A spacecraft on-orbit model construction method based on digital twinning

Abstract

The application provides a spacecraft on-orbit model construction method based on digital twinning, comprising the following steps: step one, a meta-model configuration module configures a meta-model according to task information, configures a meta-model telemetry matrix and a correlation matrix by using on-orbit telemetry data, and then outputs the configured matrix to a model calculation module; step two, the model calculation module eliminates conflicts according to a model node conflict elimination rule, and eliminates conflicts in the case that a same telemetry parameter point is used multiple times in the model matrix; step three, model correction is performed according to the meta-model matrix after conflict elimination; step four, a BP neural network is used to compare the corrected digital model with true on-orbit data to find errors, and the errors are reversely transmitted to the BP neural network for iteration, network parameters are adjusted, calculation is performed again, and the cycle is repeated until the output result reaches a model setting target.

Description

Technical Field

[0001] This invention relates to a method for constructing an on-orbit model of a spacecraft based on digital twins, belonging to the field of spacecraft engineering technology. Background Technology

[0002] Digital twins have been extensively researched and applied throughout the entire product lifecycle, from design and manufacturing to operation and maintenance. In terms of domain applications, digital twins are widely used in the aerospace field. With the significant increase in the number of satellites in orbit, research on digital model construction technology based on digital twins has become a hot topic in the forefront of simulation technology development both domestically and internationally.

[0003] Currently, spacecraft development has incorporated the concept of digital models, using them as knowledge carriers for transmission and sharing to support subsequent production, testing, and on-orbit processes. Existing models are primarily based on mechanistic modeling, using mathematical and physical formulas for description. However, due to the significant differences between the space environment and ground conditions, the on-orbit characteristics of various components exhibit unknowns, nonlinearities, and time-varying properties. These models are increasingly failing to adequately reflect the physical characteristics of the space environment in the spacecraft's on-orbit state and cannot maintain timely synchronization with the actual operating status throughout the spacecraft's entire on-orbit lifecycle.

[0004] Therefore, there is an urgent need for a dynamic on-orbit model construction method for spacecraft. This method would address the complexity of spacecraft, which involve multiple domains and subsystems, and achieve multi-domain model integration based on a meta-model. Simultaneously, it would utilize on-orbit data to dynamically correct the digital model, thereby improving its fidelity and enabling the digital twin system to continuously approximate the actual satellite system. Summary of the Invention

[0005] This invention proposes a method for constructing an on-orbit model of a spacecraft based on digital twins. By utilizing on-orbit data, a digital model of the satellite digital twin system is constructed in a data-driven manner, taking into account the complex object characteristics, mission features, and operating environment of the satellite, thereby continuously improving the approximation between the digital twin system and the actual satellite system.

[0006] The present invention is achieved through the following technical solution.

[0007] A method for constructing an on-orbit model of a spacecraft based on digital twins includes the following steps:

[0008] Step 1: The meta-model configuration module configures the meta-model based on the mission information, uses on-orbit telemetry data to configure the meta-model telemetry matrix and correlation matrix, and then outputs the configured matrix to the model calculation module.

[0009] Step 2: The model calculation module performs conflict elimination for cases where the same telemetry parameter point in the model matrix is ​​used multiple times, according to the model node conflict elimination rules.

[0010] Step 3: Based on the meta-model matrix after conflict elimination, perform model correction, compare the actual on-orbit telemetry data and the digital model data, and if the error between the two does not meet the predetermined target value, then perform correction calculations on the original digital model.

[0011] Step 4: Use a BP neural network to compare the corrected digital model with the actual on-orbit data to find the error, and then backpropagate the error to the BP neural network for iteration. Adjust the network parameters, recalculate, and repeat this process until the output result reaches the model's target.

