An AI-driven three-dimensional CAD model manufacturability detection and intelligent repair method

Abstract

This invention relates to the field of 3D modeling technology and discloses an AI-driven method for manufacturability testing and intelligent repair of 3D CAD models. The method includes: importing a solid model of a thin-walled internally cooled part to be tested and constructing a boundary representation topological attribute map; spatially segmenting the internal cavity and extracting channel centerlines, cross-sectional contours, and inner and outer wall thickness mapping data; calculating the risk distribution set of powder discharge bottlenecks based on cross-sectional geometry and objective process constraints, and mapping it to the boundary representation surface; inputting the above parameters into a pre-trained deep learning network to extract repairable surface regions and transition boundary loops; reconstructing the surface using the start and end sections as constraints to generate a repairable surface that reduces powder discharge resistance while maintaining the external functional surfaces; and finally outputting the solid model through topological stitching and manufacturability verification. This invention achieves early targeted repair of manufacturing defects and reduces the probability of powder retention after printing.

Description

Technical Field

[0001] This invention relates to the field of 3D modeling technology, and more specifically, to an AI-driven method for manufacturability testing and intelligent repair of 3D CAD models. Background Technology

[0002] In the design of thin-walled internally cooled parts for metal laser powder bed fusion molding processes, such as mold inserts with conformal cooling channels or heat exchangers with compact channels, there are typically complex protected external aerodynamic surfaces, internal meandering cooling channels, and microscopic turbulence textures distributed on the inner wall. Due to the low thermal conductivity of metal powder in the forming process, the tendency of the molten pool to sink into the overhanging area, and the inability to apply removable supports to the internal enclosed space, unmelted powder can only be discharged through a limited number of ports by gravity or high-pressure fluid after printing.

[0003] Currently, the industry primarily uses general-purpose computer-aided geometric repair tools to process imported models for compliance. Existing conventional repair tools mainly focus on checking the bounding box volume, global minimum wall thickness, or connectivity of local patches of the model, aiming to compensate for topological damage to achieve a watertight state. However, under real printing conditions, due to the spatial misalignment between the cross-sectional shrinkage of the curved section of the internal cavity and the gravity powder discharge path, as well as the objective existence of the rough thickening of the overhanging surface, hidden powder retention and channel blockage can easily occur in the internal flow channels.

[0004] When existing automatic repair algorithms smooth out minor internal defects or reconstruct transition fillets in pursuit of smooth local surfaces, they often push the already narrow throat of the cross-section further towards the side against gravity or the side of overhang deformation. These manufacturing hazards, caused by the interference of surface shape, wall thickness configuration, and powder removal path in three-dimensional space, are difficult to screen in advance using conventional volumetric measurements or global Boolean operations. This single-minded pursuit of topological closure in repair methods amplifies local thermal deformation and resistance within the cavity, leading to a large number of 3D models carrying these hazards directly entering the printing process, ultimately resulting in costly scrap of printed parts. Summary of the Invention

[0005] This invention provides an AI-driven method for manufacturability testing and intelligent repair of 3D CAD models, which solves the technical problems mentioned in the background art.

[0006] This invention provides an AI-driven method for manufacturability testing and intelligent repair of 3D CAD models, applicable to 3D computer-aided design solid models of thin-walled internally cooled structural components containing internal channels and protected external functional surfaces, including: Import the 3D computer-aided design solid model to be inspected, extract the boundary representation surfaces, geometric and topological information to construct the boundary representation topological attribute map; The internal channel in the three-dimensional computer-aided design solid model is segmented into cavity space, and the internal channel centerline, channel geometric features, cross-sectional contour set and inner and outer wall thickness mapping dataset are extracted. The risk distribution set of the powder discharge obstruction bottleneck is calculated based on the channel geometric features, and the risk distribution set of the powder discharge obstruction bottleneck is mapped onto the boundary representation surface; The boundary representation topological attribute map, the cross-sectional contour set, and the mapped risk distribution set of powder discharge obstruction bottlenecks are input into a pre-trained deep learning network model to extract repairable surface regions and transition boundary loops on the internal channels. The internal channel surface is reconstructed on the repairable region to generate a repair surface that adjusts the powder discharge resistance while maintaining the geometric features of the protected external functional surface. The repaired surface is topologically stitched and manufacturability verified with the three-dimensional computer-aided design solid model to generate a topologically closed repaired solid model. Output the repaired entity model with boundary representation topological features.

[0007] Beneficial effects include: This invention directly manipulates computer-aided design of solid surfaces and edge topological elements, mapping potential manufacturing risks caused by powder rejection, rough sagging of overhanging surfaces, and insufficient thin-wall thickness in powder bed melting processes into surface-level spatial distribution parameters; it uses artificial intelligence deep learning networks to locate high-risk continuous surface regions, and performs three-dimensional continuity reconstruction and topological stitching of internal channel surfaces while maintaining the unalterable external aerodynamic shape. This not only avoids the hidden cross-sectional shrinkage and internal wall warping amplification caused by blindly pursuing smooth surfaces in conventional repairs, but also ensures that the repaired model retains complete topological boundary expression and editability, directly meeting the process compliance verification requirements in the early stages of additive manufacturing, and effectively reducing the printing scrap rate of high-value thin-walled internally cooled parts. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the overall mechanism for AI-driven manufacturability testing and intelligent repair of 3D CAD models according to the present invention; Figure 2 This is a schematic diagram of the mechanism for identifying bottleneck risks and extracting repairable areas in powder discharge according to the present invention; Figure 3 This is a schematic diagram of the internal channel repair surface reconstruction, topology stitching and verification output mechanism of the present invention. Detailed Implementation

[0009] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0010] As an optional application scenario of this invention, the method provided by this invention can be applied to a three-dimensional computer-aided design solid model processing and repair system. The following details each execution step of the method: like Figure 1 As shown, an AI-driven method for manufacturability testing and intelligent repair of 3D CAD models is applied to the 3D computer-aided design solid model of a thin-walled internally cooled structural component containing internal channels and protected external functional surfaces, including: Import the 3D computer-aided design solid model to be inspected, extract the boundary representation surfaces, geometric and topological information to construct the boundary representation topological attribute map; The internal channel in the three-dimensional computer-aided design solid model is segmented into cavity space, and the internal channel centerline, channel geometric features, cross-sectional contour set and inner and outer wall thickness mapping dataset are extracted. The risk distribution set of the powder discharge obstruction bottleneck is calculated based on the channel geometric features, and the risk distribution set of the powder discharge obstruction bottleneck is mapped onto the boundary representation surface; The boundary representation topological attribute map, the cross-sectional contour set, and the mapped risk distribution set of powder discharge obstruction bottlenecks are input into a pre-trained deep learning network model to extract repairable surface regions and transition boundary loops on the internal channels. The internal channel surface is reconstructed on the repairable region to generate a repair surface that adjusts the powder discharge resistance while maintaining the geometric features of the protected external functional surface. The repaired surface is topologically stitched and manufacturability verified with the three-dimensional computer-aided design solid model to generate a topologically closed repaired solid model. Output the repaired entity model with boundary representation topological features.

[0011] S201, parse the topological structure of the three-dimensional computer-aided design solid model, extract the volume elements, shell elements, boundary representation surfaces, ring elements, edge elements and vertex elements, and perform consistent alignment of the surface normal vectors of all boundary representation surfaces.

