Intelligent mechanical design method and system based on learning-generation AI-AD

Abstract

The invention discloses an AI-AD intelligent machine design method and system based on a learning-generation mode. The method comprises the following steps: S1, enabling an AI-AD intelligent mechanical design system to understand design knowledge at first; s2, compiling a workflow with low codes; s3, compiling a Prompt strategy of the design Agent and a called knowledge base; s4, repeating the steps S1 to S3, and compiling various agents needing to complete various specific design functions; s5, according to the designed various agents, a multi-agent self-organizing framework is utilized, and product creation type design is completed; and S6, according to the generated design result, generating a VBA file and storing the VBA file in the cloud. And S7, the SW establishes a conversation with the agent through a back-end API of the agent, timely reads a VBA code of the cloud for modeling, and outputs a model. According to the invention, a system structure of knowledge processing, cloud data storage, SolidWorks-based secondary development modeling and a human-computer interface module is constructed, Agents are enabled to learn at first, then design is created, and a 3D model meeting user requirements is generated, so that the current situation that knowledge engineering has extremely high requirements for practitioners' admission threshold can be reduced.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and specifically relates to a learning-generative AI-AD intelligent mechanical design method and system. Background Art

[0002] Traditional parametric design tools and feature matching have certain limitations for non-standard design. Most existing technologies are based on matching design based on preset model libraries, which cannot meet the needs of non-standard design and innovative design.

[0003] With the continuous development of LLM, the trend of CAD system from parametric to intelligent system is very strong, but there is a lack of design systems with high intelligence. Therefore, it is particularly important to design a design system with a certain degree of intelligence using large language models.

[0004] The invention with publication number CN118153129A proposes a method, device and medium for generating a three-dimensional workpiece model based on a fine-tuned large model. The generation instruction input by the user is obtained through the plug-in client, and the request is sent to the cloud server. The cloud server performs parsing based on the trained workpiece generation model, generates a three-dimensional workpiece model, and displays it on the terminal device. This method uses a large model to fine-tune all parameters and combines it with sample data training. However, although this method introduces the concepts of large models and cloud servers, and uses multi-domain sample data for model training, its main function is still to generate a model through specific parameters provided by the user (such as size, number of threads, etc.). The role of the large model is limited to parsing these inputs, but it has not made substantial breakthroughs in design innovation, reasoning or truly intelligent modeling.

[0005] The invention with the publication number CN116662628A proposes a method for retrieving a three-dimensional CAD model of a complex thin-walled part by using a convolutional neural network and feature extraction technology. First, the model is normalized, preprocessed and voxelized, and then the geometric features are extracted and classified by using a convolutional neural network, and the model is matched by a two-layer mechanism of global retrieval and local retrieval. The global retrieval preliminarily determines the model category, and the local retrieval achieves the final matching through accurate feature comparison. The adjacency matrix and the optimal matching algorithm are used to accurately retrieve the model, thereby improving the efficiency of model reuse and the automation of the design process. However, the main purpose of this method is to improve the reuse rate of the model, and similarity matching is performed through geometric features. It is suitable for application in scenarios with a relatively rich model library. If the data is insufficient, the retrieval performance of the convolutional neural network will be significantly reduced, and it does not involve how to help designers innovate designs according to new requirements. It cannot perform intelligent reasoning or provide optimization suggestions based on design conditions or functional requirements, and the problem of zooming in and out has not been well solved. The applicability is limited to the retrieval and matching of existing models. This limitation is inadequate when dealing with non-standard designs or innovative designs.

[0006] The invention with publication number CN117235929A proposes a 3D CAD generative design method based on knowledge graph and machine learning. This method combines the knowledge features in product design, uses a multi-view convolutional neural network (MVCNN) for feature extraction, and integrates generative models such as generative adversarial networks (GANs) and variational autoencoders (VAEs). The design data is organized and reasoned through knowledge graphs, and finally the 3D model generation is realized. However, GCN usually relies on point clouds, meshes or voxelized data when processing 3D geometric models. Although these forms can capture the overall structure of the model, GCN may not be able to capture enough details when processing complex and fine geometric features (such as tiny surfaces and pores in workpieces). This may cause the generated model to deviate from the actual design requirements, especially in manufacturing and design scenarios that require high precision, and there may also be deviations in the representation of complex surfaces and internal structures. At the same time, the performance of GCN and MVCNN is highly dependent on the quality and diversity of training data. If the training data is insufficient, the performance of the model may drop significantly when facing unseen design requirements. Especially when dealing with innovative design requirements, the accuracy of the model may not meet expectations.

