A graph neural network prediction method for thin plate welding deformation considering uniform weld geometry deviation

Abstract

The application discloses a sheet welding deformation prediction method and system considering uniform weld geometry deviation, and belongs to the technical field of welding manufacturing and intelligent prediction. The method is aimed at the post-weld deformation prediction problem caused by the uniform weld geometry deviation introduced in the assembly stage during the butt welding process of the sheet. First, the discretized sheet structure is represented as graph topology data, and the spatial coordinates of each discrete node and the adjacent relationship thereof are taken as the basic input. Then, the spatial correlation features of the sheet structure are extracted and encoded by using a graph attention encoder, and the assembly gap and the surface difference representing the assembly state are injected into the node embedding representation as global conditions. On this basis, the post-weld final deformation prediction results of each node are output by using a graph attention decoder, and the residual stress prediction is introduced as an auxiliary task to construct a joint training mechanism, so that the representation ability of the model to the welding thermal coupling physical process is enhanced. By using the technical scheme of the application, the spatial continuous features of the sheet welding deformation and the influence of the geometry deviation can be more effectively represented, the physical consistency and the overall precision of the prediction results are improved, and the calculation efficiency of the sheet welding deformation prediction under complex working conditions is significantly improved.

Description

Technical Field

[0001] This invention relates to the fields of welding manufacturing, structural response prediction, and graph neural network intersection technology, specifically to a graph neural network prediction method for thin plate welding deformation considering uniform weld geometric deviations. This method addresses thin plate butt welding scenarios in shipbuilding, automotive, aerospace, and high-end equipment manufacturing. It utilizes the spatial topological relationships of discrete nodes in the thin plate, as well as pre-weld geometric deviations such as assembly gaps and surface differences, to rapidly predict the final deformation distribution after welding. Furthermore, residual stress-assisted learning enhances the model's ability to characterize the welding thermo-mechanical coupling process. Background Technology

[0002] Thin-plate parts, due to their large scale and low stiffness, are highly susceptible to the combined effects of heat input, boundary constraints, and changes in assembly conditions during welding, resulting in significant post-weld deformation. This deformation directly impacts the part's shape quality and assembly accuracy, further affecting subsequent final assembly matching, service performance, and manufacturing costs. Therefore, modeling, predicting, and controlling welding deformation in thin plates has always been a crucial research direction in the welding manufacturing field. In actual assembly, the interface state before welding is often not ideally fitted. Due to manufacturing errors in the initial sheet metal, fixture positioning deviations, and inconsistent external constraints, assembly gaps and surface differences typically form at the weld joint. The former reflects the width of the horizontal gap between the two thin plates at the interface, while the latter reflects the degree of misalignment of the base materials on both sides of the weld in the normal direction. For thin-plate structures, these two types of geometric deviations alter the actual heat input distribution near the weld, the molten pool formation, and the local mechanical response, further influencing the spatial evolution of residual stress and final deformation. Therefore, assembly gaps and surface differences are not simply manufacturing errors, but rather indispensable prerequisites for predicting welding deformation. Existing methods for analyzing welding deformation in thin plates can be broadly categorized into two types. One type is numerical simulation methods, represented by thermo-elastic-plastic finite element analysis. These methods can describe the thermal cycle, stress evolution, and deformation formation mechanism during welding in relatively detailed terms, but the modeling process is complex and computationally expensive, making it difficult to meet the needs of rapid evaluation under multiple working conditions. The other type is data-driven methods based on traditional regression models or general deep learning models. While these methods have certain computational efficiency advantages, they often struggle to simultaneously consider the response differences in local high-gradient regions and the continuity of the overall deformation field when dealing with structural objects like thin plates that exhibit significant spatial correlations. Especially near the weld, the interactions between nodes have obvious directionality and spatial dependence. If each node is modeled independently, or if only fully connected networks without explicit utilization of the topology are used, problems such as discontinuous prediction fields, insufficient characterization of local gradients, and inadequate representation of the symmetry features on both sides of the weld can easily arise. Furthermore, even within a graph neural network framework, using a uniform aggregation strategy with standard graph convolution will result in equal weighting of information from different neighboring nodes, making it difficult to distinguish the differentiated contributions of the weld neighborhood and the non-weld neighborhood to the target node. As the number of graph convolutional layers increases, the problem of node feature convergence may arise, thereby weakening the model's ability to identify local physical features. These shortcomings in existing technologies directly lead to two consequences: first, limited prediction accuracy, particularly in the weld seam and its adjacent high-gradient regions where errors are more pronounced; second, the model struggles to balance accuracy and efficiency, failing to achieve fast, stable, and engineering-usable welding deformation prediction under complex assembly deviations. Therefore, a new method is needed that can utilize the discrete node topology of thin plates, explicitly incorporate uniform weld seam geometric deviations into the prediction process, and balance prediction accuracy and computational efficiency. Summary of the Invention

