Large model driven industrial product design process optimization system and method

Abstract

The application relates to the technical field of three-dimensional modeling, and discloses a large model driven industrial product design process optimization system and method, the large model driven industrial product design process optimization system comprising the following steps: analyzing input text; generating a differential parameter candidate set; constructing a surface boundary entity model and constructing a surface boundary topology graph; calculating a comprehensive impedance value; generating a target geometric parameter vector; generating a final instruction sequence and a final entity model; and outputting the final instruction sequence, the final entity model, the final comprehensive impedance value, the final model quality, the final modal frequency and the impedance improvement rate. The application realizes directional optimization of parameters through a one-time finite difference candidate set, avoids randomness of a large model, and generates a model which can simultaneously meet multidimensional constraints of heat, mechanics, manufacturing and assembly, so that a large amount of manual correction is not needed, the design iteration cycle of an industrial product is shortened, and the engineering practicability and landability of a design result are improved.

Description

Technical Field

[0001] This invention relates to the field of 3D modeling technology, and more specifically, to a system and method for optimizing the industrial product design process driven by large models. Background Technology

[0002] With the rapid development of the industrial mobile robot industry, higher requirements have been placed on the design of heat dissipation and support housings for motor drivers. These housings need to simultaneously undertake multiple functions such as heat dissipation, load bearing, sealing, and assembly clearance. Traditional computer-aided design processes rely heavily on engineers' experience, requiring multiple rounds of simulation verification and structural correction. This results in long design iteration cycles, significant challenges in multi-objective collaborative optimization, and difficulty in meeting the rapidly iterating R&D needs of the industry.

[0003] Existing computer-aided design technologies driven by large models have achieved the automatic generation of 3D geometric models from natural language requirements. However, existing technologies only focus on the model's appearance and the generation effect of parametric sequences. They cannot accurately map the engineering functional requirements in natural language to the boundary representation topology and functional transfer paths of the 3D model. Furthermore, they have not established a unified dimensionless engineering performance evaluation system covering thermal, mechanical, and geometric continuity, and therefore cannot provide a clear physical basis for model optimization.

[0004] Therefore, the aforementioned existing technologies result in the generated 3D models failing to meet the rigid constraints of industrial scenarios such as heat dissipation, vibration resistance, and manufacturing assembly. This necessitates extensive manual corrections and secondary designs by engineers, failing to fundamentally improve the design efficiency of industrial products and hindering the realization of end-to-end closed-loop design from design requirements to feasible engineering models. Summary of the Invention

[0005] This invention provides a large-model-driven industrial product design process optimization system and method, which solves the technical problems mentioned in the background.

[0006] This invention provides a large-model-driven optimization system for industrial product design processes, comprising: The first module analyzes the semantic feature set and physical constraint boundary set of the input text extraction function. The second module generates a geometric parameter vector based on the functional semantic feature set and the physical constraint boundary set, and extracts the center point parameters of the geometric parameter vector to generate a candidate set of difference parameters. The third module converts the differential parameter candidate set into a surface boundary entity model, extracts the surface nodes and adjacent edges of the surface boundary entity model to construct a surface boundary topology graph, and calculates the average adjacent edge length. The fourth module matches functional surfaces in the surface boundary topology diagram, calculates the heat transfer resistance, structural flexibility and geometric bending between functional surfaces, and calculates the comprehensive impedance value by combining the average adjacent edge length. The fifth module calculates the model quality and modal frequencies of the surface boundary entity model by combining the physical constraint boundary set, and obtains the finite difference gradient of the center point parameters based on the comprehensive impedance value. Under the constraints of model quality and modal frequency, the target geometric parameter vector is generated based on the finite difference gradient. The sixth module decomposes the minimum impedance path and path impedance contribution rate from the comprehensive impedance value, and generates the final instruction sequence and final entity model by combining the target geometric parameter vector. The seventh module recalculates the final synthesized impedance value, final model quality, and final modal frequency of the final solid model. It compares the final synthesized impedance value with the synthesized impedance value to obtain the impedance improvement ratio, and outputs the final instruction sequence, final solid model, final synthesized impedance value, final model quality, final modal frequency, and impedance improvement ratio.

[0007] This invention provides a large-model-driven optimization method for industrial product design processes, comprising the following steps: Step S1: Parse the semantic feature set and physical constraint boundary set of the input text extraction function; Step S2: Generate a geometric parameter vector based on the functional semantic feature set and the physical constraint boundary set, and extract the center point parameters of the geometric parameter vector to generate a candidate set of difference parameters; Step S3: Convert the candidate set of differential parameters into a surface boundary entity model, extract the surface nodes and adjacent edges of the surface boundary entity model to construct a surface boundary topology graph, and calculate the average adjacent edge length. Step S4: Match functional surfaces in the surface boundary topology diagram, calculate the heat transfer resistance, structural flexibility and geometric bending between functional surfaces, and calculate the comprehensive impedance value in combination with the average adjacent edge length. Step S5: Calculate the model quality and modal frequencies of the surface boundary entity model by combining the physical constraint boundary set, and obtain the finite difference gradient of the center point parameters based on the comprehensive impedance value. Generate the target geometric parameter vector based on the finite difference gradient under the constraints of model quality and modal frequency. Step S6: Decompose the minimum impedance path and path impedance contribution rate from the comprehensive impedance value, and generate the final instruction sequence and final entity model by combining the target geometric parameter vector. Step S7: Recalculate the final composite impedance value, final model quality, and final modal frequency of the final solid model. Compare the final composite impedance value with the composite impedance value to obtain the impedance improvement ratio. Output the final instruction sequence, final solid model, final composite impedance value, final model quality, final modal frequency, and impedance improvement ratio.

[0008] The beneficial effects of this invention are as follows: This invention achieves closed-loop optimization of the entire process from natural language engineering requirements to parametric design of industrial products. It maps functional requirements to the boundary representation topology and functional transfer paths of a 3D model, constructing a unified comprehensive performance evaluation system and providing a clear physical direction for design optimization. This invention achieves targeted parameter optimization through a one-time finite difference candidate set, avoiding the randomness of large model generation. The generated model can simultaneously satisfy multi-dimensional constraints of thermal, mechanical, and manufacturing assembly, eliminating the need for extensive manual correction, shortening the design iteration cycle of industrial products, and improving the engineering practicality and feasibility of the design results. Attached Figure Description

[0009] Figure 1 This is a calculation flowchart of the large-model-driven industrial product design process optimization method of the present invention. Detailed Implementation

[0010] 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.

[0011] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0012] like Figure 1 As shown, the large-model-driven industrial product design process optimization system includes: The first module analyzes the semantic feature set and physical constraint boundary set of the input text extraction function. The second module generates a geometric parameter vector based on the functional semantic feature set and the physical constraint boundary set, and extracts the center point parameters of the geometric parameter vector to generate a candidate set of difference parameters. The third module converts the differential parameter candidate set into a surface boundary entity model, extracts the surface nodes and adjacent edges of the surface boundary entity model to construct a surface boundary topology graph, and calculates the average adjacent edge length. The fourth module matches functional surfaces in the surface boundary topology diagram, calculates the heat transfer resistance, structural flexibility and geometric bending between functional surfaces, and calculates the comprehensive impedance value by combining the average adjacent edge length. The fifth module calculates the model quality and modal frequencies of the surface boundary entity model by combining the physical constraint boundary set, and obtains the finite difference gradient of the center point parameters based on the comprehensive impedance value. Under the constraints of model quality and modal frequency, the target geometric parameter vector is generated based on the finite difference gradient. The sixth module decomposes the minimum impedance path and path impedance contribution rate from the comprehensive impedance value, and generates the final instruction sequence and final entity model by combining the target geometric parameter vector. The seventh module recalculates the final synthesized impedance value, final model quality, and final modal frequency of the final solid model. It compares the final synthesized impedance value with the synthesized impedance value to obtain the impedance improvement ratio, and outputs the final instruction sequence, final solid model, final synthesized impedance value, final model quality, final modal frequency, and impedance improvement ratio.

