A Digital Twin Preload Preload Prediction Method for Sealing Joints Based on Multi-Fidelity Graph Neural Networks

Abstract

This invention discloses a digital twin preload prediction method for sealing joints based on a multi-fidelity graph neural network. The bolt system is modeled as a directed graph structure based on tightening time, and a graph attention network is used to capture the elastic interaction mechanism between bolts. A three-stage multi-fidelity strategy is employed for model training: the first stage uses the elastic interaction analytical formula to generate large-scale pure elastic simulation data for pre-training; the second stage uses nonlinear finite metadata containing hyperelastic sealing rings for transfer learning fine-tuning; and the third stage constructs an auxiliary uncertainty quantification model, using sparse physical measured data and a Gaussian negative log-likelihood loss function to calibrate the prediction results. This invention can achieve high-precision prediction of the residual preload of sealing joints while significantly reducing computational and experimental costs, and simultaneously outputs the confidence level and deviation correction of the prediction results, effectively improving the monitoring level of high-reliability equipment assembly processes.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided engineering (CAE) and industrial intelligent monitoring technology, specifically relating to a method for predicting the preload of a sealing joint using a digital twin based on a multi-fidelity graph neural network. Background Technology

[0002] In the aerospace and precision equipment manufacturing fields, flange connection structures with superelastic sealing rings are one of the core components, and their sealing performance is crucial to the overall structure's performance. To ensure the integrity of the seal, the key lies in accurately predicting the final residual preload of the bolts.

[0003] However, due to the hyperelastic-elastic coupling of the sealing ring and the complexity of the tightening process, the residual preload is highly sensitive to the tightening sequence. Traditional physical measurement methods, while providing accurate results, are costly, inefficient, and produce sparse data; while simple nonlinear finite element simulation calculations are too time-consuming to meet the real-time monitoring requirements of digital twin systems. Furthermore, existing methods often struggle to effectively quantify the uncertainties caused by differences between the simulation model and the physical entity, as well as measurement errors. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems existing in the related art.

[0005] The purpose of this invention is to provide a digital twin preload prediction method for sealing joints based on multi-fidelity graph neural networks, which integrates low-cost data with high-fidelity knowledge to improve the prediction efficiency and accuracy of the final residual preload of flange connection structures.

[0006] To achieve the above objectives, the present invention provides a method for predicting the preload of a sealing joint using a digital twin based on a multi-fidelity graph neural network, comprising the following steps:

[0007] Multiple datasets were constructed, including: a pure elastic simulation dataset generated in batches based on the elastic interaction coefficient matrix method, a nonlinear coupled dataset containing hyperelastic sealing rings established through finite element simulation, and a measured dataset collected through physical experiments.

[0008] A basic prediction model is constructed, which models the bolt system of the flange sealing structure as a graph structure with bolts as nodes and the tightening sequence as directed edges, and uses a graph attention network to extract the elastic interaction features between bolts.

[0009] The model training is conducted in three stages: First, the basic prediction model is pre-trained using the pure elastic simulation dataset to learn the elastic interaction mechanism between bolts; Second, the pre-trained basic prediction model is fine-tuned using transfer learning with the nonlinear coupling dataset to obtain a high-fidelity prediction model adapted to nonlinear working conditions; Third, an auxiliary uncertainty quantification model is constructed, using the output of the high-fidelity prediction model and the hidden layer features of the model as input, and the auxiliary uncertainty quantification model is trained based on the measured dataset to output prediction bias and uncertainty parameters to calibrate the prediction results of the high-fidelity prediction model.

[0010] A further preferred embodiment of the present invention is that the pure elasticity simulation dataset is constructed in the following manner:

[0011] When measuring a single bolt on a flange sealing structure as a reference observation point, the change in preload of the bolt caused by the application of preload to adjacent bolts in the two adjacent areas is used as the local elastic interaction influence coefficient.

[0012] By utilizing the cyclic symmetry characteristics of the flange structure, the local elastic interaction influence coefficients are copied and mapped to the corresponding positions of all bolts to construct a global initial elastic interaction coefficient matrix;

[0013] For each simulation, the tightening sequence is set, and the rows and columns of the global initial elastic interaction coefficient matrix are rearranged and its upper triangular part is extracted to generate a time-dependent elastic interaction coefficient matrix.

[0014] Then, a large number of different tightening sequence codes and initial preload values ​​are randomly generated by computer program. The corresponding residual preload results are calculated in batches using the time-related elastic interaction coefficient matrix. The input initial preload, tightening sequence and output residual preload are paired and stored to form the pure elastic simulation dataset.

[0015] Preferably, the nonlinear coupling dataset is constructed as follows:

[0016] The script interface of the finite element software is called to create three-dimensional geometric models of flanges, bolts and hyperelastic seals, and to assign material properties and mesh generation rules.

