Digital Twin Testing Methods and Related Devices for Large Load-Bearing Structures

Abstract

This invention discloses a digital twin testing method and related apparatus for large load-bearing structures. It utilizes limited measurement point data to map the stress response of the entire structure, improves the model's predictive ability under complex loading conditions through a multi-task transfer learning mechanism, and proposes a visualization method to render the model's calculation results. This solves the problems of insufficient accuracy and incompleteness in predicting structural states under complex conditions by traditional methods, and significantly improves the usability of digital twin technology.

Description

Technical Field

[0001] This invention relates to the field of digital twin testing of large-scale structural static tests, specifically to a digital twin testing method and related apparatus for large-scale load-bearing structures. Background Technology

[0002] The liquid rocket engine frame is the main load-bearing component of the engine, responsible for transmitting thrust. A failure in the rocket engine frame will directly cause the rocket propulsion system to fail, leading to mission failure, significant economic losses, and even serious casualties. Therefore, static tests must be conducted on the large frame structure before a flight mission to ensure its successful completion.

[0003] However, conducting static tests using rocket engine racks is accompanied by high costs and risks, and it is difficult to obtain actual measurement data of the engine racks. These problems greatly limit the successful development of health status prediction technology for large structures.

[0004] Traditional static testing monitoring methods assess the health status of a structure by deploying strain sensors at key structural locations, simulating actual working conditions, and analyzing the collected test data. For large load-bearing structural components, this method faces two main problems: (1) The structure is large in size, requiring a large number of sensors to accurately monitor its health status, which significantly reduces the efficiency of the delivery process. (2) The structure is complex in size, making it difficult to deploy sensors or resulting in insufficient measurement accuracy on curved surfaces and in confined spaces, which seriously affects the reliability of monitoring.

[0005] Currently, the use of neural network technology to establish digital twin models of structures for structural health monitoring is widely studied. This involves mapping the stress response across the entire field by inputting sensor signals. However, such methods only consider establishing this mapping and fail to consider how to effectively optimize the overall predictive performance of the model based on measured signals in practical applications. Summary of the Invention

[0006] The purpose of this invention is to provide a digital twin testing method and related apparatus for large load-bearing structures to overcome the shortcomings of existing technologies. The transfer learning method based on the multi-task learning mechanism of this invention can simultaneously extract truly effective information from a large amount of simulation information and a small amount of measured signals, thereby achieving high accuracy and high resolution prediction of structural state response. Furthermore, it maps the full-field stress response of the structure through a small number of measurement point signals. This has important guiding significance for the rapid verification of large load-bearing structures with the same structure and the accurate identification of structural anomalies.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] The digital twin testing method for large load-bearing structures includes the following steps:

[0009] Step 1: Perform finite element numerical simulation based on the working conditions in the static test of the rocket engine frame. After obtaining the simulation results of the frame structure under different working conditions, extract the full-field stress response of the frame structure simulation data to obtain the simulation dataset. At the same time, establish the measured dataset collected in the static test of the frame. Construct the simulation measured point dataset based on the simulation dataset and the measured dataset.

[0010] Step 2: Train the autoencoder using the full-field stress response simulation data from Step 1. Then, extract the decoder part from the autoencoder to use as a generator. Next, use the simulation and measured point dataset to train a simple connector from the measured point dimension to the autoencoder latent space. Combine the generator and the simple connector to obtain a preliminary digital twin model for stress state prediction.

[0011] Step 3: Use the measured dataset from Step 1 to input, adjust, and validate the preliminary stress state prediction digital twin model to obtain the corrected digital twin model;

[0012] Step 4: Design a method for visualizing the full-field stress of the model, update the predicted stress in real time using cloud map rendering, deploy the corrected digital twin model, infer the full-field response of the corrected digital twin model based on the measured dataset, and visualize the results;

[0013] Step 5: Replace the measured dataset input to the corrected digital twin model with the strain data of the measuring points collected in real time during the static test. Then, through a multi-task transfer learning mechanism, perform a self-iterative optimization process of the model, continuously inputting the strain data of the measuring points into the corrected digital twin model, and iteratively updating the accuracy of the corrected digital twin model in real time.

