CN120257996A

CN120257996A - An AI processing method and system for long-term memory of mechanical design knowledge

Info

Publication number: CN120257996A
Application number: CN202510374068.4A
Authority: CN
Inventors: 郑柯; 周刚; 刘晓蕊; 张宇昊; 王雷奇; 古乐翔; 徐娅宁; 周胤杰
Original assignee: Zhejiang Lover Health Science and Technology Development Co Ltd
Current assignee: Zhejiang Lover Health Science and Technology Development Co Ltd
Priority date: 2025-03-27
Filing date: 2025-03-27
Publication date: 2025-07-04

Abstract

The present invention discloses an AI processing method and system for long-term memory of mechanical design knowledge, which performs semantic analysis and feature extraction on heterogeneous knowledge such as parametric formulas, engineering diagrams, process flows and design principles in the field of mechanical design through a multi-agent collaborative learning system, and establishes a multi-dimensional modular knowledge base and a context-based Prompt strategy coupling mechanism. This method solves the problems of low efficiency, insufficient analysis accuracy and limited innovation support capabilities in existing mechanical design knowledge processing technologies, especially the technical difficulties such as high computational complexity, inaccurate reasoning and poor adaptability encountered when processing multiple types of design knowledge. Through the implementation of the present invention, the processing efficiency and design automation level of design knowledge are significantly improved, the technical threshold of designers is lowered, the design cycle is shortened, and efficient analysis, accurate matching and innovative application of mechanical design knowledge are achieved, providing strong support for the intelligent development of mechanical design.

Description

AI processing method and system for long-term memory of mechanical design knowledge

Technical Field

The invention belongs to the field of artificial intelligence, and particularly relates to an AI processing method and system for long-term memory of mechanical design knowledge.

Background

The existing mechanical design knowledge processing technology has the problems of low efficiency, insufficient resolution precision, limited innovation supporting capability and the like. In the field of mechanical design, designers need to deal with a large number of multi-type knowledge, including formulas, charts, flows, principles, etc., and the efficient handling and application of these knowledge is critical to improving design efficiency and innovation level. However, existing knowledge processing techniques often have difficulty meeting this requirement, resulting in a number of challenges for the designer in the design process. Therefore, it is important to design a system for processing different types of knowledge by adopting an intelligent processing strategy.

The invention with the patent number of CN202310953110 extracts three-dimensional CAD model characteristics of complex thin-wall parts through standardization pretreatment, voxelization treatment and convolutional neural network, establishes a model part library and a characteristic library, adopts hierarchical retrieval (global retrieval and local retrieval) to process knowledge, and realizes optimal matching of model similarity through a Kuhn-Munkres algorithm. However, this method still has drawbacks. Firstly, the applicability is limited to complex thin-wall parts, the universality of other types of parts or diversified model scenes is weak, secondly, the calculation complexity of local retrieval is high, particularly, the retrieval efficiency is possibly reduced after the model library scale is enlarged, thirdly, the classification and feature extraction effects of the convolutional neural network are highly dependent on the quality and quantity of training data, if the data are insufficient or distributed unevenly, the feature extraction accuracy is affected, in addition, the accuracy requirement on model preprocessing (such as standardization and voxelization) is high, reasoning cannot be carried out once errors occur, and the accuracy of the subsequent matching result is affected. Finally, for extremely complex models or models containing special features, the method is difficult to comprehensively capture detailed features, and the reliability of the retrieval result is reduced.

The invention with the patent number of CN202311246840 acquires and processes knowledge in various modes, including extracting knowledge from text data such as product manufacturing process data, mechanical manuals and the like, converting the knowledge into structured triplet data through natural language processing technology and manual correction, and constructing a knowledge map mode layer and a data layer. In addition, the processing characteristics are extracted from the three-dimensional model by utilizing the multi-view convolutional neural network, and the expert rules are combined with the database information to realize knowledge storage, matching and reasoning. However, the knowledge acquisition and processing methods still have shortfalls. Firstly, the process of relying on expert rules and manual correction is complicated and highly dependent on rule quality, secondly, the complex association and semantic fusion problems between multi-mode data are not fully considered, and finally, the dynamic updating and expanding capabilities of the knowledge graph are limited, so that the knowledge graph is difficult to adapt to the rapidly-changing industrial design requirements. The automation degree is low, the multi-mode data fusion capability is insufficient, and the expandability of the knowledge graph is still to be improved.

