Multi-source heterogeneous data and big language driven process intelligent design system and method

Abstract

The invention relates to the technical field of artificial intelligence and intelligent manufacturing, and discloses a multi-source heterogeneous data and large language driven process intelligent design system and method, and the system comprises a multi-modal data processing module which is used for achieving the structural information extraction of multi-source heterogeneous engineering data; the multi-source heterogeneous mixed database construction module is used for realizing storage of multi-source process knowledge and joint query of a database; the large model construction and optimization module is used for constructing a large model to realize user intention understanding, process knowledge fusion and process scheme generation and optimization; the intelligent processing technology collaborative design system provides a multi-mode interaction interface for a user to realize full-process automatic technology design from demand input to scheme output; according to the system and the method, the full-process intelligence from part information extraction, process knowledge retrieval, process scheme generation to simulation verification optimization is realized, and the system and the method can be widely applied to machining process design and optimization scenes in the fields of industrial mother machines, robots, high-end equipment manufacturing and the like.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and intelligent manufacturing technology, and in particular to a process intelligent design system and method driven by multi-source heterogeneous data and large language. Background Technology

[0002] Process design is a core aspect of mechanical manufacturing, directly impacting product quality, production efficiency, and manufacturing costs. Traditional process design relies on manual experience, resulting in four major bottlenecks: First, information acquisition efficiency is low, as process data is scattered across various formats such as drawings, models, and manuals, making manual review and integration time-consuming and prone to errors. Second, knowledge reuse is difficult, as tacit experience is stored in fragments, hindering systematic accumulation and inheritance, leading to repetitive work and knowledge loss. Third, design cycles are long, with complex parts requiring repeated review and verification, making it difficult to meet rapid response requirements. Fourth, design quality is unstable, depending on individual abilities and lacking unified evaluation standards, making it difficult to guarantee optimal solutions.

[0003] In recent years, breakthroughs in large language model technology have provided new technical pathways for solving the aforementioned problems. However, the application of existing large language models in the field of process design faces multiple challenges: first, they lack the ability to process professional multimodal data such as engineering drawings, 3D models, and point cloud data; second, general-purpose models lack manufacturing expertise, and the generated process solutions often do not meet actual processing requirements; and third, they lack an effective mechanism for integration with existing enterprise process knowledge bases and equipment databases.

[0004] Therefore, there is an urgent need to develop a specialized multimodal large model system for process design that can automatically process multi-source heterogeneous data, deeply integrate domain knowledge, and intelligently generate and optimize process solutions, thereby fundamentally improving the efficiency and quality of process design. Summary of the Invention

[0005] This invention proposes a process intelligent design system and method driven by multi-source heterogeneous data and large language, in order to solve the shortcomings of the above-mentioned prior art. The system has multimodal data processing capabilities, strong model professional capabilities, and can realize the intelligent process from part information extraction and process knowledge retrieval to the generation of process schemes that meet processing requirements.

[0006] The technical solution of this invention is: a process intelligent design system driven by multi-source heterogeneous data and large language, comprising: The multimodal data processing module uses geometric feature recognition algorithms, automatic speech recognition technology, and convolutional neural networks to process 3D models, speech, natural language documents, part images, and 3D point cloud data, extracting structured information and converting it into a unified information representation format. The structured information includes part ID, material properties, geometric feature list, processing requirements, and tolerance information. A multi-source heterogeneous hybrid database includes a relational database and a graph database. The relational database stores part attribute information in the structured information, and the graph database stores process knowledge graphs. The relational database and the graph database are linked through a unified data identifier to support joint queries. The large model building module is based on a pre-trained Transformer architecture large language model. It integrates process knowledge from the multi-source heterogeneous hybrid database through a retrieval enhancement generation mechanism and uses LoRA technology to efficiently fine-tune parameters in the processing technology domain to obtain a domain-adaptive model for processing technology design. The intelligent machining process collaborative design system includes a multimodal interaction interface and an inference engine. The multimodal interaction interface is used to receive process requirements input by users, including natural language queries, drawing uploads, 3D model inputs, and voice inputs. The inference engine is used to call the joint query results of the multi-source heterogeneous hybrid database and perform inference based on the domain adaptation model to generate and output a detailed machining plan including process flow, equipment model, tooling fixtures, and tool parameters.

[0007] In at least one embodiment of the present invention, the multimodal data processing module includes a three-dimensional model processing submodule, a speech processing submodule, a natural language document processing submodule, and an image point cloud processing submodule.