[0017] Define the contact and boundary conditions between the surfaces of each component in the simulation environment to simulate the real physical installation environment;

[0018] Automated batch solving is performed according to a variety of pre-generated tightening timing schemes;

[0019] The axial residual tensile force values ​​of each bolt section at the end of the simulation are extracted, and the tightening sequence code and the residual preload value are paired and stored to form the nonlinear coupled dataset.

[0020] Preferably, the measured dataset is constructed as follows:

[0021] The test bench was built using physical metal flanges, bolts and sealing rings of the same material and size as the three-dimensional geometric model created by the finite element software, and ultrasonic length measuring sensors or patch pressure sensors were deployed at the ends of the bolts.

[0022] Perform multiple batches of tightening operations according to the preset tightening sequence code list. After each tightening action is completed, trigger the data acquisition system to record the current status of all bolts, and wait to eliminate the creep effect of the sealing ring after all bolts are tightened.

[0023] The acquired signals are denoised and converted into standard mechanical units. The final residual preload values ​​of each bolt are extracted. The tightening sequence code, the initial preload setting value and the residual preload value are aligned and matched to construct the measured dataset.

[0024] Preferably, the bolt system of the flange sealing structure is modeled as a graph structure with bolts as nodes and the tightening sequence as directed edges; specifically:

[0025] Each bolt is mapped to a corresponding node and its initial preload value is extracted as a node feature.

[0026] Read the bolt tightening process document to obtain the absolute time sequence of each bolt tightening operation;

[0027] Traverse each node, identify all nodes that have been tightened before the current node in the time series, establish unidirectional connection edges from the current node to these tightened nodes, and generate an adjacency matrix representing the time-dependent paths.

[0028] Preferably, the basic prediction model is pre-trained using the pure elastic simulation dataset to learn the elastic interaction mechanism between bolts; specifically:

[0029] Build a graph attention network architecture, instantiate graph attention network model objects in a deep learning computing framework, and configure the connection methods, computing logic and input / output interface parameters of each functional layer;

[0030] The iterative loop of forward computation and backpropagation is performed using the pure elastic simulation dataset to update the weight parameters inside the network and learn the elastic interaction mechanism between bolts.

[0031] Preferably, the step of using the nonlinear coupling dataset to perform transfer learning fine-tuning on the pre-trained basic prediction model to obtain a high-fidelity prediction model adapted to nonlinear operating conditions is as follows:

[0032] Load the pre-trained base prediction model and restore the network structure and weight parameters;

[0033] Freeze the shallow feature extraction layers or specific attention head modules of the network to keep them unchanged during fine-tuning;

[0034] Gradient updates are performed on the deep inference and output layers of the network, and the weights of the deep network are updated in a directional manner using the nonlinear coupling dataset, so that the model can adapt to nonlinear working conditions containing hyperelastic sealing rings and obtain a high-fidelity prediction model.

[0035] Preferably, the auxiliary uncertainty quantification model is constructed by taking the output of the high-fidelity prediction model and the hidden layer features of the model as input, and training the auxiliary uncertainty quantification model on the basis of the measured dataset, so that it outputs prediction bias and uncertainty parameters to calibrate the prediction results of the high-fidelity prediction model; specifically:

[0036] A lightweight multilayer perceptron network is constructed as the auxiliary uncertainty quantification model;

[0037] The output of the high-fidelity prediction model is concatenated with the hidden layer features of the model and used as the input of the auxiliary uncertainty quantification model. The tightening sequence code and the initial preload setting value in the measured dataset are used as input working condition parameters. The high-fidelity prediction model is used to generate predicted values ​​and hidden layer features, which together with the measured residual preload value in the measured dataset constitute training samples.

[0038] The auxiliary uncertainty quantification model is trained using a Gaussian negative log-likelihood loss function to output the prediction bias correction and log variance. The loss function calculates the square of the difference between the measured value of residual preload and the corrected predicted value in the measured dataset and divides it by the variance term. At the same time, the log variance term is accumulated as the optimization objective.

[0039] During training, all parameters of the high-fidelity prediction model are frozen, and only the weights of the auxiliary uncertainty quantification model are updated until the loss value converges, thereby achieving calibration of the prediction results and uncertainty assessment.

[0040] In another aspect, the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, the computer instructions causing a computer to execute the above-described method for predicting the preload of a sealing joint based on a multi-fidelity graph neural network.

[0041] In another aspect, the present invention provides an electronic device, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus, and the processor calls logical instructions in the memory to execute the above-mentioned method for predicting the preload of a sealing joint based on a multi-fidelity graph neural network.

[0042] In another aspect, the present invention provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer executes the above-described method for predicting the preload of a sealing joint based on a multi-fidelity graph neural network.

[0043] Compared with the prior art, the advantages of the present invention are as follows:

[0044] 1. This invention adopts a multi-fidelity data fusion strategy, which replaces a large number of expensive physical experiments with low-cost elastic formulas and simulation data. While significantly reducing time and economic costs, it ensures the model's prediction accuracy for nonlinear working conditions.