[0007] Although the above three patents all involve the use of large models, machine learning or convolutional neural network technology, their application scope is mostly concentrated in parametric design, feature matching and automated testing, and they focus more on the knowledge of geometric models. Mechanical CAD is not considered in the entire design and manufacturing environment, resulting in a lot of lack of knowledge and loss of information. Therefore, these methods are lacking in design innovation or the ability to perform intelligent reasoning based on functions and requirements. Therefore, there is an urgent need to focus on intelligent CAD design systems with certain thinking and reasoning capabilities from design principles, design methods to manufacturing knowledge, rather than just repeated operations on geometric models.

[0008] In addition, in the traditional design process, mechanical designers often need to rely on empirical formulas and static charts when performing tasks. For example, in V-belt drive design, designers need to select the appropriate belt type and pulley based on empirical formulas. This process involves manual calculation of multiple parameters such as power, speed, and tension. Due to the large number of design variables and their interdependence, designers often make errors when interpreting complex charts and inputting relevant parameters. In addition, complex calculation steps and the interaction of multiple design conditions increase the risk of errors. Manual calculations are not only prone to design deviations, but also greatly reduce work efficiency;

[0009] Secondly, the traditional manual modeling process is also cumbersome and inefficient. Designers often need to gradually adjust modeling parameters according to design parameters and manually create models in the CAD system. For example, in SolidWorks, manually entering design parameters, adjusting geometric dimensions, setting material properties, etc. is time-consuming and laborious, and there is a risk of human error.

[0010] In addition, the traditional design process lacks effective data traceability and version control mechanisms, and the design change records are incomplete, which often leads to the loss of information or the inability to trace back during the design process. Summary of the invention

[0011] In order to overcome the shortcomings of intelligent CAD systems in the prior art in terms of intelligent reasoning for design innovation, the present invention provides an AI-AD intelligent mechanical design method and system based on learning-generation, adopts full natural language interaction, utilizes the multi-agent underlying architecture of LLM, and through continuous human-computer interaction and context-based reasoning, allows the Agent to first learn and then generate a design and produce a 3D model that meets user requirements. Through this design method and system, the current situation of extremely high entry barriers for practitioners in knowledge engineering can be reduced, making it convenient for designers to develop and apply their own intelligent CAD.

[0012] In order to solve the above technical problems, the present invention provides the following technical solutions: an AI-AD intelligent mechanical design method based on learning-generative method, comprising the following steps:

[0013] S1: For each type of mechanical design knowledge used in product development, including formulas, standards, charts or other invisible design knowledge, it is instilled through the learning agent of the AI-AD intelligent mechanical design system, so that the AI-AD intelligent mechanical design system can first understand the design knowledge;

[0014] S2: After the AI-AD intelligent mechanical design system has completed the learning of the required mechanical design knowledge, it will distinguish various innovative design processes, provide design steps and design calculations, and use a low-code mode to write the workflow, during which code can be added and plug-ins can be called.

[0015] S3: Write the prompt strategy of the design agent and the knowledge base called to ensure that the AI-AD intelligent mechanical design system can be based on the various requirements of the designer (intelligently call the workflow written in S2);

[0016] S4: Repeat the operations of S1-S3 to write various intelligent agents that need to complete various specific design functions;

[0017] S5: Based on the various intelligent agents designed previously, the self-organizing architecture of multiple intelligent agents is used to complete the generative design of products.

[0018] S6: Based on the generated design results and parameters, the 3D model intelligent body needs to generate statements that conform to the SolidWorks macro format, generate files and store them in the cloud, and finally modeling is performed through the interface in SolidWorks.

[0019] Preferably, in S1, the AI-AD intelligent mechanical design system is allowed to understand the design knowledge first in the following cases:

[0020] S11: For formula-based knowledge, designers can directly provide formula images, call OCR to recognize the formulas in the knowledge, convert them into LaTex expressions, and then hand them over to the mature LaTex calculation plug-in deployed in the cloud to complete the calculation;

[0021] S12: If it is chart-based knowledge, the designer only needs high-definition parameters and block or curve information. The AI-AD intelligent mechanical design system can use image recognition technology to process the color blocks, curves and data information in the chart;

[0022] S13: If it is workflow process knowledge, you only need to understand the input and output of each workflow, and the workflow is a black box. When scheduling, you can combine the input and output of the workflow with the prompt strategy of the agent according to the needs of the task, and the scheduling content of the knowledge base. As long as the final requirements are put forward, the LLM of the agent can give the workflow to be executed and generate their execution links. Therefore, process knowledge is mainly determined in the prompt and knowledge base.