[0008] In at least one embodiment of the present invention, the three-dimensional model processing submodule includes a geometric feature recognition algorithm, a feature recognition rule base and a deep learning model. The geometric feature recognition algorithm is used to automatically extract the dimensional features, machining feature surfaces, geometric tolerances and surface roughness of the three-dimensional model data. The feature recognition rule base and the deep learning model are combined to realize the automatic recognition and parameterized expression of machining features such as holes, slots, steps and chamfers.

[0009] In at least one embodiment of the present invention, the speech processing submodule uses automatic speech recognition technology to convert speech signals into text, and extracts key information such as part names, processing requirements and process parameters through named entity recognition and relation extraction natural language processing technology.

[0010] In at least one embodiment of the present invention, the natural language document processing submodule is configured with a pre-trained language model. The natural language document processing submodule uses document layout analysis, table recognition and paragraph segmentation technology to perform structured parsing of process cards, process manuals and process specification data. The pre-trained language model is used to extract process knowledge elements such as process name, processing parameters, equipment requirements and tool selection from the structured parsing results.

[0011] In at least one embodiment of the present invention, the image point cloud processing submodule uses a convolutional neural network to perform part category recognition and appearance defect detection on the image of the physical part; and uses point cloud registration, segmentation and feature extraction algorithms to reconstruct the geometric features of the part and measure its dimensions based on the three-dimensional scan point cloud data.

[0012] In at least one embodiment of the present invention, the multi-source heterogeneous hybrid database includes: The structured parts information database based on MySQL uses a relational database to store the structured attribute information of parts; it adopts a star schema design, with parts as the core and linking information of various dimensions, and supports multi-condition compound queries and batch data processing; The Neo4j-based graph database process knowledge base uses a graph database to store process knowledge with complex relationships, including a processing technology knowledge graph, an equipment and tool knowledge graph, and a process rule base. A hybrid database collaboration mechanism is established to enable data synchronization and joint querying between relational databases and graph databases. A unified data identification system is used to link the parts information database and the process knowledge base, and a hybrid query engine supports cross-database queries.

[0013] In at least one embodiment of the present invention, the large model construction and optimization module includes: a user intent understanding module, a professional domain knowledge infusion module, a model fine-tuning module, a process evaluation system module, and a closed-loop process optimization module; the user intent understanding module includes a three-level intent recognition architecture of query type classification, fine-grained intent parsing, and context fusion to understand the user's process design requirements; the professional domain knowledge infusion module deeply integrates a multi-source heterogeneous hybrid database with a large language model, and adopts retrieval enhancement generation technology, a knowledge graph embedding module, and a knowledge consistency verification mechanism to improve the model's process understanding ability; the model fine-tuning module is based on LoRA parameter efficient fine-tuning technology, combined with process question-and-answer datasets and process domain-specific prompt word templates, and adopts a training strategy that combines instruction fine-tuning and reinforcement learning to achieve model domain adaptation; the process evaluation system module establishes a multi-dimensional process scheme evaluation index system, including process accuracy evaluation, process efficiency evaluation, process economic evaluation, and process stability evaluation; the closed-loop process optimization module constructs a closed-loop iterative mechanism for generation, evaluation, and optimization, adopts reinforcement learning methods for online model optimization, establishes a process scheme library, and designs multi-objective optimization algorithms to achieve process scheme trade-off optimization.

[0014] In at least one embodiment of the present invention, the intelligent manufacturing process collaborative design system includes: a system functional architecture, full-process automation capabilities, and a system self-evolution mechanism; the system functional architecture provides a visual human-computer interaction interface at the front end, supporting input methods such as drawing upload, voice input, and natural language interaction; the back end integrates a multimodal data processing engine, a knowledge base management system, and a large model inference service module; the full-process automation capabilities include requirement analysis, knowledge retrieval, process generation, and optimization output, to achieve full-process automation from requirement input to solution output; the system self-evolution mechanism, based on data accumulated during operation, realizes knowledge base updates, incremental model training, and interaction process optimization to continuously improve system performance and user experience.

