CN117235929B

CN117235929B - Three-dimensional CAD (computer aided design) generation type design method based on knowledge graph and machine learning

Info

Publication number: CN117235929B
Application number: CN202311246840.1A
Authority: CN
Inventors: 王宇; 于子淳; 王挺; 曾鹏; 武琼; 邵一凡
Original assignee: Shenyang Intelligent Robot Innovation Center Co ltd; Shenyang Institute of Automation of CAS
Current assignee: Shenyang Intelligent Robot Innovation Center Co ltd; Shenyang Institute of Automation of CAS
Priority date: 2023-09-26
Filing date: 2023-09-26
Publication date: 2024-06-04
Anticipated expiration: 2043-09-26
Also published as: CN117235929A

Abstract

The invention discloses a three-dimensional CAD (computer aided design) generation type design method based on knowledge graphs and machine learning. And analyzing the continuously fed back user design requirement through an NLP technology, and obtaining the optimal design parameters and the optimal targets until a 3D CAD model meeting the user design requirement is output. The system has an automatic design function, can improve the design efficiency and reduce the design difficulty. Meanwhile, the system can be combined with a 3D printing technology to realize one-stop service for structural design and manufacture of products. The method and the system have the advantages of high efficiency, accuracy, flexibility, intelligence and the like, can be widely applied to product design and manufacture in various fields, and have wide application prospects and market values.

Description

Three-dimensional CAD (computer aided design) generation type design method based on knowledge graph and machine learning

Technical Field

The invention belongs to the field of computer aided design, and particularly relates to a 3D CAD generation type design method and system based on knowledge graph and machine learning.

Background

The conventional 3D CAD design method is that a designer spends much time and effort manually modeling and modifying, and continuously optimizes the 3D CAD model in a trial-and-error manner by adjusting design parameters. This design method requires a lot of time and experience, the design process is cumbersome and it is difficult to ensure the correctness and consistency of the generated model. With the development of artificial intelligence technology, AI is also widely used in the field of industrial design. The AI-based generative design method can automatically generate and optimize the 3D CAD model, thereby improving the design efficiency and accuracy.

However, most of the current AI-based 3D CAD-generated design methods only consider geometric features, process requirements, and other aspects, and ignore knowledge features of product design. The knowledge points involved in product design are very wide, including aspects of materials, functions, structures, processes, etc., and complex interrelationships exist between the knowledge points. Therefore, how to use knowledge graph technology to fuse the knowledge points into the generated design method and improve the correctness and consistency of the generated 3D CAD model is a problem to be solved by the current 3D CAD generated design method and system.

Disclosure of Invention

Aiming at the requirements that the existing 3D CAD model design process is time-consuming and tedious and the correctness and consistency are difficult to ensure, the invention provides a 3D CAD generation type design method and system based on knowledge graph and machine learning. The knowledge points in the structural design of the product are fused through the knowledge graph, a machine learning algorithm is utilized to train a generated model, the design parameters and the optimization targets are automatically analyzed through continuous feedback man-machine design demand interaction by utilizing the NLP technology to improve the design result until a 3D CAD model meeting the design requirements of a user is generated, and the correctness and consistency of the generated design method are improved. The specific implementation steps of the invention are as follows:

The technical scheme adopted by the invention for achieving the purpose is as follows: the three-dimensional CAD generation type design method based on the knowledge graph and the machine learning comprises the following steps:

Step 1: acquiring product manufacturing process data containing workpiece structural design and constraint conditions;

Step 2: according to the data of the product manufacturing process, adopting logic atlas reasoning 3D model structure method, converting expert rules into triple structured data, and based on the constructed logic atlas, reasoning to obtain 3D model structure and parameters;

step 3: the method comprises the steps of adopting a CAD model matching method based on multi-view and processing feature recognition, obtaining a three-dimensional model of a part according to user requirements, calculating processing features of the part model to be matched by utilizing a multi-view convolutional neural network, and matching the part model to be matched by utilizing a three-dimensional model database to obtain processing information of the part model to be matched;

Step 4: and according to the processing information of the part model to be matched, the 3D model structure and parameters, adopting cloud containerization deployment to generate a model, and outputting a 3D CAD model.

The CAD model matching method based on multi-view and processing feature recognition comprises the following steps:

1) Acquiring a three-dimensional model of a part, constructing a three-dimensional model database, and constructing and training a ResNet-based multi-view convolutional neural network based on the database;

2) And calculating the processing characteristics of the part models to be matched by using the trained multi-view convolutional neural network, and matching the part models to be matched by using a three-dimensional model database to obtain the processing information of the part models to be matched.

Said step 1) comprises the steps of:

1.1 Acquiring a three-dimensional model of the part modeled by adopting a boundary representation method, classifying according to the attribute, taking the attribute as a classification label, constructing a three-dimensional model database, and converting all models in the model database into an STL format;

1.2 Extracting processing characteristics in each model by utilizing semantic information and topological structures in the three-dimensional model of the part;

1.3 Taking the three-dimensional model of the part as input and the classification label of the three-dimensional model as output, and training a ResNet-based multi-view convolutional neural network;

1.4 According to the trained neural network model, calculating the feature vector of each part three-dimensional model to complete the construction of a database and the training of the multi-view convolutional neural network.

Said step 2) comprises the steps of:

2.1 Converting the three-dimensional model to be matched of the boundary representation type into an STL format;

2.2 Classifying the three-dimensional model to be matched by adopting a multi-view convolutional neural network and calculating the feature vector;

2.3 Ordering the distances between the matched feature vectors in the database and the feature vectors of the models to be matched from small to large, and selecting a plurality of models before the models to finish rough matching of the three-dimensional models;

2.4 Extracting processing characteristics of the three-dimensional model according to the adjacent topological relation of the surfaces in the three-dimensional model to be matched and semantic information marked by colors and texts;

2.5 According to the extracted processing characteristics, further matching in the similar model obtained by rough matching to obtain a model with the type, the number and the size of the processing characteristics closest to the model to be matched, completing matching, and obtaining corresponding processing data from a model library.

The method for reasoning the 3D model structure by utilizing logic atlases comprises the following steps:

1) Acquiring expert rules, and automatically extracting and manually trimming and correcting by using a natural language processing technology to convert the expert rules into triple structured data;

2) Storing the triad structured data into logic atlas and visualizing it;

3) Reasoning is carried out based on the established logic map;

4) Inputting the requirements and conditions of the mechanical structure to be designed, and obtaining the 3D model structure and parameters by reasoning and visualizing the results.

Said step 1) comprises the steps of:

1.1 Text data in the mechanical manual is acquired and stored in a readable electronic document format, the text data comprising: design parameters, material characteristics, and process requirements;

1.2 Sequentially cleaning the text data;

1.3 Analyzing the cleaned text data by a natural language processing method, and identifying and extracting keywords, phrases and grammar structures in the text data as rules;

1.4 Converting the extracted rule into triple structured data and storing the triple structured data into an SQL database;

1.5 Using the actual mechanical design cases to verify and correct the rules in the SQL database.

The step 4 comprises the following steps:

(1) Packaging the generated model into a container, and deploying on a cloud server;

(2) The design requirement of the user is converted into design parameters, an optimization target and processing information of the part model to be matched after natural language processing, and the processing information is used as input of a generated model to generate and output a 3D CAD model; the design requirements of the user include text language, excel documents and 3D auxiliary models.

A three-dimensional CAD-generated design system based on knowledge graph and machine learning, comprising:

the process data acquisition module is used for acquiring product manufacturing process data containing the structural design and constraint conditions of the workpiece;

The map reasoning module is used for converting expert rules into triple structured data by adopting a logic map reasoning method of the 3D model structure according to the data of the product manufacturing process, and reasoning the three-dimensional structured data based on the constructed logic map to obtain the 3D model structure and parameters;

The feature matching module is used for acquiring a three-dimensional model of the part according to the user requirement by adopting a CAD model matching method based on multi-view and processing feature recognition, calculating the processing features of the part model to be matched by utilizing a multi-view convolutional neural network, and matching the part model to be matched by utilizing a three-dimensional model database to obtain the processing information of the part model to be matched;

The model output module is used for generating a model by adopting cloud containerization deployment according to the processing information of the part model to be matched, the 3D model structure and parameters and outputting a 3D CAD model.

A three-dimensional CAD generating type design device based on knowledge graph and machine learning comprises a memory and a processor; the memory is used for storing a computer program; the processor is configured to implement the three-dimensional CAD-generated design method based on knowledge graph and machine learning as claimed in any one of claims 1 to 7 when executing the computer program.

A computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the knowledge-graph and machine-learning based three-dimensional CAD-generated design method of claim.

A manufacturing process knowledge graph construction method for multi-mode data comprises the following steps:

s1: acquiring product manufacturing process data, and establishing an original data database;

S2: based on manufacturing process data in an original data database, building an ontology model by combining expert knowledge in the manufacturing process field, and building a knowledge graph model layer and a keyword dictionary of various entities;

s3: according to the knowledge graph mode layer, extracting manufacturing process data in an original data database by using a data extraction model, and constructing a knowledge graph data layer;

s4: combining the knowledge graph mode layer and the knowledge graph data layer to construct a knowledge graph, inputting the extracted RDF triples into a graph database, storing the knowledge graph and visually displaying.

In step S1, the product manufacturing process data includes: structured data, represented in the form of process database files, unstructured data, represented in the form of text.

The S2: based on manufacturing process data in an original data database, an ontology model is built by combining expert knowledge in the manufacturing process field, and a knowledge graph model layer and a keyword dictionary of various entities are built, comprising the following steps:

s2.1: manufacturing process data in an original data base are divided into six categories: workpiece data, feature data, manufacturing equipment data, manufacturing method data, process equipment data, and semantic relationship data;

S2.2: according to the data classification result in the original database, combining expert knowledge and the entity type of the manufacturing process database, and establishing an ontology model;

S2.3: and constructing a model layer of the knowledge graph in an ontology model mode to describe entity classes and relations among the entity classes.

The S2.2: according to the data classification result in the original database, and then combining expert knowledge and the entity type of the database of the existing manufacturing process of the enterprise, an ontology model is built, and the method comprises the following steps:

The six types of data are abstracted into six major types, wherein workpiece data, feature data, manufacturing equipment data, manufacturing method data, process equipment data and semantic relation data correspond to each other in sequence: workpiece class, feature class, equipment class, process class, tool class and relation class, wherein any class is set with subclasses according to requirements; establishing a keyword dictionary of six types of entities, wherein the dictionary comprises standard keywords and synonyms thereof;

The attributes of the class comprise class attributes and instance attributes, the class attributes are shared by various classes and subclasses thereof, all the instances share the corresponding class attributes, and the instance attributes are only all of the instances.

The ontology model of the mode layer of the knowledge graph is formally expressed as:

KGPattern＝{Entity∪Relation}

Entity＝{W∪F∪M∪O∪E}

Relation＝{(W_i,consist_of,W_j)∪(W_i,consist_of,F_j)∪(W_i,consist_of,O_j)U(O_i,

consist_of,O_j)∪(O_i,consist_of,F_j)∪(O_i,consist_of,M_j)∪(O_i,consist_of,E_j)∪(O_i,order,O_j)∪(W_i,assemble,O_j,assemble,W_j)∪}

In the above formula: KGPATTERN denotes a formalized expression model of a knowledge graph mode layer, entity denotes an Entity set of mode layer descriptions, and Relation denotes a relation set of mode layer descriptions; w represents workpiece class, F represents feature class, M represents equipment class, O represents process class, E represents tool class, and R represents relation class; the semantic relationship includes consist _ of, order, assemble; wherein consist _of represents inclusion, order represents sequence relation, assemble represents assembly relation, and i and j represent serial numbers.