[0045] 2. This invention innovatively maps the tightening process timing into a directed topology of a graph neural network, effectively capturing the complex timing dependencies and elastic crosstalk in the sealing joint, and solving the problem of variable sequence prediction that is difficult to handle by traditional methods.

[0046] 3. This invention introduces an auxiliary uncertainty quantification mechanism, which can use sparse physical measurement data to calibrate the simulation model, and provide confidence assessment while outputting high-precision prediction values, thus providing a scientific decision-making basis for the assembly quality monitoring of high-reliability equipment. Attached Figure Description

[0047] Figure 1 This is a flowchart of a digital twin preload prediction method for sealing joints based on multi-fidelity graph neural networks. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0049] The following is combined Figure 1This invention describes a digital twin preload prediction method for sealing joints based on a multi-fidelity graph neural network.

[0050] Example 1: This example provides a digital twin preload prediction method for sealing joints based on multi-fidelity graph neural networks. It maps the bolt tightening time sequence to a directed graph topology and uses a graph attention network to capture the elastic interaction mechanism between bolts. A three-stage multi-fidelity strategy is adopted: "pure elastic simulation pre-training - nonlinear simulation fine-tuning - physical data calibration". First, large-scale elastic data based on analytical formulas is used to allow the model to master basic time sequence patterns; second, transfer learning is performed using a small amount of nonlinear finite data to adapt the model to hyperelastic coupling conditions containing sealing rings; finally, sparse physical measured data is used to train an auxiliary uncertainty quantification model.

[0051] Specific steps are as follows: Figure 1 As shown, it includes:

[0052] S1. Construct an elastic residual preload generator and use the elastic interaction coefficient matrix method to generate a batch of simulation datasets under pure elastic conditions. The simulation datasets include the initial preload, tightening sequence and corresponding elastic residual preload.

[0053] S11. Define the input-output mapping relationship and establish a numerical calculation module. The function of this module is to map the input initial preload numerical vector to the output residual preload numerical vector. The core of this mapping calculation lies in calling the elastic interaction coefficient matrix and directly obtaining the numerical result of the residual preload through matrix operations.

[0054] S12. Determine the basic data of local interaction, selecting any bolt on the flange structure as a reference observation point. Apply preload to a specific number of adjacent bolts within the neighborhood on both sides of this reference observation point. During the application process, record the numerical change of the preload at the reference observation point using a measuring device. Compare the recorded change with the applied force to extract the elastic interaction coefficient within this local area. For bolts outside the neighborhood, directly set their corresponding interaction coefficient to zero; the specific method is as follows:

[0055] S121. On an actual flange or a test bench with equivalent physical properties, select any bolt as a reference monitoring point and install an ultrasonic length gauge or other real-time preload monitoring device on the bolt.

[0056] S122. First, apply a tightening torque to the bolt at the reference monitoring point to make it reach the preset preload value, and record the initial state.

[0057] S123. Tighten a specific number of adjacent bolts located in the neighborhood on both sides of the bolt at the reference monitoring point, for example, four on each side;

[0058] S124. During the tightening of each adjacent bolt, the decrease in preload on the bolt at the reference monitoring point is read and recorded simultaneously.

[0059] S125. Divide the recorded preload reduction value by the corresponding tightening force value of the adjacent bolt to obtain the elastic interaction influence coefficient in the local area; for the far bolts that have not been operated outside the neighborhood, directly set their corresponding interaction coefficient to zero.

[0060] S13. Construct a global initial coefficient matrix. Utilizing the cyclic symmetry of the flange structure, copy and map the local interaction coefficients of the reference observation points measured in step S12 to the corresponding positions of all other bolts on the flange. Through this data filling method, a complete initial elastic interaction coefficient matrix containing the mutual influence relationships of all bolts on the flange is established. The specific method is as follows:

[0061] S131. Extract a set of interaction coefficients for a single reference bolt obtained in step S125 and use it as a standard data template.

[0062] S132. Based on the cyclic symmetry characteristics of the flange structure, for each bolt position on the flange, the data template is cyclically shifted according to the relative position of the bolt index.

[0063] S133. Fill the coefficient data after cyclic shifting into the corresponding rows of the global matrix in sequence to construct the initial elastic interaction coefficient matrix containing the mutual influence relationship of all bolts on the flange.

[0064] S14. Generate the time-series correlation coefficient matrix. For each simulation's specific tightening sequence, generate corresponding row and column transformation rules. Use these rules to rearrange the rows and columns of the initial elastic interaction coefficient matrix obtained in step S13. Extract the values ​​from the upper triangular portion of the rearranged matrix and determine this upper triangular portion as the final elastic interaction coefficient matrix under the current specific tightening sequence.

