Metal fence drawing intelligent generation system

Abstract

This invention discloses an intelligent drawing generation system for metal fences, comprising a user input module, a data preprocessing and vectorization module, a cross-modal encoder, a parametric generator, a manufacturing rule base and constraint solver, a hierarchical mechanical verifier and local optimizer, and an export and integration interface. This application achieves industrial deployability through a unified multimodal intermediate representation, pre-processing and structural constraints to the parametric generator, hierarchical rapid mechanical verification and local optimization closed-loop, and the use of low-rank fine-tuning methods such as LoRA / PEFT, thereby significantly improving the efficiency, accuracy, and manufacturability of drawing generation.

Description

Technical Field

[0001] This invention relates to the field of intelligent engineering drawing generation technology, and more specifically to an intelligent drawing generation system for metal fences. Background Technology

[0002] The efficiency and quality of design drawings for components such as fence structures directly affect product delivery time and manufacturing costs. Currently, related drawing generation technologies are mainly implemented through the following three paths:

[0003] (1) Traditional manual CAD design process: Based on customer needs, designers manually draw two-dimensional construction drawings and three-dimensional component assembly drawings of the fence structure in general CAD platforms such as AutoCAD and SolidWorks, and then hand them over for processing and assembly after completing annotations, BOM list and process description. This method requires a lot of manual intervention for designs containing complex decorative elements or non-standard components, and is time-consuming and iterative.

[0004] (2) Parametric / template-driven design: Manufacturers or design teams pre-build fence parameter families, including post diameter, spacing, crossbar cross section, and decorative patterns, etc. Designers can quickly generate variants of different specifications and export engineering drawings by modifying key parameters. This method performs well in the design of standardized and serialized products and can significantly shorten the drawing time. However, the template library has limited coverage and lacks in-depth verification of manufacturing processes and structural strength.

[0005] (3) Rule-driven and semi-automated systems; some advanced systems use rule engines to encode manufacturing process constraints such as minimum bending radius, weld gap, and standard part selection rules into post-check items, providing verification prompts or semi-automatic corrections after drawing generation. In addition, some studies have attempted to introduce AI to achieve mapping from sketches or text to draft drawings. However, the following obvious shortcomings still exist: First, the multimodal input processing capability is weak: traditional template and rule systems have difficulty receiving and understanding multimodal input forms such as natural language, hand-drawn sketches and historical CAD files at the same time and with high quality, which means that manual conversion and manual proofreading are still required in the design process; Secondly, manufacturing feasibility and structural verification are delayed: the generation process often places manufacturing process and structural verification as independent steps after output. The generator itself does not embed process and mechanical constraints, which may result in outputting drawings that are "similar in form but not in manufacture", requiring rework or manual modification. This increases the verification cost before manufacturing and extends the delivery cycle. Secondly, AI solutions are disconnected from industrial-grade engineering: Most AI research solutions remain at the level of visual fitting or pattern generation, lacking the ability to couple with enterprise manufacturing preferences, standard parts libraries, and processing equipment capabilities (such as bending machines and welding process parameters), resulting in the generated results being unable to be directly implemented in terms of manufacturability and cost. Finally, the engineering deployment threshold is high: mature large models have high resource requirements when deployed in engineering scenarios. If low-rank fine-tuning or lightweight strategies are not adopted, it is difficult to deploy them universally in small and medium-sized enterprises or local server environments.

