CAD large model method and system based on AI and vector clusters

Abstract

The invention discloses a CAD large model method and system based on AI and a vector cluster, and belongs to the field of computer aided design, and the method comprises the following steps: S1, constructing a manifold space mathematical model, and mapping the three-dimensional information of an object to a linear space through the vector cluster; s2, collecting multi-industry drawing data, and constructing a multi-industry drawing data set; s3, pre-training the multi-industry drawing data set through the manifold space mathematical model, and extracting a feature template tensor corresponding to each industry drawing; s4, on the basis of the feature template tensor, coding, decoding and circularly optimizing and updating parameters by adopting a large model embedded into an M-MoE framework; and S5, mapping the manifold tensor output in the previous step to a three-dimensional space, generating a CAD drawing and a report, and circularly optimizing an output result through a loss function. According to the method, a CAD large model is constructed based on artificial intelligence and a vector cluster mathematical theory, and intelligent drawing generation, automatic auditing, language interactive editing and full-life-cycle management are achieved.

Description

Technical Field

[0001] The present invention relates to the field of computer-aided design, and in particular to a CAD large model method and system based on artificial intelligence (AI) and vector bundle mathematical theory, which is used to realize intelligent drawing generation, automatic review, language interactive editing and full life cycle management functions. Background Art

[0002] Currently, the core of CAD drawing software is parametric modeling, which includes polygonal mesh modeling, free-form surface modeling, parametric feature modeling and other methods. Among them, entering the production and manufacturing stage mainly relies on parametric feature modeling.

[0003] The core of parametric feature modeling in CAD drawing software consists of a geometric kernel and a parametric solver. The core of the geometric kernel is a library of mathematical expression functions and geometric algorithms. These include the closed-source geometric kernels of Siemens' Para-Solid and Dassault's ACIS, as well as the open-source geometric kernel of France's OCC. However, these CAD drawing software products still have the following flaws: 1. Lack of intelligent reasoning ability: Existing CAD drawing software has no thinking and reasoning ability, cannot make rule judgments and review two-dimensional and three-dimensional drawings, and is still at the lower-level drawing tool stage.

[0004] 2. Insufficient interactive capabilities: Existing CAD drawing software does not have language dialogue editing functions, and users need to manually modify and edit drawings.

[0005] 3. Unable to automatically generate drawings and reports: Existing CAD drawing software does not have the ability to automatically generate drawings and reports.

[0006] The above defects lead to the review, editing and generation of drawings requiring a lot of manpower and material resources, high cost, low efficiency and high error rate. It is urgent to improve and innovate the existing CAD drawing software to provide more intelligent CAD drawing design solutions. Summary of the Invention

[0007] The purpose of the present invention is to overcome the problems existing in the existing CAD drawing design, and provide a CAD large model method and system based on AI and vector bundles.

[0008] The object of the present invention is achieved through the following technical solutions: In a first aspect, a CAD large model method based on AI and vector bundle is provided, comprising the following steps: S1. Construct a mathematical model of manifold space and map the three-dimensional information of the object to the linear space through vector bundles; S2. Collect drawing data from multiple industries and build a multi-industry drawing dataset; S3. Pre-training the multi-industry drawing dataset using the manifold space mathematical model to extract feature template tensors corresponding to the drawings of each industry; S4. Based on the feature template tensor, a large model embedded in the M-MoE architecture is used to perform encoding, decoding, and cyclic optimization and parameter update; S5. Map the manifold tensor output in the previous step to three-dimensional space, generate CAD drawings and reports, and optimize the output results through a loss function loop.

[0009] In some embodiments, mapping the three-dimensional information of the object to a linear space via a vector bundle includes: The structure, annotation, production information and life cycle data of the object are represented by vector bundles, and the local homeomorphism of the manifold space is mapped to the linear Euclidean space R n .

[0010] In some embodiments, collecting drawing data from multiple industries and constructing a multi-industry drawing data set includes: Extract component features using a few-shot detection algorithm; The Hungarian algorithm is used to perform differential comparison of CAD drawings and establish a feature dataset.