[0003] To address the shortcomings of existing thin-plate welding deformation prediction methods, such as insufficient consideration of geometric deviations, inadequate utilization of spatial topological relationships, large prediction errors in high-gradient regions near the weld, and high computational costs of finite element simulation, this invention proposes a graph neural network prediction method and system for thin-plate welding deformation considering uniform weld geometric deviations. This method includes core processes such as discretization modeling of the thin-plate welding object, graph topology construction, input of uniform weld geometric deviation conditions, graph attention encoding, global condition injection, graph attention decoding, and joint training of deformation and residual stress. By mapping discretized thin-plate nodes to graph vertices, mapping finite element adjacency relationships to graph edges, and injecting assembly gaps and surface differences as global deviation conditions into node embeddings, the model can simultaneously characterize the spatial correlation features of the thin-plate structure and the influence of uniform weld geometric deviations on the post-weld response. Furthermore, residual stress prediction is introduced as an auxiliary task, and a multi-task learning mechanism enhances the model's ability to represent the welding thermo-mechanical coupling process, thereby improving the spatial continuity, physical consistency, and computational efficiency of the final post-weld deformation prediction for thin plates.

[0004] Step S1: Obtaining the thin plate welding object and uniform geometric deviation parameters. Obtain the geometric model, discrete node set, and node spatial coordinates of the thin plate welding object to be predicted, and obtain the uniform geometric deviation parameters used to characterize the weld assembly state. These uniform geometric deviation parameters include assembly gap and surface difference, where the assembly gap characterizes the width of the horizontal gap between the two thin plates at the weld joint, and the surface difference characterizes the degree of misalignment of the base materials on both sides of the weld in the normal direction. Under given process conditions, the assembly gap and surface difference are input into the model as important preconditions affecting welding deformation and residual stress distribution.

[0005] Step S2: Thin Plate Model Discretization and Graph Topology Construction. The thin plate welding object is discretized, transforming the thin plate structure into a graph structure composed of a set of nodes and a set of connecting edges. Geometric nodes in the finite element mesh are mapped to vertices in the graph neural network, and node pairs that together form the edge of the finite element are mapped to undirected edges in the graph. For thin plate models with a double-layer node distribution, interlayer connecting edges in the thickness direction are further established in addition to the in-plane connecting edges to preserve the in-plane geometric relationships and the thermal response coupling relationship in the thickness direction.

[0006] Step S3: Construction of Initial Node Features and Global Deviation Conditions. The three-dimensional spatial coordinates of each discrete node are used as the initial input features of the node, and the adjacency relationship between nodes in the graph is used as the structural topology input. The assembly gap and surface difference that are kept constant along the entire weld path are used as global deviation condition parameters, so that all nodes share the same set of uniform weld geometric deviation conditions. In this way, the pre-weld assembly state is transformed into a conditional input that can be processed by the graph neural network.

[0007] Step S4: Graph Attention Encoding. The node spatial coordinates and graph topology are input into the graph attention encoder. A multi-layer graph attention message passing mechanism is used to extract the spatial structural features of each node. During each layer of graph attention propagation, the target node calculates attention coefficients based on its own features and those of its neighbors, and performs differentiated weighted aggregation of neighbor node information to obtain a node embedding representation that includes local spatial features, multi-hop neighborhood information, and topological association information. This process overcomes the problem of insufficient local feature recognition ability caused by the equal-weighted aggregation of neighbor nodes in standard graph convolution.

[0008] Step S5: Global Condition Injection. The two global scalar deviation parameters, assembly gap and surface difference, are fused into the encoded embedding representation of each node to form a conditional node representation. Specifically, the node embedding output by the graph attention encoder is concatenated with the global deviation conditions, enabling the model to perceive the current weld deviation state while preserving the spatial representation of the thin plate structure. Since the assembly gap and surface difference are shared across the entire thin plate under uniform deviation conditions, the model can further learn the differentiated responses of different spatial locations to the same set of deviation conditions.