[0013] In one embodiment of the present invention, the specific calculation steps of the first module include: The formula for calculating function weights is as follows: in For functional weights, This represents the role weight vector corresponding to the functional role. It is the transpose symbol. For functional semantic embedding vectors, For comparison with the traversal numbering, The total number of functional semantic elements. To compare the role weight vectors corresponding to the functional roles, To compare the functional semantic embedding vector, It is an exponential function with the natural constant as its base; The formula for constructing the functional semantic feature set is as follows: in For functional semantic feature set, Number the elements. The total number of functional semantic elements. For functional semantic elements, For functional roles, For numerical information, For functional semantic embedding vectors, As a functional anchor point, For spatial attenuation scale, Functional weights; The formula for constructing the physical constraint boundary set is as follows: in For the physical constraint boundary set, To design the spatial shape envelope, The thermal conductivity of the material. The elastic modulus of the material. For material density, For a fixed area, This is a prohibited area. To create a constraint set, For the maximum permissible mass, This is the lower limit of the first-order natural frequency.

[0014] It should be noted that the functional semantic feature set is a structured set that carries all engineering functional requirements in the natural language input text, generated by merging multiple functional semantic elements. The physical constraint boundary set is a structured set that limits the boundary conditions of the entire industrial product design process, generated by merging the design space shape envelope, material parameters, constraint regions, manufacturing constraints, and performance thresholds. The functional role is an identifier that defines the engineering function type corresponding to a single functional semantic element. Eight core types are preset, including heat source, heat dissipation target, load input, support output, sealing, avoidance, grounding, and assembly positioning, which can be collected through semantic role annotation using a large language model. Numerical information is quantitative parameter information related to industrial product design extracted from the input text, which can be collected through named entity recognition technology using a large language model. The functional semantic embedding vector is a high-dimensional numerical vector used for semantic similarity calculation, generated after vectorizing the semantic content of a single functional semantic element. The functional anchor point is the spatial coordinate information that defines the approximate location of the engineering function corresponding to a single functional semantic element in the three-dimensional design space, which can be collected through semantic recognition mapping in conjunction with the design space shape envelope.

[0015] It should be noted that the spatial attenuation scale is a quantitative parameter defining the influence range of an engineering function corresponding to a single functional semantic element in the three-dimensional design space, which can be obtained through the semantic description and conversion of the functional influence range. The role weight vector is a weight vector calculated based on the semantic importance of different functional roles, with the preferred value being a 128-dimensional floating-point vector consistent with the dimension of the functional semantic embedding vector. The functional weight is a normalized value characterizing the importance of a single functional semantic element in the current design task. The functional semantic element is the smallest structured unit that carries the single engineering functional requirement, generated by combining functional roles, numerical information, functional semantic embedding vectors, functional anchors, spatial attenuation scale, and functional weight. The design space shape envelope is the three-dimensional spatial boundary that limits the maximum shape range of the industrial product design, which can be obtained through three-dimensional scanning of the installation space or design specifications. The material thermal conductivity is a physical parameter that determines the heat conduction capacity of the shell material selected in the design under steady-state conditions, which can be obtained through material property parameter tables or thermophysical property testing experiments. The material elastic modulus is the ratio of normal stress to corresponding normal strain of the shell material selected in the design during the elastic deformation stage, which can be obtained through material property parameter tables or mechanical property testing experiments. Material density is the mass per unit volume of the shell material selected in the design, which can be obtained through material property parameter tables or density test experiments.

[0016] It should be noted that the fixed area refers to the three-dimensional spatial area that cannot be modified during the design process and must be completely preserved; this area can be extracted and collected from the 3D design drawings. The prohibited area refers to the three-dimensional spatial area that cannot be intruded upon during the design process and must be completely avoided; this area can also be extracted and collected from the 3D design drawings. The manufacturing constraint set is the set of all manufacturing process-related constraints that must be followed during the design process; this can be collected from the processing technology design specifications or the manufacturer's processing capability parameters. The maximum permissible mass is the upper limit of the mass that the completed industrial product shell cannot exceed; this can be collected from the product weight distribution specifications or the overall machine load requirements. The lower limit of the first-order natural frequency is the lowest first-order natural frequency value that the completed industrial product shell must meet; this can be collected from operational vibration frequency testing or vibration resistance design specifications.

[0017] It should be noted that the functional semantic embedding vector is generated using a BERT encoding model pre-trained in the industrial design domain. The input is Chinese text of a single functional description statement, with a maximum input token length of 512, and the output is a 128-dimensional floating-point vector. The model is first pre-trained on a general Chinese corpus, and then fine-tuned using a contrastive learning strategy with a functional requirement corpus from the industrial design domain. During encoding, the output vector of the statement's CLStoken is extracted and reduced to 128 dimensions through a fully connected layer to obtain the final embedding vector. The BERT encoding model is a current technology and will not be elaborated upon here.

[0018] It should be noted that the role weight vector adopts the same 128-dimensional floating-point format as the functional semantic embedding vector. Each functional role corresponds to a fixed weight vector, which is generated through training on a manually labeled dataset. A dataset of labeled statements for each of the eight functional roles is pre-constructed. After generating semantic embedding vectors through the encoding model, the average value is taken to obtain the initial vector, which is then fine-tuned manually to complete the preset. The vector magnitude of the four core functional roles—heat source, heat dissipation target, load input, and support output—is set to 1.2, while the vector magnitude of the other four auxiliary functional roles is set to 0.8. For example, for the four core functional roles of heat source, heat dissipation target, load input, and support output in the shell design, the vector magnitude is set to 1.2, while for the four auxiliary functional roles of sealing, avoidance, grounding, and assembly positioning, the vector magnitude is set to 0.8. The larger the magnitude, the higher the weight of the functional role in the semantic importance calculation.

[0019] It should be noted that the design space for functional anchor points adopts a Cartesian three-dimensional coordinate system, with the origin set at the lower left corner and lower right vertex of the design space. The X-axis is along the length direction, the Y-axis along the width direction, and the Z-axis along the height direction. The functional anchor point is the geometric center coordinate of the identified structural position in this coordinate system. The spatial attenuation scale is the radius of the spherical influence area centered on the functional anchor point. For centralized functions, the priority value is 1.5 times the radius of the largest outer circle of the structure where the functional anchor point is located. For distributed functions, the priority value is 0.3 times the largest dimension of the corresponding dimension of the design space. For example, for heat source bonding and mounting hole positioning, the priority value of the spatial attenuation scale is 1.5 times the radius of the largest outer circle of the structure where the functional anchor point is located. For distributed functions, such as heat dissipation and support, the priority value of the spatial attenuation scale is 0.3 times the largest dimension of the corresponding dimension of the design space.

[0020] In one embodiment of the present invention, the specific calculation steps of the second module include: The formula for constructing the geometric parameter vector is as follows: in For geometric parameter vectors, For the shell wall thickness, The height of the heat dissipation fins The spacing between the heat dissipation fins. To increase the width of the ribs, To increase the height of the ribs, For position parameters, The radius of the fillet is 1. This refers to the opening size. It is the transpose symbol; The formulas for calculating the lower bound and upper bound of a parameter for a single dimension are as follows: in For the lower bound of parameters in a single dimension, For the upper bound of parameters in a single dimension, Generate functions for parameter boundaries. For functional semantic feature set, For the physical constraint boundary set, Number the elements. The total number of functional semantic elements. Numerical information; The formula for calculating the center point parameter for a single dimension is as follows: in For the center point parameters in a single dimension, Number the dimensions. For the lower bound of parameters in a single dimension, This is an upper bound for the parameters in a single dimension; The formula for calculating the finite difference perturbation step size for a single dimension is as follows: in For a finite difference perturbation step size in a single dimension, For the upper bound of parameters in a single dimension, For the lower bound of parameters in a single dimension, For parameter dimensions, Number the dimensions; The formula for generating the candidate set of difference parameters is as follows: in For the candidate set of difference parameters, It is a combination of multidimensional center point parameters. For a finite difference perturbation step size in a single dimension, The parameter is a unit vector in the direction of the parameter. For positive disturbance parameters, These are negative disturbance parameters. Number the dimensions. For parameter dimensions; The formula for calculating the comprehensive feasible region is as follows: in To synthesize the feasible domain, For geometric parameter vectors, For the lower bound of parameters in a single dimension, For the upper bound of parameters in a single dimension, For surface boundary solid models, To design the spatial shape envelope, This is a prohibited area. It is an empty set. For a fixed area, To create constraint functions, To create a set of constraints.

