a true label in supervised learning, used to evaluate the prediction accuracy of the mechanistic model, and to provide training targets for the data-driven model.

[0041] Understandably, the baseline satellite state value refers to the predicted satellite state value calculated by the mechanistic model at the same target time. The difference between this predicted value and the true satellite state value constitutes the bias that the data-driven model needs to learn. The bias vector is the difference between the true satellite state value and the baseline satellite state value, defined as: .

[0042] This vector quantitatively characterizes the prediction error of the mechanistic model under the current operating conditions, including the combined effects of unmodeled dynamics, parameter drift, and environmental disturbances.

[0043] In the specific implementation, an initial data-driven model is first constructed. For example, a three-layer fully connected neural network with an input dimension consistent with the number of operational features and an output dimension consistent with the dimension of the deviation vector, and model parameters initialized using Xavier or randomization; then, the target time t is extracted from the on-orbit telemetry data. i satellite state truth value (such as position and velocity), and simultaneously call the mechanistic model to calculate the basic satellite state values ​​at the same time. This ensures that the two are strictly aligned in time; then, the deviation vector is calculated point by point: This deviation vector serves as the learning objective of the data-driven model; then, the operating condition characteristics at each time point are... With the corresponding deviation vector Combine them into training sample pairs to construct a training sample set. Finally, the initial data-driven model is trained in a supervised manner using this training sample set, with mean squared error as the benchmark. As a loss function, the model parameters θ are iteratively optimized using an optimizer (such as Adam) until convergence, resulting in a well-trained data-driven model.

[0044] The digital satellite modeling method provided in this invention solves the problem of how to enable data-driven models to effectively learn unmodeled dynamics by explicitly utilizing the deviation between the output of the mechanistic model and the on-orbit true value to construct a training sample set. This technical solution achieves the effect of enabling data-driven models to focus on learning error patterns, thereby reducing learning difficulty, improving training efficiency, and increasing prediction accuracy.

[0045] Based on any of the above embodiments, constructing a training sample set based on the deviation vector includes: Obtain the operating condition characteristics corresponding to the target time, wherein the operating condition characteristics include at least the satellite attitude quaternion, solar incidence angle, geomagnetic activity index, and orbital altitude; A training sample set is constructed based on the operating condition characteristics and the deviation vector.

[0046] The digital satellite modeling method provided in this embodiment of the invention, It should be noted that operational characteristics are a set of characteristic parameters describing the environmental state and operational state of a satellite at the target time. The satellite attitude quaternion is used to uniquely and without singularities describe the satellite's orientation relative to the reference coordinate system; the solar incidence angle is the angle between sunlight and a reference surface of the satellite, such as the normal to the solar panel, which determines illumination and energy input; the geomagnetic activity index is an indicator that quantifies the degree of disturbance in the Earth's magnetic field, affecting the accuracy of atmospheric drag models and attitude control; and the orbital altitude is the distance between the satellite and the Earth's surface, determining environmental factors such as atmospheric density and gravitational field strength.

[0047] Understandably, obtaining these operational characteristics is a crucial "clue" for providing prediction biases in data-driven models. These characteristics reveal the reasons for biases in mechanistic models from different dimensions. For example, the bias of atmospheric drag models mainly varies with orbital altitude and geomagnetic activity index; the bias of solar radiation pressure models is closely related to the solar incidence angle and satellite attitude.

[0048] The digital satellite modeling method provided in this invention solves the problem of how to provide effective and sufficient input to the data-driven model to predict model deviations by clearly defining key operating condition characteristics, including satellite attitude quaternions, solar incidence angle, geomagnetic activity index, and orbital altitude. This technical solution enables the data-driven model to more accurately capture the nonlinear dependence between deviations and the operating environment and state, thereby improving the accuracy of deviation prediction.