[0014] Furthermore, the full-field stress response of the frame structure simulation data in step 1 is expressed as: The measured dataset collected from the static test of the frame is represented as follows: ,in and These represent the simulation data and measured data for each working condition, respectively, and n represents the total number of test samples; and , Where N represents a node, and the superscripts s and e represent simulation and experiment, respectively. This represents the number of nodes in the simulation results. This represents the number of measurement points of the structural sensor. .

[0015] Furthermore, the construction of the simulation and measured point datasets based on the simulation dataset and the measured dataset specifically involves:

[0016] From the simulation dataset, measurement points whose stress variation trends are consistent with those of the measured sensor points are obtained to construct a simulation-measured point dataset. .

[0017] Furthermore, the finite element numerical simulation described in step 1 specifically includes:

[0018] Based on the load magnitude, application location, and boundary conditions during the test, a finite element model of the frame structure is designed in the finite element analysis software, and the mesh size is set according to the twin requirements; the calculation job is then submitted in the finite element analysis software.

[0019] After the calculation is completed, the analysis results under each loading condition are saved as RPT files, the finite element model calculation results are saved as INP files, and then the node information is extracted using Python to obtain the simulation results of the frame structure under different working conditions. The Mises stress information extracted from the simulation results is saved as a deep learning network training format.

[0020] When extracting Mises stress information, the measurement point matching is also completed. Specifically, the simulation model nodes that match the actual sensor measurement points in the actual experiment are found in the simulation model. The nodes on the simulation model are selected from three aspects: spatial location, stress change trend and numerical magnitude, to ensure that the measurement points are correctly matched. This establishes a simulation dataset, an actual measurement dataset, and a real measurement point dataset corresponding to the actual measurement points in the simulation dataset.

[0021] Furthermore, step 2 specifically includes:

[0022] Train an autoencoder using a simulation dataset;

[0023] The decoder of the autoencoder is used as the generator;

[0024] The hidden space data of the training autoencoder part is saved. By inputting the simulated measured point dataset, the hidden space data is output to obtain a simple mapper from the low-dimensional input signal to the hidden space. Finally, the simple mapper and the generator are assembled to obtain a preliminary digital twin model for predicting the stress state from the measured point data to the full-field stress response.

[0025] Furthermore, the method for visualizing the full-field stress of the design model is specifically as follows:

[0026] The system uses Python based on the TCP / IP protocol to receive binary information streams acquired in real time and decodes these binary information streams to obtain real-time strain data.

[0027] The strain data is input into the modified digital twin model to obtain the full-field stress prediction results of the rack structure. In the visualization part, the .vtk files of the rack structure are displayed in batches in the form of Unstructured Grid, and stress values ​​are assigned to each node.

[0028] Digital twins are achieved by generating cloud map displays using custom color bars and updating the cloud map after each reading of test data.

[0029] Furthermore, the model's self-iterative optimization process specifically includes:

[0030] Each measured sensor signal is divided into three categories: input signal, adjustment signal, and verification signal. First, the full-field prediction of the finite element simulation model and the full-field prediction of the digital twin correction model are compared to construct loss function 1. At the same time, the adjustment signal, which is used to adjust the training model, is compared to construct loss function 2. Finally, the input signal in the model prediction result is compared with the collected input signal to construct loss function 3. After assigning weights to the three loss functions, they are summed to obtain the comprehensive loss function. The comprehensive loss function is used to optimize the correction digital twin model. Finally, the optimization result of the correction digital twin model is verified based on the verification signal.