[0015] This invention also proposes a process intelligent design method driven by multi-source heterogeneous data and large language, comprising the following steps: The multimodal data processing module preprocesses and extracts structured information from 3D models, speech, natural language documents, part images, and 3D point cloud data, converting them into a unified information representation format. The structured information is stored in the relational database of the multi-source heterogeneous hybrid database, and the process knowledge is stored in the graph database, and the association is established through a unified data identifier; Through the large model building module, based on the pre-trained Transformer architecture large language model, the process knowledge is integrated through the retrieval enhancement generation mechanism, and LoRA technology is used for efficient parameter fine-tuning to obtain a domain-adaptive model for processing technology design. The intelligent machining process collaborative design system receives user-inputted process requirements through its multimodal interaction interface, calls the joint query results of the multi-source heterogeneous hybrid database, and performs reasoning based on the domain adaptation model to generate and output a detailed machining scheme including process flow, equipment model, tooling fixtures, and tool parameters. Collect user feedback and actual processing data, continuously update the knowledge base and optimize the large model to achieve the system's self-evolution.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes a multi-source heterogeneous data and large language-driven intelligent process design system. By setting up a multi-modal data processing module, it achieves collaborative processing and structured information extraction of various modalities of data, including 3D models, 2D drawings, voice, natural language documents, images, and point clouds, breaking through the bottleneck of traditional process design information acquisition. Through a multi-source heterogeneous hybrid database construction module, it establishes a multi-source heterogeneous hybrid database architecture oriented towards process design, innovatively combining relational databases and graph databases to construct a hybrid database architecture that integrates part information databases and process knowledge graphs. This enables efficient storage and joint retrieval of structured part data and complex process relationship knowledge. For large-scale model construction, it proposes a professional training method for large models that integrates retrieval enhancement generation, knowledge graph embedding, and LoRA fine-tuning technologies, achieving effective migration of general-purpose large language models to specialized process design models, significantly improving the model's professional capabilities in the process field. Simultaneously, it suggests building an end-to-end intelligent system based on the above modules, encompassing requirements analysis, knowledge retrieval, process generation, simulation verification, and optimization output, and establishing a system self-evolution mechanism to achieve intelligent and continuous optimization of the entire process design process. Attached Figure Description

[0017] Figure 1 This is a flowchart of the design method of the present invention; Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0018] The accompanying drawings in this invention are not strictly drawn to scale, and the specific dimensions and quantity of each structure can be determined according to actual needs. The drawings described in this invention are merely structural schematic diagrams.

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. Based on the described 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.

[0020] Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "inner," "outer," "upper," "lower," "far," "near," "front," and "rear" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] The application of existing large language models in the field of process design faces multiple challenges: First, they lack the ability to process professional multimodal data such as engineering drawings, 3D models, and point cloud data; second, general models lack manufacturing expertise, and the generated process solutions often do not meet actual processing requirements; third, they lack an effective mechanism for connecting with the company's existing process knowledge base and equipment database; and fourth, they lack a professional evaluation system and continuous optimization capabilities for process solutions.

[0022] This invention proposes a process intelligent design system driven by multi-source heterogeneous data and big language, which integrates key technologies such as multimodal data processing, multi-source heterogeneous database construction, and large model training and fine-tuning. It realizes intelligent processing of the entire process from part information extraction, process knowledge retrieval, process scheme generation to simulation verification and optimization, and can be widely applied to the processing technology design and optimization scenarios in fields such as industrial machine tools, robots, and high-end equipment manufacturing.

[0023] Combination Figures 1 to 2 As shown, a process intelligent design system driven by multi-source heterogeneous data and large language includes: The multimodal data processing module employs geometric feature recognition algorithms, automatic speech recognition technology, and convolutional neural networks to process 3D models, speech, natural language documents, part images, and 3D point cloud data, extracting structured information and converting it into a unified information representation format. Specifically, it uses a unified data model defined by JSON Schema, containing fields such as part ID, material properties, geometric feature list, processing requirements, and tolerance information. For example: {"part_id": "xxx","material": {"type":"45 steel", "hardness": "HRC28-32"}, "features": [{"type": "hole", "diameter": 10, "depth": 20, "tolerance": "H7"}],...};

[0024] Multi-source heterogeneous hybrid databases include relational databases and graph databases. Relational databases store part attribute information in structured information, while graph databases store process knowledge graphs. Relational databases and graph databases are linked through a unified data identifier and support joint queries. The large model building module is based on a pre-trained Transformer architecture large language model. It integrates process knowledge from a multi-source heterogeneous hybrid database through a retrieval enhancement generation mechanism and uses LoRA technology to efficiently fine-tune parameters in the processing technology domain to obtain a domain-adaptive model for processing technology design. Specifically, the pre-trained model used by the large model is Qwen3.0-30b. The intelligent machining process collaborative design system includes a multimodal interaction interface and an inference engine. The multimodal interaction interface is used to receive process requirements input by users, including natural language queries, drawing uploads, 3D model inputs, and voice inputs. The inference engine is used to call the joint query results of multi-source heterogeneous hybrid databases and perform inference based on the domain adaptation model to generate and output a detailed machining plan including process flow, equipment model, tooling fixtures, and tool parameters.