The invention with the patent number of CN202410961360 obtains knowledge by recording a three-dimensional CAD modeling operation flow, and stores modeling history in the form of version nodes and operation records, thereby facilitating the generation and reproduction of test cases. The user can extract the target test case from the recorded cases through the screening instruction, execute the cases after configuring the target test environment and compare the results, thereby realizing knowledge processing and verification. However, there are limitations in the way knowledge is obtained and processed. First, the modeling operation flow has high dependency, and if the recording process is incomplete or has errors, the knowledge base may be inaccurate and cannot be accurately inferred. And secondly, the intelligent multiplexing of knowledge in the processing process is insufficient, more test scenes are difficult to adapt, and the coverage rate is limited. In addition, the result comparison is mainly based on explicit information such as examples and elements, and is difficult to capture abstract features at a deeper level.

The invention with the patent number of CN202410558221 provides a design proposal which is generated by a model pre-trained and fine-tuned by collecting the multi-mode design requirements (such as characters, pictures, audio and video) input by a user through a plug-in client, analyzing the multi-mode design requirements into a generation instruction and transmitting the generation instruction to a cloud server. However, the method has limitation on understanding capability of multi-mode input, and particularly in the field of mechanical design, it is difficult to accurately extract different types of design knowledge (such as formulas, charts, flows and principles) and realize dynamic collaborative processing. Specifically, when the technology processes multi-mode information, the relevance between complex inputs and the fusion capability of context information are not perfect, and the fine-granularity analysis requirement of multi-type knowledge in a mechanical design task is difficult to meet.

Furthermore, this technique is too dependent on predefined sample data, and it is difficult to perform efficient knowledge reasoning in the absence of high quality domain data, resulting in limited adaptability and extensibility. For dynamic adjustment and innovative design requirements in mechanical design tasks, the flexibility of the model generation results is low, and particularly, the model generation results are not enough to be performed when parameters are required to be adjusted in real time or different knowledge types (such as the cooperation of formula calculation and chart analysis) are combined.

The integrated management system of the numerical control machine tool disclosed by the invention with the patent number of CN201810873595 obtains basic information (such as machine tool and component information) of the numerical control machine tool through the equipment management module, manages the replacement period and the stock quantity of the components through the component management module, combines the data acquisition analysis module, obtains state information in real time through sensors (such as temperature, vibration, current and voltage sensors) arranged on the machine tool, and then analyzes and detects abnormality through the fault detection module. However, the system still has the defects that the comprehensive adaptability and accuracy of the data acquisition device to the complex working conditions are limited, the real-time adjustment capability of the component replacement period and the spare part stock prediction model to the dynamic working conditions is insufficient, and the intelligent level of preventive maintenance strategy and task generation is required to be improved when coping with diversified demands.

The knowledge question-answering method of the invention with the patent number of CN202411658291 analyzes user questions by hierarchically managing structured, unstructured and semi-structured knowledge bases and combining a language model, screens relevant information from a target knowledge base to generate answers, and optimizes knowledge updating and management by means of manual intervention. The method can efficiently process common problems to a certain extent, but has significant limitations in terms of complex context understanding, unstructured knowledge processing and dynamic knowledge automation updating. Especially, the method has the defects of relevance management of multi-business line knowledge, real-time dynamic knowledge expansion and adaptability in an open scene. In addition, knowledge screening mechanisms of the system generally depend on indexes such as heat regression, statistical models and the like, and can cause insufficient coverage of rare or cold knowledge, so that diversified requirements of users are difficult to comprehensively meet. Meanwhile, because the manual intervention still occupies an important position, the knowledge base updating efficiency is low, and the high-speed change of the knowledge content is difficult to adapt quickly.

