An engine part geometric information extraction and reconstruction method based on parameter rule matching

Abstract

The application discloses an engine part geometric information extraction and reconstruction method based on parameter rule matching, relates to the technical field of computer-aided design, and comprises the following steps: performing pretreatment on a CAD model to generate standard part geometric data; extracting geometric attributes and topological relations to generate a geometric primitive topological graph and a derived point set; performing a Harris 3D algorithm to generate a key geometric candidate point set; establishing an evidence index and a local geometric block; constructing an improved PointMLP model to generate a geometric parameter evidence set; constructing a reference skeleton and a parameter rule graph to generate a candidate parameter locking result; performing counter-evidence constraint verification and backtracking correction to generate an effective geometric parameter set; and performing topological preservation reconstruction to output geometric information extraction results and model reconstruction results. The application effectively improves engine part parameter extraction accuracy, functional reference consistency and reconstruction topological integrity.

Description

Technical Field

[0001] This invention relates to the field of computer-aided design technology, and in particular to a method for extracting and reconstructing geometric information of engine parts based on parameter rule matching. Background Technology

[0002] Computer-aided design (CAD) technology is applied in the design, verification, machining preparation, and digital archiving of engine parts. Engine part CAD models contain surfaces, edges, vertices, boundary loops, holes, mounting faces, load-bearing sections, and connecting transition structures. The results of geometric parameter extraction directly affect model reuse, parametric reconstruction, and subsequent engineering analysis. Existing CAD geometric information extraction methods mostly rely on geometric feature recognition, point cloud sampling, rule template matching, or manual annotation processes. These methods lack a unified expression for surface attributes, topological references, hole boundaries, and functional benchmarks in the model, making it difficult to form a parametric evidence chain for engine part structures.

[0003] Existing methods, when processing complex engine part models, often rely on local coordinates or curvature changes for geometric key point detection, lacking the involvement of CAD attributes and topological connections. This leads to missed or false detections of hole locations, boundaries, end face intersections, and transitional connection regions. Existing point cloud networks primarily focus on learning features from ordinary point sets, making it difficult to integrate surface categories, boundary types, hole location markers, and topological adjacency markers. Current parameter rule matching methods often employ static rule bases, lacking historical model data support. The separation of candidate parameter locking and structural constraint verification results in inconsistencies between parameter extraction results and the functional structure of the engine part. This leads to topological breaks, parameter omissions, and disorganized reconstruction sequences during model reconstruction.

[0004] Therefore, how to provide a method for extracting and reconstructing the geometric information of engine parts based on parameter rule matching is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for extracting and reconstructing geometric information of engine parts based on parameter rule matching. This invention combines CAD model preprocessing, Harris3D key geometric point detection, improved PointMLP evidence extraction, historical model data-driven parameter rule graphs, and contradictory evidence constraint verification to complete the locking of geometric parameters and topology-preserving reconstruction of engine parts. It has the advantages of clear parameter sources, consistent functional benchmarks, stable reconstruction topology, and high degree of automation.

[0006] A method for extracting and reconstructing geometric information of engine parts based on parameter rule matching according to an embodiment of the present invention includes:

[0007] Obtain the CAD model of the engine part to be processed, perform preprocessing on the CAD model of the engine part to be processed, and generate standard part geometric data;

[0008] Geometric attributes and topological relationships are extracted from standard part geometric data to generate a geometric primitive topology diagram and a derived point set with CAD attribute tags.

[0009] The Harris3D algorithm is used to perform key geometric point detection on the derived point set to generate a set of key geometric candidate points.

[0010] The key geometric candidate point set is back-mapped to the geometric primitive topology graph, a geometric evidence index is established, and a benchmarked local geometric block is constructed.

[0011] An improved PointMLP model is constructed by inputting a benchmarked local geometric block into the improved PointMLP model and outputting a set of geometric parameter evidence.

[0012] Based on the geometric parameter evidence set, a functional baseline skeleton of engine parts is generated, an engine part parameter rule diagram is constructed, and a chain matching is performed between the engine part functional baseline skeleton and the engine part parameter rule diagram to generate candidate parameter locking results.

[0013] By introducing structural constraints on engine parts, the candidate parameter locking results are verified by contradiction constraints, and the candidate parameter locking results that fail the verification are backtracked and corrected to generate a set of valid geometric parameters.

[0014] Based on the effective set of geometric parameters, topology-preserving reconstruction is performed in the order of functional structure to generate the reconstructed engine part model, and the geometric information extraction results and model reconstruction results are output.

[0015] Optionally, generating standard part geometry data includes:

[0016] Read the file header information, geometric entity data, and topology reference data of the CAD model of the engine part to be processed, and generate the original model parsing data;

[0017] Based on the original model, the data parsing extracts the format identifier, geometric unit, coordinate datum, and boundary representation entity to generate the model datum record;

[0018] Based on the model datum record, perform consistency processing on geometric units, coordinate datums, and boundary representation entities to generate datum-consistent geometric data;

[0019] Extract surface entities, edge entities, vertex entities, boundary loop entities, and patch topological references from the datum-consistent geometric data to generate standard part geometric data.

[0020] Optionally, generating the derived point set with CAD attribute tags includes:

[0021] Read the topological reference relationships of surface entities, edge entities, vertex entities, boundary loop entities, and patches from the standard part geometry data, and generate a geometric entity index;

[0022] Based on the geometric entity index, extract the surface category, curvature value, normal vector, boundary type, hole position mark, hole position geometric region and adjacent surface connection relationship to generate geometric attribute records;

[0023] Based on the geometric entity index and geometric attribute records, a geometric primitive topology graph is established using surface entities, boundary ring entities, and hole geometric regions as nodes, and using patch edge sharing relationships, boundary closure relationships, hole adjacency relationships, and geometric constraint relationships as connecting edges.

[0024] Based on the geometric primitive topology diagram, the surface, boundary loop and hole position are sampled, the coordinates of the sampling point are calculated, and the surface category, curvature value, normal vector, boundary type, hole position mark and topological adjacency mark are written for the sampling point to generate a derived point set with CAD attribute marks.

[0025] Optionally, generating the set of key geometric candidate points includes:

[0026] Read the coordinates, normal vectors, curvature values, surface categories, boundary types, hole position markers, and topological adjacency markers of the derived point set with CAD attribute tags, and generate key point detection input data;

[0027] Based on the key point detection input data, surface detection neighborhood, boundary detection neighborhood and hole detection neighborhood are established respectively. Covariance statistics are performed on the normal distribution of sampling points in each detection neighborhood to generate multi-scale normal covariance records.

[0028] The underlying Harris response is calculated based on the multi-scale normal covariance record, and the responses at different scales are fused to generate a multi-scale underlying response record.

[0029] CAD topological weights are generated based on surface type, boundary type, hole location markers, and topological adjacency markers. The CAD topological weights are then written into the multi-scale basic response record to generate a topologically weighted Harris response record.

[0030] Based on the topologically weighted Harris response record, the sampling points corresponding to the locations of the parametric evidence are screened, and the sampling points in the same boundary, the same hole location, and the same connection area are merged to generate a set of key geometric candidate points.

[0031] Optionally, the construction of the benchmarked local geometry block includes:

[0032] Read the set of key geometric candidate points and the topology of geometric primitives, and perform back mapping based on the sampling point coordinates, surface type, curvature value, normal vector, boundary type, hole position mark, topological adjacency mark and connecting edge to generate candidate point location records;

[0033] Establish an index relationship between the candidate point location records and surface nodes, boundary nodes, hole location nodes, and connecting edges to generate a geometric evidence index;

[0034] Based on the geometric evidence index, derived points corresponding to homologous surfaces, closed boundaries, aperture neighborhoods and topologically adjacent regions are extracted to generate evidence-constrained local geometric blocks.

[0035] The local reference coordinate records are calculated based on the hole location neighborhood, closed boundary, adjacent surface normal and connecting edge direction in the evidence-constrained local geometry block, and the coordinate reference processing is performed on the evidence-constrained local geometry block to generate a reference local geometry block.

[0036] Optionally, the output geometric parameter evidence set includes:

[0037] An improved PointMLP model is constructed. The improved PointMLP model retains the residual MLP encoding unit of the original PointMLP model and sets up a CAD attribute enhanced input layer, a geometry-topology dual neighborhood aggregation layer and a parameter evidence feedback correction layer.

[0038] Read the normalized coordinates, surface category, curvature value, normal vector, boundary type, hole position mark, topological adjacency mark and geometric evidence index from the benchmark local geometry block, and perform attribute embedding and channel splicing through the CAD attribute enhancement input layer to generate point-level CAD attribute features;

[0039] Geometric and topological neighborhood features are established based on point-level CAD attribute characteristics;

[0040] The geometric-topological dual-neighborhood aggregation layer performs bi-branch residual encoding, neighborhood weight allocation, and cross-neighborhood fusion on the geometric and topological neighborhood features to generate geometric-topological joint features.

[0041] The residual MLP encoding unit performs layer-by-layer nonlinear mapping, residual connection and local feature aggregation on the geometric-topological joint features to generate local geometric semantic features.

[0042] The parameter evidence feedback correction layer performs rule association correction on the local geometric semantic features, generating geometric primitive evidence, functional benchmark evidence, parameter candidate evidence and topological relationship evidence, and summarizing them to generate a geometric parameter evidence set;

[0043] The improved PointMLP model was trained using geometric primitive evidence classification error, functional baseline evidence localization error, parametric candidate evidence regression error, topological relationship evidence discrimination error, and rule feedback consistency error as joint optimization objectives. The parameters of the CAD attribute enhancement input layer, geometric-topological dual-neighborhood aggregation layer, residual MLP encoding unit, and parametric evidence feedback correction layer were optimized. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved PointMLP model was considered to have completed convergence training.

[0044] Optionally, the improved PointMLP model includes:

[0045] While retaining the local point set grouping, residual MLP encoding, point-by-point feature mapping, and multi-head output backbone of the PointMLP model, structural modifications are made to the point-level input, neighborhood organization, feature propagation, and evidence output around the task of extracting and reconstructing CAD geometric information of engine parts.

[0046] A CAD attribute enhancement input layer is added to the model input end, and the normalized coordinates, surface categories, curvature values, normal vectors, boundary types, hole position markers, topological adjacency markers and geometric evidence indexes in the benchmark local geometry are embedded into the evidence channel to generate point-level CAD attribute features.

[0047] A geometric-topological dual-neighborhood aggregation layer is added at the local feature aggregation location to separately encode the spatial distance neighborhood, curvature change neighborhood, common edge topology neighborhood, and aperture location association neighborhood, and perform cross-neighborhood feature fusion through neighborhood consistency weight;

[0048] The residual MLP coding unit is improved by transforming the single path point feature mapping into a residual interactive coding structure with geometric branches, topological branches and evidence fusion branches to generate local geometric semantic features.

[0049] A parameter evidence feedback correction layer is added to the output end to perform rule association correction and evidence confidence recalibration on local geometric semantic features, and output geometric primitive evidence, functional benchmark evidence, parameter candidate evidence and topological relationship evidence.

[0050] Optionally, the generation of candidate parameter locking results includes:

[0051] Read the geometric parameter evidence set and geometric primitive topology diagram, and extract the functional axis, mounting end face, hole center, load-bearing section and connection transition relationship based on geometric primitive evidence, functional benchmark evidence, parameter candidate evidence and topological relationship evidence to generate the functional benchmark skeleton of engine parts;

[0052] Retrieve and receive historical model data, extract part types, historical parameter items, geometric evidence sources, measurement relationships and structural constraints from the historical model data, and generate historical parameter rule records;

[0053] Based on historical parameter rule records, an engine part parameter rule library is generated, and an engine part parameter rule diagram is constructed by combining a geometric parameter evidence set and an engine part functional benchmark skeleton.

[0054] Perform chain-like fuzzy matching between the functional baseline skeleton of engine parts and the parameter rule diagram of engine parts, calculate the topological consistency, geometric measurability and structural constraint fit, and generate parameter matching score records;

[0055] Based on the parameter matching score records, perform association locking on the parameter candidate evidence, functional baseline nodes, and geometric evidence sources to generate candidate parameter locking results.

