Deep learning based method for grid adaptation of docking joints and storage medium

Abstract

The application provides a kind of based on deep learning's butt joint grid self-adapting method and storage medium, the method first obtains the geometric parameter and load condition of butt joint, constructs parameterized geometric model;Initial background grid is generated based on geometric model, strain energy distribution is obtained by finite element calculation, according to which the grid element size is determined and high-precision reference non-uniform grid is iteratively generated, node feature data is extracted and training data set is constructed;Subsequently, the geometric parameters, load conditions and position information are input, and the grid element size is output to train the deep learning model;Finally, input new parameters to predict grid size, generate self-adaptive non-uniform grid and complete finite element calculation and visual output.The application first determines the grid size by finite element calculation of strain energy, and then learns the mapping relationship between geometry, load, position and grid size using deep learning;Subsequently, the grid is directly predicted using the model to achieve rapid self-adaptation.

Description

Technical Field

[0001] This invention relates to the field of computational mechanics research technology, and more specifically, to a method and storage medium for mesh adaptation of docking joints based on deep learning. Background Technology

[0002] Ship structures contain numerous butt joints, which are often stress concentration points and mechanically weak points in the overall structure. Manufacturing defects and localized stress concentrations can easily lead to joint failure and damage. Traditional methods of verifying butt joint strength through physical experiments suffer from problems such as high destructiveness, high testing costs, and long testing cycles. With the rapid development of scientific computing technology, finite element simulation has become an effective alternative to physical experiments. It offers advantages such as low cost and high efficiency, and can intuitively present the stress distribution patterns of the structure, providing a reliable basis for structural design and optimization. Currently, the finite element simulation process for butt joints is quite cumbersome: engineers must first construct a geometric model in 3D modeling software, then import the model into finite element software to complete mesh generation, material parameter and boundary condition settings, and finally perform numerical calculations and visualize the results using specialized post-processing software. The entire simulation process is time-consuming and highly dependent on the engineer's professional skills and engineering experience, making it difficult to use. Therefore, it is necessary to propose a parametric butt joint simulation method that integrates geometric modeling, finite element simulation, and post-processing visualization, allowing users to complete the entire simulation process with a single click by inputting key parameters.

[0003] Patent application number 202310882839.1 discloses a finite element simulation method, system, medium, and device based on deep learning neural networks. This technology constructs the finite element mesh shape function into a neural network form and optimizes the network parameters using the minimum potential energy principle. During the solution process, it automatically optimizes the node coordinates to achieve mesh r-adaptation, achieving higher solution accuracy even with a sparse initial mesh, effectively improving the stability and accuracy of finite element calculations. However, this method belongs to a general finite element solution framework and does not involve dedicated parametric modeling. Patent application number 202510546771.9 discloses a deep learning-based method for T-joint guided mesh adaptation. This method constructs a T-joint model through parametric modeling, uses a neural network to predict mesh cell sizes to guide the generation of non-uniform meshes, and integrates finite element solution and visualization output, simplifying the joint simulation process and improving efficiency. It effectively solves the problems of high cost and long time consumption in traditional posterior error estimation mesh adaptation. In a T-joint, the two plates intersect approximately perpendicularly, and the weld cross-section is usually geometrically asymmetrical, which can easily lead to local stress concentration at the intersection of the plates. In contrast, in a butt joint, the end faces of the two plates face each other and symmetrical bevels are often used. Local stress concentration may occur at the weld toe (the junction of the base metal and the weld) and the reinforcement transition zone. Therefore, this technical solution is not applicable to butt joints.

[0004] Therefore, there is an urgent need for a mesh adaptive simulation method that is suitable for butt joints, highly efficient, highly accurate, and easy to operate. Summary of the Invention

[0005] In view of this, the present invention aims to propose a deep learning-based mesh adaptation method for docking joints to solve the problems of cumbersome traditional finite element simulation processes for docking joints, reliance on manual experience for modeling and mesh generation, and the need for repeated posterior error estimation for mesh adaptation, which leads to high computational costs and low simulation efficiency.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0007] This invention provides a deep learning-based method for adaptive meshing of docking joints, comprising the following steps:

[0008] Step S1: Obtain the geometric parameters and load conditions of the mating joint, and construct a parametric geometric model;