[0049] Figure 4 This is the second flowchart of the digital satellite modeling method provided by the present invention, as shown in Figure 4, which also includes steps 105 to 109: Step 105: Acquire real-time on-orbit telemetry data; It should be noted that real-time on-orbit telemetry data refers to the various sensor measurements transmitted from the satellite to the ground in real time through its telemetry system at the current moment or the most recent moment, such as temperature, voltage, current, attitude angle, and orbital elements. This data serves as a monitoring signal for online model updates and is used to evaluate the prediction accuracy of the current digital satellite model.

[0050] Step 106: Obtain status data at multiple time points from the multiple real-time on-orbit telemetry data, and construct a sliding time window; It should be noted that a sliding time window refers to a fixed-length time interval that includes state data from the most recent M time points. Its purpose is to limit the range of samples used for evaluation and updates, avoiding the use of outdated data.

[0051] Understandably, a sliding time window allows for continuous monitoring of recent prediction error trends. The choice of window length M needs to balance sensitivity to short-term disturbances with stability to long-term trends. Specifically, a window length of M=100 is set. Each time new telemetry data arrives, the latest sample is added to the window, and the oldest sample is removed. The window contains the true state values, mechanistic model outputs, final predicted values, and operational characteristics for the most recent 100 time points, forming the data foundation for model health monitoring and incremental learning.

[0052] Step 107: Calculate the real-time deviation between the basic satellite state value output by the mechanism model at each time point within the sliding time window and the corresponding state data; It should be noted that real-time deviation refers to the difference between the baseline satellite state value output by the mechanistic model and the actual state data at the corresponding moment, i.e.: in, Represents the actual state data; This represents the basic satellite status value.

[0053] Step 108: Based on the real-time deviation, calculate the root mean square error and fit the slope of the deviation trend through linear regression. It should be noted that the root mean square error (RMS) refers to the square root of the mean square of the deviations within the window (RMS is used to determine whether accuracy has degraded), while the deviation trend slope refers to the slope of the linear regression of the absolute value of the deviation over time (the slope is used to determine whether there is systematic drift). Their respective functions are to reflect the current overall error level and the monotonic trend of the error.

[0054] In practical implementation, the formula for calculating the root mean square error is: In the specific implementation, the slope of the deviation trend is obtained by fitting a univariate linear regression: Where k represents the linear rate of change of the deviation over time, and b is the intercept. When k is significantly positive, it indicates that the deviation is increasing; when k is significantly negative, it indicates that the deviation is decreasing.

[0055] Step 109: Update the data-driven model based on the root mean square error and the slope of the deviation trend.

[0056] It should be noted that updating the data-driven model based on RMS and slope refers to initiating an incremental learning process for the model when monitoring indicators meet preset trigger conditions, dynamically adjusting model parameters to adapt to changes in satellite status. This update mechanism includes trigger condition determination, incremental sample construction, fine-tuning of some parameters, and version management.

[0057] In the specific implementation, the data-driven model update is triggered when any of the following conditions are met: ①Accuracy degradation, ,in, The preset threshold; ②Trend drift, This indicates that the deviation shows a significant monotonically increasing or decreasing trend; ③ Major events: Upon receiving an external event flag, such as performing a track maneuver, the system automatically triggers an update.

[0058] Upon triggering, an incremental training dataset is constructed using M valid samples within a sliding window. Incremental learning is completed by freezing some network parameters and performing a small number of iterations with a low learning rate. After the update is complete, the new model replaces the currently active version, and the old version is retained as a backup. If the prediction error of the new model increases abnormally, it will automatically roll back.

[0059] The digital satellite modeling method provided in this invention introduces an online performance monitoring and update mechanism based on a sliding window. By calculating the root mean square error and the slope of the deviation trend, the real-time accuracy and degradation trend of the model are quantified, solving the technical problem that static models cannot adapt to long-term performance degradation and environmental changes in satellites in orbit. This technical solution achieves the effect of enabling digital satellite models to have online adaptive capabilities and continuously maintain high-precision prediction results.

