A CAD-Aware Edge Energy-Driven Mesh Simplification Method for Digital Twins

By using a side-level graph and a multi-task edge energy prediction network, the problem that existing mesh simplification methods cannot simultaneously take into account CAD features and topological validity is solved, achieving efficient and stable mesh simplification while maintaining CAD structure and topological validity.

CN123118968APending Publication Date: 2026-11-17JIAN COLLEGE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611388333.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing mesh simplification methods struggle to simultaneously maintain CAD features, topology validity, local mesh quality, fold position stability, and processing efficiency, and also incur high computational costs, potentially leading to non-manifold structures, boundary tearing, and feature curve drift.

Method used

We adopt an edge-centric multi-task edge energy-driven approach. By constructing an edge-level graph and extracting CAD-aware features, we use a multi-task edge energy prediction network to predict folding risk and new vertex positions. Combined with local geometric refinement and topological constraints, we achieve edge folding and simplification.

Benefits of technology

It improves mesh simplification rate, maintains CAD structure and topology validity, reduces computational overhead, avoids non-manifold structures and characteristic curve drift, and improves processing efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118968A_ABST
    Figure CN123118968A_ABST
Patent Text Reader

Abstract

This invention discloses a CAD-aware edge energy-driven mesh simplification method for digital twins. The method includes: inputting CAD mesh model data; constructing an edge-level graph and extracting CAD-aware features; using a multi-task edge energy prediction network to predict the folding risk energy and bounded position coefficients of candidate edges; constructing a comprehensive folding priority based on the learned energy and explicit CAD-aware geometric terms, and performing edge folding; performing feature-aware tangential Laplacian refinement on new vertices without geometric support constraints; and outputting a lightweight CAD mesh suitable for digital twin platforms. This invention can reduce mesh complexity while maintaining the effectiveness of CAD structures and topologies such as sharp features, cylindrical surfaces, and tori, improving the real-time rendering efficiency and geometric stability of digital twin models.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of computer graphics, computer-aided design, and lightweight digital twin models, and particularly to a CAD-aware edge energy-driven mesh simplification method, system, electronic device, and computer-readable storage medium for digital twins. Background Technology

[0002] Digital twin systems typically use high-precision 3D models to map physical entities and are used for equipment status display, operation monitoring, virtual debugging, offline programming, and virtual-real synchronization. The original CAD models of industrial equipment, robots, tooling fixtures, and complex mechanical parts, after triangulation, often contain a large number of vertices and triangular faces. Directly importing these into a real-time rendering platform can easily lead to increased loading times, higher VRAM and memory usage, lower interactive frame rates, and latency in virtual-real synchronization.

[0003] Existing mesh simplification methods typically use edge folding as the basic operation. Quadratic error metric methods can select folding edges based on local geometric errors, but they rely heavily on manually designed geometry and struggle to simultaneously identify sharp edges, hole boundaries, feature intersections, cylindrical generatrices, and other CAD engineering structures. Furthermore, the optimal folding position calculated by these methods may deviate from the original surface when the local matrix is ​​ill-conditioned or the mesh quality is poor. Some learning-based methods use vertices or triangular patches as graph nodes, making it difficult to directly represent the edge folding operation, and they typically only output a binary folding score without simultaneously learning the positions of new vertices.

[0004] Furthermore, existing learning-based simplification methods often use heuristic losses instead of real folding feedback, making it difficult for the network to obtain direct supervision of changes in normals, dihedral angles, triangle mass, and local geometric errors after folding. Frequent global graph reconstruction and repeated neural network inference during the execution phase introduce significant computational overhead. Without strict topological conditions, local quality checks, and projections of feature curves and analytical surfaces, non-manifold structures, boundary tears, cylinder radius shrinkage, or feature curve drift may occur. Using additional neural networks for global geometric correction of the simplification results further increases model complexity and may perturb already protected CAD features.

[0005] Therefore, a mesh simplification technique is needed that can jointly learn folding risk and folding location with edges as the center, and combine learning sorting with CAD feature constraints, topological constraints, local quality constraints and lightweight geometric refinement. Summary of the Invention

[0006] The purpose of this invention is to provide a CAD-aware edge energy-driven mesh simplification method for digital twins, in order to solve the problem that existing methods are unable to simultaneously take into account CAD feature preservation, topology validity, local mesh quality, fold position stability and processing efficiency.