[0009] Step S2: Generate an initial background mesh based on the geometric model, obtain the strain energy distribution of the nodes through finite element calculation, and determine the reference mesh element size of each node according to the reciprocal of the strain energy. Iteratively generate a reference non-uniform mesh that meets the preset accuracy requirements; extract the feature data of the reference non-uniform mesh nodes and construct a training dataset; wherein, the feature data includes geometric parameters, load conditions, position information and the reference mesh element size corresponding to the node;

[0010] Step S3: Using the geometric parameters, position information, and load conditions as inputs, and the reference mesh cell size at the corresponding position as output, train the model using the training dataset to construct a deep learning model;

[0011] Step S4: Input new geometric parameters and load conditions into the trained deep learning model to predict the mesh cell size of the corresponding node and drive the mesh generator to construct an adaptive non-uniform mesh.

[0012] This invention directly predicts the mesh cell size using a deep learning model, replacing the traditional iterative calculation of posterior error, significantly reducing the cost of mesh adaptation and improving the efficiency of finite element simulation.

[0013] Furthermore, in step S1, the weld area of ​​the parameterized geometric model is geometrically represented using a segmented polyline form.

[0014] This invention defines the geometric model of the welding area of ​​the butt joint through parametric geometric modeling. On the cross section perpendicular to the length of the weld, a segmented symmetrical polygonal line is used to represent the weld contour, clarifying the weld length, angle, and the dimensional ratio between the plate and the weld, thereby realizing the rapid construction and parametric adjustment of the butt joint structure.

[0015] Furthermore, in the cross-section perpendicular to the weld length direction, the weld contour of the butt joint is represented by five symmetrical broken lines, including a middle horizontal line segment and left and right side broken lines symmetrically arranged on both sides of the middle horizontal line segment; the left and right side broken lines respectively include a first broken line segment close to the middle horizontal line segment and a second broken line segment away from the middle horizontal line segment; the angle between the first broken line segment and the horizontal direction is β, and the angle between the second broken line segment and the horizontal direction is α; wherein, tanβ=2tanα and 2>tanα>0.5, 4>tanβ>1.

[0016] This invention uses a five-segment symmetrical broken line to accurately characterize the cross-sectional shape of the weld, which can better fit the structural morphology of the real welded joint and ensure the accuracy of finite element calculation. By setting the angle constraint of tanβ=2tanα, the weld contour can be made smooth, continuous and symmetrical, which conforms to the geometric characteristics of the actual weld bevel. At the same time, the geometric parameters are simplified to a single independent angle, reducing the number of input parameters and facilitating the training and convergence of the deep learning model.

[0017] Furthermore, in step S1, the geometric parameters of the butt joint include the weld width L, the tangent of the angle between the first segment of the broken line and the horizontal direction tanβ, the tangent of the angle between the second segment of the broken line and the horizontal direction tanα, and the ratio K of the plate thickness to the weld width; wherein, the weld width L is 0.2~1m, and the ratio K of the plate thickness to the weld width is 0.2~5.

[0018] In this invention, the thickness of the plate is preferably 1~3 μm, and the ratio K of the plate thickness to the weld width is preferably 0.5~2.

[0019] Furthermore, the node position information is characterized using 17-dimensional median coordinates, which include 16-dimensional planar median coordinates on the cross section perpendicular to the weld length direction and 1-dimensional depth coordinates along the weld length direction. The 16-dimensional planar median coordinates are calculated based on the polygonal region enclosed by the five symmetrical polygonal weld contours and the plate boundary within the butt joint cross section, and are used to characterize the relative topological position of the node within the cross section. The 1-dimensional depth coordinates are the axial normalized distance of the node along the weld extension direction.

[0020] This invention uses 17-dimensional median coordinates to describe node positions, accurately characterizing the topological and spatial positions of nodes within the joint, enabling deep learning models to more accurately learn the mapping relationship between position and grid size, thus improving prediction accuracy.

[0021] Furthermore, the load condition is to apply a fixed constraint to the left side of the docking joint model and apply an axial tensile load F to the right side, wherein the value of the axial tensile load F is 0~10000.

[0022] In this invention, the deep learning model is a neural network built using the Python libraries TensorFlow and Keras. The input includes geometric parameters, position information, and load parameters, and the output is the grid cell size.

[0023] In step S2 of the present invention, the size of the grid cell is defined based on the strain energy distribution, wherein the size of the grid cell is inversely proportional to the strain energy; the number of grid cells is adjusted to a preset range through a loop program.

