CAD model fillet feature recognition and removal method based on graph neural network

Abstract

The invention discloses a CAD model fillet feature recognition and removal method based on a graph neural network, and the method comprises the steps: extracting a curved surface from the boundary representation of a CAD model, and constructing a surface adjacent graph; extracting geometric features and attribute features of each curved surface, and constructing node feature vectors; performing dichotomy on each curved surface by using a graph neural network, and identifying whether the curved surface is a rounded surface or not; deleting all the fillet surfaces in the CAD model, gridding the remaining non-fillet surfaces one by one, and judging whether grid expansion, grid node index adjustment, original grid and expanded grid merging and synthesized new grid outputting are needed or not; and performing Boolean intersection and post-processing operation on the new grid to realize automatic removal of fillet features. The method is remarkably superior to an existing method in the aspects of fillet recognition accuracy, model efficiency and generalization ability, and is suitable for CAD model preprocessing in a complex engineering scene.

Description

Technical Field

[0001] This application relates to the fields of deep learning and computer graphics, and in particular to a method for recognizing and removing rounded corner features of CAD models based on graph neural networks. Background Technology

[0002] Fillets are a common geometric feature in computer-aided design (CAD) models, used to eliminate stress concentration and meet manufacturing or aesthetic requirements. However, before performing CAD engineering analysis, meshing of the CAD model is usually required, and the abundance of fillets significantly increases the complexity of mesh generation. Small-sized or complex intersecting fillets can lead to excessive local mesh refinement, increasing computational costs and potentially causing mesh quality degradation, numerical solution difficulties, and impacting simulation accuracy and convergence. Therefore, achieving automatic identification and removal of fillet features while maintaining geometric integrity is a key technical challenge for improving the efficiency of CAE analysis.

[0003] In existing technologies, methods for processing rounded corner features are mainly divided into three categories: 1. Rule-based traditional geometric methods: These methods rely on manually defined geometric attribute thresholds or predefined topological pattern matching to identify rounded corners, such as through curvature analysis or attribute adjacency graph matching. While these methods are effective in handling standard rounded corners, their recognition capability is severely limited by the predefined rule base. When faced with complex rounded corner interactions in industrial models, rule-based methods struggle to cover all situations, exhibiting poor robustness and versatility, and low automation.

[0004] 2. Discrete Mesh-Based Analysis Methods: These methods operate on a triangulated model, identifying rounded corner regions by estimating vertex curvature or analyzing normal variations. However, their effectiveness heavily depends on the quality of the input mesh and is sensitive to noise and variations in mesh density. More importantly, due to the loss of the original, accurate boundary representation information of the CAD model, their recognition accuracy and ability to handle complex features are significantly limited.

[0005] 3. Deep Learning-Based General Feature Recognition Methods: In recent years, deep learning-based feature recognition methods have brought new technical paths to CAD model understanding. However, existing general-purpose CAD feature recognition models aim to identify multiple feature types within a unified framework, failing to specifically optimize for the unique geometric and topological properties of transitional features such as fillets. This results in their accuracy and generalization ability in complex fillet recognition tasks failing to meet the requirements of high-precision CAE simplification.

[0006] In the process of realizing this invention, the inventors discovered at least the following problems in the prior art: 1. Insufficient recognition accuracy and robustness: Traditional methods rely on manual rules, which are difficult to adapt to complex and varied rounded corner structures; general deep learning models have limited ability to distinguish rounded corner features; 2. Weak generalization ability: Existing methods are sensitive to the distribution of training data, and their performance drops significantly when faced with unseen rounded corner types or complex topological interactions; 3. Limited level of automation: The entire process from identification to simplification has not yet achieved true end-to-end automation and still requires a lot of human intervention. Summary of the Invention

[0007] In view of this, embodiments of this application provide a method for identifying and removing rounded corner features of CAD models based on graph neural networks, in order to solve the problems existing in the background art.