The five patents show that the existing mechanical design knowledge processing technology has the problems of low efficiency, insufficient analysis precision, weak innovation supporting capability and the like, and particularly when processing multi-type design knowledge, the problems of high calculation complexity, inaccurate reasoning and poor adaptability are often faced. In order to solve the problems, the invention provides a design knowledge processing system based on multi-type data fusion and intelligent reasoning. The system automatically collects and analyzes various types of design requirement data (mechanical design knowledge such as formulas, charts, flow and principle) and accurately extracts knowledge features and performs intelligent reasoning and matching by utilizing deep learning and natural language processing technologies. And combining the knowledge graph and expert rules, dynamically updating the knowledge base, and supporting efficient and accurate design scheme generation. The method greatly improves the innovation capability of the design efficiency and solves the limitation of the prior art in multi-mode knowledge processing and intelligent pushing.

Disclosure of Invention

In order to overcome the defects in the prior art, the invention provides the AI processing method for the long-term memory of the mechanical design knowledge, and the problems of low knowledge processing efficiency, insufficient resolution accuracy and limited innovation supporting capability in the prior art are solved by classifying and optimizing analysis of the mechanical design knowledge. Aiming at the mechanical design knowledge of formulas, charts, processes and principles, the invention adopts an intelligent processing strategy to realize efficient and accurate knowledge analysis and dynamic application support, and the invention adopts the following specific technical scheme:

an AI processing method facing to long-term memory of mechanical design knowledge comprises the following steps:

s1, performing classified routing according to types (pictures or texts) of input data in the field of mechanical design through a multi-agent collaborative learning system, wherein the system intercepts image format errors, OCR invalidation or semantic ambiguity scenes in the processing process in real time through an anomaly detection mechanism, and realizes error correction and flow retriggering through user interaction closed loop;

s2, a multidimensional modularized knowledge base is established, and a context-sensitive Prompt strategy is combined, so that user demands and knowledge base resources are coupled, and efficient knowledge dynamic calling and multi-scene adaptation are supported;

S3, through integrating the characteristic representation of the knowledge of formulas, charts, processes and design principles, the semantic alignment and reasoning optimization of heterogeneous design knowledge are realized by taking the user requirement as a core and combining the context semantic and the domain knowledge graph, and finally a comprehensive design scheme meeting the functional, performance and constraint conditions is generated;

S4, realizing standardized conversion and classified storage of the characteristic data to the structured data through a data semantic matching and multi-type storage mechanism;

And S5, storing the analysis result in a distributed cloud architecture, so as to realize long-term memory and support multi-user concurrent access and security guarantee.

Preferably, in the intelligent processing method of S1, the processing of classified routing is divided into the following two cases:

The first case is to input aiming at the picture class, whether the picture class is a valid engineering image is verified by a format recognition module, if the picture class is verified, the structural information is extracted by adopting an OCR technology, and the structural information is further classified into a formula class (S11) or a chart class (S12) processing module according to the extracted content characteristics;

S11, constructing a Latex mathematical expression based on a symbol sequence extracted by OCR (optical character recognition) for formula knowledge, and supporting dynamic calculation and structured storage;

s12, analyzing the structure, coordinate axis data and labeling information in the image for the chart knowledge to generate a computable parameterized model;

the second case is to input text by analyzing semantic features through a large language model reasoning module, classifying the semantic features into a process class (S13) processing module if the process sequence is included, classifying the semantic features into a design principle class (S14) processing module if the process sequence is included, and marking the semantic features as invalid input and generating correction suggestions if the semantic features are failed to analyze;

S13, analyzing constraint logic of the design flow based on the process sequence and the input-output relation for the flow class knowledge;

S14, the design principle knowledge relates to implicit knowledge in mechanical design, and the implicit constraint is mined by combining a case base and a rule base to generate an optimization suggestion adapting to the current design target.

Preferably, the detailed steps of the processing of the parameterized formula of S11 are as follows:

S111, extracting characteristic representation of a formula image through a multi-layer convolution network, and optimizing the recognition accuracy of a complex formula;

S112, accurately positioning and dividing the character by using an attention mechanism to ensure the recognition integrity of the special symbol;

s113, high-reliability classification of characters is achieved based on the depth residual error network, and guarantee is provided for subsequent formula analysis.