[0011] In some embodiments, step S3 includes: Substituting the feature data set into the manifold space mathematical model for pre-training, a homology template tensor, a three-dimensional structure template tensor corresponding to the i-th level vector bundle, and a property template tensor corresponding to the i-th level vector bundle are obtained.

[0012] In some embodiments, the large model includes an encoder, an M-MoE layer, and a decoder, wherein the encoder uses four layers of multi-head attention to calculate the feature template tensor and applies 48 layers of blocks; the M-MoE layer includes 128 M-MoE architectures, each of which includes 1 shared expert and 128 routing experts.

[0013] In a second aspect, a CAD large model system based on AI and vector bundles is provided, comprising: Manifold space mathematical model construction module, used to construct the manifold space mathematical model and map the three-dimensional information of the object to the linear space through the vector bundle; Drawing dataset construction module, used to collect drawing data from multiple industries and construct drawing datasets for multiple industries; A feature template tensor extraction module is used to pre-train the multi-industry drawing dataset using the manifold space mathematical model to extract the feature template tensor corresponding to the drawings of each industry; A large model optimization module is used to perform encoding, decoding, and cyclic optimization and parameter update based on the feature template tensor using a large model embedded in the M-MoE architecture; The CAD drawing generation module is used to map the manifold tensor output by the large model optimization module into three-dimensional space, generate CAD drawings and reports, and optimize the output results through a loss function loop.

[0014] It should be further explained that the technical features corresponding to the above embodiments can be combined or replaced with each other to form a new technical solution if there is no conflict.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention builds a mathematical model based on manifold space and vector bundle theory, combining multi-industry datasets with a large-scale model of the M-MoE architecture to achieve automatic drawing generation, intelligent review, interactive language editing, and full lifecycle management. This improves the intelligence and interactive capabilities of CAD drawing software, significantly improving efficiency and reducing costs.

[0016] 2. The present invention pre-trains the multi-industry drawing dataset through the manifold space mathematical model to extract the feature template tensors corresponding to the drawings of each industry. A few-sample detection algorithm is used to extract component features, and a Hungarian algorithm is used to perform differentiated comparison of CAD drawings, thereby reducing computing power and saving costs, while taking into account both accuracy and computing speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flow chart of a CAD large model method based on AI and vector bundles is shown in an embodiment of the present invention; Figure 2 A schematic diagram illustrating a mathematical model of a manifold space according to an embodiment of the present invention; Figure 3 A schematic diagram of the overall architecture of a large model shown in an embodiment of the present invention; Figure 4 A schematic diagram of generating a large model report according to an embodiment of the present invention; Figure 5 Generate a schematic diagram for a large model drawing shown in an embodiment of the present invention; Figure 6 This is a schematic diagram of the design functions of a large model shown in an embodiment of the present invention. Figure 7 A schematic diagram of a tensor at a manifold point according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical solutions of the present invention are described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings herein can be arranged and designed in various different configurations. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0019] It should be noted that the defects existing in the solutions in the above-mentioned prior art are the results obtained by the inventor after practice and careful research. Therefore, the discovery process of the above-mentioned problems and the solutions proposed in the embodiments of this application below for the above-mentioned problems should be the contributions made by the inventor to this application in the process of invention and creation, and should not be understood as technical contents known to technical personnel in this field.

[0020] In response to the technical problems pointed out in the background technology, the embodiments provided by the present invention are as follows: In an exemplary embodiment, referring to Figure 1 This paper provides a large-scale CAD modeling method based on AI and vector bundles, introducing new mathematical tools to place all information—including 2D or 3D CAD text, drawings, and data—in manifold space. Because real designs cannot be fully represented in linearized space, and given that computers can only calculate linear numbers like 0 and 1, it is necessary to map the manifold space locally homeomorphically to linear Euclidean space. Therefore, it is necessary to find theoretical tools to homeomorphize manifold space to Euclidean space. This paper uses vector bundle theory and manifold tensor theory in manifold space to address the complex representation problem of drawing.