[0008] According to an embodiment of this application, a method for identifying and removing rounded corner features in a CAD model based on a graph neural network includes: S1: Obtain the face adjacency graph constructed from the CAD model B-Rep representation; S2: Based on the face adjacency graph, extract the geometric features and attribute features of each surface to determine the node feature vector; S3: Based on the node feature vectors and the face adjacency graph, use a graph neural network to classify each surface and determine the rounded corner recognition result; S4: Based on the rounded corner recognition results, delete all rounded corners and triangulate the remaining non-rounded corners, determine whether mesh expansion is needed, adjust the mesh node index and merge the original mesh, and determine the output composite mesh. S5: Based on the synthesized new mesh, determine the final mesh model with rounded corner features removed through Boolean intersection and post-processing operations.

[0009] The technical solutions provided by the embodiments of this application may include the following beneficial effects: As can be seen from the above embodiments, this application achieves high-precision recognition of rounded corner features in CAD models by constructing a face adjacency graph and extracting the geometric and attribute features of the surface. Employing a specially designed graph neural network architecture and training on a self-built dataset containing various rounded corner types significantly improves the model's ability to recognize complex rounded corner topologies, effectively solving the bottleneck of traditional methods when dealing with complex rounded corners.

[0010] The method proposed in this application demonstrates excellent model efficiency, with the entire recognition network containing only 0.02M parameters, far fewer than existing state-of-the-art models, achieving a balance between high performance and low computational cost. This method maintains 100% accuracy in rounded corner recognition across multiple complex test cases, exhibiting strong generalization ability and adapting to diverse rounded corner structures in real-world industrial scenarios.

[0011] This application also constructs a large-scale, diverse benchmark dataset for rounded corner feature recognition, which not only fills the gap in existing data resources, but also provides a reproducible and fair evaluation platform for subsequent research in this field.

[0012] This application innovatively integrates AI recognition with geometric simplification, achieving an automated process from rounded corner identification to removal. Through boundary-based mesh expansion and Boolean intersection operations, precise removal of rounded corner regions is achieved while maintaining geometric continuity. This method requires minimal human intervention, significantly improving the automation level of CAE preprocessing and providing a reliable solution for industrial applications.

[0013] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0014] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0015] Figure 1 This is a text flowchart illustrating a method for recognizing and removing rounded corner features in CAD models based on a graph neural network, according to an exemplary embodiment.

[0016] Figure 2 This is a schematic diagram of a method for recognizing and removing rounded corner features of a CAD model based on a graph neural network, according to an exemplary embodiment.

[0017] Figure 3 This is a schematic diagram illustrating the calculation of surface attribute features of a CAD model according to an exemplary embodiment.

[0018] Figure 4 This is a schematic diagram illustrating three rounded corner types contained in a self-built rounded corner feature dataset according to an exemplary embodiment.

[0019] Figure 5 This is a schematic diagram illustrating a method for expanding the mesh on the boundary edge after removing rounded corners, according to an exemplary embodiment.

[0020] Figure 6 This is a model of a three-hole circular mechanical connection component shown according to an exemplary embodiment, demonstrating the rounded corner features in the model.

[0021] Figure 7 This is an example of the effect of rounded corner feature recognition on a three-hole circular mechanical connection component model, as shown in an exemplary embodiment.

[0022] Figure 8This is an example of the effect of removing the rounded corners of a three-hole circular mechanical connection component model, as shown in an exemplary embodiment.

[0023] Figure 9 This is a triangulated model of a three-hole circular mechanical connection component according to an exemplary embodiment, showing some mesh points on the boundary edges.

[0024] Figure 10 This is an example of a mesh expansion effect for a three-hole circular mechanical connection component model, as shown in an exemplary embodiment.

[0025] Figure 11 This is the Boolean intersection effect of a three-hole circular mechanical connection component model shown according to an exemplary embodiment.

[0026] Figure 12 This is a non-two-dimensional manifold triangle deletion effect shown in an exemplary embodiment of a three-hole circular mechanical connection component model.

[0027] Figure 13 This is the effect of removing the fillet features of a three-hole circular mechanical connection component model according to an exemplary embodiment.

[0028] Figure 14 This is a comparison diagram of the mesh quantity before and after using the present invention to identify and remove rounded corners, according to an exemplary embodiment.