[0056] Optionally, generating the effective set of geometric parameters includes:

[0057] Read the candidate parameter locking results, engine part parameter rule diagram and historical parameter rule record, extract structural constraint terms from the engine part parameter rule diagram and historical parameter rule record, and generate engine part structural constraints;

[0058] Based on the structural constraints of engine parts, perform counter-evidence constraint verification on the parameter values, functional reference nodes, geometric evidence sources and measurement relationships in the candidate parameter locking results, and generate counter-evidence constraint verification records;

[0059] Based on the verification record of the evidence of contradiction constraint, the constraint conflict parameter, the evidence missing parameter, and the matching offset parameter are marked, anomaly parameter records are generated.

[0060] Based on the abnormal parameter records, the evidence set of geometric parameters, the functional baseline skeleton of engine parts and the parameter rule diagram of engine parts are traced back. The candidate parameter locking results are then subjected to correlation replacement, evidence supplementation and matching score update to generate corrected parameter locking results.

[0061] Perform consistency cleanup on the modified parameter locking results to generate a valid set of geometric parameters.

[0062] Optionally, the output geometric information extraction results and model reconstruction results include:

[0063] Read the effective set of geometric parameters, geometric primitive topology diagram and functional reference skeleton of engine parts, and generate a reconstruction sequence record according to functional axis, mounting end face, hole center, load-bearing section and connection transition relationship;

[0064] Based on the reconstruction order, extract parameter values, functional baseline nodes, and sources of geometric evidence from the effective geometric parameter set to generate a reconstruction parameter sequence;

[0065] Based on the reconstruction parameter sequence, topology-preserving reconstruction is performed on surface entities, edge entities, vertex entities, and boundary loop entities to generate reconstructed engine part models.

[0066] The system summarizes the effective set of geometric parameters, the sequence of reconstructed parameters, and the reconstructed engine part model, and outputs the geometric information extraction results and the model reconstruction results.

[0067] The beneficial effects of this invention are:

[0068] This invention proposes a method for extracting and reconstructing geometric information of engine parts based on parameter rule matching. Through CAD model preprocessing, joint topological expression of surface-boundary-hole positions, construction of derived point sets with CAD attribute tags, CAD topology weighted Harris3D key geometric point detection, geometric evidence indexing, and benchmarked local geometric block construction process, the method transforms surface entities, boundary ring entities, hole geometric regions, and topological connection edges in the CAD model of engine parts into computable geometric evidence. This improves the stability of key point detection for hole positions, boundaries, end face intersections, and transition connection regions, and reduces the risk of ambiguity in the source of geometric parameters and missed detection of key structures.

[0069] This invention extracts a set of geometric parameter evidence by improving the PointMLP model, generates a parameter rule diagram of engine parts by combining it with historical model data, and performs chain matching between the functional benchmark skeleton of engine parts and the parameter rule diagram of engine parts to form candidate parameter locking results. Then, it generates a set of valid geometric parameters by verifying by contradiction and backtracking correction, and performs topology preservation reconstruction based on the set of valid geometric parameters. This improves the accuracy of parameter locking, the consistency of functional benchmarks and the stability of model reconstruction, and reduces parameter omissions, topology breaks and reconstruction order disorder. Attached Figure Description

[0070] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0071] Figure 1 This is a flowchart of a method for extracting and reconstructing geometric information of engine parts based on parameter rule matching proposed in this invention;

[0072] Figure 2 This is a comparison diagram of typical test sample input and reconstruction results in an embodiment of the method for extracting and reconstructing geometric information of engine parts based on parameter rule matching proposed in this invention;

[0073] Figure 3 This is a comparison of the geometric information extraction and model reconstruction effects of the proposed method for extracting and reconstructing engine parts based on parameter rule matching and traditional methods.

[0074] Figure 4 This is a training convergence graph of the improved PointMLP model for extracting and reconstructing geometric information of engine parts based on parameter rule matching, as proposed in this invention. Detailed Implementation

[0075] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0076] refer to Figure 1 A method for extracting and reconstructing geometric information of engine parts based on parameter rule matching, comprising:

[0077] Obtain the CAD model of the engine part to be processed, perform preprocessing on the CAD model of the engine part to be processed, and generate standard part geometric data;

[0078] Geometric attributes and topological relationships are extracted from standard part geometric data to generate a geometric primitive topology diagram and a derived point set with CAD attribute tags.

[0079] The Harris3D algorithm is used to perform key geometric point detection on the derived point set to generate a set of key geometric candidate points.

[0080] The key geometric candidate point set is back-mapped to the geometric primitive topology graph, a geometric evidence index is established, and a benchmarked local geometric block is constructed.

[0081] An improved PointMLP model is constructed by inputting a benchmarked local geometric block into the improved PointMLP model and outputting a set of geometric parameter evidence.

[0082] Based on the geometric parameter evidence set, a functional baseline skeleton of engine parts is generated, an engine part parameter rule diagram is constructed, and a chain matching is performed between the engine part functional baseline skeleton and the engine part parameter rule diagram to generate candidate parameter locking results.

[0083] By introducing structural constraints on engine parts, the candidate parameter locking results are verified by contradiction constraints, and the candidate parameter locking results that fail the verification are backtracked and corrected to generate a set of valid geometric parameters.

[0084] Based on the effective set of geometric parameters, topology-preserving reconstruction is performed in the order of functional structure to generate the reconstructed engine part model, and the geometric information extraction results and model reconstruction results are output.

[0085] In this embodiment, generating standard part geometric data includes:

[0086] Read the file header information, geometric entity data, and topology reference data of the CAD model of the engine part to be processed, and generate the original model parsing data;

[0087] Based on the original model, the data parsing extracts the format identifier, geometric unit, coordinate datum, and boundary representation entity to generate the model datum record;

[0088] Based on the model datum record, perform consistency processing on geometric units, coordinate datums, and boundary representation entities to generate datum-consistent geometric data;

[0089] Extract surface entities, edge entities, vertex entities, boundary loop entities, and patch topological references from the datum-consistent geometric data to generate standard part geometric data.

[0090] In this embodiment, generating the derived point set with CAD attribute tags includes:

[0091] Read the surface solids, edge solids, vertex solids, boundary loop solids, and patch topological references from the standard part geometry data to generate a geometric solid index, where:

[0092] Generate a geometric entity index, specifically:

[0093] Read the topological references of surface solids, edge solids, vertex solids, boundary loop solids, and facets from standard part geometry data;

[0094] Extract the entity label field, entity type field, parsing order, and topological reference level for each entity;

[0095] When the entity label field exists, the entity label field is used as the entity index identifier;

[0096] When the entity label field is empty, generate entity index identifiers according to the entity type field and the parsing order;

[0097] Based on the entity index identifier, record the surface parameter expression, surface type field, surface parameter domain, edge curve parameter expression, edge curve parameter range, edge entity length, vertex coordinates, boundary loop path and patch topological reference relationship respectively;

[0098] The entity index identifier, entity type field, surface parameter expression, surface type field, surface parameter field, edge curve parameter expression, edge curve parameter range, edge entity length, vertex coordinates, boundary loop path and patch topological reference relationship are associated and stored to generate a geometric entity index;

[0099] Based on the geometric entity index, surface category, curvature value, normal vector, boundary type, hole location marker, hole location geometric region, and adjacent surface connection relationship are extracted to generate geometric attribute records, including:

[0100] Generate geometric attribute records, specifically:

[0101] Read the topological references of surface entities, edge entities, vertex entities, boundary loop entities, and patches from the geometric entity index;

[0102] Read the surface parameter representation, surface parameter domain, and surface type field from the surface entity;

[0103] Based on the surface type field, identify plane, cylindrical, conical, spherical, and freeform surfaces, and generate surface categories;

[0104] Generate a set of surface parameter evaluation points based on the surface parameter domain;

[0105] Substitute the set of evaluation points for surface parameters into the surface parameter expression to calculate the directional derivative of the first parameter, the directional derivative of the second parameter, the second derivative of the first parameter, the second derivative of the second parameter, and the derivative of the mixed parameter.

[0106] The principal curvature values ​​are calculated based on the first parameter directional derivative, the second parameter directional derivative, the first parameter second derivative, the second parameter second derivative, and the mixed parameter derivative. The absolute values ​​of the principal curvature values ​​are taken and averaged to generate the curvature values.

[0107] The cross product of the first parameter directional derivative and the second parameter directional derivative is used to obtain the initial normal vector of the surface.

[0108] Divide the initial normal vector of the surface by the magnitude of the initial normal vector of the surface to generate the normal vector;

[0109] Based on the boundary ring to which the edge entity belongs, the closure state of the boundary ring, the number of edge entities contained in the boundary ring, and the topological reference relationship of the facets, the outer contour boundary, the inner hole boundary, and the intersection line of adjacent faces are identified, and the boundary type is generated.

[0110] When the boundary type is an inner hole boundary and the associated surface type of the inner hole boundary is a cylindrical surface, a hole location marker is generated.

[0111] Based on the hole location mark, the inner hole boundary, the cylindrical surface entity, and the surface entity that shares the edge with the inner hole boundary, the hole location geometric region is generated. The hole location geometric region includes the inner hole boundary, the hole periphery surface entity, and the adjacent surface around the hole.

[0112] Based on the topological reference relationship of the facets and the shared edge entities, record the adjacent curved surface entities, the position of the shared edge, the connection direction and the normal change, and generate the connection relationship between adjacent faces;

[0113] The surface type, curvature value, normal vector, boundary type, hole position mark, hole position geometric region and adjacent surface connection relationship are associated and stored to generate geometric attribute records;

[0114] Based on the geometric entity index and geometric attribute records, a geometric primitive topology is established using surface entities, boundary loop entities, and hole geometric regions as nodes, and using patch edge-sharing relationships, boundary closure relationships, hole adjacency relationships, and geometric constraint relationships as connecting edges. Where:

[0115] The geometric primitive topology graph is constructed as follows:

[0116] Read the topological references of surface entities, boundary loop entities, edge entities, vertex entities, and facets from the geometric entity index;

[0117] Read the surface type, curvature value, normal vector, boundary type, hole mark, hole geometric region, and adjacent surface connection relationship from the geometric attribute record;

[0118] Write the surface entity into the surface node set, and write the surface category, curvature value and normal vector into the surface node attributes;

[0119] Write the boundary loop entity into the boundary node set, and write the boundary type, boundary loop path, and boundary loop closure state into the boundary node attributes;

[0120] Write the hole location geometry region into the hole location node set, and write the hole location mark, inner hole boundary, and adjacent surface around the hole into the hole location node attributes;

[0121] Establish common edge relationships for patches based on patch topological reference relationships and shared edge entities; establish boundary closure relationships based on boundary ring entities, boundary ring paths, and boundary ring closure states; establish hole adjacency relationships based on hole geometric regions, adjacent faces around holes, and inner hole boundaries; and establish geometric constraint relationships based on adjacent face connection relationships.

[0122] Write the common edge relationship of the facets, the boundary closure relationship, the hole adjacency relationship and the geometric constraint relationship into the connection edge set;

[0123] A geometric primitive topology graph is generated from the set of surface nodes, the set of boundary nodes, the set of hole nodes, and the set of connecting edges;

[0124] Based on the geometric primitive topology diagram, surfaces, boundary loops, and hole locations are sampled. The coordinates of the sampled points are calculated, and the surface category, curvature value, normal vector, boundary type, hole location marker, and topological adjacency marker are written for each sampled point. A derived point set with CAD attribute markers is generated, where:

[0125] The coordinates of the sampling points are calculated as follows:

[0126] Read surface nodes, boundary nodes, hole nodes, and connecting edges from the geometric primitive topology graph;

[0127] Read the curvature value, hole geometry region, and adjacent face connection relationship from the geometric attribute record;

[0128] Read the surface parameter representation and surface parameter domain from the surface entity associated with the surface node. The surface parameter domain includes the lower limit of the first parameter, the upper limit of the first parameter, the lower limit of the second parameter, and the upper limit of the second parameter.