Preferably, the detailed steps of the process in which the engineering chart of S12 is identified are as follows:

S121, performing high-precision pixel level segmentation on a curve and a color block area in a chart by utilizing KMeans clustering algorithm, and dynamically adjusting clustering weights;

S122, comprehensively extracting coordinate axis information and parameter labels of the chart by combining with an OCR technology enhanced by deep learning, and ensuring complete extraction of chart data;

S123, reconstructing a parameterized curve equation by using a least square polynomial fitting algorithm, and allowing a user to finely adjust a fitting result according to requirements.

Preferably, the detailed steps of the process flow of S13 are as follows:

s131, automatically extracting key logic relations and input and output parameters of a flow by a semantic analysis method based on a transducer architecture;

S132, dynamically adjusting the flow structure by using the constructed knowledge graph and logic reasoning technology to adapt to the actual design requirement.

Preferably, the detailed steps of the process in which the design rule knowledge of S14 is as follows:

S141, disassembling a core concept in a design principle into an entity-relation-attribute triple form, and constructing a domain knowledge graph;

S142, storing detailed description of design principles, expert experience and case background description in an unstructured text form, and retaining original semantic information.

Preferably, the step of matching the semantic features of S3 is as follows:

S31, analyzing semantic intent input by a user based on a natural language processing model, identifying key parameters, constraint conditions and target attributes in design requirements, and mapping the key parameters, constraint conditions and target attributes to a structural representation in a storage;

S32, extracting characteristic data from knowledge modules of formulas, charts, flow and design principles respectively, wherein the characteristic data comprises mathematical logic of parameterized formulas, relation of charts, constraint of process flows and implicit rules of design principles and the implicit rules are converted into a structured data format;

s33, dynamically matching the semantic interpretation result in S31 with the feature data extracted in S32 in a unified semantic space.

Preferably, the detailed steps of the structured data store of S4 therein are as follows:

S41, dynamically matching the feature data generated in the multi-mode collaborative reasoning stage with a preset data semantic model, extracting logical relations and context constraints in the features through a semantic alignment algorithm, and converting the logical relations and the context constraints into standardized original structure data;

s42, carrying out integrity check and conflict detection on the original structure data, and identifying redundant, missing or contradictory information in the data;

s43, after the verification is completed, performing split flow storage according to the data type and the application scene, namely storing the data (such as the node relation of the design flow) which needs high relevance and transaction support into a relational database;

S44, converting dynamic change or semi-structured data (such as chart parameters) into a lightweight JSON format for storage;

s45, for unstructured text information (such as a design principle abstract described by natural language), archiving by adopting a text file format.

Preferably, the detailed steps of the cloud storage of S5 are as follows:

S51, uniformly converting the standardized data (including JSON format parameters, relational database table items and text files) generated in the S4 stage into a cloud compatible database;

s52, long-term memory and calling functions of package design knowledge provide core data support for subsequent designs.

The beneficial effects of the invention are as follows:

1. The invention provides a systematic solution around the intelligent processing of the mechanical design knowledge, and particularly realizes important breakthrough in the recognition, classification and application of the design knowledge. The system designs a highly optimized processing method aiming at four design knowledge of formulas, charts, flows and principles, thereby remarkably improving the resolution precision and practical efficiency of the design knowledge and meeting the complex demands in nonstandard design and innovative design.

2. In the aspect of formula knowledge processing, the invention adopts an advanced OCR technology and LaTeX expression conversion method, and combines a cloud computing engine, so that structural analysis and dynamic computation of formulas can be rapidly realized, and the system can efficiently identify and generate accurate computing results no matter complex mathematical formulas or empirical formulas commonly used in engineering design, thereby providing scientific and reliable technical support for designers.

3. On the processing of graph knowledge, the invention provides a set of intelligent analysis method based on image segmentation and polynomial fitting, which segments color blocks and curve areas in a graph through a clustering algorithm, extracts character and numerical information by combining an OCR technology, and enables a system to fit scattered point data into a continuous curve and establish an accurate mapping relation with actual design parameters.

4. In the processing aspect of flow knowledge, the method fully utilizes the Prompt strategy and the knowledge base scheduling technology, can analyze the flow demand input by a user into a task link executable by a system, and for a complex flow, the system can intelligently generate a task path and dynamically optimize execution steps by analyzing the input-output relation.