[0023] S14: If it is knowledge about design principles, LLM can directly understand this knowledge expressed in natural language, even implicit knowledge; this type of knowledge is expressed in the form of examples in the knowledge base, and when the intelligent agent encounters similar design cases during operation, it will make similar corresponding choices.

[0024] Preferably, the detailed steps of S12 are as follows: the system will first pre-process the design chart, divide the image into color blocks through algorithms such as KMeans clustering, and then use OCR technology to identify the text and coordinate system of the chart; the mapping relationship between the color blocks and corresponding parameters in different charts is not fixed. In the development stage, the system uses sample data for supervised learning, and gradually establishes reliable mapping rules through multiple iterations. The system associates the extracted color point pixel values ​​with the parameter values ​​in the chart through polynomial fitting; through the fitting algorithm, the system can derive the overall equation of the curve from the known data points, so that the designer can directly query the corresponding parameter value of any coordinate point based on the chart; for charts with more curve information, the system uses image segmentation technology to separate different curves in the chart, extract the pixel coordinates of the key color points, and map these pixel coordinates with the actual parameters in the chart. At the same time, after obtaining multiple color points of the curve, the system reconstructs the trend of the curve by connecting these color points, and identifies the changing trend and key points therein. After these data information are trained by the learning model, they will be stored in the cloud as JSON files for users to query at any time.

[0025] Preferably, in step S2, in order to improve the adaptability of the workflow, the system converts fuzzy natural language into specific design steps by analyzing the user's design requirements. Each step can not only call the design formulas and standards in the existing design knowledge base, but also integrate external APIs or cloud plug-ins. This flexibility makes the workflow not only limited to simple design tasks, but also can play a role in complex design scenarios.

[0026] Preferably, the step S3 is divided into the design of prompt writing strategy and calling knowledge base, which are as follows:

[0027] S31: The writing of prompts plays a key role in the subsequent user use of the system. It determines how the intelligent agent converts the natural language input by the user into executable tasks. Through natural language processing technology, the system can recognize semantic structures and convert the user's vague description into specific design parameters. For example, the user may input "design a high-power transmission device". The intelligent agent needs to be able to extract keywords such as "high power" and "transmission device" and map them to relevant entries in the knowledge base; if the user provides preliminary design information but does not have clear detailed parameters, the intelligent agent will gradually refine the user input and obtain more information by asking questions;

[0028] S32: In the AI-AD intelligent mechanical design system, the knowledge base is the core for storing design rules, empirical formulas and standard data. The knowledge base is divided into multiple modules, each of which is responsible for processing different design tasks or design data, ensuring that the intelligent agent can call each module individually to meet the needs of specific tasks. In order to ensure that the knowledge base can respond quickly during the design process, the system can quickly find the required data by preprocessing and classifying the design parameters, standard formulas, etc. in the knowledge base through keyword searching. At the same time, the content in the knowledge base is not static. The system can dynamically update the knowledge base based on user input and feedback to ensure efficient calling and continuous updating of the knowledge base.

[0029] Preferably, in step S5, the functions of each agent are relatively independent, such as parameter design agent and geometric modeling agent, but they need to collaborate to complete the overall design task; in order to allow multiple agents to work together, a self-organizing architecture is constructed, allowing agents to communicate with each other through message passing, data sharing and dialogue interfaces. The core of the architecture is to ensure that each agent can share data seamlessly and automatically initiate dialogue requests for collaboration when necessary. When the geometric modeling agent encounters a design conflict, it can communicate with the parameter design agent to coordinate the best solution. This kind of dialogue can not only transmit data, but also exchange reasoning and decision information.

[0030] Preferably, in step S6, two different parts are divided into standard parts and non-standard parts, and there are steps and methods for classification:

[0031] S61: If the designed parts are standard or commonly used parts, such as springs, pulleys, bolts or standard gears, there will be standard drawing data and VBA conversion statements of the basic version in the cloud in the knowledge base. You only need to tell the intelligent agent to replace the original design parameters to generate a VBA text file of the design results, read it into SolidWorks, and you can generate the required 3D model;

[0032] S62: If it is a non-standard or commonly used part, generating VBA text is more complicated:

[0033] S621: First, a feature function unit that meets various functional requirements in the part is generated, and the generated data of the model, such as position, sketch plane, shape parameters and other information, is used as the attributes of the feature unit;

[0034] S622: Complete the design of connection units between characteristic functional units based on design knowledge, such as the selection of shaft shoulders, shaft rings, and shaft sleeve structures during shaft design;

[0035] S623: Connection unit structure generation and VBA text generation;

[0036] S624: Check the feasibility of the parts.