[0029] Figure 15 This is a comparison chart showing the rounded corner feature recognition performance of commercial software SolidWorks, the latest academic research Defillet, and the present invention, according to an exemplary embodiment. Detailed Implementation

[0030] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0031] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0032] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0033] A CAD model is a digital three-dimensional solid representation created using computer-aided design software. It precisely defines the product's geometry, dimensions, structure, and assembly relationships between its parts. In this invention, we primarily deal with CAD models represented using B-Rep. The B-Rep model describes a complex three-dimensional solid as a closed shell composed of "faces," each face defined by "edges," and each "edge" defined by "vertices." This representation can accurately describe complex curved surface geometry and is the core data format of standard CAD systems in the industrial field. The fillet feature recognition and removal method provided in this invention has broad application prospects in multiple industrial fields and can significantly improve the automation level and efficiency of the entire process from design to analysis and manufacturing. 1. Computer-Aided Engineering Analysis (CAE): In modern engineering simulation workflows, the CAE preprocessing stage places strict requirements on model quality. Taking mainstream CAE preprocessing software such as ANSYS SpaceClaim and Altair HyperMesh as examples, they often suffer from low automation, reliance on manual selection, and insufficient accuracy and efficiency in recognizing complex or minute fillets when handling fillet features. Therefore, this invention can be integrated as a core functional module. Before engineers import a complex part model for simulation analysis, the software can utilize the automatic fillet recognition and removal function of this invention to simplify the model into a pure geometry more suitable for mesh generation. This directly avoids generating excessively dense meshes at minute fillets, reducing the number of meshes while ensuring computational accuracy, shortening solution time, and improving overall simulation efficiency.

[0034] 2. 3D Printing: In this field, the quality of model preprocessing directly affects the final printing effect. Taking professional 3D printing preprocessing software such as MaterialiseMagics as an example, its rounded corner processing function heavily relies on the operator's experience for manual repair and removal, which is time-consuming and labor-intensive, and prone to omissions or errors when dealing with complex topologies. Therefore, this invention can serve as a key tool for model repair and weight reduction. When preparing printing data, the software can use this invention to automatically identify rounded corner features in the model and remove them in batches. This effectively simplifies the model outline, thereby reducing material costs and shortening the overall printing time while ensuring forming accuracy.

[0035] 3. Numerical Control Machining (CAM): In this field, feature recognition accuracy directly affects machining quality and efficiency. Taking CAM software such as Siemens NX CAM as an example, relying on simple rule-based filtering or manual judgment not only prolongs machining time but also leads to the omission of internal fillet areas due to insufficient experience, causing overcutting or material residue problems in the finishing stage. Therefore, this invention can be deeply integrated with its intelligent programming system. When creating finishing toolpaths for complex structural parts, this invention can automatically and accurately identify all internal fillet areas in the model that require corner clearing operations. By calculating the fillet radius value, it automatically matches the tool with the optimal diameter. Subsequently, the CAM system can automatically generate corner clearing machining trajectories for these fillet areas based on this information, ensuring machining quality and significantly shortening the programming and calculation time in the finishing stage.

[0036] This invention will use a three-hole circular mechanical connection component model with rounded corners as an exemplary embodiment to explain the specific implementation. The method of this application is applicable to CAD models with rounded corners, but is not limited to the three-hole circular mechanical connection component model described below.

[0037] Figure 1 This is a flowchart illustrating a method for recognizing and removing rounded corner features in a CAD model based on a graph neural network, according to an exemplary embodiment. Figure 1 As shown, the method may include the following steps: S1: Obtain the face adjacency graph constructed from the CAD model B-Rep representation; Specifically, the B-rep data structure of the CAD model is parsed, all surfaces are traversed, and each surface is abstracted as a node in an undirected graph. Then, based on the topological relationships between surfaces in the B-rep, undirected edges between nodes are constructed, ultimately forming a complete graph structure. .