[0021] It should be noted that the geometric parameter vector is a column vector generated from all adjustable geometric parameters related to the shell design, used for parametric design, and fixedly contains 8 core dimensions. Shell wall thickness is the thickness of the main wall surface of the heat dissipation bearing shell, preferably ranging from 1.5 mm to 5 mm. Heat dissipation rib height is the vertical height of the heat dissipation rib from its base to its top, preferably ranging from 5 mm to 30 mm. Heat dissipation rib spacing is the distance between the centerlines of two adjacent heat dissipation ribs, preferably ranging from 8 mm to 20 mm. Reinforcing rib width is the thickness of the shell reinforcing rib, preferably ranging from 2 mm to 6 mm. Reinforcing rib height is the vertical height of the shell reinforcing rib from its base to its top, preferably ranging from 5 mm to 25 mm. Position parameters are the coordinate positions of the shell mounting ears, mounting holes, and screw posts within the design space, with values ​​within the corresponding coordinate range of the design space. Corner radius is the transition corner radius of all edges of the shell, preferably ranging from 0.5 mm to 3 mm. The opening dimensions refer to the length, width, and diameter of the wiring harness opening and airflow opening on the shell, with a preferred range of 5 mm to 100 mm. The parameter boundary generation function is a piecewise linear mapping function that calculates the upper and lower bounds of each parameter dimension based on the functional semantic feature set, the physical constraint boundary set, and numerical information.

[0022] It should be noted that the lower bound of the parameter is the minimum allowed value for a single geometric parameter dimension. The upper bound of the parameter is the maximum allowed value for a single geometric parameter dimension. The center point parameter is the average of the lower and upper bounds of the parameter for a single geometric parameter dimension. The parameter dimension is the number of adjustable geometric parameters contained in the geometric parameter vector, which is fixed at 8. The finite difference perturbation step size is the numerical magnitude of the positive and negative perturbations for a single geometric parameter dimension in the finite difference calculation. The parameter direction unit vector is the unit direction vector corresponding to a single parameter dimension in the geometric parameter vector space, preferably an 8-dimensional unit vector with only the corresponding dimension having a value of 1 and the other dimensions having 0. The multidimensional center point parameter combination is the geometric parameter vector generated by combining the center point parameters of all parameter dimensions. The positive perturbation parameter is the geometric parameter vector generated by superimposing the multidimensional center point parameter combination on a single parameter dimension with the finite difference perturbation step size. The negative perturbation parameter is the geometric parameter vector generated by subtracting the finite difference perturbation step size from the multidimensional center point parameter combination on a single parameter dimension. The differential parameter candidate set is a parameter set generated by merging the multidimensional center point parameter combination, all positive perturbation parameters, and all negative perturbation parameters. The surface boundary solid model is a 3D computer-aided design solid model with complete boundary representation, generated from geometric parameter vectors. The manufacturing constraint function is a function that verifies whether the geometric parameter vectors conform to the manufacturing constraint set, and its output is the sum of the values ​​of all constraint-violating models. The comprehensive feasible region is the set of geometric parameter vectors that satisfy all geometric boundaries, spatial constraints, and manufacturing constraints.

[0023] It should be noted that the parameter boundary generation function takes as input a set of functional semantic features, a set of physical constraint boundaries, and numerical information, and outputs the upper and lower bounds for each parameter dimension. The specific execution includes the following steps: Step S101: Map the set of manufacturing constraints in the physical constraint boundary set to the lower and upper bounds of each geometric parameter. The lower and upper bounds are the absolute allowable range of the parameter under the corresponding manufacturing process. Step S102: Traverse all numerical information in the functional semantic feature set. If a certain numerical information corresponds to a specific geometric parameter, then set the range in the numerical information as the temporary upper and lower bounds of the parameter. The temporary upper and lower bounds shall not exceed the range of the basic upper and lower bounds. Step S103: Based on the design space shape envelope of the physical constraint boundary concentration, the upper bound of parameters related to the space boundary, such as heat dissipation fin height, reinforcing rib height, position parameters, and opening size, is corrected to ensure that the structure corresponding to the maximum value of the parameter does not exceed the design space. Step S104: Based on the maximum permissible mass of the physical constraint boundary set, the upper bounds of parameters affecting the shell mass, such as shell wall thickness, heat dissipation fin height, heat dissipation fin spacing, reinforcing rib width, and reinforcing rib height, are corrected. The influence coefficient of each parameter on the mass is determined through parameter sensitivity analysis to ensure that the shell mass corresponding to the maximum value of the parameter does not exceed the maximum permissible mass.

[0024] It should be noted that the output of the manufacturing constraint function is the sum of the values ​​of all manufacturing constraint items that violate the model. The core includes six types of constraint items: minimum wall thickness, minimum fillet radius, minimum rib spacing, maximum rib height-to-wall thickness ratio, maximum rib height-to-rib width ratio, and maximum opening size. When each constraint is violated, the corresponding difference is output. When there is no violation, 0 is output. A sum ≤ 0 means that all manufacturing constraints are satisfied.

[0025] It should be noted that the comprehensive feasible region projection operator uses sequential constraint projection, with constraint priorities from high to low as follows: fixed region constraints, prohibited area constraints, design space constraints, manufacturing constraints, and geometric boundary constraints. The parameter vector is checked and corrected sequentially according to priority. After each round of projection, it is checked whether all constraints are satisfied. If they are satisfied, the output is terminated; otherwise, the process is repeated. The maximum number of iterations is 20. When the upper limit is reached, the parameter vector that is closest to the feasible region is output.

[0026] In one embodiment of the present invention, the specific calculation steps of the third module include: The calculation formula for the auxiliary design construction sequence is as follows: ; in To assist in the design and construction of sequences, Generate functions for large models. For the input text requirement, For functional semantic feature set, For the physical constraint boundary set, Candidate parameters; The calculation formula for generating a surface boundary solid model is as follows: ; in For surface boundary solid models, This is a function for calculating the geometric kernel. To assist in the design and construction of sequences; The calculation formula for constructing the surface boundary topology graph is as follows: ; in For the surface boundary topology graph, For a set of face nodes, It is the set of adjacent edges; The formula for calculating the semantic geometry embedding vector is as follows: in For semantic geometry embedding vectors, For surface feature coding function, For surface area, For surface normal vectors, Let the coordinates be the centroid coordinates of the surface. For the local thickness of the surface, For surface curvature features, For face type tags, It is a set of adjacent faces; The formula for calculating the average local thickness is as follows: in For average local thickness, The local thickness of the first face node. The local thickness of the second face node; The formula for calculating the equivalent heat transfer cross section is as follows: in For the equivalent heat transfer cross section, For connection width, For average local thickness, The inter-surface continuity factor; The formula for calculating the equivalent bending section moment of inertia is as follows: in For the equivalent bending section moment of inertia, This is the cross-section correction factor. For connection width, The average local thickness; The formula for calculating the average adjacent edge length is as follows: in For the average adjacent edge length, This represents the total number of adjacent edges. Positioning and combining adjacent face nodes. For the set of adjacent edges, Equivalent path length It should be noted that the computer-aided design construction sequence is a parametric modeling history sequence consisting of a series of executable computer-aided design modeling commands used to generate 3D solid models. The large model generation function is a large language model function that generates the computer-aided design construction sequence based on the input design requirements and parameters, preferably using a CodeLlama-7B model fine-tuned from the CAD construction sequence corpus. The geometric kernel calculation function is a geometric kernel function that executes the computer-aided design modeling commands to generate a 3D solid model with a complete boundary representation structure, preferably using the OpenCASCADE open-source geometric kernel.

[0027] A surface boundary topology graph is a graph structure constructed by treating each boundary representation face of a surface boundary entity model as a node and the shared connections between adjacent faces as edges. A surface node is the basic node unit in the surface boundary topology graph, corresponding to an independent boundary representation face in the surface boundary entity model. The surface area is the numerical value of the surface area of ​​the boundary representation face corresponding to the surface node, which can be extracted using the geometric kernel surface area calculation function. The surface normal vector is the unit normal vector of the boundary representation face corresponding to the surface node, representing the orientation of the face in 3D space, and can be extracted using the geometric kernel surface normal vector calculation function.