[0037] Preferably, all the VBA code texts generated in S6 are stored in the cloud, the structured data of the special diagnostic function units and the connection units are also stored in the cloud, and the historical files of the creation process are also stored in the cloud at the same time. These data can be read and called by the functions and intelligent entities in the SolidWorks secondary development interface.

[0038] An AI-AD intelligent mechanical design system is applied to an AI-AD intelligent mechanical design method based on learning-generative approach.

[0039] The beneficial effects of the present invention are as follows:

[0040] 1. Through a number of innovative technologies such as automated design, VBA code generation for synchronous modeling, embedded dialogue interface, and cloud-based traceability of design data, it solves many limitations in the existing mechanical design process, especially in reducing design complexity, improving accuracy and automation level, and shows significant advantages;

[0041] 2. By using the mechanical design knowledge base and natural language processing technology, the automatic analysis and precise reasoning of the user's fuzzy requirements are realized. The intelligent agent can accurately calculate the complex parameters involved in the design task (such as power, speed, tension, etc.) by calling the industry standard formula in the knowledge base, ensuring that every step in the design process complies with engineering specifications. This mechanism of automated reasoning and parameter calculation greatly reduces the calculation errors caused by human factors in the design process. The image recognition and analysis function can also solve the problem of inaccurate interpretation of charts;

[0042] 3. By calling the knowledge base, generating VBA code and seamlessly integrating with SolidWorks API, real-time modeling of design parameters is realized. The intelligent agent can not only automatically generate 3D models, but also reduce human intervention in the modeling process, greatly shortening the design cycle while ensuring the accuracy of the design results;

[0043] 4. By integrating an embedded dialogue interface, designers can adjust design parameters in real time through interaction with the intelligent agent. When designers need to modify design requirements, there is no need to manually redo the design. The intelligent agent can receive input in real time, automatically regenerate the design plan and update the model in SolidWorks. This dynamic real-time feedback and adjustment mechanism significantly improves the flexibility and efficiency of the design process and reduces the time required for repeated modifications;

[0044] 5. Provides a cloud-based design data storage and traceability system. The agent automatically records the design parameters, calculation steps and model generation process at each design stage, and ensures the integrity and traceability of the design data through a version control mechanism. Through this mechanism, every modification made by the designer can be fed back to the system, triggering a new round of design optimization, forming a closed-loop design process. This not only improves the accuracy and reliability of the design, but also provides a detailed historical record for subsequent design review and improvement, ensuring the transparency and auditability of the design process;

[0045] 6. It overcomes the shortcomings of convolutional neural networks (GCN) in processing complex geometric details, and is more applicable in high-precision manufacturing and design scenarios. Traditional methods may deviate when processing complex surfaces and internal structures, while the present invention can capture design requirements more accurately, especially in complex parts and high-precision manufacturing requirements.

[0046] 7. Through the introduction of intelligent technology, the automation and intelligent optimization of the entire process have been achieved, which has greatly improved the design efficiency, flexibility and accuracy, and effectively solved the problems of complex calculations, inefficient model creation, inconvenient design adjustments and imperfect data management in the traditional mechanical design process. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the process of the present invention;

[0048] Figure 2 It is a schematic diagram of the framework of the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] Example 1

[0051] like Figure 1As shown, a learning-generative AI-AD intelligent mechanical design method includes the following steps:

[0052] S1: For each type of mechanical design knowledge used in product development, including formulas, standards, charts or other invisible design knowledge, through the AI-AD intelligent mechanical design system (such as Figure 2 The AI-AD system is instilled with a learning agent (as shown) to allow it to understand the design knowledge first.

[0053] S2: After the AI-AD system has completed the learning of the required mechanical design knowledge, it will distinguish various innovative design processes, provide design steps and design calculations, and write the workflow in almost all natural language, with code added and plug-ins called in the middle.

[0054] S3: Write and design the Agent’s prompt strategy and the calling knowledge base to ensure that AI-AD can intelligently call the workflow written in S2 according to the designer’s various requirements (natural language input).

[0055] S4: Repeat the operations of S1-S3 to write various intelligent agents that need to complete various specific design functions.

[0056] S5: Based on the various intelligent agents designed previously, use the multi-agent self-organizing architecture (dialogue between agents) to complete the generative design of the product.