[0024] Furthermore, in step S3, the deep learning model adopts a 7-layer fully connected neural network, with each layer having 32, 64, 128, 64, 32, 8, and 1 neurons respectively; the mean squared error loss function is used during training, and the Adam optimizer is used for parameter optimization.

[0025] In this invention, the training data for the deep learning model is generated in the following manner:

[0026] A docking joint model that generates random geometric model parameters;

[0027] The strain energy distribution is obtained by performing finite element analysis on a uniform background mesh.

[0028] A non-uniform mesh is generated based on the strain energy distribution, and the node location information and the corresponding mesh cell size are extracted as training data.

[0029] In this invention, in step S4, the obtained adaptive non-uniform mesh is input into the finite element solver for calculation, and the calculation results are post-processed and visualized. The mesh generator is Tetgen, which is used to generate a complete adaptive non-uniform mesh based on the predicted mesh element size distribution and the background mesh. The finite element solver is FreeFem++, which is used to perform finite element calculations on the adaptive non-uniform mesh, solve the linear elastic Lamé equations, and support users to input material parameters and boundary conditions in the interactive interface. The post-processing visualization is implemented through Paraview software, and the output includes Mises stress, X-direction normal stress, and strain energy density distribution map.

[0030] It should be noted that Mises stress is the equivalent stress used to comprehensively assess whether a metal structure has undergone plastic deformation and fracture failure; the X-direction normal stress is used to characterize the magnitude of the force along the axial tensile direction of the butt joint, and can directly reflect the risk of tensile failure in the weld area. Combining these two factors allows for a comprehensive evaluation of the mechanical properties and safety margin of the butt joint.

[0031] In this invention, in step S2, the number of reference grid cells is adjusted to a preset range through iterative looping, and the size of the reference grid cells corresponding to each node is extracted; a training dataset containing 500 sets of randomized parameterized docking joint samples is constructed, wherein a single node corresponds to one training data point.

[0032] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method described in the above technical solution.

[0033] Compared with existing technologies, the deep learning-based mesh adaptive docking joint method and storage medium described in this invention have the following advantages:

[0034] (1) The present invention uses parametric geometric modeling to realize the full automation of the simulation process. Users only need to input key geometric and load parameters to complete modeling, mesh generation, finite element calculation and post-processing, which significantly reduces the dependence on professional simulation experience. The deep learning model directly predicts the mesh element size distribution, replacing the repeated iteration of traditional posterior error estimation, which greatly improves the mesh generation efficiency and calculation speed.

[0035] (2) The present invention uses five-segment symmetrical broken lines to accurately characterize the weld cross-section shape, and uses tanβ=2tanα for geometric constraint to make the weld contour fit the real welding groove shape. While ensuring the accuracy of finite element calculation, it simplifies independent parameters and improves the training stability and prediction effect of deep learning model. It uses 17-dimensional median coordinates to accurately describe the node position information, which effectively improves the model's prediction accuracy and generalization ability for mesh size.

[0036] (3) This invention uses mature open-source tools such as Tetgen, FreeFem++, and Paraview to build a complete simulation process, which is highly compatible and easy to deploy and apply in engineering. Users only need to input key simulation parameters to complete the entire process of modeling, mesh generation, finite element calculation and result visualization with one click, which greatly reduces the threshold for use and enables people without finite element simulation experience to complete the simulation analysis of the docking joint efficiently and conveniently. Attached Figure Description

[0037] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0038] Figure 1 This is a schematic diagram of the mating joint described in this invention;

[0039] Figure 2 A schematic diagram of a uniform mesh for a butt joint;

[0040] Figure 3 A schematic diagram of strain energy density distribution corresponding to a uniform grid at a butt joint;

[0041] Figure 4 A schematic diagram of a non-uniform mesh generated by a traditional method;

[0042] Figure 5 A schematic diagram of strain energy density distribution corresponding to non-uniform meshes generated by traditional methods;

[0043] Figure 6 This is a schematic diagram of the adaptive non-uniform mesh generated by the method described in this invention;

[0044] Figure 7 This is a schematic diagram of the strain energy density distribution corresponding to the adaptive non-uniform mesh generated by the method described in this invention.