[0012] Specifically, when processing the 3D computer-aided design solid model, the topology traversal interface of the geometric modeling kernel is invoked to deconstruct the input boundary representation solid model file from top to bottom. In some optional implementations, the volume elements representing the closed 3D solid are first identified, and then the shell elements constituting the boundaries of the volume elements are obtained. Further, each independent boundary representation surface is decomposed from the shell elements. For each boundary representation surface, a closed directed loop that defines its bounded range, i.e., a loop element, is extracted. The edge elements that make up the boundary are further traced downwards through the loop elements, and the endpoints of the edge elements, i.e., vertex elements, are located. After extracting all boundary representation surfaces, in order to establish a unified spatial inside-outside relationship, this embodiment performs a consistent alignment operation of the surface normal vector. This operation selects a reference boundary representation surface and sets its surface normal vector to point outwards from the solid. Then, using the topology sharing relationship of adjacent edges, a region growing algorithm is used to propagate to adjacent boundary representation surfaces. Finally, during the propagation process, the parameterization direction of adjacent boundary representation surfaces is adjusted by calculating the normal transition state of adjacent boundary representation surfaces at the common edge, so that the surface normal vectors of all boundary representation surfaces on the entire shell element point to the external space of the solid model.

[0013] S202, extract the fluid port boundary ring, and classify and label the boundary representation surfaces corresponding to the internal channels and the boundary representation surfaces corresponding to the external functional surfaces based on the spatial bounding box nesting relationship of the topological elements and the fluid port boundary ring.

[0014] Specifically, a fluid port boundary ring refers to a boundary curve ring where the internal cavity fluid channel extends to the outer surface of the entity and forms an opening. In some optional embodiments, specific ring elements that are referenced only by one boundary representation surface or located at the junction of the outer shell and the internal cavity are identified and extracted as fluid port boundary rings. Further, the spatial bounding box nesting relationship is calculated by retrieving the triaxial minimum bounding cube of each shell element or set of independent faces in three-dimensional space. When it is detected that the spatial bounding box of an inner shell element is completely enclosed within the spatial bounding box of an outer shell element, it indicates that the topological set corresponding to the inner shell element belongs to the fluid channel or cavity inside the entity. The refinement process of classification and labeling is to use the extracted fluid port boundary ring as the starting topological node and perform a topological depth traversal along the edge elements that are topologically adjacent to it into the entity. In this embodiment, all boundary representation surfaces that can be connected to the fluid port boundary ring through topological adjacency and are located in the inner layer of the spatial bounding box nesting relationship are classified and labeled as the boundary representation surfaces corresponding to the internal channels. Correspondingly, boundary representation surfaces that cannot be connected through the fluid port boundary ring, or that are located in the outermost envelope of the entity and serve as aerodynamic shapes or assembly supports, are classified and marked as boundary representation surfaces corresponding to external functional surfaces.

[0015] S203, construct a relational graph with each boundary surface as a graph node, and record geometric and topological attributes in the graph node. The geometric and topological attributes include surface geometry type, parameter domain boundary features, adjacent edge topological relationship, edge concavity / convexity attribute, surface area, proportion of short edges, and shell identifier.

[0016] Specifically, a relational graph is a graph-structured data where each extracted boundary representation surface independently corresponds to a graph node. If two boundary representation surfaces share at least one common edge element topologically, an undirected edge is established between their corresponding graph nodes to express topological adjacency. In some optional implementations, each graph node records its specific geometric and topological attributes using digitized vectors or discrete codes. Surface geometry types are recorded using integer codes; for example, value 1 represents a plane, value 2 a cylinder, value 3 a cone, value 4 a sphere, value 5 a torus, and value 6 a non-uniform rational spline surface. The parameter domain boundary features record the domain range and periodicity characteristics of the boundary representation surface in the two-dimensional parameter space, including the maximum and minimum values ​​of the horizontal and vertical parameters. The adjacency edge topological relationship records the identifier numbers of all other surface nodes adjacent to the current surface, as well as the identifiers of their corresponding common edge elements. The convexity / concavity attribute of an edge is determined by calculating the relative angle between the surface normal vectors of the two boundary representation surfaces on the common edge at the sampling point of the common edge. If the two surface normal vectors intersect towards the outside of the entity, the adjacent edge is defined as a convex edge; if they are opposite to the intersection, it is defined as a concave edge. The surface area is a floating-point value obtained by numerically integrating the first type of surface integral of the first basic form of the surface within the domain of the two-dimensional parameter space. Further, the proportion of short edges is obtained by first setting a length threshold, which is determined based on the minimum spot diameter or minimum manufacturable feature size of laser powder bed melting, preferably ranging from 0.2 to 1.0 mm, for example, set to 0.5 mm; then counting the number of edge elements whose three-dimensional arc length is less than the length threshold among all edge elements contained in the current boundary representation surface, and using this number as the numerator; then obtaining the total number of edge elements contained in the current boundary representation surface as the denominator; the ratio obtained by dividing the numerator by the denominator is the proportion of short edges. Finally, the shell identifier is an integer label used to distinguish different connected topological blocks, recording the specific shell element number to which the graph node belongs.

[0017] S204, set the protected identifier, internal channel identifier and repairable identifier in the graph node respectively to complete the construction of the boundary representation topology attribute graph.

[0018] Specifically, the protected identifier, internal channel identifier, and repairable identifier are all state bits or multi-dimensional classification feature vectors. Based on the classification results, for topological elements marked as boundary representation surfaces corresponding to external functional surfaces and identified as fluid port boundary loops, their corresponding protected identifiers in the graph nodes are set to true. This indicates that in subsequent geometric deformation and surface reconstruction, the geometric coordinates and parameterized equations contained in these graph nodes remain fixed and cannot be tampered with or moved. In some optional implementations, for topological elements marked as boundary representation surfaces corresponding to internal channels, their corresponding internal channel identifiers in the graph nodes are set to true, indicating that they belong to the flow-through walls of fluid flow. Further, the repairable identifier is initialized in conjunction with the manufacturability bottleneck risk. For graph nodes with a true internal channel identifier and a false protected identifier, their repairable identifier is preset to true, assigning them as candidate attributes for subsequent deep learning networks to extract repairable regions and perform surface lofting reconstruction. Finally, once these three identifiers for all graph nodes are set, each graph node not only contains the underlying geometric metric attributes but also incorporates the high-level manufacturing boundary semantics, thus completing the full construction of the boundary representation topology attribute graph.

[0019] S301, a region growth algorithm with the fluid port boundary ring as the starting node is used to extract the fluid communication cavity.

[0020] Specifically, the previously identified fluid port boundary ring is used as the seed source for region growth. In some optional implementations, the region growth algorithm is executed in the interior space of a 3D solid or in the voxelized space of a 3D mesh. Starting from the topological boundary location of the fluid port boundary ring, multi-directional expansion is performed towards the hollow regions inside the solid that are not blocked by solid surfaces. Further, the algorithm searches layer by layer for spatial containing cells or topological surfaces adjacent to the currently known cavity. As long as the adjacent spatial region is not blocked by the solid walls of the solid and maintains geometric connectivity, it is included in the current growth region. This iterative process continues until the growth boundary contacts all closed boundary representations of the internal channels and can no longer extend outward. Finally, the set of all traversed and merged connected hollow spaces is extracted and defined as the fluid-connected cavity.

[0021] S302, the topological center of the fluid-connected cavity is extracted to generate the centerline of the internal channel, and a normal plane intersection profile orthogonal to the tangential vector of the centerline is established along the centerline of the internal channel to generate the cross-sectional profile set.