[0060] Based on any of the above embodiments, updating the data-driven model based on the root mean square error and the slope of the deviation trend includes: Based on the root mean square error and the slope of the deviation trend, determine whether the preset triggering condition is met; When the root mean square error is greater than a preset accuracy threshold, or the absolute value of the deviation trend slope is greater than a preset drift threshold, or when an external event flag is received, the model update process is triggered to dynamically update the parameters of the data-driven model.

[0061] It should be noted that the triggering conditions refer to pre-defined numerical judgment rules, including accuracy thresholds, drift thresholds, and external event flags. The accuracy threshold is a pre-defined upper limit of the root mean square error, used to determine whether the model accuracy has degraded to an unacceptable level; the drift threshold is a pre-defined upper limit of the absolute value of the deviation trend slope, used to determine whether the deviation exhibits a significant, systematic, monotonic change trend; and the external event flag is a Boolean signal generated by external system inputs or internal detection modules, used to identify significant events that may lead to model prediction failure (such as orbital maneuvers, attitude mode switching, onboard software upgrades, or drastic changes in the space environment).

[0062] The digital satellite modeling method provided in this invention further defines the specific judgment conditions and event types for triggering model updates. By introducing multiple threshold judgments and external event flags as trigger sources, it solves the problem of accurately and timely determining the timing of model updates under different scenarios (slow accuracy degradation, continuous trend drift, sudden major events). This technical solution achieves the effect of making the model update strategy more robust, sensitive, and comprehensive.

[0063] Based on any of the above embodiments, the dynamic updating of the parameters of the data-driven model includes: The real-time deviations corresponding to each time point within the sliding time window are used as valid samples. Based on the aforementioned valid samples, an incremental training dataset is constructed; The data-driven model is incrementally learned using the incremental training dataset, and the parameters of the data-driven model are dynamically updated.

[0064] It should be noted that a sliding time window refers to a fixed time interval of length M, which slides over time and always includes data from the most recent M moments. Valid samples refer to training data pairs extracted from this time window for subsequent incremental learning. Real-time bias refers to the difference between the final predicted satellite state and the actual on-orbit telemetry data, i.e.: This bias directly reflects the overall prediction error of the dual-drive model at the current moment and is a core indicator for measuring model performance degradation and guiding model updates.

[0065] In the specific implementation, by using real-time deviations within the sliding time window as effective samples to construct an incremental training dataset and performing incremental learning, efficient online updates of data-driven model parameters are achieved. Specifically, when the model update process is triggered, the data corresponding to the most recent M time points are first extracted from the current sliding time window. Each data point contains the working condition feature vector at that moment. and real-time deviation The data is then subjected to quality checks (such as removing missing or obvious outliers), and those that pass the checks are labeled as valid samples. These valid samples are then organized into an incremental training dataset. The dataset is much smaller than the initial training set, but it is more timely and can reflect the latest operational characteristics of the satellite. Finally, the data-driven model is incrementally learned using this incremental training dataset. The specific strategies include: freezing the parameters of the first few layers of the model to retain the learned general features, updating only the parameters of the last few layers to adapt to the new bias patterns, setting a learning rate much lower than the initial training rate (such as 1 / 10 of the initial learning rate), and performing a small number of iterative training cycles (such as 10 to 20 rounds) on the incremental dataset to avoid overfitting and catastrophic forgetting.

[0066] The digital satellite modeling method provided in this invention further defines the specific execution method of model updates, namely, incremental learning based on effective samples from a sliding window. By utilizing small batches of data for rapid fine-tuning based on a small learning rate and a finite number of iterations, the problem of how to efficiently and quickly update the model online to adapt to real-time changes in satellites while ensuring model stability and avoiding catastrophic forgetting is solved. This technical solution achieves the effect of maintaining the long-term prediction accuracy of the model with extremely low computational and storage costs.