[0038] Taking a three-hole circular mechanical connection component model as an example, such as Figure 6 As shown, the model contains 57 parametric surfaces and 126 edges. In step S1, these 57 surfaces are transformed into 57 nodes in a graph structure, and the adjacency relationships between surfaces are transformed into connecting edges between nodes, thereby constructing the face adjacency graph of the connected component model, laying the topological foundation for subsequent feature extraction and graph neural network analysis.

[0039] S2: Based on the face adjacency graph, extract the geometric and attribute features of each surface to determine the node feature vector; this step includes the following sub-steps: S21: Based on the surfaces in the face adjacency graph, the ordered sampling points are determined by uniformly sampling over the UV parameter domain of the surfaces; Specifically, for each parametric surface, its parameter domain is first normalized to... Within the range. Then generate in the UV space at equal intervals. A sampling point, for example, when At that time, 25 uniformly distributed sampling locations were obtained. For each sampling point... The corresponding three-dimensional coordinates are calculated using the parametric equations of the surface. And its differential geometric properties. This sampling strategy ensures that surfaces of different sizes and shapes can obtain feature representations of a uniform dimension, providing a standardized input format for subsequent neural network processing.

[0040] S22: Based on the ordered sampling points, calculate the Gaussian curvature and average curvature of each sampling point to determine the constructed geometric feature vector; Specifically, at each sampling point, the first fundamental form coefficients of the surface are first calculated. E, F, G Second fundamental form coefficient L, M, N Then, according to the differential geometry formulas: Calculate the Gaussian curvature K and the mean curvature H. Then, apply the curvature values ​​of all sampled points according to... Arrange, form The geometric eigenvectors of a surface. Since curvature is an intrinsic geometric quantity of a surface, this characteristic means that the surface remains undeformed under rigid body transformations such as rotation and translation, and can stably distinguish rounded surfaces from other surfaces.

[0041] S23: Based on the geometric feature vector, extract the surface width and dihedral angle attribute features of adjacent surfaces to determine the final constructed node feature vector.

[0042] Specifically, such as Figure 3 As shown, the calculation of the surface width first involves extracting all the boundary edges of the surface, and then finding the two longest boundary edges by comparing their lengths. Calculate their respective midpoints and The three-dimensional coordinates are finally calculated. The face width value is obtained. The dihedral angle between adjacent faces is calculated by finding... For other adjacent surfaces, calculate the values ​​of these two adjacent surfaces respectively. and The normal vector at a given location is used, and the angle between two normal vectors is taken as the dihedral angle between adjacent faces. These two features describe the working characteristics of the surface from different dimensions: the face width reflects the size of the feature, and the dihedral angle describes the geometric environment in which the feature is located, together providing important auxiliary information for rounded corner recognition.

[0043] S3: Based on the node feature vectors and the face adjacency graph, classify each surface using a graph neural network to determine the rounded corner recognition result; this step includes the following sub-steps: S31: Design a network architecture that includes a preprocessing layer, an embedding layer, a graph convolutional layer, a fusion layer, and a classification layer, and determine the graph neural network model for rounded corner recognition; Specifically, such as Figure 2 As shown, the preprocessing layer consists of fully connected layers, responsible for preprocessing geometric and attribute features and mapping them to a 32-dimensional latent space. The graph convolution module adopts a three-layer graph convolution structure, with each layer outputting 64 dimensions and using the Tanh function as the activation function. The remaining modules in the network use the ReLU activation function. The fusion layer is used to integrate graph convolution features and attribute features to achieve effective information integration. The classification layer outputs the probability of binary classification for final prediction. During training, batch normalization is applied after each graph convolution layer to stabilize the training dynamics and accelerate convergence. In addition, by introducing a self-loop mechanism, it is ensured that nodes can retain their own features during information aggregation, thereby alleviating the oversmoothing problem.

