Structure thermal deformation inversion method based on deep learning
By using a deep learning-based approach to decompose the structure and build a neural network model, the problems of large computational load and long time consumption of traditional methods are solved, and the rapid and accurate inversion of satellite thermal deformation in orbit is realized.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- AEROSPACE DONGFANGHONG SATELLITE
- Filing Date
- 2025-12-10
- Publication Date
- 2026-04-24
AI Technical Summary
Existing technologies cannot effectively monitor the thermal deformation of satellites in orbit. Traditional methods are computationally intensive and time-consuming, and cannot meet the real-time thermal deformation inversion requirements of complex structures.
A deep learning-based approach is adopted to decompose the structure into secondary structures, construct a neural network model, fit the displacement field mapping relationship through finite element simulation analysis and training samples, and train the neural network using a deep learning loss function to achieve rapid displacement field inversion.
It reduces computational costs while improving the accuracy of thermal deformation inversion of complex structures, enabling rapid and accurate inversion of satellite displacement fields.
Smart Images
Figure CN121920187A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of thermal deformation inversion technology, and more specifically to a method for thermal deformation inversion of complex structures based on deep learning technology. Background Technology
[0002] With the rapid development of space technology, the requirements for the dimensional stability of satellite structures are becoming increasingly stringent. However, on-orbit thermal deformation monitoring is constrained by the space environment and structural complexity, making it impossible to reliably monitor spacecraft thermal deformation in orbit around the clock. Therefore, in-orbit thermal deformation inversion for complex structures is essential.
[0003] Because the satellite structure, instruments and equipment and other heat source locations are all quite complex, the temperature field varies depending on the on-orbit operating conditions. The traditional method is to obtain the displacement field by simulating the temperature field obtained by calculation using the finite element method. This method has the disadvantages of large amount of calculation and long time consumption. Summary of the Invention
[0004] To address the problems existing in the prior art, this invention provides a method for inverting the thermal deformation of complex structures based on deep learning technology.
[0005] This application provides a deep learning-based method for structural thermal deformation inversion, including:
[0006] Step S1: Decompose the structure into multiple substructures and obtain a training sample set. The training samples include the heat source temperature, heat source location distribution and displacement field of the structure. Modal analysis results of each substructure are obtained through finite element simulation analysis.
[0007] Step S2: Based on the heat source temperature, heat source location distribution, and displacement field, construct a neural network model, and train the neural network using training samples to fit the mapping relationship between the modal analysis results of each secondary structure and the displacement field.
[0008] Step S3: Construct a deep learning loss function that can balance the loss differences among the various substructures;
[0009] Step S4: Train the neural network model using the training sample set and the deep learning loss function. Take the heat source location distribution in the training samples as input and the displacement field as output, and train the neural network model by minimizing the deep learning loss function.
[0010] Step S5: Perform displacement field inversion using the trained neural network model.
[0011] In one embodiment of the present invention, step S1 involves obtaining a training sample set using the following steps:
[0012] S11: Construct a finite element model of the structure, determine the temperature field of the structure, and calculate the displacement field of the structure under the temperature field through finite element simulation analysis to obtain a training sample including the distribution of heat source locations and displacement field.
[0013] S12: Repeat step S11 to obtain training samples until a preset number of training samples are obtained, thus obtaining a training sample set.
[0014] In one embodiment of the present invention, in step S11, the displacement field of the secondary structure is obtained by combining modal superposition and rigid body displacement.
[0015] In one embodiment of the present invention, in step S2, an LSTM model is used to construct a neural network model.
[0016] In one embodiment of the present invention, in step S3, the deep learning loss function that can balance the loss differences of each secondary structure is:
[0017]
[0018] in, This represents the output data loss corresponding to the i-th secondary structure. Indicates the weighting coefficient;
[0019] Data loss corresponding to the i-th secondary structure Calculate using the following formula:
[0020]
[0021] in, Indicates the number of training samples. Indicates the first The predicted value corresponding to each training sample Indicates the first The true value in each training sample.