[0067] The above embodiments of the present invention will be described below using a satellite orbit subsystem as an example: 1. Construct a high-precision orbit extrapolation mechanism model, and determine the satellite orbit prediction results based on this model: This step is based on Newton's law of universal gravitation and perturbation theory to construct a high-precision orbit extrapolation mechanism model, and obtains the satellite's orbit prediction results through numerical integration.

[0068] (1) Equations of satellite motion: Let the position vector of the satellite's center of mass be... The velocity vector is Then the equation of motion of the satellite in the inertial frame is: in, The distance from the Earth's center. The total perturbation acceleration term includes Earth's non-spherical gravity, atmospheric drag, third-body gravity (Sun and Moon gravity), solar radiation pressure, etc.

[0069] (2) Gravitational perturbations due to Earth's non-spherical shape: Expand using terms J2 to J4, where the formula for calculating term J2 is: in, The gravitational constant of Earth, This represents the Earth's second-order zone harmonic coefficient. The average radius of the Earth's equator. The distance from the Earth's center. Position vector Z-axis component, is the unit vector of the Z-axis in the geocentric inertial coordinate system.

[0070] (3) Atmospheric drag perturbation: in, The drag coefficient, Let be the projected area of ​​the satellite in the direction of relative velocity. For satellite quality, The local atmospheric density, Let V be the satellite's velocity vector relative to the atmosphere. .

[0071] (4) Gravitational perturbation of the third body (gravity of the sun and moon) in, The gravitational constant of the moon, The gravitational constant of the Sun, Let be the position vector of the Moon in the Earth-centered inertial frame. Let be the position vector of the Sun in the Earth-centered inertial frame.

[0072] (5) Solar radiation pressure perturbation: in, This is the shaded function, with a value range of [0,1]. The radiation pressure reflectance is the coefficient of reflection. For the effective projected area, For satellite quality, The solar radiation pressure constant, The unit is the astronomical unit. The distance between the Earth and the Sun. This is the unit vector pointing from the satellite to the sun.

[0073] (6) Numerical integration and orbit prediction: By integrating the above differential equation using a numerical integrator, the future time can be obtained. Mechanism for predicting orbital state .

[0074] This result serves as the basis for subsequent bias compensation in data-driven models.

[0075] 2. Construct a bias training sample set based on satellite on-orbit telemetry data: This step utilizes real telemetry data transmitted by the satellite during its on-orbit operation to analyze the deviation between the output of the computer model and the actual state, and extracts current operating condition features to construct a sample set for training the data-driven model.

[0076] (1) Obtain the true value of the orbital state: The true value of the orbital state from the received satellite on-orbit telemetry data is denoted as: .

[0077] Among them, t i For the i-th telemetry sampling time, r true and v true These are the actual position vector and the velocity vector, respectively.

[0078] (2) Output and deviation vector of the computer model: Call the orbit extrapolation mechanism model to calculate t at the same time as the telemetry data. i Mechanism model output And define the deviation vector as: .

[0079] (3) Extracting operating condition characteristics: Simultaneously, the current operating condition characteristics are extracted as input variables for the data-driven model, including: Satellite attitude quaternions ; Sun's angle of incidence ; Geomagnetic activity index ; orbital altitude ; Timestamp .

[0080] (4) Construct the training sample set: The above operating condition characteristics are combined with the deviation vector to construct a training sample set: in, The input vector is composed of the above-mentioned operating condition characteristics. For the corresponding output labels.

[0081] 3. Build a data-driven model and train it using the training sample set: This step builds a data-driven model to learn the mapping relationship between operating condition characteristics and deviations in the mechanism model.

[0082] (1) Constructing a data-driven model structure: Construct a time series neural network structure as a data-driven model Its goal is to learn the mapping relationship. .

[0083] (2) Training data-driven model: Set the loss function to Mean Squared Error (MSE). Use an optimizer (such as Adam or SGD) to train the model parameters θ until the loss function converges, and obtain a trained data-driven model.