[0045] Figure 8 This is a schematic diagram of α and β as described in the present invention, wherein the angle between the first broken line and the horizontal direction is β, and the angle between the second broken line and the horizontal direction is α;

[0046] Figure 9 This is a schematic diagram of the weld width and plate thickness according to the present invention, wherein the weld width is L and the plate thickness is D;

[0047] Figure 10 This is a schematic diagram of the 17-dimensional median coordinates described in this invention. Detailed Implementation

[0048] The present invention will be further described below with reference to specific embodiments. First, it should be noted that the data in the following experimental examples were obtained by the inventors through numerous experiments. Due to space limitations, only a portion of these data is shown in the specification, and those skilled in the art can understand and implement the present invention based on this data. These embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the contents of this invention, those skilled in the art can make various modifications or alterations to the invention, and these modifications or alterations also fall within the scope of protection of this application.

[0049] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0050] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0051] Finite element method (FEM) simulation is a core numerical method that discretizes a continuous physical structure into a finite number of simple elements and solves for the mechanical, thermal, and other physical field responses through numerical calculations. It can replace traditional physical experiments, providing a direct view of the distribution of structural stress, strain, and displacement, and has become the mainstream method for strength analysis of engineering structures. Mesh generation is one of the core technologies of FEM simulation; its function is to divide the continuous solution domain into a finite number of elements, achieving spatial discretization. Unstructured meshes are widely used in engineering simulations because they can be automatically generated and locally refined. Generally, the smaller the mesh element size, the higher the computational accuracy, but the computation time increases superlinearly with the number of elements. To balance computational accuracy and efficiency, non-uniform meshes are widely used; adaptive non-uniform meshes can be refined in critical areas and sparsed in non-critical areas. The current technical challenge lies in determining the optimal refinement location to achieve the highest computational efficiency. Conventional methods involve estimating the posterior error based on the FEM calculation results and refining the mesh in areas with large errors to achieve a uniform error distribution; however, this method requires completing the FEM solution first and then performing error analysis, and the computational cost of generating non-uniform meshes is extremely high.

[0052] This embodiment provides a deep learning-based method for adaptive meshing of docking joints, including the following steps:

[0053] Step S1: Obtain the geometric parameters and load conditions of the mating joint, and construct a parametric geometric model;

[0054] Step S2: Generate an initial background mesh based on the geometric model, obtain the strain energy distribution of the nodes through finite element calculation, and determine the reference mesh element size of each node according to the reciprocal of the strain energy. Iteratively generate a reference non-uniform mesh that meets the preset accuracy requirements; extract the feature data of the reference non-uniform mesh nodes and construct a training dataset; wherein, the feature data includes geometric parameters, load conditions, position information and the reference mesh element size corresponding to the node;

[0055] Step S3: Using the geometric parameters, position information, and load conditions as inputs, and the reference mesh cell size at the corresponding position as output, train the model using the training dataset to construct a deep learning model;

[0056] Step S4: Input new geometric parameters and load conditions into the trained deep learning model to predict the mesh cell size of the corresponding node and drive the mesh generator to construct an adaptive non-uniform mesh.

[0057] This embodiment obtains a large number of samples through finite element simulation, uses the traditional adaptive mesh result as the standard answer, and trains a deep learning model to fit the above input-output mapping relationship, thereby realizing the direct and fast prediction of mesh size and the completion of adaptive mesh generation under new working conditions without repeating posterior error estimation.

[0058] It should be noted that, as Figure 1 As shown, the butt joint described in this embodiment is a welded joint formed by two plates with their end faces facing each other and connected by a weld. It mainly bears axial loads and is a typical load-bearing component in the fields of shipbuilding and steel structures.

[0059] like Figure 8 As shown, in this embodiment, the weld contour of the butt joint on a cross section perpendicular to the weld length direction is represented by five symmetrical broken lines, including a middle horizontal line segment and left and right side broken lines symmetrically arranged on both sides of the middle horizontal line segment; the left and right side broken lines respectively include a first broken line segment close to the middle horizontal line segment and a second broken line segment away from the middle horizontal line segment; the angle between the first broken line segment and the horizontal direction is β, and the angle between the second broken line segment and the horizontal direction is α; wherein, tanβ=2tanα and 2>tanα>0.5, 4>tanβ>1.

