Parametric axisymmetric structure modeling system and method based on ai decision

Abstract

This invention discloses a parameterized axisymmetric structure modeling system and method based on AI decision-making, belonging to the technical field of graphics data processing and computational modeling for high-end equipment. The system includes a user interaction layer, an AI perception layer, a logic verification layer, a geometric execution layer, and an output layer. The AI ​​perception layer only extracts and outputs structured JSON parameters, without outputting geometric data. The logic verification layer completes physical constraint verification and generates multiple optimization schemes. The geometric execution layer adopts a macroscopic topology plus precision feature toolbox separation architecture, with built-in edge caching mechanism, ray Fabry-Persistent operation engine, NURBS curve units, and Hermite smooth transition units. The output layer controls the normal orientation and exports STEP / STL files. This invention completely decouples AI semantic parsing from geometric calculation, overcoming the millimeter-level accuracy and topological damage defects of existing AI-CAD solutions, achieving micrometer-level IT5 industrial precision. It supports the injection of various precision features, has engineering compliance verification capabilities, is completely autonomous and controllable, and can significantly shorten the design cycle of mechanical parts.

Description

Technical Field

[0001] This invention belongs to the technical field of graphic data processing and computational modeling for high-end equipment, and particularly relates to a parameterized axisymmetric structure modeling system and method based on AI decision-making. Background Technology

[0002] Computer-aided design (CAD) and computer-aided engineering (CAE) are core tools for modern industrial design and intelligent manufacturing. Currently, mainstream commercial CAD systems (such as SolidWorks, CATIA, and NX) rely on parametric axisymmetric structure modeling and feature history trees to complete model construction, and depend on third-party geometric kernels such as Parasolid, ACIS, and OpenCASCADE to achieve complex topological operations and Boolean operations.

[0003] With the development of large-scale modeling technology, AI has begun to intervene in industrial 3D design. Existing AI-assisted CAD solutions fall into three categories: The first is traditional parametric CAD systems, which rely on engineers manually inputting dimensions. These systems cannot parse natural language input, and the rationality of part parameters depends on manual verification, resulting in long design iteration cycles. The second category involves AI directly generating 3D models, outputting CAD scripts from large models or directly outputting STL or OBJ mesh models. This type of solution only achieves millimeter-level accuracy, and the generated models lack a complete BRep (boundary representation) topology, exhibiting topological defects such as overlapping faces, non-manifold edges, and normal flipping. It also cannot perform Boolean operations or engineering simulations. The third category is hybrid AI-CAD systems, where AI extracts parameters and calls preset templates to generate models. This method only supports simple preset parts, has poor scalability, lacks engineering constraint verification, and is prone to surface breakage and geometric self-intersection when modeling complex and precise features such as splines, internal gears, threads, and undercuts.

[0004] As can be seen from the above, the main shortcomings of existing technologies include: the random generation characteristics of AI contradict the determinism and micron-level precision required for industrial modeling; AI directly participates in geometric generation, bringing semantic illusions into geometric calculations; there is a lack of a separate processing architecture for macroscopic topology and microscopic precision features; most rely on third-party geometric kernels, which poses licensing risks; and there is a lack of engineering constraint verification after AI parameter extraction and before geometric modeling, which only triggers modeling when the parameters are unreasonable, resulting in a waste of computing resources. Summary of the Invention

[0005] The purpose of this invention is to provide a parametric axisymmetric structure modeling scheme that completely decouples AI semantic understanding from geometric computation, while also possessing high-precision modeling, complete BRep topology, engineering parameter verification, scalable precision features, and an autonomous and controllable geometric kernel. This invention is achieved through the following technical solution:

[0006] A first aspect of the present invention provides a parameterized axisymmetric structure modeling system based on AI decision-making, comprising:

[0007] The user interaction layer is used to receive user input information, the input information being in the form of at least one of natural language description, sketch, and parameter table;

[0008] The AI ​​perception layer is used to call a multimodal large model to perform intent recognition, part type recognition, geometric parameters, and physical condition parameter extraction on the input information, and output structured JSON parameters that do not contain geometric data.

[0009] The logic verification layer includes a physical constraint verification module and a multi-scheme optimization module. The physical constraint verification module receives structured JSON parameters output by the AI ​​perception layer, performs root cut verification, strength verification, transmission ratio matching, and national standard compliance verification, intercepts unreasonable parameters, and outputs modification suggestions to the user interaction layer for the user to modify. The multi-scheme optimization module generates at least two optional design schemes based on the verified parameters, including a standard scheme, a lightweight scheme, and a high-strength scheme, and pushes scheme selection information to the user interaction layer for the user to choose from.

[0010] The geometric execution layer includes the BRep module, the precision feature toolbox, and the Boolean operation engine. The BRep module has a built-in edge caching mechanism, which uses a hash table to cache vertex and edge objects and standardizes geometric parameters to generate hash keys, thereby constructing the macroscopic topological skeleton of the part. The precision feature toolbox encapsulates a variety of atomic-level precision feature algorithms and injects microscopic precision features at specified positions in the macroscopic topological skeleton through a callback mechanism. The Boolean operation engine uses a hybrid algorithm of edge classification and ray casting to complete entity union, intersection, and difference operations. Rays use non-principal axis direction vectors to assemble and generate complete BRep entities.

[0011] The output layer controls the orientation of the face normals of the BRep entity and outputs STEP and / or STL format files.

[0012] As a preferred technical solution, the Boolean operation engine specifically includes:

[0013] Edge classification unit: For each edge of the entity participating in the operation, take the midpoint of the edge and emit a ray, and count the number of intersections; odd number of intersections determine that the edge is inside the entity, and even number of intersections determine that the edge is outside the entity;

[0014] Face classification unit: classify all edges associated with a face by statistical analysis; if both internal and external labels exist, the face is marked as a cross-face; for the remaining faces, the label with the highest percentage of edges is used as the dominant face classification.

[0015] Face selection unit: Performs face selection based on Boolean operation type and face category;

[0016] Topology collection unit: collects the retained faces, associated edges, and vertices, and assembles them to generate a complete BRep entity.

[0017] As a preferred technical solution, the AI-based parameterized axisymmetric structure modeling system further includes a finite element simulation module that interfaces with the logic verification layer. The logic verification layer receives the finite element simulation feedback results from the finite element simulation module and iteratively optimizes the parameters until a preset safety factor is met.