[0031] A large-scale load-bearing structure digital twin testing system includes:

[0032] Dataset creation module: It is used to perform finite element numerical simulation based on the working conditions in the static test of rocket engine frame, obtain the simulation results of the frame structure under different working conditions, extract the full field stress response of the frame structure simulation data, and obtain the simulation dataset. At the same time, it establishes the measured dataset collected in the static test of the frame, and constructs the simulation measured point dataset based on the simulation dataset and the measured dataset.

[0033] Model building module: Used to train an autoencoder with full-field stress response simulation data. Then, the decoder part of the autoencoder is extracted to form a generator. The simulated measured point dataset is then used to train a simple connector from the measured point dimension to the autoencoder latent space. The generator and the simple connector are combined to obtain a preliminary digital twin model for stress state prediction.

[0034] Model correction module: Used to input, adjust and verify the preliminary stress state prediction digital twin model using measured datasets to obtain a corrected digital twin model;

[0035] Visualization module: Used to design a method for visualizing the full-field stress of the model, update the predicted stress in real time by rendering cloud maps, deploy the corrected digital twin model, infer the full-field response of the corrected digital twin model based on the measured dataset, and visualize the results;

[0036] The self-iterative optimization module is used to replace the measured dataset input to the corrected digital twin model with the strain data of the measuring points collected in real time during the static test. Then, through a multi-task transfer learning mechanism, the model performs a self-iterative optimization process, continuously inputting the strain data of the measuring points into the corrected digital twin model and iteratively updating the accuracy of the corrected digital twin model in real time.

[0037] A computer device includes 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 steps of the digital twin testing method for large load-bearing structures.

[0038] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the digital twin testing method for large load-bearing structures.

[0039] Compared with the prior art, the present invention has the following beneficial technical effects:

[0040] 1) Based on the deep learning multi-task transfer learning mechanism, this invention proposes to use simulation datasets and measured datasets to jointly optimize the model, achieving high-accuracy and high-dimensional global stress prediction, and solving the phenomenon that the overall prediction accuracy decreases instead of increases during the optimization process of general digital twin models.

[0041] 2) The full-field stress state prediction digital twin model constructed in this invention can invert the full-domain stress response of the structure through finite measurement points, which solves the problem of the curse of node dimension in general digital twin models. At the same time, it solves the problems of low verification efficiency and insufficient accuracy in the early stage of engineering delivery, and reduces the difficulty of establishing a model that learns low-dimensional mapping to high-dimensional modeling.

[0042] 3) This invention proposes a method for visualizing twin results. It extracts grid information from the structural model INP file and aligns it with the prediction results of the spliced ​​model. Finally, it visualizes the results through Pyvista to obtain a digital twin model that can respond in real time. Attached Figure Description

[0043] The accompanying drawings are provided to further understand the invention and constitute a part of this invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0044] Figure 1 This is a flowchart of the method of the present invention;

[0045] Figure 2 This is a schematic diagram illustrating the principle of measurement point matching in the method of this invention;

[0046] Figure 3 This is a diagram illustrating the predicted effects of the present invention;

[0047] Figure 4 The results are the predictions of the twin model, where (a) is the prediction result of the component structure at the 1st loading level, (b) is the prediction result of the component structure at the 7th loading level, (c) is the prediction result of the component structure at the 17th loading level, and (d) is the prediction result of the component structure at the 24th loading level.

[0048] Figure 5 A comparison chart showing the predicted input signal results of the twin model and the measured sensor signals;

[0049] Figure 6 The prediction results of the measured input signal input to the preliminary component-level twin model at three measurement points are shown in the figure.