[0065] S15. Batch synthesis of elastic simulation datasets: A large number of different tightening sequence codes and corresponding initial preload values ​​are randomly generated using a computer program. Each set of random parameters is input into the calculation process determined in step S14, and the corresponding residual preload results are calculated in batches. The input initial preload, tightening sequence, and output residual preload results are paired and stored to form a large-scale pure elastic simulation dataset for subsequent model training.

[0066] S2. Construct a basic preload prediction model based on a graph attention network; model the bolt system as a directed graph structure, where the feature matrix is ​​the initial preload vector and the adjacency matrix is ​​constructed based on the tightening time sequence; use the simulation dataset in step S1 to pre-train the basic preload prediction model so that it learns the elastic interaction mechanism between bolts.

[0067] S21. Construct a time-driven graph data structure, perform data mapping operations, convert the physical entity parameters in the flange structure into a graph theory data format in computer memory, and establish a digital representation of node indices and time-series connection relationships; specifically:

[0068] S211. Establish a node index list and map each physical bolt on the flange structure to a virtual node in the list according to its spatial location or process number.

[0069] S212. Construct a node feature matrix, extract the pre-set initial preload value of each bolt, encapsulate it into a one-dimensional feature vector, and assign it to the corresponding virtual node as the initial state attribute of the graph.

[0070] S213. Read and parse the bolt tightening process document to extract the absolute time sequence information of all bolts being tightened.

[0071] S214. Execute the directed edge generation program, traverse each node, identify all tightened nodes that are located before the current node in the time series, and establish unidirectional connection edges from the current node to these tightened nodes.

[0072] S215. Based on the established connection relationship, generate an adjacency matrix data structure, mark the positions of node pairs with connected edges as valid values, and thus generate a numerical matrix representing temporally dependent paths.

[0073] S22. Build a graph attention network architecture, instantiate a network model object in the deep learning computing framework, and configure the connection method, computing logic and input / output interface parameters of each functional layer.

[0074] S23. Perform pre-training for the pure elastic working condition, load the simulation data generated in step S1 into the model training environment, execute the iterative loop of forward calculation and backpropagation, and update the weight parameters inside the network numerically.

[0075] S3. By developing finite element simulation software, a nonlinear finite element model containing a hyperelastic sealing ring is established; and hyperelastic-elastic coupling datasets under different tightening sequences are generated in batches.

[0076] S31. Construct a parametric nonlinear geometric model, call the script interface of the finite element software, execute geometric modeling instructions, create three-dimensional solids of flanges, bolts and sealing rings, and assign corresponding material properties and mesh generation rules.

[0077] S32. Configure contact and boundary conditions, define the physical interaction rules between the surfaces of each component in the simulation environment, and set constraints to simulate the real physical installation environment;

[0078] S33. Execute the automated batch solution script, write the loop control program, drive the finite element solver to continuously submit calculation jobs and monitor the solution status according to the pre-generated multiple tightening timing schemes.

[0079] S34. Extract data and construct a coupled dataset, access the simulation results database, parse the results files, extract key mechanical response data, and store them in a structured manner; specifically:

[0080] S341. Call the post-processing interface to open the simulation result database file after the solution is completed;

[0081] S342. Traverse all bolt nodes and extract the axial residual tensile force values ​​on each bolt section at the end of the simulation.

[0082] S343. Use the current tightening sequence code as the input feature and the extracted residual preload values ​​of each bolt as the output label, and write them into the data file in pairs.

[0083] S344. Repeat the above extraction process until all simulation jobs are completed, and then compile and generate a hyperelastic-elastic coupling dataset containing nonlinear contact features.

[0084] S4. Using the coupled dataset in step S3, the pre-trained basic preload prediction model in step S2 is fine-tuned through transfer learning to obtain a high-fidelity preload prediction model that adapts to nonlinear working conditions.

[0085] S41. Load and partition the nonlinear dataset, call the data processing interface, read the hyperelastic-elastic coupling dataset generated in step S3, and perform data shuffling and partitioning operations.

[0086] S42. Load the pre-trained model and freeze the parameters. Read the model file saved in step S2, restore the network structure and weight parameters, and modify the trainable properties of some network layers; specifically:

[0087] S421. Instantiate the graph attention network model object, call the parameter loading function, and write the weight parameters saved in the pre-training stage of step S2 into the current model structure.

[0088] S422. Traverse the network layer structure and select the shallow feature extraction layer or a specific attention head module.

[0089] S423. Execute the parameter freeze command to set the gradient calculation attribute of the selected layer to the off state, and lock its weight values ​​so that they will not change in subsequent training.

[0090] S424. Confirm that the gradient calculation attribute of the deep inference and output layers of the network is enabled to ensure that these layers can receive gradient updates.

[0091] S43. Perform transfer learning fine-tuning training, input nonlinear data into the model, perform backpropagation while freezing some parameters, and update the weights of the deep network in a targeted manner.

