Digital prototype lightweight system based on feature extraction and adaptive grid optimization

Abstract

The application provides a digital prototype lightweight system based on feature extraction and adaptive grid optimization, and relates to the technical field of computers, comprising a model access and preprocessing module and a file analysis unit; the file analysis unit is used for receiving an original digital prototype model, and analyzing geometric topology data and design tree information of the model; a feature collaborative decision unit receives output results of the three units, realizes screening and classification of multi-dimensional features through feature influence weight calculation, and generates a feature mapping diagram containing key feature areas and non-key feature areas; through multi-dimensional collaborative extraction of geometric analysis, topology analysis and semantic analysis, and in combination with weight coefficient calculation of feature comprehensive influence degree determined according to downstream CAE analysis sensitivity, key areas having influences on product performance and assembly precision can be accurately identified, compared with an existing single-dimensional feature extraction scheme, loss of functional key features is avoided, and a basis guarantee for subsequent CAE analysis precision is provided.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a lightweight digital prototype system based on feature extraction and adaptive mesh optimization. Background Technology

[0002] As product design complexity continues to increase, the volume of geometric topology data and product manufacturing information in digital prototype models is growing exponentially. Such large-volume models are prone to problems such as low computational efficiency, long transmission time, and excessive hardware resource consumption in subsequent CAE (computer-aided engineering) analysis, cross-departmental data transmission, and collaborative design processes. Therefore, lightweighting of digital prototypes has become a key technical link connecting CAD (Computer Aided Design) and CAE. The core requirement is to significantly reduce the amount of model data while fully retaining the key features that affect product performance and assembly accuracy.

[0003] Existing digital prototype lightweighting technologies have the following significant drawbacks:

[0004] The extraction dimension is singular. Most solutions only identify key areas based on geometric curvature or topological mating relationships, ignoring semantic information in PMI (Product and Manufacturing Information), such as geometric tolerances, datum markings, and surface roughness. This leads to the easy loss of key functional features after lightweighting, affecting the accuracy of subsequent CAE analysis.

[0005] Mesh optimization lacks a dynamic priority mechanism and adopts a uniform mesh density strategy for all regions of the model. This either results in excessive densification of non-critical regions to ensure the accuracy of critical regions, or sacrifices the accuracy of critical regions in pursuit of compression ratio, leading to mesh quality that does not meet CAE requirements. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a lightweight digital prototype system based on feature extraction and adaptive mesh optimization.

[0007] A lightweight digital prototype system based on feature extraction and adaptive mesh optimization includes a model access and preprocessing module and a file parsing unit;

[0008] The file parsing unit is used to receive the original digital prototype model and parse the geometric topology data and design tree information of the model;

[0009] The multi-dimensional feature collaborative extraction module includes a geometric analysis unit, a topological analysis unit, a semantic parsing unit, and a feature collaborative decision-making unit;

[0010] The feature collaborative decision-making unit receives the output results of the geometric analysis unit, the topology analysis unit, and the semantic parsing unit, and realizes the screening and classification of multi-dimensional features through feature influence weight calculation, generating a feature mapping map containing key feature regions and non-key feature regions.

[0011] The dynamic priority mesh optimization module includes a feature priority calculation unit, an adaptive mesh generation unit, and a mesh quality optimization unit. The feature priority calculation unit dynamically calculates the optimization priority, geometric complexity, and mean curvature change rate of each feature region based on the geometric complexity, semantic tolerance level, and topological fit strength of the features in the feature map.

[0012] The closed-loop accuracy verification module integrates a sampling calculation unit and a feedback adjustment unit. The sampling calculation unit dynamically adjusts the sampling strategy based on feature priority.

[0013] Preferably, the model access and preprocessing module includes a format parsing unit, a defect detection unit, and an automatic healing unit;

[0014] The format parsing unit supports CAD file parsing, reads model data through an adapted interface, and outputs geometric data containing topological relationships of faces, edges, and vertices.

[0015] The defect detection unit identifies defects such as gaps, geometric overlaps, and free surface patches using a geometric topology verification algorithm; the automatic healing unit executes an automatic healing algorithm based on tolerance.

[0016] Preferably, the geometric analysis unit of the multi-dimensional feature collaborative extraction module includes a curvature calculation subunit, an anomaly removal subunit, and a geometric feature recognition subunit;

[0017] The curvature calculation subunit uses the local surface fitting method to calculate the principal curvature and Gaussian curvature of each vertex on the model surface; the anomaly removal subunit removes abnormal patches using the dynamic standard deviation algorithm; and the geometric feature recognition subunit identifies geometric features based on curvature parameters.

[0018] Preferably, the topology analysis unit of the multi-dimensional feature collaborative extraction module includes a bounding box tree construction subunit, a feature pre-filtering subunit, and a cooperating region identification subunit. The bounding box tree construction subunit constructs a hierarchical AABB bounding box tree based on the functional grouping of model components.

[0019] The feature pre-filtering subunit filters out non-matching related components; the matching region identification subunit identifies matching regions in the assembly.

[0020] Preferably, the semantic parsing unit of the multi-dimensional feature collaborative extraction module includes a PMI data extraction subunit, a semantic association subunit, and a semantic feature labeling subunit, wherein the PMI data extraction subunit extracts product manufacturing information from CAD files;

[0021] The semantic association subunit establishes the association between PMI annotations and geometric elements through a domain-adapted pre-trained BERT model; the semantic feature labeling subunit labels functional features based on the association results.

[0022] Preferably, the feature priority calculation unit of the dynamic priority grid optimization module includes a factor extraction subunit, a weight determination subunit, and a priority calculation subunit, wherein the factor extraction subunit extracts the calculation factors of feature influence.