[0006] In summary, how to deeply integrate multimodal input understanding, manufacturing process constraints, and structural strength closed-loop verification during the design and generation phase, and provide a lightweight and feasible industrial deployment solution, is a technical challenge that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of the above problems, the present invention is proposed to provide a smart metal fence drawing generation system that overcomes or at least partially solves the above problems.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides an intelligent metal fence drawing generation system, comprising: The input module is used to receive multimodal design data input by the user; The preprocessing and vectorization module is used to extract semantic information and geometric elements from the multimodal design data, obtain the corresponding constraint specifications, and jointly map them into a structured intermediate representation. A cross-modal encoder is used to align the structured intermediate representation with text geometric features and generate a conditional vector; A parameterization generator is used to match and generate initial design parameters from a preset parameter family library based on the condition vector, and to generate a three-dimensional geometric model of the fence structure based on the initial design parameters. The manufacturing rule base and constraint solver are used to call the constraint solver and use the preset manufacturing process rule base to verify and correct the manufacturability of the initial design parameters, and generate the first design parameters that meet the process constraints. The hierarchical mechanical verification and optimization unit is used to perform closed-loop verification and optimization of the structural mechanical performance of the three-dimensional geometric model corresponding to the first design parameters, and generate second design parameters that simultaneously meet the process constraints and mechanical performance indicators. An export and integration interface is used to output engineering deliverables based on the second design parameters.

[0009] Preferably, the multimodal design data includes at least one or more of the following: natural language description, hand-drawn sketches / images, and existing computer-aided design (CAD) files.

[0010] Preferably, when the multimodal design data contains multiple types of data, they are fused according to a preset priority rule, wherein the priority from high to low is as follows: structural parameters explicitly specified by the user, geometric topology information parsed from CAD files, primitive information of hand-drawn sketches, and semantic information described in natural language. When different types of data conflict, the highest priority information source is retained, and the stability of the design results is ensured through consistency checks.

[0011] Preferably, the structured intermediate representation includes at least a geometric element table, a part attribute table, a constraint set, and style / function tags.

[0012] Preferably, the cross-modal encoder performs domain adaptation through low-rank adaptive LoRA or parameter-efficient fine-tuning PEFT; and includes a text branch and a geometric branch, wherein the text branch and the geometric branch achieve text geometric feature alignment through cross attention.

[0013] Preferably, it also includes a condition vector and task management module, which is used to cache condition vectors and task status, and is responsible for version control and audit log management.

[0014] Preferably, the generation of the first design parameters that satisfy the process constraints includes: The constraint solver formalizes the rules in the manufacturing process rule base into algebraic equations and / or inequality constraints, and uses a hybrid solution strategy to solve the initial design parameters; if the initial design parameters violate strong constraints, they are corrected by penalty function method or projection method, or a priority backoff strategy is triggered.

[0015] Preferably, closed-loop verification and optimization of structural mechanical properties are performed, including: Based on the three-dimensional geometric model corresponding to the first design parameters, a simplified rod or shell element model is constructed for linear static rapid analysis to identify hot spots where stress or deformation exceeds a preset threshold. The hotspot region is meshed and a high-precision finite element analysis solver is called for accurate calculation. If the accurate calculation result does not meet the preset mechanical performance requirements, a local optimizer is triggered to optimize the three-dimensional geometric model. After optimization, the model is verified through closed-loop iteration using the manufacturing rule base and constraint solver.

[0016] Preferably, the local optimizer optimizes the three-dimensional geometric model, including: Density topology optimization based on sensitivity analysis is performed on the hotspot region to achieve material distribution optimization at the micro level; and cross-sectional dimensions or support arrangement optimization based on parameter search is performed on the global structure to achieve structural parameter optimization at the macro level; and the optimized topological features are mapped into parameterizable engineering features to reconstruct a three-dimensional geometric model that meets manufacturing process requirements.

[0017] Preferably, it also includes a user interaction and iteration recording module, used to record the user's acceptance, fine-tuning or regeneration of the generated results, forming labeled sample data for model iterative training.

[0018] This invention provides an intelligent metal fence drawing generation system, which aims to significantly improve the efficiency, accuracy, and manufacturability of drawing generation by using a unified multimodal intermediate representation, pre-parameterizing manufacturing process and structural constraints to a parametric generator, hierarchical rapid mechanical verification and local optimization closed loop, and employing low-rank fine-tuning methods such as LoRA / PEFT to achieve industrial deployability.