[0025] As an alternative embodiment, the multimodal data processing module includes a 3D model processing submodule, a speech processing submodule, a natural language document processing submodule, and a part image recognition and 3D point cloud processing submodule. Each submodule contains two core steps: data preprocessing and structured information extraction. The preprocessing step performs standardized processing such as format conversion, noise filtering, and missing value imputation for different data types. The structured information extraction step converts the processed data into a unified structured representation, facilitating subsequent storage, retrieval, and model training.

[0026] As an alternative implementation, the 3D model processing submodule incorporates a geometric feature recognition algorithm and constructs a feature recognition rule base and a deep learning model. The geometric feature recognition algorithm automatically extracts dimensional features, machining feature surfaces, geometric tolerances, and surface roughness from the 3D model data. The feature recognition rule base and deep learning model are combined to achieve automatic recognition and parameterized representation of machining features such as holes, slots, steps, and chamfers. For complex curved surface features, curvature analysis and feature decomposition techniques are used for classification and recognition. Further, taking a 3D model of a power tool holder as an example: the format of this 3D model is step, and the part features include: end faces, threaded holes, stepped holes inside the housing, through holes, keyways, and irregular shapes. Specifically, the processing procedure of the 3D model processing submodule is as follows:

[0027] (1) STEP model reading and geometric data extraction: STEP file is read using STEPControl_Reader of OCCT library and converted into TopoDS_Shape topology; based on B-Rep boundary representation, the hierarchical structure of face, edge and vertex is traversed to build a geometric feature database.

[0028] (2) Identify and determine the reference plane, identify all features in the axis direction based on the spindle axis, then use the intersection of point-plane projection and straight plane to calculate and determine the machining origin, and then determine the radial axis by coordinate system orthogonalization, and identify all features of the axis.

[0029] (3) Feature Classification and Recognition Algorithm: A multi-level feature classification strategy combining dynamic proportional thresholding and rule matching is adopted. First, the average radius and average surface area of ​​the solid cylinder are calculated by statistically analyzing the cylindrical surface features in the model. Based on the interquartile range method, outliers are filtered out, and the proportional threshold for hole recognition is dynamically set: a specific proportion of the hole radius being smaller than the average radius of the solid cylinder, a hole depth-to-diameter ratio being greater than a set threshold, and a proportional threshold of the hole surface area being smaller than the average surface area of ​​the solid cylinder. If two of these conditions are met, the hole is determined to be a hole feature. The recognition of threaded holes is further combined with the thread minor diameter database matching method. Based on the proportional judgment, the minimum difference is matched with the standard thread specification to achieve automatic recognition of thread parameters. For the distribution of hole groups, this invention adopts the symmetry detection method and the circumferential uniform distribution analysis method. The symmetry of the hole features about the origin, X-axis, and Y-axis is judged by coordinate transformation, and the uniform hole pattern is identified by polar coordinate transformation and angle uniformity test, providing a basis for machining path planning.

[0030] (4) Geometric Calculation and Topological Analysis: A coaxiality judgment algorithm is adopted. Parallelism is judged by calculating the dot product of the direction vectors of the two axes, and the distance between the axes is calculated by combining the point-to-line distance formula to achieve coaxiality judgment. For hole connectivity, a coaxiality and distance joint judgment method is adopted. The coaxiality test and the center distance analysis are combined to identify the characteristics of coaxial adjacent holes. The cylinder depth calculation adopts the axial projection range method. The boundary points of the cylinder surface are projected onto the axis, and the depth is calculated by the maximum and minimum projection difference. The cross-section analysis is combined to distinguish between through holes and blind holes. The point-plane relationship judgment adopts the projection intersection method. The point is projected onto the plane or the intersection of the line and the plane is calculated. The positional relationship of the point is verified by combining the tolerance. In addition, the present invention also realizes a feature deduplication and merging mechanism based on geometric attributes. By comparing geometric attributes such as axis, radius, and position, the duplicate identification of the same feature is avoided.

[0031] As an alternative implementation, the voice processing submodule uses automatic speech recognition technology to convert speech signals into text, and extracts key information such as part names, processing requirements, and process parameters through named entity recognition and relation extraction natural language processing technology. This module supports multi-turn dialogue interaction and can accurately understand user intent based on context.

[0032] As an alternative implementation, the natural language document processing submodule has a built-in pre-trained language model. The natural language document processing submodule uses document layout analysis, table recognition, and paragraph segmentation technology to perform structured parsing of process cards, process manuals, and process specification documents. The pre-trained language model is used to extract process knowledge elements such as process name, processing parameters, equipment requirements, and tool selection from the structured parsing results.