[0023] The weight determination subunit determines the weight coefficient of each factor; the priority calculation subunit calculates the priority of each feature region.

[0024] Preferably, the adaptive mesh generation unit of the dynamic priority mesh optimization module includes a density calculation subunit, a mesh generation subunit, and a transition layer generation subunit. The density calculation subunit determines the mesh unit size based on feature priority. The mesh generation subunit calls Gmsh to generate a tetrahedral mesh.

[0025] The transition layer generates sub-units that generate gradient transition meshes in regions with different densities.

[0026] Preferably, the mesh quality optimization unit of the dynamic priority mesh optimization module includes a quality evaluation sub-unit, a smoothing optimization sub-unit, and a re-partitioning sub-unit, wherein the quality evaluation sub-unit calculates the quality parameters of the mesh cells;

[0027] The smoothing optimization subunit executes the Laplace smoothing algorithm; the repartitioning subunit repartitions unqualified units.

[0028] Preferably, the sampling calculation unit of the closed-loop accuracy verification module includes a sampling planning subunit, a distance calculation subunit, and an error statistics subunit, wherein the sampling planning subunit determines the sampling strategy based on feature priority;

[0029] The distance calculation subunit calculates the distance error of the sampling points; the error statistics subunit statistically analyzes the error results.

[0030] Preferably, the feedback adjustment unit of the closed-loop accuracy verification module includes a feature threshold adjustment subunit and a grid parameter adjustment subunit. The feature threshold adjustment subunit adjusts the feature screening threshold based on the error result; the grid parameter adjustment subunit adjusts the grid density parameter based on the error result.

[0031] This invention provides a lightweight digital prototype system based on feature extraction and adaptive mesh optimization. It offers the following advantages:

[0032] By using multi-dimensional collaborative extraction of geometric analysis, topological analysis, and semantic parsing, and combining the weight coefficients determined by downstream CAE analysis sensitivity to calculate the comprehensive influence of features, key areas that affect product performance and assembly accuracy can be accurately identified. Compared with existing single-dimensional feature extraction schemes, this avoids the loss of key functional features and provides a basic guarantee for the accuracy of subsequent CAE analysis. Attached Figure Description

[0033] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] like Figure 1 As shown, this invention proposes a lightweight digital prototype system based on feature extraction and adaptive mesh optimization, including a model access and preprocessing module and a file parsing unit;

[0036] The file parsing unit supports STEPAP242, IGES5.3, and CATIAV5 / V6CGR formats. It is used to receive the original digital prototype model and parse the geometric topology data and design tree information of the model.

[0037] The geometric topology data includes the coordinates of faces, edges, and vertices, as well as their topological relationships. The design tree information includes the feature creation order and assembly constraint type. An automatic healing algorithm that dynamically determines the tolerance based on the model's geometric accuracy requirements repairs the gaps in faces and geometric overlaps in the model, outputting a repaired complete geometric model.

[0038] The multi-dimensional feature collaborative extraction module includes a geometric analysis unit, a topology analysis unit, a semantic parsing unit, and a feature collaborative decision-making unit. The geometric analysis unit calculates the surface curvature parameters of the model using a local surface fitting method. The curvature parameters include principal curvature and Gaussian curvature. The topology analysis unit identifies the mating regions of the assembly. The mating regions include the interface regions corresponding to bolted connections, welding, and interference fits. The semantic parsing unit parses the product manufacturing information of the CAD file based on a domain-adapted pre-trained BERT model.

[0039] Product manufacturing information includes geometric tolerances, datum, surface roughness, and critical dimensions;

[0040] The feature collaborative decision-making unit receives the output results of the geometric analysis unit, the topology analysis unit, and the semantic parsing unit, and realizes the screening and classification of multi-dimensional features through feature influence weight calculation, generating a feature mapping map containing key feature regions and non-key feature regions.

[0041] The dynamic priority mesh optimization module includes a feature priority calculation unit, an adaptive mesh generation unit, and a mesh quality optimization unit. The feature priority calculation unit dynamically calculates the optimization priority, geometric complexity, and mean curvature change rate of each feature region based on the geometric complexity, semantic tolerance level, and topological fit strength of the features in the feature map.

[0042] In the semantic tolerance levels, IT5 corresponds to 1.0, IT6 corresponds to 0.9, IT7 corresponds to 0.8, and IT8 corresponds to 0.7 (after normalization). The topological fit strength is characterized by the proportion of the contact area of ​​the mating surface. The adaptive mesh generation unit calls GmshSDK to generate tetrahedral meshes with different densities based on the priority results. The mesh quality optimization unit generates gradient transition meshes at the boundaries of different density regions through a joint algorithm of Delaunay refinement and Laplace smoothing, and outputs a lightweight mesh model.

[0043] The closed-loop accuracy verification module integrates a sampling calculation unit and a feedback adjustment unit. The sampling calculation unit dynamically adjusts the sampling strategy based on feature priority.

[0044] The sampling density is 50 points / mm² in the core area, 30 points / mm² in the ordinary area, and 15 points / mm² in the non-critical area. The size error and surface approximation error between the lightweight mesh model and the original model are calculated. The feedback adjustment unit feeds the error results back to the multi-dimensional feature collaborative extraction module and the dynamic priority mesh optimization module, respectively. The feedback to the multi-dimensional feature collaborative extraction module is used to adjust the feature selection threshold, and the feedback to the dynamic priority mesh optimization module is used to adjust the mesh cell size.