The S3: according to the knowledge graph mode layer, extracting manufacturing process data in an original data database by using a data extraction model, and constructing a knowledge graph data layer;

S3.1: the structural data processing module converts manufacturing process data of a relational database in an original data database into RDF triples by adopting D2R according to a preset knowledge graph mode layer;

S3.2, respectively extracting data contained in the workpiece model and drawing of the original database and text data in the original database by the unstructured data processing module to generate RDF triples;

S3.3: the three-dimensional CAD model and drawing processing module extracts text information, structure tree information and characteristic information in the three-dimensional CAD model according to the knowledge pattern layer and the keyword dictionary, and establishes triplet expression of workpiece class and characteristic class entities, attributes and relations;

s3.4: the natural language processing module carries out entity labeling on unstructured text according to information in a keyword dictionary, converts labeled text data into a data set with word segmentation labels BIO, and divides the data set into a training set, a testing set and a verification set for training an entity recognition model; extracting workpiece class, feature class, process class, equipment class and tool class entities in unstructured data by adopting a trained entity recognition model; determining the relation among the entities according to the knowledge graph mode layer to obtain a triplet composed of the entities and the relation;

S3.5: the knowledge fusion module fuses the knowledge which is extracted by the structured data processing module and the unstructured data processing module and possibly has overlapping.

The step S3.5 comprises the following steps:

Carrying out semantic acquaintance calculation on the entity, the relation and the attribute extracted by the structured data processing module and the unstructured data processing module and standard keywords and synonyms thereof in a keyword dictionary, and replacing the extracted entity, relation and attribute words with the standard keywords corresponding to the synonyms according to a calculation result to realize cleaning alignment;

Traversing the three-dimensional CAD model and the drawing processing module for the unstructured data processing module, forming entity pairs by the extracted workpiece class and feature class entities and the workpiece class and feature class entities extracted by the natural language processing module, scoring the entity pairs according to a bilinear matching algorithm, and unifying the names of the two entities in the entity pairs according to a score result so as to finish the alignment of the extracted data of the unstructured data processing module;

And the structured data processing module and the unstructured data processing module are aligned, the aligned workpiece class and feature class entity attributes keep consistent with the workpiece class and feature class entity attributes extracted by the three-dimensional CAD model and the drawing processing module, and entity fusion is completed.

A manufacturing process knowledge graph construction system facing multi-mode data comprises:

the database module is used for acquiring the data of the manufacturing process of the product and establishing an original data database;

the model layer construction module is used for constructing an ontology model based on manufacturing process data in an original data database and combining expert knowledge in the field of manufacturing processes to construct a knowledge graph model layer and a keyword dictionary of various entities;

The data layer construction module is used for utilizing the data extraction model to extract manufacturing process data in the original data database according to the knowledge graph mode layer and constructing a knowledge graph data layer;

The knowledge graph construction module is used for combining the knowledge graph model layer and the knowledge graph data layer to construct a knowledge graph, inputting the extracted RDF triples into a graph database, storing the knowledge graph and visually displaying

The data extraction model includes:

The structured data processing module is used for converting the data in the relational database into RDF triples by adopting D2R according to a preset knowledge graph mode layer;

an unstructured data processing module comprising:

The three-dimensional CAD model and drawing processing module is used for extracting data contained in the workpiece model and drawing of the original database and generating RDF triples;

The natural language processing module is used for extracting text data in the original database and generating RDF triples;

and the knowledge fusion module is used for fusing the possibly overlapped knowledge extracted by the structured data processing module and the unstructured data processing module.

A CAD model matching method based on multi-view and processing feature recognition comprises the following steps:

Said step 1) comprises the steps of:

The format of the three-dimensional model of the part comprises: STEP, iges or other CAD software specific data formats.

Said step 1.2) comprises the steps of:

1.2.1 Traversing all geometric surfaces in the model and the geometric surfaces adjacent to each geometric surface to construct an adjacent graph of the model part;

1.2.2 Aiming at the adjacency relation of the geometric surface in the target to be searched, constructing a feature adjacency graph of the target feature;

1.2.3 Selecting the geometric surface with the greatest adjacency or the first geometric surface in the traversing sequence of the adjacency surfaces as an initial point in the geometric surface of the model, searching sub-graphs in the adjacency graph of the part, and inquiring the minimum condition sub-graph matched with the characteristic adjacency graph to be used as a processing characteristic;

1.2.4 Extracting the processing characteristics of the three-dimensional model of the part through semantic information of color and text labeling.

Said step 1.3) comprises the steps of:

1.3.1 ResNet-based multi-view convolutional neural network deep learning model is built, and a classification label of the three-dimensional model of the part is used as a classification label of the multi-view convolutional neural network deep learning model;

1.3.2 Multi-view of the three-dimensional model of the component in the STL format is obtained, and the multi-view convolutional neural network deep learning model is trained based on the multi-view and the labels.

Said step 2) comprises the steps of:

A CAD model matching system based on multi-view and machined feature recognition, comprising:

The model and database construction module is used for acquiring a three-dimensional model of the part, constructing a three-dimensional model database, constructing and training a ResNet-based multi-view convolutional neural network based on the database;

And the matching module is used for calculating the processing characteristics of the part models to be matched by using the trained multi-view convolutional neural network, and matching the part models to be matched by using the three-dimensional model database to obtain the processing information of the part models to be matched.

The model and database construction module comprises:

The database construction module is used for acquiring a three-dimensional model of the part modeled by adopting the boundary representation method, classifying according to the attribute, taking the attribute as a classification label, constructing a three-dimensional model database, and converting all models in the model database into STL format;

the feature vector extraction module is used for extracting processing features in each model by utilizing semantic information and topological structures in the three-dimensional model of the part;

The model training module is used for taking a three-dimensional model of the part as input, taking a classification label of the three-dimensional model as output, and training a ResNet-based multi-view convolutional neural network;

And the feature vector calculation module is used for calculating the feature vector of each part three-dimensional model according to the trained neural network model, and completing the construction of a database and the training of the multi-view convolutional neural network.

The feature vector extraction module includes:

The model adjacency graph construction module is used for traversing all geometric surfaces in the model and the geometric surfaces adjacent to each geometric surface to construct an adjacency graph of the model part;

the feature adjacency graph construction module is used for constructing a feature adjacency graph of the target feature aiming at the adjacency relation of the geometric surface in the target to be searched;

The processing feature calculation module is used for selecting the geometric surface with the greatest adjacency or the first geometric surface in the traversing sequence of the adjacency surfaces from the geometric surfaces of the model as a starting point, searching sub-graphs in the adjacency graph of the part, and inquiring the minimum condition sub-graph matched with the feature adjacency graph to be used as a processing feature;

and the processing feature extraction module is used for extracting the processing features of the three-dimensional model of the part through semantic information of color and text labeling.

The matching module comprises:

The format conversion module is used for converting the three-dimensional model to be matched of the boundary representation type into an STL format;

The model feature vector calculation module to be matched is used for classifying and calculating feature vectors of the three-dimensional model to be matched by adopting the multi-view convolutional neural network;

The rough matching module is used for sequencing the distances between the matching feature vectors in the database and the feature vectors of the models to be matched from small to large, selecting a plurality of models before, and completing rough matching of the three-dimensional models;

The processing feature extraction module of the model to be matched is used for extracting processing features of the model to be matched according to the adjacent topological relation of the surfaces in the three-dimensional model to be matched and semantic information marked by colors and texts;

And the secondary matching module is used for further matching in the similar model obtained by rough matching according to the extracted processing features, obtaining a model with the type, the number and the size of the processing features closest to the model to be matched, completing matching, and obtaining corresponding processing data from a model library.

A method for reasoning about 3D model structure using logic atlases, comprising the steps of:

2) Storing the triad structured data into logic atlas and visualizing it;

3) Reasoning is carried out based on the established logic map;

Said step 1) comprises the steps of:

1.2 Sequentially cleaning the text data;

Said step 2) comprises the steps of:

2.1 Defining node and relation types of the triple structured data model and node attribute types;

2.2 Using Python to parse the triplet structured data into the form of nodes and relationships;

2.3 Creating nodes and relations, designating node types, attributes and labels, designating relation types, attributes and labels, and designating relations between a starting node and an ending node by using a Python driver;

2.4 Using Python to store the nodes and relationships into a graph database, mapping the data structure of the triplet structured data in logic atlas.

Constructing the logic map includes: data reading, ontology construction and map storage, wherein the ontology construction specifically comprises the following steps: abstracting the processed triple structured data into different node classes and relation classes, wherein the node classes comprise:

basic entity class, the content is basic noun, including part, component name, part attribute name;

The judgment/formula class is a group of names of judgment selection or formula calculation targets in the design flow and is used for connecting judgment selection or formula calculation;

Judging the calculation class, wherein the content is the actual content of judgment selection and formula calculation in the design flow;

the reasoning result class, the content is the characteristic, the size, the model and the structure of the part;

event class, the content is the name of the design part, design structure or work piece;

The output class, which is a marker node, marks the end of the partial inference.

The relationship class node comprises:

The semantic relation class is used for connecting nodes and expressing the relation among the nodes, and specifically comprises the following steps: a causal relationship class, a context relationship class, a compliance relationship class, and a conditional relationship class.

The step 3) is specifically as follows: based on the established logic map, the system converts the map into an expert system for reasoning, and the method comprises the following steps:

3 a.1) inquiring logic the subgraphs of the part nodes to be inquired in the map;

3 a.2) inquiring all nodes with the attribute of 'inference results' in the subgraph, inquiring all paths with the starting points of event names and the end point of each path of 'inference results';

3 a.3) respectively reading and recording all ' formula/judging ' nodes in each path, taking the ' formula/judging ' nodes as the condition of each part type selection, and storing all the formula/judging ' nodes into the necessary condition of the reasoning result;

3 a.4) repeating the steps 3 a.2) to 3 a.3) for all the nodes of the 'reasoning result' until all paths of each node of the 'reasoning result' are inquired;

3 a.5) taking a 'formula/judgment' node in each path of each reasoning result as a rule of the reasoning result, and converting the use of the rule into an expert system;

3 a.6) using an inference engine in the expert system to infer and interpret the necessary conditions of each inference result;

3 a.7) carrying out visual display on the reasoning result.