[0019] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: 1. Through a unified multimodal IR and cross-modal coding mechanism, the designer's natural language description, hand-drawn sketches and historical CAD files are directly and accurately mapped to engineering parameters, which greatly shortens the human-computer interaction and information translation process, significantly improves the speed of first draft generation and reduces communication errors.

[0020] 2. By adopting low-rank fine-tuning schemes such as LoRA / PEFT, high-performance cross-modal understanding and generation capabilities can be deployed with limited GPU memory and computing power, lowering the threshold for enterprises to go to the cloud or local deployment, thereby expanding the application scenarios and popularity of the system.

[0021] 3. By embedding manufacturing process rules and structural mechanics constraints into the parametric generator and solver in the form of mathematical constraints, candidate designs that are unmanufacturable or do not meet strength standards can be eliminated or automatically corrected during the generation stage. This avoids the rework and additional costs caused by the traditional process of "generating first and then verifying", thereby generating direct savings in time and production costs.

[0022] 4. The layered mechanical verification strategy ensures both engineering safety and computational efficiency, quickly filters and performs high-precision solutions only on hot spots, ensuring reliable structural evaluation results even in resource-constrained environments.

[0023] 5. Based on ezdxf's programmatic DXF output and Web plugin display, the generated drawings can be directly used for material cutting, bending, and assembly, thus achieving a true "from idea to manufacturing" closed loop. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the intelligent metal fence drawing generation system provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] This invention discloses an intelligent metal fence drawing generation system, which tightly couples multimodal input understanding, parametric generation, manufacturing process constraint solving, and structural verification and optimization into an end-to-end automated closed loop. Furthermore, it achieves industrial-grade deployment and direct manufacturing delivery through lightweight large model adaptation and engineering export interfaces.

[0028] The overall implementation of the system consists of nine interconnected modules: a user input module, a data preprocessing and vectorization module, a cross-modal encoder, a parameterization generator, a manufacturing rule base and constraint solver, a hierarchical mechanics checker, and a local optimizer. Preferably, it also includes a condition vector pool and task management module, an export and integration interface, and a user interaction and iteration recording module.

[0029] In one specific embodiment, based on Figure 1 The flowchart in the document describes each module and its implementation details step by step. Figure 1 middle," This indicates that the parameterized generator, manufacturing rule base, constraint solver, hierarchical mechanics verifier, and local optimizer form a closed loop and can be iterated. Specifically, In one embodiment, the input module is configured to receive multimodal design data input by a user; In this embodiment, the multimodal design data includes at least one or more of the following: natural language descriptions, hand-drawn sketches / images, and existing computer-aided design (CAD) files. In one implementation, the natural language is parsed using an NLU module based on a pre-trained large model (combined with an entity dictionary from the engineering domain); the sketches are processed through an image processing pipeline (U-Net performs pixel-level semantic segmentation → Hough / vectorization algorithm converts the contours into line segments / arcs → dimension annotations are recognized); and the CAD files are parsed using OpenCASCADE or the CAD SDK to parse geometric entities and attributes.

[0030] In a preferred embodiment, when the multimodal design data contains multiple data types, a format validation and priority fusion strategy is executed. Specifically, in this embodiment, after receiving CAD files (including DWG, DXF, STEP, IGES, etc.) uploaded by the user, the system first parses the geometric entities and attributes using OpenCASCADE or the enterprise's internal CAD SDK, and simultaneously performs format validity and consistency checks to ensure the engineering quality and stability of the input data. This includes systematically verifying whether the file structure is complete, whether geometric entities are degraded or distorted, whether topological relationships are closed, whether units and coordinate systems are consistent, and whether semantic attributes are missing. For missing or abnormal data, the system will automatically execute repair or completion strategies and retain the original attributes for traceability.