[0129] The edge entity lengths contained in the boundary ring entity are read by connecting the boundary nodes through the surface nodes, and the average edge entity lengths are calculated to generate the average edge entity length.

[0130] The average length of the edge entity is used as the basic sampling interval of the surface. The average length of the edge entity is divided by a curvature value to obtain the curvature-adjusted sampling interval. The smaller of the two sampling intervals is used as the surface sampling interval.

[0131] The first parameter span and the second parameter span are calculated based on the surface parameter domain, and the first parameter sequence and the second parameter sequence are generated by combining the surface sampling interval;

[0132] The first parameter sequence and the second parameter sequence are combined to generate surface sampling parameter pairs, and then substituted into the surface parameter expression to calculate the coordinates of the surface sampling points;

[0133] Read the boundary loop path, edge curve parameter expression, edge curve parameter range, and edge entity length from the boundary loop entity associated with the boundary node;

[0134] The number of boundary samples is calculated based on the edge entity length and the average edge entity length, and the edge curve parameter interval is divided equally to generate a boundary sampling parameter sequence.

[0135] Substitute the boundary sampling parameter sequence into the edge curve parameter expression to calculate the coordinates of the boundary sampling points;

[0136] Read the inner hole boundary and the hole periphery surface entity from the hole geometric region associated with the hole node, average the boundary sampling points corresponding to the inner hole boundary, and generate the center coordinates of the hole.

[0137] Based on the surface parameter domain and surface sampling interval of the hole periphery surface entity, hole periphery sampling parameters are generated and substituted into the hole periphery surface entity parameter expression to calculate the coordinates of the hole position sampling point;

[0138] The coordinates of surface sampling points, boundary sampling points, hole center coordinates, and hole sampling points are merged to generate a set of sampling point coordinates.

[0139] Generate a set of derived points with CAD attribute tags, specifically as follows:

[0140] Read the surface nodes, boundary nodes, hole nodes, and connecting edges corresponding to the set of sampling point coordinates from the geometric primitive topology graph;

[0141] Read the surface type, curvature value, normal vector, boundary type, hole position mark, and adjacent surface connection relationship from the geometric attribute record;

[0142] Write the surface category, curvature value, and normal vector according to the surface node to which the sampling point belongs;

[0143] Write the boundary type based on the boundary node to which the sampling point belongs, and write the hole position mark based on the hole position node to which the sampling point belongs;

[0144] Based on the surface nodes, boundary nodes, hole nodes, and connecting edges to which the sampling point belongs, record the identifiers of the surface nodes, boundary nodes, hole nodes, connecting edges, adjacent surface nodes, adjacent boundary nodes, and adjacent hole nodes to generate topological adjacency markers;

[0145] The sampling point coordinates, surface category, curvature value, normal vector, boundary type, hole position mark, and topological adjacency mark are associated and stored to generate a derived point set with CAD attribute marks.

[0146] In this embodiment, generating the set of key geometric candidate points includes:

[0147] Read the coordinates, normal vectors, curvature values, surface categories, boundary types, hole location markers, and topological adjacency markers of the derived point set with CAD attribute tags to generate key point detection input data, including:

[0148] The keypoint detection input data is generated as follows:

[0149] Read the sampling point coordinates, normal vectors, curvature values, surface categories, boundary types, hole location markers, and topological adjacency markers from the derived point set with CAD attribute tags;

[0150] The sampling point coordinates, normal vectors, and curvature values ​​are combined into geometric detection features; the surface category, boundary type, and hole position markers are combined into CAD attribute features; and the topological adjacency markers are used as topological constraint features.

[0151] Based on the sampling point arrangement order in the derived point set with CAD attribute tags, the geometric detection features, CAD attribute features, and topological constraint features are associated and stored to generate key point detection input data;

[0152] Based on the keypoint detection input data, surface detection neighborhoods, boundary detection neighborhoods, and hole detection neighborhoods are established respectively. Covariance statistics are performed on the normal distribution of sampling points within each detection neighborhood to generate multi-scale normal covariance records, where:

[0153] Generate multi-scale normal covariance records, specifically as follows:

[0154] Read the sampling point coordinates, normal vector, curvature value, surface category, boundary type, hole position marker, and topological adjacency marker from the key point detection input data;

[0155] Based on the same surface category, the same surface node identifier in the topological adjacency marker, and the connection of adjacent surface node identifiers, sampling points in the neighborhood of the surface are selected to establish the surface detection neighborhood;

[0156] Boundary neighborhood sampling points are selected based on the same boundary type, the same boundary node identifier in the topological adjacency tag, and the connection of adjacent boundary node identifiers to establish a boundary detection neighborhood.

[0157] Based on the same hole location marker, the same hole location node identifier in the topological adjacency marker, and the connection of adjacent hole location node identifiers, sampling points in the neighborhood of the hole location are selected to establish the hole location detection neighborhood;

[0158] Read the normal vectors of the sampling points in the surface detection neighborhood, boundary detection neighborhood and hole detection neighborhood respectively. Average all normal vectors in the same detection neighborhood to obtain the mean normal vector. Subtract the mean normal vector from each normal vector in the same detection neighborhood to obtain the normal deviation vector.

[0159] Multiply each normal deviation vector by its corresponding transpose vector to obtain the single-point normal deviation matrix;

[0160] The average of the normal deviation matrices of all single points within the same detection neighborhood is used to obtain the normal covariance matrix of the detection neighborhood.

[0161] The normal covariance matrices corresponding to the surface detection neighborhood, boundary detection neighborhood, and hole detection neighborhood are associated and stored to generate a multi-scale normal covariance record.

[0162] The underlying Harris response is calculated based on the multi-scale normal covariance record, and the responses at different scales are fused to generate a multi-scale underlying response record, wherein:

[0163] Generate multi-scale baseline response records, specifically as follows:

[0164] Read the normal covariance matrices corresponding to the surface detection neighborhood, boundary detection neighborhood, and hole detection neighborhood from the multi-scale normal covariance record;

[0165] Calculate the determinant of each normal covariance matrix to obtain the volume of change in the neighborhood normal, calculate the trace of each normal covariance matrix, and square the trace to obtain the volume of change in the principal direction of the neighborhood normal.

[0166] Divide the trace of the normal covariance matrix by the trace of the added normal covariance matrix to obtain the response adjustment coefficient. Multiply the response adjustment coefficient by the change in the principal normal direction of the neighborhood to obtain the response suppression term.

[0167] Subtracting the response suppression term from the neighborhood normal variation volume yields the Harris response of the surface foundation, the Harris response of the boundary foundation, and the Harris response of the hole foundation.

[0168] Response normalization is performed on the Harris responses of the surface foundation, boundary foundation, and hole foundation respectively to generate the surface normalized response, boundary normalized response, and hole normalized response.

[0169] Read the number of sampling points in the surface detection neighborhood, boundary detection neighborhood, and hole detection neighborhood, and divide the number of sampling points in each detection neighborhood by the total number of sampling points in the three types of detection neighborhoods to obtain the scale fusion weight;

[0170] The surface normalized response, boundary normalized response, and aperture normalized response are multiplied by their respective scale fusion weights and then summed to generate the fused basic Harris response.

[0171] The Harris responses of the surface foundation, boundary foundation, aperture foundation, and fusion foundation are associated and stored to generate a multi-scale foundation response record.

[0172] CAD topological weights are generated based on surface type, boundary type, hole location markers, and topological adjacency markers. These CAD topological weights are then written into the multi-scale basic response record to generate a topologically weighted Harris response record, where:

[0173] Generate a topology-weighted Harris response record, specifically as follows:

[0174] Read the surface category, boundary type, hole position mark, and topological adjacency mark corresponding to the sampling point from the key point detection input data;

[0175] Read the surface-based Harris response, boundary-based Harris response, aperture-based Harris response, and fusion-based Harris response from the multi-scale fundamental response record;

[0176] For each sampling point, count the number of sampling points with the same surface category, the same boundary type, the same hole position mark, and connected by topological adjacency mark in the corresponding detection neighborhood;

[0177] Divide each statistical quantity by the total number of sampling points in the detection neighborhood to obtain the surface category consistency weight, boundary type consistency weight, hole position label consistency weight, and topological adjacency weight.

[0178] The average of the four weights is used to generate the CAD topology weights.

[0179] The CAD topology weights are multiplied by the Harris responses of the surface foundation, boundary foundation, hole foundation, and fusion foundation, respectively, to generate a topology-weighted Harris response record.

[0180] Based on the topologically weighted Harris response records, sampling points corresponding to the locations of parametric evidence are selected. Candidate merging is then performed on sampling points within the same boundary, the same aperture location, and the same connection region to generate a set of key geometric candidate points, where:

[0181] Generate a set of key geometric candidate points, specifically as follows:

[0182] Read the coordinates of each sampling point, the topological weighted Harris response value, the surface node to which it belongs, the boundary node to which it belongs, the hole node to which it belongs, the connecting edge, the boundary type, the hole marker, and the topological adjacency marker for each sampling point from the topological weighted Harris response record;

[0183] Sort the topology-weighted Harris response values ​​in descending order to generate a sorted response record;

[0184] Based on the coordinates of the sampling points and the topological adjacency markers, adjacent sampling points are determined. Neighborhood extreme value judgment is performed on the sampling points in the response sorting record. Sampling points whose response values ​​are greater than the response values ​​of their adjacent sampling points are retained, and an initial candidate point set is generated.

[0185] Read the boundary nodes, hole nodes, and connecting edges of the sampling points in the initial candidate point set;

[0186] Sampling points that are associated with the same hole location node and the same hole location marker are divided into the same hole location candidate group; sampling points that are associated with the same boundary node and have the same boundary type are divided into the same boundary candidate group; and sampling points that are associated with the same connecting edge and whose topological adjacency markers point to the same adjacent surface node are divided into the same connecting region candidate group.

[0187] Within the same boundary candidate group, the same aperture candidate group, and the same connection region candidate group, merge the sampling points that are connected by topological adjacency markers and located on the continuous path of the same candidate group, retain the sampling point with the highest topological weighted Harris response value, and generate a set of key geometric candidate points.

[0188] In this embodiment, the construction of the benchmarked local geometry block includes:

[0189] Read the key geometric candidate point set and geometric primitive topology graph, and perform back mapping based on the sampling point coordinates, surface type, curvature value, normal vector, boundary type, hole position marker, topological adjacency marker, and connecting edge to generate candidate point location records, where:

[0190] Generate candidate point location records, specifically as follows:

[0191] Read the sampling point coordinates, surface category, curvature value, normal vector, boundary type, hole location marker, and topological adjacency marker of the candidate sampling points from the set of key geometric candidate points;

[0192] Read the set of surface nodes, the set of boundary nodes, the set of hole nodes, and the set of connecting edges from the geometric primitive topology graph;

[0193] Based on the identifier of the surface node in the topological adjacency tag, the candidate sampling points are backmapped to the corresponding surface node;

[0194] Based on the boundary node identifier in the topological adjacency tag, the candidate sampling points are backmapped to the corresponding boundary node;

[0195] Based on the identifier of the hole node in the topological adjacency tag, the candidate sampling points are backmapped to the corresponding hole node;

[0196] Based on the connection edge identifier in the topological adjacency tag, the candidate sampling points are backmapped to the corresponding connection edges;

[0197] The candidate sampling point coordinates, surface nodes, boundary nodes, hole nodes, connecting edges, surface type, curvature value, normal vector, boundary type, topological adjacency marker, and hole marker are associated and stored to generate a candidate point positioning record;

[0198] The candidate point location records are indexed with surface nodes, boundary nodes, hole location nodes, and connecting edges to generate a geometric evidence index, where:

[0199] Generate a geometric evidence index, specifically as follows:

[0200] Read the coordinates of candidate sampling points, surface nodes, boundary nodes, hole nodes, connecting edges, surface categories, boundary types, hole markers, and topological adjacency markers from the candidate point location records;

[0201] Read the set of surface nodes, the set of boundary nodes, the set of hole nodes, and the set of connecting edges from the geometric primitive topology graph;

[0202] Based on the coordinates of the candidate sampling points and the identifiers of the surface nodes to which they belong, a surface evidence index is established between the candidate sampling points and the surface nodes.