[0050] Figure 7 The figures show a comparison of the prediction results at the measurement points for the two adjustment methods, where (a) is the traditional transfer learning method and (b) is the adjustment method using a multi-task learning mechanism. Detailed Implementation

[0051] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0052] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0053] Example 1

[0054] A digital twin testing method for large load-bearing structures includes the following steps:

[0055] Step 1: In the offline phase, finite element numerical simulation is performed based on the working conditions in the static test of the rocket engine frame. After obtaining the simulation results of the frame structure under different working conditions, the full-field stress response of the frame structure simulation data is extracted. Simulation datasets were obtained. Simultaneously, measured datasets collected from the static load test of the frame were established. .in, and These represent simulation data and experimental data for each operating condition, respectively. , ,in This represents the number of nodes in the simulation results. This represents the number of measurement points of the structure sensor, generally speaking. Here, n represents the total number of samples in the experiment, N represents the number of nodes, and the superscripts s and e represent simulation and experiment, respectively. Subsequently, it is necessary to complete the measurement point matching, obtaining measurement points from the simulation dataset that match the stress change trends of the measured sensor points, thus constructing a simulation-measured point dataset. It can be seen that the number of working condition nodes and the number of actual measurement points in each simulation and measurement point dataset are consistent.

[0056] Finite element analysis specifically includes: establishing a finite element analysis model based on the geometric dimensions of the research object; defining complete assembly and constraint relationships between components; determining the load magnitude, application location, and boundary conditions during the experiment; designing the finite element model of the structure in finite element analysis software such as ABAQUS or ANSYS; setting the mesh size according to twin requirements; submitting the calculation job in the finite element analysis software; saving the analysis results under each loading condition as an RPT file after the calculation is completed; saving the finite element model information as an INP file; then using Python to extract node information to obtain simulation results of the frame structure under different working conditions; extracting Mises stress information from the simulation results and saving it as a deep learning network training format; when extracting Mises stress information, it is necessary to complete the measurement point matching; finding the measurement points in the simulation model and selecting them from three aspects: spatial location, stress change trend, and numerical magnitude to ensure correct matching of measurement points; and establishing a simulation dataset, a measured dataset, and a dataset of real measurement points corresponding to the measured points in the simulation dataset.

[0057] Step 2: Construct a digital twin model for stress state prediction based on an autoencoder neural network. Specifically, an autoencoder is trained using the full-field stress response simulation data from Step 1. Then, the decoder part of the autoencoder is extracted to form a generator. A simple connector (Simple Mapper) from the dimension of the measured points to the latent space of the autoencoder is trained using the simulation and measured point dataset. The generator and the simple connector are combined to complete the initial construction of the digital twin model for stress state prediction.

[0058] First, an autoencoder is trained using the full-field stress response of the input simulation data. The autoencoder automatically compresses high-dimensional data into a low-dimensional space, then maps the low-dimensional space back to the input. The decoder part of this autoencoder, trained using the full-field stress response of the input simulation data, becomes the generator we need. The data type consists of one-dimensional data sequences. Since the generator's input dimension differs from the measurement point dimension, a simple mapping space from the measurement point dimension to the input dimension needs to be established. We save the hidden space data from the training of the autoencoder part. By inputting the measurement point data and outputting the hidden space data, we obtain such a simple connector. Finally, assembling the generator and the simple connector yields a preliminary digital twin model for stress state prediction, mapping the measurement point data to the full-field stress response. The training process for this preliminary digital twin model can use each part of the frame structure as a basic unit, or the entire model can be directly input.

[0059] The aforementioned stress-corrected full-field stress state digital twin prediction model integrates multiple neural network models to establish a stress prediction tool with high robustness and generalization ability. First, to establish a neural network model that maps low-dimensional stress to high-dimensional stress, a generative model is prioritized, employing an autoencoder architecture to automatically construct a mapping model from the measurement point dimension to the prediction dimension. It should be noted that since the number of nodes in the predicted stress field is often in the tens of thousands, while the number of measured points is relatively small, the latent representation dimension of the autoencoder cannot directly match the measurement point dimension. Therefore, convolutional layers with better feature extraction are used in the network layers, and the convolutional kernel parameters are used to process different parts of the input data, reducing the number of model parameters and lowering the model complexity. Simultaneously, a simple mapper is trained in the mapping from the model's measurement point dimension to the latent space representation to fully extract and represent complex features in the data. Each layer in the encoder part uses a downsampling operation to achieve sufficient dimensionality reduction of the data. In the decoder part, deconvolutional layers are used to implement the model's generation function. The specific model parameters are as follows:

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072]

[0073] in, Represents a convolutional layer LeakyReLU and Softplus are activation functions. This represents the deconvolution operation, where y is the final output. Indicates a linear layer. It is a max pooling layer operation. This represents the output of each layer. A simple mapper model is as follows:

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080] in, Indicates the input dimension. Indicates the output dimension.

[0081] Step 3: Improve the accuracy of the initial stress state prediction digital twin model based on the measured dataset to enhance the overall stress state prediction accuracy. Specifically, the measured dataset from Step 1 is used to input, adjust, and validate the initial stress state prediction digital twin model to obtain the corrected digital twin model. During the correction process, the measured dataset is also divided into three types of signals: input signal, adjustment signal, and validation signal. The input signal is input into the model to obtain the overall response. The adjustment signal is then used to improve the model's prediction accuracy, and the validation signal is used to verify the adjustment effect. If the optimization effect is insufficient, the model is readjusted until it reaches the optimal state.

[0082] Step 4: Design a method for visualizing the full-field stress of the model, update the predicted stress using cloud map rendering, deploy the digital twin model, infer and correct the full-field response of the digital twin model based on the measured dataset, and visualize the results.

[0083] The system uses Python based on the TCP / IP protocol to receive and decode binary information streams acquired in real time to obtain real-time strain data. This strain data is then input into a modified digital twin model to obtain the full-field stress prediction results for the rack structure. In the visualization section, the .inp files obtained from the finite element model in step 1 are processed using Python to extract the component mesh node coordinates and element information. Multiple elements are reconstructed into triangular meshes and saved as .vtk files. The .vtk files of the rack structure are displayed in batches as an Unstructured Grid, with each node assigned the stress prediction results obtained in step 3. A custom color bar is used to generate a contour map display effect, which is updated after each test data reading, thus achieving digital twin functionality.

[0084] The results are stitched together to form a global stress prediction result. After aligning the result with the model nodes, the model is rendered to obtain the global stress display result.

[0085] Step 5: During the landing test, the measured dataset input to the corrected digital twin model is replaced with the strain data of the measuring points collected in real time during the static test. Then, the model self-iterative optimization process is carried out through a multi-task transfer learning mechanism, and the strain data of the measuring points is continuously input into the corrected digital twin model to update the model accuracy in real time.

[0086] The self-iterative optimization process of the model adaptively assigns weights to simulation information and measured point information through a constructed loss function, maximizing the preservation of the effective value of the two different information sources and making the model's prediction results more accurate. Each measured sensor signal is divided into three categories: input signal, adjustment signal, and verification signal. First, the full-field prediction of the finite element simulation model and the full-field prediction of the digital twin correction model are compared to construct loss function 1. Simultaneously, the adjustment signal, which is used to adjust the training model, is compared to construct loss function 2. Finally, the input signal in the model prediction result is compared with the collected input signal to construct loss function 3. After assigning weights to the three loss functions, they are summed to obtain the comprehensive loss function. The comprehensive loss function is used to optimize the correction digital twin model. Finally, the optimization results of the correction digital twin model are verified based on the verification signal.

[0087] This invention utilizes simulation and measured data, with simulation data as the foundation. Through a data fusion algorithm, it combines the advantages of the broad dimensionality of simulation data and the high precision of measured data to establish a digital twin model. This improves prediction accuracy, ensures the structural health of the structure, facilitates task execution, and enhances project delivery efficiency. Furthermore, this digital twin model can continuously optimize and update its prediction accuracy using simulation data. To avoid negative transfer, a multi-task transfer learning mechanism is employed, ensuring effective optimization of the overall prediction accuracy during model adjustments. Additionally, a full-field state visualization scheme for the model is proposed during the testing phase, improving the model's real-time display performance.