[0018] As a preferred technical solution, the AI ​​perception layer includes a prompt word engineering unit and a RAG retrieval enhancement unit; the prompt word engineering unit limits the multimodal large model to only output structured JSON parameters, and the RAG retrieval enhancement unit assists in parameter extraction by retrieving knowledge information from an external mechanical design knowledge base.

[0019] As a preferred technical solution, the atomic-level precision feature algorithm includes keyway cutting, thread generation, spline cutting, undercut groove, and cam contour algorithms; when adding a new atomic-level precision feature algorithm, only the algorithm function is added to the precision feature toolbox.

[0020] As a preferred technical solution, the geometry execution layer further includes a NURBS curve generation unit, which generates discrete points and constructs cubic NURBS curves based on mathematical equations for the contours of gear involutes and bearing raceways.

[0021] As a preferred technical solution, the geometric execution layer further includes a smooth transition unit, which uses cubic Hermite interpolation at the boundaries of features including keyways and relief grooves to generate a C1 continuous transition region at the feature boundaries.

[0022] A second aspect of the present invention provides a parameterized axisymmetric structure modeling method based on AI decision-making, applied to the system described above, comprising the following steps:

[0023] S100: The user interaction layer receives input information from the user in at least one form, including natural language, sketches, and parameter tables;

[0024] S200: The AI ​​perception layer parses user input information, identifies part type, geometric dimensions, material, and working condition constraints, and outputs structured JSON parameters that do not contain geometric data;

[0025] S300: The logic verification layer performs engineering constraint verification. If the structured JSON parameters do not meet the design standards, it outputs modification suggestions to the user interaction layer for the user to modify. After the verification is successful, multiple candidate solutions are generated for the user to choose from.

[0026] S400: After the user selects a scheme, the geometry execution layer performs deterministic modeling.

[0027] (1) The BRep module enables the edge caching mechanism based on the target scheme parameters to ensure the uniqueness of vertex and edge objects and construct the macroscopic topological skeleton of the parts;

[0028] (2) Call back the precision feature toolbox to inject micro-precision features at the specified position of the macro topological skeleton;

[0029] (3) Call the Boolean operation engine to complete the entity Boolean operation and obtain the complete BRep entity;

[0030] S500: Output layer controls the orientation of the face normals of the BRep entity, and exports STEP and / or STL files.

[0031] As a preferred technical solution, the Boolean operation engine performs the following steps:

[0032] Edge classification: For each edge of the entity participating in the operation, take the midpoint of the edge and emit a ray, and count the number of intersections; odd number of intersections determine that the edge is inside the entity, and even number of intersections determine that the edge is outside the entity;

[0033] Face classification: Classify all edges associated with a face by their classification labels; if both internal and external labels exist, the face is labeled as a cross-face; for the remaining faces, the label with the highest percentage of edges is used as the dominant face classification.

[0034] Face selection: Face selection is completed based on Boolean operation type and face classification; for cross-faces, the center point ray method of the sampled face is used to determine whether the face is inside or outside; during the difference operation, the normals of the faces of the subtrahend entity are flipped;

[0035] Topology collection: Collect the retained faces, associated edges, and vertices, and assemble them to generate a complete BRep entity.

[0036] As a preferred technical solution, the parameterized axisymmetric structure modeling method based on AI decision-making further includes a simulation closed-loop iteration step: sending the generated BRep entity into the finite element simulation module to obtain simulation results; sending the simulation safety factor and stress results back to the logic verification layer, and the logic verification layer automatically adjusts the geometric parameters and re-executes the modeling process until the simulation indicators meet the preset safety factor.

[0037] This invention provides a parameterized axisymmetric structure modeling system and method based on AI decision-making. The AI ​​only outputs structured parameters and does not participate in geometric calculations, eliminating the interference of large model illusions on geometric entities. Modeling accuracy is improved from millimeter-level to micrometer-level by traditional AI-generated models. An independent logic verification layer is set up to complete engineering constraint verification. A separate architecture of macroscopic topology skeleton and precision feature toolbox is adopted. The BRep module is responsible for overall topology construction, and the toolbox injects local precision features through callbacks. The BRep module kernel introduces an edge caching mechanism to achieve unique instance storage of shared edges, solving the defects of broken shared edges and topology damage in traditional BRep modeling, reducing memory usage, and optimizing STEP export quality. The precision feature toolbox supports various complex precision features, has strong extensibility, and adding new features only extends the toolbox functions without modifying the main engine core code. Boolean operations and an edge caching topology maintenance mechanism are used to achieve topological integrity, improving the automation level of modeling and shortening the part design cycle. A Boolean operation engine using "edge classification + non-principal axis ray method" is adopted, eliminating reliance on third-party geometry libraries such as OpenCASCADE, achieving complete system autonomy and control. The intersection failure problem caused by parallel rays and planes is avoided by using non-principal axis rays. Attached Figure Description

[0038] Figure 1 The diagram shows the architecture of a parameterized axisymmetric structure modeling system based on AI decision-making, which is provided for a specific embodiment of the present invention.

[0039] Figure 2 A flowchart of a parameterized axisymmetric structure modeling method based on AI decision-making provided for a specific embodiment of the present invention.

[0040] Figure 3 This is a bar chart showing the end-to-end full-link performance of five test scenario examples in a specific embodiment of the present invention.

[0041] Figure 4 This is a comparison chart of workflow efficiency between a specific example of the present invention and a traditional CAD+CAE solution.

[0042] Figure 5 This is a performance comparison chart of basic geometry creation using a specific example of the present invention and the OpenCASCADE method.

[0043] Figure 6 The bar chart shows a comparison of Boolean operation performance between the present invention and the OpenCASCADE method, using a specific example.

[0044] Figure 7 This is a radar comparison chart of the geometric kernel capability dimensions of the OpenCASCADE and ACIS / Parasolid schemes, serving as a specific example of the present invention.

[0045] Figure 8This is a comparison of the mesh quality of the tooth root transition section and the FEA mesh convergence curve in a specific example of the present invention.

[0046] Figure 9 This is a bar chart showing the automatic verification coverage of national standards in five areas: steel structure, pressure vessel, gear, key connection, and general machinery, as specific examples of this invention.

[0047] Figure 10 The quality comparison of the rule engine plus multimodal large model hybrid diagnostic scheme and the pure multimodal large model scheme, which are specific examples of the present invention, is shown in the corresponding AI diagnostic quality multidimensional radar comparison chart.