[0022] Specifically, the topology center extraction employs a mid-axis transformation algorithm or a three-dimensional distance field refinement algorithm to shrink and skeletonize the fluid-connected cavity inward, stripping away the outer space and ultimately leaving a one-dimensional geometric center curve that characterizes the cavity's extension trend, thus generating the internal channel centerline. In some optional embodiments, discrete point sampling is then performed on the internal channel centerline at a set step size. The sampling step size is dynamically adjusted according to the channel's curvature, decreasing in areas of high curvature and increasing in flat areas. The sampling step size ranges from 0.1 to 1.0 mm, and the product of the sampling step size and the local curvature should be less than 0.1. Further, at each sampling point, the local tangent direction of the centerline at that point is calculated to obtain the centerline tangential vector. Using this sampling point as the origin, a two-dimensional plane orthogonal to the centerline tangential vector is constructed, i.e., the normal plane. The normal plane is then subjected to three-dimensional geometric intersection calculation with the inner wall surface of the fluid-connected cavity, and the resulting closed intersection profile is the normal plane intersection profile. Finally, all sampling points on the centerline of the internal channel are traversed, and all the resulting intersecting contours of the normal planes are sequentially arranged and combined to generate the set of cross-sectional contours.

[0023] S303, for each section in the set of cross-sectional profiles, calculate the cross-sectional area, perimeter, hydraulic diameter, local tangential vector and path direction vector pointing to the corresponding fluid port boundary ring, and extract the boundary representation surface that intersects with the cross-section.

[0024] Specifically, for each independent intersecting contour of the normal plane in the set of cross-sectional contours, the area of ​​the closed two-dimensional region enclosed by it is calculated using Green's formula or the numerical integration method of polygon area to obtain the cross-sectional area. The perimeter is calculated by integrating and summing the arc lengths of each segment of the contour line. In some optional embodiments, the hydraulic diameter is calculated by dividing four times the cross-sectional area by the sum of the perimeter and the geometric calculation tolerance. Here, the geometric calculation tolerance is a preset floating-point number to prevent the denominator from being zero, and its preferred value range is set to 10. -8 Up to 10 -5 mm, for example, set to 0.000001 mm. The local tangential vector directly inherits the centerline tangential vector at the point where the cross-section resistance is calculated. Further, the path direction vector pointing to the corresponding fluid port boundary ring is obtained by constructing a topological path along the centerline of the internal channel from the current cross-section sampling point to the nearest fluid port boundary ring, and calculating the unit direction vector of this path at the current sampling point. Finally, by searching the geometric topological relationships, it is found which boundary representation surfaces intersect the normal plane of the current cross-section at spatial points or lines, and these traversed surfaces are extracted and recorded as boundary representation surfaces intersecting the cross-section.

[0025] S304, find the intersection of the surface normal vectors of the intersecting boundary representation surfaces with the outside of the solid, and calculate the matching solid surfaces to generate the inner and outer wall thickness mapping dataset.

[0026] Specifically, for each boundary surface intersecting the cross-section, point sampling is performed on its intersection line with the cross-sectional profile. At each sampling point, the surface normal vector of the boundary surface is obtained. Since the surface normal vector has been uniformly aligned, it points inward to the solid metal material and towards the external functional surface. In some optional implementations, the algorithm uses the sampling point as the starting point and the surface normal vector as the ray emission direction to emit a geometric ray deep into the internal material of the solid. This ray passes through the solid wall thickness of the thin-walled part and eventually intersects with the outermost surface of the part or the surface of the adjacent cavity, producing an intersection point. This outer surface hit by the ray is calculated and defined as the matched solid surface. Further, the three-dimensional Euclidean distance between the starting point and the intersection point is calculated, and this distance value represents the local solid wall thickness at that location. Finally, the local solid wall thickness values ​​calculated from all sampling points on all cross-sections, the corresponding starting point coordinates, the intersection point coordinates, and the matched solid surface identifiers are associated and integrated to form a structured data matrix, thereby generating the inner and outer wall thickness mapping dataset.

[0027] S305, based on the set construction direction vector, calculate the curved surface region on the cross-sectional boundary that satisfies the set included angle threshold and faces the powder support side, and mark it as the overhanging surface region.

[0028] Specifically, the set build direction vector refers to the unit direction vector of the worktable layer by layer during printing in a metal laser powder bed fusion forming device, represented as a unit vector along the vertical axis in a three-dimensional spatial coordinate system. For any position on the cross-sectional contour boundary, the surface normal vector of the corresponding intersecting boundary surface is obtained. In some optional embodiments, the spatial angle between this surface normal vector and the set build direction vector is calculated. Since there is no solid metal support below the molten metal pool in the overhanging position, it is only supported by loose powder with low thermal conductivity, which easily leads to sinking. Further, the set angle threshold is a critical angle predetermined based on the forming limit of a specific metal material and printing equipment, preferably ranging from 20 degrees to 45 degrees, for example, set to 30 degrees. When the angle between the surface normal vector and the opposite direction of the set build direction vector is less than or equal to the set angle threshold, it indicates that the surface slope of this area is relatively gentle and faces downwards, belonging to the area that requires powder pad support during the printing process. Finally, all spatial ranges on the cross-sectional boundary that satisfy this geometric condition are calculated and uniformly identified as the overhanging surface region.

[0029] S401, for each cross section, obtain the hydraulic diameter of the adjacent cross sections of the current cross section, add the hydraulic diameters of the adjacent cross sections together and divide by the value 2 to obtain the reference hydraulic diameter of the adjacent cross sections.

[0030] Specifically, if the current section is the starting or ending section, the hydraulic diameter of the missing adjacent section is directly taken as the hydraulic diameter of the current section. In some optional embodiments, the hydraulic diameter of the current section is then subtracted from the reference hydraulic diameter of the adjacent section to obtain the difference. This difference is compared with the value 0, and the larger of the two values ​​is taken as the shrinkage numerator. Further, the reference hydraulic diameter of the adjacent section is then added to the geometric calculation tolerance to obtain the shrinkage denominator. Finally, the shrinkage numerator is divided by the shrinkage denominator to obtain the local section shrinkage characteristic. The geometric calculation tolerance is a preset minimum floating-point number, preferably within the range of 10. -8 Up to 10 -5 mm, for example, a value of 0.000001 mm.

[0031] S402, the gravity-assisted powder removal direction vector is the three-dimensional spatial vector of the gravity direction of the auxiliary powder falling when removing unmelted powder after the metal forming process is completed.

[0032] Specifically, during the calculation, the local tangential vector is multiplied by the set gravity-assisted powder removal direction vector, and the result of the multiplication is inversely represented to obtain the powder removal resistance assessment value. In some optional embodiments, this powder removal resistance assessment value is then compared with the value 0, and the larger of the two values ​​is taken to calculate and generate the anti-gravity powder removal resistance characteristic.

[0033] S403, the set critical self-supporting angle refers to the minimum support angle required for a specific metal material to not collapse locally when it is melted and formed in a laser powder bed, and its preferred value range is between 20 degrees and 45 degrees.