The step 3) is specifically as follows: reasoning is carried out based on the constructed logic map, and the reasoning method comprises the following steps:

3 b.1) selecting an event node corresponding to logic patterns of the part to be inferred, selecting a logic pattern subgraph of the event node, searching all basic entity nodes of the subgraph, reading the content, and obtaining all data corresponding to the basic entity nodes;

3 b.2) searching and selecting a judgment computing node connected with the event node in the step 3 b.1), respectively inquiring the basic entity node connected with the node again, matching the basic entity node with the data acquired in the step 3 b.1), and recording;

3 b.3) searching for a "judgment/formula" node connected with the "judgment calculation" in step 3 b.2), reading the content of a single "judgment/formula" node according to a random sequence, and bringing the corresponding data in step 3 b.2) into the content of the "judgment/formula" node;

3 b.4) recording the content of the node which accords with the judgment/formula in the step 3 b.3), and terminating the reading and inquiring process of the step 3 b.3);

3 b.5) searching the downstream node of the step 3 b.4) and recording, and continuing searching the downstream node;

3 b.6) searching the downstream "judgment calculation" node of the step 3 b.5), repeating the steps 3 b.2) to 3 b.5), terminating reasoning when the "output" node is inquired, and storing the names of all the recorded nodes;

3 b.7) searching for a path with a starting point being a selected event node, wherein the path contains all the recorded nodes in the step 3 b.6), the end point is a path of an output node, inquiring the path of an inference result node, and outputting the content of the inference result node, namely the inference result;

3 b.8) generating natural language by using the node stored in the step 3 b.6) by using a natural language processing technology, and outputting the reasoning process and logic of the natural language as the reason and rule support of the reasoning result.

A system for reasoning about 3D model structures using logic atlases, comprising:

the rule extraction processing and structuring module is used for obtaining expert rules and converting the expert rules into triple structured data by using a mode of combining natural language processing technology automatic extraction and manual fine adjustment correction;

The structural data is converted into logic map module which is used for storing the triad structural data into a logic map and visualizing the triad structural data;

the map reasoning module is used for reasoning based on the established logic maps;

and the reasoning result visualization output module is used for inputting the requirements and conditions of the mechanical structure to be designed, and visualizing the result through reasoning to obtain the 3D model structure and parameters.

The structured data is converted into logic map modules comprising:

the node definition module is used for defining nodes and relation types of the triple structured data model and node attribute types;

The data analysis module is used for analyzing the triple structured data into the form of nodes and relations by using Python;

The node creation module is used for connecting the graph database, creating nodes and relations by using a Python driver, designating node types, attributes and labels, designating relation types, attributes and labels, and designating the relation between a starting node and an ending node;

And the data mapping module is used for storing the nodes and the relations into a graph database by using the Python and mapping the data structure of the triple structured data in the logic map.

The map reasoning module comprises:

The automatic conversion module is used for inquiring the subgraph of the part node to be inquired in the logic map; inquiring all nodes with the attribute of 'reasoning result' in the subgraph, inquiring all paths with the initial point of event name and the end point of each path of 'reasoning result'; reading and recording all 'formula/judging' nodes in each path respectively, taking the 'formula/judging' nodes as the condition of each part type selection, and storing all the nodes into the necessary condition of the reasoning result; taking a formula/judgment node in each path of each reasoning result as a rule of the reasoning result, and converting the rule into an expert system based on the rule;

the reasoning module is used for reasoning and explaining the necessary conditions of each reasoning result by using a reasoning machine in the expert system;

And the inquiry and display module is used for visually displaying the reasoning results.

The map reasoning module comprises:

The reasoning process triggering module is used for selecting an event node corresponding to logic patterns of the part to be reasoning, selecting a logic pattern sub-graph of the event node, searching all basic entity nodes of the sub-graph, reading the content and obtaining all data corresponding to the basic entity nodes;

The reasoning module is used for searching and selecting the 'judgment calculation' nodes connected with the 'event' nodes, respectively inquiring the 'basic entity' nodes connected with the nodes again, matching the data acquired by the reasoning process triggering module and recording; searching a judgment/formula node connected with the judgment calculation, reading the content of a single judgment/formula node according to a random sequence, and bringing corresponding data into the content of the judgment/formula node; recording the content conforming to the node of the judgment/formula, and terminating the reading and inquiring process; continuing to search for the downstream node; searching a downstream judgment calculation node, terminating reasoning when an output node is inquired, and storing the names of all recorded nodes; searching for event class nodes with a starting point being selected, wherein the path contains all recorded nodes, the end point is a path of an output node, inquiring the inference result class node of the path, and outputting the content of the inference result class node, namely, the inference result;

And the query and display module is used for generating natural language by using a natural language processing technology by using the stored nodes, and outputting the reasoning process and logic of the natural language as the reason and rule support of the reasoning result.

The invention has the following beneficial effects and advantages:

1. The method utilizes the knowledge graph to fuse the knowledge characteristics of the product design, including knowledge points of materials, functions, structures, processes and the like related to the product structural design. Compared with the traditional generating type design method only considering geometric features, the accuracy and generalization of the generating type design method are improved.

2. The 3D CAD model output by the generated model not only has the appearance design of the product, but also comprises the internal structural design of the product, thereby ensuring the integrity and manufacturability of the product design in actual manufacture.

3. According to the system provided by the invention, the output result of the 3D CAD model can be continuously improved according to the design requirement fed back by the user, so that the professional skill requirement of the user is reduced, the user experience of man-machine interaction is improved, and more accurate and efficient design service is provided for the user.

4. The system can be combined with a 3D printing technology to realize rapid manufacturing, and the product design and manufacturing period is greatly shortened.

Drawings

FIG. 1 is a technical roadmap of the design method of the invention;

FIG. 2 is a flow schematic and functional block diagram of the design system of the present invention;

FIG. 3 is a flow chart of one embodiment of a weld fixture design implemented using the design method and system of the present invention;

fig. 4 is a flowchart of a knowledge graph construction method according to an embodiment of the present invention;

FIG. 5 is a flow chart for constructing a knowledge-graph model layer according to an embodiment of the present invention;

FIG. 6 is a diagram of the mind of a knowledge graph entity model provided by an embodiment of the present invention;

FIG. 7 is a flow chart of the construction of the knowledge graph data layer provided by the embodiment of the invention;

fig. 8 is a schematic structural diagram of a knowledge graph construction system according to an embodiment of the present invention;

FIG. 9 is a flow chart of the method of the present invention;

FIG. 10 is a block diagram of the present system;

FIG. 11a is a schematic diagram I of matching processing features based on topology information;

the serial numbers 1 to 8 respectively represent different surfaces, wherein the cylindrical surface 9, the plane 11 and the cylindrical surface 10 jointly form a countersunk hole machining characteristic, as shown in a dotted line surrounding;

FIG. 11b is a schematic diagram II of matching processing features based on topology information;

FIG. 12 is a schematic diagram of a ResNet-based MVCNN structure;

FIG. 13 is a schematic diagram of the system of the present invention;

FIG. 14 is a system architecture, functionality, and flow diagram for use;

FIG. 15 is a diagram of an exemplary triplet data set after NLP processing in an automobile welding fixture design manual;

FIG. 16 logic is a graph node relationship diagram;

The triplet data of fig. 17 is converted into logic map representation.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the described embodiments are some, but not all embodiments of the present invention. Accordingly, all other embodiments that may be made by one of ordinary skill in the art without undue burden are within the scope of the present invention.

A3D CAD generation type design method and system based on knowledge graph and machine learning includes:

step 1: structural design related data and constraints are collected, including design parameters, optimization objectives, constraints, and the like.

Step 2: and automatically extracting and constructing a knowledge graph, integrating knowledge points of related materials, functions, structures, processes and the like into the knowledge graph, and forming a knowledge base taking product design as a core.

Step 3: the generative model is trained using a machine learning algorithm. The training process is divided into a plurality of stages, each stage employing a different strategy and algorithm.

Step 4: the cloud containerized deployment generates a model, and the system analyzes design parameters and optimization targets through Natural Language Processing (NLP) technology according to design intent input by a user, and infers and outputs a 3D CAD model.

Step 5: and the user continuously corrects the design parameters and the optimization targets through continuous design demand feedback until a 3D CAD model meeting the design requirements of the user is output.

Step 6: the system is directly connected with a 3D printer to print the model, so that one-stop service for designing and manufacturing the product is realized.

The step 2 specifically includes:

step 2.1: the knowledge extraction comprises knowledge points of materials, functions, structures, processes and the like related to the structural design of the product.

Step 2.2: the extraction of knowledge adopts a mode of combining rule-based, instance-based and semantic-based to construct a map of defined rules and logics, existing design cases and semantic information between entities and relations.

Step 2.3: the knowledge graph constructed by the invention is a knowledge graph taking the structural design knowledge of the product as a entity, and is one of training data of a follow-up generation algorithm.

The step 3 specifically includes:

Step 3.1: machine learning algorithms employ a Generation Antagonism Network (GAN), a variational self-encoder (VAE), etc. in deep learning.

Step 3.2: the training data of the generated algorithm of the invention comprises a design case besides the knowledge graph provided in the step 2. When another 3D auxiliary model, for example, a workpiece model (clamped object) is used as a precondition required for designing a clamp model (clamped object), the training data also includes processing characteristic data identified by the 3D auxiliary model, and the like.

Step 3.3: the 3D CAD model output by the generating algorithm not only has the appearance of the product, but also comprises the internal structure of the product. The difference between the two is that the 3D CAD model file containing the internal structure information of the product can be directly used for processing and manufacturing, but the 3D CAD model file containing the appearance information of the product cannot be directly used for processing and manufacturing, and further design work is needed to be manually completed by engineers.

The step 4 specifically includes:

step 4.1: the method has the advantages that the generated model is packed into the container, the container arrangement technology is used for deployment on the cloud server, and the method has highly portable and elastic computing resources and can reduce the deployment cost.

Step 4.2: the design requirement input by the user is converted into design parameters and an optimization target as input of the generative model after being analyzed by a natural language processing technology, and the 3D CAD model file is output of the generative model.

Step 4.3: the design intent entered by the user is primarily in text language, aided by Excel documents and other 3D aided models formulated in templates provided by the system.

The step 5 specifically includes:

Step 5.1: the system receives the design requirement input by a user, analyzes the design parameters and the optimization target through an NLP technology, inputs the design parameters and the optimization target into a generated model, and outputs the generated model to the system after reasoning. The system will show the generated 3D CAD model to the user and wait for user feedback. If the user is not satisfied with the generated 3D CAD model, the design requirement details can be further input, the system can correct the design parameters and the optimization targets, and the 3D CAD model is output again. This process may iterate repeatedly until a 3D CAD model is output that meets the user design requirements.

Step 5.2: in the process of continuous feedback of the user, the system stores memory of the input information of the previous user, and outputs new design parameters and optimization targets after comprehensive analysis. By the mode, the system can improve the user experience of man-machine interaction, and avoid repeated input of the user, so that more accurate and efficient design service is provided for the user.

Step 5.3: the generated model supports the output of the 3D CAD model under different materials and process requirements, and ensures that the output 3D CAD model is feasible in actual manufacturing.

The step 6 specifically includes:

Step 6.1: the 3D printing support module provided by the system can automatically convert the generated 3D CAD model file into a file format readable by a 3D printer, so that the operation and time cost of a user are reduced, and the manufacturing efficiency is improved.

Step 6.2: in the system, a user can directly use the 3D printer connected with the system to design and manufacture products and parts, so that the system is suitable for rapidly manufacturing small-batch or customized products, and the manufacturing efficiency of the products can be improved.

As shown in fig. 1, a technical roadmap of the design method of the invention is depicted. Step 1, collecting design-related data and constraint conditions, and step 2, automatically extracting and constructing a knowledge graph, wherein the connection relation between the data and the knowledge is a conversion relation of the data and the knowledge. In step 1, the collected design data and constraints provide the basis for knowledge-graph construction. These data and conditions include design parameters, optimization objectives, constraints, etc. of the product, reflecting the requirements and limitations of the design.