[0060] In existing technologies, butt joint welds are mostly characterized by simple straight lines, single bevels, conventional grooves, or complex curved surfaces, without a dedicated five-segment symmetrical polygonal line parameterization method. This invention uses a five-segment symmetrical polygonal line to represent the weld section, which can not only describe the geometry of the butt joint concisely and accurately and satisfy the force symmetry characteristics, but also form a regular polygonal region to calculate the 17-dimensional median coordinates. At the same time, it facilitates parametric modeling and automated mesh generation, providing stable input and standardized boundaries for deep learning mesh prediction.

[0061] like Figure 10 As shown, in this embodiment, the node position information is represented by 17-dimensional median coordinates. The 17-dimensional median coordinates include 16-dimensional planar median coordinates on the cross section perpendicular to the weld length direction and 1-dimensional depth coordinates along the weld length direction. The 16-dimensional planar median coordinates are calculated based on the polygonal region enclosed by the five symmetrical broken line weld contours and the plate boundary within the butt joint cross section, and are used to represent the relative topological position of the node within the cross section. The 1-dimensional depth coordinates are the axial normalized distance of the node along the weld extension direction.

[0062] Median coordinates are a type of generalized centroid coordinates that can represent any point within a planar polygonal region as a weighted combination of the vertices of the polygon's boundary. The coordinate values ​​satisfy properties such as smoothness, non-negativity, piecewise linearity, and interpolation boundary. They can be stably calculated for any simple convex / non-convex polygon and are the standard method for describing the relative position of a point with respect to a region in geometric modeling and numerical computation.

[0063] This embodiment uses 17-dimensional median coordinates as the relative position encoding of node locations. The cross-sectional portion uses 16-dimensional planar median coordinates, calculated from the polygon enclosed by five symmetrical polygonal weld seams and the plate boundary. The depth direction uses 1-dimensional weld extension direction normalized coordinates. This coordinate system can uniformly describe butt joints of different sizes and proportions, ensuring consistent feature representation for the same critical point / stress concentration point on the structure, thereby significantly improving the generalization ability of the deep learning model. Compared with rectangular coordinates (x, y, z), median coordinates do not depend on global position and mesh topology, do not require manual feature extraction, and can be directly used as standardized input for neural networks. They are more robust and can provide stable and reliable positional features for mesh size prediction, making them key to achieving high-precision, automated mesh adaptation.

[0064] It should be noted that, in this embodiment, the input to the deep learning model includes three parts:

[0065] Geometric parameters: weld width L, the ratio K of plate thickness D to weld width W (e.g., Figure 9 As shown in the figure, the angle between the first broken line and the horizontal direction is β, and the angle between the second broken line and the horizontal direction is α;

[0066] Node location information: It is characterized by 17-dimensional median coordinates, including 16-dimensional planar median coordinates on the cross section perpendicular to the weld length direction, and 1-dimensional depth-normalized coordinates along the weld length direction;

[0067] Load information: Axial tensile load F applied to the right side of the butt joint.

[0068] It should be noted that, in this embodiment, the value ranges and settings of key parameters during the construction of the parametric geometric model, generation of random finite element examples, and training dataset are as follows: geometric parameters include weld width L of 0.2~1m, plate thickness of 1~3m, plate thickness to weld width ratio K of 0.5~2, tanα of 0.5~2, and tanβ of 1~4; load parameters include axial tensile load F less than 10000; background mesh size and total number of non-uniform mesh elements are adaptively determined according to model size and computational resources; deep learning model training uses 30 iterations with a batch size of 128. All parameters are randomly sampled within their respective value ranges according to a uniform distribution to construct 500 sets of finite element examples and form a training dataset.

[0069] In this embodiment, the obtained adaptive non-uniform mesh is input to the finite element solver for calculation, and the calculation results are post-processed and visualized. The finite element solution, Tetgen mesh generation, Paraview post-processing, deep learning network structure and training method are all existing technologies, and will not be described in detail here.

[0070] Example 1

[0071] A deep learning-based method for adaptive mating joint mesh includes the following steps:

[0072] Step S1: Obtain the geometric parameters and load conditions of the butt joint, and construct a parametric geometric model. The butt joint is a welded joint formed by two plates with their end faces facing each other and connected by a weld, primarily bearing axial loads. On a cross-section perpendicular to the weld length, the weld profile is represented by five symmetrical broken lines, including a central horizontal line segment and two symmetrically distributed broken lines on either side. Specific geometric parameters include the weld width L, the tangents of the angles between the two broken lines and the horizontal direction (tanα and tanβ), and the ratio K between the plate thickness and the weld width. The load conditions are set as follows: a fixed constraint is applied to the left side of the butt joint, and an axial tensile load F is applied to the right side.