5. For principle knowledge, the invention realizes the deep analysis and reasoning of the implicit knowledge in the design principle by integrating the large language model, the fuzzy or natural language description input by the designer can automatically match the example cases in the knowledge base and generate the design suggestion meeting the actual requirement. The reasoning capability is particularly suitable for solving the problem of special constraint in design, provides scientific decision support for designers, and effectively reduces the dependence of high-end design tasks on experience.

In general, the invention takes the mechanical design knowledge as the core, converts the complex knowledge analysis task into efficient and visual automatic operation through classifying and optimizing the different types of knowledge, and the technical innovation of the invention in the aspects of formula identification, chart analysis, flow generation and principle reasoning effectively breaks through the technical bottleneck of the traditional design tool, thereby providing important support for the intelligent and efficient development of mechanical design.

Drawings

FIG. 1 is a schematic flow chart of the present invention;

Fig. 2 is a schematic diagram of the framework of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

In the intelligent processing method of S1, the processing of classified routing is divided into the following two cases:

Example 1 Intelligent resolution and dynamic computation based on formula-like knowledge

In the mechanical design process, complex mathematical formula calculation is a common difficulty, and particularly in the scenes of transmission design, mechanical analysis and the like, a large number of parameterized formulas are required to be processed. The method specifically comprises the following steps:

(1) Formula input and recognition

If the user uploads a picture containing a formula, such as a bearing load calculation formula, a tooth surface fatigue strength formula, etc., through the interface. The system calls an OCR module to automatically identify mathematical expressions in the pictures and convert the mathematical expressions into LaTeX format;

(2) Structured storage and cloud computing

The identified formula is analyzed through a cloud LaTeX computing plug-in, the system automatically reads parameters provided by a user, and a result is dynamically computed;

(3) Dynamic feedback of results

The system returns the calculation result to the user.

Through the embodiment, a user does not need to manually input a complex formula or perform complicated manual calculation, and the efficient identification and dynamic calculation capability of the system greatly improve the design efficiency.

Example 2 Intelligent resolution and modeling support based on graph class knowledge

In mechanical design, engineers often need to extract parameter values from charts, such as selecting belt types in V belt transmission, and the invention provides a chart knowledge analysis method to realize seamless conversion from chart data to mathematical models. The method specifically comprises the following steps:

(1) Chart input and preprocessing

A user uploads a V-belt transmission type selection chart, wherein the chart comprises a plurality of color block curves which respectively correspond to power-rotating speed relations of different belt types, and the system optimizes the input quality of chart data through image enhancement and noise suppression technologies;

(2) Region segmentation and parameter extraction

The system calls an image processing module, a KMeans clustering algorithm is adopted to divide the chart, pixel coordinates and color block areas of each curve are extracted, and then coordinate axis information of the chart is recognized through an OCR module;

(3) Curve fitting and model reconstruction

The system converts curve data into a mathematical equation through a polynomial fitting algorithm;

(4) Dynamic interaction and data export

The user can inform the mapping of the belt type and the coordinates of the intelligent body through the intelligent body interaction interface, the system automatically matches, and meanwhile, the result is exported to be structured data for subsequent modeling and analysis.

The embodiment solves the inefficiency problem of manual graph checking and manual parameter extraction, realizes automatic analysis and modeling support of complex chart data, and is widely applicable to scenes requiring chart parameters such as transmission design.

Example 3 workflow Generation and invocation based on flow class knowledge

The invention realizes the automation and optimization of the design flow by the deep analysis of flow knowledge and the intelligent call of dynamic workflow, thereby improving the design efficiency and accuracy. The method specifically comprises the following steps:

(1) Demand input and process disassembly

The user inputs the design requirement through natural language, namely, designing an efficient transmission device with the output power of 10 kW. The system call LLM module analyzes the requirements, extracts key information (such as 'power=10kW', 'transmission device'), and the key information is related to a standardized design flow template in a knowledge base, and starts corresponding workflow call;

(2) Workflow invocation and execution

The system calls a preset design workflow based on a flow template and user input to realize automatic task processing, wherein the example flow comprises the following steps:

(21) Material matching

The system calls a material knowledge base, and the design parameters are combined to automatically match the applicable materials, so that the mechanical strength, the cost and the service life are comprehensively considered;