[0092] S44. Verify and save the high-fidelity model, monitor the error curve during the fine-tuning process, and screen and persistently store the model parameters with the best performance.

[0093] S5. Construct an auxiliary uncertainty quantification model, wherein the auxiliary model is a multilayer perceptron; obtain sparse preload data measured in actual physical experiments as the measured dataset.

[0094] S51. Construct a physical experimental test bench, assemble a physical test device with parameters consistent with the simulation model, and deploy high-precision measuring equipment to obtain real mechanical response data; specifically:

[0095] S511. Select metal flanges, bolts, fasteners and sealing rings that are completely consistent with the material and size of the finite element model, and perform assembly operations according to standard processes.

[0096] S512. Install an ultrasonic length measuring sensor or a patch pressure sensor at the end of the bolt, connect the data acquisition card to the industrial control computer, and verify the stability of the signal transmission channel.

[0097] S513, equipped with a digital torque wrench, set the target tightening torque value to ensure that the applied initial load is consistent with the simulation setting.

[0098] S52. Conduct multiple batches of physical tightening tests, performing tightening operations according to various preset process schemes, and collecting data on preload changes throughout the process; specifically:

[0099] S521. Read the preset tightening sequence code list, which includes typical timing sequences such as star, circle, and diagonal, and determine the current experimental operation procedure;

[0100] S522. The operator holds a torque wrench and applies torque to the bolts in strict sequence. After each tightening action is completed, the data acquisition system is triggered to record the current status of all bolts.

[0101] S523. After all bolts have been tightened, wait for a period of time to eliminate the creep effect of the seal ring, and then record the final stable residual preload value of each bolt.

[0102] S524. Perform disassembly and reset operations, replace the sealing ring with a new one to avoid cumulative errors due to plastic deformation, and repeat the above until all preset timing experiments are completed.

[0103] S53. Construct a sparse physical measurement dataset, clean and format the original acquired signals, and construct a ground truth table to assist model training; specifically:

[0104] S531. Export the original waveform file from the data acquisition system, execute the signal denoising program, and filter out high-frequency electromagnetic interference and abnormal abrupt changes.

[0105] S532. Extract the final residual preload value of each bolt, convert it into standard mechanical units, and use it as label data for physical experiments;

[0106] S533. Use the corresponding tightening sequence code and initial preload setting value as index data, and perform row alignment matching with the label data;

[0107] S534. Store the organized data pairs into the physical measurement database. Due to experimental cost limitations, the number of samples in this dataset is significantly less than that in the simulation dataset.

[0108] S54. Build an auxiliary neural network model architecture. Instantiate a lightweight multilayer perceptron model in the deep learning framework, which is specifically used to learn prediction bias and uncertainty.

[0109] S6. The output of the high-fidelity preload prediction model in step S4 is concatenated with the node features of the model's hidden layer and used as the input of the auxiliary uncertainty quantification model. Based on the measured dataset, the auxiliary uncertainty quantification model is trained to output the prediction bias and logarithmic variance, thus completing the final prediction and uncertainty assessment of the residual preload.

[0110] S61. Construct a hybrid feature input vector, execute forward inference instructions, use a high-fidelity model to process the working parameters of the physical experiment, extract deep features and shallow prediction values ​​and perform data fusion.

[0111] S62. Configure the probability loss calculation logic, and define a special loss function calculation module in the training environment to handle regression problems without direct variance labels; specifically:

[0112] S621. Write a program to calculate the Gaussian negative log-likelihood loss function, defining that the function accepts three input variables: the physical measured true value, the bias correction amount of the auxiliary model output, and the log-variance of the auxiliary model output;

[0113] S622. Configure the first calculation logic to calculate the square of the difference between the physical true value and the corrected predicted value, and divide it by the variance term to penalize the prediction bias;

[0114] S623. Configure the second calculation logic to directly accumulate the logarithmic variance term to prevent the model from outputting an infinitely large variance;

[0115] S624. The sum of the above two items is taken as the final optimization target value and registered into the automatic differentiation engine of the deep learning framework.

[0116] S63. Perform joint training of the auxiliary model, freeze the backbone network parameters, and perform optimization iterations only on the auxiliary model, using sparse physical data to calibrate the prediction results; specifically:

[0117] S631. Set the parameter update filter to lock all weight parameters of the preceding high-fidelity graph neural network to ensure that they remain unchanged during the training of physical data.

[0118] S632, Activate the parameter update permission for the auxiliary uncertainty quantification model;

[0119] S633. Start the training loop, input the hybrid feature vector constructed in step S61 into the auxiliary model, and output the prediction bias and variance.

[0120] S634. Call the loss function configured in step S62 to calculate the loss value, perform backpropagation, and drive the optimizer to update the weights of the auxiliary model until the loss value converges.