In step 2, a knowledge graph can be automatically extracted and constructed according to the data and conditions collected in step 1. Knowledge maps fuse the knowledge points of related materials, functions, structures, processes and the like, and provide useful information and data for training a generated model. Step 1 provides data and conditions for knowledge graph construction, and step 2 integrates the data and conditions into the knowledge graph to become knowledge, which provides a basis for subsequent training and reasoning.

The connection between the step 2 and the step 3 is that the knowledge automatically extracted and constructed in the step 2 is input into the step 3 as training data. In step 2, the knowledge graph includes knowledge points of materials, functions, structures, processes and the like related to design, and certain association relations exist between the knowledge points.

In step 3, knowledge points with association relations in the knowledge graph are obtained by using a machine learning algorithm to train, and a generated model comprehensively considering geometric features, process requirements and product design knowledge features is formed.

The connection relation between the step 3 and the step 4 is that the step 4 deploys the generated model in a cloud containerization mode, after the system acquires the design intent input by the user, the system analyzes the design parameters and the optimization targets by utilizing a natural language processing technology, the design parameters and the optimization targets are transmitted to the cloud generated model through a network, a 3D CAD model file is output after model calculation and is transmitted to the system through the network, and the system opens the file in 3D CAD software for the user to view.

And 5, 6, namely, on the basis of the output result of the step 4, the user can modify or supplement the detailed description of the design requirement again, the design parameters and the optimization targets are re-analyzed through the NLP technology, and the iterative process has no time limit requirement until the user is satisfied with the generated 3D CAD model or the user abandons to continue modification.

In step 7, the 3D CAD model may be inferred and output, and saved as a file for use in subsequent design, analysis, manufacturing, and other links. And also support direct transfer to a 3D printer for manufacturing.

As shown in fig. 2, a flow schematic and functional block diagram of the design system of the present invention is depicted. The process specifically comprises the following steps: the method comprises the steps of starting a software environment, inputting external limiting conditions, inputting design requirements of a user, generating model reasoning, feeding back and optimizing, outputting 3D model appearance, structure and 3D printing, and closing software, and belongs to a full life cycle operation flow of a system. The functional module specifically comprises: the system comprises an interactive design module, a data preprocessing module, a knowledge graph construction module, a machine learning training module, a CAD model generation module and a 3D printing support module.

S1: and starting the software environment, and enabling the user to start the software environment and enter the system interface.

S2: external constraints are entered, and the user enters external constraints, such as an auxiliary model file, excel written in system templates, and the like. Different design requirements and different templates are required.

S3: the user's design requirements, such as textual description of the appearance, structure, function, process, materials, etc., are entered in textual form. The S1 and S2 processes mainly comprise the steps that an interactive design module provides a user interface, and the system analyzes design parameters and optimization targets according to design requirements input by an S3 user and external limiting conditions input by an S2 user.

S4: the flow is mainly supported by a data preprocessing module, a knowledge graph construction module and a machine learning training module. The data preprocessing module converts text information input by a user into design parameters and optimization targets, the knowledge graph construction module is used for storing collected data and converting the collected data into knowledge, the machine learning training module allows configuration of training parameters, selection of algorithm networks and setting of evaluation indexes, and the training module outputs a generated model after iterative training. Knowledge graph construction and machine learning training are needed to be completed in advance. In the daily design process of the user, only the generated model reasoning is needed.

S5: and outputting the appearance and the structure of the 3D model, and generating a visualized 3D model in a standard format which is deduced by the model according to the acquired design and parameters and the optimization target, so that a user can intuitively know the appearance and the structure of the model. The process mainly comprises the steps of providing a user interface by a CAD model generation module, wherein the system supports the output 3D model standard format to comprise: obj, stl and fbx.

S6: and continuously updating the design requirement by the user according to the visual 3D CAD model fed back by the output, and continuously correcting the design parameters and the optimization target until the 3D CAD model meeting the design requirement of the user is output.

S7:3D prints, and the 3D prints support module outputs the optimized 3D CAD model as a file, provides support for being imported into a 3D printer, and is convenient for a user to print. At the same time, this module also provides some 3D printing related settings, such as printing resolution, support structure, etc.

S8: the software is turned off and the user can save the 3D CAD model or directly turn off the software.

In the whole design flow, the external limiting conditions and the user design requirements are information input by a user, and preprocessing of the input information, generating model reasoning and outputting of the appearance and the structure of the 3D model are completed by the system. 3D printing is an optional process, and if a 3D printer is not needed or connected, the 3D model file can be directly saved, and then the software is closed.

As shown in fig. 3, a flowchart of this embodiment is described, illustratively, with the automotive industry weld fixture as the generative design object. The welding fixture is used for guaranteeing the fixing performance and stability of the position of an automobile in the welding process, various workpieces (front longitudinal beams, front bumpers, fenders, front coamings and the like) of the automobile are changed, and the appearance and the structure of the fixture are different. There is a need for small-lot customization and fast design, requiring the following flow operations:

Step 1: the knowledge operators are responsible for collecting design related data and constraints, including but not limited to, fixture design expert knowledge, fixture design industry standards, fixture design past cases, and fixture design user behavior operations.

Step 11: the clamp design expert knowledge and industry standard are mainly derived from books, industry manuals, journal papers, experiment reports and other documents. The text data sources are extracted into structured data by means of Natural Language Processing (NLP) technology and data mining technology, and a manual extraction mode is adopted for formulas or complex chart data which exist in a picture form and are difficult to extract. Maintaining the automatic extraction to manual extraction ratio at about 9:1.

Step 12: the fixture design past cases are a large number of historical case libraries meeting design requirements and standards, which are completed and verified by manual design. Each set of cases includes: workpiece model, base plate model, fixture model, welding gun working track, RPS point, etc.

Step 13: the feature data extraction of the past case is divided into two cases. One is that the model has a standardized MBD structure, and mainly uses special CAD software formats such as Catia, UG, solidworks and the like, the model is often derived from enterprise internal data, and the data can be intelligently extracted based on standard specifications; the other model is not provided with a standardized MBD structure, and mainly takes common formats such as stl, obj, fbx and the like, the model is often derived from internet retrieval data, and can carry out finite element feature extraction on geometric topology information through grid division, and each feature has a special extraction algorithm, for example: a slot feature extraction algorithm, a hole feature extraction algorithm, an outline feature extraction algorithm, or a chamfer feature extraction algorithm.

Step 14: the workpiece is a clamped object of the designed clamp, and the appearance shape and the structural design of the clamping head of the clamp are determined. The Base plate is connected with the designed clamp Base, and the position of the Base plate determines the direction and the height of the clamp. The welding gun and the welding gun working track are a series of interference conditions which the clamp needs to avoid, and influence the appearance shape, structure and direction design of the clamp. The RPS input document should include at least the clamping point information of the workpiece, determining the relative position design of the fixture and the workpiece.

Step 15: the user behavior data is used for recording manual modification behavior data of a user in the 3D CAD design software, the data potentially comprises operation against the design intention of the user and modification points unsatisfactory to the generated design result of the user, and the generated model can be better optimized and updated through analysis of the user behavior data.

Step 2: and automatically extracting and constructing knowledge maps, including but not limited to constructing clamp past case knowledge maps, clamp design specification knowledge maps and clamp structure reasoning logic knowledge maps.

Step 21: the knowledge spectrum of the past case of the clamp surrounds two recommended dimensions of the whole clamp and a single clamp unit, and semantic design and automatic extraction of the spectrum are carried out. In this embodiment, only the construction of the unit map of the workpiece, the clamping point and the clamp is taken as an example, and the body design is as follows:

The workpiece is taken as an object of the body, and the attribute of the workpiece can be expressed as id, name, symmetry plane, shape, belonged vehicle model, belonged vehicle name and manufacturer.

The clamping points are the body objects and their properties can be expressed as id, weld point association, node type, position, flange depth, function, symmetry plane, workpiece edge, support or clamping direction requirements.

The clamp unit is used as a body object, and the attribute of the clamp unit can be expressed as id, name, node type, installation type, standard type, function, pressing block offset, clamping point height, cylinder pressing mode, cylinder swing arm position and rotation interference risk.

Step 22: the fixture design standardizes the knowledge graph, and semantic design and automatic extraction of the graph are carried out around four types of bodies of design activities, standards (versions), chapters and pages.

The relationship of the standard (version), chapter, page and design activity is: the former is the design basis of the latter, i.e. the former is used for the completion of the design activity. The attributes of the relationship are either strictly in accordance with the standard or with reference to the standard.

The relation between design activities is lower level, namely, the upper level design activity is completed, all the design activities of the lower level to which the design activities belong are required to be completed, and the relation also exists between the sub-level design activities, in the embodiment, the detection holes are required to be punched on the base, so that the detection hole design is also related to the base design.

The centers of the four types of ontologies are standard versions, chapters and pages, and the main relations are as follows: the version of the standard is the standard version and the relationship attribute is the effective time; chapters and pages are part of the standard version (partof), and there are attributes (start page and end page) of the chapter's relationship to the standard version; the modification relation exists among different versions of the same standard, and the attribute is modification content or degree; the versions of different standards have references and require similar relations; the storage address has a relationship with the standard version as an ontology.

The standards can be divided into design standards and standard component standards, and the attributes are names, standard numbers, categories (factories, national standards, etc.), organization units and priorities (i.e. different standards have different requirements on the same design activity, and then it is necessary to determine which standard to perform the design activity according to the priorities)

Step 23: the structure of the fixture infers logic knowledge-graph, in this embodiment, taking locating pin type inference as an example, the steps are as follows:

And reasoning the RPS characteristics of the positioning points to be designed, the RPS characteristics of the other positioning points and the geometrical characteristic attributes of the workpiece according to the bearing relation.

And (5) inferring the angle of the positioning hole, the variable part, the non-variable part and the part taking direction according to the upper-lower relation.

And reasoning the judgment of the telescopic pin/locating pin according to the cis-bearing relation and the causal relation.

Step 3: and training the generated model by using the knowledge stored in the knowledge graph as a training set and using a machine learning algorithm to form an intelligent reasoning engine.

The above is a design preparation phase of a solder clip design embodiment, followed by a system application phase of this embodiment, as follows:

Step 4: user input requirements, including but not limited to text, models, documents, and the like. Taking the design embodiment of the welding fixture as an example, the content of the input requirement is as follows:

step 41: and (3) inputting design requirements, wherein the system directly inputs text on an interactive information input interface. The system analyzes the text to design parameters and optimize targets through a natural language processing algorithm.

Step 42: and inputting the workpiece model, wherein the system provides an uploading interface of the workpiece model, and after the uploading is successful, the system automatically performs feature extraction.

Step 43: and the PRS document is input, the system provides an uploading interface of the RPS input document, and after the uploading is successful, the system automatically extracts the data of the RPS point.

Step 5: the generated design output model includes, but is not limited to, generated design, intelligent checksum interaction feedback. Taking the embodiment of welding fixture design as an example, the steps are as follows:

Step 51: the generated design is based on the generated model trained in the step 3, and a 3D CAD model conforming to design parameters and optimization targets is output.