[0009] Step S6: Graph Attention Decoding and Multi-Task Output. The conditional node representations are input into the graph attention decoder. During decoding, the spatial adjacency relationships of the graph are used for information propagation to recover the post-weld response field. After the decoder outputs the shared node decoding features, a dual-branch structure is used to predict the final post-weld deformation and residual stress of each node. The deformation prediction branch, as the main task, outputs the final post-weld deformation of each node in the thin plate, while the residual stress prediction branch, as an auxiliary task, outputs the equivalent residual stress of each node.

[0010] Step S7: Joint Loss Training. A joint loss function is constructed based on the predicted final deformation and residual stress after welding. The joint loss function includes a deformation prediction loss term and a residual stress prediction auxiliary loss term, both of which are constrained by mean square error. The influence of the residual stress prediction task on model training is controlled by the auxiliary task weight coefficient. During training, the prediction of the final deformation after welding is the core objective, while the physical correlation between the residual stress field and welding deformation is used to enhance the feature learning capabilities of the encoder and decoder.

[0011] Step S8: Model Training Sample Generation and Parameter Optimization. Training samples are generated based on a thermo-elastic-plastic finite element simulation model consistent with actual welding conditions. The training samples include nodal welding deformation data and nodal equivalent residual stress data under different assembly gaps and surface differences. The parameters of the graph attention encoder, global conditional injection module, and graph attention decoder are optimized using the training set, and the generalization performance during the training process is monitored using the validation set. After training, a graph neural network model that can quickly predict the final deformation after welding based on the thin plate diagram structure and uniform weld geometric deviation parameters is obtained.

[0012] Step S9: Rapid Prediction and Evaluation of Post-Weld Deformation. For the new uniform deviation condition, input the topology of the thin plate diagram, node spatial coordinates, assembly gaps, and surface difference parameters. The trained model outputs the predicted final deformation values ​​of each node after welding, and can simultaneously output the predicted residual stress values. By comparing with the actual welding experimental measurement results or finite element simulation results, the model's prediction performance is evaluated using indicators such as mean absolute error, mean square error, and maximum absolute error. This method can significantly shorten the single-condition analysis time while maintaining high prediction accuracy, and is suitable for rapid deformation prediction under multi-deviation conditions in thin plate welding. Attached Figure Description

[0013] Figure 1 This is an overall flowchart of the thin plate welding deformation prediction method described in this invention; Figure 2 This is a schematic diagram of the thin plate butt welding structure and uniform weld geometric deviation described in this invention; Figure 3 This is a schematic diagram of the thin-plate model discretization and graph topology model construction process described in this invention; Figure 4 This is a structural block diagram of the graph attention prediction model described in this invention; Figure 5 This is a flowchart of the sample dataset construction, model training, and rapid prediction process described in this invention. Detailed Implementation

[0014] To enable those skilled in the art to clearly understand the implementation process of the present invention, the technical solution of the present invention will be further described below in conjunction with specific embodiments. This embodiment takes the post-weld deformation prediction of a butt-welded structure of two thin plates under the conditions of uniform assembly gap and uniform surface difference as the application scenario, and illustrates how to achieve rapid prediction of the final deformation of the thin plate after welding through graph topology modeling, graph attention encoding, global condition injection, multi-task decoding, and joint loss training.

[0015] In the stage of defining the welding objects and deviation parameters for thin plates, two thin plates to be welded are selected to form a butt weld structure. The assembly state near the weld is described by two parameters: assembly gap and surface difference. The assembly gap represents the width of the horizontal gap between the two thin plates at the weld joint, and the surface difference represents the degree of misalignment of the base materials on both sides of the weld in the normal direction. This embodiment is for the case of uniform weld geometric deviation, that is, the assembly gap and surface difference are kept constant along the entire weld path. Therefore, the assembly gap and surface difference can be used as two global scalar inputs to the model. With this setting, the pre-weld assembly error is explicitly incorporated into the welding deformation prediction process.

[0016] In the discretization stage of the thin plate model, the thin plate geometric model is transformed into discrete nodes and element connection relationships. First, the node coordinates in the thin plate finite element model are obtained, and the three-dimensional spatial coordinates of each node are used as the initial node features of the graph neural network. Then, the node adjacency relationships are extracted based on the edge relationships between nodes in the finite element elements, and the node pairs that together constitute the edges of the finite element elements are transformed into undirected edges in the graph structure. In this way, the thin plate structure is transformed from a continuous geometric model into graph topological data that can be processed by the graph neural network.