[0034] Specifically, the set of intersection lines between the cross section and the intersecting boundary representation surfaces is extracted, and multiple spatial sampling points are selected on these intersection lines. In some optional embodiments, for each spatial sampling point, the surface normal vector at that point is multiplied by the construction direction vector, and the minimum value is found from the multiplication results of all spatial sampling points and defined as the minimum normal projection. Further, before calculation, the critical self-supporting angle is converted from angle to radians, the cosine value of the set critical self-supporting angle is calculated, the minimum normal projection is subtracted from the cosine value, and the result of the subtraction is compared with the value 0. The larger of the two values ​​is taken as the defect tendency numerator. Finally, the cosine value of the set critical self-supporting angle is added to the value 1, and the numerical calculation tolerance is added to obtain the defect tendency denominator. Finally, the defect tendency numerator is divided by the defect tendency denominator to obtain the forming defect tendency characteristics of the overhanging surface region. Here, the numerical calculation tolerance is a preset minimum floating-point number, preferably within the range of 10. -8 Up to 10 -5 For example, the value can be 0.000001.

[0035] S404, the set allowable wall thickness threshold refers to the minimum solid thin-wall size that can be stably printed in the metal forming process without thermal breakdown or deformation instability. Its preferred value is usually set to 0.2 mm to 0.5 mm.

[0036] Specifically, based on the inner and outer wall thickness mapping dataset, all wall thickness data corresponding to the current cross-section are extracted, and the average value is calculated to obtain the average wall thickness. In some optional implementations, the average wall thickness is subtracted from the set allowable wall thickness threshold to obtain the wall thickness difference. This wall thickness difference is compared with the value 0, and the larger of the two values ​​is taken as the out-of-tolerance numerator. Further, the set allowable wall thickness threshold is added to the geometric calculation tolerance to obtain the out-of-tolerance denominator. Finally, the out-of-tolerance numerator is divided by the out-of-tolerance denominator to calculate the wall thickness out-of-tolerance feature.

[0037] S405, extract the local curvature of the centerline of the internal channel at the current section, and calculate the absolute value of the local curvature.

[0038] Specifically, the local curvature is the reciprocal of the radius of curvature. The absolute value of this local curvature is multiplied by the hydraulic diameter to obtain the bending product term. In some optional embodiments, this bending product term is used as the bending numerator, and 1 is added to the bending product term to obtain the bending denominator. Finally, the bending numerator is divided by the bending denominator to calculate the flow channel bending characteristics.

[0039] S406, the shrinkage feature (i.e., local cross-sectional shrinkage feature), the resistance feature (i.e., anti-gravity powder discharge resistance feature), and the defect tendency feature (i.e., forming defect tendency feature) are multiplied together, and the product is defined as the spatial coupling feature.

[0040] Specifically, this spatial coupling feature is used to capture the associated physical risks of individual indicators not exceeding limits but combining to form hidden channel blockages. In some optional implementations, one bias weight coefficient and six feature weight coefficients given by a pre-trained deep learning network are obtained in advance. Further, the local cross-sectional contraction feature, anti-gravity powder discharge resistance feature, forming defect tendency feature, flow channel bending feature, wall thickness deviation feature, and spatial coupling feature are multiplied by their corresponding feature weight coefficients. All the resulting products are accumulated and then added to the bias weight coefficient to obtain the feature fusion value, thereby fusing the contraction feature, resistance feature, defect tendency feature, flow channel bending feature, and wall thickness deviation feature into the channel geometric feature. Finally, the feature fusion value is processed by a nonlinear mapping function. Specifically, the inverse of the feature fusion value is obtained, and a power function value with the natural constant as the base and the inverse as the exponent is calculated. The power function value is added to the value of 1 to obtain a nonlinear denominator. The value of 1 is divided by the nonlinear denominator to convert the output into a cross-sectional comprehensive risk coefficient.

[0041] S407, for each boundary representation surface, traverse all sections contained in the internal channel.

[0042] Specifically, the shortest three-dimensional spatial distance from the current boundary surface to the current cross-section is calculated. In some optional embodiments, this shortest three-dimensional spatial distance is divided by the sum of the hydraulic diameter of the current cross-section and the geometric calculation tolerance to obtain the distance attenuation index. Here, the geometric calculation tolerance is a preset, extremely small floating-point number, preferably within the range of 10. -8 Up to 10 -5 mm, for example, a value of 0.000001 mm. Further, the inverse of the distance attenuation exponent is taken, and an attenuation weight term is calculated with the natural constant as the base and the inverse as the exponent. This process constitutes the spatial distance attenuation function. The cross-sectional level comprehensive risk coefficient of the current cross-section is multiplied by this attenuation weight term to obtain the risk mapping projection value of the current boundary representation surface relative to the current cross-section. Finally, among all the risk mapping projection values ​​corresponding to all cross-sections, the one with the largest value is selected and assigned to the current boundary representation surface as the surface-level risk feature value. After processing all boundary representation surfaces, the generated surface-level risk feature values ​​are summarized to construct the risk distribution set of the powder discharge obstruction bottleneck.

[0043] S501: For each graph node in the constructed relational graph, construct a multidimensional feature vector for each boundary representation surface.

[0044] Specifically, the multidimensional feature vector is a one-dimensional feature array containing multiple scaling values. The system retrieves the recorded face geometry type, surface area, proportion of short sides, concavity / convexity attributes of adjacent edges, and preliminarily calculated face-level risk feature values ​​from the current graph node. In some optional implementations, the surface curvature is calculated by uniformly sampling spatial grid points within the two-dimensional parameter domain of the corresponding boundary representation surface, calculating the first and second principal curvatures of each spatial grid sampling point, and taking the arithmetic mean of the absolute values ​​of the principal curvatures of all spatial grid sampling points. This arithmetic mean is used as the surface curvature value of the boundary representation surface. Further, the method for obtaining the assigned channel number is to perform a depth-first topological traversal of the connected topological blocks inside the three-dimensional entity to find independent branch channels, assigning a unique positive integer index to each unbranched connected cavity branch, finding the specific branch region in three-dimensional space where the current boundary representation surface falls, and defining the positive integer index assigned to that specific branch region as the assigned channel number. The path length parameter from the fluid port boundary ring is calculated as follows: The geometric centroid coordinates of the current boundary representation surface are extracted. The nearest orthogonal projection point along the internal channel centerline is found. Through three-dimensional curve integration, the arc length of the three-dimensional curve extending from this orthogonal projection point along the internal channel centerline to the nearest fluid port boundary ring is obtained as a floating-point value. This floating-point value is defined as the path length parameter from the fluid port boundary ring. Finally, all extracted items and calculated values ​​are concatenated according to a fixed array index order to generate the multidimensional feature vector of the corresponding boundary representation surface.

[0045] S502, after all nodes have completed the construction of multidimensional feature vectors, uses a graph attention mechanism to perform information transfer and aggregation of multidimensional feature vectors between adjacent boundary representation surface graph nodes with shared edge topological relationships.

[0046] Specifically, Z-score standardization is applied to all multidimensional feature vectors, transforming features of different magnitudes into a distribution with a mean of 0 and a variance of 1. In the graph neural network computational architecture, a set of learnable weight matrix variables is pre-set. In some optional implementations, for any central target graph node in the relational graph, all associated topological neighbor graph nodes are indexed through undirected edges. Multidimensional feature vectors of the central target graph node and each topological neighbor graph node are extracted. Further, the learnable weight matrix variables are used to perform a linear projection transformation on all extracted multidimensional feature vectors. The transformed features of the central target graph node and the features of the specific topological neighbor graph node are concatenated, and the resulting combined input features are fed into a single-layer feedforward neural network. The nonlinear activation function value of the output of this single-layer feedforward neural network is calculated, and this activation value is used as an attention coefficient to measure the importance of the spatial geometry and pollutant rejection risk association between the central target graph node and the specific topological neighbor graph node. Finally, using all the calculated corresponding attention coefficients, a weighted summation operation is performed on the projected transformation features of all topological neighbor graph nodes. The accumulated feature vector generated by weighted summation is added to the self-projected feature of the central target graph node to generate an aggregated multidimensional feature vector. This completes the information transfer and aggregation of the multidimensional feature vector in a single round, and the surface-level risk feature value carried by the graph node is updated according to the corresponding dimension value in the aggregated feature vector.