[0121] S64. Generate the final prediction and confidence report, deploy the complete joint model system, and perform end-to-end prediction and evaluation for unknown new operating conditions.

[0122] Example 2: This example provides a digital twin preload prediction device for sealing joints based on multi-fidelity graph neural networks. The device includes:

[0123] The elastic simulation data generation module is used to build an elastic residual preload generator. It uses the elastic interaction coefficient matrix method to generate simulation datasets under pure elastic conditions in batches. The dataset includes the initial preload, tightening sequence and corresponding elastic residual preload, providing the model with a basic understanding of physical laws.

[0124] The graph neural network construction and pre-training module is used to build a basic preload prediction model based on graph attention network; the bolt system is modeled as a directed graph structure, where the feature matrix is ​​composed of the initial preload vector, the adjacency matrix is ​​constructed based on the tightening time sequence, and the later tightened bolt points to the earlier tightened bolt; the model is pre-trained using elastic simulation data to enable it to master the elastic interaction mechanism between bolts.

[0125] The nonlinear fine-tuning module is used to generate nonlinear coupled datasets containing hyperelastic sealing rings in batches through secondary development of finite element simulation software. The pre-trained basic model is then fine-tuned using this dataset through transfer learning to correct the nonlinear deviation caused by the compression and rebound of the sealing ring, thereby obtaining a high-fidelity preload prediction model.

[0126] The uncertainty quantification module is used to construct an auxiliary multilayer perceptron network. It concatenates the output prediction value of the high-fidelity model with the hidden layer node features as input. Based on sparse physical measurement data, the auxiliary model is trained through the Gaussian negative log-likelihood loss function, and the prediction bias correction and log variance are output to complete the final prediction and confidence assessment of the residual preload.

[0127] Example 3: This example provides a non-transitory computer-readable storage medium storing computer instructions that cause a computer to execute a digital twin preload prediction method for sealing joints based on a multi-fidelity graph neural network. The method includes the following steps:

[0128] S1. Construct an elastic residual preload generator and use the elastic interaction coefficient matrix method to generate a batch of simulation datasets under pure elastic conditions. The simulation datasets include the initial preload, tightening sequence and corresponding elastic residual preload.

[0129] S2. Construct a basic preload prediction model based on a graph attention network; model the bolt system as a directed graph structure, where the feature matrix is ​​the initial preload vector and the adjacency matrix is ​​constructed based on the tightening time sequence; use the simulation dataset in step S1 to pre-train the basic preload prediction model so that it learns the elastic interaction mechanism between bolts;

[0130] S3. By developing finite element simulation software, a nonlinear finite element model containing a hyperelastic sealing ring is established; and hyperelastic-elastic coupling datasets under different tightening sequences are generated in batches.

[0131] S4. Use the coupled dataset in step S3 to perform transfer learning fine-tuning on the pre-trained basic preload prediction model in step S2 to obtain a high-fidelity preload prediction model that adapts to nonlinear working conditions.

[0132] S5. Construct an auxiliary uncertainty quantification model, wherein the auxiliary model is a multilayer perceptron; obtain sparse preload data measured in actual physical experiments as the measured dataset;

[0133] S6. The output of the high-fidelity preload prediction model in step S4 is concatenated with the node features of the model's hidden layer and used as the input of the auxiliary uncertainty quantification model. Based on the measured dataset, the auxiliary uncertainty quantification model is trained to output the prediction bias and logarithmic variance, thus completing the final prediction and uncertainty assessment of the residual preload.

[0134] Example 4: This example provides an electronic device that may include a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The processor can call logical instructions from the memory to execute a digital twin preload prediction method for sealing joints based on a multi-fidelity graph neural network. This method includes the following steps:

[0135] S1. Construct an elastic residual preload generator and use the elastic interaction coefficient matrix method to generate a batch of simulation datasets under pure elastic conditions. The simulation datasets include the initial preload, tightening sequence and corresponding elastic residual preload.

[0136] S2. Construct a basic preload prediction model based on a graph attention network; model the bolt system as a directed graph structure, where the feature matrix is ​​the initial preload vector and the adjacency matrix is ​​constructed based on the tightening time sequence; use the simulation dataset in step S1 to pre-train the basic preload prediction model so that it learns the elastic interaction mechanism between bolts;

[0137] S3. By developing finite element simulation software, a nonlinear finite element model containing a hyperelastic sealing ring is established; and hyperelastic-elastic coupling datasets under different tightening sequences are generated in batches.

[0138] S4. Use the coupled dataset in step S3 to perform transfer learning fine-tuning on the pre-trained basic preload prediction model in step S2 to obtain a high-fidelity preload prediction model that adapts to nonlinear working conditions.