[0044] S32: Based on the graph neural network model used for rounded corner recognition, train the network using a self-built labeled dataset. During the training process, use the Adam optimizer, L2 regularization, and Dropout strategy to determine the trained rounded corner recognition model. Specifically, the training dataset contains 4486 carefully annotated CAD models, where the rounded corner features are divided into three categories based on structural complexity: regular rounded corners with the same radius (all rounded corners are regular rounded corners with the same radius), regular rounded corners with different radii (all rounded corners are irregular rounded corners with different radii), and irregular rounded corners with varying radii (all rounded corners are variable radius rounded corners). Figure 4 As shown, the dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio. During training, the Adam optimizer, L2 regularization, and Dropout strategy were used to prevent overfitting and improve the model's generalization ability. After training, the parameters of the model that performed best on the validation set were saved and identified as the successfully trained rounded corner recognition model.

[0045] S33: Based on the trained rounded corner recognition model and the node feature vector, obtain the probability that each surface belongs to a rounded corner surface, and determine the rounded corner surface recognition result.

[0046] Specifically, the node feature vector is input into the trained rounded corner recognition model, and the model's classification layer outputs the probability that each surface belongs to a rounded corner surface or a non-rounded corner surface. Based on the probability, a corresponding classification label is assigned to each surface, and the classification labels of all surfaces are summarized to form a complete surface-level semantic segmentation result of the B-rep representation of the CAD model, thereby determining the final rounded corner recognition result.

[0047] Taking a three-hole circular mechanical connection component model as an example, after the rounded corner recognition model was trained and predicted, 29 of the 57 curved surfaces on the model were accurately identified. Figure 7 The visualization of the model's rounded corner recognition effect is shown: the green marked areas in the figure are all the rounded corners recognized by the model, and the blue marked areas are the non-rounded corners, which intuitively verifies the invention's ability to accurately recognize rounded corner features in complex parts.

[0048] S4: Based on the rounded corner recognition results, delete all rounded corners and triangulate the remaining non-rounded corners, determine whether mesh expansion is needed, adjust the mesh node indices, and merge the original meshes to determine the output composite mesh; this step includes the following sub-steps: S41: Based on the rounded corner recognition results, determine the new boundary edge generated at the original connection point by deleting all rounded corners; Specifically, based on the rounded corner recognition results, all surfaces identified as rounded corners are traversed, and these surfaces are removed from the model's topology. This operation generates boundary edges at the original locations where rounded and non-rounded corners connected, and these boundary edges accurately identify the geometric gaps that need to be repaired.

[0049] Taking a three-hole circular mechanical connection component model as an example, delete the 29 curved surfaces identified as rounded corners. For example... Figure 8 As shown, the model has many geometric gaps, and the outlines of these gaps are the newly generated boundary edges after the deletion operation.

[0050] S42: Based on the new boundary edge, traverse all remaining non-rounded faces. For non-rounded faces without boundary edges, perform triangulation directly. For non-rounded faces with boundary edges, first find the boundary edges with continuous positions on the curved surface to form a boundary edge group. After triangulation, locate all grid points on the boundary edge group and determine the grid points on the boundary edge group. Specifically, the system traverses all remaining non-rounded surfaces and categorizes them into two types based on whether the new boundary edge exists: for complete surfaces without a new boundary edge, triangular meshing is directly performed to generate the initial mesh of the surface; for surfaces with a new boundary edge, the boundary edges that are continuously arranged in space are first identified and extracted, and they are combined to define a boundary edge group. Then, the surface is triangularly meshed, and after the meshing is completed, the coordinates of all the mesh vertices on the boundary edge group are accurately located to determine the complete mesh point set of the boundary edge group.

[0051] Taking a three-hole circular mechanical connection component model as an example, Figure 9 The results of triangular meshing of the model after removing rounded corner features are shown. Some mesh points located on the boundary edges have been highlighted. These boundary point sets clearly define the geometric gap boundaries that need to be repaired later.

[0052] S43: Based on the grid points on the boundary edge group, calculate the offset direction by calculating the normal information of the triangular grid surface where the grid points are located and the tangent information of the grid edge, and determine the newly generated grid points; Specifically, such as Figure 5 As shown, for each boundary point P Find the boundary point P Given the boundary edges at the endpoints and the triangular facets containing those edges, based on the normal vector of the facets... N Tangent vector to the boundary edge T Through cross product operation Obtain the outward normal vector of the boundary edge of the mesh. D Boundary points P The external normal vector of all connected edges D average As a point P The offset direction. The offset distance d is set by the user to ensure that the expanded mesh fully covers the original rounded corner area to obtain new points. .