[0033] As an alternative implementation, the part image recognition and 3D point cloud processing submodule uses a convolutional neural network to identify the part category and detect appearance defects in images of physical parts; it uses point cloud registration, segmentation and feature extraction algorithms to reconstruct the geometric features of parts and measure their dimensions from 3D scanned point cloud data, providing data support for reverse engineering and quality inspection.

[0034] As an alternative embodiment, the multi-source heterogeneous hybrid database building module includes: This structured parts information database, based on MySQL, uses a relational database to store the structured attribute information of parts. It employs a star schema design, linking various dimensions of information around the part, and supports multi-condition composite queries and batch data processing. Specifically, the structured attribute information includes, but is not limited to: a basic parts information table, storing basic attributes such as part number, name, material, and batch size; a dimensional feature table, storing various dimensional parameters, tolerance requirements, and datum information; a machining feature table, storing feature types, feature parameters, and spatial relationships between features; and a multimedia data table, storing and linking files such as 2D drawings, 3D models, point cloud data, and physical images of the parts.

[0035] The Neo4j-based graph database process knowledge base stores process knowledge with complex relationships, including a machining process knowledge graph, an equipment and tool knowledge graph, and a process rule base. Specifically, the machining process knowledge graph uses process elements as nodes and process relationships as edges to construct a knowledge graph containing multiple types of entities such as materials, features, processes, parameters, equipment, and tools, as well as their relationships. The equipment and tool knowledge graph uses machining equipment, tooling fixtures, and measuring tools as core nodes to establish a knowledge network of equipment capabilities, applicable scope, and matching relationships. The process rule base stores process design criteria, constraints, and empirical rules in a graph structure, supporting rule reasoning and inheritance.

[0036] A hybrid database collaboration mechanism establishes a data synchronization and joint query mechanism between relational databases and graph databases. It achieves the connection between the part information database and the process knowledge base through a unified data identification system, and supports cross-database queries through a hybrid query engine. For example, it can query all possible process routes and corresponding equipment applicable to a certain material and characteristic. Specifically, the hybrid query engine parses user queries into subqueries through a unified query interface layer (custom API gateway), routing them to MySQL and Neo4j for execution respectively; it uses unified identifiers such as part numbers for cross-database association; and it employs a query orchestrator to coordinate the query order and result merging strategy between the two databases (first querying MySQL to obtain basic part information, then querying Neo4j based on feature type to obtain process routes).

[0037] As an alternative implementation, the large model building and optimization module includes: User Intent Understanding Module: Constructs a multi-layered intent recognition architecture to achieve accurate understanding of user process design needs. The first layer classifies query types, distinguishing between different request types such as information queries, solution generation, solution optimization, and problem consultation; the second layer performs fine-grained intent parsing, extracting key elements such as part type, processing requirements, constraints, and optimization goals; the third layer performs context fusion, combining dialogue history and user profiles to achieve accurate intent inference and ambiguity resolution.

[0038] Domain-specific knowledge infusion module: This module deeply integrates the aforementioned multi-source heterogeneous hybrid database with the large language model. It employs retrieval-enhanced generation technology to dynamically retrieve relevant process knowledge during model inference, injecting the retrieval results as contextual information into the model. A knowledge graph embedding module is constructed to convert process knowledge in the graph database into vector representations, enhancing the model's understanding of process relationships. A knowledge consistency verification mechanism is designed to ensure consistency between the model output and the domain knowledge base.

[0039] Model Fine-tuning Module: Based on efficient parameter fine-tuning techniques such as LoRA, this module injects specialized capabilities in the machining process domain while maintaining the basic model's performance. Specifically, it includes: constructing a high-quality process question-and-answer dataset covering typical scenarios such as process solution generation, parameter recommendation, and fault diagnosis; designing process-specific prompt templates to guide the model in outputting structured solutions according to process design specifications; and employing a training strategy combining instruction fine-tuning and reinforcement learning to improve the model's task compliance and output quality. Specifically, the LoRA rank (r) is 16, scaling factor α is 32, dropout rate is 0.05, learning rate is 2e-5, training epochs are 500, and the optimizer is AdamW. Specifically, the process-specific prompt template establishes a professional background by assigning a senior engineer role, thereby clarifying specific task objectives (such as process design or problem analysis). The key to the template is requiring users to provide detailed part technical data and production conditions as input, driving AI to generate accurate solutions. Subsequently, by specifying formatted output requirements (such as process route tables and key process analysis), it ensures that the answers are directly actionable.

[0040] The process evaluation system module establishes a multi-dimensional evaluation index system for process schemes, including: process accuracy evaluation, verifying whether the scheme meets the processing feasibility requirements and process specification constraints; process efficiency evaluation, assessing efficiency indicators such as processing time, tool change frequency, and clamping frequency; process economic evaluation, calculating cost factors such as material utilization rate, tool consumption, and energy consumption; and process stability evaluation, analyzing the sensitivity of process parameters and quality consistency. The evaluation results are used to guide the continuous optimization and training of the model.