[0057] S6: Based on the generated design results and parameters, the 3D model intelligent body needs to generate statements that conform to the SolidWorks macro format, generate files and store them in the cloud, and finally modeling is performed through the interface in SolidWorks.

[0058] Furthermore, in S1, we let AI-AD understand the design knowledge in the following situations:

[0059] S11: For formula-based knowledge, designers can directly provide formula images, first call OCR to recognize the formulas in the knowledge, convert them into LaTex expressions, and then hand them over to the mature LaTex calculation plug-in deployed in the cloud to complete the calculation.

[0060] S12: If it is chart-based knowledge, the designer only needs to display parameters and block or curve information. AI-AD can use image recognition technology to process color blocks, curves and data information in the chart. The system will first pre-process the design chart, divide the image into color blocks through algorithms such as KMeans clustering, and then use OCR technology to recognize the text and coordinate system of the chart. The mapping relationship between color blocks and corresponding parameters in different charts is not fixed. Therefore, in the development stage, the system conducts supervised learning through sample data, and gradually establishes reliable mapping rules through multiple iterations. The system associates the extracted color point pixel values ​​with the parameter values ​​in the chart through polynomial fitting. Through the fitting algorithm, the system can derive the overall equation of the curve from known data points, so that designers can directly query the corresponding parameter values ​​of any coordinate points based on the chart. For charts with more curve information, the system uses image segmentation technology to separate different curves in the chart, extract the pixel coordinates of key color points, and map these pixel coordinates with the actual parameters in the chart. At the same time, after obtaining multiple color points of the curve, the system reconstructs the trend of the curve by connecting these color points and identifies the changing trends and key points. After these data information are trained by the learning model, they will be stored in the cloud as JSON files for users to query at any time.

[0061] S13: If it is workflow process knowledge, you only need to understand the input and output of each workflow, and the workflow is a black box. When scheduling, you can combine the input and output of the workflow with the prompt strategy of the agent according to the needs of the task, the scheduling content of the knowledge base, and the final requirements. The LLM of the agent can give the workflow to be executed and generate their execution links. Therefore, process knowledge is mainly determined in the prompt and the knowledge base.

[0062] S14: If it is design principle knowledge, LLM can directly understand this natural language expression, even implicit knowledge. In the knowledge base, this type of knowledge is expressed as an example case. When the intelligent agent encounters similar design cases during operation, it will make similar corresponding choices.

[0063] Furthermore, in S2, in order to improve the adaptability of the workflow, the system converts fuzzy natural language into specific design steps by analyzing the user's design requirements. Each step can not only call the design formulas and standards in the existing design knowledge base, but also integrate external APIs or cloud plug-ins. This flexibility makes the workflow not only limited to simple design tasks, but also able to play a role in complex design scenarios.

[0064] Furthermore, in S3, the design is divided into prompt writing strategy and calling knowledge base, as follows:

[0065] S31: The writing of prompts plays a key role in the subsequent user use of the system, and determines how the agent converts the natural language input by the user into executable tasks. Through natural language processing technology, the system can recognize semantic structures and convert the user's vague description into specific design parameters. For example, the user may enter "design a high-power transmission device", and the agent needs to be able to extract keywords such as "high power" and "transmission device" and map them to relevant entries in the knowledge base. If the user provides preliminary design information but does not have clear detailed parameters, the agent will gradually refine the user input and obtain more information by asking questions.

[0066] S32: In the AI-AD system, the knowledge base is the core of storing design rules, empirical formulas and standard data. The knowledge base is divided into multiple modules, each of which is responsible for processing different design tasks or design data, ensuring that the intelligent agent can call each module individually to meet the needs of specific tasks. In order to ensure that the knowledge base can respond quickly during the design process, by pre-processing and classifying the design parameters, standard formulas, etc. in the knowledge base, the system can quickly find the required data by searching for keywords, etc. At the same time, the content in the knowledge base is not static. The system can dynamically update the knowledge base based on user input and feedback to ensure efficient calling and continuous updating of the knowledge base.

[0067] Furthermore, in S5, the functions of each agent are relatively independent, such as parameter design agent, geometric modeling agent, etc., but they need to work together to complete the overall design task. In order to allow multiple agents to work together, a self-organizing architecture is constructed, allowing agents to communicate with each other through message passing, data sharing, and dialogue interfaces. The core of the architecture is to ensure that each agent can share data seamlessly and automatically initiate dialogue requests for collaboration when necessary. When the geometric modeling agent encounters a design conflict, it can communicate with the parameter design agent to coordinate the best solution. This dialogue can not only transmit data, but also exchange reasoning and decision information.