Step 52: and after each design iteration is completed, the intelligent verification is automatically performed once. In the welding fixture design embodiment, it is mainly checked whether the generated fixture model envelope coincides with the welding gun model envelope under different trajectories, and if so, the design is re-performed.

Step 53: the interactive feedback is feedback and adjustment of the design result by a user in the design process, so as to achieve a better design effect. Such feedback may be problems and demands found by the user when using the design results, or may be suggestions and demands actively made by the user, such as adjustments to design details, improvements to functional performance, etc. After the user feeds back further, the system calculates the iteration again, and outputs the 3D CAD model which accords with the latest design parameters and the optimization targets.

Step 6:3D printing, in the implementation of the present invention, existing 3D printing techniques may be used, such as: FDM (fused deposition modeling), SLA (light curing modeling) or SLS (laser sintering modeling), and the optimized 3D CAD model is output as a 3D print file in a standard format such as STL, and manufactured using a 3D printer. Because the optimized 3D CAD model already meets the constraint conditions and achieves the optimal optimization target, the manufactured 3D printing product can meet the expected performance and quality requirements. Meanwhile, due to the flexibility and high automation of the 3D printing technology, complex shapes and structures can be manufactured rapidly, and the requirements of individuation and small-batch production are met.

Table 1 is an input Excel document template for one embodiment of a weld fixture design using the design method and system of the present invention. Illustratively, the RPS input document template of this embodiment is described with the automotive manufacturing industry welding fixture as the generative design object. Templates are another means by which users provide design requirements or constraints to supplement the design requirements that are not readily described in text.

Table 1RPS input document template

The RPS document of this embodiment includes the following fields: name, x, y, z coordinates, locating hole diameter, vamp, locating hole circumferential direction, work piece thickness, etc. The information in the RPS input document is extracted, so that the accurate clamping position information and the dimensional accuracy requirement of the part can be obtained, and the method is a key constraint target for realizing good operability and matching of the clamp.

When the clamped workpiece is subjected to process change, design parameters in an RPS input document are modified, so that the welding fixture can adapt to actual requirements, such as reinforcing rigidity, reducing weight, optimizing positioning hole layout and the like. And (5) re-entering the requirements, and finally generating and optimizing the 3D model.

It should be noted that the structure of the RPS document is not fixed, the requirements and the modification fields can be added by user definition according to the user requirements, the system automatically extracts the information in the cells through an Excel secondary development technology, and the information is input into the design engine for reasoning after being identified by using a natural language processing technology.

Example 1

Fig. 4 is a flowchart of a knowledge graph construction method based on manufacturing process data, specifically including the following steps:

S1: acquiring data information related to a product manufacturing process, establishing an original data database, realizing the function of primarily screening data irrelevant to the manufacturing process in a manufacturing enterprise, and simplifying the complexity of subsequent data extraction; the step of establishing the product manufacturing process data in the original data database comprises the following steps: enterprise manufacturing process database files, three-dimensional CAD models, drawing data, process files, processing equipment data and process equipment data; and storing the data which are preliminarily screened and related to the manufacturing process into a local hard disk or a network cloud disk as an original data database.

S2: based on the original data database, an ontology model is built by combining expert knowledge in the field of manufacturing process, and a knowledge graph model layer and a keyword dictionary of various entities are built. The flow of constructing the knowledge graph mode layer is shown in fig. 5, and the specific steps include:

s2.1: analyzing data; the manufacturing process data in the induction database is analyzed, the processing object of the manufacturing industry comprises components and parts, CAD models, drawings and process files of the components and the manufacturing process information contained in the enterprise database are analyzed, and the components and the manufacturing process information can be classified into six types of data: workpiece data, feature data, manufacturing equipment data, manufacturing method data, process equipment data, and semantic relationship data; the workpiece data describes information of names, numbers, versions, overall geometric features, materials and the like of the workpieces; the characteristic data refer to specific processing objects and requirements of each process in the manufacturing process of the workpiece and characteristics of clamping positions of each process, such as machined surfaces, holes, shafts, threads and the like, wherein the requirements comprise processing size, precision, heat treatment, surface treatment, welding quality and the like, and the characteristics of the clamping positions comprise positioning hole diameter, positioning hole shape, positioning surface shape and the like; the manufacturing equipment data mainly refer to equipment required by product manufacturing, such as lathes, milling machines, punching machines, welding robots and the like; manufacturing method data refers to a method and operation for manufacturing a workpiece, and steps for materials to pass from a blank to a product, particularly refer to working procedures and working steps; the process equipment data comprise a clamp, a die, a cutter, a gauge and other auxiliary tools in the manufacturing process; the semantic relation data refers to the relation among the previous data and represents the relation among the data of different categories;

s2.2: constructing a body model; combining expert knowledge and the entity type of the existing manufacturing process database of the enterprise to establish an ontology model;

Fig. 6 is a diagram illustrating the thinking of a knowledge graph entity model according to an embodiment of the present invention. In order to maintain the accuracy and completeness of the ontology model, the above six types of data are abstracted into six types of entities, wherein workpiece data, feature data, manufacturing equipment data, manufacturing method data, process equipment data, processing feature data and semantic relation data correspond to each other in sequence: workpiece class, feature class, equipment class, process class, tool class and relation class, and establishing a keyword dictionary of six types of entities, wherein the dictionary comprises standard keywords and synonyms of the six types of entities and attributes thereof;

Workpiece class includes two subclasses of components and parts; the feature class includes two subclasses of manufacturing features and clamping features; the process class comprises two subclasses of working procedures and working steps; the tool class comprises a clamp, a die, a checking fixture, an auxiliary tool, a cutter and the like; the relationship class includes consist _ of, order, assemble three subclasses;

The class attributes comprise class attributes and instance attributes, wherein the class attributes are various types and subclasses thereof, all instances share the corresponding class attributes, and the instance attributes are only all of the instances; the six types of data all have name type attributes, the workpiece type has drawing numbers, overall characteristics and management characteristic type attributes, parts in the workpiece type have material type attributes, the characteristic type has one or more of size, position, tolerance, surface roughness, geometric tolerance, technical requirements and other attributes, the equipment type has model type attributes, and other instance attributes are added according to different entities;

S2.3: establishing and displaying a knowledge graph mode layer; the knowledge graph mode layer is constructed in an ontology model mode, and the ontology model is described in an ontology language OWL; the ontology model is a model of a mode layer, and the ontology language is a language describing the mode layer; the mode layer is established according to the abstract entity class in the S2.2, the abstract entity class and the relationship are described through an expression, an ontology model is established, the abstract entity class and the relationship are established and displayed by utilizing prot software, the knowledge of the mode layer is represented by the ontology model, and the ontology model can be described by using owl language; the ontology model of the model layer of the knowledge graph is formally expressed as:

KGPattern＝{Entity∪Relation}

Wherein:

Entity＝{W∪F∪M∪O∪E}

Relation＝{R}

R＝{(W_i,consist_of,W_j)∪(W_i,consist_of,F_j)∪(W_i,consist_of,O_j)U(O_i,consist_of,O_j)∪(O_i,consist_of,F_j)∪(O_i,consist_of,M_j)∪(O_i,consist_of,E_j)∪(O_i,order,O_j)∪(W_i,assemble,O_j,assemble,W_j)∪}

In the above formula: KGPATTERN denotes a formalized expression model of a knowledge graph mode layer, entity denotes an Entity set of mode layer descriptions, and Relation denotes a relation set of mode layer descriptions; w represents workpiece class, F represents feature class, M represents equipment class, O represents process class, E represents tool class, and R represents relation class; the semantic relationship R comprises consist _ of, order, assemble, consist _of relationship is defined as the inclusion relationship between the workpiece class, the inclusion relationship between the workpiece class and the feature class, the inclusion relationship between the process class and the feature class, the inclusion relationship between the equipment class and the tool class, and the inclusion relationship between the process class, the relationship is expressed and described as: (W _i,consist_of,W_j) representing that workpiece W _i contains workpiece W _j,(W_i,consist_of,F_j) representing that workpiece W _i contains feature F _j,(W_i,consist_of,O_j) representing that workpiece W _i contains process O _j,(O_i,consist_of,O_j) representing that process O _i contains process O _j,(O_i,consist_of,F_j) representing that process O _i contains feature F _j,(O_i,consist_of,M_j) representing that process O _i contains equipment M _j,(O_i,consist_of,E_j) representing that process O _i contains tooling E _j; order relations are defined as sequential relations between processes or steps in a process class, and the relations are expressed and described as: (O _i,order,O_j) means that the process (step) O _j is performed after the process (step) O _i is completed; assemble relationships are defined as assembly relationships between components or parts in a workpiece class, and the relationships are expressed and described as: (W _i,assemble,O_j,assemble,W_j) represents the assembly of the component (part) W _i to the component (part) W _j by the process O _j; as shown in table 2;

Table 2 semantic relationship representation and description

At this time, the knowledge pattern layer is constructed, and the process goes to the next step S3.

S3: according to the knowledge graph mode layer, extracting relevant data in an original data database by using a data extraction model, and constructing a knowledge graph data layer; the data extraction model comprises a structured data processing module, an unstructured data processing module and a knowledge fusion module, as shown in fig. 7, and the specific operation flow is as follows:

S3.1: the structured data processing module converts the data in the relational database into RDF triples by adopting D2R according to a preset knowledge graph mode layer; the relation between the extracted data is defined according to the relation existing in the knowledge graph mode layer, and the extracted data and the relation between the extracted data are expressed in the form of RDF triples;

S3.2, the data extraction model comprises a three-dimensional CAD model, a drawing processing module and a natural language processing module, wherein the three-dimensional CAD model and the drawing processing module are used for extracting data contained in a workpiece model and a drawing of an original database, such as workpiece class data and feature class data, and generating RDF triples, the natural language processing module is used for extracting text data in the original database and generating RDF triples, and the knowledge fusion module is used for fusing entities extracted by the two processing modules;

S3.3: the three-dimensional CAD model and drawing processing module extracts attribute information, structure tree information and feature information in the model according to a knowledge graph mode layer by utilizing CAD software secondary development and multi-view feature extraction technology, the three-dimensional CAD model and drawing are converted into MBD models conforming to national standards, enterprise standards or other standards before extraction, and three-dimensional CAD software secondary development is carried out through python programming, so that three-dimensional CAD software secondary development is carried out, and three-dimensional group expression of workpiece types, feature type entities and attributes thereof and relations among workpieces, workpieces and features is established; the attribute information comprises workpiece names, numbers, materials, technical requirements and the like, the structural tree information refers to the containing relation between the component workpieces and the components forming the component, the characteristic information comprises workpiece overall characteristics and workpiece local characteristics, and the local characteristics comprise information such as names, sizes, positions, tolerances, surface roughness, geometric tolerances and the like;

Attribute information and structure tree information can be directly extracted from the standard three-dimensional MBD model; extracting overall feature information of the workpiece by adopting a multi-view feature extraction technology, obtaining feature vectors of the workpiece model after training by using a CNN (computer numerical network) network by utilizing multi-view pictures of the three-dimensional model, and storing the feature vectors as overall feature attributes of the workpiece entity; workpiece management features such as key components, important components, general components and the like can be obtained through field matching in the technical requirements of the MBD model;

The method comprises the steps that a CAD software secondary development technology is adopted for reading topological information of a model to indirectly extract local features of a workpiece, feature types are determined by a topological structure, for example, a simple through hole and a shaft are formed by two parallel planes and cylindrical surfaces perpendicular to the planes, if an intersection point exists between the axis of the cylindrical surface and the plane, the feature types are the shafts, otherwise, the feature types are holes, the sizes, the positions, the tolerances, the surface roughness and the geometric tolerance are determined by marked references and contents on the model, for example, the size of the simple through hole comprises the diameter and the depth, the mark of the cylindrical surfaces which are referred to form the holes is the diameter mark of the holes, and the mark of the two planes which are referred to form the holes is the length mark of the holes.