[0031] During the fusion of multi-source design information, the fusion is performed according to a preset priority rule: the structural parameters explicitly input by the user have the highest priority, followed by the geometric and topological structures reflected in the CAD files, then the information expressed in sketches, textual semantic descriptions, and the system's reasoning and completion results. When there are conflicts between different information sources, the system will prioritize retaining the data source with the strongest engineering constraints and ensure the stability and reliability of the design results through consistency checks and version rollback mechanisms.

[0032] Through the above-mentioned format validation and priority fusion strategy, the system can still accurately reproduce the true design intent in complex design input environments and robustly support the subsequent parametric modeling of fence structures and automatic drawing generation process.

[0033] In one embodiment, the preprocessing and vectorization module is used to extract semantic information and geometric elements from the multimodal design data, and obtain corresponding constraint specifications, which are then mapped together into a structured intermediate representation (IR, JSON Schema). In this embodiment, all modules use IR as the exchange format to ensure module substitutability and traceability. The IR includes a geometric element table (line segments, arcs, surfaces, volumes), a part attribute table (material, thickness, processing technology), a constraint set (minimum bending radius, hole spacing, tolerance, etc.), and style / functional tags. IR emphasizes the reversible mapping between geometry and semantics to facilitate manual review and iterative training. In this invention, data from text descriptions, design sketches / CAD drawings, and industry standards are not directly processed in the generated model, but are first uniformly mapped into an intermediate representation (IR, defined using JSON Schema).

[0034] In the data mapping stage, this application employs a "three-stage mapping mechanism" to construct the intermediate representation (IR). First, in the geometric analysis stage, the input file is topologically decomposed using a CAD parser or image vectorization algorithm to extract line segments, arcs, surfaces, and solid structures, generating a geometric element table while preserving topological relationships and coordinate system information. Second, in the semantic analysis stage, a domain text understanding model is used to structure natural language descriptions and style keywords, extracting material types, dimensional intentions, functional uses, and style tags to generate a part attribute table and a style / functional tag set. Finally, in the constraint matching and normalization stage, the system performs conditional retrieval and rule matching from the constraint database based on the identified material categories, component types, and structural forms, transforming the matched process constraints and structural specifications into unified formal constraint expressions and writing them into the constraint set field.

[0035] The above three-stage mapping results are uniformly encapsulated into a structured intermediate representation (IR) (defined using JSON Schema), which includes four core parts: geometric element table, part attribute table, constraint set, and style / function tags. A bidirectional index mapping relationship is established between geometric elements and semantic tags to support manual review, version tracking, and subsequent model training.

[0036] The material and process constraints in this invention are mainly obtained from three types of data resources: (1) national and industry standards (such as bending radius, weld type, hole diameter and spacing restrictions related to metal structure processing); (2) the long-term accumulated process database of the enterprise (such as minimum forming radius corresponding to different material thicknesses, recommended welding parameters, etc.); (3) historical project parameters and user feedback data that are continuously accumulated during the operation of the system.

[0037] During the data access phase, the system will perform unified modeling and version management of the above-mentioned standard information. That is, after modeling, it will be stored as a rule entry database. The rules are defined in the form of parameter ranges, logical expressions or algebraic equations / inequalities, and an association index with material type, component category and processing method will be established. This will form a "material-process-structure constraint library" that can be understood and reasoned by machines, and serve as an important part of the constraint set in IR.

[0038] To ensure the transparency and verifiability of the technical implementation, the system will simultaneously display the extracted geometric entities and attribute information in the preprocessing and vectorization modules, including line segments, arcs, surfaces, solid structures, material types, thickness information, and associated processing technologies. This allows designers to promptly verify whether the system's understanding aligns with their design intentions. Figure 1 To.