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

[0008] A CAD-aware edge energy-driven mesh simplification method for digital twins includes:

[0009] S1. Input CAD mesh model data; obtain CAD model data corresponding to the digital twin object, convert the CAD model data into a triangular mesh model, and read the vertex coordinates, edge connection relationships, triangular face index, vertex normal and face normal information of the triangular mesh model;

[0010] S2. Construct a side-level graph and extract CAD-aware features; use the edges in the triangular mesh model as graph nodes of the side-level graph, and use the edges that share at least one mesh vertex as adjacent graph nodes to construct edge-edge adjacency relationships, and extract CAD-aware edge feature vectors containing geometric and structural descriptions as well as local triangle quality descriptions for each candidate edge.

[0011] S3. Perform multi-task edge energy prediction; input the CAD-sensing edge feature vector and the edge-edge adjacency relationship into the multi-task edge energy prediction network. The multi-task edge energy prediction network includes a shared encoder, an energy head, and a position head. The energy head outputs the folding risk energy of each candidate edge, and the position head outputs the bounded three-dimensional position coefficients used to determine the position of the new vertex after folding.

[0012] S4. Perform energy-driven constraint edge folding; normalize the folding risk energy and combine the normalized folding risk energy with CAD-aware geometry terms to obtain the comprehensive folding priority of candidate edges; select edges to be folded based on the comprehensive folding priority, determine the positions of candidate new vertices using the bounded 3D position coefficients, and perform edge folding when topological validity, local triangle quality, sharp feature protection, feature curve constraints, and analytical surface constraints are all satisfied; after folding, inherit the basic energy of related neighboring edges, and update the comprehensive folding priority in combination with the recalculated local geometry terms, repeating the process until the target vertex retention rate is reached or there are no foldable edges that satisfy the constraints;

[0013] S5. Perform geometric refinement after folding; perform feature-aware tangential Laplacian refinement on new vertices generated during edge folding that are not constrained by feature curves or analytical surfaces, and check the local triangle quality and surface normal changes at candidate refinement locations to obtain the geometrically refined CAD mesh model.

[0014] S6. Output the simplified CAD mesh model; delete the degenerate triangles formed during the simplification process, and reconstruct the effective vertex index and triangle index; finally, output a lightweight CAD mesh model suitable for loading, real-time rendering, and virtual-real synchronization on the digital twin platform.

[0015] Furthermore, the CAD-sensing edge feature vector is a 10-dimensional feature vector, including: normalized edge length, endpoint normal consistency, normalized dihedral angle, endpoint sharpness, local area, cylinder generatrix indication, minimum quality risk of associated surface, minimum quality risk of a ring, average quality risk of a ring, and proportion of low-quality surfaces in a ring.

[0016] Furthermore, the normalized side length is obtained by standardizing the logarithm of the side length using Z-score; the cylinder generatrix indicator is used to mark sides whose direction is approximately perpendicular to the average normal of the endpoints and which have at least two adjacent sides with the same height as their direction within a ring neighborhood; the local triangle mass is calculated using the following formula:

[0017]

[0018] in, Triangular face The quality value; Triangular face area, , and Triangular face The three side lengths; the minimum quality risk of the associated surface, the minimum quality risk of a ring, the average quality risk of a ring, and the proportion of low-quality surfaces in a ring are all obtained by calculating the quality of the local triangle.

[0019] Furthermore, the shared encoder includes a feature embedding multilayer perceptron and at least one EdgeConv graph convolutional module; the feature embedding multilayer perceptron maps the CAD-sensing edge feature vectors to the latent feature space; the EdgeConv graph convolutional module, based on the edge-to-edge adjacency relationship, concatenates, nonlinearly transforms, and aggregates the feature differences between the central edge node representation and the adjacent edge nodes relative to the central edge node; when the EdgeConv graph convolutional module is multilayer, residual connections are set between adjacent layers.

[0020] Furthermore, the multi-task edge energy prediction also includes local folding simulation, which at least evaluates the optimal position of the edge midpoint, the two edge endpoints, and the numerically valid local quadratic error metric; the position with the highest folding preference is selected as the pseudo-target position from candidate folding positions with acceptable geometric risk, and the target position coefficient is generated according to the following formula:

[0021]

[0022] in, Candidate edges Target position coefficient; Candidate edges The location of the false target. and Let be the coordinates of the endpoints of the candidate edge. is the length of the candidate edge.