[0028] It should be noted that the centroid coordinates of a face are the coordinates of the geometric centroid of the boundary representation face corresponding to the face node in three-dimensional space, which can be extracted using the geometric kernel face centroid calculation function. The local thickness of a face is the thickness of the shell entity at the location of the boundary representation face corresponding to the face node, which can be measured and extracted using the geometric kernel wall thickness analysis function. The surface curvature features are the average curvature and Gaussian curvature values ​​of the boundary representation face corresponding to the face node, characterizing the degree of curvature of the face, which can be extracted using the geometric kernel surface curvature calculation function. The face type label is a label identifying the geometric type and engineering function type of the boundary representation face corresponding to the face node, which can be classified and labeled using the geometric kernel face type recognition function combined with geometric features. The adjacent face set is the set of all adjacent face nodes that share a connection with the current face node, which can be extracted using the geometric kernel face adjacency query function. The face feature encoding function is an 8-layer graph neural network encoder that encodes the geometric and topological features of face nodes into semantic geometric embedding vectors. The semantic geometric embedding vector is a high-dimensional numerical vector generated after vectorizing and encoding the geometric, topological, and engineering function features of face nodes.

[0029] It should be noted that adjacent edges are the basic edge units in the surface boundary topology graph, corresponding to the shared connection relationship between two adjacent surface nodes. The equivalent path length is the straight-line distance between the centroids of the two adjacent surface nodes corresponding to the adjacent edge, which can be obtained by calculating the Euclidean distance using the coordinates of the two surface centroids. The connection width is the total length of the shared boundary of the two adjacent surface nodes corresponding to the adjacent edge, which can be obtained by summing the lengths using the geometric kernel edge length calculation function. The inter-surface continuity factor is a quantitative parameter characterizing the smoothness of the connection between two adjacent surface nodes, with a preferred value range of 0 to 1. The average local thickness is the average of the local thicknesses of the two adjacent surface nodes corresponding to the adjacent edge. The equivalent heat transfer section is the equivalent cross-sectional area characterizing the heat transfer capacity between two adjacent surface nodes. The section correction coefficient is a quantitative parameter reflecting the effect of the local structure on enhancing the bending resistance of the section, with a preferred value range of 1 to 2.5. The equivalent bending section moment of inertia is a geometric quantity characterizing the ability of the connection structure between two adjacent surface nodes to resist bending deformation. The average adjacent edge length is the average of the equivalent path lengths corresponding to all adjacent edges in the face boundary topology graph.

[0030] It should be noted that the basic architecture of the large model generation function is the CodeLlama-7B model, with the addition of a CAD command validation adapter and constraint alignment module. The input is a fixed-structure prompt word template, including four modules: design object, functional requirements, constraints, and geometric parameters. The output is Python modeling code conforming to the OpenCASCADE specification. The model is first pre-trained on a general code corpus, and then fine-tuned using an industrial shell CAD modeling code corpus, with syntax validation loss and constraint alignment loss added to complete the training.

[0031] It should be noted that the surface type label is divided into two levels: geometric type and engineering function type. Geometric type includes plane, cylindrical surface, conical surface, spherical surface, and freeform surface, which are determined by curvature characteristics. Engineering function type includes heat source bonding surface, heat dissipation fin side, mounting ear root surface, support bottom surface, sealing groove side wall, wire harness opening surface, grounding boss surface, and appearance surface, which are determined by the position, geometric characteristics, and adjacency relationship of the bonding surface.

[0032] It should be noted that the inter-surface continuity factor is fixed in the range of 0 to 1, the value of tangent surfaces for smooth connection is 1, the value of fillet transition connection is 0.7 to 0.95 according to the ratio of fillet radius to connection width, the value of straight vertical connection is 0.5, and the value of sharp bend connection less than 90 degrees is 0.1 to 0.3.

[0033] It should be noted that the correction method for the cross-section correction factor is as follows: the value is 1 for a straight connection without reinforcement, 1.1 to 1.3 for a connection with a rounded corner transition, 1.4 to 1.8 for a connection with a single-sided reinforcing rib, 1.8 to 2.2 for a connection with double-sided reinforcing ribs, and 2.2 to 2.5 for a connection with a thickened boss.

[0034] In one embodiment of the present invention, the specific calculation steps of the fourth module include: The formula for calculating the normalized area value is as follows: in This is the area normalized value. For surface area, The median of the surface area. For a set of face nodes, It is the smallest positive number in the area; The formula for calculating the area adjustment function is as follows: in This is an area adjustment function. This is the area normalized value. For target functional roles; The formula for calculating the face matching score is as follows: in For face matching score, To satisfy the functional semantic elements of functional roles, For the matched functional role, For the target functional role, For functional weights, For cosine similarity calculation, It is an exponential function with the natural constant as its base. For functional semantic embedding vectors, For semantic geometry embedding vectors, Let the coordinates be the centroid coordinates of the surface. As a functional anchor point, For spatial attenuation scale, This is an area adjustment function; The calculation formulas for the set of location heat source surfaces, the set of heat dissipation target surfaces, the set of load input surfaces, and the set of support output surfaces are as follows: in For heat source surface collection, For face nodes, The surface matching score belongs to the heat source surface attribute. Select a threshold for the source surface. For the set of heat dissipation target surfaces, The surface matching score belongs to the heat dissipation target surface attribute. Select a threshold for the target surface. For the set of load input surfaces, The surface matching score is a property belonging to the load input surface. To support the output surface set, The face matching score belongs to the supporting output face attribute; The formulas for calculating thermal path strength and force path strength are as follows: in For thermal path intensity, For the source surface thermal matching score, A score is assigned to the target surface for heat dissipation. For force path strength, For the source surface load matching score, Matching scores to the target surface; The formula for calculating path weight is as follows: in For path weights, For the global weight of the hot path, For thermal path intensity, For the global weight of the force path, For force path strength, This represents the combination of traversal of the source and target faces. To calculate the thermal path intensity, To calculate the force path strength, The value is a positive number with a very small weight. The formula for calculating the original thermal resistance is as follows: in This is the original value of thermal resistance. The equivalent path length, The thermal conductivity of the material. For the equivalent heat transfer cross section, It is a positive number that represents the minimum heat transfer value; The formula for calculating the heat transfer resistance is as follows: in This is the heat transfer resistance value. This is the original value of thermal resistance. This is the median of the original thermal resistance. For the adjacent edges being traversed, For the set of adjacent edges, It is a dimensionless positive number for thermal resistance; The formula for calculating the elastic compliance factor is as follows: in For elastic flexibility, The equivalent path length, The elastic modulus of the material. For the equivalent bending section moment of inertia, It is a positive number with a minimum inertia; The formula for calculating the structural flexibility value is as follows: in This represents the structural flexibility value. This is the original amount of elastic flexibility. This represents the median of the original elastic flexibility. It is the smallest positive number in the dimension of flexibility; The formula for calculating geometric bend is as follows: in For geometric bending degree, Let be the normal vector of the first face. Let be the surface normal vector of the second face. The equivalent path length, For the average adjacent edge length, It is a minimum positive number for bending; The formula for calculating the bending amount of the combined impedance is as follows: in The bending amount is the combined impedance of the edge. As thermal resistance weight, This is the heat transfer resistance value. For elasticity and flexibility weight, This represents the structural flexibility value. For geometric bending weight, Geometric bending degree; The formula for calculating the minimum impedance path is as follows: in For the path with minimum impedance, This indicates the path traversal from the source face to the target face. To iterate through the adjacent edges within the path. The bending amount of the edge composite impedance; The formula for calculating the overall impedance is as follows: in This is the combined impedance value. For path weights, For the path with minimum impedance, The bending amount is the combined impedance of the edge. The equivalent path length, For the average adjacent edge length, It is a dimensionless, minimal positive number.