[0012] The beneficial effects of this invention are:

[0013] 1. This invention achieves universality in the construction of spacecraft simulation models through the configuration of meta-models, thereby improving development and usage efficiency;

[0014] 2. This invention effectively solves the problem of low accuracy of satellite digital models by using a model correction method based on on-orbit data, and improves the realism of the system model through intelligent iterative learning;

[0015] 3. This invention integrates the model and assembles and constructs it in a modular way, like assembling building blocks. After constructing the meta-model, the entire satellite system simulation model can be divided into three levels, from high to low: meta-model, component model and system model. The high-level model defines the basic functional elements required for the description of the low-level model, while the low-level model is an instantiation of the high-level model based on language elements.

[0016] 4. This invention adopts a combination of Docker and microservices, which has significant advantages in independent deployment, operation and maintenance management, and rapid delivery;

[0017] 5. During the model integration process, it is easy for the same model to be integrated multiple times or for different models to perform the same sub-model functions multiple times. This invention limits the complexity of the model tree structure by setting an upper limit on the number of models in the model tree.

[0018] 6. This invention obtains a large amount of on-orbit telemetry data through sensors, and then uses real data to correct the digital twin system digital model, so as to promote the continuous evolution of the digital twin as the satellite is actually in orbit, gradually approaching the real physical state of the satellite product and maintaining its high fidelity. Attached Figure Description

[0019] Figure 1This is a flowchart of the spacecraft on-orbit model construction method based on digital twins according to the present invention;

[0020] Figure 2 This is a schematic diagram of the meta-model configuration of the present invention;

[0021] Figure 3 This is a schematic diagram of the meta-model structure of the present invention. Detailed Implementation

[0022] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the drawings are merely exemplary and are intended to illustrate the principles and spirit of the present invention, and are not intended to limit the scope of the present invention.

[0023] like Figure 1 As shown, the present invention provides a method for constructing an on-orbit model of a spacecraft based on digital twins, which specifically includes the following steps:

[0024] Step 1: The meta-model configuration module configures the meta-model based on the mission information, uses on-orbit telemetry data to configure the meta-model telemetry matrix and correlation matrix, and then outputs the configured matrix to the model calculation module.

[0025] like Figure 2 As shown in this embodiment, the on-orbit telemetry data includes changes in telemetry parameters controlled by remote control commands, changes in telemetry parameters controlled by component functions, and changes in telemetry parameters controlled by the environment.

[0026] like Figure 3 As shown in this embodiment, the configuration of the correlation matrix using on-orbit telemetry data is specifically carried out in the following way: a cube model is used to represent the correlation features and functions between telemetry parameters, the nodes in the meta-model constitute the related telemetry of the model, the node construction adopts a multi-mode model, and the relationship between nodes is constructed using the edges of the cube.

[0027] The cubic meta-model described in this embodiment is a quantifiable N-dimensional telemetry space, consisting of N-dimensional parameter coordinates with dimension identifiers and vectors or matrices within the corresponding dimensions. The vectors, matrices, and vectors and matrices are directly or indirectly associated and calculated through directed transformation relationships, adjoint relationships, and symbiotic relationships. The correlation relationships between various telemetry points in the model are characterized through relational calculations.

[0028] For example, Figure 3The example shown is a 3×3×3 meta-model, a simulation matrix consisting of 27 telemetry points and 6 sections. Each section represents a functional attribute of the meta-model, including data management attributes, energy attributes, thermal control attributes, structural attributes, and fault attributes. The entire meta-model describes the basic simulation function of a component. Structures 1, 2, and 3 are strongly correlated functional telemetry points, each consisting of 3 telemetry points (i.e., correlated 1×3 vectors), forming a 3×3 planar matrix. Structures 1, 4, and 5 are also strongly correlated functional telemetry points, each consisting of 3 telemetry points (i.e., correlated 1×3 vectors), forming a 3×3 planar matrix. Structures 2, 3, 4, and 5 are weakly correlated, and their relationship attributes need to be connected and calculated through Structure 1, which represents the data communication attribute in the meta-model.

[0029] Step 2: The model calculation module performs conflict elimination for cases where the same telemetry parameter point in the model matrix is ​​used multiple times, according to the model node conflict elimination rules.