[0039] Furthermore, in terms of the mapping implementation, this invention adopts a three-stage process of "geometric analysis, semantic analysis, and constraint normalization": First, the geometric analysis submodule performs topological and geometric decomposition of the CAD file based on OpenCASCADE, generating a node-based geometric element table while maintaining the topological relationships unchanged; second, the semantic analysis submodule uses a domain text understanding model to perform structured analysis of natural language descriptions and style keywords, forming style / functional labels and parameter dictionaries; finally, the constraint normalization submodule transforms material, process, and specification data into formal constraint expressions (such as "R ≥ k·t", "hole spacing ≥ Pmin", "weld grade ≥ W2", etc.), and uniformly incorporates them into the constraint set of IR. The system ensures a reversible mapping relationship between geometric elements and semantic labels when constructing IR, enabling the design results to have both structured expressive capabilities and support for manual review and subsequent continuous model training and optimization.

[0040] In one embodiment, a cross-modal encoder is used to align the structured intermediate representation with text geometric features and generate conditional vectors to ensure an accurate mapping from description to parameters. In this embodiment, the cross-modal encoder adopts a dual-branch encoder architecture: the text branch uses a pre-trained large model (e.g., a visual-language large model chosen in practice), and the geometry branch uses an image / vector encoder (CNN + point / line / topological feature extractor or graph neural network-based encoder). Constraint specifications participate in parameter generation and solving as independent structured constraint channels. Alignment between text and geometric features is achieved through multiple cross-attention layers, outputting a conditional vector. Preferably, the encoder can adopt a Transformer-based architecture, or a convolutional + RNN hybrid structure, a graph neural network (GNN) to encode vectorized geometry, or a CNN-based feature pyramid plus an attention fusion unit to achieve feature alignment.

[0041] Furthermore, to facilitate engineering implementation, this invention prioritizes low-rank fine-tuning using LoRA / PEFT on the encoder, keeping the main model weights unchanged and training only a small number of adaptation layers. This significantly reduces training / deployment costs and enables domain adaptation even with limited GPU memory. Preferably, the LoRA rank is between 8 and 16, and training employs mixed precision (FP16 forward, FP32 gradient) and the AdamW optimizer combined with warmup and cosine decay. In some implementations, Adapter or BitFit can also be chosen as fine-tuning strategies, or only the final classification / mapping layer can be fine-tuned.

[0042] As a preferred implementation, a condition vector pool and a task management module are set up to cache condition vectors and task states. At the same time, this module supports parallel tasks, asynchronous queues and long task scheduling (e.g., FEA simulation may be a long task); and is responsible for version control (condition vector, parameter family version) and audit logs to meet the needs of engineering change management.

[0043] Furthermore, in one embodiment, a parametric generator, as the core of the system, is used to generate initial design parameters from a preset parameter family library based on the condition vector, and to generate a three-dimensional geometric model of the fence structure based on the initial design parameters. In one implementation, the parametric generator consists of a parameter family library (predefined component families: columns, crossbars, connectors, decorative patterns, etc.) and a mapping engine. The mapping engine transforms the condition vector (based on the parameter family library) into a unified vector space through embedding mapping, performs similarity matching with the standardized parameter family models in the parameter family library to obtain candidate parameter solutions, and instantiates the geometric model according to parametric rules.

[0044] Unlike existing template-based methods, the parameterized generator calls the manufacturing rule base and constraint solver in real time during the instantiation process to verify and correct parameters, thereby ensuring that the instantiation results are feasible at the manufacturing level.

[0045] In one embodiment, a manufacturing rule base and a constraint solver are used to call the constraint solver to perform manufacturability verification and correction on the initial design parameters using a preset manufacturing process rule base, thereby generating first design parameters that satisfy process constraints.