[0045] The specific steps for calculating the feature influence weight of the multi-dimensional feature collaborative decision-making unit are as follows:

[0046] The mean curvature change rate output by the geometric analysis unit is extracted as the geometric complexity factor, the contact area ratio of the mating surfaces output by the topology analysis unit is used as the topology mating strength factor, and the tolerance level quantification value output by the semantic parsing unit is used as the semantic tolerance factor.

[0047] The weight coefficients of the three factors were determined based on the sensitivity test results of the downstream CAE analysis. The downstream CAE analysis adopted static analysis and tested 10 different weight combinations. Finally, the weights of the geometric complexity factor (0.4), topological fit strength factor (0.3), and semantic tolerance factor (0.3) were determined. This combination minimizes the error of the CAE analysis.

[0048] The comprehensive influence of each feature region is calculated as follows: geometric complexity factor × corresponding weight + topological fit strength factor × corresponding weight + semantic tolerance factor × corresponding weight. Regions with a comprehensive influence of ≥0.7 are marked as key feature regions. The key feature regions have an impact of ≥5% on the accuracy of CAE analysis results. This is verified based on tests of 20 sets of standard parts.

[0049] The domain-adapted pre-trained BERT model is specifically designed as follows: a BERT-base model pre-trained on the publicly available CAD annotation dataset CADNet. CADNet contains 100,000 sets of PMI annotations for mechanical parts. LoRA layered fine-tuning is performed using industry-specific CAD annotation data, which consists of 50,000 sets of PMI annotations for aero-engine parts. The parameters of the first 8 pre-training layers are frozen, and only the last 4 domain adaptation layers are fine-tuned. The learning rate is 0.00005, and the training epochs are 15. The model output is the association confidence between semantic annotations and geometric elements. A confidence score ≥ 0.85 is considered a valid association.

[0050] As an optional embodiment: the model access and preprocessing module includes a format parsing unit, a defect detection unit, and an automatic healing unit;

[0051] The format parsing unit supports the parsing of CAD files in STEPAP242, IGES5.3, and CATIAV5 / V6CGR formats. It reads model data through the adapted interface and outputs geometric data containing the topological relationships of faces, edges, and vertices. The geometric data includes the number of faces and the edge association list.

[0052] The defect detection unit identifies defects such as gaps, geometric overlaps, and free surface patches using a geometric topology verification algorithm; the automatic healing unit executes an automatic healing algorithm based on tolerance.

[0053] The specific steps of the geometric topology verification algorithm are as follows:

[0054] Traverse the edges of all facets in the model, calculate the spatial distance and normal angle between adjacent facet edges. The spatial distance is calculated using the Euclidean distance formula, and the normal angle is calculated using the vector dot product. Edges with a distance greater than a preset distance threshold or a normal angle greater than a preset angle threshold are marked as facet gaps. The threshold is set to 0.01mm for IT7 level accuracy models and 0.02mm for IT8 level accuracy models. The preset angle threshold is 15°, based on the facet connection smoothness requirements.

[0055] Calculate the spatial intersection of any two faces using Boolean operations. Faces with an intersection area greater than a preset area threshold and whose normals are consistent are marked as geometrically overlapping. The preset area threshold is 0.5% of the total area of ​​the faces to avoid misjudging minor overlaps.

[0056] The percentage of edges without adjacent faces is calculated, and the edge percentage is the ratio of the length of the free edge to the total edge length of the face. Faces with a percentage greater than a preset threshold of 5% are marked as free faces, based on model integrity requirements. The tolerance-based automatic healing algorithm is as follows:

[0057] Based on the geometric accuracy requirements in the model design document, the healing tolerance is dynamically determined. The tolerance is 1 / 5 of the geometric accuracy requirement and not less than 0.005mm to avoid healing failure due to excessively small tolerance.

[0058] For gaps in the face plates, if the gap width is less than the tolerance, the edges of adjacent face plates are extended and stitched together. The extension length is equal to the gap width × 1.2, ensuring complete coverage of the gap. If the gap width is greater than the tolerance, a transition face plate is created and stitched together. The transition face plate is fitted with a B-spline surface. The curvature of the transition face plate is calculated by interpolating the curvature of adjacent face plates. The interpolation method is linear interpolation, ensuring that the curvature continuity meets the C1 level, i.e., the second derivative is continuous. This is verified based on 10 sets of surface models.

[0059] For geometric overlaps, retain the facets created earlier in the design tree, prioritize the preservation of basic features such as the base and main structure, delete the overlapping parts, use Boolean subtraction operations, and ensure the topological integrity of the facets through edge reconstruction after deletion, and call the interface to perform edge reconstruction;

[0060] For free facets, if their area is less than a preset area threshold, they are directly deleted. The preset area threshold is 0.1% of the total surface area of ​​the model. Otherwise, they are re-associated based on the topological relationship of their surrounding facets. The topological relationship includes adjacent vertices and edge overlap. After association, geometric verification is performed to ensure that no new defects are generated.

[0061] As an optional embodiment: the geometric analysis unit of the multi-dimensional feature collaborative extraction module includes a curvature calculation subunit, an anomaly removal subunit, and a geometric feature recognition subunit;

[0062] The curvature calculation subunit uses the local surface fitting method to calculate the principal curvature and Gaussian curvature of each vertex on the model surface; the anomaly removal subunit removes abnormal patches using the dynamic standard deviation algorithm; and the geometric feature recognition subunit identifies geometric features based on curvature parameters.