[0203] Based on the boundary node identifier and boundary type, establish a boundary evidence index between candidate sampling points and boundary nodes;

[0204] Based on the corresponding well location node identifier and well location marker, establish a well location evidence index between candidate sampling points and well location nodes;

[0205] Based on the edge identifiers, a topological connection evidence index is established between candidate sampling points and edge connections;

[0206] The surface evidence index, boundary evidence index, hole location evidence index, and topological connection evidence index are associated and stored to generate a geometric evidence index;

[0207] Based on the geometric evidence index, derived points corresponding to homologous surfaces, closed boundaries, aperture neighborhoods, and topologically adjacent regions are extracted to generate evidence-constrained local geometric blocks, where:

[0208] Generate evidence-constrained local geometric blocks, specifically:

[0209] Read the surface evidence index, boundary evidence index, aperture evidence index, and topological connection evidence index from the geometric evidence index;

[0210] Read the sampling point coordinates, surface type, curvature value, normal vector, boundary type, hole position mark and topological adjacency mark from the derived point set with CAD attribute marks;

[0211] Based on the surface node identifiers in the surface evidence index, extract the derived points with the same surface node identifiers to generate a set of derived points of the same source surface;

[0212] Based on the boundary node identifier and boundary type in the boundary evidence index, extract the derived points that have the same boundary node identifier, are located on the boundary loop path, and are in the boundary loop closed state, and generate a set of closed boundary derived points;

[0213] Based on the hole location node identifier and hole location mark in the hole location evidence index, extract the derived points that have the same or adjacent hole location node identifiers, and generate a set of hole location neighborhood derived points.

[0214] Based on the connection edge identifiers in the topological connectivity evidence index, extract the derived points connected to adjacent surface node identifiers, adjacent boundary node identifiers, or adjacent hole node identifiers to generate a set of derived points in the topologically adjacent region.

[0215] The set of derived points from the same source surface, the set of derived points from the closed boundary, the set of derived points from the pore location neighborhood, and the set of derived points from the topologically adjacent region are merged, and duplicate sampling points are removed to generate an evidence-constrained local geometric block.

[0216] Based on the hole location neighborhood, closed boundary, adjacent surface normal, and connecting edge direction in the evidence-constrained local geometry, local reference coordinate records are calculated, and coordinate benchmarking processing is performed on the evidence-constrained local geometry to generate a benchmarked local geometry, wherein:

[0217] Generate a benchmarked local geometry block, specifically as follows:

[0218] Read the set of derived points from the same origin surface, the set of derived points from the closed boundary, the set of derived points from the pore location neighborhood, and the set of derived points from the topologically adjacent region from the evidence-constrained local geometry block;

[0219] Read the sampling point coordinates, normal vector, boundary type, hole position marker and topological adjacency marker corresponding to each derived point;

[0220] Average the coordinates of the sampled points in the set of derived points in the neighborhood of the hole location to generate a local reference origin;

[0221] Based on the boundary node identifier and boundary ring path order of the topological adjacency markers in the set of derived points of the closed boundary, the derived points of the closed boundary are sorted.

[0222] The average of the coordinate differences between adjacent closed boundary derived points after sorting is calculated and then normalized to generate the first local coordinate axis.

[0223] Based on the edge identifiers in the set of derived points of topologically adjacent regions, the derived points of topologically adjacent regions are divided into a set of derived points at the first end of the edge and a set of derived points at the second end of the edge.

[0224] The average coordinates of the sampled points in the two sets of end derived points are calculated to obtain the coordinates of the center of the first end and the center of the second end.

[0225] Subtract the center coordinates of the first end from the center coordinates of the second end and perform normalization to generate the direction of the connecting edge;

[0226] The average normal vectors of the derived point set of the same origin surface and the derived point set of the topologically adjacent region are averaged and normalized to generate the average normal vector of the adjacent surface.

[0227] The second local coordinate axis is generated by cross-product of the first local coordinate axis and the average normal of the adjacent surface and normalized.

[0228] The third local coordinate axis is generated by cross-product of the first and second local coordinate axes and normalized.

[0229] The local reference origin, three local coordinate axes, and the direction of the connecting edges are associated and stored to generate a local reference coordinate record;

[0230] Based on the local reference coordinate record, the coordinates of the sampling points in the evidence-constrained local geometry are transformed to the local reference coordinate system, and the surface type, curvature value, normal vector, boundary type, hole position mark and topological adjacency mark are preserved to generate the referenced local geometry.

[0231] In this embodiment, the output geometric parameter evidence set includes:

[0232] An improved PointMLP model is constructed. The improved PointMLP model retains the residual MLP encoding unit of the original PointMLP model and sets up a CAD attribute enhanced input layer, a geometry-topology dual neighborhood aggregation layer and a parameter evidence feedback correction layer.

[0233] The normalized coordinates, surface type, curvature value, normal vector, boundary type, hole location marker, topological adjacency marker, and geometric evidence index of the benchmarked local geometry are read. Attribute embedding and channel splicing are then performed through the CAD attribute enhancement input layer to generate point-level CAD attribute features, where:

[0234] The CAD attribute enhancement input layer specifically includes:

[0235] Geometry block reading end: Reads normalized coordinates, surface type, curvature value, normal vector, boundary type, hole position mark, topological adjacency mark and geometric evidence index from the benchmark local geometry block, and divides the input channels according to point coordinates, surface attributes, boundary hole position attributes and topological evidence source;

[0236] Attribute Slot Embedding Area: Write the surface category into the surface attribute slot, the boundary type into the boundary attribute slot, the hole position mark into the hole position function slot, and the topology adjacency mark into the topology association slot to generate CAD attribute embedding.

[0237] Evidence weighted splicing area: Based on the geometric evidence index, evidence guiding weights are generated, and CAD attributes are embedded, normalized coordinates, curvature values ​​and normal vectors are weighted and spliced ​​to generate point-level CAD attribute features;

[0238] Generate point-level CAD attribute features, specifically:

[0239] By enhancing the geometry block reading end in the input layer with CAD attributes, the normalized coordinates, surface category, curvature value, normal vector, boundary type, hole position mark, topological adjacency mark, and geometric evidence index corresponding to each sampling point are read from the benchmark local geometry block.

[0240] Normalized coordinates are input into the coordinate mapping channel to generate coordinate codes. The curvature value is divided by one plus the absolute value of the curvature value to obtain the curvature modulation value. The normal vector is divided by the normal vector magnitude to obtain the unit normal code. The curvature modulation value and the unit normal code are concatenated to generate continuous geometric features.

[0241] By using the attribute slot embedding area, the surface category is written into the surface attribute slot, the boundary type is written into the boundary attribute slot, the hole position mark is written into the hole position function slot, and the topological adjacency mark is written into the topological association slot, and these are respectively mapped to surface attribute embedding, boundary attribute embedding, hole position attribute embedding and topological association embedding.

[0242] Through the evidence weighting splicing area, the surface evidence index, boundary evidence index, hole location evidence index, and topological connection evidence index are read from the geometric evidence index;

[0243] Count the number of evidence corresponding to each type of evidence index, divide the number of evidence in a single type by the total number of evidence, and generate the corresponding evidence weight.

[0244] After multiplying the coordinate encoding, continuous geometric features, surface attribute embedding, boundary attribute embedding, hole location attribute embedding and topological association embedding by their respective evidence weights, channel splicing is performed to generate point-level CAD attribute features.

[0245] Geometric and topological neighborhood features are established based on point-level CAD attribute features, where:

[0246] Geometric and topological neighborhood features are established based on point-level CAD attribute features, specifically as follows:

[0247] Read the coordinate code, continuous geometric features, surface attribute embedding, boundary attribute embedding, hole location attribute embedding, and topological association embedding corresponding to each sampling point from the point-level CAD attribute features;

[0248] The chain path is established through the geometric neighborhood in the geometric-topological dual-neighborhood aggregation layer. The Euclidean distance between sampling points is calculated based on the coordinate encoding. The curvature difference and normal angle are calculated based on continuous geometric features.

[0249] Establish geometric adjacency relationships for sampling points with an Euclidean distance of no more than 0.08, a curvature difference of no more than 0.12, and a normal angle of no more than 15°;

[0250] When the number of sampling points that satisfy the geometric adjacency relationship exceeds 32, the first 32 sampling points are retained in ascending order of Euclidean distance to generate geometric neighborhood features;

[0251] By establishing a chain path in the topological neighborhood of the geometric-topological dual-neighborhood aggregation layer, the corresponding surface node identifier, boundary node identifier, hole node identifier, and connecting edge identifier of the topological association embedding are read.

[0252] Establish topological adjacency relationships between sampling points that have the same surface node identifier, the same boundary node identifier, the same hole node identifier, or are connected by a single hop in the connecting edge identifier;

[0253] When the number of sampling points that satisfy the topological adjacency relationship exceeds 32, the first 32 sampling points are retained in ascending order of the number of hops of the connecting edges to generate topological neighborhood features;

[0254] Geometric neighborhood features and topological neighborhood features are associated and stored as input features for the geometric-topological dual neighborhood aggregation layer;

[0255] The geometric-topological dual-neighborhood aggregation layer performs bi-branch residual encoding, neighborhood weight allocation, and cross-neighborhood fusion on the geometric and topological neighborhood features to generate joint geometric-topological features, where:

[0256] The geometric-topological dual-neighborhood aggregation layer specifically includes:

[0257] Geometric neighborhood chaining path: establish spatial neighborhood, curvature neighborhood, and normal variation neighborhood based on normalized coordinates, curvature values, and normal vectors;

[0258] Topological neighborhood chain building path: Based on the surface node identifier, boundary node identifier, hole position node identifier, and connecting edge identifier in the topological adjacency tag, establish surface common-source neighborhood, boundary same-loop neighborhood, hole same-domain neighborhood, and connecting edge neighborhood;

[0259] Cross-neighborhood fusion zone: Calculate the node correspondence and connection edge correspondence between geometric neighborhood and topological neighborhood, generate neighborhood consistency weight, and output geometric-topological joint features based on neighborhood consistency weight;

[0260] The geometric-topological dual-neighborhood aggregation layer performs bi-branch residual encoding, neighborhood weight allocation, and cross-neighborhood fusion on geometric and topological neighborhood features, specifically as follows:

[0261] Read the geometric neighbor sampling points, Euclidean distance, curvature difference, normal angle and geometric adjacency relationship corresponding to each center sampling point from the geometric neighbor features;

[0262] Read the topological neighborhood sampling points, the corresponding surface node identifiers, the corresponding boundary node identifiers, the corresponding hole node identifiers, the connecting edge identifiers, and the topological adjacency relationships corresponding to each central sampling point from the topological neighborhood features;

[0263] By establishing a chain path in the geometric-topological dual-neighborhood aggregation layer, the coordinate codes and continuous geometric features of the central sampling point and the geometric neighbor sampling points are read.

[0264] The coordinate code of the center sampling point is generated by subtracting the coordinate code of the geometric neighborhood sampling point from the coordinate code of the center sampling point.

[0265] The relative coordinate encoding, Euclidean distance, curvature difference, and normal angle are concatenated to generate a geometric relationship vector.

[0266] The geometric relationship vector is concatenated with the continuous geometric features of the geometric neighborhood sampling points and then input into the residual mapping to generate single-point geometric residual features.

[0267] Average pooling is performed on all single-point geometric residual features of the same central sampling point to generate geometric branch coding features;

[0268] By establishing a chain path through topological neighborhood, the surface node identifier, boundary node identifier, and hole location node identifier of the central sampling point and the topological neighborhood sampling point are compared to generate surface homogeneity markers, boundary homo-loop markers, and hole location homo-domain markers.

[0269] Calculate the number of jumps on the connecting edge based on the connecting edge identifier, and generate the distance marker for the connecting edge.