[0023] Furthermore, the training loss of the multi-task edge energy prediction network includes folding cost loss, pairwise preference ranking loss, position regression loss, and feature energy loss.

[0024] Furthermore, the bounded three-dimensional position coefficients and the candidate new vertex positions satisfy:

[0025]

[0026] in, Candidate edges Bounded three-dimensional position coefficients, For the multilayer perceptron mapping of the position head. Candidate edges Shared potential representation, For position boundary parameters, This represents the candidate new vertex position.

[0027] Furthermore, the feature is that, when considering candidate new vertex positions in CAD-sensitive areas, feature chains are first extracted from the high dihedral angle edges, and the feature chains are fitted as straight lines or circles; cylindrical or toroidal geometric supports are established based on compatible circular features; folding is rejected when the two endpoints of a candidate edge belong to different geometric supports or only one endpoint belongs to a geometric support; when the candidate edge belongs to the same geometric support, the candidate new vertex position is projected onto the geometric support.

[0028] Furthermore, the multi-task edge energy prediction network performs a global forward inference on the current grid at the start of simplification; when the edge... Folding forms adjacent sides At that time, the basic energy of the adjacent side is determined by the folded front side. With edge The normalized learning energy is inherited, and the smaller value is taken when both exist, the value is reused when only one exists, and a preset neutral value is used when neither exists. The inherited basic energy is combined with the recalculated local geometry to update the local synthesis folding priority.

[0029] Compared with the prior art, the present invention has at least the following beneficial effects:

[0030] 1. This invention uses grid edges as graph nodes, making the network output directly correspond to the edge folding operation; it uses 10-dimensional CAD to perceive edge features and simultaneously express local geometric scale, normal and dihedral angle changes, sharp edge intersections, surface discrete density, cylinder semantics, and the mass context of a ring triangle.

[0031] 2. This invention employs a multi-task EdgeEnergyNet composed of a shared encoder, an energy head, and a position head, which simultaneously obtains the folding risk energy and bounded position coefficients of candidate edges in a single forward inference, avoiding the separation of edge selection and vertex placement.

[0032] 3. This invention generates continuous folding costs, preference rankings, and target location pseudo-labels through local folding simulation, enabling the network to directly learn the geometric and quality effects after folding, rather than relying solely on manual heuristic classification rules.

[0033] 4. This invention combines learning energy with dihedral angles, cylindrical generatrices, endpoint normal curvature, short side and triangle aspect ratio terms, and superimposes characteristic curves, straight lines, circles, cylindrical surfaces, toroidal surfaces, topological chain conditions and local quality checks, thereby improving the simplification rate while maintaining the validity of the CAD structure and topology.

[0034] 5. This invention employs local energy inheritance and local geometry update to reduce global graph reconstruction and repetitive neural network inference during the simplification process; it also adopts a batch folding strategy that does not share endpoints, balancing execution efficiency and local update security.

[0035] 6. This invention improves local triangle regularity without compromising the protected CAD structure by replacing an additional global neural refinement network with deterministic, interpretable feature-aware tangential Laplacian refinement and excluding new vertices already constrained by geometric support. (See attached figures)

[0036] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0037] Figure 1 This is a schematic diagram of the overall process of a CAD-aware edge energy-driven mesh simplification method for digital twins according to an embodiment of this application;

[0038] Figure 2 This is a schematic diagram of a multi-task edge energy prediction network structure according to an embodiment of this application. Detailed Implementation

[0039] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0040] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0042] See Figure 1 A CAD-aware edge energy-driven mesh simplification method for digital twins includes the following steps:

[0043] S1: Input CAD mesh model data;

[0044] See Figure 1 In step S1, CAD model data corresponding to the digital twin object is acquired. The CAD model data is triangulated, and vertex coordinates, triangular facet indices, edge connectivity, vertex normals, and face normals are read in. A set of non-repeating edges and a vertex ring neighborhood are established based on the triangular faces. The mesh is then translated to the center of the bounding box, and scale normalization is performed according to the maximum external dimensions to reduce the scale differences between different digital twin objects.

[0045] Let the input triangular mesh be

[0046]

[0047] in, , and They are the vertex set, edge set, and face set, respectively. , and These represent the number of vertices, edges, and faces, respectively. The goal is to reduce the number of vertices from... Reduce to ,in To achieve the target vertex retention rate, and to maintain CAD geometric features and topological validity during the simplification process.