[0088] Example 2

[0089] To better illustrate the technical effects of this invention, experiments were conducted at different levels using a specific example to verify the invention. This embodiment is based on a digital twin experiment conducted on a static test of a certain type of liquid rocket engine frame. The loading conditions of this static test are shown in the table below:

[0090]

[0091] To verify the accuracy of the proposed digital twin technology in inferring the full-field response of a structure based on limited measured data, a component-level digital twin model was first established. Based on the above loading conditions, finite element simulation was performed in Abaqus to obtain the ODB result file and output the RPT file for each loading level. Python was then used to extract data from a specified component and save it as a txt file. The number of lines in each file represents the number of nodes in that component, and the data type of each line is as follows:

[0092]

[0093] Simultaneously, this component has three measurement sensors, namely S1, SS1, and S2. Next, measurement point matching needs to be completed. In the simulation model, the node IDs corresponding to these three sensors are determined to be 4816, 2699, and 9387. Data corresponding to these IDs is extracted from a TXT file using Python and saved as a TXT file. This experiment mainly adopts the neural network architecture of an autoencoder. In model correction, the negative transfer phenomenon that occurs in traditional transfer learning is demonstrated. After the model training is completed, the real-time response of the structure can be calculated by inputting the model with the actual measured dataset and then visualized. In the visualization part, by assigning index values ​​to the VTK file of each component and expressing and rendering it using the Pyvista library, a digital twin model that can reflect the real-time state of the structure is obtained. The specific model parameters are as follows:

[0094]

[0095] The preliminary twin model prediction results are as follows: Figure 4 As shown. Figure 4 The results show the effectiveness of the proposed model in fitting limited metadata, with a fitting error of less than 0.02%.

[0096] We continue to construct a digital twin model that maps a small number of sensors to the global signal. This invention uses sensor S1 as the input signal, sensor S2 as the fine-tuning signal, and sensor S3 as the verification signal. First, we compare the differences between the model predictions and the measured signals, such as... Figure 5 and Figure 6 As can be seen, the model has high prediction accuracy at the input measurement points. However, the prediction accuracy at the other two measurement points is unsatisfactory.

[0097] Simple migration effect using SS1 sensor as shown Figure 7 As shown in (a), Figure 7 (b) This method was used. Specific errors are shown in the table below. It can be seen that when using the SS1 sensor for transfer learning, the error of S2 increases instead of decreasing, indicating that model optimization did not have a positive impact on the overall model when using new sensor information. However, this method can effectively acquire useful knowledge from the new sensor, reducing the error of the SS1 sensor while increasing the prediction accuracy of S2.

[0098]

[0099] Meanwhile, based on the error table below, this invention also employed a complete rack for testing and achieved good results. The rack contains 72 sensors, which are divided into three signal groups. It is evident that this method still effectively improves the global prediction accuracy of the twin model.

[0100]

[0101] Example 3

[0102] This invention provides a digital twin testing system for large-scale load-bearing structures, comprising:

[0103] Dataset creation module: It is used to perform finite element numerical simulation based on the working conditions in the static test of rocket engine frame, obtain the simulation results of the frame structure under different working conditions, extract the full field stress response of the frame structure simulation data, and obtain the simulation dataset. At the same time, it establishes the measured dataset collected in the static test of the frame, and constructs the simulation measured point dataset based on the simulation dataset and the measured dataset.

[0104] Model building module: Used to train an autoencoder with full-field stress response simulation data. Then, the decoder part of the autoencoder is extracted to form a generator. The simulated measured point dataset is then used to train a simple connector from the measured point dimension to the autoencoder latent space. The generator and the simple connector are combined to obtain a preliminary digital twin model for stress state prediction.