[0046] In this embodiment, the manufacturing process rule library includes minimum bending radius, weld allowance, standard part model and matching tolerance, material processing limit, cutting and stamping tolerance, etc., which are derived from national industry standards, enterprise processing specifications and historical project data. The constraint solver formalizes process rules into algebraic equations / inequalities and employs a hybrid solution strategy. This includes using heuristic search (genetic algorithm, particle swarm optimization) or constraint programming to find parameter solutions that satisfy the rules, or using a pre-trained learner (neural network regressor) as an approximate solver to quickly provide a set of feasible parameters. Alternatively, equality constraints can be solved using the Lagrange multiplier method or nonlinear least squares, while inequality constraints can be solved using the penalty function method or projection method. Specifically, the system substitutes candidate parameters into the constructed algebraic equations / inequalities, solving and verifying geometric feasibility, manufacturability, and structural safety one by one. If the initial design parameters violate strong constraints, they are corrected using the penalty function, projection method, or Lagrange method, or a priority backoff strategy is triggered to reselect parameters. Ultimately, only the set of design parameters that satisfies all strong constraints and is optimal under the comprehensive objective function is retained. In this invention, the manufacturing rule base and constraint solver make manufacturing process constraints hard constraints or real-time checkpoints for the generator, rather than post-verification.

[0047] In one embodiment, a hierarchical mechanical verification and optimization unit is used to provide a structural verification closed loop during the generation process; specifically, for the three-dimensional geometric model corresponding to the first design parameters, closed-loop verification and optimization of the structural mechanical performance are performed to generate second design parameters that simultaneously meet process constraints and mechanical performance indicators; the steps include: The three-dimensional geometric model corresponding to the first design parameter is simplified into a rod or shell element model and subjected to linear static rapid analysis. After identifying hot spots, the mesh of the hot spot region is refined, and a high-precision finite element analysis solver (open source such as CalculiX or commercial such as Abaqus) is called for accurate calculation. If the strength / deflection does not meet the set threshold (e.g., maximum allowable deflection of 5 mm, safety factor ≥ 2.0), the local optimizer is triggered to optimize the three-dimensional geometric model. After optimization, the closed-loop iterative verification is performed through the manufacturing rule base and constraint solver.

[0048] It should be noted that the present invention can also employ a fast structural evaluator based on machine learning (training a neural network to predict stress and deflection), or use empirical formulas / engineering verification tables as a preliminary screening method, and then call a more refined finite element analysis solver when necessary, thus providing a functionally equivalent fast safety determination when computational resources are limited.

[0049] The local optimizer provides two types of strategies: Density topology optimization based on sensitivity analysis is performed on the hotspot region to achieve microscopic material distribution optimization. For hotspot regions with prominent local stress, density topology optimization discretizes the region into finite element units, using minimizing structural flexibility or maximizing stiffness as the objective function and the overall material volume fraction as a constraint. Formal models such as SIMP are used to establish the relationship between element density and equivalent elastic modulus. The system calculates the partial derivative of the objective function with respect to element density based on sensitivity analysis, guiding the material to iteratively migrate between elements, thereby gradually forming a more reasonable material distribution path and achieving optimized adjustment of the local microstructure shape.

[0050] The system optimizes the cross-sectional dimensions or support layout of the global structure based on parameter search to achieve macro-level structural parameter optimization. The cross-section / support optimization mainly focuses on macro-structural parameters such as the cross-sectional dimensions of fence posts, the cross-sectional form of crossbars, the support layout, and the support angle. The system uses strength, deflection, mass, or cost as objective functions and material specifications, processing capabilities, and standard specifications as constraints. It uses parameter optimization methods such as gradient optimization, genetic algorithms, or grid search to systematically search the parameter space, thereby obtaining a better combination of structural parameters while meeting the requirements of processes and specifications.

[0051] In this application, two types of optimization methods are applied to different levels of the structural design process, working together to ensure that the generated fence structure has both superior mechanical properties and meets the requirements of actual manufacturing processes.