[0063] The specific steps of the local surface fitting method are as follows:

[0064] The number of neighborhood patch layers is dynamically determined based on the feature type to which the vertex belongs. Feature types include holes, curved surfaces, and sharp edges. Hole feature vertices are assigned 3 neighborhood patch layers because the curvature of the hole wall varies greatly, requiring more patches to ensure fitting accuracy. Curved surface feature vertices are assigned 2 neighborhood patch layers, and sharp edge feature vertices are assigned 1 neighborhood patch layer to avoid blurring of edge information. The number of neighborhood layers is determined through testing 10 sets of standard parts with different feature types.

[0065] Construct the quadratic surface equation z = ax² + by² + cxy + dx + ey + f, substitute the vertex coordinates (x, y) of the neighborhood patch, take the mean of the vertex coordinates as the fitting sample points, and solve the least squares solution using the LeastSquaresConjugateGradient solver of the Eigen 3.4.0 linear algebra library to obtain the coefficients of a, b, c, d, e, and f in the surface equation. The coefficients are kept to 8 decimal places.

[0066] The principal curvature and Gaussian curvature of each vertex are calculated based on the coefficients of the surface equation. The curvature matrix is ​​solved by eigenvalues. The deviation between the fitted curvature and the theoretical curvature is tested based on the standard spherical model.

[0067] The dynamic standard deviation algorithm is specifically as follows:

[0068] The set of distances from the neighborhood patch to the fitted surface is calculated using the shortest distance formula from a point to the surface. The standard deviation of the set of distances is then calculated using the sample standard deviation formula.

[0069] The standard deviation factor is adjusted based on the geometric consistency of neighboring patches. Geometric consistency is the angle between the patch normal and the fitted surface normal. An angle <10° indicates high consistency, and ≥10° indicates low consistency. The factor for high consistency is 2.3, and the factor for low consistency is 2.8. This factor is determined through anomaly patch removal accuracy testing, with a target removal accuracy of ≥99%.

[0070] Surface patches with a distance greater than a multiple of the standard deviation are marked as abnormal surfaces. After removing abnormal surfaces, local surface fitting is re-executed until the proportion of abnormal surfaces is less than a preset proportion threshold of 0.5%, based on fitting accuracy requirements. The geometric feature recognition subunit specifically comprises:

[0071] Based on the product of Gaussian curvature k and principal curvature difference k1-k2, geometric features such as sharp edges, rounded corners, bosses, and grooves are identified. The principal curvature is the maximum and minimum value of the normal curvature in all possible directions at that point, with the maximum value being k1 and the minimum value being k2.

[0072] Sharp edges correspond to Gaussian curvature K<0 and |k1-k2|>0.5mm, rounded corners correspond to K>0 and |k1-k2|<0.2mm, bosses correspond to K>0 and k1>0 and k2>0, and grooves correspond to K>0 and k1<0 and k2<0. The preset difference threshold is determined by testing the geometric feature recognition accuracy of 50 sets of mechanical parts.

[0073] As an optional embodiment: the topology analysis unit of the multi-dimensional feature collaborative extraction module includes a bounding box tree construction subunit, a feature pre-filtering subunit, and a cooperative region identification subunit. The bounding box tree construction subunit constructs a hierarchical AABB bounding box tree based on the functional grouping of model components.

[0074] The feature pre-filtering subunit filters out non-matching related components; the matching region identification subunit identifies matching regions in the assembly.

[0075] The specific steps for constructing sub-units using the hierarchical AABB bounding box tree are as follows:

[0076] Read the component function classification in the CAD model design tree, extract the classification information from the PartBodyProduct node of the CATIA design tree. The component types include bolt components, welded brackets, and decorative cover plates. Group the components according to their functions into connection component group, support component group, and decorative component group.

[0077] Build a root bounding box for each group of components. Calculate the minimum bounding box of the components using BndBox, with coordinate precision retained to 6 decimal places. The root bounding box contains all components in that group.

[0078] For each component, a sub-bounding box is constructed, calculated using BndBox, which contains all the geometric elements of the component.

[0079] A leaf bounding box is constructed for each geometric element. The geometric element includes a face, an edge, and a vertex. The leaf bounding box is attached to the surface of the geometric element. The maximum distance between the bounding box and the geometric element is <0.005mm, based on the collision detection accuracy requirements. The feature pre-filtering sub-unit is specifically: extracting attributes from the Attribute node of the CAD file based on the component attributes in the design tree.

[0080] The attributes include HasAssemblyConstraint and HasTolerance, which filter components without assembly constraints and without mating tolerance annotations. No assembly constraints correspond to HasAssemblyConstraint=False, and no mating tolerance annotations correspond to HasTolerance=False. Components include decorative covers, nameplates, etc. Only connecting components, supporting components, etc., that may have mating relationships are retained. Connecting components include bolts and nuts, and supporting components include brackets and bases. The filtering accuracy is ≥99%, based on testing with 100 assembly models.

[0081] The specific steps of the cooperation region identification subunit are as follows:

[0082] For the retained components, collision detection is performed using a hierarchical AABB bounding box tree. The BulletPhysics engine's btBroadphaseAabbPairCache is used for Broadphase detection, and btCollisionDispatcher is used for Narrowphase detection. The minimum spatial distance between components is calculated with an accuracy of ≤0.001mm.

[0083] For component pairs with a minimum spatial distance less than a preset distance threshold, read the assembly constraint type in the design tree and extract it from the Constraint node. The constraint type includes FastenerWeldInterferenceFit. The preset distance threshold is set to 0.1mm for bolted connections, 0.2mm for welding, and 0.05mm for interference fits. Verification is based on assembly connection strength test.