[0035] It should be noted that the area normalization value is a dimensionless value obtained by dividing the area of ​​a single face node by the sum of the median of the areas of all faces in the face node set and the smallest positive value of the area. The smallest positive value of the area is a very small value of 0.001 square millimeters to avoid the denominator of the area normalization calculation being zero. The area adjustment function is a piecewise function that corrects the functional face matching score based on the face area. The face matching score is a quantitative value that characterizes the degree of matching between a single face node and the target functional role; the higher the value, the higher the degree of matching. The source face selection threshold is the critical value of the face matching score for screening source face nodes, preferably 0.5. The target face selection threshold is the critical value of the face matching score for screening target face nodes, preferably 0.5. The heat source face set is the set of face nodes that perform the heat source bonding function and whose face matching score exceeds the source face selection threshold. The heat dissipation target face set is the set of face nodes that perform the heat dissipation function and whose face matching score exceeds the target face selection threshold. The load input surface set consists of surface nodes whose surface matching scores exceed the source surface selection threshold and which perform the load input function. The support output surface set consists of surface nodes whose surface matching scores exceed the target surface selection threshold and which perform the support output function.

[0036] It should be noted that thermal path intensity is a quantitative value characterizing the importance of the heat transfer path between a single heat source surface and a single heat dissipation target surface. Force path intensity is a quantitative value characterizing the importance of the force transfer path between a single load input surface and a single support output surface. Global weight of thermal path is a weighting parameter that adjusts the proportion of thermal paths in the overall evaluation, preferably set to 0.5. Global weight of force path is a weighting parameter that adjusts the proportion of force paths in the overall evaluation, preferably set to 0.5. Minimum positive weight is a minimum value to avoid the denominator of path weight calculation being zero, preferably set to 0.0001. Path weight is a normalized value characterizing the proportion of the path from a single source surface to the target surface in the overall evaluation. Minimum positive heat transfer is a minimum value to avoid the denominator of thermal resistance calculation being zero, preferably set to 0.001 watts per meter Kelvin multiplied by square millimeters. Thermal resistance is a physical quantity characterizing the difficulty of heat transfer between two adjacent surface nodes, with units of Kelvin per watt. To avoid calculating the thermal resistance value with a zero denominator, a minimum positive value of 0.001 Kelvin per watt is preferred. The thermal resistance value is a dimensionless value obtained by normalizing the original thermal resistance value to the median; the higher the value, the more difficult the heat transfer.

[0037] It should be noted that the minimum positive inertia number is a minimum value to avoid the denominator of the elastic compliance original quantity calculation being zero, and is preferably taken as 0.001 gigapascals multiplied by millimeters to the fourth power. The elastic compliance original quantity is a physical quantity characterizing the ease with which the connection structure between two adjacent surface nodes deforms under load. The minimum positive dimensionality number is a minimum value to avoid the denominator of the structural compliance value calculation being zero, and is preferably taken as 0.001 millimeters per Newton. The structural compliance value is a dimensionless value obtained by normalizing the elastic compliance original quantity to the median; the higher the value, the easier the structure is to deform. The minimum positive bending number is a dimensionless minimum value to avoid the denominator of the geometric bending degree calculation being zero, and is preferably taken as 0.0001. The geometric bending degree is a dimensionless value characterizing the degree of bending between two adjacent surface nodes; the higher the value, the more severe the bending. The thermal resistance weight is a dimensionless weighting parameter that adjusts the proportion of the thermal resistance value in the edge composite impedance bending amount, and is preferably taken as 0.4. The elastic compliance weight is a dimensionless weight parameter that adjusts the proportion of structural compliance value in the edge composite impedance bending amount, preferably set to 0.4. The geometric bending weight is a dimensionless weight parameter that adjusts the proportion of geometric bending degree in the edge composite impedance bending amount, preferably set to 0.2. The edge composite impedance bending amount is a dimensionless value generated by weighting the thermal resistance value, structural compliance value, and geometric bending degree; a higher value results in a higher composite impedance. The minimum impedance path is the path with the minimum sum of edge composite impedance bending amounts for all adjacent edges between the source and target surfaces. The dimensionless minimum positive number is a dimensionless minimum value to avoid zero denominators in the composite impedance calculation, preferably set to 0.0001. The composite impedance value characterizes the overall engineering performance of the shell model; a lower value indicates better performance.

[0038] It should be noted that the semantic similarity is calculated using the cosine similarity function to measure the similarity between the semantic embedding vector and the semantic geometric embedding vector. First, the two vectors are L2 normalized, and then the dot product is calculated to obtain the similarity value, which ranges from -1 to 1. During numerical processing, only the positive correlation results are retained, and the negative correlation values ​​are uniformly set to 0.

[0039] It should be noted that the minimum impedance path is solved using Dijkstra's algorithm. The inputs are the surface boundary topology, source surface nodes, target surface nodes, and the combined impedance bending of adjacent edges. The outputs are the minimum impedance path and the total impedance value. The algorithm initializes with the cumulative impedance of the source surface nodes as 0. Each time, it selects the unvisited node with the minimum cumulative impedance, updates the cumulative impedance of adjacent nodes and the predecessor node, and terminates when the target surface node is visited or all reachable nodes have been traversed. Finally, the minimum impedance path is obtained by backtracking through the predecessor nodes.

[0040] It should be noted that the adjustment rules for the weight parameters are as follows: the global weights of the thermal path and the force path are summed to 1. In scenarios prioritizing heat dissipation, the weight of the thermal path can be adjusted to 0.6 to 0.7, and in scenarios prioritizing vibration resistance, the weight of the force path can be adjusted to 0.6 to 0.7. The weights of thermal resistance, flexibility, and bending are summed to 1. In high-power scenarios, the weight of thermal resistance can be adjusted to 0.5 to 0.6, and in high-vibration scenarios, the weight of flexibility can be adjusted to 0.5 to 0.6. The threshold values ​​for the source surface and the target surface are in the range of 0.2 to 0.8, and can be adjusted according to the number of functional surfaces required.

[0041] In one embodiment of the present invention, the specific calculation steps of the fifth module include: The formula for calculating model quality is as follows: in For model quality, For material density, For volume calculation function, For surface boundary solid models; The formula for calculating the dimensionless mass constraint value is as follows: in This is a dimensionless mass constraint value. For model quality, For the maximum permissible mass, It is a dimensionless positive number of mass. The formula for calculating modal frequencies is as follows: in For modal frequencies, For modal analysis functions, For surface boundary solid models, The elastic modulus of the material. For material density, For fixed areas; The formula for calculating the dimensionless frequency constraint value is as follows: in The value is a dimensionless frequency constraint. This is the lower limit of the first-order natural frequency. For modal frequencies, It is a positive number with minimal frequency dimensions; The formula for calculating the finite difference gradient is as follows: in For finite difference gradients, This is the combined impedance value for positive disturbances. This represents the combined impedance value for negative disturbances. For finite difference perturbation step size, The parameter is a unit vector in the direction of the parameter. The formula for calculating the parameter update step size coefficient is as follows: in Update the step size coefficients for the parameters. Let L be the L2 norm of the difference between the upper bound and the lower bound of the parameter. For parameter dimensions, The 2-norm of the finite difference gradient vector, It is a positive number that minimizes the gradient; The formula for calculating the target geometric parameter vector is as follows: in For the target geometric parameter vector, For feasible region projection operators, For center point parameters, Update the step size coefficients for the parameters. It is a finite difference gradient vector.

[0042] It should be noted that the volume calculation function is used to calculate the volume of the 3D solid model, which can be achieved through the volume calculation function of the computer-aided design geometry kernel. The model mass is the total mass of the surface boundary solid model, obtained by multiplying the material density by the model volume. The dimensionlessly small positive mass number is a minimum value to avoid the denominator being zero in the calculation of the dimensionless mass constraint value; a value of 0.001 grams is preferred. The dimensionless mass constraint value is the model mass normalized to the maximum allowable mass; a value ≤1 satisfies the mass constraint requirements. The modal analysis function is a finite element analysis function for calculating the first natural frequency of the shell structure, preferably using a linear perturbation solver. The modal frequency is the first natural frequency of the surface boundary solid model, characterizing the vibration resistance of the shell structure. The dimensionlessly small positive frequency number is a minimum value to avoid the denominator being zero in the calculation of the dimensionless frequency constraint value; a value of 0.1 Hz is preferred.