[0270] By concatenating the surface homogeneity markers, boundary homo-loop markers, hole location homo-domain markers, connecting edge distance markers, and topological adjacency relationships, and inputting them into the topological residual mapping, a single-point topological residual feature is generated.

[0271] Average pooling is performed on all single-point topological residual features of the same central sampling point to generate topological branch coding features;

[0272] Calculate the cosine similarity between the geometric branch coding features and the topological branch coding features, add one to the cosine similarity and divide by two to generate the initial weight for neighborhood consistency.

[0273] The neighborhood consistency weight is generated by dividing the initial neighborhood consistency weight by the sum of all the initial neighborhood consistency weights of the same central sampling point.

[0274] By multiplying the geometric branch encoding features by the neighborhood consistency weight through the cross-neighborhood fusion region, geometric weighted features are generated.

[0275] Multiply the topological branch encoding features by one minus the neighborhood consistency weight to generate topological weighted features;

[0276] The geometrically weighted features and the topologically weighted features are concatenated through channels, and the features of the central sampling point are written back through residual connections to generate geometrically-topological joint features.

[0277] The residual MLP encoding unit performs layer-by-layer nonlinear mapping, residual connection, and local feature aggregation on the joint geometric-topological features to generate local geometric semantic features, where:

[0278] Layer-by-layer nonlinear mapping, specifically:

[0279] Read the joint feature vector, neighborhood sample point joint feature vector, and geometric evidence index corresponding to each central sampling point from the geometric-topological joint features;

[0280] The joint feature vector of the center sampling points is input into the first feature mapping layer through the residual MLP encoding unit;

[0281] Multiply the joint feature vector of the central sampling point with the first mapping weight matrix and add the first bias vector to generate the first layer of mapping features;

[0282] Perform nonlinear activation on the first layer of mapped features to generate the first layer of activated features;

[0283] The activation features of the first layer are input into the second feature mapping layer;

[0284] The first layer activation features are multiplied by the second mapping weight matrix, and the second bias vector is added to generate the second layer mapping features.

[0285] Residual join, specifically:

[0286] Read the joint feature vector of the second-layer mapping features and the center sampling point;

[0287] When the number of channels of the joint feature vector of the second-layer mapping feature and the center sampling point is the same, the joint feature vector of the second-layer mapping feature and the center sampling point is added channel by channel to generate a single-point residual coding feature;

[0288] When the number of channels of the joint feature vector of the second layer mapping feature and the center sampling point is inconsistent, the joint feature vector of the center sampling point is input into the channel alignment mapping layer to generate the aligned joint feature vector;

[0289] The second-layer mapping features are added channel by channel to the aligned joint feature vector to generate single-point residual coding features;

[0290] Local feature aggregation, specifically:

[0291] Read the joint feature vector, single-point residual coding features, and geometric evidence index of the neighboring sampling points corresponding to the same central sampling point;

[0292] Perform average pooling on the joint feature vector of the neighborhood sampling points to generate the neighborhood average feature;

[0293] Max pooling is performed on the joint feature vector of the neighborhood sampling points to generate salient neighborhood features;

[0294] The single-point residual coding features, neighborhood average features, and neighborhood salient features are concatenated through channels to generate local aggregated features;

[0295] Read the surface evidence index, boundary evidence index, hole location evidence index, and topological connection evidence index corresponding to the center sampling point from the geometric evidence index;

[0296] Count the number of index entries in the surface evidence index, boundary evidence index, hole location evidence index, and topological connection evidence index respectively;

[0297] The surface evidence count is obtained by incrementing the number of surface index entries by one; the boundary evidence count is obtained by incrementing the number of boundary index entries by one; the hole location evidence count is obtained by incrementing the number of hole location index entries by one; and the topology connection evidence count is obtained by incrementing the number of topology connection index entries by one.

[0298] Add the surface evidence count, boundary evidence count, hole location evidence count, and topological connection evidence count to generate the total evidence count;

[0299] Divide the surface evidence count, boundary evidence count, hole location evidence count, and topological connection evidence count by the total evidence count to generate the surface evidence weight, boundary evidence weight, hole location evidence weight, and topological connection evidence weight, respectively.

[0300] The weights of surface evidence, boundary evidence, hole location evidence, and topological connection evidence are associated and stored to generate an evidence channel weight record corresponding to the central sampling point.

[0301] Local aggregation features are divided into surface semantic channels, boundary semantic channels, aperture semantic channels, and topological connection semantic channels according to surface evidence, boundary evidence, aperture evidence, and topological connection evidence.

[0302] Multiply the surface semantic channel by the surface evidence weight, multiply the boundary semantic channel by the boundary evidence weight, multiply the hole position semantic channel by the hole position evidence weight, and multiply the topological connection semantic channel by the topological connection evidence weight.

[0303] The weighted surface semantic channel, boundary semantic channel, aperture semantic channel and topological connection semantic channel are concatenated to generate local geometric semantic features.

[0304] The parameter evidence feedback correction layer performs rule-based association correction on local geometric semantic features, generating geometric primitive evidence, functional baseline evidence, parameter candidate evidence, and topological relationship evidence, and summarizing them to generate a geometric parameter evidence set, in which:

[0305] The parameter evidence feedback correction layer specifically includes:

[0306] Evidence distribution output head: Reads local geometric semantic features and writes them into the output paths of geometric primitive evidence, functional baseline evidence, parameter candidate evidence and topological relationship evidence respectively;

[0307] Confidence recalibration gate: Recalibrate the confidence of each evidence output path based on the surface evidence index, boundary evidence index, hole location evidence index, and topological connection evidence index;

[0308] Correlation Correction Output Area: Based on the correspondence between geometric primitive evidence, functional benchmark evidence, parameter candidate evidence, and topological relation evidence, correlation correction is performed to generate a set of geometric parameter evidence;

[0309] The parameter evidence feedback correction layer performs rule-based association correction on local geometric semantic features, specifically as follows:

[0310] The local geometric semantic features, geometric evidence index, topological adjacency markers, and evidence channel weight records are read through the parameter evidence feedback correction layer.

[0311] By using the evidence distribution output head, local geometric semantic features are written into the geometric primitive evidence output path, functional benchmark evidence output path, parameter candidate evidence output path, and topological relationship evidence output path, respectively.

[0312] In the geometric primitive evidence output path, surface primitives, boundary primitives, aperture primitives and connecting transition primitives are identified based on the surface semantic channel and the boundary semantic channel to generate geometric primitive evidence;

[0313] In the functional benchmark evidence output path, the hole center, mounting end face, functional axis and bearing section are extracted based on the hole position semantic channel, topological connection semantic channel and normal direction features to generate functional benchmark evidence;

[0314] In the parameter candidate evidence output path, the radius, hole spacing, end face spacing, axial direction and transition connection dimensions are calculated based on the surface primitive, boundary primitive, hole position primitive and functional benchmark evidence to generate parameter candidate evidence;

[0315] In the topological relationship evidence output path, topological relationship evidence is generated based on the surface node identifier, boundary node identifier, hole node identifier, and connecting edge identifier in the topological adjacency marker.

[0316] The confidence recalibration gate is used to read the weights of surface evidence, boundary evidence, hole location evidence, and topological connection evidence.

[0317] Multiply the confidence level of geometric primitive evidence by the weights of surface evidence and boundary evidence to generate recalibrated geometric primitive evidence.

[0318] The confidence level of the functional baseline evidence is multiplied by the weights of the aperture evidence and the topological connectivity evidence to generate recalibrated functional baseline evidence.

[0319] Multiply the confidence scores of the parameter candidate evidence by the surface evidence weight, boundary evidence weight, and hole location evidence weight to generate recalibrated parameter candidate evidence.

[0320] Multiply the confidence score of the topological relationship evidence by the weight of the topological connection evidence to generate recalibrated topological relationship evidence;

[0321] By associating and correcting the output area, association correction is performed based on the node correspondence, boundary correspondence, hole position correspondence, and connecting edge correspondence between geometric primitive evidence, functional benchmark evidence, parameter candidate evidence, and topological relationship evidence.

[0322] The revised geometric primitive evidence, functional benchmark evidence, parameter candidate evidence, and topological relationship evidence are associated and stored to generate a geometric parameter evidence set.

[0323] The improved PointMLP model was trained using geometric primitive evidence classification error, functional baseline evidence localization error, parametric candidate evidence regression error, topological relationship evidence discrimination error, and rule feedback consistency error as joint optimization objectives. The parameters of the CAD attribute enhancement input layer, the geometric-topological dual-neighborhood aggregation layer, the residual MLP encoding unit, and the parametric evidence feedback correction layer were optimized. The improved PointMLP model was considered to have completed convergence training when the change in the joint loss value over five consecutive training rounds was less than 0.001.

[0324] The improved PointMLP model is trained as follows:

[0325] The training sample set extracts benchmarked local geometric blocks, true geometric primitive labels, true functional benchmark locations, true parameter candidate values, and true topological relation labels. The training sample set includes 10,000 sets of engine part CAD training samples. Each set of engine part CAD training samples contains 256 to 2048 sampling points, 4 to 12 surface nodes, 2 to 10 boundary nodes, 1 to 8 hole nodes, and 3 to 20 connecting edges. The benchmarked local geometric blocks are input into the improved PointMLP model. Point-level CAD attribute features are output by the geometric block reading end, attribute slot embedding area, and evidence weighting splicing area in the CAD attribute enhancement input layer. Geometric-topological joint features are output by the geometric neighborhood chaining path, topological neighborhood chaining path, and cross-neighborhood fusion area in the geometric-topological dual neighborhood aggregation layer. Local geometric semantic features are output by the residual MLP encoding unit. Predicted geometric primitive evidence, predicted functional benchmark evidence, predicted parameter candidate evidence, and predicted topological relation evidence are output by the evidence assignment output head, confidence recalibration gate, and association correction output area in the parameter evidence feedback correction layer.

[0326] The classification error of geometric primitive evidence is obtained by calculating the category difference between the surface primitive category, boundary primitive category, hole location primitive category, and connection transition primitive category in the predicted geometric primitive evidence and the true geometric primitive label. The average of the squared category difference is then taken. The positioning error of functional benchmark evidence is obtained by calculating the coordinate difference between the hole location center, installation end face, functional axis, and bearing section in the predicted functional benchmark evidence and the true position of the functional benchmark. The average of the squared coordinate difference is then taken. The regression error of parameter candidate evidence is obtained by calculating the difference between the radius, hole spacing, end face spacing, axis direction, and transition connection dimension in the predicted parameter candidate evidence and the true value of the parameter candidate. The discrimination error of topological relationship evidence is obtained by calculating the difference between the surface common edge relationship, boundary closure relationship, hole location adjacency relationship, and connection edge relationship in the predicted topological relationship evidence and the true topological relationship label.

[0327] The consistency error is fed back based on the node correspondence, boundary correspondence, hole location correspondence, and connecting edge correspondence among the predicted geometric primitive evidence, predicted functional baseline evidence, predicted parameter candidate evidence, and predicted topological relationship evidence.

[0328] The classification error of geometric primitive evidence is multiplied by 0.25, the localization error of functional baseline evidence is multiplied by 0.20, the regression error of parameter candidate evidence is multiplied by 0.25, the discrimination error of topological relationship evidence is multiplied by 0.15, and the consistency error of rule feedback is multiplied by 0.15. These are then summed to generate a joint loss value. The rate of change of the joint loss value relative to the trainable parameters in the CAD attribute enhancement input layer, the geometric-topological dual neighborhood aggregation layer, the residual MLP encoding unit, and the parameter evidence feedback correction layer is calculated to obtain the gradient value. The gradient value is multiplied by the learning rate of 0.001 and then subtracted from the current value of the corresponding trainable parameter to obtain the updated model parameters. The updated model parameters are then written into the corresponding structure. The process of sample input, evidence prediction, error calculation, joint loss value calculation, and parameter update is repeated until the change in the joint loss value in 5 consecutive training rounds is less than 0.001. This indicates that the improved PointMLP model has completed convergence training.