[0053] S44: Determine the constructed extended triangular mesh based on the newly generated mesh points and the mesh points on the original boundary edge group; Specifically, such as Figure 5 As shown, the original boundary edge points and newly generated points are used as the input point set, and the boundary edges are used as constraint edges to ensure that the generated triangular mesh perfectly connects with the original mesh at the boundary edges. During the merging process, the global indexes of the mesh points are first renumbered to avoid index conflicts. Then, a mapping relationship is established between the old and new mesh points to ensure the correctness of the topological connections. The final synthesized mesh is geometrically continuous, topologically consistent, and completely covers the space occupied by the original rounded corners.

[0054] S45: Based on the extended triangular mesh and the original mesh, integrate them through global node index reorganization to determine the final output composite new mesh.

[0055] Specifically, by reorganizing the global node index, the original meshes of each non-rounded face are integrated with all extended triangular meshes to ensure seamless connection between different extended mesh patches, thereby determining a new synthetic mesh that is geometrically and topologically complete and consistent.

[0056] Taking a three-hole circular mechanical connection component model as an example, Figure 10 The diagram illustrates the synthesized new mesh of the model. In this figure, the green mesh areas represent the newly generated triangular meshes through boundary expansion, which precisely fill the geometric gaps created at the original connections due to the removal of rounded faces; while the blue mesh areas represent the retained meshes obtained by directly subdividing the original non-rounded faces. After global indexing and reorganization, these two types of meshes form a complete, continuous, and closed synthesized new mesh, whose geometry has had its rounded features removed, preparing it for subsequent Boolean intersection operations.

[0057] S5: Based on the synthesized new mesh, determine the final mesh model with rounded corner features removed through Boolean intersection and post-processing operations. This step includes the following sub-steps: S51: Based on the synthesized new mesh, perform Boolean intersection operation and automatically detect and process the intersection relationship between different extended mesh patches, perform mesh trimming and reconstruction, and determine the mesh after Boolean intersection; Specifically, Boolean intersection calculation employs a precise geometric computation process: First, a spatial acceleration structure is established based on an AABB tree to quickly detect intersecting triangle pairs. For each pair of intersecting triangles, precise intersection points and lines are calculated using rational numbers to avoid floating-point errors. Then, the intersecting triangles are divided into several sub-triangles based on the intersection points and lines, ensuring that all intersection points are located at the endpoints of the new mesh edges, and that all intersection lines become edges of the new mesh. Finally, for coplanar intersection scenarios, duplicates in the common intersecting regions are removed to ensure the correctness and consistency of the Boolean intersection calculation results.

[0058] Taking a three-hole circular mechanical connection component model as an example, Figure 11 The results of Boolean intersection operations on the synthesized new mesh are shown. It can be seen that multiple extended mesh patches that originally intersected or overlapped have, after precise Boolean intersection operations, had their intersections automatically detected, re-trimmed, segmented, and reconstructed, forming a complete and unified mesh structure without internal interference. This crucial step eliminates geometric conflicts between different extended meshes, ensuring that the model maintains strict geometric closure and topological correctness even after corner rounding removal.

[0059] S52: Based on the mesh obtained by Boolean intersection, iteratively delete all boundary triangles to determine the mesh that conforms to the two-dimensional manifold structure; Specifically, boundary cleaning employs multiple iterations: First, the number of shared triangles for all edges in the mesh is counted, and edges used by only one triangle are marked. Then, all triangles containing at least one boundary edge, i.e., boundary triangles, are found and removed from the mesh. Removal may generate new boundary triangles, so the above process needs to be repeated until no boundary triangles remain in the mesh. This process ensures that each edge in the final mesh is shared by exactly two triangles, satisfying the basic requirements of a two-dimensional manifold mesh.