[0048] S2: Boundary-level graph construction and CAD-aware feature extraction;

[0049] For the edge ,by and Representing the coordinates of the two endpoints, in and This represents the unit normal to both endpoints. A boundary-level graph treats each grid edge as a graph node; if two grid edges share at least one vertex, an adjacency relationship is established between the corresponding graph nodes.

[0050] The CAD-aware feature vector of each candidate edge is defined as follows:

[0051]

[0052] in, to For geometric and structural description, to This is a local triangle mass descriptor.

[0053] set up This represents the Z-score normalization over all candidate edges in the current grid, applied to scalar edge properties. have

[0054]

[0055] in, and They are respectively The mean and standard deviation.

[0056] Normalized side length is

[0057]

[0058] in, Logarithmic transformation is used to compress the dynamic range of side lengths in CAD meshes with different discrete densities.

[0059] Endpoint normal consistency is

[0060]

[0061] set up and For the edge The unit normal, dihedral angle, and normalized characteristics of two adjacent triangular faces are:

[0062]

[0063] The sharpness of the endpoint is

[0064]

[0065] in, and Respectively with the vertex and Connect and normalize the number of edges whose dihedral angles exceed the acute edge threshold.

[0066] set up As vertex The set of related surfaces, For noodles The area, then

[0067]

[0068] in, and Vertices and vertex The cumulative added value of the related surfaces, This represents the area of ​​a local region.

[0069] Cylindrical busbar indicator quantity is

[0070]

[0071] in, For the detected set of cylinder generatrix edges, This is an indicative function. When the direction of a candidate edge is approximately perpendicular to the average normal of its two endpoints, and there are at least two adjacent edges in a ring neighborhood that have the same height as the candidate edge, the candidate edge is marked as a cylindrical generatrix edge.

[0072] set up For the edge The set of related surfaces, For the edge The set of toruses corresponding to two endpoints. For an area of... The side length is , and triangular face Its quality is

[0073]

[0074] This leads to four characteristics of quality risk:

[0075]

[0076]

[0077]

[0078]

[0079] in, In one alternative implementation, a low quality threshold is set to... . Reflects the minimum quality risk of the related surface. Extend the worst-case quality risk to one ring surface, i.e., the minimum quality risk of one ring. This reflects the overall regularity of a triangular ring, i.e., the average quality risk of the ring. This reflects the proportion of low-quality surfaces in the first ring.

[0080] S3: Multi-task edge energy prediction;

[0081] See Figure 2 The multi-task edge energy prediction network includes a shared encoder, an energy head, and a location head. The input edge feature matrix is ​​denoted as... Its shape is The sparse index of the boundary graph topology is denoted as . Its shape is ,in, Let be the number of directed adjacency relationships between edge nodes.

[0082] The shared encoder includes a feature embedding multilayer perceptron and at least one EdgeConv graph convolutional module.

[0083] The embedded multilayer perceptron maps 10-dimensional edge features to a latent representation:

[0084]

[0085] In one alternative implementation, the embedded multilayer perceptron sequentially includes Linear ReLU and Linear .

[0086] For the central edge node in the boundary graph ,set up Let be the set of adjacent edge nodes, the first... The EdgeConv layer output is

[0087]

[0088] in, For EdgeConv layer indexing, This refers to the number of EdgeConv layers; and The first Layer center edge node and its adjacent edge nodes The expression, ;symbol This represents vector concatenation. This represents a convolutional mapping multilayer perceptron.

[0089] When EdgeConv layer number When using residual updates:

[0090]

[0091] Shared encoder output , its first OK Candidate edges Shared potential representation.

[0092] The energy head outputs raw folding risk energy:

[0093]

[0094] The position head outputs bounded three-dimensional position coefficients:

[0095]

[0096] in, These are the position boundary parameters. The position of the candidate new vertex is...

[0097]

[0098] in, and Let be the coordinates of the endpoints of the candidate edge. The length of the candidate edge;

[0099] By restricting the location coefficients to a bounded range and scaling them locally by side length, we can avoid the predicted location from deviating too much from the original edge.

[0100] The simulator-in-the-loop training of the multi-task side energy prediction network includes:

[0101] Before training, a local folding simulation is performed on each candidate edge. Candidate folding locations must include at least the edge midpoint, both endpoints, and a numerically valid local QEM location. For each candidate location, the affected torus is temporarily updated, and the maximum change in face normal, the maximum dihedral angle of the smooth region, the average and minimum triangle mass changes, and the local QEM error are calculated after folding.