[0068] Further in S6, two different parts are divided into standard parts and non-standard parts, and there are steps and methods for classification:

[0069] S61: If the designed parts are standard or commonly used parts, such as springs, pulleys, bolts or standard gears, there will be standard drawing data and basic cloud-based VBA conversion statements in the knowledge base. Just tell the intelligent agent to replace the original design parameters to generate a VBA text file of the design results. After reading it into SolidWorks, the required 3D model can be generated.

[0070] S62: If it is a non-standard or commonly used part, generating VBA text is more complicated:

[0071] S621: First, characteristic function units that meet various functional requirements in the part are generated, and the generation data of the model, such as position, sketch plane, shape parameters and other information, are used as attributes of the characteristic units.

[0072] S622: Complete the design of connection units between characteristic functional units based on design knowledge, such as the selection of shaft shoulders, shaft rings, shaft sleeves and other structures during shaft design.

[0073] S623: Generate the structure of the connection unit and generate the VBA text.

[0074] S624: Check the feasibility of the parts.

[0075] All VBA code texts generated in S6 are stored in the cloud (Alibaba Cloud Server), the structured data of special diagnosis function units and connection units are also stored in the cloud, and the historical files of the creation process are also stored in the cloud. These data can be read and called by functions and intelligent bodies in the SolidWorks secondary development interface.

[0076] Example 2

[0077] This embodiment provides a complete system that allows users to obtain the complete design process and the final three-dimensional model results through natural language, that is, a three-dimensional model that meets the user's design requirements can be generated with high quality and efficiency based on the generation instructions input by the user.

[0078] S1: Initialization and user interaction

[0079] The agent receives the user's design requirements through the dialogue interface, such as the required cylinder diameter and other basic parameters. The user enters the initial value through the dialog box, and the agent automatically records and verifies the completeness of the input parameters. If the user is missing some necessary parameters, the agent will prompt the user to complete them to ensure that the design input is accurate.

[0080] S2: Call workflow analysis requirements

[0081] The intelligent agent automatically calls the preset workflow and calls the relevant modules according to the input initial parameters. After completing all the parametric designs, such as piston ring design, piston skirt design, piston top design, piston pin design, etc.

[0082] S3: Perform parameter verification

[0083] After the initial design is completed, the verification workflow is called to verify the strength of the piston ring and skirt dimensions, the stress of the piston pin, etc. to ensure the accuracy of the data. If the designer wants to modify or propose new requirements after reviewing, the intelligent agent will recalculate the part that needs to be corrected and regenerate the design report.

[0084] S4: 3D model generation

[0085] The system uses geometric modeling agents to build 3D models based on the previous design parameters, and completes the modeling by calling modularized VBA macro codes. The generation process and results of the target 3D model are embedded in SolidWorks, without the need for users to switch applications, achieving a seamless process from user needs to model generation, improving work efficiency and design consistency.

[0086] S5: Save and export

[0087] The final 3D piston model will be saved in a specified format so that users can perform further design or simulation in other systems. At the same time, the model's parametric data will also be saved for subsequent design calls or modifications. During the output process, the agent will also automatically record the key parameters and constraints of the modeling to ensure the traceability of the design data.

[0088] Example 3

[0089] In actual engineering applications, complex chart data such as V-belt selection diagrams often contain a lot of information, which is extremely tedious to search and analyze manually. To this end, the intelligent agent has designed an automated processing flow that can extract key information from the selection diagram through image recognition and OCR technology, and realize belt type query.

[0090] S1: Image extraction and preprocessing

[0091] The user first provides a picture of the V-belt selection diagram and an example of the coordinate values ​​of the belt type in each area. The agent automatically extracts the image data through the URL of the picture. The agent preprocesses the image, including grayscale and denoising, to ensure that the image has high clarity in the subsequent recognition process to improve the accuracy of the OCR and clustering algorithms.

[0092] S2: Region segmentation and color clustering: Call the KMeans algorithm to divide the color region, extract the color points of the curve, perform polynomial fitting, and generate a mapping table of belt type and power-speed coordinates.

[0093] The agent uses the KMeans clustering algorithm to analyze the pixel values ​​of the image and divides different areas in the image into several categories by color. This segmentation method helps to identify different functional areas in the chart, such as coordinate axes, curves, data areas, etc. After clustering, the agent identifies the main parts of the chart based on color differences, including the band distribution area and the coordinate system area, preparing for subsequent OCR recognition and data extraction.