[0084] 4. Integrate the mechanistic model and the data-driven model for final prediction: This step integrates the basic predictions from the mechanistic model with the deviations from the data-driven model to obtain the final prediction of the satellite's state.

[0085] (1) Obtain the basic forecast value and the deviation forecast value At any predicted time t k First, the basic predicted value is output by the orbit extrapolation mechanism model. Simultaneously extract current operating condition features This data is then input into a pre-trained data-driven model to obtain the bias prediction value: (2) The final predicted value is obtained by fusion. The deviation prediction value is compensated into the basic prediction value of the mechanism model to obtain the final orbital state prediction value: .

[0086] The fusion result retains the physical constraints of the mechanistic model while incorporating the self-learning and correction capabilities of the data-driven model.

[0087] 5. Online updates to the data-driven model: This step involves real-time in-orbit telemetry data acquisition and dynamic updates of data-driven model parameters using a small-sample incremental learning approach, enabling the digital satellite model to adapt to satellite performance degradation and environmental changes.

[0088] (1) Calculate the prediction deviation and monitoring indicators Whenever new telemetry data arrives, calculate the deviation between the final predicted value and the actual value for the most recent M time points: Calculate the root mean square error (RMS) of the deviation: And the slope of the deviation trend was fitted using linear regression: .

[0089] (2) Determine whether a model update is triggered. The model update process is triggered when any of the following conditions are met: ①Accuracy degradation, ,in, The preset threshold; ②Trend drift, This indicates that the deviation shows a significant monotonically increasing or decreasing trend; ③ Major events: Upon receiving an external event flag, such as performing a track maneuver, the system will automatically trigger an update.

[0090] (3) Perform incremental learning and updating Once the model update process is triggered, the latest The following valid samples constitute the incremental training dataset: Update the data-driven model using an incremental learning strategy: Freeze some network parameters (such as the first few layers) and only update the last few layers; Use a small learning rate (e.g., 1 / 10 of the initial learning rate); Perform a small number of iterative training sessions (e.g., 10-20 rounds).

[0091] (4) Model version management and rollback mechanism After the update is complete, save the new model parameters. Use the currently active version, while retaining the previous version. As a backup, if the new model causes an abnormal increase in prediction error on the validation set, it will automatically roll back to the old version and record the failure log for manual analysis.

[0092] The digital satellite modeling apparatus provided by this invention is described below. The digital satellite modeling apparatus described below can be referred to in correspondence with the digital satellite modeling method described above. For example... Figure 5 As shown, the digital satellite modeling device includes: Module 10 is used to build a mechanistic model of the satellite; The acquisition module 20 is used to acquire on-orbit telemetry data measured during the satellite's on-orbit operation, and to construct a data-driven model based on the on-orbit telemetry data. The prediction module 30 is used to predict the basic satellite state prediction value at a future time through the mechanism model, and to predict the deviation compensation value at the future time through the data-driven model. The correction module 40 corrects the basic satellite state prediction value based on the deviation compensation value to obtain the final satellite state prediction value.

[0093] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640, wherein the processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a digital satellite modeling method, which includes: constructing a mechanistic model of the satellite; collecting on-orbit telemetry data measured during the satellite's on-orbit operation and constructing a data-driven model based on the on-orbit telemetry data; predicting a basic satellite state prediction value for future times using the mechanistic model and predicting a deviation compensation value for the future times using the data-driven model; and correcting the basic satellite state prediction value based on the deviation compensation value to obtain a final satellite state prediction value.

[0094] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0095] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the digital satellite modeling method provided by the above methods. The method includes: constructing a mechanistic model of the satellite; collecting on-orbit telemetry data measured during the satellite's on-orbit operation and constructing a data-driven model based on the on-orbit telemetry data; predicting a basic satellite state prediction value at a future time through the mechanistic model and predicting a deviation compensation value at the future time through the data-driven model; and correcting the basic satellite state prediction value based on the deviation compensation value to obtain a final satellite state prediction value.