[0102] Select the position with the highest folding preference from candidate positions where geometric risk is acceptable. As the pseudo-target location, the corresponding target location coefficient is

[0103]

[0104] The norm of the target position coefficient is constrained by a preset upper limit. The total training loss of the multi-task edge energy prediction network is:

[0105]

[0106] in, For the cost of folding, For pairwise preference ranking loss, For position regression loss, Characteristic energy loss; , , , These are the corresponding weighting coefficients.

[0107] set up To achieve the goal of continuous folding cost, the folding cost loss uses binary cross-entropy with logits:

[0108]

[0109] set up For the high-preference edge and low preference edge The set of preference pairs, the pairwise preference ranking loss is

[0110]

[0111] in, This represents the sorting interval.

[0112] The position regression weights can be set to ,in To simulate folding risk; position regression loss is

[0113]

[0114] in, Indicates smoothness loss

[0115] set up The set of CAD-sensitive edges consists of dihedral angle edges and cylindrical generatrice edges, with characteristic energy loss being...

[0116]

[0117] in, The characteristic energy interval is denoted by .

[0118] The aforementioned loss assigns higher energy to high-cost edges and lower energy to high-preference edges relative to low-preference edges, enabling the location head to learn safe candidate target locations and preventing CAD-sensitive edges from being assigned excessively low energy. Specific weights and intervals can be adjusted based on the type and discrete density of the training mesh.

[0119] S4: Energy-driven constraint edge folding execution;

[0120] At the start of the simplification, a global forward inference of the multi-task edge energy prediction network is performed on the current grid. The raw energy is mapped using the Sigmoid function. The overall folding priority is defined as follows:

[0121]

[0122] in, For the normalized dihedral angle term, For cylindrical busbar indicator items, For the endpoint normal bending term, For the short-side bias term, This is the aspect ratio offset term. The average side length of the current grid. The average aspect ratio of the candidate edge-associated triangles is the value after logarithmic transformation and standardization. The smaller the value, the higher the priority for folding the candidate edge. , , , and These are the weights of the dihedral angle term, the cylinder generatrix term, the endpoint normal curvature term, the short side offset term, and the triangle aspect ratio offset term, respectively.

[0123] The positions of candidate new vertices are initially determined by the position head output. For CAD-sensitive regions, feature chains are first constructed from the high dihedral corner edges and fitted to straight lines or circles; compatible circular features are used to establish cylindrical or toroidal geometric supports. Candidate edges that cross different geometric supports are rejected from collapsing; candidate edges belonging to the same geometric support are excluded. When, execute

[0124]

[0125] in, , , and These represent geometric supports for a straight line, a circle, a cylindrical surface, and a torus, respectively. For the projection operator corresponding to the geometric support;

[0126] For straight lines, orthogonal projection can be used; for circles, projection onto the fitted circular plane and radial normalization can be used; for cylindrical surfaces, axial parameters can be preserved and radial distance can be restored to the fitted radius; for toroidal surfaces, candidate positions can be projected onto the fitted toroidal surface.

[0127] Before performing the fold, check the topological chain loop condition. For internal edges... ,Require

[0128]

[0129] For boundary edges, the requirements are as follows:

[0130]

[0131] Subsequently, the neighborhood of the candidate edge is temporarily updated, and it is checked whether a duplicate or degenerate surface is formed. The minimum triangle mass, surface normal change, and dihedral angle change of the smooth region are also checked after folding. If any of the checks fails to meet the preset conditions, folding is rejected.

[0132] The executor can select a batch of valid candidate edges that do not share endpoints from a priority queue to perform folding, in order to avoid simultaneous modification of overlapping local neighborhoods within the same batch. After folding, the vertices, faces, and adjacency relationships are updated, and invalid triangles are deleted.

[0133] For the adjacent sides formed after folding Its basic energy comes from the folded front. and Inheritance is used; the smaller normalized energy is taken when both exist, the value is reused when only one exists, and a neutral value (0.5) is used when neither exists. Subsequently, local dihedral angles, normals, side lengths, and aspect ratios are recalculated, and local priorities are updated. This avoids rebuilding the global edge graph and repeatedly performing network inference after each fold.

[0134] S5: Geometric refinement after folding;

[0135] set up The new set of vertices generated by edge folding. For a new set of vertices constrained by a characteristic curve, cylindrical surface, or torus, refine the vertex set as follows:

[0136]

[0137] The symbol “∖” represents the set difference operation. This represents a newly generated and unconstrained set of vertices.