[0041] The closed-loop process optimization module constructs a closed-loop iterative mechanism for generation, evaluation, and optimization. Based on the feedback results of the evaluation system, reinforcement learning methods are used to optimize the model online; a process solution library is established to accumulate verified feasible solutions as reference benchmarks for subsequent generation; a multi-objective optimization algorithm is designed to support trade-off optimization among multiple objectives such as efficiency, cost, and quality. Furthermore, the reinforcement learning method is RLHF (Reinforcement Learning Based on Human Feedback), with the matching degree between the process solution and the standard solution, whether the processing feasibility constraints are met (negative rewards are given for violating constraints), processing efficiency scores, and user feedback scores as the core parameters of the reward function.

[0042] As an alternative embodiment, the intelligent manufacturing process collaborative design system includes: System functional architecture: The front end provides a visual human-computer interaction interface, supporting multiple input methods such as drawing upload, voice input, and natural language interaction; the back end integrates core functional modules such as a multimodal data processing engine, a knowledge base management system, and large model inference services.

[0043] Full-process automation capability: The system achieves full-process automation from requirement input to solution output, specifically including: requirement parsing, automatically understanding the part information and processing requirements input by the user; knowledge retrieval, retrieving relevant process knowledge and historical cases from multi-source knowledge bases; process generation, automatically generating complete process solutions such as process routes, process contents, and process parameters based on a large model; and optimization output, automatically optimizing process parameters and outputting the final solution based on simulation results and evaluation feedback.

[0044] System self-evolution mechanism: Establish a continuous learning and self-optimization system evolution mechanism. User feedback, solution evaluations, and actual processing results accumulated during operation are fed back into the knowledge base after quality review, enriching the process knowledge reserve; the model is incrementally trained regularly based on newly accumulated data to continuously improve its professional capabilities; and user behavior analysis is used to continuously optimize interaction processes and recommendation strategies to enhance user experience.

[0045] This invention also proposes a process intelligent design method driven by multi-source heterogeneous data and large language, comprising the following steps: The multimodal data processing module preprocesses and extracts structured information from 3D models, speech, natural language documents, part images, and 3D point cloud data, and converts them into a unified structured representation. The multimodal data processing module preprocesses and extracts structured information from 3D models, speech, natural language documents, part images, and 3D point cloud data, converting them into a unified information representation format. Structured information is stored in a relational database with a multi-source heterogeneous hybrid database, and process knowledge is stored in a graph database, with connections established through unified data identifiers; Through the large model building module, based on the pre-trained Transformer architecture large language model, the process knowledge is integrated through the retrieval enhancement generation mechanism, and LoRA technology is used for efficient parameter fine-tuning to obtain a domain-adaptive model for processing technology design. The intelligent machining process collaborative design system receives user-inputted process requirements through its multimodal interactive interface, calls the joint query results of multi-source heterogeneous hybrid databases, and performs reasoning based on the domain adaptation model to generate and output a detailed machining plan including process flow, equipment model, tooling fixtures, and tool parameters.

[0046] Collect user feedback and actual processing data, continuously update the knowledge base and optimize the large model to achieve the system's self-evolution.

[0047] Specific Implementation Example 1 of this Invention: Example of a power tool holder housing part from a certain company. The 3D model of this power tool holder is in the .step format. The part features include: end face, threaded hole, stepped hole inside the housing, through hole, keyway, and irregular shape.

[0048] The extraction effect of the system according to the present invention is as follows: (1) Reference plane information, for example: Principal reference plane: [Plane_001]; Global center: (30.200,0.000,30.736); Local coordinates: (-0.000,15.264,21.800); Area: 152.34mm²; Normal vector: (-1.00,0.00,-0.00).

[0049] (2) Axial feature distribution (sorted by Z-axis coordinate) Example: Location: 001: Z=0.40mm; Entity Feature: [Cylinder_023]; Diameter: 54.00mm (Radius: 27.00mm); Length: 0.80mm; Type: Blind Hole; Surface Area: 67.9mm²; Local Coordinates: (-17.19, -0.00, 0.40); Location: 027: Z = 29.02mm; Cylindrical Hole Feature: [Hole_009]; Diameter: 3.30mm (Radius: 1.65mm); Depth: 10.10mm; Type: Through Hole; Aspect Ratio: 6.12; Radius Ratio: 0.15; Local Coordinates: (-24.03, -31.95, 29.02).