[0094] S3: OCR recognizes coordinate system content

[0095] The agent uses OCR technology to recognize text information in the image and extracts the values ​​and labels on the coordinate axis, such as the numerical range of speed (n) and power (P). The agent further matches the text content recognized by OCR with the chart structure to construct the coordinate system framework of the chart.

[0096] S4: Mapping of band type and coordinates

[0097] After identifying the coordinate system, the agent uses a mapping algorithm to match the band types in the chart with the coordinate values ​​and establishes the relationship between the (P,n) coordinates and the corresponding band types.

[0098] S5: Result storage and cloud synchronization

[0099] After the mapping is completed, the agent stores the relationship data between the belt type and the coordinates in the cloud database for subsequent query. The detailed data of each belt type, including the corresponding (P,n) coordinate range, will be stored in a structured manner to ensure the integrity and accessibility of the data.

[0100] S6: Band type inquiry and feedback

[0101] When the user needs to query a specific belt type, the intelligent body will retrieve the corresponding (P,n) coordinate value from the cloud database and feedback the corresponding belt type information to the user. If the belt type queried by the user does not exist in the cloud database, the intelligent body will report an error prompt to remind the user to re-enter or check the data to improve the accuracy of use.

[0102] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A learning-generative AI-AD intelligent mechanical design method, characterized by: The following steps are involved: S1: For each type of mechanical design knowledge used in product development, including formulas, standards, charts or other invisible design knowledge, it is instilled through the learning agent of the AI-AD intelligent mechanical design system, so that the AI-AD intelligent mechanical design system can first understand the design knowledge; S2: After the AI-AD intelligent mechanical design system has completed the learning of the required mechanical design knowledge, it will distinguish various innovative design processes, provide design steps and design calculations, and use a low-code mode to write the workflow, during which code can be added and plug-ins can be called. S3: Write the prompt strategy and the calling knowledge base of the design agent to ensure that the AI-AD intelligent mechanical design system can intelligently call the workflow written in S2 according to the various requirements of the designer; S4: Repeat the operations of S1-S3 to write various intelligent agents that need to complete various specific design functions; S5: Based on the various intelligent agents designed previously, the self-organizing architecture of multiple intelligent agents is used to complete the generative design of products. S6: Based on the generated design results and parameters, the 3D model intelligent body needs to generate statements that conform to the SolidWorks macro format, generate files and store them in the cloud, and finally modeling is performed through the interface in SolidWorks.

2. The AI-AD intelligent mechanical design method based on learning-generative approach according to claim 1, characterized in that: In S1, the AI-AD intelligent mechanical design system is allowed to understand the design knowledge first, which is divided into the following situations: S11: For formula-based knowledge, designers can directly provide formula images, call OCR to recognize the formulas in the knowledge, convert them into LaTex expressions, and then hand them over to the mature LaTex calculation plug-in deployed in the cloud to complete the calculation; S12: If it is chart-based knowledge, the designer only needs high-definition parameters and block or curve information. The AI-AD intelligent mechanical design system can use image recognition technology to process the color blocks, curves and data information in the chart; S13: If it is workflow process knowledge, you only need to understand the input and output of each workflow, and the workflow is a black box. When scheduling, you can combine the input and output of the workflow with the prompt strategy of the agent according to the needs of the task, and the scheduling content of the knowledge base. As long as the final requirements are put forward, the LLM of the agent can give the workflow to be executed and generate their execution links. Therefore, process knowledge is mainly determined in the prompt and knowledge base. S14: If it is knowledge about design principles, LLM can directly understand this knowledge expressed in natural language, even implicit knowledge; this type of knowledge is expressed in the form of examples in the knowledge base, and when the intelligent agent encounters similar design cases during operation, it will make similar corresponding choices.

3. The AI-AD intelligent mechanical design method based on learning-generative approach according to claim 2, characterized in that: The detailed steps of S12 are as follows: the system will first pre-process the design chart, divide the image into color blocks through algorithms such as KMeans clustering, and then use OCR technology to identify the text and coordinate system of the chart; the mapping relationship between the color blocks and the corresponding parameters in different charts is not fixed. In the development stage, the system uses sample data for supervised learning, and gradually establishes reliable mapping rules through multiple iterations. The system associates the extracted color point pixel values ​​with the parameter values ​​in the chart through polynomial fitting; through the fitting algorithm, the system can derive the overall equation of the curve from the known data points, so that the designer can directly query the corresponding parameter value of any coordinate point based on the chart; for charts with more curve information, the system uses image segmentation technology to separate different curves in the chart, extract the pixel coordinates of key color points, and map these pixel coordinates with the actual parameters in the chart. At the same time, after obtaining multiple color points of the curve, the system reconstructs the trend of the curve by connecting these color points and identifies the changing trend and key points. After these data information are trained by the learning model, they will be stored in the cloud as JSON files for users to query at any time.