[0084] The mating area range is determined based on the assembly constraint type: for bolted connections, the mating area is the bolt hole and surrounding area, with a range equal to twice the bolt diameter; for welded connections, the mating area is the weld and heat-affected zone, with a range equal to 1 / 5 of the weld length; and for interference fits, the mating area is the mating surface and surrounding area, with a range equal to 1 / 3 of the mating surface width. This unit, through clearly defined bounding box construction steps, filtering rules, and mating area range calculation methods, ensures the reproducibility of module functions and resolves the issue of ambiguous module interaction logic.

[0085] As an optional embodiment: the semantic parsing unit of the multi-dimensional feature collaborative extraction module includes a PMI data extraction subunit, a semantic association subunit, and a semantic feature labeling subunit, wherein the PMI data extraction subunit extracts product manufacturing information from CAD files;

[0086] The semantic association subunit establishes the association between PMI annotations and geometric elements through a domain-adapted pre-trained BERT model; the semantic feature labeling subunit labels functional features based on the association results.

[0087] The PMI data extraction subunit is specifically as follows:

[0088] The PMI data structure of the CAD file is parsed through the PMIReader interface of the OSDA2.3 component to extract annotation information such as geometric tolerances, datums, surface roughness, and important dimensions. Geometric tolerances include cylindricity, flatness, and circular runout, and the tolerance values ​​and datum references are recorded. Datums include datum planes and datum axes, and the datum identifier and geometric element ID are recorded. Surface roughness records the Ra value and evaluation length. Important dimensions record the basic dimension, upper deviation, and lower deviation. The numerical values ​​of each annotation are recorded, and the precision is retained to 4 decimal places. The annotation position is recorded in three-dimensional coordinates, and the associated geometric element ID is recorded, which corresponds one-to-one with the geometric element ID output by the format parsing unit.

[0089] The specific steps of the semantic association subunit are as follows: input the PMI labeled text and the geometric element description text into the domain-adapted pre-trained BERT model. The PMI labeled text is, for example, cylindricity tolerance 0.01mm, and the geometric element description text is, for example, hole φ10mm. The text preprocessing uses WordPiece word segmentation, and the vocabulary includes CAD domain-specific terms, such as cylindricity interference fit.

[0090] The model outputs the association confidence score between PMI labels and geometric elements, with the confidence score ranging from 0 to 1. Association pairs with a confidence score ≥ 0.85 are considered valid associations. Based on semantic association accuracy testing, this threshold corresponds to an accuracy of ≥ 98.5%.

[0091] For invalid association pairs, with a confidence level <0.85, supplementary rules are applied: cylindricity and circular runout annotations are preferentially associated with hole and shaft geometric elements, with the geometric element ID containing hole and shaft keywords; flatness and parallelism annotations are preferentially associated with planar geometric elements, with the ID containing planar keywords. After supplementary rules, the association accuracy is improved to 99.2%. The semantic feature labeling subunit is specifically: based on the valid association results, the area where geometric elements labeled with geometric tolerances, datums, surface roughness, or important dimensions are located is marked as a functional feature area. The tolerance level of the geometric tolerance is ≥IT7, for example, IT7 grade cylindricity tolerance is 0.01mm, the datum level is A / B, the surface roughness Ra≤1.6μm, and the dimensional tolerance level of important dimensions is ≥IT8.

[0092] As an optional embodiment: the feature priority calculation unit of the dynamic priority grid optimization module includes a factor extraction subunit, a weight determination subunit, and a priority calculation subunit, wherein the factor extraction subunit extracts the calculation factor of feature influence degree;

[0093] The weight determination subunit determines the weight coefficient of each factor; the priority calculation subunit calculates the priority of each feature region.

[0094] The factor extraction subunit is specifically:

[0095] The mean curvature change rate of the feature region output by the geometric analysis unit is extracted as a geometric complexity factor. The value of this factor ranges from 0 to 1, reflecting the geometric fineness of the feature. The larger the value, the higher the mesh precision is required.

[0096] The proportion of the contact area of ​​the mating surfaces of the feature regions output by the topology analysis unit is extracted as the topology mating strength factor. The value of this factor ranges from 0 to 1, reflecting the assembly importance of the feature. The larger the value, the higher the mesh accuracy is required.

[0097] The semantic tolerance level quantization value of the feature region output by the semantic parsing unit is extracted as the semantic tolerance factor. The factor ranges from IT8 to IT5, reflecting the accuracy requirements of the feature. The larger the value, the higher the mesh accuracy is required. The weight determination sub-unit is specifically: the accuracy of the downstream CAE analysis results under different weight combinations is tested through orthogonal experiments. The stress error of static analysis is used as the indicator. The experiment is designed with 3 factors and 3 levels. The geometric weight is set to three levels: 0.3, 0.4, and 0.5. The topological weight is set to three levels: 0.2, 0.3, and 0.4. The semantic weight is set to three levels: 0.2, 0.3, and 0.4. A total of 27 sets of experiments were conducted. The results show that when the geometric weight is 0.4, the topological weight is 0.3, and the semantic weight is 0.3, the CAE analysis error is the smallest, with an error of <2%. Therefore, this weight combination is selected.

[0098] The priority calculation subunit specifically calculates the comprehensive priority of each feature region as follows: Geometric complexity factor × 0.4 + Topological fit strength factor × 0.3 + Semantic tolerance factor × 0.3. The comprehensive priority is divided into three levels: high, medium, and low. High priority corresponds to a comprehensive priority ≥ 0.7, medium priority corresponds to 0.5-0.7, and low priority corresponds to < 0.5. Based on the CAE analysis accuracy requirements, high priority corresponds to an error < 2%, medium priority corresponds to an error < 3%, and low priority corresponds to an error < 5%. The three levels correspond to key feature regions, ordinary key regions, and non-key regions, respectively.