[0021] Step S1 specifically includes: The structure, annotation, production information and life cycle data of the object are represented by vector bundles, and the local homeomorphism of the manifold space is mapped to the linear Euclidean space R n According to the conclusion of the object representation in the three-dimensional world, the spatial information of the object's three-dimensional space is constructed into a manifold basis space M, which is inversely mapped to the vector bundle E1 (representing the three-dimensional structure of the object), and then the vector bundle E1 is smoothly mapped to the vector bundle E2 (representing all the annotation information such as the object rules), and then the vector bundle E2 is smoothly mapped to the vector bundle E3 (representing the manufacturing and construction information of the object), and then the vector bundles E1 and E3 are smoothly mapped to the vector bundle E4 (representing the full life cycle information of the object). The vector bundle E i (The value of i can be increased and adjusted according to the situation) is homeomorphic to the linear Euclidean space R n The mathematical model of the manifold space constructed is as follows Figure 2 In practical applications, according to the requirements of the representation of object drawings and reports, the vector bundle tool of the manifold space is selected, and the M manifold basis space is used as the vector bundle space Ei The basis of , the vector bundle can be operated as follows: Direct sum operation, scalar multiplication operation, existence of dual space, tensor product operation, algebraic geometry operation and topological operation.

[0022] If the above operation rules are met, the inner product dot multiplication can be performed, which can be directly used as the input of various models. Define the local coordinate card ₯ (p, Ψ) of the manifold space and map it to R n The space is transformed by the general linear group GL(n), where P is any manifold point in the popular space, Φ3(π -1 ),Φ2(π -1 ),Φ1(π -1 ) is the bundle space mapping shown, and finally transformed to Euclidean space through the general linear group.

[0023] The tensor on the manifold space is a partial linear function θ, which represents the (r, s) type tensor at the manifold point P. The tensor is represented as follows Figure 7 shown.

[0024] Among them, T P *M represents the dual tensor at the manifold point P, T P M represents the tensor at the manifold point P. This (r, s) type tensor satisfies the vector bundle operation rules and can be used for subsequent data analysis and processing.

[0025] Furthermore, the E1 space is used as the function of the geometric kernel and the geometric algorithm representation space to meet the two-dimensional or three-dimensional representation of the object world.

[0026] Step S2 specifically includes: We collect basic text data and drawing data from dozens of industries, including the oil and gas chemical industry, biopharmaceutical industry, tobacco industry, automotive industry, semiconductor electronics industry, robotics and drone industry, industrial design industry, e-cigarette industry, spare parts industry, and construction machinery industry, and store them in a local database. Using vector database tools and feature extraction tool models as needed, we construct feature datasets for each industry, combining the multimodal object drawing information and full lifecycle data in the database.

[0027] Due to format differences among the aforementioned industry databases, detailed information in the databases is interpreted to obtain a feature dataset. A few-shot detection algorithm is used to extract component features, including component identification, line clustering, and legend matching, to automatically identify key components in drawings. The few-shot detection algorithm includes the following steps: (1) Generate region candidates for CAD drawings based on a deep feature extraction network and construct a feature embedding space between component classes using a small sample support set; (2) Using the metric learning framework, each candidate region is mapped into a high-dimensional feature vector, and the few-shot classification is achieved by matching the distance function to obtain the category labeling and location information of each component in the drawing; (3) Further construct a connectivity structure graph for the identified component set, and apply a density clustering algorithm to spatially cluster the lines in the graph to identify the connection relationship; (4) Combined with the legend template in the graphic library, the legend symbols in the drawing are accurately matched based on image matching and geometric correction strategies.

[0028] Furthermore, a differential comparison of CAD drawings is performed using an improved Hungarian algorithm to establish a feature data set. The improved Hungarian algorithm includes the following steps: (1) Convert the two CAD drawings to be compared into graph structures and ,in Represents a collection of component nodes, Represents the topological connection between components; (2) Calculate the matching cost between each pair of nodes. The cost function comprehensively considers the visual feature distance, spatial position deviation and local topological structure similarity. Specifically, the cost matrix can be defined as follows: in is an adjustable weighting parameter, represents the Euclidean distance of the component visual embedding, is the node position coordinate difference, Jaccard distance representing the local topology of a node; (3) Improve the traditional Hungarian algorithm by introducing a fault-tolerant mechanism, including adding empty nodes to match missing or newly added components to avoid unreasonable forced matching; dynamically adjusting the cost threshold to adapt to component layout changes and annotation differences; and combining graph convolutional neural networks to calculate node embedding features to improve topological matching accuracy; (4) Finally, the optimal matching pair is output and the set of unmatched nodes is extracted as the candidate component pairs of the differences between the drawings, and the differential feature map is automatically generated.