[0047] S503, obtain a preset floating-point constant as the set risk threshold. The preferred value range of the set risk threshold is between 0.6 and 0.8, for example, set the value to 0.7.

[0048] Specifically, all graph nodes carrying internal channel-related topological markers are traversed in the relational graph. The surface-level risk feature value of each graph node, updated through a graph attention mechanism, is read. In some optional implementations, the read surface-level risk feature value is compared with a set risk threshold. When the surface-level risk feature value is greater than the set risk threshold, the boundary representation surface corresponding to the graph node is extracted as a high-risk candidate surface. Further, among all candidate surfaces, connectivity component analysis is performed using their shared edge topological relationships on the 3D solid structure to merge spatially connected and uninterrupted geometric surfaces into a continuous channel inner wall surface. Association retrieval is performed in the 3D solid boundary representation topology to find other graph nodes that share common edge elements with the extracted continuous channel inner wall surfaces. Graph nodes exhibiting excessively rounded corner geometric features are selected, and their corresponding spatial surfaces are identified as rounded boundary representation surfaces. Graph nodes with convex micro-geometric features that undertake forced fluid heat transfer are selected, and their corresponding spatial surfaces are identified as turbulence boundary representation surfaces. Finally, the continuous channel inner wall surfaces, rounded boundary representation surfaces, and turbulence boundary representation surfaces that satisfy the common edge topology relationship are combined and merged in the topology data structure to form a continuous set of high-risk topology surfaces. The high-risk topology surface set after the overall combination is uniformly marked to form the repairable surface region.

[0049] S504, Locate the two end boundaries of the repairable surface region in the entire internal cavity interconnected space.

[0050] Specifically, starting from the two boundary positions of the repairable surface region, the system extends spatially to both sides along the tangential vector direction of the internal channel centerline. In some optional embodiments, during the segmental extension along the tangential vector direction, the surface-level risk characteristic values ​​of the intersection graph nodes associated with the passed positions are read in real time, and their numerical relationship with a set risk threshold is compared. The system continues to traverse away from the repairable surface region until a specific spatial cross-section position is found. At this specific spatial cross-section position, all surface-level risk characteristic values ​​associated with its surroundings decrease and are less than or equal to the set risk threshold. Further, at this spatial position where the powder retention risk is within the set range, the spatial topological connectivity of its corresponding cross-section is examined. When the edge elements constituting this cross-section are connected end-to-end in three-dimensional space, forming a closed shape without gaps or breaks, this position is determined to be the location where the surface-level risk characteristic value drops to within the set risk threshold and the cross-section topology is closed. At this location that meets specific conditions, all peripheral edge elements constituting the closed geometric boundary of the cross-section are extracted. These end-to-end edge elements are then topologically merged and extracted to obtain the cross-sectional boundary ring at the corresponding location as the start and end cross-sectional boundary rings. Finally, a start and end cross-sectional boundary ring is extracted at each of the two ends of the extension of the repairable region. These topological boundary rings that are closed in three-dimensional space and located in low-risk geometric regions are paired and collectively defined as the transition boundary ring.

[0051] S601, extract the two end face contours in the geometric region with low powder retention risk, and set them as the start and end section boundary rings.

[0052] Specifically, at the start of the repair operation on the internal cavity, two specific boundary loops are assigned fixed and immovable attribute parameters in the data structure of the 3D solid model, serving as fixed constraint boundaries. In some optional implementations, under the constraint of these fixed constraint boundaries, a sequence of continuous edge elements connecting these two boundary loops is extracted from the inner surface topology of the 3D solid, and combined and spliced ​​to form spatial boundary curves that constitute the closed contour of the corresponding internal channel. Finally, these spatial boundary curves close in 3D space, defining the exact physical range of the surface reconstruction required, restricting the geometric modification to the local channel where there is a manufacturing hazard, and preventing the modification range from extending outward and affecting the external functional surface.

[0053] S602 invokes a cubic polynomial interpolation algorithm to perform curve generation calculations in the three-dimensional space between two fixed constraint boundaries.

[0054] Specifically, multiple two-dimensional or three-dimensional closed loops are generated by calculating the set offset step length, forming a set of reconstructed cross-sectional curves. In some optional embodiments, two numerical constraints are applied when calculating and generating this set of reconstructed cross-sectional curves. The first constraint is that the hydraulic diameter of the reconstructed cross-section curves varies with a gentle gradient along the centerline of the internal channel, and the ratio of the change in hydraulic diameter of adjacent reconstructed cross-sections to the centerline step length does not exceed 0.1, avoiding abrupt local contraction. Further, the second constraint is to calculate the shortest vertical distance from the channel boundary to the centerline of the internal channel, thereby locating the narrowest spatial node, i.e., the minimum gap position. The coordinates of this minimum gap position are translated in the two-dimensional parameter space of the cross-section, constraining the minimum gap position to shift from the overhanging surface region to the non-overhanging surface region. This geometric translation shifts the narrow throat within the channel, which is most prone to powder blockage, from the printing support surface side, which is prone to sinking and slag buildup, to a lateral or upper region with better forming quality.

[0055] S603, locate the inner wall surface on the side of the printing support surface, and extract the topology node data corresponding to these surfaces.

[0056] Specifically, the boundary representation surface of the overhanging area is topologically reconstructed to eliminate microscopic pits, arrays of protrusions, or sharp transition chamfers present in the original design. In some optional embodiments, a smooth three-dimensional patch is reconstructed through surface fitting to generate a guide surface with a monotonically transitioning direction along the powder discharge path. Further, monotonically transitioning means that the spatial partial derivative at each point on this guide surface is calculated along the gravity-driven powder discharge tangent, with the direction of the partial derivative remaining constant and without sign reversal. This spatial geometry eliminates all local adverse obstruction structures during powder descent, providing a smooth downward channel and reducing the probability of unmelted metal particles adhering to and remaining on the overhanging surface.

[0057] S604, the reconstructed cross-sectional curve is used as a guide line, and the generated guide surface data is combined and input into the surface lofting calculator.

[0058] Specifically, a lofted surface is generated by lofting a non-uniform rational spline surface. In some optional implementations, the lofting calculation uses a continuously arranged cross-sectional profile in space as the control reference to calculate and generate node vectors and control point matrices, constructing a continuous three-dimensional parametric surface. Further, a tangent matching operation is introduced at the edge joints of the lofting calculation to constrain the lofted surface to satisfy first-order geometric continuity with the original internal channel boundary representation surface at the boundary. Finally, by extracting the normal and tangent vectors of the original internal channel boundary representation surface at the splicing point, the tangent plane of the newly generated lofted surface is forced to coincide with the tangent plane of the original surface on the boundary line. This tangential coincidence eliminates sharp angles or creases that may occur at the surface splicing point, ensuring that the fluid channel surface maintains a high standard of geometric smoothness.

[0059] S605 passes the lofted surface containing parametric equations to the underlying 3D geometry calculation engine.