[0138] for ,set up For a vertex neighborhood of a ring, the candidate refinement positions are:

[0139]

[0140] matrix Remove the normal component of the Laplacian displacement to reduce out-of-plane shrinkage. for Identity matrix. Damping weights. The weighting is determined by the local average dihedral angle and the normalized curvature of the surface: regions with strong curvature or sharp features are given smaller weights or frozen, while smooth regions are given larger weights.

[0141] set up and These represent the minimum quality of the associated surfaces before and after the candidate update. and Before and after the candidate update The unit normal. Let... To refine the vertices A set of associated triangle faces. Only if the condition is met...

[0142]

[0143] When accepting candidate refinement positions, among which As the lower limit of quality, This represents the maximum allowable change angle of the surface normal. The refinement process can be iterated a preset number of times.

[0144] S6: Outputs simplified CAD meshes to the digital twin platform;

[0145] After achieving the target vertex retention rate or eliminating the presence of collapsible edges that satisfy the constraints, degenerate triangular patches formed during the simplification process are deleted, and valid vertex indices and triangular patch indices are reconstructed. An inverse transformation corresponding to the translation and scale normalization described in step S1 is performed on the geometrically refined CAD mesh model to restore the original spatial scale and position of the CAD mesh model. The output lightweight CAD mesh can be applied to digital twin models of industrial robot bodies, robot links, end effectors, gripper mechanisms, tooling fixtures, machine tool components, conveying equipment, housings, disc covers, shafts, and other mechanical products. It can also be imported into Unity, Unreal Engine, WebGL, or other digital twin visualization platforms for real-time display, collision visualization, runtime mapping, virtual debugging, and virtual-real synchronization.

[0146] This invention can also be implemented by software modules, hardware acceleration modules, or a combination of both. The corresponding system includes a CAD mesh input module, a side-level graph and 10D CAD-aware feature construction module, a multi-task edge energy prediction module, an energy-driven constraint folding module, a post-folding geometry refinement module, and a lightweight mesh output module. Simulator-in-the-loop training of the multi-task edge energy prediction module can be pre-executed as part of the model training process for the multi-task prediction module. The method can be implemented by a processor of an electronic device executing a program in memory, or it can be stored in a non-transitory computer-readable storage medium.

[0147] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A CAD-aware edge energy-driven mesh simplification method for digital twins, characterized in that, include: S1. Input CAD mesh model data; Obtain CAD model data corresponding to the digital twin object, convert the CAD model data into a triangular mesh model, and read the vertex coordinates, edge connection relationships, triangular face indexes, vertex normals, and face normals of the triangular mesh model; S2. Construct a side-level graph and extract CAD-aware features; use the edges in the triangular mesh model as graph nodes of the side-level graph, and use the edges that share at least one mesh vertex as adjacent graph nodes to construct edge-edge adjacency relationships, and extract CAD-aware edge feature vectors containing geometric and structural descriptions as well as local triangle quality descriptions for each candidate edge. S3. Perform multi-task edge energy prediction; input the CAD-sensing edge feature vector and the edge-edge adjacency relationship into the multi-task edge energy prediction network. The multi-task edge energy prediction network includes a shared encoder, an energy head, and a position head. The energy head outputs the folding risk energy of each candidate edge, and the position head outputs the bounded three-dimensional position coefficients used to determine the position of the new vertex after folding. S4. Perform energy-driven constraint edge folding; The folding risk energy is normalized, and the normalized folding risk energy is combined with the CAD-perceived geometry term to obtain the comprehensive folding priority of the candidate edges. The edges to be folded are selected based on the comprehensive folding priority. The positions of candidate new vertices are determined using the bounded 3D position coefficients. Edge folding is performed when the topological validity, local triangle quality, sharp feature protection, feature curve constraints, and analytical surface constraints are all satisfied. After folding, the basic energy of the relevant neighboring edges is inherited, and the comprehensive folding priority is updated in combination with the recalculated local geometry terms. This process is repeated until the target vertex retention rate is reached or there are no foldable edges that satisfy the constraints. S5. Perform geometry refinement after folding; Feature-aware tangential Laplacian refinement is performed on new vertices generated during edge folding that are not constrained by feature curves or analytical surfaces. Local triangle quality and face normal changes are checked at candidate refinement locations to obtain a geometrically refined CAD mesh model. S6. Output the simplified CAD mesh model; delete the degenerate triangles formed during the simplification process, and reconstruct the effective vertex index and triangle index; finally, output a lightweight CAD mesh model suitable for loading, real-time rendering, and virtual-real synchronization on the digital twin platform.