[0030] In this embodiment, the correlation is represented and calculated by correlation rules, that is, according to the bidirectional positive correlation between the parameter coordinates of the meta-model adjoint matrix and the grid combination relationship between satellite telemetry, that is, under the condition of N-dimensional coordinate parameters, the correlation degree value between the N-dimensional coordinate parameters and the K-dimensional coordinate parameters is used as the matrix element value, and the edge between the nodes represents the relationship between the telemetry of the two nodes.

[0031] (1) If telemetry nodes TM1 and TM2 are functionally related, then the boundary value between the connection is set to 1, and each attribute interface can be represented by a relation matrix composed of 0s and 1s; for each telemetry node TM in the telemetry set TM i On the set TM, there exists a mapping as:

[0032] f TM :TM→[0,1],tm→f TM (tm), where

[0033] (2) The current changes in telemetry parameters are determined through a grid combination algorithm between models. The grid is divided into three sets: remote control command control (C), component function control (G), and environmental control (H). The meta-model telemetry change matrix is ​​E = C × G × H, where...

[0034]

[0035] Step 3: Based on the meta-model matrix after conflict elimination, perform model correction, compare the actual on-orbit telemetry data and the digital model data, and if the error between the two does not meet the predetermined target value, then perform correction calculations on the original digital model.

[0036] In this embodiment, the correction calculation uses the error norm as the objective function, transforming the correction problem into a single-objective nonlinear constrained programming problem; the specific steps are as follows:

[0037] 3.1 Set constraint functions;

[0038] Let f(p) be the objective function, g j (p) represents the constraint condition, p i For the design parameters that need to be corrected, p i1 and p i2 Let ||TM| be the lower and upper limits of the design parameters, respectively. Then the objective function is ||TM|. L -TM S ||;

[0039] Wherein, the superscript L represents the theoretical value of the telemetry parameter, the superscript S represents the actual on-orbit value of the telemetry parameter, TM is the selected telemetry matrix of the satellite digital model to be corrected, and the constraint condition is that the satellite's index Q remains unchanged before and after the model correction, i.e., ∑ΔQ i If = 0, then the constraint function is:

[0040] minf(p)

[0041]

[0042] 3.2 Determine the upper and lower limits of the model parameters;

[0043] The model parameters p that need to be corrected i The upper and lower limits of the change are set according to the actual on-orbit empirical values ​​of the original variable or the on-orbit operational requirements; in this embodiment, 0.6p is selected. i ≤p i ≤1.4p i The selected modified objective function values ​​include the telemetry time-domain eigenvalue vector, the telemetry correlation matrix, and their weighted linear combination; the final objective function is obtained using the Lagrange multiplier method:

[0044]

[0045] Where G is the telemetry correlation vector of the satellite model to be corrected, and TM j Let ω be the j-th order telemetry time series feature vector of the model, Q be the model revision constraint index, and ω be the weighting coefficient.

[0046] 3.3 Iterate through the telemetry parameters and make corrections;

[0047] During model correction, the upper and lower limits of the correlation degree of the telemetry correlation matrix G are selected as thresholds in the one-dimensional search. If the threshold is exceeded, the correction is considered unreasonable, and the parameter is not corrected. The initial correction direction is to correct only one telemetry parameter TM in the telemetry matrix each time.j Each round of correction modifies the telemetry parameters of the analog quantity based on time-series changes.

[0048] In this embodiment, the search for each telemetry parameter correction is performed only once. That is, after determining the two endpoints and the initial value, the objective function value is calculated to obtain three points. The quadratic trinomial is determined using the Lagrange interpolation function, and the minimum point of the quadratic trinomial is determined, which is also the final minimum point. In specific correction, the value of the weighting coefficient is guaranteed to make the error of this term at least one order of magnitude higher than the sum of the other terms.