[0073] Step S2: Within the set range of geometric and load parameters, parameters are randomly generated using a uniform distribution, resulting in 500 sets of random finite element simulation examples. The computational domain of a single simulation example contains multiple nodes, with each node corresponding to one training data point.

[0074] To obtain the standard mesh element size required for training, this method generates a high-precision non-uniform mesh according to the following process: First, a uniform and relatively sparse initial background mesh is constructed; finite element analysis is performed based on this background mesh to obtain the strain energy distribution within the computational domain; the mesh element size is determined according to the reciprocal of the strain energy, i.e., the mesh element size is smaller where the strain energy is larger, and larger where the strain energy is smaller; then, a non-uniform mesh is generated based on the geometric boundaries and the above element size distribution rules. This method uses a loop program to determine the total number of elements in the current non-uniform mesh and compares it with a preset target number, repeatedly performing mesh refinement or sparsification operations until the total number of meshes meets the requirements, finally obtaining the reference non-uniform mesh generated by the traditional method.

[0075] Finally, the grid cell size corresponding to each node of the background grid is extracted. The node position information represented by 3 geometric parameters, 17-dimensional median coordinates, and 1 load parameter are used as the training data input, and the grid cell size at the corresponding position is used as the output to form a complete training dataset.

[0076] Step S3: A deep learning network was built using the Python libraries TensorFlow and Keras. The network structure was a 7-layer fully connected neural network, with the number of neurons in each layer being 32, 64, 128, 64, 32, 8, and 1, respectively. The mean squared error was used as the loss function during training, and the Adam optimizer was used for parameter updates. The model was trained for 30 epochs, with each batch containing 128 data points.

[0077] After training, adaptive mesh generation can be performed quickly for new working conditions: the geometric model of the docking joint is regenerated based on the size parameters input by the user, and a sparse and uniform background mesh is constructed on this geometry; the median coordinates and geometric and load parameters of each node of the background mesh are read to form the input tensor of the neural network; the tensor is input into the trained model, and the mesh size corresponding to each node position is directly output; finally, the mesh generator Tetgen generates a deep learning adaptive non-uniform mesh based on the background mesh and the cell size distribution predicted by the model.

[0078] Step S4: Using the trained deep learning model to predict the mesh size distribution under the new working condition, the mesh generator is driven to construct an adaptive non-uniform mesh. Then, the finite element solver is called for calculation and post-processing: After the user inputs key parameters and starts the calculation, the program automatically generates an executable file that FreeFem++ can recognize. The file contains boundary conditions and solution settings. FreeFem++ reads the file and combines it with the adaptive mesh generated by the neural network to quickly complete the finite element solution. After the calculation is completed, the program calls the open-source post-processing software Paraview to visualize the strain energy, mesh morphology, Mises stress, and normal stress in the x-direction, completing the entire simulation process.

[0079] Comparative Example 1

[0080] The butt joint model is meshed using a uniform-sized mesh, such as... Figures 2-3 As shown, the mesh element size remains consistent throughout the entire computational domain, without any local refinement, sparsification, or adaptive adjustment. The mesh generation process does not consider strain energy, stress gradient, or posterior error distribution; it only completes mesh construction with uniform element size, followed by conventional finite element solution and result output.

[0081] Comparative Example 2

[0082] Non-uniform meshes are generated using the traditional finite element adaptive meshing method, such as... Figures 4-5As shown, the specific process is as follows: finite element solution is performed based on the initial mesh, and the calculation error distribution is obtained through posterior error estimation; mesh refinement is performed on high strain energy and stress concentration areas according to the error magnitude, while a coarser mesh is maintained in low error areas; the mesh is continuously optimized through multiple iterations of "solution - error estimation - mesh refinement" until the accuracy requirements are met, and finally a traditional adaptive non-uniform mesh is generated.

[0083] As shown in Table 1 below, the energy densities of Comparative Example 2 and this embodiment were obtained using the above method under typical input parameters.

[0084] Table 1 Maximum Mises stress values ​​under typical input parameters.