[0048] Figure 11 This is a bar chart comparing the time taken to export a STEP format file with the file size, as a specific example of the present invention. Detailed Implementation

[0049] To make the technical solution of the present invention clearer and its technical advantages more apparent, the technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of the present invention.

[0050] It should be noted that although the steps in the flowcharts of this application's embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0051] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0052] This invention innovatively decouples AI semantic understanding from geometric computation, employing a three-layer architecture of "AI perception layer + logic verification layer + geometric execution layer." This leverages AI's powerful natural language understanding capabilities while ensuring industrial-grade accuracy and determinism in geometric modeling. To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be further described in detail below with reference to the accompanying drawings.

[0053] See Figure 1 As shown in the figure, this embodiment of a parameterized axisymmetric structure modeling system based on AI decision-making includes: a user interaction layer, an AI perception layer, a logic verification layer, a geometric execution layer, and an output layer.

[0054] The user interaction layer receives user input information, which may take the form of at least one of natural language descriptions, sketches, or parameter tables. The AI ​​perception layer connects to the user interaction layer and utilizes the powerful semantic understanding capabilities of multimodal large models (such as Hunyuan and GPT-4o) to perform intent recognition, part type recognition, geometric parameters, and physical condition parameter extraction on the input information. This enables automatic extraction from descriptions in natural language and other forms to structured parameters, and outputs structured JSON parameters that do not contain geometric data.

[0055] In this process, parameter extraction does not rely on the "native" output of the large model. Instead, the core parameters are deterministically extracted by the rule engine (keyword mapping + regular expression + unit conversion). The large model performs semantic enhancement only under six constraints (few sample hints / low temperature / schema enumeration whitelist / failure degradation). The logic verification layer provides fallback verification. The three layers work together to transform the probabilistic output into deterministic parameters. In other words, the large model outputs the required structured JSON parameters based on the data in the library and the user's semantic understanding.

[0056] All extracted parameters are uniformly converted to standard JSON format, with numerical units standardized to millimeters and Newtons, ensuring direct parsing and use downstream. Furthermore, the output of the AI ​​perception layer contains no geometric information, only structured parameter definitions, thus completely isolating the uncertainty of AI from the deterministic requirements of geometric computation. This results in a natural language parsing accuracy exceeding 95%, parameter extraction time less than 2 seconds, and support for multimodal input (text descriptions, sketch annotations, parameter tables). Moreover, the AI ​​and geometry engine are completely decoupled, guaranteeing the determinism of the modeling results; the accuracy is ensured by a combination of "100% rule engine + schema whitelist verification + validation layer rejecting unreasonable parameters."

[0057] Considering the potential engineering inconsistencies in the parameters extracted by AI (such as insufficient gear teeth leading to undercut, or insufficient shaft diameter to withstand load), this embodiment adds a logic verification layer between AI parameter extraction and geometric execution to ensure the engineering rationality of the generated model. The logic verification layer includes a physical constraint verification module and a multi-scheme optimization module. The physical constraint verification module receives structured JSON parameters output by the AI ​​perception layer and verifies the rationality of the extracted parameters based on the part type and engineering domain knowledge. Specifically, it performs undercut verification, strength check, transmission ratio matching, and national standard compliance verification, intercepting unreasonable parameters and outputting modification suggestions to the user interaction layer for user modification. For example, for gear parts, the system verifies whether the number of teeth is greater than the minimum number of teeth to avoid undercut, calculates whether the bending stress is less than the allowable stress of the material, and verifies whether the transmission ratio meets the design requirements. If the parameters do not meet the constraints, the system provides modification suggestions. In addition, the multi-scheme optimization module generates at least two optional design schemes based on the verified parameters, including a standard scheme, a lightweight scheme, and a high-strength scheme, and pushes scheme selection information to the user interaction layer for user selection. The standard solution meets all engineering constraints; the lightweight solution reduces material usage while meeting strength requirements; and the high-strength solution improves the safety factor. Users can choose the optimal solution based on their actual needs. As can be seen, the configured logic verification layer can automatically intercept unreasonable designs and avoid invalid modeling; it can achieve multi-solution comparison and optimization to improve design quality; and it can also solidify engineering design experience into the system, reducing reliance on engineers' experience.

[0058] The geometry execution layer includes a BRep module, a precision feature toolbox, and a Boolean operation engine. The BRep module has a built-in edge caching mechanism, using a hash table to cache vertex and edge objects, and standardizing geometric parameters to generate hash keys, thereby constructing the macroscopic topological skeleton of the part. The precision feature toolbox encapsulates various atomic-level precision feature algorithms, injecting microscopic precision features at specified positions on the macroscopic topological skeleton through a callback mechanism. The Boolean operation engine uses a hybrid algorithm of edge classification and ray casting to perform entity union, intersection, and difference operations. Rays use non-principal axis direction vectors to assemble and generate complete BRep entities. The output layer controls the orientation of the face normals of the BRep entities and outputs STEP and / or STL format files.

[0059] The BRep module described above adopts a boundary representation topology structure, comprising four levels: faces, loops, edges, and vertices. Faces define geometric surfaces (planes, cylinders, cones, etc.) and boundary loops; loops are divided into outer loops and inner loops (holes), containing directed edge sequences; edges define curved geometry (straight lines, arcs, etc.) and start and end vertices; vertices define 3D coordinates. In this embodiment, the edge caching mechanism addresses the common issue in BRep modeling where the same geometric edge is often shared by multiple faces (e.g., the seam edge on the side of a cylinder). This embodiment uses a hash table to cache edge and vertex objects, ensuring that edges with the same geometric parameters have only one instance in memory. This maintains topological integrity (avoiding broken shared edges), reduces memory usage, and avoids duplicate geometric definitions when exporting output files.

[0060] Furthermore, traditional CAD systems often use the same algorithm to process both the overall topology and local features when handling complex parts, leading to high computational complexity and a high risk of errors. This embodiment innovatively separates precision feature processing into a toolbox module, providing over twenty atomic-level geometric feature algorithms, including keyway cutting, thread generation, spline cutting, undercutting, eccentric machining, cam contouring, and spoke hollowing. The BRep module's macro-topology generator handles the overall rotation / extension structure, calling the toolbox algorithm for radius correction at locations requiring precision features. For example, when generating a stepped shaft with a keyway, the basic stepped shaft topology is generated first, and then the keyway cutting algorithm is called at a specified angle to correct the radius. This architecture ensures both the integrity of the overall topology and precise control of local features.