[0049] Step 4: Use a backpropagation (BP) neural network to compare the corrected digital model with the actual on-orbit data to find the error. Then, backpropagate the error to the BP neural network for iteration, adjust the network parameters, and recalculate. This process is repeated until the output result meets the model's target. Specific steps include:

[0050] 4.1 Starting from the second layer of the BP neural network, all weights and biases of the layers that take the second layer as input are set to 0. At this point, the weights are a zero matrix and the biases are a zero vector; where x i Y is the input parameter. i For the output result, θ i For the output result of the layer above the hidden layer, when The connection weights between the j-th neuron in layer L-1 and the i-th neuron in layer L are... Let i be the bias of the i-th neuron in the L-th layer. The learning function of the neural network is:

[0051]

[0052] in, This is the input to the i-th neuron in layer L. is the activation function of the neuron;

[0053] 4.2 Calculate the gradient matrix Δω and vector Δβ of the weights and biases of each layer of neurons using the backpropagation method; update the weights and biases according to the iterative method: s is the total number of layers whose weights and biases need to be updated.

[0054]

[0055]

[0056] 4.3 In practical implementation, establishing a model correction network based on a BP neural network requires a large amount of satellite on-orbit telemetry data for training. Therefore, in this embodiment, during automatic iteration, the number of network layers is set to 3. The three-layer network consists of an input node x, a hidden layer θ, and a final output layer Y. The weight between the input layer and the hidden layer is ω, and the weight between the hidden layer and the output layer is γ. The iterative calculation method is as follows:

[0057]

[0058]

[0059] The error calculation is based on According to the error calculation formula, the error is a function of the weights ω and γ. When the error is minimized, the weights of each layer of the network are established, and the simulation model is corrected.

[0060] After completing the above model construction, the present invention further integrates and encapsulates the constructed model:

[0061] The principle of model integration is as follows: based on the modeling and correction of the meta-model, modular assembly and construction are carried out in accordance with the integration method, like piecing together building blocks. After the meta-model is constructed, the entire satellite system simulation model can be divided into three levels, from high to low: meta-model, component model and system model. The high-level model defines the basic functional elements required for the description of the low-level model, while the low-level model is an instantiation of the high-level model based on language elements.

[0062] The integration includes top-level model integration and domain model integration. In this embodiment, the top-level model refers to the coupling mechanism of system models with different properties (such as discrete models, continuous models, qualitative models, quantitative models, etc.); the domain model refers to the coupling mechanism of heterogeneous models modeled by different meta-model modeling mechanisms, and the coupling mechanism is added to each granularity model. Wherein:

[0063] The top-level model integration is the integration of the top-level static structural model and dynamic behavior model, used to verify the top-level behavior and function of the spacecraft; the domain model integration is the integration of heterogeneous models of different subsystems, used to verify the joint simulation of two or more subsystems or submodules. When performing the simulation task, different models running concurrently at the same time T resolve conflicts and integrate the models through independent processes.

[0064] In practice, model integration can easily lead to situations where the same model is integrated multiple times or different models repeatedly perform the same sub-model functions. This results in sub-model nodes with identical functions in the model tree, creating loops and causing cyclical conflicts during model execution. Consequently, the model tree becomes excessively large, requiring significant computational resources for model operation, making model execution difficult, and even preventing model integration from being completed. To address this technical problem, this embodiment adopts the following integration method:

[0065] Set an upper limit on the number of models in the model tree to limit the complexity of the model tree structure. Retain child model nodes that are far from the root node of the model tree and delete other nodes built with the same model or functionally similar child models and their subsequent nodes. This ensures that child models closer to the leaf nodes should be run first. The output model is integrated into the intermediate node and stored in the database. At the same time, the running priority of this model is defined to be higher than that of the child models corresponding to the parent nodes of all deleted nodes.

[0066] In this embodiment, the model integration rules are as follows: child node models in the model set are run and calculated first; among child nodes of the same layer, independent node child models have higher priority than reference node child models, and among reference node child models, independent constraint unit child models have higher priority than reference constraint unit child models; the parent node model of the deleted model has lower running priority than the retained identical or functionally similar child model nodes; the model corresponding to each node is processed from bottom to top in the model set to achieve model integration and solution.