[0329] In this embodiment, the improved PointMLP model includes:

[0330] While retaining the local point set grouping, residual MLP encoding, point-by-point feature mapping, and multi-head output backbone of the PointMLP model, structural modifications are made to the point-level input, neighborhood organization, feature propagation, and evidence output around the task of extracting and reconstructing CAD geometric information of engine parts.

[0331] A CAD attribute enhancement input layer is added to the model input end, and the normalized coordinates, surface categories, curvature values, normal vectors, boundary types, hole position markers, topological adjacency markers and geometric evidence indexes in the benchmark local geometry are embedded into the evidence channel to generate point-level CAD attribute features.

[0332] A geometric-topological dual-neighborhood aggregation layer is added at the local feature aggregation location to separately encode the spatial distance neighborhood, curvature change neighborhood, common edge topology neighborhood, and aperture location association neighborhood, and perform cross-neighborhood feature fusion through neighborhood consistency weight;

[0333] The residual MLP coding unit is improved by transforming the single path point feature mapping into a residual interactive coding structure with geometric branches, topological branches and evidence fusion branches to generate local geometric semantic features.

[0334] A parameter evidence feedback correction layer is added to the output end to perform rule association correction and evidence confidence recalibration on local geometric semantic features, and output geometric primitive evidence, functional benchmark evidence, parameter candidate evidence and topological relationship evidence.

[0335] In this embodiment, generating the candidate parameter locking result includes:

[0336] Read the geometric parameter evidence set and geometric primitive topology diagram. Based on the geometric primitive evidence, functional reference evidence, parameter candidate evidence, and topological relationship evidence, extract the functional axis, mounting end face, hole center, load-bearing section, and connection transition relationship to generate the functional reference skeleton of the engine part, wherein:

[0337] Generate the functional baseline skeleton for engine parts, specifically as follows:

[0338] Extract geometric primitive evidence, functional benchmark evidence, parameter candidate evidence, and topological relation evidence from the geometric parameter evidence set;

[0339] Read the set of surface nodes, the set of boundary nodes, the set of hole nodes, and the set of connecting edges from the geometric primitive topology graph;

[0340] Based on the cylindrical surface evidence, the hole position evidence, and the hole position center evidence in the geometric element evidence, the coordinates of the hole position center point and the hole diameter parameters are read.

[0341] Calculate the direction vector between any two hole center points, and classify the hole center points whose direction vector angle is no greater than 5° into the coaxial candidate group;

[0342] Perform linear fitting on the center point of the hole position within the coaxial candidate group and calculate the distance from the center point of the hole position to the fitted line;

[0343] When the distance from the center point of the hole to the fitted straight line is no greater than 0.05 times the hole diameter parameter, the coaxial candidate group is determined as the coaxial hole group.

[0344] Average the coordinates of the center points of the holes within the coaxial hole group to generate the axis reference point;

[0345] Perform principal direction calculation on the coordinates of the center points of the holes within the coaxial hole group to generate the axis direction vector;

[0346] Functional axes are generated from axis reference points and axis direction vectors;

[0347] Based on the end face datum mark in the functional datum evidence, the planar element evidence in the geometric element evidence, and the surface node normal vector in the geometric element topology diagram, the installation end face curved surface node is selected;

[0348] Read the coordinates of the sampling points associated with the nodes of the mounting end face, average the coordinates of the sampling points, and generate the center coordinates of the mounting end face;

[0349] Use the normal vector of the surface node of the mounting end face as the normal vector of the mounting end face, and use the boundary node associated with the surface node of the mounting end face as the boundary node of the mounting end face.

[0350] The mounting end face is generated from the center coordinates of the mounting end face, the normal of the mounting end face, and the boundary nodes of the mounting end face.

[0351] Based on the borehole center evidence in the functional benchmark evidence, read the coordinates of the borehole center point, and associate and store the coordinates of the borehole center point with the borehole node identifier to generate the borehole center;

[0352] Based on the cross-sectional dimensions in the candidate parameter evidence, the load-bearing surface element evidence in the geometric element evidence, and the connection edge relationship in the topological relationship evidence, the surface nodes and boundary nodes connected to the functional axis or the mounting end face are selected.

[0353] The selected surface nodes, boundary nodes, and cross-sectional dimensions are associated and stored to generate a load-bearing cross-section;

[0354] Based on the surface-shared edge relationship, boundary closure relationship, hole adjacency relationship and connecting edge relationship in the topological relationship evidence, and combined with the connecting edge set in the geometric primitive topology diagram, the connection sequence, connection direction and transition surface node between the functional axis, mounting end face, hole center and bearing section are extracted to generate the connection transition relationship;

[0355] Functional axes, mounting faces, hole centers, and load-bearing sections are used as functional reference nodes, connection transition relationships are used as skeleton connection edges, and geometric element evidence, functional reference evidence, parameter candidate evidence, and topological relationship evidence are recorded as evidence sources to generate functional reference skeletons for engine parts.

[0356] Retrieve and receive historical model data, extract part types, historical parameter items, geometric evidence sources, measurement relationships, and structural constraints from the historical model data, and generate historical parameter rule records, including:

[0357] The rules for generating historical parameter records are as follows:

[0358] Historical model data is retrieved and received from the historical model database. The historical model data includes historical engine part CAD models, historical part names, historical parameter tables, historical measurement records, and historical structural constraint records.

[0359] Extract part types from historical part names and model baseline records of historical engine part CAD models;

[0360] Read the hole diameter, hole spacing, end face spacing, axial direction, cross-sectional dimensions, and transition connection dimensions from the historical parameter table to generate historical parameter items;

[0361] Geometric evidence sources are generated by reading the surface nodes, boundary nodes, hole nodes, and connecting edges corresponding to the parameters from historical measurement records.

[0362] Based on the starting point, ending point, measurement direction, and measurement benchmark in historical measurement records, measurement relationships are generated;

[0363] Read coaxial constraints, parallel constraints, perpendicular constraints, hole adjacency constraints, and end face connection constraints from historical structural constraint records to generate structural constraints;

[0364] The system associates and stores part types, historical parameter items, sources of geometric evidence, measurement relationships, and structural constraints to generate historical parameter rule records.

[0365] An engine part parameter rule library is generated based on historical parameter rule records, and an engine part parameter rule diagram is constructed by combining a geometric parameter evidence set and an engine part functional baseline skeleton, wherein:

[0366] Constructing the engine part parameter rule diagram, specifically:

[0367] Read part type, historical parameter items, geometric evidence source, measurement relationship and structural constraint from historical parameter rule records;

[0368] Historical parameter items are categorized according to part type to generate a part type rule table;

[0369] By associating historical parameter items with corresponding geometric evidence sources, measurement relationships, and structural constraints, parameter source rules, parameter measurement rules, and parameter constraint rules are generated.

[0370] The part type rule table, parameter source rule, parameter measurement rule, and parameter constraint rule are summarized to generate an engine part parameter rule library;

[0371] Extract geometric primitive evidence, functional benchmark evidence, parameter candidate evidence, and topological relation evidence from the geometric parameter evidence set;

[0372] Read the functional axes, mounting faces, hole centers, load-bearing sections, and connection transitions from the functional reference skeleton of engine parts;

[0373] Functional axis, mounting end face, hole center, load-bearing section, parameter candidate evidence, and historical parameter items are used as rule nodes;

[0374] Geometric evidence sources, measurement relationships, structural constraints, and connection transition relationships are used as rule edges;

[0375] By associating rule nodes with rule edges, a rule graph of engine part parameters is constructed;

[0376] A chain-like fuzzy matching process is performed between the functional baseline skeleton of the engine parts and the parameter rule diagram of the engine parts. Topological consistency, geometric measurability, and structural constraint fit are calculated to generate parameter matching score records, including:

[0377] Generate parameter matching score records, specifically as follows:

[0378] Read the functional axes, mounting faces, hole centers, load-bearing sections, and connection transitions from the functional reference skeleton of engine parts;

[0379] Read the rule nodes, rule edges, geometric evidence sources, measurement relationships, and structural constraints from the engine part parameter rule diagram;

[0380] Using the functional axis as the starting point of the chain matching, a skeleton matching chain is generated according to the connection sequence from the functional axis to the mounting end face, hole center, bearing section and connection transition relationship;

[0381] The functional baseline nodes in the skeleton matching chain are matched with the rule nodes in the engine part parameter rule diagram by name type matching, geometric evidence source matching, and node connection relationship matching to generate node matching results.

[0382] The connection transition relationship in the skeleton matching chain is matched with the regular edge in the engine part parameter rule diagram by matching the connection object and the connection direction to generate the edge matching result.

[0383] The number of successful matches in the node matching results and edge matching results is counted, and the number of successful matches is divided by the total number of nodes and rule edges involved in the matching to generate the topology consistency.

[0384] Based on the measurement relationship, read the corresponding starting point, ending point, measurement direction and measurement reference, and determine whether the candidate evidence of parameters simultaneously has the measurement starting point, measurement ending point and measurement reference;

[0385] The number of candidate evidence for parameters with complete measurement conditions is counted, and the number of complete measurement conditions is divided by the total number of candidate evidence for parameters to generate geometric measurability.

[0386] Based on the structural constraints, read the coaxial constraints, parallel constraints, perpendicular constraints, hole adjacency constraints, and end face connection constraints, and compare the corresponding node relationships in the functional datum skeleton with the structural constraints item by item;

[0387] The number of constraint terms that satisfy the structural constraints is counted, and the number of constraint terms that satisfy the structural constraints is divided by the total number of structural constraints participating in the comparison to generate the structural constraint fit degree.

[0388] The parameter matching score is generated by averaging the topological consistency, geometric measurability, and structural constraint fit.

[0389] The node matching results, edge matching results, topological consistency, geometric measurability, structural constraint fit, and parameter matching score are associated and stored to generate parameter matching score records.

[0390] Based on the parameter matching score records, association locking is performed on the parameter candidate evidence, functional baseline nodes, and geometric evidence sources to generate candidate parameter locking results, where:

[0391] The candidate parameter locking results are generated as follows:

[0392] Read the node matching results, edge matching results, topological consistency, geometric measurability, structural constraint fit, and parameter matching score from the parameter matching score record;

[0393] Filter candidate evidence for parameters with a parameter matching score of not less than 0.80;

[0394] The selected candidate evidence for parameters is associated with the corresponding functional benchmark nodes, sources of geometric evidence, and measurement relationships.

[0395] When multiple parameter candidate evidences exist at the same functional baseline node, retain the parameter candidate evidence with the highest parameter matching score;

[0396] The retained candidate evidence for parameters, functional baseline nodes, sources of geometric evidence, measurement relationships, and parameter matching scores are associated and stored to generate candidate parameter locking results.

[0397] In this embodiment, generating the effective geometric parameter set includes:

[0398] Read the candidate parameter locking results, engine part parameter rule diagram and historical parameter rule record, extract structural constraint terms from the engine part parameter rule diagram and historical parameter rule record, and generate engine part structural constraints;

[0399] Based on the structural constraints of engine parts, the parameter values, functional reference nodes, geometric evidence sources, and measurement relationships in the candidate parameter locking results are subjected to contradictory constraint verification, generating a contradictory constraint verification record, wherein:

[0400] Based on the structural constraints of engine parts, the parameter values, functional reference nodes, sources of geometric evidence, and measurement relationships in the candidate parameter locking results are verified by contradiction constraints, specifically as follows:

[0401] Read parameter values, functional baseline nodes, sources of geometric evidence, measurement relationships, and parameter matching scores from the candidate parameter locking results;

[0402] Read coaxial constraints, parallel constraints, perpendicular constraints, hole adjacency constraints, and end face connection constraints from the structural constraints of engine parts;

[0403] Based on the coaxial constraint, the aperture parameter D is read, and the coaxial deviation threshold is set to 0.05D. The vertical distance from the hole center to the functional axis is calculated and used as the coaxial deviation value. When the coaxial deviation value is not greater than 0.05D, the coaxial verification status is passed and the coaxial conflict is marked as no conflict. When the coaxial deviation value is greater than 0.05D, the coaxial verification status is failed and the coaxial conflict is marked as coaxial deviation conflict.