[0061] In summary, the aforementioned geometry execution layer follows a design philosophy of macroscopic dependence on topology and microscopic dependence on toolboxes, separating overall topology construction from local precision feature processing. This layer does not rely on third-party geometry libraries (such as OpenCASCADE), ensuring the system's autonomy and controllability, as well as its optimization capabilities for precision parts. Furthermore, adding atomic-level precision feature algorithms only requires extending toolbox functions, without modifying the BRep module's main engine code, demonstrating strong extensibility.

[0062] Considering that existing open-source solutions (such as OpenCASCADE) are powerful but bulky and difficult to customize, this embodiment uses a Boolean operation engine that employs a hybrid algorithm of "edge classification + ray casting" to maintain code simplicity and controllability while ensuring functional integrity. Therefore, as an optional implementation, the Boolean operation engine includes:

[0063] Edge Classification Unit: For each edge of the participating entities, emit a ray from the midpoint of the edge and count the number of intersections. An odd number of intersections indicates the edge is inside the entity, and an even number indicates the edge is outside the entity. For example, for each edge of entity A, determine its positional relationship relative to entity B: First, check if the midpoint of the edge is on the surface of entity B (distance less than the geometric tolerance value). If so, mark it as "on the surface"; otherwise, emit a ray from the midpoint of the edge and count the number of intersections with the surface of entity B. An odd number indicates the edge is inside entity B, and an even number indicates it is outside. Perform the same logic for the edges of entity B.

[0064] Face classification unit: The classification labels of all edges associated with a face are statistically analyzed; if both internal and external labels exist, the face is labeled as a crossing face; for other faces, the label with the highest percentage of edges is taken as the dominant face classification. Specifically, if the edge classifications are mixed (both internal and external), the face is a "crossing face" (partially internal and partially external); otherwise, the classification with the highest percentage of edges is taken as the dominant face classification.

[0065] Face selection unit: Performs face selection based on Boolean operation type and face classification. Specifically, for cross-cutting faces, the center point of the sampled face is determined using the ray method to determine whether it is inside or outside; for non-cross-cutting faces, whether to retain them is determined based on the Boolean operation type (union / intersection / difference) and the dominant classification; for difference operations, faces from the subtrahend entity need to have their normal direction flipped.

[0066] Topology Collection Unit: Collects the retained faces along with their associated edges and vertices, assembling them to generate a complete BRep entity. Specifically, it adds the retained faces and their associated edges and vertices to the resulting entity, ensuring the integrity of the topological reference relationships.

[0067] The key to the aforementioned computational engine lies in: emitting a ray from the point to be judged, counting the number of intersections with the solid surface, with odd numbers indicating the point is inside the solid and even numbers indicating the point is outside. The crucial technique is the selection of the ray direction: using a non-principal axis direction vector (e.g., (1.0, 0.001, 0.002)) to avoid parallelism with the geometric surface, which would lead to intersection failure. Furthermore, when a ray intersects a geometric surface, for a plane, the dot product of the ray direction and the plane normal is calculated to determine if it is parallel; if not, the intersection parameters are calculated. For a cylindrical surface, the ray and the cylinder axis are projected onto a plane perpendicular to the axis, transforming it into a two-dimensional circle and line intersection problem, solving a quadratic equation to obtain the intersection point. Moreover, when judging a region, after the ray intersects the surface, it is necessary to determine if the intersection point is within the surface's boundary. For a planar surface, the intersection point is projected onto the surface's local UV coordinate system, and the ray method is used to determine if it is inside the ring; for a cylindrical surface, the axial position of the intersection point is determined to be within the surface's axial range.

[0068] As an optional specific implementation, the AI-based parameterized axisymmetric structure modeling system described above also includes a finite element simulation module. This finite element simulation module interfaces with the logic verification layer, receives finite element simulation feedback results, and iteratively optimizes parameters until a preset safety factor is met.

[0069] As an optional implementation, the AI ​​perception layer described above includes a prompt word engineering unit and a RAG retrieval enhancement unit. The prompt word engineering unit limits the multimodal large model to output only structured JSON parameters. Specifically, through the prompt word engineering unit, the large model is guided to identify part types (gears / shafts / bearings / flanges, etc.), extract geometric parameters (module, number of teeth, diameter, length, etc.), resolve physical constraints (material, load, speed, etc.), and determine accuracy requirements (tolerance grade, surface roughness, etc.) from user input information. The RAG retrieval enhancement unit retrieves mechanical design knowledge bases to assist in parameter extraction.

[0070] As a further preferred implementation, the geometry execution layer described above also includes a NURBS curve generation unit. This unit generates discrete points and constructs cubic NURBS curves based on mathematical equations for contours including gear involutes and bearing raceways. Specifically, for complex contours such as gear involutes and bearing raceways, this implementation uses NURBS (Non-Uniform Rational B-Spline) curves for precise mathematical definition. NURBS curves have advantages such as strong local controllability, accurate representation of conic sections, and good computational stability. The system generates high-precision discrete points (typically 50 sampling points) based on the mathematical equations of the involute and then constructs cubic NURBS curves. Compared with traditional discrete point approximation methods, NURBS curves can accurately express tooth profiles with an accuracy of IT5, meeting the machining requirements of high-precision gears.

[0071] As a further preferred embodiment, the geometry execution layer described above also includes a smooth transition unit; the smooth transition unit uses cubic Hermite interpolation at the boundaries of features including keyways and relief grooves to generate a C1 continuous transition region at the feature boundaries. Specifically, at the boundaries of features such as keyways and relief grooves, abrupt radius changes can lead to a decrease in mesh quality and visual discontinuities. This invention sets a transition region at the feature boundaries and uses a cubic Hermite interpolation function to calculate the transition radius. The first derivative of this interpolation function is zero at the boundaries, achieving C1 continuity (continuous first derivative), which avoids visual abrupt changes, ensures mesh quality, and is computationally simple and suitable for real-time modeling.

[0072] As an optional implementation, the output layer described above outputs STEP and STL format files in the following specific way:

[0073] STEP format file export: The export process converts the BRep topology (vertices, edges, faces) into STEP entity definitions (CARTESIAN_POINT, EDGE_CURVE, ADVANCED_FACE, etc.) while maintaining topological references; it also converts geometric definitions (planes, cylindrical surfaces, conical surfaces, arcs, lines, etc.) into corresponding STEP geometric entities. The key lies in normal orientation control: the same_sense attribute of ADVANCED_FACE is used to control the normal direction of the face, ensuring correct entity recognition when imported into other CAD systems.