[0067] Model encapsulation refers to dividing simulation functions into granular parts according to the needs of simulation business scenarios, and then encapsulating the decomposed functions in Docker containers to build microservices.

[0068] In this embodiment, the microservice interface adopts the REST protocol and parallel simulation data format. According to the microservice interface calling standard, all service request and response data are encoded in the simulation system control information format.

[0069] In this embodiment, the microservice interface call error information is encapsulated in the simulation container log. The date format of all requests and responses is yyyy-MM-dd HH:mm:ss. After the microservice is encapsulated, it is registered to the service discovery component cluster to provide support for system business call control. The encapsulated microservice is deployed on the server in the form of a jar package.

[0070] In practice, after the services are started, they register their configuration information with the deployed service cluster. All requests are processed by the service scheduling and control module. The development environment, code, and configuration files of the microservice application are packaged into containers for deployment and isolation. The entire functionality is encapsulated in Docker. The image built from the application and Dockerfile generates a new image and is stored in an image repository. To run a microservice instance on another machine, simply pull the image from the image repository and run it. This combination of Docker and microservices offers significant advantages in independent deployment, operation and maintenance management, and rapid delivery.

[0071] Example 1:

[0072] The following analysis examines the model construction process of a satellite simulation system's power subsystem, using it as an example. The power subsystem simulation model comprises a solar array simulation model, a battery simulation model, a power controller simulation model, and a load simulation model. Each component model is composed of meta-models. First, meta-models are created. Then, the models are iteratively revised based on on-orbit telemetry data. Finally, the different meta-models are integrated to form the power subsystem simulation model.

[0073] 1) First, configure the initial parameters of the solar cell array model, including I. SC For short-circuit current, V OC I is the open-circuit voltage. mp For the maximum operating current, V mp The maximum operating voltage is given by Pm, the maximum power is given by Pm, the output current of the TM1 solar array is given by Pm, the output voltage of the TM2 solar array is given by Pm, and the output power of the TM3 solar array is given by Pm. The short-circuit current I is used as the reference. sc Open circuit voltage V OC Maximum power point voltage Vmp, maximum power point current I mp Four characteristic parameters can be used to determine the correlation matrix G1 between the output current TM1 and voltage TM2 of the solar cell:

[0074] 2) Construct the correlation matrix G2 of the solar cell array TM3 based on the output voltage and current telemetry. The correlation matrix G2 of the solar cell array's output power can be obtained from the relationships between the output voltage and current of the solar cell array under different light intensities and temperatures:

[0075] P = I × s × V × F1 × F2 × X S ×X T ×Fm×η×cosθ

[0076] 3) Based on the determined meta-model matrix, complete the model establishment of the entire power subsystem according to the meta-model establishment process. Decompose the meta-model relation matrix into matrix sets G1, G2, etc., corresponding to the multi-dimensional spatiotemporal matrix. Formally, the number of rows and columns is completely consistent with the original telemetry parameter element matrix, that is, the correspondence is completed in the mapping of correlation relationships;

[0077] 4) The battery charging process consists of a first-stage high-current charging and a second-stage trickle charging. The charging process begins after receiving the power controller's operating status information (charging). Let the battery capacity at the discharge termination point be Q0. During the first-stage charging process, the battery charging current I... 充 If the current is 6.6A and the charging time is t, then the charging capacity is Q. 充 =6.6*t, the battery capacity is Q = Q0 + Q 充The battery voltage at different capacities is obtained from the battery parameter table. In the two-stage trickle charging stage, the trickle charging capacity mainly compensates for the self-discharge capacity, and the trickle charging capacity is basically constant with the self-discharge capacity. Therefore, during the trickle charging stage, the battery pressure and capacity are considered to remain unchanged.