[0050] (3) Feature classification statistics Planar features: 30; Solid cylinders: 22; Cylindrical holes: 12; Threaded holes: 7; Curved surface features: 2; Total: 73 features.

[0051] As can be seen from the above extraction results, the system of the present invention extracts vital signs accurately and completely.

[0052] Specific Implementation Example 2 of this Invention: Taking a shaft-type part as an example: The input processing information is: process a solid shaft-type part. The characteristics of the part are: the size is φ59*330; a B-type center hole needs to be processed; the outer circle is a stepped shaft, with the dimensions being φ46*15, φ50*76, φ53*34, ​​and φ59*205 respectively, and the φ46 end face needs to be chamfered with C8.

[0053] The processing steps for this part, based on the system of this invention, are as follows: Process - Material cutting: Material dimensions: 45 steel round bar 60*340.

[0054] Process - Heat treatment: The parts are heat treated to HRC28-32.

[0055] Process - Turning 1: Clamp one end of φ60*300 with a hard-jaw chuck, turn the unclamped end face by 1mm, and machine the B-type center hole. The puller clamps the φ60 outer diameter, the hard-jaw chuck clamps the φ60*7 outer diameter, and the servo tailstock holds the center hole. Turn the φ46*15, φ50*76, φ53*34, ​​and φ59*205 outer diameters in sequence, and turn the C8 chamfer on the φ46 end face.

[0056] Process - Turning 2: Change the clamping edge, clamp and process one end of φ59*200, and turn the unclamped end face to the required part size of 330.

[0057] Process - Inspection: Upon receipt of goods, each purchased part must be inspected by an inspector and pass inspection. Only after the inspector issues an inspection report can the part be put into storage.

[0058] Process - Warehousing: After passing inspection, the parts are put into storage, and the inspection report is kept on file.

[0059] As can be seen from the above process, the intelligent design system of the present invention has a clear process design and can directly produce shaft parts.

[0060] The novelty of this invention includes the following aspects: (1) Multimodal Engineering Data Unified Processing Framework. This invention proposes for the first time a multimodal data unified processing framework for process design, which realizes collaborative processing and structured information extraction of multiple modal data such as 3D models, 2D drawings, voice, natural language documents, images, and point clouds, breaking through the bottleneck of traditional process design information acquisition.

[0061] (2) Multi-source heterogeneous hybrid database architecture for process design. It innovatively combines relational databases and graph databases to construct a hybrid database architecture that integrates part information database and process knowledge graph, so as to realize efficient storage and joint retrieval of structured part data and complex process relationship knowledge.

[0062] (3) Specialized training method for large-scale process design models. A specialized training method for large-scale models is proposed, which integrates technologies such as retrieval enhancement generation, knowledge graph embedding, and LoRA fine-tuning. This method enables the effective transfer of general-purpose large-scale language models to specialized process design models, significantly improving the professional capabilities of the models in the process field.

[0063] (4) Multi-dimensional process evaluation and closed-loop optimization mechanism. Establish a multi-dimensional process evaluation system covering accuracy, efficiency, economy and stability, and build a closed-loop optimization mechanism based on evaluation feedback to achieve continuous improvement of process schemes and continuous enhancement of model capabilities.

[0064] (5) End-to-end intelligent collaborative design system. Construct an end-to-end intelligent system from requirements analysis, knowledge retrieval, process generation, simulation verification to optimization output, and establish a system self-evolution mechanism to realize the intelligentization and continuous optimization of the entire process of processing technology design.

[0065] The above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions implemented in the present invention, and should all be covered within the protection scope of the present invention.

Claims

1. A process intelligent design system driven by multi-source heterogeneous data and large language, characterized in that, include: The multimodal data processing module uses geometric feature recognition algorithms, automatic speech recognition technology, and convolutional neural networks to process 3D models, speech, natural language documents, part images, and 3D point cloud data, extracting structured information and converting it into a unified information representation format. The structured information includes part ID, material properties, geometric feature list, processing requirements, and tolerance information. A multi-source heterogeneous hybrid database includes a relational database and a graph database. The relational database stores part attribute information in the structured information, and the graph database stores process knowledge graphs. The relational database and the graph database are linked through a unified data identifier to support joint queries. The large model building module is based on a pre-trained Transformer architecture large language model. It integrates process knowledge from the multi-source heterogeneous hybrid database through a retrieval enhancement generation mechanism and uses LoRA technology to efficiently fine-tune parameters in the processing technology domain to obtain a domain-adaptive model for processing technology design. The intelligent machining process collaborative design system includes a multimodal interaction interface and an inference engine. The multimodal interaction interface is used to receive process requirements input by the user. The inference engine is used to call the joint query results of the multi-source heterogeneous hybrid database and perform inference based on the domain adaptation model to generate and output a machining scheme including process flow, equipment model, tooling fixtures, and tool parameters.