[0074] STL format file export: The export process triangulates each face of BRep (using a fan-shaped triangulation algorithm), calculates the normal and vertex coordinates of each triangle, and outputs it in Binary STL format (80-byte header + 4-byte number of triangles + 50 bytes per triangle).

[0075] Combination Figure 2 As shown, this embodiment of the invention also provides a parameterized axisymmetric structure modeling method based on AI decision-making, applied to the above-described parameterized axisymmetric structure modeling system based on AI decision-making, including the following steps:

[0076] S100: The user interaction layer receives input information from the user in at least one form, including natural language, sketches, and parameter tables;

[0077] S200: The AI ​​perception layer parses user input information, identifies part type, geometric dimensions, material, and working condition constraints, and outputs output information containing only structured JSON parameters;

[0078] S300: The logic verification layer performs engineering constraint verification. If the structured JSON parameters do not meet the design standards, it outputs modification suggestions to the user interaction layer for the user to modify. After the verification is successful, it generates multiple candidate solutions for the user to choose from and receives the target solution selected by the user.

[0079] S400: After the user selects a scheme, the geometry execution layer performs deterministic modeling.

[0080] (1) The BRep module enables the edge caching mechanism based on the target scheme parameters to ensure the uniqueness of vertex and edge objects and construct the macroscopic topological skeleton of the parts;

[0081] (2) Call back the precision feature toolbox to inject micro-precision features at the specified position of the macro topological skeleton;

[0082] (3) Call the Boolean operation engine to complete the entity Boolean operation and obtain the complete BRep entity;

[0083] S500: Output layer controls the orientation of the face normals of the BRep entity, and exports STEP and / or STL files.

[0084] The Boolean operation engine performs the following steps:

[0085] Edge classification: For each edge of the entity participating in the operation, take the midpoint of the edge and emit a ray, and count the number of intersections; odd number of intersections determine that the edge is inside the entity, and even number of intersections determine that the edge is outside the entity;

[0086] Face classification: Classify all edges associated with a face by their classification labels; if both internal and external labels exist, the face is labeled as a cross-face; for the remaining faces, the label with the highest percentage of edges is used as the dominant face classification.

[0087] Face selection: Face selection is completed based on Boolean operation type and face classification; for cross-faces, the center point ray method of the sampled face is used to determine whether the face is inside or outside; during the difference operation, the normals of the faces of the subtrahend entity are flipped;

[0088] Topology collection: Collect the retained faces, associated edges, and vertices, and assemble them to generate a complete BRep entity.

[0089] Furthermore, the AI-based parameterized axisymmetric structure modeling method provided in the above embodiments also includes a simulation closed-loop iteration step: the generated BRep entity is sent to the finite element simulation module to obtain simulation results; the simulation safety factor and stress results are sent back to the logic verification layer, the logic verification layer automatically adjusts the geometric parameters, and re-executes the modeling process until the simulation indicators meet the preset safety factor.

[0090] In summary, the main innovations and beneficial effects of the AI-based parameterized axisymmetric structure modeling system and method provided in the above embodiments of the present invention include:

[0091] 1. A layered architecture that completely decouples AI from geometry. Existing AI-aided design systems attempt to allow AI to directly generate geometric models or scripts, but the uncertainty of AI leads to unreliable accuracy. This invention innovatively confines AI to the parameter extraction level, while geometric calculations are performed entirely by a deterministic algorithm engine. This leverages AI's semantic understanding advantages while ensuring industrial-grade accuracy. Accuracy is improved from millimeter-level to micrometer-level, a 1000-fold increase.

[0092] Second: A modeling architecture that separates macroscopic topology from microscopic features. Traditional CAD systems use the same algorithm to process both overall topology and local features, resulting in complex calculations and a high risk of errors. This invention separates the processing of precise features into a toolbox module, with the main engine handling only the macroscopic topology and calling the toolbox algorithm where precise features are needed. This architecture supports multiple types of precise features, and adding new features only requires expanding the toolbox, making its scalability far exceed that of traditional solutions.

[0093] Thirdly, the innovative BRep module and Boolean operation engine. Existing solutions rely on third-party geometry libraries (such as OpenCASCADE), which are bulky and difficult to customize. This invention develops a lightweight BRep module and Boolean operation engine, employs an edge caching mechanism to ensure topological integrity, and uses ray casting to achieve accurate internal / external judgment. It has no external dependencies, and the system is completely autonomous and controllable.

[0094] Fourthly, the addition of NURBS precision curve elements and smooth transition processing elements. For complex contours such as gear involutes, NURBS curves are used for precise representation, achieving an accuracy of IT5 level. Hermite interpolation is used at feature boundaries to achieve continuous C1 transitions, avoiding mesh quality degradation caused by abrupt radius changes, resulting in a mesh quality improvement of over 60%.

[0095] 5. Engineering Constraint Verification and Multi-Solution Optimization. A physical constraint verification layer is added after AI parameter extraction to automatically intercept unreasonable designs (such as root cuts, insufficient strength, etc.) and generate multiple optimization schemes for users to choose from. This solidifies engineering design experience into the system, reducing reliance on engineers' experience.

[0096] Compared with traditional CAD systems and existing AI-assisted systems, the relevant quantitative indicators of the technical effects of this invention are shown in Table 1 below:

[0097]

[0098] Table 1

[0099] The technical solution, application scenarios, and beneficial effects of the present invention will be further illustrated below with specific examples:

[0100] I. Architecture Process and Simple Example:

[0101] The explanation is illustrated by a specific instruction: "Design a standard spur gear with a module of 3, 20 teeth, a pressure angle of 20 degrees, a tooth width of 30 mm, made of 45 steel with heat treatment, a power transmission of 5 kW, a speed of 1500 rpm, a center hole diameter of 25 mm, a fit tolerance of H7, and a safety factor of not less than 1.5."

[0102] The complete execution flow of this example is as follows:

[0103] At the perception level: the requirement analysis module of the multimodal large model identifies the design type as "spur gear". The artificial intelligence analysis module loads preset prompts (positioning itself as a mechanical design expert, prohibiting the output of geometric coordinates) and sends them to the multimodal large model along with the user input, returning structured parameters "module 3.0 mm, number of teeth 20, pressure angle 20 degrees, tooth width 30 mm, safety factor 1.5, center bore diameter 25 mm, accuracy level 7". The knowledge base module queries and returns the material properties of 45 steel (yield strength 355 MPa) and applicable national standards. This layer embodies the decoupling concept of the first key point: the output object does not contain any geometric information, only providing structured input for the subsequent verification layer; at the same time, it provides the source of national standard clauses for subsequent constraint verification.