[0060] Specifically, the geometric kernel is invoked to perform boundary trimming and topological stitching on the lofted surface. In some optional implementations, the geometric kernel first calculates the spatial intersection of the lofted surface and the fixed constraint boundary, and removes the redundant parametric surface regions that exceed the constraint boundary. Then, the trimmed boundary edge elements are extracted and subjected to tolerance matching algorithm calculations with the reserved adjacent face edge elements on the original solid model. Further, when the spatial distance between adjacent edges falls within a set small numerical tolerance range, a topological edge merging operation is performed. Next, face nodes and their associated elements that initially posed a manufacturing risk are stripped and deleted from the 3D solid's data tree, and the repairable surface region is replaced with the stitched new surface to obtain the repaired surface. At this point, the geometry of the internal powder discharge channel of the entire solid model has been adjusted and solidified while maintaining no coordinate displacement of the external structure.

[0061] S701, Perform a topology element replacement operation on the repaired surface and the three-dimensional computer-aided design solid model.

[0062] Specifically, in the hierarchical topological data structure of the entity model, the old patch nodes corresponding to the repairable regions originally marked as having high powder removal resistance are precisely located. In some optional implementations, these old patch nodes and their associated bottom-level edges and vertex elements are stripped of data and unregistered from memory, while the newly generated patch nodes corresponding to the repaired surface are attached to the original parent shell element node hierarchy. Further, after the replacement operation is completed, in the local surface reconstruction region and its adjacent space, self-intersecting overlapping surfaces and non-manifold patches generated in the topological element replacement operation are retrieved and removed. By calculating the intersection test of all adjacent or close patches in the local region in three-dimensional space, it is determined whether there are self-intersecting overlapping surfaces where different regions of the same patch intersect spatially, or overlapping surfaces where different patches intersect each other at non-boundary points. At the same time, the number of patches sharing the same common edge element in the local topology is counted. If it is detected that an edge is simultaneously associated with and referenced by three or more independent patches, these regions that cause geometric ambiguity are identified as non-manifold patches. Finally, for the detected self-intersecting overlapping surfaces and non-manifold surfaces, their corresponding topological surface nodes and their dependent association attributes are directly removed from the geometric data tree.

[0063] S702 calls the geometry operation kernel to perform topological stitching on adjacent boundaries of the surface.

[0064] Specifically, the underlying geometric operation kernel reads the outer boundary curves of the newly added repaired surface, as well as the boundary curves of adjacent facets preserved on the original 3D computer-aided design solid model. In some optional implementations, a stitching distance tolerance representing the computational accuracy is set here. This tolerance is preferably between 0.001 and 0.01 mm, and not greater than 1 / 100 of the model's minimum feature size. When the absolute distance between the old and new boundary curves in 3D space is less than this stitching distance tolerance, the geometric operation kernel merges these two originally independent boundary curves, reducing their dimension to a shared common edge element, thereby anchoring and stitching adjacent spatial surfaces together at the topological level. Furthermore, after successful stitching, this embodiment unifies the surface normal vector. The normals of the unreplaced peripheral surfaces of the original solid model are extracted as reference directions. These directions are then transferred to the newly stitched repair surface. If the normals of the new surface and the reference directions reverse at a common edge, the coordinates of the new surface are interchanged or their signs are reversed in its horizontal and vertical definitions in its parameter space, ensuring that all surface normal vectors after stitching point towards the outside of the metal solid material. Finally, for small geometric cracks larger than the stitching distance tolerance caused by model accuracy differences or surface approximation errors, a surface projection extension algorithm is used to close the topological tolerance gap at the intersection. A surface on one side of the crack is selected, and following its local tangential plane and curvature variation trend, the boundary contour is extrapolated and smoothly extended towards the opposite side of the crack within its own two-dimensional parameter space. The spatial intersection curve between the virtual extended surface formed after the extension and the original preserved surface on the opposite side of the crack is calculated. This spatial intersection curve is used as a new and rigorous common boundary. The excess extended surface portion is cut off and discarded in reverse, thereby eliminating the original topological tolerance gap and forming a seamless watertight shell feature.

[0065] S703, extract the reconstructed internal channel centerline, hydraulic diameter, and inner and outer wall thickness mapping dataset.

[0066] Specifically, for the replaced and repaired fluid cavity region, based on the new boundary representation surface, spatial distance field skeletonization calculation is initiated to extract the reconstructed internal channel centerline. An orthogonal normal plane is established along this updated centerline, and a series of cross-sectional areas and perimeters of the new channel are recalculated to calculate the new hydraulic diameter. In some optional implementations, simultaneously, rays are emitted from the normal vector of the newly generated repair surface to the outer surface of the entity to find intersections, and the new local thickness is calculated, thereby updating the generated inner and outer wall thickness mapping dataset. Based on these updated underlying geometric parameters, spatial connectivity detection is performed. Further, starting from one of the fluid port boundary rings, a connectivity roaming test is performed along the reconstructed internal channel centerline to the fluid port boundary ring at the other end of the space to detect whether there are blocking nodes on this roaming path where the hydraulic diameter minimum value approaches 0. Simultaneously, the set allowable minimum process wall thickness is compared with the new inner and outer wall thickness mapping dataset to screen for local wall thickness breakdown or excessively weak points caused by surface inversion. Finally, the spatial roaming Boolean judgment results and the wall thickness extreme value calculation data are combined and packaged to generate spatial connectivity and wall thickness verification state parameters. By verifying that the values ​​of all indicators in these state parameters are within the set forming safety range, it is confirmed that there are no fluid closed blind zones in the reconstruction range and that the fluid connectivity topology has not been broken.

[0067] S704, convert the spatial connectivity and wall thickness verification status parameters into entity metadata attributes.

[0068] Specifically, these entity metadata attributes are constructed in a digital key-value pair format, where the tag key name is defined as a specific manufacturing compliance inspection identifier, and the corresponding tag key value is filled with calculated state parameter Boolean and floating-point values. In some optional implementations, these converted entity metadata attributes are written into the corresponding internal boundary representation surface data structure using the underlying data structure editing interface. That is, information such as verified wall thickness compliance markers and connectivity status unobstructed markers are deeply bound as auxiliary digital tags to the reconstructed internal fluid channel diagram nodes. Finally, the entire 3D model eliminates the risk of powder residue retention in terms of geometry, eliminates all geometric gaps and logical conflicts in terms of topology, and integrates manufacturing inspection imprints in terms of underlying data, thereby forming a continuous and unbroken spatial 3D set at the data structure level, generating the repaired entity model with the closed topology.

[0069] S801, export the repaired solid model with topological closure as a three-dimensional model data format that supports boundary representation of topology.

[0070] Specifically, a 3D model data format that supports boundary representation topology refers to a digital file standard that can contain and record geometric connections and topological hierarchies such as volume elements, shell elements, face elements, ring elements, edge elements, and vertex elements, as well as other elements. This differs from surface mesh formats that only record discrete spatial triangular faces. In some optional implementations, the file writing interface of the 3D geometry operation kernel is invoked to serialize the processed solid geometric tree data structure in memory into a long-term stored binary or text data stream. This format not only preserves the static geometric appearance of the repaired state but also retains the mathematical continuity definition, parametric equations, and spatial adjacency relationships between surfaces. This allows the exported data to be directly read and edited by 3D computer-aided manufacturing software or additive manufacturing equipment slicing preprocessing software in industrial manufacturing enterprises.

[0071] S802, constrain the standardization of the output model data structure, and retain the reconstructed internal channel boundary representation surface topology, inner and outer wall thickness mapping dataset, the construction direction vector and the protected external functional surface.