[0139] S5. Construct an auxiliary uncertainty quantification model, wherein the auxiliary model is a multilayer perceptron; obtain sparse preload data measured in actual physical experiments as the measured dataset;

[0140] S6. The output of the high-fidelity preload prediction model in step S4 is concatenated with the node features of the model's hidden layer and used as the input of the auxiliary uncertainty quantification model. Based on the measured dataset, the auxiliary uncertainty quantification model is trained to output the prediction bias and logarithmic variance, thus completing the final prediction and uncertainty assessment of the residual preload.

[0141] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0142] Example 5: This example provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a digital twin preload prediction method for sealing joints based on a multi-fidelity graph neural network. This method includes the following steps:

[0143] S1. Construct an elastic residual preload generator and use the elastic interaction coefficient matrix method to generate a batch of simulation datasets under pure elastic conditions. The simulation datasets include the initial preload, tightening sequence and corresponding elastic residual preload.

[0144] S2. Construct a basic preload prediction model based on a graph attention network; model the bolt system as a directed graph structure, where the feature matrix is ​​the initial preload vector and the adjacency matrix is ​​constructed based on the tightening time sequence; use the simulation dataset in step S1 to pre-train the basic preload prediction model so that it learns the elastic interaction mechanism between bolts;

[0145] S3. By developing finite element simulation software, a nonlinear finite element model containing a hyperelastic sealing ring is established; and hyperelastic-elastic coupling datasets under different tightening sequences are generated in batches.

[0146] S4. Use the coupled dataset in step S3 to perform transfer learning fine-tuning on the pre-trained basic preload prediction model in step S2 to obtain a high-fidelity preload prediction model that adapts to nonlinear working conditions.

[0147] S5. Construct an auxiliary uncertainty quantification model, wherein the auxiliary model is a multilayer perceptron; obtain sparse preload data measured in actual physical experiments as the measured dataset;

[0148] S6. The output of the high-fidelity preload prediction model in step S4 is concatenated with the node features of the model's hidden layer and used as the input of the auxiliary uncertainty quantification model. Based on the measured dataset, the auxiliary uncertainty quantification model is trained to output the prediction bias and logarithmic variance, thus completing the final prediction and uncertainty assessment of the residual preload.

[0149] This invention can achieve high-precision prediction of residual preload while significantly reducing computation and experimental costs, and simultaneously output the deviation correction amount and confidence level of the prediction results, directly improving the monitoring level of high-reliability assembly processes.

[0150] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0151] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0152] 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 method for predicting the preload of a sealing joint using a digital twin based on a multi-fidelity graph neural network, characterized in that, It has the following steps: Multiple datasets were constructed, including: a pure elastic simulation dataset generated in batches based on the elastic interaction coefficient matrix method, a nonlinear coupled dataset containing hyperelastic sealing rings established through finite element simulation, and a measured dataset collected through physical experiments. A basic prediction model is constructed, which models the bolt system of the flange sealing structure as a graph structure with bolts as nodes and the tightening sequence as directed edges, and uses a graph attention network to extract the elastic interaction features between bolts. The model training is conducted in three stages: First, the basic prediction model is pre-trained using the pure elastic simulation dataset to learn the elastic interaction mechanism between bolts; Second, the pre-trained basic prediction model is fine-tuned using transfer learning with the nonlinear coupling dataset to obtain a high-fidelity prediction model adapted to nonlinear working conditions; Third, an auxiliary uncertainty quantification model is constructed, using the output of the high-fidelity prediction model and the hidden layer features of the model as input, and the auxiliary uncertainty quantification model is trained based on the measured dataset to output prediction bias and uncertainty parameters to calibrate the prediction results of the high-fidelity prediction model.

2. The method for predicting the preload of a sealing joint based on a multi-fidelity graph neural network according to claim 1, characterized in that, The pure elasticity simulation dataset is constructed as follows: When measuring a single bolt on a flange sealing structure as a reference observation point, the change in preload of the bolt caused by the application of preload to adjacent bolts in the two adjacent areas is used as the local elastic interaction influence coefficient. By utilizing the cyclic symmetry characteristics of the flange structure, the local elastic interaction influence coefficients are copied and mapped to the corresponding positions of all bolts to construct a global initial elastic interaction coefficient matrix; For each simulation, the tightening sequence is set, and the rows and columns of the global initial elastic interaction coefficient matrix are rearranged and its upper triangular part is extracted to generate a time-dependent elastic interaction coefficient matrix. Then, a large number of different tightening sequence codes and initial preload values ​​are randomly generated by computer program. The corresponding residual preload results are calculated in batches using the time-related elastic interaction coefficient matrix. The input initial preload, tightening sequence and output residual preload are paired and stored to form the pure elastic simulation dataset.

3. The method for predicting the preload of a sealing joint based on a multi-fidelity graph neural network according to claim 1, characterized in that, The specific method for constructing the nonlinear coupling dataset is as follows: The script interface of the finite element software is called to create three-dimensional geometric models of flanges, bolts and hyperelastic seals, and to assign material properties and mesh generation rules. Define the contact and boundary conditions between the surfaces of each component in the simulation environment to simulate the real physical installation environment; Automated batch solving is performed according to a variety of pre-generated tightening timing schemes; The axial residual tensile force values ​​of each bolt section at the end of the simulation are extracted, and the tightening sequence code and the residual preload value are paired and stored to form the nonlinear coupled dataset.