[0017] In the graph topology construction phase, for a thin plate model with a two-layer node distribution, the upper and lower layer nodes are each constructed as node sets in the graph. In-plane connections are established between adjacent nodes in the same layer, and inter-layer connections are established between corresponding nodes in the upper and lower layers. In-plane connections represent local geometric relationships within the thin plate plane, while inter-layer connections represent thermal response coupling along the thickness direction. This graph topology construction method preserves the physical continuity of the finite element model, enabling subsequent graph attention models to propagate spatial information between nodes.

[0018] In the training sample construction phase, the reliability of the thermo-elastic-plastic finite element model was first verified through real welding experiments. Q235 steel thin plates were used in the experiment; both plates were 200 mm × 150 mm in size and 1 mm thick. The two workpieces were butt-welded along their long sides, with a weld length of 200 mm. A dedicated mechanical fixture was used to position and hold the workpieces, and the assembly gap and surface difference were controlled using a shim method. A structural blue light 3D scanner was used to measure the morphology of the thin plates before and after welding, and an X-ray stress meter was used to measure the internal stress. The welding process employed pulsed laser welding with a scanning speed of 20 mm / s, a scanning width of 1.8 mm, a laser power of 500 W, a duty cycle of 100%, and a pulse frequency of 2000 Hz.

[0019] In the finite element simulation phase, a welding thermo-mechanical coupling model consistent with experimental conditions was established based on ABAQUS 6.14. The material parameters, boundary conditions, and heat source parameters in the model were set according to the actual experimental conditions. The thermal efficiency coefficient was set to 0.75 in the finite element simulation, the welding process duration was 10 s, and the fixture constraints were released 10 s after the welding thermal cycle ended to simulate the unloading process in actual production. The numerical results after welding and sufficient cooling for 1000 s were selected as the final post-weld response, outputting the welding deformation and equivalent residual stress of the thin plate joints. By comparing the results of four sets of real welding experiments with the finite element simulation results, the simulation model was verified to provide a reliable source of samples for subsequent graph neural network training.

[0020] During the dataset generation phase, after verifying the finite element simulation model, a total of 500 welding simulation analyses were conducted. Assembly deviations were set for each simulation, with assembly gaps ranging from 0 mm to 3 mm and assembly surface differences ranging from 0 mm to 2 mm. Except for geometric deviations, the welding process configuration and mesh generation scheme remained consistent for each simulation. After each simulation was completed, the nodal welding deformation data and nodal equivalent stress data after cooling to room temperature were exported as input-output samples for training the graph neural network model.

[0021] During the dataset partitioning phase, the 500 simulation samples were divided into training, validation, and test sets in a 7:1.5:1.5 ratio, with 350 sets in the training set, 75 sets in the validation set, and 75 sets in the test set. The training set was used for updating model parameters, the validation set was used to monitor generalization performance during training and assist in hyperparameter tuning, and the test set was used to evaluate the final prediction accuracy of the model. The input for each sample set included node spatial coordinates, graph topology connectivity, assembly gaps, and surface difference parameters, and the output included the final deformation and residual stress of each node after welding.

[0022] In the graph attention encoding stage, a GAT encoder is constructed to extract features from the thin-plate graph structure. The GAT encoder consists of three graph attention convolutional layers, with hidden feature dimensions of 64, 128, and 256 for each layer. Each layer is configured with four attention heads, and the activation function is ReLU with a negative slope of 0.2. The graph attention encoder takes node spatial coordinates and graph adjacency relationships as input and adaptively calculates the contribution weights of different neighboring nodes to the target node through a multi-head attention mechanism, thereby obtaining the spatial structure embedding representation of each node. This node embedding representation includes the node's spatial location, local connectivity patterns, and multi-hop neighborhood structure information.

[0023] In the global condition injection stage, the assembly gap and surface difference, two scalar parameters, are input into the model as global deviation conditions. Specifically, after linear mapping expansion, these two deviation parameters are concatenated with the node embedding representation output by the GAT encoder, resulting in a feature dimension of 258. Since the assembly gap and surface difference under the same working conditions are shared by all nodes, this condition injection method enables all nodes to form differentiated response expressions based on their own spatial structural features under unified deviation conditions. This allows the model to perceive the impact of uniform weld geometric deviations on the post-weld deformation of the entire plate.