[0043] It should be noted that the dimensionless frequency constraint value is the value obtained after normalizing the model modal frequencies to the lower limit of the first-order natural frequency; a value ≤1 satisfies the vibration resistance constraint requirements. The positive perturbation composite impedance value is the model composite impedance value corresponding to the positive perturbation parameters. The negative perturbation composite impedance value is the model composite impedance value corresponding to the negative perturbation parameters. The finite difference gradient is an approximation of the partial derivative of the composite impedance value with respect to a single geometric parameter dimension, characterizing the degree and direction of the parameter's influence on the composite impedance value. The parameter update step size coefficient is a coefficient controlling the parameter update amplitude, ensuring that the update amplitude matches the parameter space scale. The gradient minimum positive number is a minimum value to avoid the denominator of the parameter update step size coefficient calculation being zero; a value of 0.001 per millimeter is preferred. The feasible region projection operator is a function that projects the parameter vector into the comprehensive feasible region, ensuring that the updated parameters satisfy all constraints. The target geometric parameter vector is the optimal geometric parameter vector obtained after optimization and feasible region projection.

[0044] It should be noted that the modal analysis function is implemented using the finite element method, and the solver is a linear perturbation frequency solver; the boundary conditions are to apply completely fixed constraints to the surface of the fixed region, and the remaining surfaces are free boundaries; the mesh generation uses tetrahedral second-order elements, the global element size is 0.5 times the minimum wall thickness, and the stress concentration region is refined to 0.3 times the global size; after solving, the first natural frequency is extracted as the modal frequency output.

[0045] It should be noted that mass and frequency constraints are hard constraints, and their projection priority is second only to the fixed region and prohibited region constraints. When the mass exceeds the limit, parameters that significantly affect the mass, such as wall thickness, rib height, and rib height, should be reduced first. When the frequency does not meet the limit, parameters that significantly affect the stiffness, such as wall thickness, rib width, and rib height, should be increased first. When the two constraints conflict, the frequency constraint takes priority, while ensuring that the model mass does not exceed 1.1 times the maximum allowable mass.

[0046] In one embodiment of the present invention, the specific calculation steps of the sixth module include: The formula for calculating the path impedance contribution rate is as follows: in Contribution to path impedance The bending amount of the combined impedance of a single, independently existing edge. This is the sum of the combined impedance deflections of all sides within the path of minimum impedance. It is a positive number representing the minimum impedance; The formula for calculating the final instruction sequence is as follows: in This is the final instruction sequence. For large model modules, For the input text requirement, For functional semantic feature set, For the physical constraint boundary set, For the target geometric parameter vector, For the path with minimum impedance, Contribution to path impedance; The calculation formula for generating the final solid model by calling the geometry kernel is as follows: in For the final physical model, For geometric kernel, This is the final instruction sequence.

[0047] It should be noted that the minimum positive impedance number is preferred to be 0.0001 to avoid a dimensionless minimum value where the denominator of the path impedance contribution rate calculation is zero. The path impedance contribution rate is the proportion of the edge-combined impedance bending amount of a single adjacent edge in the total impedance value of the path with the minimum impedance, representing the degree of contribution of that adjacent edge to the total path impedance. The large model module is a large language model module that generates the final auxiliary design construction sequence. It is consistent with the architecture of the large model generation function, and the input includes the minimum impedance path and path impedance contribution rate information. The final instruction sequence is the final version of the auxiliary design construction sequence, generated by the optimal geometric parameters and performance bottleneck information, and can be executed to generate the final solid model. The final solid model is a three-dimensional boundary representation solid model obtained after full-process optimization, which meets all design requirements and constraints.

[0048] It should be noted that the prompt template content of the large model module is a performance bottleneck optimization module added on the basic template. It includes three parts: core minimum impedance path description, core bottleneck edge description, and bottleneck optimization instructions. It clearly requires that the optimization operation must not change the core parameter values, must not violate the constraints, and must reduce the corresponding type of impedance in a targeted manner.

[0049] It should be noted that the optimization direction for thermal resistance bottleneck is to increase the wall thickness at the corresponding location, widen the rib connection, and increase the transition fillet; the optimization direction for flexibility bottleneck is to increase the reinforcing ribs, increase the fillet, and locally thicken the shell; the optimization direction for bending bottleneck is to adjust the direction of the ribs, increase the transition fillet, and reduce sharp bend surfaces.

[0050] In one embodiment of the present invention, the specific calculation steps of the seventh module include: The final formula for calculating the overall impedance value is as follows: ; in This is the final composite impedance value. For the comprehensive impedance accounting system function, For the final physical model, For functional semantic feature set, For the physical constraint boundary set; The final formula for calculating model quality is as follows: ; in For the quality of the final model, For material density, For volume calculation function, For the final physical model; The formula for calculating the final modal frequency is as follows: ; in For the final modal frequency, For modal analysis functions, For the final physical model, The elastic modulus of the material. For material density, For fixed areas; The formula for calculating the impedance improvement ratio is as follows: in For impedance improvement ratio, The composite impedance value corresponding to the center point parameter. This is the final composite impedance value. The improvement rate is a very small positive number; The formula for calculating the final output set is as follows: in For the final output set, For the final physical model, This is the final instruction sequence. For the target geometric parameter vector, This is the final boundary topology diagram. This is the final composite impedance value. For impedance improvement ratio, For the quality of the final model, For the final modal frequency, Contribution to total path impedance.

[0051] It should be noted that the comprehensive impedance calculation system function is the same function used to calculate the comprehensive impedance value of the final physical model, and is completely consistent with the comprehensive impedance value calculation method. The final comprehensive impedance value is the dimensionless comprehensive impedance value corresponding to the final physical model, characterizing the comprehensive engineering performance of the final design scheme. The final model mass is the total mass value corresponding to the final physical model. The final modal frequency is the first-order natural frequency value corresponding to the final physical model. The minimum positive improvement rate is a dimensionless minimum value to avoid the denominator of the impedance improvement ratio calculation being zero, and is preferably taken as 0.0001. The impedance improvement ratio is the relative improvement magnitude of the comprehensive impedance value, quantifying the performance improvement effect of the final design scheme relative to the baseline model. The final output set is a structured output set containing the entire data chain of the final design, representing a complete engineering design deliverable.

[0052] It should be noted that the final output set is delivered in four folders: 3D model files, design parameter files, performance analysis files, and executable code files. All files use industry-standard formats and can be directly imported into mainstream CAD / CAE software.

[0053] In one embodiment of the present invention, a large-model-driven industrial product design process optimization method includes the following steps: Step S1: Parse the semantic feature set and physical constraint boundary set of the input text extraction function; Step S2: Generate a geometric parameter vector based on the functional semantic feature set and the physical constraint boundary set, and extract the center point parameters of the geometric parameter vector to generate a candidate set of difference parameters; Step S3: Convert the candidate set of differential parameters into a surface boundary entity model, extract the surface nodes and adjacent edges of the surface boundary entity model to construct a surface boundary topology graph, and calculate the average adjacent edge length. Step S4: Match functional surfaces in the surface boundary topology diagram, calculate the heat transfer resistance, structural flexibility and geometric bending between functional surfaces, and calculate the comprehensive impedance value in combination with the average adjacent edge length. Step S5: Calculate the model quality and modal frequencies of the surface boundary entity model by combining the physical constraint boundary set, and obtain the finite difference gradient of the center point parameters based on the comprehensive impedance value. Generate the target geometric parameter vector based on the finite difference gradient under the constraints of model quality and modal frequency. Step S6: Decompose the minimum impedance path and path impedance contribution rate from the comprehensive impedance value, and generate the final instruction sequence and final entity model by combining the target geometric parameter vector. Step S7: Recalculate the final composite impedance value, final model quality, and final modal frequency of the final solid model. Compare the final composite impedance value with the composite impedance value to obtain the impedance improvement ratio. Output the final instruction sequence, final solid model, final composite impedance value, final model quality, final modal frequency, and impedance improvement ratio.

[0054] Specifically, the actual deployment of this invention can be divided into two modes: offline deployment and cloud deployment. Offline deployment adopts a local workstation deployment method, which is equipped with a graphics processor with no less than 16G of video memory, 32G of RAM and 2T of storage space, and installs an adapted computer-aided design geometry kernel, finite element analysis solver and pre-trained large language model. Cloud deployment adopts a containerized deployment method, which provides elastic computing resources through cloud servers and provides design services to the outside world through interface calls, supporting concurrent access by multiple users.