4. The method for predicting the preload of a sealing joint based on a multi-fidelity graph neural network according to claim 3, characterized in that, The specific method for constructing the measured dataset is as follows: The test bench was built using physical metal flanges, bolts and sealing rings of the same material and size as the three-dimensional geometric model created by the finite element software, and ultrasonic length measuring sensors or patch pressure sensors were deployed at the ends of the bolts. Perform multiple batches of tightening operations according to the preset tightening sequence code list. After each tightening action is completed, trigger the data acquisition system to record the current status of all bolts, and wait to eliminate the creep effect of the sealing ring after all bolts are tightened. The acquired signals are denoised and converted into standard mechanical units. The final residual preload values ​​of each bolt are extracted. The tightening sequence code, the initial preload setting value and the residual preload value are aligned and matched to construct the measured dataset.

5. The method for predicting the preload of a sealing joint based on a multi-fidelity graph neural network according to claim 1, characterized in that, The bolt system of the flange sealing structure is modeled as a graph structure with bolts as nodes and the tightening sequence as directed edges; specifically: Each bolt is mapped to a corresponding node and its initial preload value is extracted as a node feature. Read the bolt tightening process document to obtain the absolute time sequence of each bolt tightening operation; Traverse each node, identify all nodes that have been tightened before the current node in the time series, establish unidirectional connection edges from the current node to these tightened nodes, and generate an adjacency matrix representing the time-dependent paths.

6. The method for predicting the preload of a sealing joint based on a multi-fidelity graph neural network according to claim 5, characterized in that, The basic prediction model is pre-trained using the pure elastic simulation dataset to learn the elastic interaction mechanism between bolts; specifically: Build a graph attention network architecture, instantiate graph attention network model objects in a deep learning computing framework, and configure the connection methods, computing logic and input / output interface parameters of each functional layer; The iterative loop of forward computation and backpropagation is performed using the pure elastic simulation dataset to update the weight parameters inside the network and learn the elastic interaction mechanism between bolts.

7. The method for predicting the preload of a sealing joint based on a multi-fidelity graph neural network according to claim 1, characterized in that, The pre-trained basic prediction model is fine-tuned using the nonlinear coupling dataset through transfer learning to obtain a high-fidelity prediction model adapted to nonlinear operating conditions; specifically: Load the pre-trained base prediction model and restore the network structure and weight parameters; Freeze the shallow feature extraction layers or specific attention head modules of the network to keep them unchanged during fine-tuning; Gradient updates are performed on the deep inference and output layers of the network, and the weights of the deep network are updated in a directional manner using the nonlinear coupling dataset, so that the model can adapt to nonlinear working conditions containing hyperelastic sealing rings and obtain a high-fidelity prediction model.

8. The method for predicting the preload of a sealing joint based on a multi-fidelity graph neural network according to claim 1, characterized in that, The auxiliary uncertainty quantification model is constructed by taking the output of the high-fidelity prediction model and the hidden layer features of the model as inputs, and training the auxiliary uncertainty quantification model with the actual test dataset as a benchmark, so that it outputs prediction bias and uncertainty parameters to calibrate the prediction results of the high-fidelity prediction model. Specifically: A lightweight multilayer perceptron network is constructed as the auxiliary uncertainty quantification model; The output of the high-fidelity prediction model is concatenated with the hidden layer features of the model and used as the input of the auxiliary uncertainty quantification model. The tightening sequence code and the initial preload setting value in the measured dataset are used as input working condition parameters. The high-fidelity prediction model is used to generate predicted values ​​and hidden layer features, which together with the measured residual preload value in the measured dataset constitute training samples. The auxiliary uncertainty quantification model is trained using a Gaussian negative log-likelihood loss function to output the prediction bias correction and log variance. The loss function calculates the square of the difference between the measured value of residual preload and the corrected predicted value in the measured dataset and divides it by the variance term. At the same time, the log variance term is accumulated as the optimization objective. During training, all parameters of the high-fidelity prediction model are frozen, and only the weights of the auxiliary uncertainty quantification model are updated until the loss value converges, thereby achieving calibration of the prediction results and uncertainty assessment.

9. A non-transitory computer-readable storage medium, characterized in that, It stores computer instructions that cause the computer to execute the digital twin preload prediction method for sealing joints based on multi-fidelity graph neural networks as described in any one of claims 1-8.

10. An electronic device, characterized in that, include: The system includes a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The processor calls logical instructions from the memory to execute the digital twin preload prediction method for sealing joints based on multi-fidelity graph neural networks as described in any one of claims 1-8.

Images