[0060] Taking a three-hole circular mechanical connection component model as an example, Figure 12 This visually illustrates the mesh portions that do not meet the requirements of a 2D manifold structure and need to be deleted. The boundary triangles highlighted in green are the ones to be deleted. These triangles are typically located in the transition zone between the original extended meshes, and their boundary edges are not shared by the two triangles after Boolean operations. Through iterative cleanup, these triangles that do not meet the requirements of a 2D manifold are removed one by one.

[0061] S53: Based on the mesh conforming to the two-dimensional manifold structure, identify and remove [the affected structures] based on connectivity analysis. All isolated mesh fragments and dangling mesh cells separated from the main mesh volume are used to determine the final mesh model with rounded corner features removed.

[0062] Specifically, connectivity analysis is based on a breadth-first search algorithm: starting with any triangle, it traverses its adjacent triangles, marking all connected triangle sets. The number of triangles contained in each connected region is counted, and only the largest connected region is retained; the remaining mesh regions are considered noise or invalid debris and are deleted. Simultaneously, the algorithm automatically handles dangling triangle groups connected to the main mesh only by a single vertex or edge, ensuring the removal of all non-manifold structures. The final output mesh maintains geometric integrity while eliminating topological noise and isolated debris, generating a clean mesh with a complete structure and clear boundaries—the result of rounded corner removal.

[0063] Taking a three-hole circular mechanical connection component model as an example, Figure 13 The final mesh model with rounded corners removed is shown. As you can see, the original rounded corner areas on the model have been completely removed and replaced with sharp edges defined by precise geometry. The entire mesh maintains a complete geometric outline and a strict two-dimensional manifold structure, without any isolated fragments or non-manifold elements.

[0064] Furthermore, to verify the practical effect of this method in simplifying the complexity of the mesh model, we remeshed the mesh before and after removing the fillets. Taking a three-hole circular mechanical connection component model as an example, as follows... Figure 14 As shown, under the condition that the mesh quality of the two mesh models is approximately consistent, the statistical data shows that the mesh model before removing the fillets contains 102,862 triangular mesh faces, while the mesh model after removing the fillets contains only 48,066 triangular mesh faces, a reduction of approximately 53.3% in the number of mesh faces. This result fully demonstrates that the proposed method can significantly reduce the model complexity and greatly improve the computational efficiency of subsequent CAE analysis while maintaining geometric accuracy.

[0065] As demonstrated by the above embodiments, this application effectively solves the technical challenge of automated identification and removal of complex rounded corner features in CAD models. By innovatively constructing a dedicated graph neural network architecture for rounded corner features, and combining a multi-source feature fusion mechanism of geometric features and engineering attributes, high-precision identification of various rounded corner features is achieved. The combination of a boundary-aware mesh expansion algorithm and precise Boolean intersection operations fully preserves the original model's design intent and geometric features while achieving accurate removal and geometric reconstruction of rounded corner regions, forming an automated solution from feature recognition to geometric simplification, significantly improving the efficiency and quality of CAE preprocessing.

[0066] In our verification experiments, we conducted a comprehensive comparative evaluation of this invention with mainstream commercial software Solidworks and the latest academic achievement, Defillet. For example... Figure 15 As shown, the left side presents three CAD models with typical rounded corner features: a porous spherical structure model, a support-type mechanical connection model, and an electromechanical control component housing model. The right side visually demonstrates the rounded corner recognition performance of the three methods on these three models. For the first porous spherical structure model, its rounded corner features are relatively simple. SolidWorks failed to accurately identify any rounded corners, while the Defilet method and the present invention accurately located the rounded corner features. For the latter two models, due to their complex rounded corner features, such as multiple intersecting rounded corners and multi-shaped transition rounded corners, the recognition accuracy of the present invention is much higher than that of SolidWorks and the Defilet method. This indicates that the present invention exhibits stronger robustness and adaptability when handling complex rounded corner scenarios, effectively solving the problem of low recognition accuracy of existing methods in such scenarios, and providing reliable technical support for the integrated CAD / CAE workflow in industry.

[0067] In summary, this invention, while maintaining a lightweight algorithm, achieves a comprehensive improvement in recognition accuracy, processing efficiency, and output quality, providing reliable technical support for integrated CAD / CAE workflows in the industrial field. The verification in Example 1 fully demonstrates the significant advancements and practical value of the method of this invention compared to existing technologies.