[0104] Verification Layer: After receiving parameters from the perception layer, the engine sequentially performs undercut checks (minimum number of teeth without undercut is 17.1, actual number is 20), tooth surface contact strength checks (stress 365.4 MPa is less than the allowable 550 MPa), and tooth root bending strength checks (44.8 MPa is much less than the allowable 220 MPa). These three checks are the core functions of the constraint verification engine. Then, using module, number of teeth, and tooth width as variables, a mesh search is performed for optimization, generating three comparative schemes for the user to choose from: a standard scheme (weight 1.08 kg), a lightweight scheme (0.71 kg, weight reduction 34%), and a reinforced scheme (safety factor 2.10). This layer demonstrates the multi-scheme optimization capability.

[0105] Execution layer: Receives verified pure numerical parameters and routes them to the gear builder. This layer embodies the following characteristics: layered modeling first constructs the tooth tip cylindrical skeleton and then injects the involute tooth profile tooth by tooth; edge hashing ensures that there are no repeated vertices or edges throughout the topology construction; Boolean operations complete the subtraction of the center hole solid; B-spline curve fitting of the involute and Hermite interpolation achieve a smooth transition at the tooth root; standard format export completes the output with consistent normals.

[0106] From this point onward, artificial intelligence no longer participates in any calculations.

[0107] II. Overall Architecture: A three-layer system decoupling artificial intelligence and geometry.

[0108] The system adopts a three-layer decoupled architecture of "artificial intelligence perception layer, logic verification layer, and geometric execution layer". All interfaces between the three layers are structured numerical data. The perception layer outputs a set of parameters in the form of key-value pairs, the verification layer outputs the verification results and optimization suggestions, and the execution layer drives the entire process of geometric construction solely with numerical values ​​without passing any geometric objects, thus fundamentally cutting off the path for the illusion of artificial intelligence to penetrate into the geometric layer.

[0109] To verify the overall performance of the architecture, end-to-end full-link experiments were conducted on five typical design scenarios: "drive shaft", "cantilever bracket", "thick plate", "support column", and "perforated bracket". See Table 2 for details. Figure 3 :

[0110]

[0111] Table 2

[0112] in, Figure 3 The vertical axis in the figure is a segmented, non-uniform scale (segmented scale: 0ms→1ms→10ms→15ms→50ms→75ms→100ms→125ms→150ms→175ms), which allows the blue bars with a time of less than 1ms in the modeling stage to be enlarged and displayed at the bottom; in the figure, "total time approximately 60ms" and "total time approximately 177ms" represent the total time spent on "modeling", "simulation" and "export"; and the "approximately" means slightly greater than but not greater than 1ms.

[0113] Furthermore, comparative experiments with traditional CAD+CAE workflows show that this system reduces a single design evaluation from hours to milliseconds (see Table 3). Figure 4 :

[0114]

[0115] Table 3

[0116] The term "millisecond level" here refers to the end-to-end time of 15-300ms for sub-millisecond extraction and geometry / simulation / export by the rule engine. The time for large model network inference is in the second level, which has been measured separately and is not included in the total time of millisecond level.

[0117] III. Layered Modeling: Separation and processing of macroscopic skeleton and microscopic precision features.

[0118] In this example, the modeling of the geometry execution layer adopts a layered approach of "building the skeleton first, then adding details".

[0119] Macro layer: Construct a cylinder with a tooth tip circle radius of 33 mm and a height of 30 mm as the topological skeleton, containing only 3 faces and 18 edges.

[0120] The micro-layer is injected in four steps: First, the pitch circle radius is calculated to be 30 mm, the base circle radius is 28.19 mm, and the root circle radius is 26.25 mm. Involute tooth profile curves are generated tooth by tooth and stretched to form 40 tooth surfaces.

[0121] The second step is to apply a fillet with a radius of 1.14 mm at the intersection of the tooth root and the tooth surface.

[0122] The third step is to apply a 0.5 mm chamfer to the tooth tip.

[0123] The fourth step involves creating a cylinder with a radius of 12.5 mm and subtracting the center hole using Boolean difference. When adding a new feature type, only the processing function needs to be added and the injection location specified; no modification to the macro layer is required.

[0124] The performance of this architecture was verified through experiments creating eight basic geometries, as shown in Table 4:

[0125]

[0126] Table 4

[0127] Experimental testing revealed a 91% accuracy rate in face count calculation, with an average creation time of 88 microseconds. Figure 5 This is a performance comparison chart of the basic geometry creation method of the present invention compared with the OpenCASCADE 7.7 (C++) method.

[0128] IV. Edge hashing for deduplication: the infrastructure for ensuring topological uniqueness.

[0129] The BRep module maintains vertex and edge hash tables. When creating a vertex, the coordinates are quantized to integers with a tolerance of one millionth of a millimeter to generate a key; when creating an edge, the endpoints are standardized, sorted, and then concatenated with the curve content to generate a key. If the query cache is hit, an existing instance is returned.

[0130] Taking the example of the involute face of the first tooth and the top face of the gear sharing an edge: when the top face first creates this edge, it is stored in the cache; when the involute face is called again, the endpoints are passed in the reverse order, but after normalization and sorting, the same hash key is generated, hitting the cache, and the two faces share the same edge object. In this example, the vertex deduplication rate is 75%, the edge deduplication rate is 65%, and on average, each edge is shared by 2.87 faces. This mechanism provides the underlying support for Boolean operation stitching and pairing and deduplication of entities exported in standard format.

[0131] Experimental tests revealed that all pure basic shells and simple Boolean results achieved closure, the ratio of unique edges to faces conformed to the expectations of B-spline surface theory, and there were no repeated edges defined.

[0132] V. Boolean Operation Engine: Four-stage pipeline, the creation of the center hole (gear entity minus hole cylinder) is completed in four stages.

[0133] 1. Intersection stage: The intersection register distributes dedicated intersection units according to the "plane-cylindrical surface" combination, resulting in two circular intersection lines with a radius of 12.5 mm and a two-dimensional projection of the surface parameter domain.