[0105] Model correction module: Used to input, adjust and verify the preliminary stress state prediction digital twin model using measured datasets to obtain a corrected digital twin model;

[0106] Visualization module: Used to design a method for visualizing the full-field stress of the model, update the predicted stress in real time by rendering cloud maps, deploy the corrected digital twin model, infer the full-field response of the corrected digital twin model based on the measured dataset, and visualize the results;

[0107] The self-iterative optimization module is used to replace the measured dataset input to the corrected digital twin model with the strain data of the measuring points collected in real time during the static test. Then, through a multi-task transfer learning mechanism, the model performs a self-iterative optimization process, continuously inputting the strain data of the measuring points into the corrected digital twin model and iteratively updating the accuracy of the corrected digital twin model in real time.

[0108] Example 4

[0109] The present invention provides a computer 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 steps of the digital twin testing method for large load-bearing structures.

[0110] Example 5

[0111] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the digital twin testing method for large load-bearing structures.

[0112] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0113] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0114] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0115] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0116] 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 its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A digital twin testing method for large load-bearing structures, characterized in that, Includes the following steps: Step 1: Perform finite element numerical simulation based on the working conditions in the static test of the rocket engine frame. After obtaining the simulation results of the frame structure under different working conditions, extract the full-field stress response of the frame structure simulation data to obtain the simulation dataset. At the same time, establish the measured dataset collected in the static test of the frame. Construct the simulation measured point dataset based on the simulation dataset and the measured dataset. Step 2: Train the autoencoder using the full-field stress response simulation data from Step 1. Then, extract the decoder part from the autoencoder to use as a generator. Next, use the simulation and measured point dataset to train a simple connector from the measured point dimension to the autoencoder latent space. Combine the generator and the simple connector to obtain a preliminary digital twin model for stress state prediction. Step 3: Use the measured dataset from Step 1 to input, adjust, and validate the preliminary stress state prediction digital twin model to obtain the corrected digital twin model; Step 4: Design a method for visualizing the full-field stress of the model, update the predicted stress in real time using cloud map rendering, deploy the corrected digital twin model, infer the full-field response of the corrected digital twin model based on the measured dataset, and visualize the results; Step 5: Replace the measured dataset input to the corrected digital twin model with the strain data of the measuring points collected in real time during the static test. Then, through a multi-task transfer learning mechanism, perform a self-iterative optimization process of the model, continuously inputting the strain data of the measuring points into the corrected digital twin model, and iteratively updating the accuracy of the corrected digital twin model in real time.

2. The digital twin testing method for large load-bearing structures according to claim 1, characterized in that, The full-field stress response of the frame structure simulation data in step 1 is expressed as follows: The measured dataset collected from the static test of the frame is represented as follows: ,in and These represent the simulation data and measured data for each working condition, respectively, and n represents the total number of test samples; and , Where N represents a node, and the superscripts s and e represent simulation and experiment, respectively. This represents the number of nodes in the simulation results. This represents the number of measurement points of the structural sensor. .

3. The digital twin testing method for large load-bearing structures according to claim 2, characterized in that, The construction of the simulation and measured point datasets based on the simulation dataset and the measured dataset specifically involves: From the simulation dataset, measurement points whose stress variation trends are consistent with those of the measured sensor points are obtained to construct a simulation-measured point dataset. .

4. The digital twin testing method for large load-bearing structures according to claim 1, characterized in that, The finite element numerical simulation described in step 1 specifically refers to: Based on the load magnitude, application location, and boundary conditions during the test, a finite element model of the frame structure is designed in the finite element analysis software, and the mesh size is set according to the twin requirements; Submit the calculation job in the finite element analysis software; After the calculation is completed, the analysis results under each loading condition are saved as RPT files, the finite element model calculation results are saved as INP files, and then the node information is extracted using Python to obtain the simulation results of the frame structure under different working conditions. The Mises stress information extracted from the simulation results is saved as a deep learning network training format. When extracting Mises stress information, the measurement point matching is also completed. Specifically, the simulation model nodes that match the actual sensor measurement points in the actual experiment are found in the simulation model. The nodes on the simulation model are selected from three aspects: spatial location, stress change trend and numerical magnitude, to ensure that the measurement points are correctly matched. This establishes a simulation dataset, an actual measurement dataset, and a real measurement point dataset corresponding to the actual measurement points in the simulation dataset.