[0078] 5) Select telemetry data of the charging process after satellite emergence; set the charging mode conversion model to be corrected as: I 充 =p1×t 4 +p2×t 3 +p3×t 2 +p4×t+C, where p1, p2, p3, and p4 are the model parameters to be corrected, and the parameters are dynamically corrected based on the read on-orbit telemetry data. The constraint for model correction is that the battery capacity of the power subsystem remains unchanged, i.e., ∑ΔQ i =0. Simultaneously, the correction parameter range is set to 0.6p. i ≤p i ≤1.4p i Charging current I 充 It is related to the battery voltage, battery capacity, and battery temperature. Let I be... 充 Relationship matrix According to formulas 1 and 2, I 充 Make corrections, where ω λ =0.3, ω TM =0.6, ω Q =0.1;

[0079] 6) Based on the set correction conditions, the model parameters are iteratively corrected using a dynamic correction method. This paper selects a three-layer BP neural network, with the sigmoid function chosen as the automatic learning function f, and the minimum error E... 最小 =0.1. Formula 3 is used for system error calculation. After four iterations of learning, the charging current I of the power supply subsystem is... 充 The errors between the trained neural network and the actual charging current values ​​for each correction model are shown in Table 1, based on the 210 sample telemetry data.

[0080] Table 1 Comparison of parameter values ​​before and after correction

[0081] Number of iterations <![CDATA[p1]]> <![CDATA[p2]]> <![CDATA[p3]]> <![CDATA[p4]]> error 1 0.0151 0.0201 -0.155 1.2881 5.42％ 2 -0.0073 0.0154 -0.101 0.2443 2.77％ 3 -0.0025 0.0092 -0.081 -0.3481 1.02％ 4 -0.0002 0.0054 -0.038 -0.6579 0.55％

[0082] Based on the dynamic simulation model construction method proposed in this invention, a preliminary version of digital model construction software for a satellite power subsystem was developed using the C# platform. The dynamic model correction theory and software were initially verified using a simulation engineering example of the charging current model of a satellite power subsystem.

[0083] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for constructing an on-orbit model of a spacecraft based on digital twins, characterized in that, Includes the following steps: Step 1: The meta-model configuration module configures the meta-model based on the mission information, uses on-orbit telemetry data to configure the meta-model telemetry matrix and correlation matrix, and then outputs the configured matrix to the model calculation module. Step 2: The model calculation module performs conflict elimination on the case where the same telemetry parameter point in the meta-model telemetry matrix is ​​used multiple times, according to the model node conflict elimination rules. Step 3: Based on the telemetry matrix of the meta-model after conflict elimination, perform model correction, compare the actual on-orbit telemetry data and the digital model data, and if the error between the two does not meet the predetermined target value, then perform correction calculations on the original digital model. Step 4: Use a BP neural network to compare the corrected digital model with the actual on-orbit data to find the error, and then backpropagate the error to the BP neural network for iteration. Adjust the network parameters, recalculate, and repeat this process until the output result reaches the model's target.

2. The method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 1, characterized in that, The on-orbit telemetry data includes changes in telemetry parameters controlled by remote control commands, changes in telemetry parameters controlled by component functions, and changes in telemetry parameters controlled by the environment.

3. A method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 1 or 2, characterized in that, The correlation matrix is ​​configured using the on-orbit telemetry data, specifically in the following manner: a cubic meta-model is used to represent the correlation characteristics and functions between telemetry parameters. The nodes in the cubic meta-model constitute the related telemetry of the model. The node construction adopts a multi-mode model, and the relationship between nodes is constructed using the edges of the cube.

4. The method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 3, characterized in that, The cubic meta-model is quantifiable and computable. dimensional telemetry space, consisting of dimensions identified by... The model consists of dimensional parameter coordinates and corresponding vectors or matrices. The vectors, matrices, and vectors and matrices are directly or indirectly associated and calculated through directed transformation relationships, adjoint relationships, and symbiotic relationships. The relationship calculation represents the correlation between various telemetry points in the model.