[0024] In the graph attention decoding stage, the embedded representations of the conditionally injected nodes are input into the GAT decoder. The GAT decoder also employs a three-layer graph attention structure, with hidden dimensions set to 128, 64, and 32 respectively. Unlike the node-by-node independent decoding method, the GAT decoder continues to utilize graph adjacency relationships for information propagation between nodes during the prediction process, maintaining spatial coherence in the prediction results between adjacent nodes. After the decoder outputs the node decoding features, it is further input into a bi-branch prediction structure to obtain the final post-weld deformation prediction result and the post-weld residual stress prediction result, respectively.

[0025] In the multi-task prediction stage, a deformation prediction branch and a residual stress prediction branch are set at the end of the decoder. Both branches adopt a two-layer multilayer perceptron with a hidden layer dimension of 16 and share the node feature representations extracted by the graph attention decoder. Specifically, the deformation prediction branch outputs the final deformation of each node after welding, and the residual stress prediction branch outputs the residual stress of each node after welding. Since both the post-weld deformation and residual stress originate from the welding thermo-mechanical coupling process, the residual stress prediction branch participates in model optimization as an auxiliary task during the training phase to encourage the shared feature layer to learn a more physically meaningful intermediate representation.

[0026] During the joint loss training phase, a joint loss function is constructed using a deformation prediction loss term and a residual stress prediction auxiliary loss term. Both the deformation prediction loss term and the residual stress prediction loss term are constrained by mean squared error, and the auxiliary task weight coefficient is set to 0.2 to ensure that the model training focuses on predicting the final deformation after welding, while leveraging the physical correlation information contained in the residual stress field to enhance the model's feature learning ability. During training, a Dropout regularization mechanism with a dropout rate of 0.1 is applied after each attention layer to suppress model overfitting.

[0027] During the model parameter optimization phase, the Adam optimizer was used to perform gradient descent with an initial learning rate of 0.001. A cosine annealing learning rate scheduling strategy was introduced to dynamically adjust the learning rate in the later stages of training. The model training batch size was set to 16, the maximum number of iterations was set to 800, and an early stopping mechanism was implemented to terminate training when the validation set loss did not decrease for 50 consecutive iterations. The relevant training and inference tasks were executed on a computing platform equipped with an Intel Core i9-11900H processor and an Nvidia GeForce RTX4090 graphics card. The model was implemented using the PyTorch and PyTorch Geometric libraries.

[0028] During the model inference phase, the topological data of the thin plate under the target working condition, the spatial coordinates of the nodes, the assembly gap, and the surface difference parameters are input into the trained model. The model outputs the predicted final deformation value of the corresponding node after welding, and can simultaneously output the predicted value of residual stress. In this embodiment, after the model training is completed, the inference time for a single set of samples is less than 0.2 seconds. Compared with the traditional finite element simulation that takes about 4 hours for a single working condition, this method can significantly improve the computational efficiency of thin plate welding deformation prediction under multi-deviation working conditions.

[0029] In the real-world testing phase, four sets of actual welding experimental measurement data that were not used in model training were employed as evaluation benchmarks. The assembly gap and surface difference corresponding to each set of experiments were used as model inputs to obtain the predicted deformation values ​​for each node after welding, which were then compared node-by-node with the deformation data measured by a blue light scanner. Experimental results show that the model can reflect typical welding deformation characteristics such as symmetrical deformation on both sides of the weld, a large deformation gradient in the weld area, and a gradual reduction in deformation away from the weld. The average absolute errors of the four sets of experiments were 0.24 mm, 0.31 mm, 0.35 mm, and 0.25 mm, respectively, while the maximum absolute errors were 1.12 mm, 1.48 mm, 1.32 mm, and 1.38 mm, respectively.

[0030] In the comparative verification phase, the graph attention network model described in this invention was compared with Kriging, support vector regression, MLP, and GCN models. Test results show that the MAE, MSE, and MaxAE of the GAT model are 0.29, 1.31, and 1.48, respectively; the SVR model is 0.41, 1.96, and 1.92; the Kriging model is 0.46, 2.21, and 2.05; the MLP model is 0.52, 2.68, and 2.23; and the GCN model is 0.37, 1.62, and 1.71. Therefore, this invention, by introducing graph topology modeling, graph attention differential aggregation, uniform deviation condition injection, and residual stress-assisted learning, can improve the prediction accuracy of thin plate welding deformation and enhance the spatial continuity and physical rationality of the prediction field.