2. The method according to claim 1, characterized in that, The CAD-sensing edge feature vector is a 10-dimensional feature vector, including: normalized edge length, endpoint normal consistency, normalized dihedral angle, endpoint sharpness, local area, cylinder generatrix indication, minimum quality risk of associated surface, minimum quality risk of a ring, average quality risk of a ring, and proportion of low-quality surfaces in a ring.

3. The method according to claim 2, characterized in that, The normalized side length is obtained by standardizing the logarithm of the side length using the Z-score; the cylinder generatrix indicator is used to mark sides whose direction is approximately perpendicular to the average normal of the endpoints and which have at least two adjacent sides with the same height as their direction within a ring neighborhood; the local triangle mass is calculated using the following formula: in, Triangular face The quality value; Triangular face area, , and Triangular face The three side lengths; the minimum quality risk of the associated surface, the minimum quality risk of a ring, the average quality risk of a ring, and the proportion of low-quality surfaces in a ring are all obtained by calculating the quality of the local triangle.

4. The method according to claim 1, characterized in that, The shared encoder includes a feature embedding multilayer perceptron and at least one EdgeConv graph convolutional module; the feature embedding multilayer perceptron maps the CAD-sensing edge feature vectors to a latent feature space; the EdgeConv graph convolutional module, based on the edge-to-edge adjacency relationship, concatenates, performs nonlinear transformation, and means aggregation on the feature difference between the central edge node representation and the adjacent edge nodes relative to the central edge node; when the EdgeConv graph convolutional module is multilayer, residual connections are set between adjacent layers.

5. The method according to claim 1, characterized in that, Performing multi-task edge energy prediction also includes local folding simulation, which evaluates at least the optimal positions of the edge midpoint, the two edge endpoints, and the numerically valid local quadratic error metric. The position with the highest folding preference is selected from candidate folding positions with acceptable geometric risk as the pseudo-target position, and the target position coefficient is generated according to the following formula: in, Candidate edges Target position coefficient; Candidate edges The location of the false target. and Let be the coordinates of the endpoints of the candidate edge. is the length of the candidate edge.

6. The method according to claim 1 or 5, characterized in that, The training loss of the multi-task edge energy prediction network includes folding cost loss, pairwise preference ranking loss, position regression loss, and feature energy loss.

7. The method according to claim 1, characterized in that, The bounded three-dimensional position coefficients and candidate new vertex positions satisfy: in, Candidate edges Bounded three-dimensional position coefficients, For the multilayer perceptron mapping of the position head. Candidate edges Shared potential representation, For position boundary parameters, This represents the candidate new vertex position.

8. The method according to claim 1 or 7, characterized in that, When considering candidate new vertex positions in CAD-sensitive areas, feature chains are first extracted from the high dihedral corner edges, and these feature chains are fitted to straight lines or circles. Cylindrical or toroidal geometric supports are then established based on compatible circular features. Folding is rejected when the two endpoints of a candidate edge belong to different geometric supports or when only one endpoint belongs to a geometric support. When a candidate edge belongs to the same geometric support, the candidate new vertex position is projected onto that geometric support.

9. The method according to claim 1, characterized in that, The multi-task edge energy prediction network performs a global forward inference on the current grid at the start of simplification; when the edge... Folding forms adjacent sides At that time, the basic energy of the adjacent side is determined by the folded front side. With edge The normalized learning energy is inherited, and the smaller value is taken when both exist, the value is reused when only one exists, and a preset neutral value is used when neither exists. The inherited basic energy is combined with the recalculated local geometry to update the local synthesis folding priority.

10. A CAD-aware edge energy-driven mesh simplification system for digital twins, characterized in that, include: The system comprises a CAD mesh input module, a side-level graph and CAD-aware feature construction module, a multi-task edge energy prediction module, an energy-driven constraint folding module, a folded geometry refinement module, and a lightweight mesh output module; wherein the multi-task edge energy prediction network in the multi-task edge energy prediction module is pre-trained in a simulator-in-the-loop manner; each module is configured to perform the method described in any one of claims 1 to 9.