[0022] In one embodiment of the present invention, step S4 includes:
[0023] S401: Input the heat source location distribution and heat source temperature from multiple training samples into the neural network model in sequence to obtain the corresponding displacement field output by the neural network model;
[0024] S402: Compare the displacement field output by the neural network model with the displacement fields of each secondary structure in the training samples, and calculate the prediction accuracy of the neural network model.
[0025] S403: Determine whether the prediction accuracy is greater than the preset accuracy threshold. If yes, use the current neural network model as the completed neural network model. If no, adjust the parameters of the neural network model using the deep learning loss function and return to step S401.
[0026] In one embodiment of the present invention, in step S401, the heat source location distribution in the training samples is input from the neural network input terminal, and the displacement field is obtained by optimizing the modal displacement superposition and rigid body displacement superposition coefficients.
[0027] In one embodiment of the present invention, in step S402, the difference between the predicted and actual displacement field values of each secondary structure is calculated, and the prediction accuracy of the neural network model is comprehensively evaluated.
[0028] In one embodiment of the present invention, in step S403, an accuracy threshold is set for each secondary structure.
[0029] In one embodiment of the present invention, the method is used to perform displacement field inversion on a satellite.
[0030] The complex structure thermal deformation inversion method based on deep learning technology of the present invention transforms the problem of obtaining the displacement field by temperature field simulation calculation into an optimization problem of minimizing the loss function by combining modal displacement superposition and rigid body displacement. The deep learning neural network is trained by minimizing the loss function. Based on the trained deep learning neural network, the displacement field of complex structures can be quickly inverted using only the temperature and location distribution data of the heat source, reducing the computational cost while achieving high inversion accuracy. Attached Figure Description
[0031] The following description, in conjunction with the accompanying drawings, will further illustrate the above-mentioned features, technical characteristics, advantages, and implementation methods of this application in a clear and understandable manner. The accompanying drawings are for illustrative and explanatory purposes only and do not limit the scope of this application. Wherein:
[0032] Figure 1 This is a flowchart of a method for thermal deformation inversion of complex structures based on deep learning technology, according to an embodiment of the present invention. Detailed Implementation
[0033] 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 in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. 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.
[0034] This invention provides a method for inverting the thermal deformation of complex structures based on deep learning technology, comprising:
[0035] Constructing finite element models of complex structures;
[0036] The temperature field data obtained through temperature field analysis is used to obtain the total displacement field of the complex structure through simulation analysis. The heat source temperature and the displacement field obtained through simulation analysis are used as training data for deep learning.
[0037] Decompose complex structures into several simple substructures;
[0038] The modal analysis results of the secondary structure were obtained through finite element simulation analysis.
[0039] A deep learning neural network is constructed, and the training data is used to train the deep learning neural network to fit the mapping relationship between the modal analysis results of each secondary structure and the displacement field.
[0040] The heat source temperature data is input into the trained deep learning neural network to obtain the displacement field of the complex structure.
[0041] The technical solutions provided by the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0042] refer to Figure 1 An embodiment of the present invention provides a method for inverting the thermal deformation of complex structures based on deep learning technology, the method comprising the following steps S1 to S5:
[0043] Step S1: Decompose the structure into multiple substructures and obtain a training sample set. The training samples include the heat source temperature, heat source location distribution and displacement field of the structure. Modal analysis results of each substructure are obtained through finite element simulation analysis.
[0044] In one embodiment of the present invention, for a satellite with a determined structural form and a determined distribution of instruments and equipment, the heat source temperature, the heat source location distribution, and the displacement field corresponding to its structure are determined as a training sample to obtain a sufficient number of sample sets.
[0045] In one embodiment of the present invention, the number of training samples is determined according to actual needs. Generally speaking, the more effective training samples there are, the higher the prediction accuracy of the trained neural network model.