[0085] Serial Number Weld width L / m tanα tanβ The ratio K of plate thickness to weld width Axial tensile load F / N Energy density of deep learning grids / Pa Energy density / Pa of traditional adaptive grid 1 1.23 0.17 0.34 2.24 1888 2.9 2.6 2 0.87 0.23 0.46 1.42 6451 12.5 10.8 3 2.56 0.45 0.90 1.88 4974 5.8 5.4

[0086] As can be seen from Table 1, the maximum stress obtained by the traditional adaptive non-uniform mesh used in Comparative Example 2 is similar to that of the present technical solution; at the same time, the mesh generation speed of Example 1 is about ten times faster than the traditional non-uniform mesh generation method of Comparative Example 2.

[0087] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various alterations and modifications without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A method for adaptive meshing of docking joints based on deep learning, characterized in that, Includes the following steps: Step S1: Obtain the geometric parameters and load conditions of the mating joint, and construct a parametric geometric model; Step S2: Generate an initial background mesh based on the geometric model, obtain the strain energy distribution of the nodes through finite element calculation, and determine the reference mesh cell size of each node according to the reciprocal of the strain energy. Iteratively generate a reference non-uniform mesh that meets the preset accuracy requirements. Extract the feature data of the reference non-uniform mesh nodes and construct a training dataset. The feature data includes geometric parameters, load conditions, location information, and the reference mesh cell size corresponding to the node. Step S3: Use the geometric parameters, location information, and load conditions as inputs, and the reference mesh cell size at the corresponding position as output, to train the model using the training dataset to construct a deep learning model. Step S4: Input new geometric parameters and load conditions into the trained deep learning model to predict the mesh cell size of the corresponding node and drive the mesh generator to construct an adaptive non-uniform mesh.

2. The method according to claim 1, characterized in that, In step S1, the weld region of the parametric geometric model is geometrically represented using a segmented polyline form.

3. The method according to claim 2, characterized in that, The weld contour of the butt joint is characterized by five symmetrical broken lines on a cross section perpendicular to the weld length direction, including a middle horizontal line segment and left and right side broken lines symmetrically arranged on both sides of the middle horizontal line segment; the left and right side broken lines respectively include a first broken line segment close to the middle horizontal line segment and a second broken line segment away from the middle horizontal line segment; the angle between the first broken line segment and the horizontal direction is β, and the angle between the second broken line segment and the horizontal direction is α; wherein, tanβ=2tanα and 2>tanα>0.5, 4>tanβ>1.

4. The method according to claim 3, characterized in that, In step S1, the geometric parameters of the butt joint include the weld width L, the tangent of the angle between the first segment of the broken line and the horizontal direction tanβ, the tangent of the angle between the second segment of the broken line and the horizontal direction tanα, and the ratio K of the plate thickness to the weld width; wherein the weld width L is 0.2~1m, and the ratio K of the plate thickness to the weld width is 0.2~5.

5. The method according to claim 4, characterized in that, In step S2, the position information is characterized using 17-dimensional median coordinates. The 17-dimensional median coordinates include 16-dimensional planar median coordinates on the cross section perpendicular to the weld length direction and 1-dimensional depth coordinates along the weld length direction. The 16-dimensional planar median coordinates are calculated based on the polygonal region enclosed by the five symmetrical polygonal weld contours and the plate boundary within the cross section of the butt joint, and are used to characterize the relative topological position of the node within the cross section. The 1-dimensional depth coordinates are the axial normalized distance of the node along the weld extension direction.

6. The method according to claim 1, characterized in that, The load condition is to apply a fixed constraint to the left side of the docking joint model and apply an axial tensile load F to the right side, wherein the value of the axial tensile load F is 0~10000N.

7. The method according to claim 1, characterized in that, In step S3, the deep learning model uses a 7-layer fully connected neural network, with each layer having 32, 64, 128, 64, 32, 8, and 1 neurons respectively; The mean squared error loss function is used during training, and the Adam optimizer is used for parameter optimization.

8. The method according to claim 1, characterized in that, In step S4, the obtained adaptive non-uniform mesh is input into the finite element solver for calculation, and the calculation results are post-processed and visualized. The mesh generator is Tetgen, the finite element solver is FreeFem++, and the post-processing visualization is implemented using Paraview software.

9. The method according to claim 1, characterized in that, In step S2, the number of reference grid cells is adjusted to a preset range through iterative looping, and the size of the reference grid cells corresponding to each node is extracted; a training dataset containing 500 sets of randomized parameterized docking joint samples is constructed, wherein a single node corresponds to one training data point.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 9.

Images