S3.4: the natural language processing module carries out entity marking on unstructured text by utilizing brat marking tools according to information in a keyword dictionary, converts marked text data into a data set with word segmentation labels BIO through a python program, and simultaneously divides the data set into a training set, a testing set and a verification set for training a Bi-LSTM-CRF entity recognition model; extracting workpiece class, feature class, process class, equipment class and tool class entities from unstructured data by adopting a trained Bi-LSTM-CRF entity recognition model, wherein the extracted workpiece class and feature class entities are not comprehensive and accurate enough and are only used for aligning the workpiece class and feature class entities extracted by the three-dimensional CAD model and the drawing processing module; and determining the relation among the specific entities according to the knowledge graph mode layer to obtain a triplet composed of the entities and the relation.

S3.5: the knowledge fusion module is used for fusing the knowledge which is extracted by the two processing modules and possibly has overlapping; carrying out semantic acquaintance calculation on the entity, the relation and the attribute extracted by the structured data processing module and the unstructured data processing module and standard keywords and synonyms thereof in a keyword dictionary, and replacing the extracted entity, relation and attribute words with the standard keywords corresponding to the synonyms according to a calculation result to realize cleaning alignment; traversing the three-dimensional CAD model and the workpiece class and feature class entities extracted by the drawing processing module, forming entity pairs with the workpiece class and feature class entities extracted by the natural language processing module, scoring the entity pairs according to a bilinear matching algorithm, arranging the scoring results in ascending order, enabling the entity pairs with lower score to represent higher alignment degree of the two entities, unifying the names of the two entities in the entity pair with the lowest score according to the scoring results, thereby completing the alignment of extracted data of the unstructured data processing module, and then completing the alignment of various entities extracted by the structured and unstructured data processing module according to the flow; and keeping the entity attributes of the workpiece class and the feature class after alignment consistent with the entity attributes of the workpiece class and the feature class extracted by the three-dimensional CAD model and the drawing processing module, and finishing entity fusion.

S4: combining the mode layer and the data layer to construct a knowledge graph, inputting the extracted RDF triples into a Neo4j graph database, storing the knowledge graph and visually displaying.

Example two

Fig. 8 is a schematic structural diagram of a manufacturing process knowledge graph construction system for multi-mode data according to an embodiment of the present invention, including:

The database module is used for acquiring data information related to the field of the product manufacturing process and establishing an original data base, wherein the data information of the product manufacturing process comprises the following components: structured enterprise manufacturing process database files, unstructured three-dimensional CAD models and drawing data, process files, processing equipment data, process equipment data, normative files, professional books and other data;

the model layer construction module is used for obtaining expert knowledge in the field of the manufacturing process to build an ontology model based on the original data database and constructing a knowledge graph model layer and a keyword dictionary of various entities;

The data layer construction module specifically comprises a structured data extraction module, an unstructured data extraction module and a knowledge fusion module; the structured data extraction module converts data in the relational database into RDF triples by adopting D2R according to a preset knowledge graph mode layer; the unstructured data extraction module comprises a three-dimensional CAD model and drawing processing module, a natural language processing module and a knowledge fusion module; the three-dimensional CAD model and drawing processing module is used for extracting data contained in the workpiece model and drawing of the original database and generating an RDF triplet, the three-dimensional CAD model and drawing are converted into MBD models conforming to national standards, enterprise standards or other standards before extraction, and the natural language processing module is used for extracting text data in the original database and generating the RDF triplet; the knowledge fusion module is used for fusing the knowledge which is extracted by the two processing modules and possibly has overlapping;

and the knowledge graph construction module is used for inputting the extracted RDF triples into a Neo4j graph database, storing the knowledge graph and visually displaying.

The knowledge graph construction system based on the manufacturing process data in this embodiment is basically identical to the working process of the knowledge graph construction method based on the manufacturing process data in the first embodiment, and is not described herein again.

The invention adopts CAD software secondary development technology and multi-view feature extraction technology to directly extract the information such as workpiece attribute information, overall features, workpiece local features and the like from the three-dimensional model and drawing of the workpiece, constructs the workpiece and the feature entities contained therein under the constraint of a mode layer, comprehensively and accurately describes the workpiece and the features thereof, expands the data types for constructing the map, and improves the accuracy and the availability of the constructed knowledge map. The knowledge graph for manufacturing process data can accurately and reasonably store a large number of existing products and process routes of the characteristics of the existing products and processing equipment and tools used by the existing products, and the recommendation of the new work process routes, public equipment and tools is realized by using a graph algorithm according to the matching degree of the new work and the characteristics of the new work and the historical work and the characteristics of the historical work stored in the graph, so that data support is provided for the process design of the products.

As shown in fig. 9, a CAD model matching method based on multi-view and machined feature recognition includes the steps of:

S1, acquiring a three-dimensional model of a part modeled by a B-REP method, wherein the model format comprises STEP, iges or other special data formats of CAD software; classifying according to the attributes of processing requirements, geometric shapes, purposes and the like, constructing a three-dimensional model database, and adding model processing and manufacturing related data;

S2, converting all models in a model library into STL format for deep learning training;

S3, matching processing features in each model by utilizing semantic information and topological structures in the models;

s3.1, traversing all geometric surfaces in the model and the geometric surfaces adjacent to each geometric surface by using a B-REP type three-dimensional model to construct an adjacent graph of the model part;

s3.2, constructing a feature adjacency graph of the target feature aiming at the adjacency relation of the surface in the target to be searched;

S3.3, searching subgraphs in the adjacent graphs of the parts by taking the geometric surface with the greatest adjacent or the first geometric surface in the adjacent surface traversing sequence as a starting point, and inquiring the minimum condition subgraph matched with the characteristic adjacent graph to acquire processing characteristics;

s3.4, extracting processing characteristics such as threaded holes, finish machining surfaces and the like by utilizing semantic information marked by colors and texts in the B-REP type CAD model;

S4, training a ResNet-based multi-view convolutional neural network (MVCNN) to enable the multi-view convolutional neural network to have the functions of model classification and feature vector calculation;

s4.1, constructing a ResNet-based MVCNN deep learning model, wherein a model classification label in the S1 step is a MVCNN classification label;

S4.2, obtaining multiple views of the model according to the STL format model; using multiple views and labels for MVCNN training;

S5, calculating feature vectors of each model according to the trained model; completing construction of a database and training of MVCNN;

s6, converting the B-REP type three-dimensional model to be matched into an STL format;

S7, classifying the three-dimensional model to be matched and calculating the feature vector by adopting MVCNN;

S8, matching a plurality of models with minimum distances between the feature vectors and the model feature vectors to be matched in a database, so as to finish rough matching of the three-dimensional model;

s9, extracting processing characteristics of the B-REP type three-dimensional model to be matched according to the adjacent topological relation of the surfaces in the model and semantic information marked by colors and texts, wherein the method is the same as that in the step S3;

And S10, further matching in the similar model obtained by rough matching in S8 according to the processing features extracted in S9, obtaining models with the types, the numbers and the sizes of the processing features close to the models to be matched, completing matching, and obtaining corresponding processing data from a model library for subsequent operation.

As shown in fig. 10, a CAD model matching system based on multi-view and processing feature recognition, based on a CAD model matching method based on multi-view and processing feature recognition, inputs a CAD model to be retrieved, outputs a CAD model having a shape and structure similar to those of the input and its auxiliary processing information, the CAD model matching system should include:

the file management module realizes the functions of file reading, writing, storage, format conversion and the like;

The 3D model database comprises a large amount of 3D models and corresponding information such as feature vectors, processing features and the like, and can realize simple information matching such as processing features, material properties and the like;

The deep learning reasoning module adopts MVCNN of training to convert the classification and the feature vector of the 3D model to be matched in the calculation of the STL format;

The processing feature extraction module is used for extracting processing features of the 3D model to be matched of the input BREP type;

And the vector retrieval module is used for matching the classification and the feature vector obtained by calculation in the deep learning reasoning module in the feature vector data of the 3D model database.

Examples are as follows:

S1, in CATIA software, a plurality of (more than 4000) three-dimensional models of machined parts in CATPart format are obtained, classified according to purposes, shapes, machining characteristics and the like, a 3D model database is constructed, and model machining and manufacturing related data are added;

s2, converting CATPart format software into STL format by using CATIA software through a file management module and VBA script;

S3, extracting the processing characteristics in the three-dimensional model through a processing characteristic extraction module, wherein the processing characteristics are specifically as follows: s3.1, searching for the characteristics of a finish machining hole, a threaded hole, a finish machining surface and the like in CATIA software by adopting a search command according to the properties of color, name and the like;

S3.2, extracting a topological structure of the model based on CATIA CAA secondary development, and generating a adjacency graph of the surface; for a CAD model shown on the left side of FIG. 11a, generating an adjacency graph is shown in FIG. 11b, where gray nodes represent planes, black nodes represent cylindrical surfaces, and inter-node links represent two surface adjacencies; in fig. 11a and 11b, for an L-shaped plate with a countersunk through hole, fig. 11a shows the serial numbers of the faces, fig. 11b shows the adjacent relationship between the faces by straight line connection with the faces as nodes, wherein gray nodes represent flat surfaces and black nodes represent cylindrical surfaces. Cylindrical surface 9-planar surface 11-cylindrical surface 10 together form a counterbore machining feature, as shown within the dashed enclosure.