[0134] 2. Segmentation stage: Add a circular inner ring to the top and bottom surfaces of the gear using projection.

[0135] 3. Classification and selection stage: Take the sampling points of the surface and emit rays along three non-axial directions. Odd-numbered intersections are judged as internal and even-numbered intersections are judged as external. The majority vote is taken in the three directions.

[0136] 4. Stitching stage: The preserved faces are stitched together into a closed shell by pairing vertex and curve hash keys. The entire pipeline has zero external dependencies.

[0137] To objectively evaluate the performance of Boolean operations, this system underwent a comparative experiment with the mainstream international open-source kernel OpenCASCADE under the same conditions. See Table 5 for details.

[0138]

[0139] Table 5

[0140] The experimental data above show that this system achieves complete autonomy with zero dependencies by trading for approximately a 2x performance difference. Figure 6 This is a bar chart comparing the Boolean operation performance of the technical solution of this invention with that of OpenCASCADE.

[0141] Furthermore, from the perspective of intersection locator coverage, the invented technical solution is significantly superior to the OpenCASCADE and ACIS / Parasolid solutions in terms of geometric kernel capability (see Table 6). Figure 7 As shown:

[0142]

[0143] Table 6

[0144] VI. Spline curves and smooth transitions: precise geometric representation.

[0145] Involute representation: 50 sampling points are generated at equal intervals along the tooth profile range on the base circle of 28.19 mm. The chord length parameterization is used to fit a 3rd order B-spline curve with an accuracy of 0.0008 mm, reaching IT5 level.

[0146] Smooth transition at the tooth root: The curvature at the end of the involute is zero, and the radius of curvature of the tooth root arc is 26.25 mm. There is a jump in curvature at the two points. Cubic Hermite interpolation is used in the approximately 2 mm transition section, and the position and tangent direction are matched to achieve continuity of the first derivative.

[0147] As the most complex test part (182 faces), the gear's standard format exported data fully verified the correctness and performance of the spline curve representation, as shown in Table 7:

[0148]

[0149] Table 7

[0150] In addition, see Figure 8 The comparison of mesh quality before and after, along with the FEA convergence curve, shows that the minimum Jacobian value of the transition section increased from 0.41 to 0.72 (an increase of 75%), the proportion of qualified elements in the entire tooth increased from 61% to 97% (an increase of 59%), curvature abrupt changes were completely eliminated, and the first-order continuity met the engineering requirements.

[0151] VII. Constraint Verification and Multi-Solution Optimization: Ensuring the Rationality of the Project.

[0152] Located between the perception layer and the execution layer, it performs engineering verification on the parameters extracted by artificial intelligence.

[0153] The verification results for this example are as follows: root cut test passed (20 > 17.1), contact strength passed (safety factor 1.51 > 1.5), and bending strength passed (safety factor 4.91).

[0154] After verification, three comparison schemes were output: standard scheme (weight 1.08 kg), lightweight scheme (0.71 kg, weight reduction 34%), and reinforced scheme (safety factor 2.10). The user selected the standard scheme.

[0155] If simulation is configured, the safety factor at the root fillet of the first gear is only 1.04, which fails. The system automatically provides feedback in three rounds: increasing the fillet radius, increasing the tooth width, and upgrading the material to 40Cr steel, ultimately achieving a safety factor of 2.79.

[0156] To verify the practical engineering value of constraint verification and AI diagnostics, two sets of experiments were conducted.

[0157] The first group consists of comprehensive tests, with 30 test cases spanning five areas: steel structures, pressure vessels, gears, key connections, and general machinery. These tests evaluate 401 national standard clauses. (See Table 8 for details.) Figure 9 :

[0158]

[0159] Table 8

[0160] As can be seen, out of a total of 30 cases and 401 clauses, 342 clauses were automatically compliant, with an average compliance rate of 85.3%.

[0161] The second group compared the quality of the hybrid diagnostic scheme of rule engine plus multimodal large model in this system with the pure multimodal large model scheme. The results of 13 cases are shown in Table 9. Figure 10 :

[0162]

[0163] Table 9

[0164] As can be seen from the above, this system is significantly superior to the pure large model solution in four dimensions: deterministic verification, national standard coverage, interpretability traceability, and offline availability.

[0165] 8. Standard format export and normal control.

[0166] The export module traverses the topology, benefiting from edge hashing for deduplication; for example, 602 vertices and 1260 unique edges are automatically written without duplication. Coordinates, direction vectors, curves, surfaces, directed edges, edge loops, high-level faces, closed shells, and unit definitions are output sequentially according to international standards. Face normal orientation is controlled by Boolean flags: outer surfaces are marked as true (normal facing outwards), while the inner wall of the center hole is flipped to false (normal facing inwards) during Boolean difference calculations, maintaining consistency throughout the entire link. The exported files have been verified with multiple commercial software programs: solids are closed, all normals face outwards, and there are no topology errors.

[0167] To verify the integrity and interoperability of the standard format export, export tests were performed on all six surface types one by one (see Table 10).

[0168]

[0169] Table 10

[0170] Experimental testing revealed that, under engineering constraints, the pass rate of the parts was 96%, with some parts exhibiting surface loss. Figure 11 This is a bar chart comparing the time taken to export a STEP format file with the file size for the technical solution of this invention.

[0171] IX. Summary of Examples

[0172] In this example, the overall architecture establishes a core route of three-layer decoupling: AI perception, logic verification, and geometric execution. Constraint verification completes the engineering rationality review before geometric construction. Layered modeling first builds the topological skeleton and then injects precise features. The entire construction process relies on edge hashing to ensure topological uniqueness. Boolean operations process entity subtraction in a four-stage pipeline. Spline curves and smooth transitions ensure involute accuracy and tooth root continuity. Standard export completes the output of consistent normals. Six types of surfaces have passed 100% international standard verification.

[0173] Experimental data shows that the system's single design time is 15 to 300 milliseconds, which is about 100,000 times faster than traditional workflows. The national standard verification coverage exceeds 85%, and the intersection parser coverage is 86% with zero external dependencies. The deterministic values ​​upstream drive the execution downstream, and the downstream topology infrastructure feeds back upstream, forming a tightly connected organic whole.