(22) Selection calculation

The system calls a formula module, calculates the optimal transmission ratio according to the requirement, and generates a preliminary model selection scheme;

(23) Parameter checking

The system calls corresponding check workflow, plug-ins and the like, so that the design scheme is ensured to meet the standard requirement;

(3) Flow adjustment and optimization

The user can dynamically adjust design parameters (such as transmission ratio, material priority and the like) on the interactive interface, the system can recall corresponding modules according to new input, parameter checking and optimization are completed, and process rationality and reliability of design results are ensured;

(4) Result output and report generation

And (3) the system calls a document generation module, integrates the selection, the material and the check result, and outputs a complete design report.

Through the embodiment, a user can quickly complete a complex mechanical design flow, reduce dependence on manual experience, and simultaneously improve consistency of design quality and document output.

Example 4 optimization reasoning and case support based on principle class knowledge

In mechanical design, design principles typically rely on engineer experience and case accumulation. According to the method, the Large Language Model (LLM) and the knowledge base are combined, the principle knowledge is intelligently analyzed and optimized, the targeted design suggestion is provided, and the design efficiency and the scheme reliability are effectively improved. The method specifically comprises the following steps:

(1) Design objective resolution and principle matching

The user inputs design goals through natural language, such as "design a bearing system, require high rigidity and low cost. The system analyzes the description by utilizing the LLM module, extracts key requirements ('high rigidity', 'low cost'), and searches a design principle and an optimization method related to the key requirements in a knowledge base, thereby providing a basis for subsequent reasoning;

(2) Case support and optimization reasoning

The system provides a reference scheme for a user based on design cases in a knowledge base, and displays different design directions according to specific requirements:

(21) In a certain lightweight bearing system, aluminum alloy is adopted as a main material;

(22) In the design of a certain high-rigidity bearing, the high-strength requirement is met by optimizing the structural design;

(23) In some low cost design cases, cost control is achieved by using conventional steels;

the system provides the background and the application condition of the cases according to the user target, and helps the user to quickly screen the suitable scheme;

(3) Dynamic adjustment and parameter support

The user can adjust design parameters (such as bearing size, material selection and the like) in a system interface according to specific requirements, and the system updates relevant reference cases and parameter ranges in real time so as to ensure the applicability and accuracy of design principles;

(4) Result output and aided design

The system integrates design goals, matching principles and case information to generate a detailed design reference report.

Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims

1. An AI processing method for long-term memory of mechanical design knowledge, characterized by comprising the following steps:

S1. For different knowledge in the field of mechanical design, data is input through a multi-agent collaborative learning system, and a classification routing processing method is executed for images or texts according to the data type; wherein the system intercepts image format errors, OCR failure scenarios and semantic ambiguity scenarios in the process of the classification routing processing method in real time through an anomaly detection mechanism, and realizes error correction and process re-triggering through a user interaction closed loop;

S2. Establish a multi-dimensional modular knowledge base, combine context-sensitive prompt strategies, couple user needs with knowledge base resources, and support dynamic knowledge call and multi-scenario adaptation;

S3, by integrating the feature representation of formula, chart, process and design principle knowledge, taking user needs as the core, combining contextual semantics and domain knowledge graph, to achieve semantic alignment and reasoning optimization of heterogeneous design knowledge;

S4. Through data semantic matching and multi-type storage mechanism, the standardized conversion and classified storage of feature data into structured data is realized;

S5. Store the analysis results in a distributed cloud architecture to achieve long-term memory, support multi-user concurrent access and security assurance.