[0099] As an optional embodiment: the adaptive mesh generation unit of the dynamic priority mesh optimization module includes a density calculation subunit, a mesh generation subunit, and a transition layer generation subunit. The density calculation subunit determines the mesh unit size based on feature priority. The mesh generation subunit calls Gmsh to generate a tetrahedral mesh.

[0100] The transition layer generates sub-units that generate gradient transition meshes in regions with different densities.

[0101] The density calculation subunit is specifically:

[0102] Based on the feature priority level and the mesh accuracy requirements of downstream CAE analysis, the mesh element size range corresponding to each priority is determined. For example, the downstream CAE analysis requires a minimum element size of 0.1 mm in ANSYS Workbench static analysis. High priority corresponds to critical feature areas, and the element size range is 0.8-1.0 times the minimum element size required by CAE. Medium priority corresponds to ordinary critical areas, and the element size range is 1.0-1.5 times. Low priority corresponds to non-critical areas, and the element size range is 1.5-2.5 times.

[0103] The unit size is dynamically adjusted based on the geometric complexity of the feature region. High geometric complexity corresponds to an average curvature change rate > 0.5, in which case the lower limit of the size range is taken. Low geometric complexity corresponds to an average curvature change rate < 0.3, in which case the upper limit of the size range is taken. The adjustment is based on the mesh accuracy test of 10 sets of features with different geometric complexity to ensure that the feature mesh accuracy of high geometric complexity is sufficient.

[0104] The specific mesh generation sub-unit is as follows: integrate Gmsh4.11.1SDK, create a model through the gmsh:model:add interface, and call gmsh:model:geo:addPoint, addLine, addSurface, and addVolume to construct geometric entities;

[0105] Set the mesh generation parameters using gmsh:model:mesh:setTransfiniteAutomatic, passing in the cell size parameters for each feature region. For example, set the cell size parameter for the core region to 0.08mm, and then call gmsh:model:mesh:generate.

[0106] Tetrahedral meshes are generated, and deformed cells are filtered in real time during the generation process. Deformed cells correspond to cells with a Jacobian coefficient <0.6, and the filtering rate is <0.1%. The specific steps for generating sub-cells in the transition layer are as follows: 1. Calculate the cell size ratio of adjacent feature regions. The size ratio is the ratio of the large size to the small size. If the size ratio is >2.0, a transition layer is set to avoid abrupt size changes that could lead to a decrease in mesh quality.

[0107] The number of transition layers is determined based on the size ratio: 3 layers for a size ratio of 2.0-3.0, 5 layers for 3.0-4.0, and 8 layers for a size ratio greater than or equal to 4.0. This number of layers is determined through mesh transition smoothness testing, and the size gradient change of the transition layer elements is ≤50%.

[0108] The gmsh:model:mesh:refine interface of Gmsh is called. The size of the transition layer unit increases from the high priority region to the low priority region in a geometric sequence with a common ratio of 1.2 to ensure smooth size transition. The common ratio of 1.2 corresponds to a size gradient of 20%, which meets the gradient requirement of ≤50%.

[0109] As an optional embodiment: the mesh quality optimization unit of the dynamic priority mesh optimization module includes a quality evaluation sub-unit, a smoothing optimization sub-unit, and a re-partitioning sub-unit, wherein the quality evaluation sub-unit calculates the quality parameters of the mesh unit;

[0110] The smoothing optimization subunit executes the Laplace smoothing algorithm; the repartitioning subunit repartitions unqualified units.

[0111] The quality assessment subunit is specifically:

[0112] Calculate the Jacobian coefficient, aspect ratio, and warpage for each tetrahedral mesh element. The Jacobian coefficient is the ratio of the minimum to the maximum determinant of the element's Jacobian matrix, ranging from 0 to 1. The aspect ratio is the ratio of the longest to the shortest side of the element. The warpage is the angle of deviation of the element from the plane containing its four vertices. The calculation is performed by fitting the plane equation. Standard values ​​include Jacobian coefficient ≥ 0.7, aspect ratio ≤ 5, and warpage ≤ 15°. Element that is not qualified is marked if Jacobian coefficient < standard value, aspect ratio > standard value, or warpage > standard value.

[0113] The specific steps of the Laplace smoothing optimization subunit are as follows:

[0114] The smoothing step size is dynamically adjusted based on the element distortion degree. The element distortion degree is calculated as 1 - element Jacobian coefficient / 0.7, with a value range of 0-1. When the distortion degree is <0.2, the step size is 0.3; when the distortion degree is between 0.2 and 0.4, the step size is 0.2; and when the distortion degree is greater than 0.4, the step size is 0.1. The step size setting is based on a test that balances mesh smoothing efficiency and feature preservation, avoiding feature loss due to an excessively large step size and low efficiency due to an excessively small step size.

[0115] Laplacian smoothing is applied to the vertices of the mesh cells. The new coordinates of the vertices are equal to the weighted average of the coordinates of the neighboring vertices, with the weight being the area of ​​the neighboring cells. The vertex coordinates are adjusted to bring the cell quality parameters closer to the standard values.

[0116] The smoothing operation is performed iteratively until the rate of change of the mesh quality parameters between two adjacent iterations is less than 0.01 or the number of iterations reaches the preset upper limit. The upper limit of the number of iterations is dynamically determined based on the number of cells.

[0117] In this embodiment, the number of times can be set to 10 times for up to 100,000 units, 15 times for 100,000 to 500,000 units, and 20 times for more than 500,000 units.