[0055] Specifically, the data acquisition of this invention is divided into three stages. The first stage is design requirement data acquisition, which collects the user's natural language design requirements through a text input interface, and supports the import of 3D design space models, 3D files of fixed areas and prohibited areas. The second stage is basic parameter data acquisition, which collects material properties such as thermal conductivity, elastic modulus, and density through a preset material library, relevant parameters of manufacturing constraint sets through a manufacturing process library, and performance thresholds such as maximum allowable mass and lower limit of first-order natural frequency through a design specification library. The third stage is model feature data acquisition, which extracts geometric and topological features such as surface area, normal vector, centroid coordinates, local thickness, and adjacency relationships from the generated 3D model through a geometric kernel, without the need for additional manual measurement and annotation.

[0056] The final output of this invention is a complete set of deliverables, comprising four parts: The first part is an editable parametric 3D model, which can be directly imported into mainstream computer-aided design software for secondary editing and modification; the second part is a list of optimal geometric parameters, including the final values ​​of eight core geometric parameters and explanations of constraint fulfillment; the third part is a complete performance analysis report, including the final integrated impedance value, the impedance improvement ratio relative to the baseline model, mass and modal frequency verification results, and explanations of core performance bottleneck location and optimization; the fourth part is executable modeling code, which can be directly run to generate the corresponding 3D model. For example, for the design of a motor driver housing for an industrial mobile robot, after inputting natural language design requirements, the final model output by this invention has a 32% lower integrated impedance value compared to the baseline model, a housing mass of 285 grams that meets the maximum allowable mass requirement of 300 grams, and a first-order natural frequency of 135 Hz that meets the lower limit requirement of 120 Hz. The generated model can be directly used for mold manufacturing.

[0057] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0058] 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. A large-model-driven industrial product design process optimization system, characterized in that, include: The first module analyzes the semantic feature set and physical constraint boundary set of the input text extraction function. The second module generates a geometric parameter vector based on the functional semantic feature set and the physical constraint boundary set, and extracts the center point parameters of the geometric parameter vector to generate a candidate set of difference parameters. The third module converts the differential parameter candidate set into a surface boundary entity model, extracts the surface nodes and adjacent edges of the surface boundary entity model to construct a surface boundary topology graph, and calculates the average adjacent edge length. The fourth module matches functional surfaces in the surface boundary topology diagram, calculates the heat transfer resistance, structural flexibility and geometric bending between functional surfaces, and calculates the comprehensive impedance value by combining the average adjacent edge length. The fifth module calculates the model quality and modal frequencies of the surface boundary entity model by combining the physical constraint boundary set, and obtains the finite difference gradient of the center point parameters based on the comprehensive impedance value. Under the constraints of model quality and modal frequency, the target geometric parameter vector is generated based on the finite difference gradient. The sixth module decomposes the minimum impedance path and path impedance contribution rate from the comprehensive impedance value, and generates the final instruction sequence and final entity model by combining the target geometric parameter vector. The seventh module recalculates the final synthesized impedance value, final model quality, and final modal frequency of the final solid model. It compares the final synthesized impedance value with the synthesized impedance value to obtain the impedance improvement ratio, and outputs the final instruction sequence, final solid model, final synthesized impedance value, final model quality, final modal frequency, and impedance improvement ratio.

2. The large-model-driven industrial product design process optimization system according to claim 1, characterized in that, Extract functional roles, corresponding functional roles, numerical information, functional semantic embedding vectors, corresponding functional semantic embedding vectors, functional anchors, and spatial decay scales from the input text. Extract the role weight vector corresponding to the functional role and the corresponding role weight vector of the functional role respectively. Then calculate the dot product of the transpose of the role weight vector corresponding to the functional role and the functional semantic embedding vector, and calculate the exponential function value of the dot product with the natural constant as the base. By combining the comparison traversal number and the total number of functional semantic elements, the dot product of the transpose of the role weight vector corresponding to the comparison functional role and the comparison functional semantic embedding vector is calculated. Then, the exponential function value of the dot product with the natural constant as the base is calculated and accumulated to generate the sum. Divide the exponential function value of the dot product (base of the natural constant) by the sum to obtain the extraction function weight; Functional semantic elements are generated by combining functional roles, numerical information, functional semantic embedding vectors, functional anchors, spatial decay scales and functional weights, and then all functional semantic elements are merged to generate a functional semantic feature set. Extract the design space shape envelope, material thermal conductivity, material elastic modulus, material density, fixed area, prohibited area, manufacturing constraint set, maximum allowable mass and lower limit of first-order natural frequency, and merge them to generate a physical constraint boundary set.

3. The large-model-driven industrial product design process optimization system according to claim 1, characterized in that, Combine the shell wall thickness, heat dissipation fin height, heat dissipation fin spacing, reinforcing rib width, reinforcing rib height, position parameters, corner radius and opening size to generate a geometric parameter vector; By combining the functional semantic feature set, the physical constraint boundary set, and numerical information, the parameter boundary generation function is used to obtain the lower bound and upper bound of the parameters for a single dimension. Extract the lower bound of the parameter for a single dimension and add it to the upper bound of the parameter for a single dimension to obtain the sum. Then divide the sum by two to obtain the center point parameter for a single dimension. Extract the upper bound of the parameter for a single dimension and subtract the lower bound of the parameter for a single dimension to obtain the range value. Then extract twice the value of the parameter dimension and divide the range value by twice the value to obtain the finite difference perturbation step size for a single dimension. Along the unit vector of the parameter direction, first extract the multidimensional center point parameter combination and add the finite difference perturbation step size for a single dimension to generate positive perturbation parameters. Then extract the multidimensional center point parameter combination and subtract the finite difference perturbation step size for a single dimension to generate negative perturbation parameters. Combine the multidimensional center point parameter combination, positive perturbation parameters and negative perturbation parameters to generate a candidate set of difference parameters. By combining the surface boundary entity model and the manufacturing constraint function, the following constraints are imposed: the geometric parameter vector is greater than or equal to the lower bound of the parameter for a single dimension and less than or equal to the upper bound of the parameter for a single dimension; the surface boundary entity model is contained within the shape envelope of the design space; the intersection of the surface boundary entity model and the prohibited area is an empty set; the fixed area is contained within the surface boundary entity model; and the manufacturing constraint function is less than or equal to zero. A comprehensive feasible region is then established.

4. The large-model-driven industrial product design process optimization system according to claim 1, characterized in that, Input the input text requirements, functional semantic feature set, physical constraint boundary set and candidate parameters into the large model generation function to obtain the auxiliary design construction sequence; input the auxiliary design construction sequence into the geometric kernel calculation function to calculate and generate the surface boundary solid model; Extract the set of face nodes and the set of adjacent edges contained in the face boundary entity model, and then construct the face boundary topology graph; extract the face area, face normal vector, face centroid coordinates, face local thickness, face curvature features, face type label and the set of adjacent faces for the face nodes in the face node set, and then input them into the face feature encoding function to generate semantic geometric embedding vectors; For adjacent edges in the adjacent edge set, extract the equivalent path length, connection width and inter-face continuity factor; extract the local thickness of the first face node and the local thickness of the second face node, add them together to obtain the sum, and then divide the sum by two to obtain the average local thickness. The equivalent heat transfer cross section is calculated by multiplying the connection width, average local thickness and inter-surface continuity factor. Extract the section correction factor, then multiply the connection width by the cube of the average local thickness, multiply the result by the section correction factor and divide by twelve to obtain the equivalent bending section moment of inertia. For the location combination of adjacent face nodes, first accumulate the equivalent path length, then divide the accumulated result by the total number of adjacent edges to obtain the average adjacent edge length.