2. The AI processing method for long-term memory of mechanical design knowledge according to claim 1, characterized in that, in S1, the classification routing processing method includes a picture input processing method and a text input processing method;

The image input processing method verifies whether it is a valid engineering image through a format recognition module; if the verification is successful, the OCR technology is used to extract structured information, and jumps to the formula knowledge processing module S11 or the chart knowledge processing module S12 according to the extracted content features; if the verification fails, the exception handling mechanism is triggered to terminate the process and feedback the error type;

S11. For formula knowledge, Latex mathematical expressions are constructed based on the symbol sequences extracted by OCR, supporting dynamic calculation and structured storage;

S12. For chart-related knowledge, analyze the structure, coordinate axis data and annotation information in the image to generate a computable parameterized model;

The text input processing method analyzes semantic features through a large language model reasoning module. If it contains a process step sequence, it jumps to a process knowledge processing module S13; if it contains implicit design rules or empirical design criteria, it jumps to a design principle knowledge processing module S14; if the analysis fails, it is marked as invalid input and a correction suggestion is generated;

S13. For process knowledge, analyze the constraint logic of the design process based on the sequence of work steps and the relationship between input and output;

S14. For design principle knowledge, which involves implicit knowledge in mechanical design, the implicit constraints are mined by combining the case library and rule library to generate optimization suggestions that are suitable for the current design goals.

3. The AI processing method for long-term memory of mechanical design knowledge according to claim 2, characterized in that in S11, the processing of parameterized formulas includes:

S111, extracting feature representation of formula images through multi-layer convolutional networks to optimize the recognition accuracy of complex formulas;

S112. Use the attention mechanism to accurately locate and segment characters to ensure the recognition integrity of special symbols;

S113. High-reliability character classification is achieved based on deep residual network, providing guarantee for subsequent formula parsing.

4. The AI processing method for long-term memory of mechanical design knowledge according to claim 2, characterized in that in said S12, the processing of the recognition of engineering diagrams comprises:

S121. Use KMeans clustering algorithm to perform high-precision pixel-level segmentation on curves and color block areas in the chart, and dynamically adjust clustering weights;

S122. Combined with deep learning enhanced OCR technology, the coordinate axis information and parameter annotations of the chart are fully extracted to ensure the complete extraction of chart data;

S123. Reconstruct the parameterized curve equation using a least squares polynomial fitting algorithm, while allowing the user to fine-tune the fitting result according to needs.

5. The AI processing method for long-term memory of mechanical design knowledge according to claim 2, characterized in that in said S13, the processing of the process flow includes:

S131, semantic parsing method based on Transformer architecture, automatically extracting key logical relationships and input and output parameters of the process;

S132. Use the constructed knowledge graph and logical reasoning technology to dynamically adjust the process structure to adapt to actual design needs.

6. The AI processing method for long-term memory of mechanical design knowledge according to claim 2, characterized in that in said S14, the processing of design principle knowledge includes:

S141. Decompose the core concepts in the design principles into "entity-relationship-attribute" triples to build a domain knowledge graph;

S142. Store detailed descriptions of design principles, expert experience, and case background information in unstructured text form, retaining the original semantic information.

7. The AI processing method for long-term memory of mechanical design knowledge according to claim 1, characterized in that S3 comprises:

S31. Parse the semantic intent of user input based on the natural language processing model, identify key parameters, constraints and target attributes in the design requirements, and map them to structured representations in storage;

S32, extracting characteristic data from the formula, chart, process and design principle knowledge modules respectively, including converting the mathematical logic of parameterized formulas, the relationship of charts, the constraints of process flows and the implicit rules of design principles into structured data formats;

8. The AI processing method for long-term memory of mechanical design knowledge according to claim 1, characterized in that said S4 comprises:

S41, dynamically matching the feature data generated in the multimodal collaborative reasoning stage with the preset data semantic model, extracting the logical relationship and context constraints in the features through the semantic alignment algorithm, and converting them into standardized original structure data;

S42, perform integrity check and conflict detection on the original structure data to identify redundant, missing or contradictory information in the data;

S43. After verification is completed, perform offload storage according to the data type and application scenario - for data that requires high relevance and transaction support, store it in a relational database;

S44. For dynamically changing or semi-structured data, convert it into lightweight JSON format for storage;

S45. For unstructured text information, it is archived in text file format.

9. The AI processing method for long-term memory of mechanical design knowledge according to claim 1, characterized in that said S5 comprises:

S51: Convert the standardized data generated in S4 into a cloud-compatible database;

S52: The long-term memory and call function of packaging design knowledge provides core data support for subsequent design.

10. An intelligent processing system for multiple types of mechanical design knowledge, characterized in that the system adopts the AI processing method for long-term memory of mechanical design knowledge as described in any one of claims 1 to 9.