[0029] Step S3 specifically includes: Based on the vector bundle and tensor theory of manifold space, the vector bundle E is represented by four sets of vector bundle bases of manifold basis space M. i The tensor of CAD information of the homologous object, the i-th vector bundle, and the i-th level corresponding representation.

[0030] After the manifold space mathematical model is established, each feature data set established in step S2 is substituted into the manifold space mathematical model pre-training to obtain a homologous template tensor, a corresponding i-th vector bundle, and a corresponding i-th level corresponding representation.

[0031] The overall structure of the large model of the present invention is as follows Figure 3 As shown, it includes an encoder, M-MoE layer, and decoder. In step S3, multiple template datasets are pre-trained. The large model inputs multimodal CAD drawing information (images, text, IoT data, etc.) and the feature dataset established in step S2. The output model is converted into four corresponding vector bundles.

[0032] The core of the large model's encoder is the improved Transformer. The encoder uses four layers of multi-head attention to calculate the feature template tensor and applies 48 layers of blocks. It outputs three four-level tensors that serve as input for the next decoding layer. Specifically, the multiple tensors obtained in step S3 are fed into the encoder for homology, structure, and physical and chemical property comparison learning. The encoder cyclically adjusts the parameters of these multiple tensors, generating new tensors that serve as input to the M-MoE layer.

[0033] The large model adds 128 M-MoE layers. Each MoE architecture includes one shared expert and 128 routing experts. Each blueprint token activates multiple routing experts, building a multi-agent system that shares rules, knowledge, and blueprints across various industries. In practice, the number of expert layers is adjusted based on industry needs, allowing different industries to call upon expert agents from different industries. Each agent represents the knowledge and expertise corresponding to each industry, and calling upon an expert agent invokes the corresponding judgment rules and specifications.

[0034] Step S6 specifically includes: The decoder uses the multiple tensors output by the M-MoE layer as input, performing homology adjustment, structural adjustment, and physical and chemical property adjustment learning. It repeatedly adjusts the parameters of these multiple tensors, generating multiple new manifold space tensors that serve as input for the next step of mapping into a three-dimensional linear space. Specifically, the tensor output from step S5 is used as input to update the structure of each of the four levels. Using physical and chemical properties as constraints (e.g., minimizing potential energy and optimizing stability at each level), the geometric kernel and constraints are updated, and the parameters are updated through an eight-block optimization cycle.

[0035] Furthermore, in step S7, the manifold space tensor output in the previous step is projected onto the three-dimensional space structure data and the data of related functions, converted into two-dimensional or three-dimensional drawings and reports, and then looped and optimized. The three-dimensional structure output is optimized by the minimum loss function, and the loop is continuously performed. In this embodiment, the loop is designed to be repeated for 5 rounds, and the encoding and decoding process is executed in a loop. After multiple adjustments, end-to-end two-dimensional or three-dimensional drawings and reports are output. In addition, a language interaction mechanism is introduced in the loop optimization process, allowing users to interactively edit and fine-tune the generated drawings and reports through natural language, such as adjusting structural details, modifying dimensioning, or adding explanatory text, thereby realizing human-computer collaborative optimization of design output. CAD industry reports generated based on large CAD models, such as Figure 4 As shown, the drawings generated based on the CAD large model are as follows Figure 5 shown.