[0096] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the digital satellite modeling method provided by the above methods. The method includes: constructing a mechanistic model of the satellite; collecting on-orbit telemetry data measured during the satellite's on-orbit operation and constructing a data-driven model based on the on-orbit telemetry data; predicting a basic satellite state prediction value at a future time using the mechanistic model and predicting a deviation compensation value at the future time using the data-driven model; and correcting the basic satellite state prediction value based on the deviation compensation value to obtain a final satellite state prediction value.

[0097] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0098] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A digital satellite modeling method, characterized in that, include: Construct a mechanistic model of the satellite; Collect on-orbit telemetry data measured during the satellite's on-orbit operation, and construct a data-driven model based on the on-orbit telemetry data; The underlying satellite state prediction value for future times is predicted using the aforementioned mechanism model, and the deviation compensation value for future times is predicted using the aforementioned data-driven model. Based on the deviation compensation value, the basic satellite state prediction value is corrected to obtain the final satellite state prediction value.

2. The digital satellite modeling method according to claim 1, characterized in that, The data-driven model built based on the on-orbit telemetry data includes: Build the initial data-driven model; The true value of the satellite state at the target time is obtained from the on-orbit telemetry data, and the basic satellite state value at the same target time is determined through the mechanism model. Based on the true satellite state value and the basic satellite state value, determine the deviation vector; Based on the deviation vector, a training sample set is constructed, and the initial data-driven model is trained using the training sample set to obtain a trained data-driven model.

3. The digital satellite modeling method according to claim 2, characterized in that, The construction of the training sample set based on the deviation vector includes: Obtain the operating condition characteristics corresponding to the target time, wherein the operating condition characteristics include at least the satellite attitude quaternion, solar incidence angle, geomagnetic activity index, and orbital altitude; A training sample set is constructed based on the operating condition characteristics and the deviation vector.

4. The digital satellite modeling method according to claim 1, characterized in that, Also includes: Acquire real-time on-orbit telemetry data; The status data at multiple time points are obtained from the multiple real-time on-orbit telemetry data, and a sliding time window is constructed; Calculate the real-time deviation between the basic satellite state value output by the mechanism model at each time point within the sliding time window and the corresponding state data; Based on the real-time deviation, the root mean square error is calculated, and the slope of the deviation trend is fitted by linear regression. The data-driven model is updated based on the root mean square error and the slope of the deviation trend.

5. The digital satellite modeling method according to claim 4, characterized in that, The step of updating the data-driven model based on the root mean square error and the slope of the deviation trend includes: Based on the root mean square error and the slope of the deviation trend, determine whether the preset triggering condition is met; When the root mean square error is greater than a preset accuracy threshold, or the absolute value of the deviation trend slope is greater than a preset drift threshold, or when an external event flag is received, the model update process is triggered to dynamically update the parameters of the data-driven model.

6. The digital satellite modeling method according to claim 5, characterized in that, The dynamic updating of the parameters of the data-driven model includes: The real-time deviations corresponding to each time point within the sliding time window are used as valid samples. Based on the aforementioned valid samples, an incremental training dataset is constructed; The data-driven model is incrementally learned using the incremental training dataset, and the parameters of the data-driven model are dynamically updated.

7. A digital satellite modeling device, characterized in that, include: Building blocks are used to construct mechanistic models of satellites; The acquisition module is used to acquire on-orbit telemetry data measured during the satellite's on-orbit operation and to build a data-driven model based on the on-orbit telemetry data. The prediction module is used to predict the basic satellite state prediction value at a future time through the mechanistic model, and to predict the deviation compensation value at the future time through the data-driven model. The correction module corrects the basic satellite state prediction value based on the deviation compensation value to obtain the final satellite state prediction value.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the digital satellite modeling method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the digital satellite modeling method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the digital satellite modeling method as described in any one of claims 1 to 6.

Images