[0068] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0069] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for identifying and removing rounded corner features in CAD models based on graph neural networks, characterized in that, include: S1: Obtain the face adjacency graph constructed from the CAD model B-Rep representation; S2: Based on the face adjacency graph, extract the geometric features and attribute features of each surface to determine the node feature vector; S3: Based on the node feature vectors and the face adjacency graph, use a graph neural network to classify each surface and determine the rounded corner recognition result; S4: Based on the rounded corner recognition results, delete all rounded corners and triangulate the remaining non-rounded corners, determine whether mesh expansion is needed, adjust the mesh node index and merge the original mesh, and determine the output composite mesh. S5: Based on the synthesized new mesh, determine the final mesh model with rounded corner features removed through Boolean intersection and post-processing operations.

2. The method according to claim 1, characterized in that, Based on the face adjacency graph, the geometric and attribute features of each surface are extracted to determine the node feature vector, including: S21: Based on the surfaces in the face adjacency graph, the ordered sampling points are determined by uniformly sampling over the UV parameter domain of the surfaces; S22: Based on the ordered sampling points, calculate the Gaussian curvature and average curvature of each sampling point to determine the constructed geometric feature vector; S23: Based on the geometric feature vector, extract the surface width and dihedral angle attribute features of adjacent surfaces to determine the final constructed node feature vector.

3. The method according to claim 1, characterized in that, Based on the node feature vectors, a graph neural network is used to classify each surface and determine the rounded corner recognition result, including: S31: Design a network architecture that includes a preprocessing layer, an embedding layer, a graph convolutional layer, a fusion layer, and a classification layer, and determine the graph neural network model for rounded corner recognition; S32: Based on the graph neural network model for rounded corner recognition, train the network using a self-built labeled dataset. During the training process, use the Adam optimizer, L2 regularization, and Dropout strategy to determine the trained rounded corner recognition model. S33: Based on the trained rounded corner recognition model and the node feature vector, obtain the probability that each surface belongs to a rounded corner surface, and determine the rounded corner surface recognition result.

4. The method according to claim 1, characterized in that, Based on the rounded corner recognition results, all rounded corners are deleted, and the remaining non-rounded corners are triangulated. It is then determined whether mesh expansion is needed, the mesh node indices are adjusted, and the original meshes are merged to determine the output composite mesh, including: S41: Based on the rounded corner recognition results, determine the new boundary edge generated at the original connection point by deleting all rounded corners; S42: Based on the new boundary edge, traverse all remaining non-rounded faces. For non-rounded faces without boundary edges, perform triangulation directly. For non-rounded faces with boundary edges, first find the boundary edges with continuous positions on the curved surface to form a boundary edge group. After triangulation, locate all grid points on the boundary edge group and determine the grid points on the boundary edge group. S43: Based on the grid points on the boundary edge group, calculate the offset direction by calculating the normal information of the triangular grid surface where the grid points are located and the tangent information of the grid edge, and determine the newly generated grid points; S44: Determine the constructed extended triangular mesh based on the newly generated mesh points and the mesh points on the original boundary edge group; S45: Based on the extended triangular mesh and the original mesh, integrate them through global node index reorganization to determine the final output composite new mesh.

5. The method according to claim 1, characterized in that, Based on the synthesized new mesh, the final mesh model with rounded corner features removed is determined through Boolean intersection and post-processing operations, including: S51: Based on the synthesized new mesh, perform Boolean intersection operation and automatically detect and process the intersection relationship between different extended mesh patches, perform mesh trimming and reconstruction, and determine the mesh after Boolean intersection; S52: Based on the mesh obtained by Boolean intersection, iteratively delete all boundary triangles to determine the mesh that conforms to the two-dimensional manifold structure; S53: Based on the mesh conforming to the two-dimensional manifold structure, identify and remove [the affected structures] based on connectivity analysis. All isolated mesh fragments and dangling mesh cells separated from the main mesh volume are used to determine the final mesh model with rounded corner features removed.

Images