[0404] Based on the parallel constraint, the parallel angle threshold is set to 3°. The parallel angle deviation value corresponding to the angle between the normal vectors of the two mounting end faces is calculated. When the parallel angle deviation value is not greater than 3°, the parallel verification status is passed and the parallel conflict is marked as no conflict. When the parallel angle deviation value is greater than 3°, the parallel verification status is failed and the parallel conflict is marked as parallel direction conflict.

[0405] Based on the vertical constraint, the vertical angle threshold is set to 3°. The absolute value of the difference between the angle between the normal vector of the installation end face and the direction of the functional axis relative to 90° is calculated as the vertical angle deviation value. When the vertical angle deviation value is not greater than 3°, the vertical verification status is passed and the vertical conflict is marked as no conflict. When the vertical angle deviation value is greater than 3°, the vertical verification status is failed and the vertical conflict is marked as vertical direction conflict.

[0406] The center-to-center distance between adjacent holes is calculated based on the hole adjacency constraint, and the connection relationship between the hole nodes is compared to generate the hole spacing deviation value, hole adjacency verification status, and hole adjacency conflict flag.

[0407] Based on the end-face connection constraints, compare the connection objects and connection directions between the mounting end face, connection edge and transition surface nodes, and generate end-face connection verification status, end-face connection deviation items and end-face connection conflict flags.

[0408] The coaxial verification status, coaxial deviation value, parallel verification status, parallel angle deviation value, perpendicular verification status, perpendicular angle deviation value, hole adjacency verification status, hole spacing deviation value, end face connection verification status, end face connection deviation item and corresponding conflict mark are associated and stored with parameter values, functional reference nodes, geometric evidence sources and measurement relationships to generate a counter-evidence constraint verification record.

[0409] Based on the verification record of the evidence of contradiction constraint, the constraint conflict parameter, the evidence missing parameter, and the matching offset parameter are marked, anomaly parameter records are generated.

[0410] Based on the abnormal parameter records, the evidence set of geometric parameters, the functional baseline skeleton of engine parts, and the parameter rule diagram of engine parts are traced back. The candidate parameter locking results are then subjected to correlation replacement, evidence supplementation, and matching score update to generate corrected parameter locking results, where:

[0411] The results of generating the corrected parameter locking are as follows:

[0412] Read the abnormal parameter identifier, abnormal type, conflict marker, associated functional baseline node, geometric evidence source, measurement relationship and parameter matching score from the abnormal parameter record, and locate the abnormal parameter item in the candidate parameter locking result;

[0413] When the anomaly type is a constraint conflict parameter, backtrack to the engine part parameter rule diagram, read the corresponding rule node, rule edge and structural constraint item, select the parameter candidate evidence that satisfies the structural constraint and has the highest parameter matching score, and replace the abnormal parameter item;

[0414] When the anomaly type is missing parameter evidence, backtrack to the geometric parameter evidence set, read the corresponding geometric primitive evidence, functional benchmark evidence, parameter candidate evidence and topological relationship evidence, and supplement the source of missing geometric evidence.

[0415] When the anomaly type is a matching offset parameter, backtrack to the functional reference skeleton of the engine part and redetermine the functional reference nodes and measurement relationships;

[0416] Based on the replacement results, supplementary selection results, newly determined functional baseline nodes and measurement relationships, the parameter matching score is recalculated, and the corrected parameter locking result is generated.

[0417] Perform consistency cleanup on the modified parameter locking results to generate a valid set of geometric parameters.

[0418] In this embodiment, the output geometric information extraction results and model reconstruction results include:

[0419] Read the effective set of geometric parameters, geometric primitive topology diagram and functional reference skeleton of engine parts, and generate a reconstruction sequence record according to functional axis, mounting end face, hole center, load-bearing section and connection transition relationship;

[0420] Based on the reconstruction order, extract parameter values, functional baseline nodes, and sources of geometric evidence from the effective geometric parameter set to generate a reconstruction parameter sequence;

[0421] Based on the reconstruction parameter sequence, topology-preserving reconstruction is performed on the surface entity, edge entity, vertex entity, and boundary loop entity to generate the reconstructed engine part model, where:

[0422] Based on the reconstruction parameter sequence, topology-preserving reconstruction is performed on surface entities, edge entities, vertex entities, and boundary loop entities, specifically as follows:

[0423] Read parameter values, functional baseline nodes, geometric evidence sources, surface entity identifiers, edge entity identifiers, vertex entity identifiers, and boundary loop entity identifiers from the reconstructed parameter sequence;

[0424] Establish a reconstructed entity index based on functional axes, mounting end faces, hole centers, load-bearing sections, and connection transition relationships;

[0425] Calculate the reconstructed vertex coordinates based on the functional axis, hole center, and parameter values, write the vertex entity identifier, and generate the reconstructed vertex entity;

[0426] Based on the reconstructed vertex entities, edge curve parameter expressions, edge curve parameter ranges, and connection transition relationships, reconstructed edge entities are generated, and reference relationships between the starting vertex entities and the ending vertex entities are established.

[0427] Based on the boundary ring entity identifier, boundary ring path, and edge entity connection order, the edge entities are reconstructed to generate the reconstructed boundary ring entity;

[0428] Generate a reconstructed surface entity based on the surface entity identifier, surface type field, surface parameter expression, and parameter values;

[0429] Write the reconstructed boundary loop entity into the boundary reference relationship of the reconstructed surface entity, while maintaining the topological reference hierarchy between entities;

[0430] Based on the comparison of node connection relationships using geometric evidence sources, the inconsistencies between the first and last vertices of the boundary loop and the inconsistencies in the references between the surface entity and the boundary loop entity are corrected.

[0431] The corrected surface entities, edge entities, vertex entities, and boundary loop entities are associated and stored to generate the reconstructed engine part model;

[0432] The system summarizes the effective set of geometric parameters, the sequence of reconstructed parameters, and the reconstructed engine part model, and outputs the geometric information extraction results and the model reconstruction results.

[0433] Example 1: This example is conducted in a digital reuse verification scenario for engine parts. The system receives historical archived CAD models and CAD models to be reconstructed. All models are engine accessory mounting parts, including mounting faces, coaxial hole groups, load-bearing sections, stiffening surfaces, and transition connection areas. Some archived models lack a complete modeling history tree; hole boundaries are only preserved as boundary ring entities; transition fillets are tightly connected to freeform surfaces; traditional geometric recognition methods easily identify the inner boundaries of holes as ordinary closed contours and the intersection lines of end faces as noise edges, resulting in errors in extracting hole spacing, end face spacing, and axial direction.

[0434] The training sample set contains 10,000 sets of CAD training samples for engine parts, and the test sample set contains 1,200 sets of test samples. Each training sample contains 256 to 2,048 sampling points, 4 to 12 surface nodes, 2 to 10 boundary nodes, 1 to 8 hole nodes, and 3 to 20 connecting edges. A typical training sample contains 132 surface entities, 418 edge entities, 286 vertex entities, 54 boundary loop entities, 16 hole nodes, and 231 connecting edges. It includes manually labeled hole diameter (16 items), hole spacing (24 items), end face spacing (6 items), axial direction (3 items), cross-sectional dimensions (5 items), and transition connection dimensions (18 items). In the typical training sample, four holes are located at the intersection of a freeform surface and a cylindrical surface. Traditional static templates initially missed two holes and identified three end face intersection lines as ordinary boundaries.

[0435] During training, benchmark local geometry blocks are input into the improved PointMLP model. The CAD attribute enhancement input layer reads normalized coordinates, surface category, curvature value, normal vector, boundary type, hole location markers, topological adjacency markers, and geometric evidence indexes to generate point-level CAD attribute features. A geometry-topology dual-neighborhood aggregation layer establishes geometric and topological neighborhoods, with a maximum neighborhood size of 32, an Euclidean distance threshold of 0.08, a curvature difference threshold of 0.12, and a normal angle threshold of 15°. A parameter evidence feedback correction layer outputs geometric primitive evidence, functional benchmark evidence, parameter candidate evidence, and topological relationship evidence. The improved PointMLP model converged after 60 training rounds, with the training loss decreasing from 1.26 to 0.082 and the validation loss decreasing from 1.34 to 0.091. The joint loss variation was less than 0.001 for five consecutive rounds.

[0436] During the testing phase, the system received a 21.4MB CAD model of an engine part to be processed. After parsing, the model yielded 188 surface entities, 642 edge entities, 456 vertex entities, 76 boundary loop entities, and 1512 patch topological references. After model baseline recording and standardization, standard part geometric data was generated, identifying 46 planar surfaces, 31 cylindrical surfaces, 5 conical surfaces, 4 spherical surfaces, and 102 freeform surfaces. The constructed geometric primitive topology graph contained 188 surface nodes, 76 boundary nodes, 26 hole nodes, and 372 connecting edges.

[0437] Sampling was performed based on the geometric primitive topology diagram, resulting in a set of 52,360 sampling point coordinates, including 42,180 surface sampling points, 7,216 boundary sampling points, 26 hole center points, and 2,938 hole sampling points. Each sampling point was assigned a surface category, curvature value, normal vector, boundary type, hole location marker, and topological adjacency marker, forming a derived point set with CAD attribute markers. After CAD topology-weighted Harris 3D detection, 1,968 initial response points were obtained. After merging candidates from the same boundary, the same hole location, and the same connection region, 336 key geometric candidate points were retained. Traditional methods retained 389 key points, with 4 missed detections in the hole location region and 61 false detections in the transition region; the method of this invention had 0 missed detections in the hole location region and 13 false detections in the transition region.

[0438] After backmapping key geometric candidate points, the system establishes surface evidence indexes, boundary evidence indexes, hole location evidence indexes, and topological connectivity evidence indexes, and constructs benchmarked local geometric blocks. The improved PointMLP model outputs 188 sets of geometric primitive evidence, 71 sets of functional benchmark evidence, 136 sets of parameter candidate evidence, and 372 sets of topological relationship evidence. Traditional methods output 177, 59, 119, and 318 sets respectively. After manual review, the accuracy rate of geometric primitive evidence in this invention is 97.6%, the accuracy rate of functional benchmark evidence is 96.1%, and the accuracy rate of parameter candidate evidence is 95.4%; the corresponding accuracy rates of traditional methods are 89.2%, 84.7%, and 82.5%.

[0439] The parameter rule matching stage generates a functional baseline skeleton for engine parts, resulting in 4 functional axes, 7 mounting end faces, 26 hole centers, 9 load-bearing sections, and 52 connection transition relationships. The topological consistency after chain-like fuzzy matching is 0.962, the geometric measurability is 0.978, the structural constraint fit is 0.951, and the average parameter matching score is 0.964. Traditional methods yield corresponding values ​​of 0.884, 0.861, 0.827, and 0.857.

[0440] Based on a parameter matching score threshold of 0.80, the method of this invention identifies 26 aperture parameters, 38 hole spacing parameters, 13 end face spacing parameters, 4 axial direction parameters, 9 cross-sectional dimension parameters, and 34 transition connection dimensions. Verification using contradictory constraints revealed 2 conflicting parameters, 2 missing evidence parameters, and 1 matching offset parameter. After correlation replacement, evidence supplementation, and score updates, the average parameter matching score improved to 0.979. Traditional methods revealed 11 incorrectly identified parameters and 13 unlocked parameters.

[0441] like Figure 2 As shown, the CAD model to be processed contains 188 surface entities, 642 edge entities, 456 vertex entities, and 76 boundary loop entities. The number of entities in the reconstructed model remains the same: 188, 642, 456, and 76, respectively. The number of entities remains consistent before and after reconstruction, and the topological reference integrity rate reaches 100%. After reconstruction using the traditional method, there are 4 instances of inconsistent boundary loop beginnings and ends, 5 instances of inconsistent surface and boundary loop references, and 3 instances of hole center offset, with a topological reference integrity rate of 93.8%.