[0052] Finally, the optimized topological features are mapped to parameterizable engineering features to reconstruct a 3D geometric model that meets manufacturing requirements. Since the density distribution results obtained from topology optimization typically lack clear structural boundaries and are difficult to manufacture directly, the system employs a manufacturable post-processing workflow. First, geometric features are extracted from the topology optimization results to identify the main force paths, high-density regions, and key structural connections. Then, these topological features are mapped to parameterizable engineering features, such as stiffeners, support rods, and cross-sectional reinforcement areas. Next, a parameterized generator is used to reconstruct a structural model that meets standard processing capabilities, and welding, bending, and cutting processes are adapted accordingly. Finally, the reconstructed model is re-entered into the manufacturing rule base and constraint solver for process and specification verification, retaining only design schemes that simultaneously meet manufacturing capabilities and structural performance requirements.

[0053] It is important to note that in this application, the hierarchical mechanical verification and optimizer are not sequential with the manufacturing rule base and constraint solver. Instead, they perform mechanical feasibility verification and performance optimization on the same parametric model after the manufacturing rule base and constraint solver have completed the process and geometric constraint verification. If the optimization results involve adjustments to geometric or structural parameters, the parametric model generated by the parametric generator is updated through a parameter backflow mechanism, and the manufacturing rule base and constraint solver are re-triggered to perform rule verification, thus forming a closed-loop design process.

[0054] In this application, the combination of a parametric generator and a hierarchical mechanical verifier with an optimizer, a manufacturing rule base and a constraint solver can achieve simultaneous structural safety and manufacturability.

[0055] As a preferred embodiment of this invention, the intelligent metal fence drawing generation system of the present invention further includes: an export and integration interface and / or a user interaction and iteration recording module.

[0056] The export and integration interface is responsible for exporting the final qualified model to commonly used manufacturing and design formats: DWG, DXF, STEP, IGES, SVG, and PDF, and generating engineering deliverables such as BOM, blanking / bending instructions, and welding process specifications. Implementation details include: using the ezdxf library to generate DXF files in code on the backend (generating 2D line segments, text annotations, layers, and blocks via Python's ezdxf library), and providing a web plugin (based on WebAssembly or server-side rendering and frontend Three.js / Canvas) to display the generated DXF / DWG in a browser for review and online fine-tuning. The system also provides a RESTful API and CAD plugins (AutoCAD / Inventor / SolidWorks) for enterprise PLM / ERP integration and production deployment, ensuring closed-loop executability from design to manufacturing. This application supports cloud services (GPU elastic scaling) and enterprise local deployment (supporting single-card RTX3090 level servers). Task scheduling uses asynchronous queues (such as Celery / RabbitMQ) to handle long-term simulation and offline training tasks. Version control is used for parameter families, rule bases and model weight management.

[0057] As an optional implementation, you can also directly call the official CAD API, use the Python interface of OpenCascade / FreeCAD to generate STEP / IGES, and then export it as DXF by the CAD software, or use vector rendering based on SVG / Canvas and then convert it to DXF; or use vector rendering based on SVG / Canvas and then convert it to DXF.

[0058] The user interaction and iteration recording module provides an interactive interface that supports three types of user actions: acceptance, local fine-tuning, and regeneration. The system records user actions and converts them into labeled samples to be added to the training pool for offline retraining or online fine-tuning. This module is a key data loop for the system's adaptation and personalization.

[0059] The difference and improvement logic between this application and existing technologies lies in two aspects: First, manufacturing processes and mechanical constraints are pre-defined as solvable constraints within the generator, ensuring that the output is directly manufacturable. Second, through lightweight fine-tuning strategies such as LoRA and the integration of ezdxf automated drawing generation with web plugins, large model capabilities can be deployed on-site or on local servers and directly output engineering drawings for material cutting and processing, achieving an end-to-end automated closed loop from creative text / sketch to manufacturable CAD drawings. These technological differences collectively bring about objective improvements in engineering feasibility and efficiency, thereby meeting the differentiated needs of enterprises in terms of data security, real-time performance, and computing resources.