5. The method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 3, characterized in that, The correlation relationships are represented and calculated through correlation rules, namely, according to the bidirectional positive correlation between parameter coordinates of the meta-model adjoint matrices and the grid combination relationship of the correlation between satellite telemetry, that is, in the ... Under the condition of dimensional coordinate parameters, the first Dimensional coordinate parameters and the first The correlation values ​​between the coordinate parameters are used as matrix element values, and the edges between nodes represent the relationship between the telemetry of two nodes.

6. A method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 1 or 2, characterized in that, The correction calculation uses the error norm as the objective function, transforming the correction problem into a single-objective nonlinear constrained programming problem.

7. A method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 1 or 2, characterized in that, Step four specifically includes: 4.1 Starting from the second layer of the BP neural network, all weights and biases of layers that take the second layer as input are set to 0. At this point, the weights are a zero matrix and the biases are a zero vector; where For input parameters, For the output results, For the output result of the layer above the hidden layer, when The connection weights between the j-th neuron in layer L-1 and the i-th neuron in layer L are... Let i be the bias of the i-th neuron in the L-th layer. The learning function of the neural network is: in, This is the input to the i-th neuron in layer L. is the activation function of the neuron; 4.2 Calculate the gradient matrix of weights and biases of neurons in each layer using the backpropagation method. sum vector Update the weights and biases using an iterative method: It is the total number of layers that need to have their weights and biases updated. 4.3 Set the network to 3 layers, with the input node in the three-layer network. Hidden layer and the final output layer The weights between the input layer and the hidden layer are The weights between the hidden layer and the output layer are The iterative calculation method is as follows: The error calculation is based on According to the error calculation formula, the error is a weight. and The function is used to establish the weights of each layer of the network when the error is minimized, thus completing the correction of the simulation model.

8. The method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 7, characterized in that, Following step four, further model integration is performed, including top-level model integration and domain model integration.

9. The method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 8, characterized in that, The top-level model integration is the integration of the top-level static structural model and dynamic behavior model, used to verify the top-level behavior and function of the spacecraft; the domain model integration is the integration of heterogeneous models of different subsystems, used to verify the joint simulation of two or more subsystems or submodules. During the joint simulation, different models running concurrently at the same time T resolve conflicts and integrate the models through independent processes.

10. The method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 9, characterized in that, The integration process employs the following approach: setting an upper limit on the number of models in the model tree to limit the complexity of the model tree structure; retaining child model nodes that are far from the root node of the model tree; and deleting other nodes built with the same model or functionally similar child models and their subsequent nodes. This ensures that child models closer to the leaf nodes are run first. The output model integration intermediate nodes are stored in the database, and the running priority of this model is defined to be higher than that of the child models corresponding to the parent nodes of all deleted nodes.

11. A method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 8, 9, or 10, characterized in that, The rules for model integration are as follows: child node models in the model set are run and calculated first; among child nodes of the same layer, independent node sub-models have higher priority than reference node sub-models, and among reference node sub-models, independent constraint element sub-models have higher priority than reference constraint element sub-models; the parent node model of a deleted model has lower running priority than the retained identical or functionally similar child model nodes; the model corresponding to each node is processed from bottom to top in the model set to achieve model integration and solution.

12. The method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 8, characterized in that, After step four, the model is further encapsulated. The model encapsulation is based on the simulation business scenario requirements, which divides the simulation functions into granular parts and encapsulates the split functions in Docker containers to build microservices.

13. The method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 12, characterized in that, The microservice interface adopts the REST protocol and parallel simulation data format. According to the microservice interface calling standard, all service request and response data are encoded in the simulation system control information format.

14. A method for constructing an on-orbit model of a spacecraft based on digital twins as described in claim 12 or 13, characterized in that, Microservice API call error messages are encapsulated in the simulation container logs. The date format for all requests and responses is yyyy-MM-dd HH:mm:ss. After encapsulation, the microservice is registered with the service discovery component cluster to provide support for system business call control. The encapsulated microservice is deployed on the server as a JAR package.

Images