S3.3, generating a feature adjacency graph aiming at the topological structure of the middle surface of the processing feature (such as a stepped hole);

S3.4, in the adjacent graph of the model, using the geometric surface with the largest adjacent or the first geometric surface in the traversing sequence as a starting point, and matching the minimum condition subgraph matched with the feature adjacent graph in the adjacent graph of the model to acquire the processing feature; in the models shown in fig. 11a and 11b, the cylindrical surface 9-plane 11-cylindrical surface 10 is retrieved to form a counter-sunk hole topological relation, such as the inner part of a broken line in fig. 11 b;

s3.5, the processing features extracted in the S3.3 and the S3.4 are stored in a database, so that subsequent matching is facilitated;

S4, training a deep learning reasoning module, namely MVCNN based on ResNet, so that the deep learning reasoning module has the functions of model classification and feature vector calculation, and specifically comprises the following steps: s4.1, constructing a MVCNN based on ResNet18 based on Pytorch, namely, disposing a model on a display card server according to the classification label of the MVCNN, namely, the classification label in the step S1; selecting a cross entropy function as a loss function, and adopting a Dropout method to alleviate overfitting in the training process; the pre-training data of ResNet are recommended to be loaded, so that the iteration times of training are reduced; the MVCNN frame is shown in fig. 12, and the principle is as follows: for a three-dimensional model, obtaining views under multiple views, and respectively inputting each view into ResNet deep learning networks to respectively obtain a plurality of tensors; inputting the tensors into a view pool layer to obtain feature vectors of the three-dimensional model, further inputting the feature vectors into a full-connection layer, and calculating classification results of the model;

s4.2, acquiring multi-view images (12 views) of the STL format model in the step S2 by adopting a secondary development function of Blender software; cutting redundant parts of each image, wherein the size of the redundant parts is 224 to 224, and the number of channels is 3; converting the plurality of adjusted images into a tensor structure under Pytorch frames; inputting the data into MVCNN models in batches, and completing training after multiple iterations;

S5, calculating the feature vector of each model in S1 according to MVCNN for training in S4.3, and storing the feature vector in a database; the construction of a 3D model database and the training of a deep learning reasoning module are completed;

s6, converting the three-dimensional model to be matched in STEP or CATPart format into STL format by using CATIA program through a file management module;

S7, obtaining multiple views from the three-dimensional model to be matched in the STL format by utilizing secondary development of the Blender program in the step S4.2, and cutting to finally generate corresponding tensors; carrying out reasoning calculation by using the deep learning reasoning module trained in the step S4.3 to obtain classification and feature vectors of the deep learning reasoning module;

S8, matching the first 10 models with smaller feature vector distances between the models to be matched in a database by adopting a vector retrieval module, so as to finish rough matching of the three-dimensional models;

S9, adopting a processing feature extraction module, extracting processing features of a STEP-format model to be matched according to adjacent topological relations of surfaces in the model and semantic information marked by colors and texts, wherein the method is the same as that in the STEP S3;

and (3) further matching in a similar model obtained by rough matching in S8 according to the processing features extracted in S9 by utilizing the matching capability of the 3D model database, obtaining models with the types, the numbers and the sizes of the processing features close to those of the models to be matched, completing matching, and obtaining corresponding design processing data from a model library as reference to assist subsequent operation.

The three-dimensional model matching method based on geometric multi-view and processing feature recognition can effectively solve the problem of three-dimensional model matching in the manufacturing industry, improves the computer aided design efficiency, reduces the labor burden of staff, and has wide application prospect. The advantages of the two ideas of the geometric shape and the topological relation are combined, the calculated amount is low in the matching process, the calculating speed is high, and the matching time can be greatly shortened while the accuracy is improved. Is particularly suitable for three-dimensional model matching with typical processing characteristics, such as: high-precision parts with chamfer, groove and hole machining characteristics commonly used in the fields of automobiles, semiconductors, military industry and the like. The invention can better describe and capture the characteristics of the 3D model, thereby improving the reusability of the 3D model. For example, when one 3D model is used, other 3D models having similar features may be found by the matching system and used as references, thereby improving reusability of the 3D model.

A method and system for reasoning 3D model structure using logic atlases, comprising:

Rule extraction processing and structuring: based on expert rules of a mechanical design manual or a specification, the expert rules are converted into triple structured data by using a natural language processing technology to automatically extract and combine with manual fine tuning correction.

The structured data is converted into logic patterns: the structured data is stored in logic atlas and visually presented to the machine design rules and their flow, and the expert further fine-tunes the correction rules.

Expert system construction and reasoning based on logic atlas: based on the established logic map, the system converts the map into an expert system, and the system supports the reasoning of the expert system and the logic map.

Based on the built logic map, the system directly uses the rational map to make reasoning.

Inference results of the system: and inputting the requirements and conditions of the mechanical structure to be designed in an input interface provided by the system, visually presenting the design result by the system, and outputting the 3D model structure and parameters.

The rule extraction processing and structuring specifically comprises the following steps:

And (3) data acquisition: relevant data in the machine design manual is collected, including design parameters, material properties, process requirements, and the like.

And (3) data processing: preprocessing the acquired data, including data cleaning, data analysis, feature extraction and the like.

Rule extraction: based on expert knowledge and machine learning algorithm, extracting expert rules in the mechanical design manual, including design parameter selection, material selection, model selection and the like.

Rule verification: and verifying the extracted rule by using an actual mechanical design case, and correcting the rule.

the data acquisition is performed by manual input, scanning, OCR recognition and other methods.

The data acquisition module comprises a sensor, a scanner, image acquisition equipment, computer aided design software and the like.

The data processing comprises preprocessing operations such as data cleaning, denoising, feature extraction, dimension reduction and the like.

Rule extraction is based on machine learning algorithms, including decision trees, convolutional neural networks, antagonistic neural networks, and the like. And (3) automatically extracting and mining domain knowledge by adopting machine learning and natural language processing technologies.

Rule validation is performed by actual mechanical design cases, including rationality testing and usability testing.

The structural data is converted into logic patterns specifically as follows:

the structured data is stored in logic maps using a graph database technique.

An initial model of the machine design rules and their flows is defined, abstracting the rules and flows into a data model that can be inferred computationally.

The meta-model of the mechanical design rules and their flow is mapped to the data structure stored in logic's map.

And visually presenting the mechanical design rule and the flow thereof by using a visual technology.

The logic atlas visually presented is endowed with an editable function, and an automatically generated mechanical design logic atlas interface is adjusted or upgraded for iteration, and meanwhile manual construction is supported.

The structural data is converted into logic patterns specifically as follows:

Graph database techniques may include Neo4j, janus graph, hua cloud, pandaDB, protege, and the like.

Visualization techniques may include various forms of graphical user interfaces, visual editors, and the like.

Map database reading can realize rapid storage and query of logic maps through encapsulation of the map database.

Visual presentations may include a variety of functions such as graphical presentation of mechanical design rules and their flows, interactive queries, statistical analysis of data, etc.

Establishing a map in the 3D model structure field logic, which comprises data reading, ontology construction and map storage. Wherein the body is constructed by: the processed triplet structured data is generalized into several classes, as shown in fig. 4: basic entity, judgment/formula, judgment calculation, reasoning result, event, output and semantic relation. The data of the above classes are abstracted into logic classes of atlases, wherein the basic entity, the judgment/formula, the judgment calculation, the reasoning result, the event and the semantic relation respectively correspond to the basic entity class, the judgment/formula class, the judgment calculation class, the reasoning result class, the event class and the semantic relation class. Based on logic characteristics of map reasoning, an output class is independently added.

Logic the map consists of node classes and relationship classes.

The basic entity class is a node class, and the content is a basic noun comprising a part, a component name and a part attribute name.

The judgment/formula class is a node class, and the content is the name of a group of judgment selection or formula calculation targets in the design flow and is used for connecting the judgment selection or formula calculation targets.

The judgment calculation class is a node class, and the content is the actual content of judgment selection and formula calculation in the design flow.

The inference result class is a node class, the content of which includes the feature, size, model and structural composition of the part.

The event class is a node class, the content including the name of the design part, design structure, or artifact.

The output class is a node class, and the content has no practical meaning as a marked node, and marks the end of the partial reasoning.

The semantic relation class is a relation class and is used for connecting nodes and expressing the relation among the nodes. The relationship class comprises a causal relationship class, an upper relationship class, a lower relationship class, a cis-bearing relationship class and a conditional relationship class. The causal class relationship is a strong logical relationship, and the next node is inferred because of the content of one node; the context class is an inclusive relationship, with the next node being part of the last node, or the last node being part of the next node. The cis-bearing relationship class is a precedence relationship, and the last node finishes the next node; the conditional relationship class is a strong logical relationship, with the last node being a requirement for the next node.

The manual atlas editing realizes functions of remote collaboration, knowledge sharing and the like based on cloud computing and other technologies.

The expert system based on logic atlas is constructed specifically as follows:

and constructing knowledge representation and reasoning models of the expert system according to the concept system and rules of the mechanical design map. The reasoning and query functions of the expert system are realized, and the reasoning and query functions comprise a reasoning algorithm based on rules, models and statistical methods.

The knowledge representation and reasoning model for constructing the expert system is specifically as follows:

The expert system construction system based on the map comprises an automatic conversion module, an reasoning module, a query module and a display module.

The three-dimensional model reasoning process based on the atlas and the three-dimensional group data conversion method based on the natural language generation technology NLG comprise the following steps: analyzing the input triplet data through a natural language processing technology; generating a natural language sentence by utilizing a predefined grammar and rules according to the analysis result; outputting the converted natural language sentences. Wherein the predefined grammars and rules include: a grammar structure of subjects, predicates, objects; determining parts of speech and quantity of subjects and objects according to the triplet relation; corresponding verbs, adjectives and adverbs are selected according to a predefined dictionary.

Based on the established logic map, the system directly uses the rational map to infer specifically:

and by utilizing the semantic representation capability of logic atlases, high-efficiency reasoning and analysis are carried out on large-scale complex data, so that high-efficiency 3D model structure reasoning analysis is realized.

Design of the inference engine: the node type and the attribute of the map of the query logic are traversed by using Python and Cypher languages, the queried nodes are respectively processed in different modes according to basic entity, judgment/formula, judgment calculation, reasoning result, event and semantic relation, and then are combined with the data input by the module 5 to be processed in a comparison, calculation, judgment and the like, so that the 3D model structure is finally deduced.

The automatic design 3D model structure based on the expert system and the knowledge graph is specifically as follows:

and automatically deducing the structure and the characteristics of the 3D model meeting the requirements according to the requirements of user input in a question-answer input mode.

Wherein the display module may include editing software, rendering software, 3D printing software, etc.

And outputting the generated 3D model structure and characteristics in the form of files or APIs, so as to realize automatic design.

The invention is implemented with reference to fig. 13. S1, a module 1, based on expert rules of a mechanical design manual or a specification, automatically extracting and combining with manual fine adjustment correction by using a natural language processing technology to convert the expert rules into triple structured data; s2, a module 2 stores the structured data into logic maps and visually displays mechanical design rules and flow thereof, and an expert further fine-tunes the correction rules; s3, a module 3, based on the established logic atlas, converting the atlas into an expert system by the background, and realizing an reasoning process; s4, logic a map 3D model reasoning module, wherein the part uses automatic reasoning of a 3D model of a rational map row; s5, inputting the requirements and conditions of the mechanical structure to be designed in an input interface provided by input, and visually presenting the design result by the system.

The steps are described in detail with reference to fig. 14:

S1 module 1, the system processes and constructs the rules from the given machine design manual, machine design instruction manual and other data extraction

S1.1, collecting relevant text data in a mechanical design manual, including design parameters, material characteristics, process requirements and the like, and storing the relevant text data in a readable electronic document format, such as PDF (portable document format), word documents and the like.

S1.2, data processing: and cleaning and formatting the data, cleaning the data, analyzing the data and extracting the characteristics. The method comprises the steps of removing stop words, normalizing texts, removing special characters, segmenting words, removing HTML labels and other formatting marks, removing punctuation marks and the like.

S1.3, rule extraction: the text is analyzed using NLP techniques to identify keywords, phrases, and grammatical structures therein, such as sentences, paragraphs, and the like. Data collection using existing NLP libraries and tools, such as NLTK and SpaCy.

S1.4, generating triple structured data, and converting the extracted rules in the text into the triple structured data. As shown in fig. 15.

S1.5, storing and managing the triple data, and storing the triple structured data in an SQL database.