2. The process intelligent design system driven by multi-source heterogeneous data and large language as described in claim 1, characterized in that, The multimodal data processing module includes a 3D model processing submodule, a speech processing submodule, a natural language document processing submodule, and an image point cloud processing submodule.

3. The process intelligent design system driven by multi-source heterogeneous data and large language as described in claim 2, characterized in that, The 3D model processing submodule includes a geometric feature recognition algorithm, a feature recognition rule base, and a deep learning model. The geometric feature recognition algorithm is used to automatically extract the dimensional features, machining feature surfaces, geometric tolerances, and surface roughness of the 3D model data. The feature recognition rule base and the deep learning model are combined to achieve automatic recognition and parameterized expression of machining features such as holes, slots, steps, and chamfers.

4. The process intelligent design system driven by multi-source heterogeneous data and large language as described in claim 2, characterized in that, The speech processing submodule uses automatic speech recognition technology to convert speech signals into text, and extracts key information such as part names, processing requirements and process parameters through named entity recognition and relation extraction natural language processing technology.

5. The process intelligent design system driven by multi-source heterogeneous data and large language as described in claim 2, characterized in that, The natural language document processing submodule is equipped with a pre-trained language model. The natural language document processing submodule uses document layout analysis, table recognition, and paragraph segmentation technology to perform structured parsing of process cards, process manuals, and process specification documents. The pre-trained language model is used to extract process knowledge elements such as process name, processing parameters, equipment requirements, and tool selection from the structured parsing results.

6. The process intelligent design system driven by multi-source heterogeneous data and large language as described in claim 2, characterized in that, The image point cloud processing submodule uses a convolutional neural network to identify the type of the physical part and detect appearance defects in the image; it uses point cloud registration, segmentation and feature extraction algorithms to reconstruct the geometric features of the part and measure its dimensions based on the 3D scan point cloud data.

7. The process intelligent design system driven by multi-source heterogeneous data and large language as described in claim 1, characterized in that, The multi-source heterogeneous hybrid database includes: The structured parts information database based on MySQL uses a relational database to store the structured attribute information of parts; it adopts a star schema design, with parts as the core and linking information of various dimensions, and supports multi-condition compound queries and batch data processing; The Neo4j-based graph database process knowledge base uses a graph database to store process knowledge with complex relationships, including a processing technology knowledge graph, an equipment and tool knowledge graph, and a process rule base. A hybrid database collaboration mechanism is established to enable data synchronization and joint querying between relational databases and graph databases. A unified data identification system is used to link the parts information database and the process knowledge base, and a hybrid query engine supports cross-database queries.

8. The process intelligent design system driven by multi-source heterogeneous data and large language as described in claim 1, characterized in that, The large model construction module includes: The user intent understanding module includes a three-tiered intent recognition architecture: query type classification, fine-grained intent parsing, and context fusion, to achieve an understanding of user process design requirements. The knowledge infusion module deeply integrates the multi-source heterogeneous hybrid database with the large language model, and uses retrieval enhancement generation technology, knowledge graph embedding module and knowledge consistency verification mechanism to improve the model's process understanding ability; The model fine-tuning module, based on LoRA parameter efficient fine-tuning technology, combines process question-and-answer datasets and process-specific prompt word templates, and adopts a training strategy that combines instruction fine-tuning and reinforcement learning to achieve model domain adaptation.

9. A process intelligent design method driven by multi-source heterogeneous data and large language, based on the process intelligent design system driven by multi-source heterogeneous data and large language as described in claim 1, characterized in that, Includes the following steps: The multimodal data processing module preprocesses and extracts structured information from 3D models, speech, natural language documents, part images, and 3D point cloud data, converting them into a unified information representation format. The structured information is stored in the relational database of the multi-source heterogeneous hybrid database, and the process knowledge is stored in the graph database, and the association is established through a unified data identifier; Through the large model building module, based on the pre-trained Transformer architecture large language model, the process knowledge is integrated through the retrieval enhancement generation mechanism, and LoRA technology is used for efficient parameter fine-tuning to obtain a domain-adaptive model for processing technology design. The intelligent machining process collaborative design system receives user-inputted process requirements through its multimodal interaction interface, calls the joint query results of the multi-source heterogeneous hybrid database, and performs reasoning based on the domain adaptation model to generate and output a machining scheme including process flow, equipment model, tooling fixtures, and tool parameters.

Images