[0036] In another exemplary embodiment, a CAD large model system based on AI and vector bundle is provided, comprising: Manifold space mathematical model construction module, used to construct the manifold space mathematical model and map the three-dimensional information of the object to the linear space through the vector bundle; Drawing dataset construction module, used to collect drawing data from multiple industries and construct drawing datasets for multiple industries; A feature template tensor extraction module is used to pre-train the multi-industry drawing dataset using the manifold space mathematical model to extract the feature template tensor corresponding to the drawings of each industry; A large model optimization module is used to perform encoding, decoding, and cyclic optimization and parameter update based on the feature template tensor using a large model embedded in the M-MoE architecture; The CAD drawing generation module is used to map the manifold tensor output by the large model optimization module to three-dimensional space, generate CAD drawings and reports, and optimize the output results through a loss function loop; in addition, the CAD drawing generation module also includes a language interaction sub-module, which is used to interactively edit the generated CAD drawings and reports based on natural language instructions, to realize the modification and adjustment of drawing structure, dimensions, symbol annotations and other contents, thereby improving human-computer interaction efficiency and design flexibility.

[0037] The large model platform and design functions built based on this system are as follows: Figure 6 As shown, the system can automatically generate and review drawings and reports, as well as edit them interactively using language. This reduces costs and improves efficiency. Furthermore, the system's large model platform enables production management.

[0038] The above specific implementation methods are detailed descriptions of the present invention. It cannot be considered that the specific implementation methods of the present invention are limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, they can make several simple deductions and substitutions without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A CAD large model method based on AI and vector bundle, characterized in that: The following steps are involved: S1. Construct a mathematical model of manifold space and map the three-dimensional information of the object to the linear space through vector bundles; S2. Collect drawing data from multiple industries and build a multi-industry drawing dataset; S3. Pre-training the multi-industry drawing dataset using the manifold space mathematical model to extract feature template tensors corresponding to the drawings of each industry; S4. Based on the feature template tensor, a large model embedded in the M-MoE architecture is used to perform encoding, decoding, and cyclic optimization and parameter update; S5. Map the manifold tensor output in the previous step to three-dimensional space, generate CAD drawings and reports, and optimize the output results through a loss function loop.

2. The CAD large model method based on AI and vector bundle according to claim 1 is characterized in that: Mapping the three-dimensional information of the object to the linear space through the vector bundle includes: The structure, annotation, production information and life cycle data of the object are represented by vector bundles, and the local homeomorphism of the manifold space is mapped to the linear Euclidean space R n .

3. The CAD large model method based on AI and vector bundle according to claim 1 is characterized in that: The process of collecting drawing data from multiple industries and constructing a drawing data set for multiple industries includes: Extract component features using a few-shot detection algorithm; The Hungarian algorithm is used to perform differential comparison of CAD drawings and establish a feature dataset.

4. The CAD large model method based on AI and vector bundle according to claim 3 is characterized in that: The step S3 comprises: Substituting the feature data set into the manifold space mathematical model for pre-training, a homology template tensor, a three-dimensional structure template tensor corresponding to the i-th level vector bundle, and a property template tensor corresponding to the i-th level vector bundle are obtained.

5. The CAD large model method based on AI and vector bundle according to claim 1 is characterized in that: The large model includes an encoder, an M-MoE layer, and a decoder. The encoder uses four layers of multi-head attention to calculate the feature template tensor and applies 48 layers of blocks. The M-MoE layer includes 128 M-MoE architectures, each of which includes 1 shared expert and 128 routing experts.

6. A CAD large model system based on AI and vector bundle, characterized by: include: Manifold space mathematical model construction module, used to construct the manifold space mathematical model and map the three-dimensional information of the object to the linear space through the vector bundle; Drawing dataset construction module, used to collect drawing data from multiple industries and construct drawing datasets for multiple industries; A feature template tensor extraction module is used to pre-train the multi-industry drawing dataset using the manifold space mathematical model to extract the feature template tensor corresponding to the drawings of each industry; A large model optimization module is used to perform encoding, decoding, and cyclic optimization and parameter update based on the feature template tensor using a large model embedded in the M-MoE architecture; The CAD drawing generation module is used to map the manifold tensor output by the large model optimization module into three-dimensional space, generate CAD drawings and reports, and optimize the output results through a loss function loop.

Images