[0118] The re-division of sub-units specifically involves: for sub-units that are still unqualified after smoothing optimization, reducing the sub-unit size by a reduction ratio of 1 - 0.2 × distortion degree, with a maximum reduction of 0.3. Then, the mesh is regenerated and sub-units are re-divided. After re-division, a quality assessment is performed again until the proportion of unqualified sub-units is less than 1%. Based on the overall mesh quality requirements, this proportion can meet the accuracy of CAE analysis. This unit ensures the reproducibility of the mesh quality optimization process and resolves issues of terminological ambiguity and logical contradictions through clear quality assessment standards, smoothing step size adjustment logic, and re-division rules.

[0119] As an optional embodiment: the sampling calculation unit of the closed-loop accuracy verification module includes a sampling planning subunit, a distance calculation subunit, and an error statistics subunit, wherein the sampling planning subunit determines the sampling strategy based on feature priority;

[0120] The distance calculation subunit calculates the distance error of the sampling points; the error statistics subunit statistically analyzes the error results.

[0121] The specific steps of the sampling planning subunit are as follows:

[0122] For key feature regions, the comprehensive priority is ≥0.7. The sampling density is determined according to the geometric complexity of the features. High geometric complexity corresponds to a mean curvature change rate >0.5, at which point the sampling density is 50 points / mm². Medium geometric complexity corresponds to 0.3-0.5, at which point the sampling density is 35 points / mm². Low geometric complexity corresponds to <0.3, at which point the sampling density is 20 points / mm². The sampling density is determined through error convergence testing. Error convergence occurs at 50 points / mm², and further increasing the density results in an error change of <0.1%.

[0123] For ordinary key regions, the overall priority is 0.5-0.7, and the sampling density is 0.6 times that of key feature regions.

[0124] For non-critical regions, the overall priority is <0.5, and the sampling density is 0.3 times that of critical feature regions.

[0125] Latin hypercube sampling method is used to generate sampling points in each feature region. The LatinHypercube sampling function of SciPy library is used to ensure that the sampling points are evenly distributed and the sampling point spacing deviation is <10%. The uniformity of sampling point distribution is tested.

[0126] The distance calculation subunit specifically performs the following steps: For each sampling point, calculate the shortest distance to the surface of the lightweight mesh model using a point-to-surface projection algorithm. This is achieved by calling the OpenCASCADE BRepProjProjectPointOnShape interface, with a calculation accuracy ≤0.001mm. Based on a standard spherical model test, the distance error from the sampling point to the sphere is <0.001mm, which is the surface approximation error of the sampling point. For hole and shaft features, extract sampling points, with ≥30 sampling points at the hole edge. Fit feature dimensions, such as the hole diameter, using the least squares method to fit a circle equation. Calculate the difference between the fitted dimension and the original model's labeled dimension, which is the size error. The error statistics subunit specifically performs the following steps: Statistically calculate the surface approximation error and feature size error for all sampling points. The surface approximation error includes the maximum deviation and RMS deviation, with accuracy retained to 4 decimal places. The feature size error includes the maximum deviation and average deviation. Compare these with a preset error tolerance, which is the maximum error allowed by downstream CAE analysis.

[0127] As an optional embodiment: the feedback adjustment unit of the closed-loop accuracy verification module includes a feature threshold adjustment subunit and a grid parameter adjustment subunit. The feature threshold adjustment subunit adjusts the feature screening threshold based on the error result; the grid parameter adjustment subunit adjusts the grid density parameter based on the error result.

[0128] The specific steps of the feature threshold adjustment subunit are as follows:

[0129] If the error of a certain feature area is greater than the preset tolerance, for example, the tolerance is set to 0.02mm, the feature screening process of that area is analyzed. If the feature influence value is close to the screening threshold, for example, when the comprehensive value is close to the threshold of 0.7, the comprehensive value is 0.68, and it is misjudged as low priority, then the screening threshold is reduced. The reduction amount is 0.05×(error-tolerance) / tolerance. For example, when the error is 0.03mm, the reduction amount is 0.05×(0.03-0.02) / 0.02=0.025, and the threshold is reduced from 0.7 to 0.675, ensuring that the key area can be correctly identified after adjustment.

[0130] If the semantic association confidence is insufficient, for example, 0.83, resulting in the omission of feature labels, then the semantic association confidence threshold is reduced by 0.03 × (error - tolerance) / tolerance. For example, when the error is 0.03 mm, the reduction is 0.03 × 0.5 = 0.015, and the threshold is reduced from 0.85 to 0.835.

[0131] After adjustments, multi-dimensional feature collaborative extraction was re-executed to generate a new feature map. The accuracy of key feature recognition in the new map was improved by ≥1.5%.

[0132] The specific steps of the grid parameter adjustment subunit are as follows: if the error of a certain feature region is greater than the preset tolerance, calculate the error excess ratio, where excess ratio = (error - tolerance) / tolerance. For example, when the error is 0.03mm, the excess ratio is 50%.

[0133] Reduce the size of the grid cells in this area according to the excess tolerance ratio, with a reduction amount of 0.1 × excess tolerance ratio and a maximum reduction amount of 0.3, to avoid excessive densification that could lead to a decrease in efficiency;

[0134] If the size ratio of adjacent areas is too large, for example, size ratio > 4.0, causing the transition layer error to exceed the tolerance, then the number of transition layers will be increased. The increase number = 1 × floor (tolerance ratio / 0.2). For example, if the tolerance ratio is 50%, 2 layers will be added.