[0046] In some possible implementations, the heat source locations are distributed as follows: , This represents the coordinates of the heat source in the X direction of the defined three-dimensional coordinate system. This represents the coordinates of the heat source in the Y direction of the defined three-dimensional coordinate system. This represents the coordinates of the heat source in the Z direction of the defined three-dimensional coordinate system.
[0047] In some possible implementations, thermal deformation is , This represents the thermal displacement component in the X direction of a defined three-dimensional coordinate system. This represents the thermal displacement component in the Y direction of the defined three-dimensional coordinate system. This represents the thermal displacement component in the Z direction of the defined three-dimensional coordinate system.
[0048] In some possible implementations, the training sample set is obtained in the following ways:
[0049] The temperature field is determined based on the satellite's on-orbit operating conditions and the distribution of instruments and equipment. The displacement field under this temperature field is calculated using the finite element simulation analysis method, resulting in a training sample that includes the distribution of heat source locations and the displacement field.
[0050] Repeat the process of obtaining training samples until a preset number of training samples are obtained, thus obtaining a training sample set.
[0051] In some possible implementations, the displacement field of the secondary structure is obtained by combining modal superposition and rigid body displacement:
[0052]
[0053] in, For thermal deformation field, Let i be the displacement field of the i-th mode of the secondary structure. These are the modal displacement field weighting coefficients. Let be the rigid body displacement field.
[0054] Step S2: Based on the heat source temperature, heat source location distribution, and displacement field, construct a neural network model, and train the neural network using training samples to fit the mapping relationship between the modal analysis results of each secondary structure and the displacement field.
[0055] In one embodiment of the present invention, an LSTM model is used to construct a neural network model.
[0056] Step S3: Construct a deep learning loss function that can balance the loss differences among the substructures.
[0057] In one embodiment of the present invention, the training objective of the neural network model is to minimize the difference between the predicted value and the true value.
[0058] Specifically, in one embodiment of the present invention, the deep learning loss function that can balance the loss differences of each secondary structure is:
[0059]
[0060] in, This represents the output data loss corresponding to the i-th secondary structure. This represents the weighting coefficient.
[0061] In some possible implementations, the data loss corresponding to the i-th secondary structure Calculate using the following formula:
[0062]
[0063] in, Indicates the number of training samples. Indicates the first The predicted value corresponding to each training sample Indicates the first The true value in each training sample.
[0064] Step S4: Train the neural network model using the training sample set and the deep learning loss function. The heat source coordinates in the training samples are used as input, and the normalized displacement fields of each secondary structure in the training samples are used as output. The neural network model is trained by minimizing the deep learning loss function, including the following steps S401~S403:
[0065] Step S401: Input the coordinates and temperatures of the heat sources in multiple training samples into the neural network model in sequence to obtain the corresponding displacement field output by the neural network model.
[0066] In one embodiment of the present invention, the heat source temperature and coordinates in the training samples are input from the input end of the neural network, and the displacement field is obtained by optimizing the modal displacement superposition and rigid body displacement superposition coefficients.
[0067] Step S402: Compare the predicted displacement field output by the neural network model with the displacement fields of each secondary structure in the training samples, and calculate the prediction accuracy of the neural network model.
[0068] In one embodiment of the present invention, the difference between the predicted and actual displacement field values of each secondary structure is calculated, and the prediction accuracy of the neural network model is comprehensively evaluated.
[0069] Step S403: Determine whether the prediction accuracy is greater than the preset accuracy threshold. If yes, use the current neural network model as the completed neural network model. If no, adjust the parameters of the neural network model using the loss function and return to step S401.
[0070] In one embodiment of the present invention, an accuracy threshold is set for each substructure. If the prediction accuracy obtained at least 5 times is greater than the preset accuracy threshold of the substructure, the current neural network model is used as the completed neural network model and the neural network model training ends; otherwise, the parameters of the neural network model are updated using the loss function and the neural network model training continues.
[0071] In some possible implementations, the method further includes:
[0072] Before training the neural network model using the training sample set, the displacement fields of each secondary structure are normalized.