[0174] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A parameterized axisymmetric structure modeling system based on AI decision-making, characterized in that, include: The user interaction layer is used to receive user input information, which includes natural language descriptions, sketches, and parameter tables. The AI ​​perception layer calls a multimodal large model to perform intent recognition, part type recognition, geometric parameters, and physical condition parameter extraction on the input information, and outputs structured JSON parameters that do not contain geometric data. The logic verification layer includes a physical constraint verification module and a multi-scheme optimization module. The physical constraint verification module receives structured JSON parameters output by the AI ​​perception layer, performs root cut verification, strength verification, transmission ratio matching, and national standard compliance verification, intercepts unreasonable parameters, and outputs modification suggestions to the user interaction layer for the user to modify. The multi-scheme optimization module generates at least two optional design schemes based on the verified parameters, including a standard scheme, a lightweight scheme, and a high-strength scheme, and pushes scheme selection information to the user interaction layer for the user to choose from. The geometric execution layer includes the BRep module, the precision feature toolbox, and the Boolean operation engine. The BRep module has a built-in edge caching mechanism, which uses a hash table to cache vertex and edge objects and standardizes geometric parameters to generate hash keys, thereby constructing the macroscopic topological skeleton of the part. The precision feature toolbox encapsulates a variety of atomic-level precision feature algorithms and injects microscopic precision features at specified positions in the macroscopic topological skeleton through a callback mechanism. The Boolean operation engine uses a hybrid algorithm of edge classification and ray casting to complete entity union, intersection, and difference operations. Rays use non-principal axis direction vectors to assemble and generate complete BRep entities. The output layer controls the orientation of the face normals of the BRep entity and outputs STEP and / or STL format files.

2. The parameterized axisymmetric structure modeling system based on AI decision-making according to claim 1, characterized in that, The Boolean operation engine includes: Edge classification unit: For each edge of the entity participating in the operation, take the midpoint of the edge and emit a ray, and count the number of intersections; odd number of intersections determine that the edge is inside the entity, and even number of intersections determine that the edge is outside the entity; Face classification unit: classify all edges associated with a face by statistical analysis; if both internal and external labels exist, the face is marked as a cross-face; for the remaining faces, the label with the highest percentage of edges is used as the dominant face classification. Face selection unit: Performs face selection based on Boolean operation type and face classification; for cross-faces, uses the center point ray method to determine whether the face is inside or outside; during difference operations, the faces of the subtrahend entity are flipped according to their normals; Topology collection unit: collects the retained faces, associated edges, and vertices, and assembles them to generate a complete BRep entity.

3. The parameterized axisymmetric structure modeling system based on AI decision-making according to claim 1, characterized in that, It also includes a finite element simulation module that interfaces with the logic verification layer. The logic verification layer receives the finite element simulation feedback results from the finite element simulation module and iteratively optimizes the parameters until a preset safety factor is met.

4. The parameterized axisymmetric structure modeling system based on AI decision-making according to claim 1, characterized in that, The AI ​​perception layer includes a prompt word engineering unit and a RAG retrieval enhancement unit; the prompt word engineering unit limits the multimodal large model to only output structured JSON parameters, and the RAG retrieval enhancement unit assists in parameter extraction by retrieving knowledge information from an external mechanical design knowledge base.

5. The parameterized axisymmetric structure modeling system based on AI decision-making according to claim 1, characterized in that, The atomic-level precision feature algorithms include keyway cutting, thread generation, spline cutting, undercut groove, and cam contour algorithms; when adding a new atomic-level precision feature algorithm, only the algorithm function is added to the precision feature toolbox.

6. The parameterized axisymmetric structure modeling system based on AI decision-making according to claim 1, characterized in that, The geometry execution layer also includes a NURBS curve generation unit, which generates discrete points and constructs cubic NURBS curves based on mathematical equations for the contours of gear involutes and bearing raceways.

7. The parameterized axisymmetric structure modeling system based on AI decision-making according to claim 1, characterized in that, The geometry execution layer also includes a smooth transition unit, which uses cubic Hermite interpolation at the boundaries of features including keyways and relief grooves to generate a C1 continuous transition region at the feature boundaries.

8. A parameterized axisymmetric structure modeling method based on AI decision-making, characterized in that, The system applied to any one of claims 1-7 includes the following steps: S100: Receives user input information in at least one form, including natural language, sketches, and parameter tables; S200: The AI ​​perception layer parses user input information, identifies part type, geometric dimensions, material, and working condition constraints, and outputs structured JSON parameters that do not contain geometric data; S300: The logic verification layer performs engineering constraint verification. If the structured JSON parameters do not meet the design standards, it outputs modification suggestions to the user interaction layer for the user to modify. After the verification is successful, it generates multiple candidate solutions for the user to choose from and receives the target solution selected by the user. S400: The geometry execution layer performs deterministic modeling, specifically including: (1) The BRep module constructs the macroscopic topology skeleton of the part based on the target scheme parameters; (2) Call back the precision feature toolbox to inject micro-precision features at the specified position of the macro topological skeleton; (3) Enable the edge caching mechanism to ensure the uniqueness of vertex and edge objects and construct the macroscopic topological skeleton of the parts; (4) Call the Boolean operation engine to complete the entity Boolean operation and obtain the complete BRep entity; S500: Output layer controls the orientation of the face normals of the BRep entity, and exports STEP and / or STL files.

9. The parameterized axisymmetric structure modeling method based on AI decision-making according to claim 8, characterized in that, The Boolean operation engine performs the following steps: Edge classification: For each edge of the entity participating in the operation, take the midpoint of the edge and emit a ray, and count the number of intersections; odd number of intersections determine that the edge is inside the entity, and even number of intersections determine that the edge is outside the entity; Face classification: Classify all edges associated with a face by their classification labels; if both internal and external labels exist, the face is labeled as a cross-face; for the remaining faces, the label with the highest percentage of edges is used as the dominant face classification. Face selection: Face selection is completed based on Boolean operation type and face classification; for cross-faces, the center point ray method of the sampled face is used to determine whether the face is inside or outside; during the difference operation, the normals of the faces of the subtrahend entity are flipped; Topology collection: Collect the retained faces, associated edges, and vertices, and assemble them to generate a complete BRep entity.

10. The parameterized axisymmetric structure modeling method based on AI decision-making according to claim 8 or 9, characterized in that, It also includes a simulation closed-loop iteration step: the generated BRep entity is sent to the finite element simulation module to obtain simulation results; the simulation safety factor and stress results are sent back to the logic verification layer, the logic verification layer automatically adjusts the geometric parameters, and re-executes the modeling process until the simulation indicators meet the preset safety factor.

Images