5. The digital twin testing method for large load-bearing structures according to claim 1, characterized in that, Step 2 specifically involves: Train an autoencoder using a simulation dataset; The decoder of the autoencoder is used as the generator; The hidden space data of the training autoencoder part is saved. By inputting the simulated measured point dataset, the hidden space data is output to obtain a simple mapper from the low-dimensional input signal to the hidden space. Finally, the simple mapper and the generator are assembled to obtain a preliminary digital twin model for predicting the stress state from the measured point data to the full-field stress response.

6. The digital twin testing method for large load-bearing structures according to claim 1, characterized in that, The method for visualizing the full-field stress of the design model is as follows: The system uses Python based on the TCP / IP protocol to receive binary information streams acquired in real time and decodes these binary information streams to obtain real-time strain data. The strain data is input into the modified digital twin model to obtain the full-field stress prediction results of the rack structure. In the visualization part, the .vtk files of the rack structure are displayed in batches in the form of Unstructured Grid, and stress values ​​are assigned to each node. Digital twins are achieved by generating cloud map displays using custom color bars and updating the cloud map after each reading of test data.

7. The digital twin testing method for large load-bearing structures according to claim 1, characterized in that, The model's self-iterative optimization process is specifically as follows: Each measured sensor signal is divided into three categories: input signal, adjustment signal, and verification signal. First, the full-field prediction of the finite element simulation model and the full-field prediction of the digital twin correction model are compared to construct loss function 1. At the same time, the adjustment signal, which is used to adjust the training model, is compared to construct loss function 2. Finally, the input signal in the model prediction result is compared with the collected input signal to construct loss function 3. After assigning weights to the three loss functions, they are summed to obtain the comprehensive loss function. The comprehensive loss function is used to optimize the correction digital twin model. Finally, the optimization result of the correction digital twin model is verified based on the verification signal.

8. A digital twin testing system for large-scale load-bearing structures, characterized in that, include: Dataset creation module: It is used to perform finite element numerical simulation based on the working conditions in the static test of rocket engine frame, obtain the simulation results of the frame structure under different working conditions, extract the full field stress response of the frame structure simulation data, and obtain the simulation dataset. At the same time, it establishes the measured dataset collected in the static test of the frame, and constructs the simulation measured point dataset based on the simulation dataset and the measured dataset. Model building module: Used to train an autoencoder with full-field stress response simulation data. Then, the decoder part of the autoencoder is extracted to form a generator. The simulated measured point dataset is then used to train a simple connector from the measured point dimension to the autoencoder latent space. The generator and the simple connector are combined to obtain a preliminary digital twin model for stress state prediction. Model correction module: Used to input, adjust and verify the preliminary stress state prediction digital twin model using measured datasets to obtain a corrected digital twin model; Visualization module: Used to design a method for visualizing the full-field stress of the model, update the predicted stress in real time by rendering cloud maps, deploy the corrected digital twin model, infer the full-field response of the corrected digital twin model based on the measured dataset, and visualize the results; The self-iterative optimization module is used to replace the measured dataset input to the corrected digital twin model with the strain data of the measuring points collected in real time during the static test. Then, through a multi-task transfer learning mechanism, the model performs a self-iterative optimization process, continuously inputting the strain data of the measuring points into the corrected digital twin model and iteratively updating the accuracy of the corrected digital twin model in real time.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the digital twin testing method for large load-bearing structures as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the digital twin testing method for large load-bearing structures as described in any one of claims 1 to 7.

Images