[0073] Step S5: Perform displacement field inversion using the trained neural network model.
[0074] Specifically, when satellite displacement field inversion is required, the satellite's heat source temperature and position data are input into a trained neural network model to obtain the structural displacement field output by the neural network model.
[0075] An embodiment of the present invention provides a method for inverting the thermal deformation of complex structures based on deep learning technology. By combining modal displacement superposition and rigid body displacement, the problem of obtaining the displacement field from temperature field simulation is transformed into an optimization problem of minimizing the loss function. The deep learning neural network is trained by minimizing the loss function. Based on the trained deep learning neural network, the displacement field of complex structures can be quickly inverted using only the temperature and location distribution data of the heat source, reducing the computational cost while achieving high inversion accuracy.
[0076] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Additionally, the terms "front," "back," "left," "right," "upper," and "lower" in this document refer to the placement shown in the accompanying drawings.
[0077] 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 deep learning-based method for structural thermal deformation inversion, comprising: Step S1: Decompose the structure into multiple substructures and obtain a training sample set. The training samples include the heat source temperature, heat source location distribution and displacement field of the structure. Modal analysis results of each substructure are obtained through finite element simulation analysis. Step S2: Based on the heat source temperature, heat source location distribution, and displacement field, construct a neural network model, and train the neural network using training samples to fit the mapping relationship between the modal analysis results of each secondary structure and the displacement field; Step S3: Construct a deep learning loss function that can balance the loss differences among the various substructures; Step S4: Train a neural network model using the training sample set and a deep learning loss function. Take the heat source location distribution in the training samples as input and the displacement field as output, and train the neural network model by minimizing the deep learning loss function. Step S5: Perform displacement field inversion using the trained neural network model.
2. The method according to claim 1, wherein, In step S1, the training sample set is obtained using the following steps: S11: Construct a finite element model of the structure, determine the temperature field of the structure, and calculate the displacement field of the structure under the temperature field through finite element simulation analysis to obtain a training sample including the distribution of heat source locations and displacement field. S12: Repeat step S11 to obtain training samples until a preset number of training samples are obtained, thus obtaining a training sample set.
3. The method according to claim 2, wherein, In step S11, the displacement field of the secondary structure is obtained by combining modal superposition and rigid body displacement.
4. The method according to claim 1, wherein, In step S2, an LSTM model is used to construct a neural network model.
5. The method according to claim 1, wherein, In step S3, the deep learning loss function that can balance the loss differences among the secondary structures is: in, This represents the output data loss corresponding to the i-th secondary structure. Indicates the weighting coefficient; Data loss corresponding to the i-th secondary structure Calculate using the following formula: in, Indicates the number of training samples. Indicates the first The predicted value corresponding to each training sample Indicates the first The true value in each training sample.
6. The method according to claim 1, wherein, Step S4 includes: S401: Input the heat source location distribution and heat source temperature from multiple training samples into the neural network model in sequence to obtain the corresponding displacement field output by the neural network model; S402: Compare the displacement field output by the neural network model with the displacement fields of each secondary structure in the training samples, and calculate the prediction accuracy of the neural network model. S403: Determine whether the prediction accuracy is greater than the preset accuracy threshold. If yes, use the current neural network model as the completed neural network model. If no, adjust the parameters of the neural network model using the deep learning loss function and return to step S401.
7. The method according to claim 6, wherein, In step S401, the heat source location distribution in the training samples is input from the neural network input end, and the displacement field is obtained by optimizing the modal displacement superposition and rigid body displacement superposition coefficients.
8. The method according to claim 6, wherein, In step S402, the difference between the predicted and actual displacement field values of each secondary structure is calculated, and the prediction accuracy of the neural network model is comprehensively evaluated.
9. The method according to claim 6, wherein, In step S403, an accuracy threshold is set for each secondary structure.
10. The method according to claim 1, wherein, The method is used to perform displacement field inversion on satellites.