[0060] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0061] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A metal fence drawing intelligent generation system, characterized in that, include: The input module is used to receive multimodal design data input by the user; The preprocessing and vectorization module is used to extract semantic information and geometric elements from the multimodal design data, obtain the corresponding constraint specifications, and jointly map them into a structured intermediate representation. A cross-modal encoder is used to align the structured intermediate representation with text geometric features and generate a conditional vector; A parameterization generator is used to match and generate initial design parameters from a preset parameter family library based on the condition vector, and to generate a three-dimensional geometric model of the fence structure based on the initial design parameters. The manufacturing rule base and constraint solver are used to call the constraint solver and use the preset manufacturing process rule base to verify and correct the manufacturability of the initial design parameters, and generate the first design parameters that meet the process constraints. The hierarchical mechanical verification and optimization unit is used to perform closed-loop verification and optimization of the structural mechanical performance of the three-dimensional geometric model corresponding to the first design parameters, and generate second design parameters that simultaneously meet the process constraints and mechanical performance indicators. An export and integration interface is used to output engineering deliverables based on the second design parameters.

2. The intelligent metal fence drawing generation system as described in claim 1, characterized in that, The multimodal design data includes at least one or more of the following: natural language descriptions, hand-drawn sketches / images, and existing computer-aided design (CAD) files.

3. The intelligent metal fence drawing generation system as described in claim 1 or 2, characterized in that, When the multimodal design data contains multiple types of data, they are fused according to a preset priority rule. The priority rule, from high to low, is as follows: user-specified structural parameters, geometric topology information parsed from CAD files, primitive information from hand-drawn sketches, and semantic information described in natural language. When different types of data conflict, the highest priority information source is retained, and the stability of the design results is ensured through consistency checks.

4. The intelligent metal fence drawing generation system as described in claim 1, characterized in that, The structured intermediate representation includes at least a geometric element table, a part attribute table, a constraint set, and style / function tags.

5. The intelligent metal fence drawing generation system as described in claim 1, characterized in that, The cross-modal encoder performs domain adaptation through low-rank adaptive LoRA or parameter-efficient fine-tuning PEFT; and includes a text branch and a geometric branch, which achieve text geometric feature alignment through cross attention.

6. The intelligent metal fence drawing generation system as described in claim 1, characterized in that, It also includes a condition vector and task management module, which is used to cache condition vectors and task status, as well as manage version control and audit logs.

7. The intelligent metal fence drawing generation system as described in claim 1, characterized in that, The first design parameters that satisfy the process constraints include: The constraint solver formalizes the rules in the manufacturing process rule base into algebraic equations and / or inequality constraints, and uses a hybrid solution strategy to solve the initial design parameters; if the initial design parameters violate strong constraints, they are corrected by penalty function method or projection method, or a priority backoff strategy is triggered.

8. The intelligent metal fence drawing generation system as described in claim 1, characterized in that, Perform closed-loop verification and optimization of structural mechanical properties, including: Based on the three-dimensional geometric model corresponding to the first design parameters, a simplified rod or shell element model is constructed for linear static rapid analysis, hot spots are identified and the mesh of the hot spots is refined, and a high-precision finite element analysis solver is called for accurate calculation; if the accurate calculation results do not meet the preset mechanical performance requirements, a local optimizer is triggered to optimize the three-dimensional geometric model, and the optimized model is verified through closed-loop iteration using the manufacturing rule base and constraint solver.

9. The intelligent metal fence drawing generation system as described in claim 8, characterized in that, The local optimizer optimizes the three-dimensional geometric model, including: Density topology optimization based on sensitivity analysis is performed on the hotspot region, and cross-sectional dimensions or support arrangement optimization based on parameter search is performed on the global structure; the optimized topological features are then mapped to parameterizable engineering features to reconstruct a three-dimensional geometric model that meets manufacturing process requirements.

10. The intelligent metal fence drawing generation system as described in claim 1, characterized in that, It also includes a user interaction and iteration recording module, which records the user's acceptance, fine-tuning or regeneration of the generated results, forming labeled sample data for model iterative training.

Images