4. The AI-AD intelligent mechanical design method based on learning-generative approach according to claim 1, characterized in that: In step S2, in order to improve the adaptability of the workflow, the system analyzes the user's design requirements and converts vague natural language into specific design steps. Each step can not only call the design formulas and standards in the existing design knowledge base, but also integrate external APIs or cloud plug-ins. This flexibility makes the workflow not only limited to simple design tasks, but also able to play a role in complex design scenarios.

5. The AI-AD intelligent mechanical design method based on learning-generative approach according to claim 1, characterized in that: In step S3, the design of prompt writing strategy and calling knowledge base is as follows: S31: The writing of prompts plays a key role in the subsequent user use of the system. It determines how the intelligent agent converts the natural language input by the user into executable tasks. Through natural language processing technology, the system can recognize semantic structures and convert the user's vague description into specific design parameters. For example, the user may input "design a high-power transmission device". The intelligent agent needs to be able to extract keywords such as "high power" and "transmission device" and map them to relevant entries in the knowledge base; if the user provides preliminary design information but does not have clear detailed parameters, the intelligent agent will gradually refine the user input and obtain more information by asking questions; S32: In the AI-AD intelligent mechanical design system, the knowledge base is the core for storing design rules, empirical formulas and standard data. The knowledge base is divided into multiple modules, each of which is responsible for processing different design tasks or design data, ensuring that the intelligent agent can call each module individually to meet the needs of specific tasks. In order to ensure that the knowledge base can respond quickly during the design process, the system can quickly find the required data by preprocessing and classifying the design parameters, standard formulas, etc. in the knowledge base through keyword searching. At the same time, the content in the knowledge base is not static. The system can dynamically update the knowledge base based on user input and feedback to ensure efficient calling and continuous updating of the knowledge base.

6. The AI-AD intelligent mechanical design method based on learning-generative method according to claim 1, characterized in that: In step S5, the functions of each agent are relatively independent, such as the parameter design agent and the geometric modeling agent, but they need to collaborate to complete the overall design task. In order to allow multiple agents to work together, a self-organizing architecture is constructed, allowing the agents to communicate with each other through message passing, data sharing, and dialogue interfaces. The core of the architecture is to ensure that each agent can share data seamlessly and automatically initiate dialogue requests for collaboration when necessary. When the geometric modeling agent encounters a design conflict, it can communicate with the parameter design agent to coordinate the best solution. This kind of dialogue can not only transmit data, but also exchange reasoning and decision information.

7. The AI-AD intelligent mechanical design method based on learning-generative approach according to claim 1, characterized in that: In step S6, there are two different parts, standard parts and non-standard parts, and there are steps and methods for classification: S61: If the designed parts are standard or commonly used parts, such as springs, pulleys, bolts or standard gears, there will be standard drawing data and VBA conversion statements of the basic version in the cloud in the knowledge base. You only need to tell the intelligent agent to replace the original design parameters to generate a VBA text file of the design results, read it into SolidWorks, and you can generate the required 3D model; S62: If it is a non-standard or commonly used part, generating VBA text is more complicated: S621: First, a feature function unit that meets various functional requirements in the part is generated, and the generated data of the model, such as position, sketch plane, shape parameters and other information, is used as the attributes of the feature unit; S622: Complete the design of connection units between characteristic functional units based on design knowledge, such as the selection of shaft shoulders, shaft rings, and shaft sleeve structures during shaft design; S623: Connection unit structure generation and VBA text generation; S624: Check the feasibility of the parts.

8. The AI-AD intelligent mechanical design method based on learning-generative approach according to claim 7, characterized in that: All the VBA code texts generated in S6 are stored in the cloud, the structured data of the special diagnostic function units and connection units are also stored in the cloud, and the historical files of the creation process are also stored in the cloud. These data can be read and called by the functions and intelligent entities in the SolidWorks secondary development interface.

9. A learning-generative AI-AD intelligent mechanical design system, characterized by: The system is applied to the AI-AD intelligent mechanical design method based on learning-generation described in claims 1-8.

Images