[0135] After adjustment, dynamic priority mesh optimization is re-executed to generate a new lightweight mesh model. The error of the new model is reduced by ≥30%. The feedback adjustment process is executed iteratively until the error of all feature regions is ≤ preset tolerance or the number of iterations reaches 3. This avoids infinite iteration and ensures system efficiency. The probability of still exceeding the tolerance after 3 iterations is <1%. If the error still exceeds the tolerance after 3 iterations, an out-of-tolerance region report is output. The report includes the coordinates of the out-of-tolerance region, the error value, and the suggested adjustment direction.

[0136] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A lightweight digital prototype system based on feature extraction and adaptive mesh optimization, characterized in that, Includes a model access and preprocessing module, and a file parsing unit; The file parsing unit is used to receive the original digital prototype model and parse the geometric topology data and design tree information of the model; The multi-dimensional feature collaborative extraction module includes a geometric analysis unit, a topological analysis unit, a semantic parsing unit, and a feature collaborative decision-making unit; The feature collaborative decision-making unit receives the output results of the three units, and realizes the screening and classification of multi-dimensional features through feature influence weight calculation, generating a feature mapping map containing key feature regions and non-key feature regions. Dynamic priority grid optimization module, It includes a feature priority calculation unit, an adaptive mesh generation unit, and a mesh quality optimization unit. The feature priority calculation unit dynamically calculates the optimization priority of each feature region based on the geometric complexity, semantic tolerance level, and topological fit strength of the features in the feature map. The geometric complexity is represented by the mean curvature change rate. The closed-loop accuracy verification module integrates a sampling calculation unit and a feedback adjustment unit. The sampling calculation unit dynamically adjusts the sampling strategy based on feature priority. The geometric analysis unit of the multi-dimensional feature collaborative extraction module includes a curvature calculation subunit, an anomaly removal subunit, and a geometric feature recognition subunit; The curvature calculation subunit uses the local surface fitting method to calculate the principal curvature and Gaussian curvature of each vertex on the model surface; the anomaly removal subunit removes abnormal patches using the dynamic standard deviation algorithm; and the geometric feature recognition subunit identifies geometric features based on curvature parameters. The topology analysis unit of the multi-dimensional feature collaborative extraction module includes a bounding box tree construction subunit, a feature pre-filtering subunit, and a cooperative region identification subunit. The bounding box tree construction subunit constructs a hierarchical AABB bounding box tree based on the functional grouping of model components. The feature pre-filtering subunit filters out non-fitting related components; the fitting region identification subunit identifies fitting regions in the assembly. The semantic parsing unit of the multi-dimensional feature collaborative extraction module includes a PMI data extraction subunit, a semantic association subunit, and a semantic feature labeling subunit. The PMI data extraction subunit extracts product manufacturing information from CAD files. The semantic association subunit establishes the association between PMI annotations and geometric elements through a domain-adapted pre-trained BERT model; the semantic feature labeling subunit labels functional features based on the association results.

2. The lightweight digital prototype system based on feature extraction and adaptive mesh optimization according to claim 1, characterized in that: The model access and preprocessing module includes a format parsing unit, a defect detection unit, and an automatic healing unit; The format parsing unit supports CAD file parsing, reads model data through the adapted interface, and outputs geometric data containing topological relationships of faces, edges, and vertices, including the number of faces and the edge association list. The defect detection unit identifies defects such as gaps, geometric overlaps, and free surface patches using a geometric topology verification algorithm; the automatic healing unit executes an automatic healing algorithm based on tolerance.

3. The lightweight digital prototype system based on feature extraction and adaptive mesh optimization according to claim 1, characterized in that: The feature priority calculation unit of the dynamic priority grid optimization module includes a factor extraction subunit, a weight determination subunit, and a priority calculation subunit. The factor extraction subunit extracts the calculation factors for the feature influence degree. The weight determination subunit determines the weight coefficient of each factor; the priority calculation subunit calculates the priority of each feature region.

4. The lightweight digital prototype system based on feature extraction and adaptive mesh optimization according to claim 1, characterized in that: The adaptive mesh generation unit of the dynamic priority mesh optimization module includes a density calculation subunit, a mesh generation subunit, and a transition layer generation subunit. The density calculation subunit determines the mesh unit size based on feature priority. The mesh generation subunit calls Gmsh to generate a tetrahedral mesh. The transition layer generates sub-units that generate gradient transition meshes in regions with different densities.

5. The lightweight digital prototype system based on feature extraction and adaptive mesh optimization according to claim 4, characterized in that: The dynamic priority mesh optimization module includes a mesh quality optimization unit comprising a quality assessment subunit, a smoothing optimization subunit, and a repartitioning subunit. The quality assessment subunit calculates the quality parameters of the mesh unit. The smoothing optimization subunit executes the Laplace smoothing algorithm; the repartitioning subunit repartitions unqualified units.

6. The lightweight digital prototype system based on feature extraction and adaptive mesh optimization according to claim 1, characterized in that: The sampling calculation unit of the closed-loop accuracy verification module includes a sampling planning subunit, a distance calculation subunit, and an error statistics subunit. The sampling planning subunit determines the sampling strategy based on feature priority. The distance calculation subunit calculates the distance error of the sampling points; the error statistics subunit statistically analyzes the error results.

7. The lightweight digital prototype system based on feature extraction and adaptive mesh optimization according to claim 1, characterized in that: The feedback adjustment unit of the closed-loop accuracy verification module includes a feature threshold adjustment subunit and a grid parameter adjustment subunit. The feature threshold adjustment subunit adjusts the feature screening threshold based on the error result; the grid parameter adjustment subunit adjusts the grid density parameter based on the error result.

Images