[0072] Specifically, the standardized operation of this constraint output model data structure refers to establishing independent data blocks and attribute dictionaries according to the set industry standard protocol during the serialization and writing process, without discarding various process verification status tags generated during the detection and repair process. In some optional implementations, a metadata storage area is defined within the data body, and the reconstructed internal channel boundary representation surface topology is written in the serialized form of a node graph, maintaining the original traceability relationship of the connected cavities. Further, the inner and outer wall thickness mapping dataset is converted into attribute key-value pairs attached to the nodes of the specific internal boundary representation surface and packaged together for writing, providing thickness distribution base map attributes for subsequent thermal stress analysis or process planning. The construction direction vector representing the spatial attitude of metal forming is encoded as a global environment variable or entity attribute and saved. Finally, when writing the protected external functional surface, all its surface control point matrices, node vectors, and weight parameters are frozen, and no order reduction, meshing, or approximate fitting operations are performed on it, maintaining its original design-set analytical accuracy and morphological boundaries, thereby conveying a three-dimensional solid geometric asset with a real manufacturable state.

[0073] S803, by comparing the three-dimensional coordinate matrix, it is confirmed that the surface parameterized coordinate system of the protected external functional surface and fluid port boundary ring in the output model maintains the coordinate axis and coincidence with the origin of the imported three-dimensional computer-aided design solid model.

[0074] Specifically, two independent data comparison views are established during readback verification before or after generating the final file data. In some optional implementations, the local spatial transformation matrix of the corresponding face elements in the initially imported 3D computer-aided design solid model is extracted, and the local spatial transformation matrix of the same face elements in the model to be output or already output is also extracted. Further, the translation vector components and rotation quaternion components of these two 3D coordinate matrices are compared, and the numerical difference generated by the comparison is quantized and compared with a minimal floating-point tolerance, which is typically set to the level of 10 to the power of -6 millimeters. Only when the numerical difference between the two matrices in all dimensions of 3D space is less than this minimal floating-point tolerance is it determined that there has been no relative deflection or spatial translation positioning error in the direction of the origin and all three spatial coordinate axes. Finally, this coordinate system comparison mechanism plays a crucial role in ensuring that when thin-walled metal structural components with internal powder discharge risks are successfully cleared and placed back into their original, large mechanical assembly environment, their protected external functional surfaces used for assembly positioning and the fluid port boundary rings used for fluid medium docking can still accurately dock with other surrounding mechanical parts. This eliminates the potential for global coordinate system drift or destructive external geometric deformation induced by internal channel geometric reconstruction operations, and outputs a three-dimensional solid model with high assembly datum consistency.

[0075] like Figure 1 As shown, in this embodiment, the solid model of the thin-walled internal cooling part to be tested is first imported, and the protected external functional surface, internal channel, and fluid port are distinguished in a two-dimensional cross-sectional view. Then, the boundary representation topology attribute map, the center line of the internal channel, the cross-sectional profile, and the mapping relationship between the inner and outer wall thicknesses are superimposed on the same model to form a combined expression of surface nodes, connecting edges, cross-sectional geometry, and wall thickness status. Further, based on the construction direction, gravity powder discharge direction, and the necking, bending, and overhanging positions of the internal channel, the risk distribution of powder discharge obstruction bottleneck is obtained. Then, the topology attribute map, cross-sectional profile, and risk distribution are input into a graph attention network to identify repairable surface regions and corresponding transition boundary loops in the internal channel. Finally, the repairable surface regions are reconstructed by repairing the surface, and through topological stitching, wall thickness constraints, powder discharge unobstructedness, and self-support verification, a repaired solid model with unchanged external functional surfaces and improved powder discharge performance of the internal channel is output.

[0076] like Figure 2As shown, in this embodiment, for the internal fluid channel of the thin-walled internal cooling structure, the channel centerline is extracted along the boundary ring from the left port to the boundary ring from the right port, and cross-sectional profiles are established at different positions of the centerline. At each cross-section, the local wall thickness state is determined by the wall thickness mapping relationship from the inner wall to the outer wall. At the same time, combined with the construction direction and the gravity-assisted powder discharge direction, high-risk bottleneck areas such as channel necking areas, bending areas, and overhanging sides that are prone to powder retention are identified. Furthermore, each boundary surface is abstracted into nodes in the surface topology diagram, and regular surface nodes, high-risk surface nodes, and high-risk critical surface nodes are distinguished. The surface topology diagram and risk characteristics are aggregated and analyzed by an AI network model to extract continuous high-risk inner wall surfaces as repairable regions. At both ends of the repairable region, positions where the risk is reduced and the cross-sectional topology is closed are selected as transition boundary rings.

[0077] like Figure 3 As shown in this embodiment, in the two-dimensional cross-sectional view before repair, there are overhanging side depressions, narrow bottlenecks, and unfavorable powder discharge dead angles in the internal channel, and this area is identified as a repairable surface region, with its two ends defined by the starting transition boundary ring and the ending transition boundary ring, respectively. During the repair process, the two transition boundary rings are used as fixed constraint boundaries, and multiple reconstructed cross-sectional curves are generated between them. The reconstructed cross-sections are smoothly transitioned along the centerline direction, and the minimum gap position is transferred from the overhanging side to the non-overhanging side to form a guiding surface and a repair surface that reduce powder discharge resistance. After repair, the external functional surface remains unchanged as indicated by the dashed line, while the internal channel presents a continuous, smooth shape that is conducive to powder discharge. Subsequently, topological stitching is performed on the boundaries of the new and old surfaces to merge the separated boundaries into shared edges. The continuity of the centerline, the wall thickness qualification, and the channel connectivity are verified to confirm that there are no closed blind areas or wall thickness failures in the repaired area. Finally, metadata such as the repair surface identifier, the starting and ending boundary ring identifier, the minimum gap position, and process parameters are written into the internal surface structure, and the repaired solid model that can be used for subsequent process processing is output.

[0078] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. An AI-driven three-dimensional CAD model manufacturability detection and intelligent repair method applied to a thin-walled inner cooling structural part three-dimensional computer-aided design solid model containing internal channels and protected external functional surfaces, characterized in that, include: Import the 3D computer-aided design solid model to be inspected, extract the boundary representation surfaces, geometric and topological information to construct the boundary representation topological attribute map; The internal channel in the three-dimensional computer-aided design solid model is segmented into cavity space, and the internal channel centerline, channel geometric features, cross-sectional contour set and inner and outer wall thickness mapping dataset are extracted. The risk distribution set of the powder discharge obstruction bottleneck is calculated based on the channel geometric features, and the risk distribution set of the powder discharge obstruction bottleneck is mapped onto the boundary representation surface; The boundary representation topological attribute map, the cross-sectional contour set, and the mapped risk distribution set of powder discharge obstruction bottlenecks are input into a pre-trained deep learning network model to extract repairable surface regions and transition boundary loops on the internal channels. The internal channel surface is reconstructed on the repairable region to generate a repair surface that adjusts the powder discharge resistance while maintaining the geometric features of the protected external functional surface. The repaired surface is topologically stitched and manufacturability verified with the three-dimensional computer-aided design solid model to generate a topologically closed repaired solid model. Output the repaired entity model with boundary representation topological features.