[0442] like Figure 3 As shown, the method of this invention has 0 missed hole locations, while the traditional method has 4; the number of incorrectly locked parameters is 1, while the traditional method has 11; the topology reference integrity rate is 100%, while the traditional method is 93.8%; the maximum error in hole diameter is 0.016 mm, while the traditional method is 0.064 mm; the maximum error in hole spacing is 0.024 mm, while the traditional method is 0.117 mm; the maximum error in axial direction is 0.19°, while the traditional method is 0.83°; the total time of this invention is 60.0 s, while the traditional method is 214.7 s. The results indicate that the method of this invention can reduce missed hole locations, incorrectly locked parameters, and reconstruction errors, and improve the accuracy of geometric parameter extraction and the topology reference integrity rate.

[0443] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for extracting and reconstructing geometric information of engine parts based on parameter rule matching, characterized in that, include: Obtain the CAD model of the engine part to be processed, perform preprocessing on the CAD model of the engine part to be processed, and generate standard part geometric data; Geometric attributes and topological relationships are extracted from standard part geometric data to generate a geometric primitive topology diagram and a derived point set with CAD attribute tags. The Harris3D algorithm is used to perform key geometric point detection on the derived point set to generate a set of key geometric candidate points. The key geometric candidate point set is back-mapped to the geometric primitive topology graph, a geometric evidence index is established, and a benchmarked local geometric block is constructed. An improved PointMLP model is constructed by inputting a benchmarked local geometric block into the improved PointMLP model and outputting a set of geometric parameter evidence. Based on the geometric parameter evidence set, a functional baseline skeleton of engine parts is generated, an engine part parameter rule diagram is constructed, and a chain matching is performed between the engine part functional baseline skeleton and the engine part parameter rule diagram to generate candidate parameter locking results. By introducing structural constraints on engine parts, the candidate parameter locking results are verified by contradiction constraints, and the candidate parameter locking results that fail the verification are backtracked and corrected to generate a set of valid geometric parameters. Based on the effective set of geometric parameters, topology-preserving reconstruction is performed in the order of functional structure to generate the reconstructed engine part model, and the geometric information extraction results and model reconstruction results are output.

2. The method for extracting and reconstructing geometric information of engine parts based on parameter rule matching according to claim 1, characterized in that, The generation of standard part geometric data includes: Read the file header information, geometric entity data, and topology reference data of the CAD model of the engine part to be processed, and generate the original model parsing data; Based on the original model, the data parsing extracts the format identifier, geometric unit, coordinate datum, and boundary representation entity to generate the model datum record; Based on the model datum record, perform consistency processing on geometric units, coordinate datums, and boundary representation entities to generate datum-consistent geometric data; Extract surface entities, edge entities, vertex entities, boundary loop entities, and patch topological references from the datum-consistent geometric data to generate standard part geometric data.

3. The method for extracting and reconstructing geometric information of engine parts based on parameter rule matching according to claim 1, characterized in that, The generation of derived point sets with CAD attribute tags includes: Read the topological reference relationships of surface entities, edge entities, vertex entities, boundary loop entities, and patches from the standard part geometry data, and generate a geometric entity index; Based on the geometric entity index, extract the surface category, curvature value, normal vector, boundary type, hole position mark, hole position geometric region and adjacent surface connection relationship to generate geometric attribute records; Based on the geometric entity index and geometric attribute records, a geometric primitive topology graph is established using surface entities, boundary ring entities, and hole geometric regions as nodes, and using patch edge sharing relationships, boundary closure relationships, hole adjacency relationships, and geometric constraint relationships as connecting edges. Based on the geometric primitive topology diagram, the surface, boundary loop and hole position are sampled, the coordinates of the sampling point are calculated, and the surface category, curvature value, normal vector, boundary type, hole position mark and topological adjacency mark are written for the sampling point to generate a derived point set with CAD attribute marks.

4. The method for extracting and reconstructing geometric information of engine parts based on parameter rule matching according to claim 1, characterized in that, The generation of the key geometric candidate point set includes: Read the coordinates, normal vectors, curvature values, surface categories, boundary types, hole position markers, and topological adjacency markers of the derived point set with CAD attribute tags, and generate key point detection input data; Based on the key point detection input data, surface detection neighborhood, boundary detection neighborhood and hole detection neighborhood are established respectively. Covariance statistics are performed on the normal distribution of sampling points in each detection neighborhood to generate multi-scale normal covariance records. The underlying Harris response is calculated based on the multi-scale normal covariance record, and the responses at different scales are fused to generate a multi-scale underlying response record. CAD topological weights are generated based on surface type, boundary type, hole location markers, and topological adjacency markers. The CAD topological weights are then written into the multi-scale basic response record to generate a topologically weighted Harris response record. Based on the topologically weighted Harris response record, the sampling points corresponding to the locations of the parametric evidence are screened, and the sampling points in the same boundary, the same hole location, and the same connection area are merged to generate a set of key geometric candidate points.

5. The method for extracting and reconstructing geometric information of engine parts based on parameter rule matching according to claim 1, characterized in that, The construction of the benchmarked local geometry block includes: Read the set of key geometric candidate points and the topology of geometric primitives, and perform back mapping based on the sampling point coordinates, surface type, curvature value, normal vector, boundary type, hole position mark, topological adjacency mark and connecting edge to generate candidate point location records; Establish an index relationship between the candidate point location records and surface nodes, boundary nodes, hole location nodes, and connecting edges to generate a geometric evidence index; Based on the geometric evidence index, derived points corresponding to homologous surfaces, closed boundaries, aperture neighborhoods and topologically adjacent regions are extracted to generate evidence-constrained local geometric blocks. The local reference coordinate records are calculated based on the hole location neighborhood, closed boundary, adjacent surface normal and connecting edge direction in the evidence-constrained local geometry block, and the coordinate reference processing is performed on the evidence-constrained local geometry block to generate a reference local geometry block.

6. The method for extracting and reconstructing geometric information of engine parts based on parameter rule matching according to claim 1, characterized in that, The output geometric parameter evidence set includes: An improved PointMLP model is constructed. The improved PointMLP model retains the residual MLP encoding unit of the original PointMLP model and sets up a CAD attribute enhanced input layer, a geometry-topology dual neighborhood aggregation layer and a parameter evidence feedback correction layer. Read the normalized coordinates, surface category, curvature value, normal vector, boundary type, hole position mark, topological adjacency mark and geometric evidence index from the benchmark local geometry block, and perform attribute embedding and channel splicing through the CAD attribute enhancement input layer to generate point-level CAD attribute features; Geometric and topological neighborhood features are established based on point-level CAD attribute characteristics; The geometric-topological dual-neighborhood aggregation layer performs bi-branch residual encoding, neighborhood weight allocation, and cross-neighborhood fusion on the geometric and topological neighborhood features to generate geometric-topological joint features. The residual MLP encoding unit performs layer-by-layer nonlinear mapping, residual connection and local feature aggregation on the geometric-topological joint features to generate local geometric semantic features. The parameter evidence feedback correction layer performs rule association correction on the local geometric semantic features, generating geometric primitive evidence, functional benchmark evidence, parameter candidate evidence and topological relationship evidence, and summarizing them to generate a geometric parameter evidence set; The improved PointMLP model was trained using geometric primitive evidence classification error, functional baseline evidence localization error, parametric candidate evidence regression error, topological relationship evidence discrimination error, and rule feedback consistency error as joint optimization objectives. The parameters of the CAD attribute enhancement input layer, geometric-topological dual-neighborhood aggregation layer, residual MLP encoding unit, and parametric evidence feedback correction layer were optimized. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved PointMLP model was considered to have completed convergence training.

7. The method for extracting and reconstructing geometric information of engine parts based on parameter rule matching according to claim 6, characterized in that, The improved PointMLP model includes: While retaining the local point set grouping, residual MLP encoding, point-by-point feature mapping, and multi-head output backbone of the PointMLP model, structural modifications are made to the point-level input, neighborhood organization, feature propagation, and evidence output around the task of extracting and reconstructing CAD geometric information of engine parts. A CAD attribute enhancement input layer is added to the model input end, and the normalized coordinates, surface categories, curvature values, normal vectors, boundary types, hole position markers, topological adjacency markers and geometric evidence indexes in the benchmark local geometry are embedded into the evidence channel to generate point-level CAD attribute features. A geometric-topological dual-neighborhood aggregation layer is added at the local feature aggregation location to separately encode the spatial distance neighborhood, curvature change neighborhood, common edge topology neighborhood, and aperture location association neighborhood, and perform cross-neighborhood feature fusion through neighborhood consistency weight; The residual MLP coding unit is improved by transforming the single path point feature mapping into a residual interactive coding structure with geometric branches, topological branches and evidence fusion branches to generate local geometric semantic features. A parameter evidence feedback correction layer is added to the output end to perform rule association correction and evidence confidence recalibration on local geometric semantic features, and output geometric primitive evidence, functional benchmark evidence, parameter candidate evidence and topological relationship evidence.

8. The method for extracting and reconstructing geometric information of engine parts based on parameter rule matching according to claim 1, characterized in that, The generated candidate parameter locking result includes: Read the geometric parameter evidence set and geometric primitive topology diagram, and extract the functional axis, mounting end face, hole center, load-bearing section and connection transition relationship based on geometric primitive evidence, functional benchmark evidence, parameter candidate evidence and topological relationship evidence to generate the functional benchmark skeleton of engine parts; Retrieve and receive historical model data, extract part types, historical parameter items, geometric evidence sources, measurement relationships and structural constraints from the historical model data, and generate historical parameter rule records; Based on historical parameter rule records, an engine part parameter rule library is generated, and an engine part parameter rule diagram is constructed by combining a geometric parameter evidence set and an engine part functional benchmark skeleton. Perform chain-like fuzzy matching between the functional baseline skeleton of engine parts and the parameter rule diagram of engine parts, calculate the topological consistency, geometric measurability and structural constraint fit, and generate parameter matching score records; Based on the parameter matching score records, perform association locking on the parameter candidate evidence, functional baseline nodes, and geometric evidence sources to generate candidate parameter locking results.

9. The method for extracting and reconstructing geometric information of engine parts based on parameter rule matching according to claim 1, characterized in that, The generation of the effective geometric parameter set includes: Read the candidate parameter locking results, engine part parameter rule diagram and historical parameter rule record, extract structural constraint terms from the engine part parameter rule diagram and historical parameter rule record, and generate engine part structural constraints; Based on the structural constraints of engine parts, perform counter-evidence constraint verification on the parameter values, functional reference nodes, geometric evidence sources and measurement relationships in the candidate parameter locking results, and generate counter-evidence constraint verification records; Based on the verification record of the evidence of contradiction constraint, the constraint conflict parameter, the evidence missing parameter, and the matching offset parameter are marked, anomaly parameter records are generated. Based on the abnormal parameter records, the evidence set of geometric parameters, the functional baseline skeleton of engine parts and the parameter rule diagram of engine parts are traced back. The candidate parameter locking results are then subjected to correlation replacement, evidence supplementation and matching score update to generate corrected parameter locking results. Perform consistency cleanup on the modified parameter locking results to generate a valid set of geometric parameters.

10. The method for extracting and reconstructing geometric information of engine parts based on parameter rule matching according to claim 1, characterized in that, The output geometric information extraction results and model reconstruction results include: Read the effective set of geometric parameters, geometric primitive topology diagram and functional reference skeleton of engine parts, and generate a reconstruction sequence record according to functional axis, mounting end face, hole center, load-bearing section and connection transition relationship; Based on the reconstruction order, extract parameter values, functional baseline nodes, and sources of geometric evidence from the effective geometric parameter set to generate a reconstruction parameter sequence; Based on the reconstruction parameter sequence, topology-preserving reconstruction is performed on surface entities, edge entities, vertex entities, and boundary loop entities to generate reconstructed engine part models. The system summarizes the effective set of geometric parameters, the sequence of reconstructed parameters, and the reconstructed engine part model, and outputs the geometric information extraction results and the model reconstruction results.

Images