5. The large-model-driven industrial product design process optimization system according to claim 1, characterized in that, Divide the area of ​​a face by the sum of the median of the face areas in the set of face nodes and the smallest positive number of face areas to obtain the normalized area value. For the target functional role, first compare the normalized area value with one and extract the minimum value, then construct the area adjustment function; For functional semantic elements that satisfy functional roles, firstly, cosine similarity is calculated by combining functional semantic embedding vector and semantic geometric embedding vector and the maximum non-negative value is extracted. Then, the exponential function value with the natural constant as the base is calculated by combining the surface centroid coordinates, functional anchor points and spatial decay scale. Finally, the surface matching score is obtained by performing a product of functional weight, maximum non-negative value, exponential function value with the natural constant as the base and area adjustment function. The surface matching score belonging to the heat source surface attribute is compared with the source surface selection threshold to locate the surface nodes contained in the heat source surface set. Then, the surface matching score belonging to the heat dissipation target surface attribute is compared with the target surface selection threshold to locate the surface nodes contained in the heat dissipation target surface set. The surface matching score belonging to the load input surface attribute is compared with the source surface selection threshold to locate the surface nodes contained in the load input surface set. Then, the surface matching score belonging to the support output surface attribute is compared with the target surface selection threshold to locate the surface nodes contained in the support output surface set. The thermal path strength is obtained by multiplying the source surface thermal matching score by the target surface heat dissipation matching score, and the force path strength is obtained by multiplying the source surface load matching score by the target surface support matching score. The numerator is obtained by multiplying the thermal path strength by the global weight of the thermal path and the force path strength by the global weight of the force path, and summing the results. For the combination of traversal of the source and target surfaces, the product of the thermal path strength calculated by traversal and the global weight of the thermal path, and the product of the force path strength calculated by traversal and the global weight of the force path are summed and the minimum positive number of the weight is added to obtain the denominator. The numerator is then divided by the denominator to obtain the path weight.

6. The large-model-driven industrial product design process optimization system according to claim 5, characterized in that, The thermal resistance is calculated by dividing the equivalent path length by the sum of the product of the material's thermal conductivity and the equivalent heat transfer cross section and the minimum positive number of heat transfer. Then, the median of the thermal resistance corresponding to the traversed adjacent edges is extracted, and the thermal resistance is divided by the sum of the median of the thermal resistance and the minimum positive number of the thermal resistance dimension to obtain the heat transfer resistance value. The original elastic flexibility is calculated by dividing the cube of the equivalent path length by the sum of the product of the material's elastic modulus and the equivalent bending section's moment of inertia and the minimum positive number of inertia. Then, the median of the original elastic flexibility corresponding to the traversed adjacent edges is extracted, and the original elastic flexibility is divided by the sum of the median of the original elastic flexibility and the minimum positive number of the flexibility dimension to obtain the structural flexibility value. Subtract the dot product of the surface normal vector of the first face and the surface normal vector of the second face from the first face, and square the difference to generate the first operator. Then divide the equivalent path length by the average adjacent edge length, square the difference, and add the minimum positive number of bend to generate the second operator. Divide the first operator by the second operator to obtain the geometric bend. Extract the product of heat transfer resistance value and thermal resistance weight, the product of structural flexibility value and elastic flexibility weight, and the product of geometric bending degree and geometric bending weight, and then sum the three to calculate the edge composite impedance bending amount; for the path traversal from the source surface to the target surface, find the path with the minimum sum of edge composite impedance bending amounts corresponding to adjacent edges in the traversal path to construct the minimum impedance path. The impedance numerator is obtained by multiplying the edge combined impedance deflection amount within the path with the path weight and then summing the results globally. The equivalent path lengths are summed within the path with the minimum impedance. The summation result is divided by the average adjacent edge length and a dimensionless minimum positive number is added to obtain the reduction result. The reduction result is multiplied by the path weight and summed to obtain the impedance denominator. The impedance numerator is divided by the impedance denominator to calculate the comprehensive impedance value.

7. The large-model-driven industrial product design process optimization system according to claim 1, characterized in that, Input the surface boundary solid model into the volume calculation function to calculate the volume, and then multiply the material density by the calculation result to obtain the model mass; The dimensionless mass constraint value is obtained by dividing the model mass by the sum of the maximum permissible mass and the smallest positive number of the mass dimension. The modal frequencies are obtained by inputting the surface boundary solid model, material elastic modulus, material density, and fixed region into the modal analysis function; then the dimensionless frequency constraint value is obtained by dividing the lower limit of the first natural frequency by the sum of the modal frequency and the smallest positive number of the frequency dimension. Extract the difference between the positive disturbance integrated impedance value and the negative disturbance integrated impedance value, and then divide the difference by twice the finite difference disturbance step size to calculate the finite difference gradient. Extract the L2 norm of the difference between the upper and lower bounds of the parameters, and then divide it by twice the product of the parameter dimension, the L2 norm of the finite difference gradient vector, and the sum of the minimum positive gradient number to obtain the parameter update step size coefficient. Under the constraints that the dimensionless mass constraint value is less than or equal to one and the dimensionless frequency constraint value is less than or equal to one, the parameter update step size coefficient is first multiplied with the finite difference gradient vector, and the result of the multiplication is subtracted from the center point parameter to generate the iterative value. Then, the iterative value is input into the feasible region projection operator and projected onto the constraints in the comprehensive feasible region to generate the target geometric parameter vector.

8. The large-model-driven industrial product design process optimization system according to claim 1, characterized in that, Extract the bending amount of the combined impedance of a single independent edge; extract the sum of the bending amounts of all edges in the path with the minimum impedance, and add the minimum positive impedance number to generate the total resistance. Divide the bending amount of the combined impedance of a single independent edge by the total resistance to obtain the path impedance contribution rate. The input text requirements, functional semantic feature set, physical constraint boundary set, target geometric parameter vector, minimum impedance path and path impedance contribution rate are input into the large model module to obtain the final instruction sequence. The geometric kernel is then called to load the final instruction sequence and calculate to generate the final entity model.

9. The large-model-driven industrial product design process optimization system according to claim 1, characterized in that, The final solid model, functional semantic feature set, and physical constraint boundary set are input into the comprehensive impedance accounting system function to generate the final comprehensive impedance value; the final solid model is input into the volume accounting function, and the accounting result is multiplied by the material density to calculate the final model mass; For the final solid model, the material elastic modulus, material density, and fixed region are input into the modal analysis function to calculate the final modal frequencies; Extract the comprehensive impedance value corresponding to the center point parameter and subtract the final comprehensive impedance value to obtain the improvement difference. Then, add the minimum positive number of the improvement rate to the comprehensive impedance value corresponding to the center point parameter to obtain the base value. Finally, divide the improvement difference by the base value to calculate the impedance improvement ratio. The final output set is generated and output by merging the final entity model, final instruction sequence, target geometric parameter vector, final surface boundary topology, final integrated impedance value, impedance improvement ratio, final model quality, final modal frequency and total path impedance contribution rate.

10. A large-model-driven optimization method for industrial product design processes, characterized in that, Implementing the large-model-driven industrial product design process optimization system as described in any one of claims 1 to 9 includes the following steps: Step S1: Parse the semantic feature set and physical constraint boundary set of the input text extraction function; Step S2: Generate a geometric parameter vector based on the functional semantic feature set and the physical constraint boundary set, and extract the center point parameters of the geometric parameter vector to generate a candidate set of difference parameters; Step S3: Convert the candidate set of differential parameters into a surface boundary entity model, extract the surface nodes and adjacent edges of the surface boundary entity model to construct a surface boundary topology graph, and calculate the average adjacent edge length. Step S4: Match functional surfaces in the surface boundary topology diagram, calculate the heat transfer resistance, structural flexibility and geometric bending between functional surfaces, and calculate the comprehensive impedance value in combination with the average adjacent edge length. Step S5: Calculate the model quality and modal frequencies of the surface boundary entity model by combining the physical constraint boundary set, and obtain the finite difference gradient of the center point parameters based on the comprehensive impedance value. Generate the target geometric parameter vector based on the finite difference gradient under the constraints of model quality and modal frequency. Step S6: Decompose the minimum impedance path and path impedance contribution rate from the comprehensive impedance value, and generate the final instruction sequence and final entity model by combining the target geometric parameter vector. Step S7: Recalculate the final composite impedance value, final model quality, and final modal frequency of the final solid model. Compare the final composite impedance value with the composite impedance value to obtain the impedance improvement ratio. Output the final instruction sequence, final solid model, final composite impedance value, final model quality, final modal frequency, and impedance improvement ratio.

Images