2. The AI-driven 3D CAD model manufacturability detection and intelligent repair method of claim 1, wherein, Import the 3D computer-aided design solid model to be inspected, extract boundary representation surfaces, geometric and topological information to construct a boundary representation topological attribute map, including: The topological structure of the three-dimensional computer-aided design solid model is analyzed, and the volume elements, shell elements, boundary representation surfaces, loop elements, edge elements, and vertex elements are extracted. The surface normal vectors of all boundary representation surfaces are then aligned in a consistent manner. Extract the fluid port boundary ring, and classify and label the boundary representation surfaces corresponding to the internal channels and the external functional surfaces based on the spatial bounding box nesting relationship of the topological elements and the fluid port boundary ring. Construct a relational graph with each boundary surface as a graph node, and record geometric and topological attributes in the graph node. The geometric and topological attributes include surface geometry type, parameter domain boundary features, adjacent edge topological relationship, edge concavity and convexity attributes, surface area, proportion of short edges, and shell identifier. By setting protected identifiers, internal channel identifiers, and repairable identifiers in the graph nodes, the construction of the boundary representation topology attribute graph is completed.

3. The AI-driven 3D CAD model manufacturability detection and intelligent repair method of claim 2, wherein, Extract the internal channel centerline, channel geometric features, cross-sectional profile set, and inner and outer wall thickness mapping dataset, including: A region growth algorithm with the fluid port boundary ring as the starting node is used to extract the fluid connectivity cavity; The topological center of the fluid-connected cavity is extracted to generate the centerline of the internal channel, and a normal plane intersection profile orthogonal to the tangential vector of the centerline is established along the centerline of the internal channel to generate the cross-sectional profile set. For each section in the set of cross-sectional profiles, calculate the cross-sectional area, perimeter, hydraulic diameter, local tangential vector, and path direction vector pointing to the corresponding fluid port boundary ring, and extract the boundary representation surface that intersects with the section; The surface normal vectors along the intersecting boundary representation surfaces are intersected with the outside of the solid, and the matching solid surfaces are calculated to generate the inner and outer wall thickness mapping dataset; Based on the set construction direction vector, calculate the curved surface region on the cross-sectional boundary that satisfies the set included angle threshold and faces the powder support side, and mark it as the overhanging surface region.

4. The AI-driven 3D CAD model manufacturability detection and intelligent repair method of claim 3, wherein, Based on the channel geometric features, a risk distribution set of bottlenecks hindering powder discharge is calculated, and the risk distribution set of bottlenecks hindering powder discharge is mapped onto the boundary representation surface, including: Calculate the local cross-sectional contraction characteristics based on the hydraulic diameter; The anti-gravity powder discharge resistance characteristics are calculated based on the angle between the local tangential vector and the set gravity-assisted powder discharge direction vector. By combining the surface normal vector, the construction direction vector, and the set critical self-supporting angle, the forming defect tendency characteristics of the overhanging surface region are calculated; By comparing the set allowable wall thickness threshold with the inner and outer wall thickness mapping dataset, the wall thickness deviation characteristics are calculated; The channel bending characteristics are calculated by combining the local curvature of the centerline of the internal channel with the hydraulic diameter. The contraction characteristics, resistance characteristics, defect tendency characteristics, flow channel bending characteristics, and wall thickness deviation characteristics are fused into the channel geometric characteristics, and then converted into a cross-sectional comprehensive risk coefficient through a nonlinear mapping function. By combining the spatial distance attenuation function, the cross-sectional level comprehensive risk coefficient is mapped to the corresponding boundary representation surface, generating surface-level risk feature values ​​to construct the risk distribution set of the powder discharge obstruction bottleneck.

5. The AI-driven 3D CAD model manufacturability detection and intelligent repair method of claim 4, wherein, Extracting the repairable surface region and transition boundary loop on the internal channel, including: A multidimensional feature vector is constructed for each boundary representation surface. The multidimensional feature vector includes the surface geometry type, surface curvature, surface area, proportion of short sides, concavity and convexity attributes of adjacent sides, surface-level risk feature value, belonging channel number, and path length parameter from the fluid port boundary loop. A graph attention mechanism is used to perform information transfer and aggregation of multi-dimensional feature vectors between nodes of adjacent boundary representation surfaces with shared edge topological relationships; Extract the continuous channel inner wall surfaces whose post-aggregation risk feature values ​​exceed the set risk threshold, and merge them with the rounded boundary representation surface and the turbulence boundary representation surface that have topological adjacency to form the repairable surface region; Extending along the centerline of the internal channel to the position where the surface-level risk characteristic value drops to within the set risk threshold and the cross-sectional topology is closed, extract the cross-sectional boundary ring at the corresponding position as the start and end cross-sectional boundary ring, and define it as the transition boundary ring.

6. The AI-driven method for manufacturability testing and intelligent repair of 3D CAD models according to claim 5, characterized in that, Perform internal channel surface reconstruction on the repairable region to generate a repair surface that adjusts the powder discharge resistance while maintaining the geometric features of the protected external functional surface, including: Using the starting and ending section boundary rings as fixed constraint boundaries, extract the spatial boundary curves that constitute the closed contour of the corresponding internal channel; An interpolation algorithm is used to generate a set of reconstructed cross-sectional curves between the starting and ending cross-sectional boundary rings. The hydraulic diameter of the reconstructed cross-sectional curves is constrained to change with a gradient along the centerline of the internal channel, and the minimum gap position is constrained to shift from the overhanging surface region to the non-overhanging surface region. The boundary representation surface on the side of the overhanging surface region is topologically reconstructed to generate a guide surface that transitions monotonically along the powder discharge path direction; The reconstructed cross-sectional curve is used as a guide line to generate a lofted surface by non-uniform rational B-spline surface lofting, and the lofted surface is constrained to satisfy first-order geometric continuity with the original internal channel boundary representation surface at the boundary. The geometric kernel is invoked to perform boundary trimming and topology stitching on the lofted surface, and the repairable surface is replaced to obtain the repaired surface.

7. The AI-driven method for manufacturability testing and intelligent repair of 3D CAD models according to claim 6, characterized in that, The repaired surface is topologically stitched and manufacturability verified with the three-dimensional computer-aided design solid model to generate a topologically closed repaired solid model, including: Perform a topology element replacement operation on the repaired surface and the three-dimensional computer-aided design solid model, and retrieve and remove self-intersecting overlapping surfaces and non-manifold patches generated in the topology element replacement operation; The geometric operation kernel is invoked to perform topological stitching on adjacent boundaries of the surface, unify the surface normal vector, and use the surface projection extension algorithm to close the topological tolerance gap at the intersection. Extract the reconstructed internal channel centerline, hydraulic diameter, and inner and outer wall thickness mapping dataset, perform spatial connectivity detection, generate spatial connectivity and wall thickness verification state parameters, and verify that there are no fluid closed blind zones in the reconstructed interval and that the fluid connectivity topology has not been broken. The spatial connectivity and wall thickness verification state parameters are converted into entity metadata attributes and written into the corresponding internal boundary representation surface data structure to generate the repaired entity model with topological closure.

8. The AI-driven method for manufacturability testing and intelligent repair of 3D CAD models according to claim 7, characterized in that, The output includes the repaired entity model with boundary representation topological features, including: The repaired solid model with the topological closure is exported as a three-dimensional model data format that supports boundary representation of topology; Constrain the standardization of the output model data structure, and retain the reconstructed internal channel boundary representation surface topology, inner and outer wall thickness mapping dataset, the construction direction vector and the protected external functional surface; By comparing the three-dimensional coordinate matrices, it was confirmed that the surface parameterized coordinate system of the protected external functional surface and fluid port boundary ring in the output model maintains the coordinate axes and coincides with the origin of the imported three-dimensional computer-aided design solid model.

Images