Claims

1. A graph neural network prediction method for thin plate welding deformation considering uniform weld geometric deviations, characterized in that, The method includes: acquiring a set of discrete nodes and a set of connecting edges of a thin-plate welding object, and constructing a thin-plate graph structure based on the spatial coordinates and adjacency relationships of the nodes; acquiring assembly gap parameters and surface difference parameters characterizing the weld assembly state, and combining the assembly gap parameters and surface difference parameters into a global deviation condition; inputting the thin-plate graph structure into a graph attention encoder to obtain a spatial structure embedding representation of each discrete node; injecting the global deviation condition into the spatial structure embedding representation of each discrete node to obtain a conditional node representation; inputting the conditional node representation into a graph attention decoder to output the predicted final deformation value after welding for each discrete node, and outputting the predicted residual stress value after welding for each discrete node; constructing a joint loss function based on the predicted final deformation value after welding and the predicted residual stress value after welding, and training or updating the graph attention encoder and the graph attention decoder.

2. The method for predicting welding deformation of thin plates according to claim 1, characterized in that, The construction of the thin plate diagram structure includes: mapping the geometric nodes in the finite element mesh to vertices in the diagram, mapping the node pairs that together constitute the edges of the finite element unit to undirected edges in the diagram, and setting interlayer connection edges in the thickness direction of the thin plate to preserve the geometric association in the thin plate surface and the thermal response coupling relationship in the thickness direction.

3. The method for predicting welding deformation of thin plates according to claim 1, characterized in that, The initial input feature of each discrete node is the three-dimensional spatial coordinates of the node. The assembly gap parameter is used to characterize the width of the horizontal gap between the two thin plates at the weld joint. The surface difference parameter is used to characterize the degree of misalignment of the base materials on both sides of the weld in the normal direction.

4. The method for predicting welding deformation of thin plates according to claim 1, characterized in that, The graph attention encoder adopts a multi-layer graph attention convolution structure. By normalizing the attention coefficients between the target node and its neighboring nodes, it achieves differentiated aggregation of the contributions of different neighboring nodes, so as to obtain a node embedding representation containing local spatial features and multi-hop neighborhood topology information.

5. The method for predicting welding deformation of thin plates according to claim 1, characterized in that, The step of injecting global deviation conditions into the spatial structure embedding representation of each discrete node includes: performing a linear mapping on the assembly gap parameter and surface difference parameter, and splicing the mapping result with the spatial structure embedding representation of each discrete node, so that all discrete nodes share the same set of global deviation conditions.

6. The method for predicting welding deformation of thin plates according to claim 1, characterized in that, The graph attention decoder adopts a multi-layer graph attention convolution structure and sets a deformation prediction branch and a residual stress prediction branch at the output end. The deformation prediction branch is used to output the final post-weld deformation prediction value of each discrete node, and the residual stress prediction branch is used to output the post-weld residual stress prediction value of each discrete node.

7. The method for predicting welding deformation of thin plates according to claim 1, characterized in that, The joint loss function is obtained by weighted summation of the deformation prediction loss term and the residual stress prediction loss term. Both the deformation prediction loss term and the residual stress prediction loss term are constructed using mean square error, and the weight coefficient of the residual stress prediction loss term is less than 1.

8. The method for predicting welding deformation of thin plates according to claim 1, characterized in that, Before training or updating the graph attention encoder and the graph attention decoder, the method further includes: generating training samples based on a thermo-elastic-plastic finite element simulation model consistent with actual welding conditions. The training samples include nodal equivalent stress data and nodal welding deformation data under different assembly gap parameters and surface difference parameters.

9. A thin plate welding deformation prediction system considering uniform weld geometric deviation, characterized in that, The system includes a graph structure construction module, a deviation condition acquisition module, an encoding module, a condition injection module, a decoding module, and a training module. The graph structure construction module constructs a thin-plate graph structure based on a set of discrete nodes and a set of connected edges. The deviation condition acquisition module acquires assembly gap parameters and surface difference parameters. The encoding module extracts node spatial structure features using a graph attention encoder. The condition injection module injects assembly gap parameters and surface difference parameters into the node spatial structure features. The decoding module outputs the predicted final deformation and predicted residual stress after welding for each discrete node. The training module trains or updates the system based on a joint loss function.

Images