S1.6 rule verification, namely taking the stored data in the SQL database as a rule, wherein the step function is to realize visualization and editing iteration, the part uses a written customized program, and the main function is to improve the accuracy and avoid semantic drift generated during NLP processing as much as possible.

And S2, the module 2 stores the structured data into logic maps and visually presents the mechanical design rules and the flow thereof, and supports the expert to further fine-tune the correction rules.

S2.1 defines the data model in S1.5, including node and relationship types, and node attribute types, using Cypher and Python statements.

S2.2, analyzing the structured data, and analyzing the structured data into the form of nodes and relations by using Python.

S2.3, establishing nodes and relations, connecting the Neo4j database by the system, and then using a Python driver to establish the nodes and the relations. The node type, attribute, label are specified, the relationship type, attribute, label are specified, and the relationship between the start node and the end node are specified as shown in fig. 16.

S2.4, using Python to store the nodes and relations in the Neo4j database, as in the example shown in FIG. 17, the data structure of the triplet model of the mechanical design rules and their flow in logic maps.

The S2.5 system provides a visual configuration interface and is in bidirectional connection with the Neo4j database. The system uses a visualization technology to visually present logic maps representing the mechanical design rules and the flow thereof, and simultaneously provides various functions such as interactive inquiry, data statistics analysis and the like. The visual interface supports the functions of adding, modifying and deleting, so as to be used for manual correction or knowledge iterative upgrade. While performing the operation, the Neo4J logic map at the rear end also performs the corresponding operation and stores.

S2.6 map database Neo4J reading can realize rapid storage and query of logic maps on the Neo4J level through encapsulation of the map database.

And S3, based on the established logic atlas, the system converts the atlas into an expert system and performs reasoning. And constructing knowledge representation and reasoning models of the expert system according to the concept system and rules of the logic atlas. The reasoning and query functions of the expert system are realized, and the reasoning and query functions comprise a reasoning algorithm based on rules, models and statistical methods. The system comprises an automatic conversion module, an reasoning module, a query module and a display module.

S3.1, an automatic conversion module: the knowledge in the map section read logic by the system, which provides an automated transformation module and interface to transform the stored logic map into a system including, but not limited to, clips, jboss Rules, drools expert systems, is expressed in logic or rules. Querying a certain part or structure in the domain and category of 3D model reasoning in logic patterns selected in FIG. 16, and querying the related subgraphs of the node. Then, all nodes with the attribute of 'reasoning result' in the subgraph are queried, all starting points are queried as event names, end point end nodes are path paths of each 'reasoning result', all 'formula/judgment' nodes in each path are respectively read and recorded, the 'formula/judgment' nodes are used as conditions of each part type, and all the conditions of the reasoning result are stored. The process is repeated for all the nodes of the reasoning result until all the paths of the nodes of each reasoning result are queried. The "formula/judgment" node in each path of each inference result is used as the rule of the inference result, and based on the rule, the use is converted into an expert system.

S3.2, an inference module: the inference engine in the expert system performs reasoning and interpretation on the necessary conditions of each reasoning result in S3.1.

S3.3, a query module and a display module: the system provides an expert system query module and a display module based on a local or WEB side, and supports the query source code and structure of the expert system in S3.1. The expert system user selects the model to be inferred, then inputs necessary parameters and characteristics, and the expert system automatically outputs the inference result, wherein the inference result is the 3D model structure and the design information parameters.

S4, 3D model reasoning of logic atlas: the system can use two modes of a rational map and an expert system to realize the reasoning process, wherein the logic map method has the advantages of rapidness and faster realization of simpler reasoning; the advantage of expert system reasoning is that the reasoning process can be implemented faster for a large number of complex reasoning than if the rational atlas is used directly. Logic map realizes the reasoning process:

S4.1, triggering an reasoning process: as shown in FIG. 16, when a user selects a domain and category of a certain 3D model inference, and selects a specific inference of a certain part or structure, an "event" class node corresponding to logic atlas, and a logic atlas subgraph of the event node is selected. And searching all 'basic entity' nodes of the sub-graph and reading the content, and obtaining corresponding data of all 'basic entities' through a method of user input or automatic acquisition.

S4.2 reasoning process: the system uses Python to search and select the 'judgment calculation' node connected with the event, and inquires the 'basic entity' node connected with the node again respectively, matches the corresponding data acquired in S4.1 and records.

S4.3, searching a judgment/formula node connected with judgment calculation in S4.2 by using Python, reading the content of a single judgment/formula node according to a random sequence, and bringing corresponding data in S4.2 into the content of the judgment/formula node.

S4.4 is in accordance with the content record of the node of the judgment/formula in S4.3, and the reading and inquiring process of S4.3 is terminated.

And S4.5, searching and recording the downstream node of the S4.4, and continuing to search the downstream node.

S4.6 looks up the downstream "judgment calculation" node of S4.5. The process of S4.2-S4.5 is repeated, and when the system queries the "output" node in fig. 4, reasoning is terminated and the names of all recorded nodes are stored.

S4.7, searching path, wherein the path is characterized in that: the starting point is the event class node selected by the user, the path contains all the recorded nodes in S4.6, and the end point is the 'output' node. And queries the "inference results" class nodes of the path. Outputting the content of the node of the 'reasoning result' class, namely the reasoning result

S4.8, generating natural language by using NLG by the nodes stored in S4.6, and outputting the reasoning process and logic of the natural language as the reason and rule support of the reasoning result in 4.7.

S5, a module 5, wherein the reasoning result of the system uses an interface.

S5.1 system provides a system reasoning result use interface, and the user selects the field and category needing mechanical reasoning at the interface.

S5.2, the system prompts the name of the condition to be input, and the user inputs related indexes such as design requirements, parameters, using conditions and the like to be filled in.

S5.3, automatically reasoning based on the expert system and logic map module at the back end, and outputting the characteristics and attributes of the reasoning result.

The S5.4 display interface comprises editing software and rendering software.

S5.5, outputting the generated 3D model structure and the characteristic attribute in a file or API form, and realizing automatic design.

The flow of this embodiment is described for locating pin sizing, illustratively with the automotive industry weld fixture as the generative design object. The automobile welding fixture is a fixture used in a welding process on an automobile production line and used for clamping an automobile body or an automobile body part so as to facilitate welding and assembly. The design of automotive welding fixtures is difficult because they need to take into account many factors and must meet stringent requirements. The welding fixture must be perfectly matched to the geometry and dimensions of the vehicle model and body components. The welding fixture must have sufficient rigidity and stability to ensure that no deviation or deformation occurs in the welding process, so that the design difficulty of the automobile welding fixture is very high, a designer needs to have deep expertise and experience, and meanwhile, the following flow operations are performed with small-batch customization and rapid design requirements:

step 1: the developer inputs necessary design data such as fixture design specifications, fixture design manuals, fixture design industry standards and the like of the automobile welding fixture.

Step 11: text data is extracted, including design parameters, material characteristics, process requirements, etc., and stored as word.

Step 12: and (3) data processing, namely cleaning and formatting the data, removing stop words, normalizing texts, removing special characters, word segmentation, punctuation marks and the like.

Step 13: the text is analyzed using NLP to identify keywords, phrases, and grammatical structures therein, such as sentences, paragraphs, and the like. Library data collection was used NLTK.

Step 14: the triplet structured data is generated and the rules extracted from the text as shown in fig. 15 are converted into triplet structured data.

Step 15: the triplet structured data is stored in an SQL database.

Step 16: and (3) rule verification, namely performing preliminary verification on the processed structured data, and improving accuracy.

Step 2: the structured data is stored in logic map and the rule of the automobile welding fixture and the flow thereof are visually presented.

Step 21: the data model is defined using the Cypher and Python statements, including node and relationship types, and node attribute types.

And step 22, analyzing the structured data, and analyzing the structured data into the form of nodes and relations by using Python. The nodes and relationships are established, the system connects to the Neo4j database, and then the Python driver is used to create the nodes and relationships. Specifying a node type, attribute, label, specifying a relationship between a start node and an end node. The basic entity node can be expressed as: { < id >:366< name >: flexible element < attribute of attribute >: workpiece geometry attribute }, the formula judgment entity node may be expressed as { < id >:367< name >: workpiece feature = flexible element < attribute >: telescopic pin/fixed pin judgment }

Step 23: the nodes and relationships are stored in Neo4j database using Python, and the data structure of the triplet model of the mechanical design rule and its flow in logic map is mapped, as shown in fig. 17.

Step 24: the logic map of the automobile welding fixture is operated on a visual interface provided by the system and used for checking and confirming the correctness of the map, and the adjustment rule can be modified on the interface according to special requirements or updated rules.

Step 3: the function of reasoning the 3D model is realized, the system converts the map into an expert system based on the built logic map, and both the logic map and the expert system support the function of independently completing reasoning.

Step 31: according to the map of the automobile welding fixture logic, the automatic conversion module is as follows: and constructing a knowledge representation and an inference model of the expert system. The stored map of the automobile welding fixture logic is converted to include, but is not limited to, a Clips expert system.

Step 32: inference, the inference engine in the Clips expert system infers and interprets knowledge.

Step 33: and providing an expert system query module and a display module based on a local or WEB side in the system, and directly querying the source code and structure of the expert system.

And 4, inputting conditions and outputting an reasoning result by the system.

Step 41: and (3) selecting the 'automobile welding fixture-locating pin selection type' reasoning in the step 1 in a reasoning result use interface provided by the system.

Step 42: the system prompts the name of the condition to be input, and the user inputs according to the prompt:

locating hole angle a= (0, 1)

Locating hole angle b= (0, 1)

Non-deformable part

Pick-up direction= (0, 1)

Setpoint hole diameter = 25mm

Pin working length = 50mm

Step 43: the system automatically infers based on a back-end expert system or logic map module and displays the result of the inference result:

{ the following 2 parts all meet the design requirements

(1) "Dowel type 39d 20614, d=25 mm, l=60 mm";

(2) "dowel type 39d 20613, d=25 mm, l=60 mm". }

Step 44: the system will output:

Because the locating hole angle is not equal to get a direction, so select the telescopic pin, the car welding jig locating pin model of reasoning is: 39d 20614, d=25 mm, l=60 mm or 39d 20613, d=25 mm, l=60 mm.

The intelligent design of the system improves the mechanical design efficiency: the traditional mechanical design 3D model structure needs a plurality of related field experts to conduct manual calculation and reasoning design, and the 3D model structure can be automatically deduced and generated based on logic atlas reasoning 3D model structure system, so that the design efficiency is greatly improved, the design time and cost are saved, the design production period of a terminal product is directly reduced, the workload of engineers is greatly reduced, and the efficiency is increased while the manpower is reduced.

The design accuracy is improved: based on logic atlas reasoning 3D model structure system, can convert the rules and knowledge in the mechanical design manual into a computable form, and store and manage in atlas form. Through deep mining and automatic reasoning of mechanical design knowledge, the design thought and flow are more accurate and clear, the design precision can be improved, the possibility of design errors is reduced, the repeated design risk is reduced, and the design production period of the terminal product is further reduced.

Support multi-domain design: the logic atlas reasoning 3D model structure system can integrate knowledge in multiple fields, including knowledge in the aspects of mechanical design specification, material mechanics, fluid mechanics, theoretical mechanics, mechanical manufacturing and the like, can effectively integrate knowledge in various fields, and can support complex 3D model structure design in multiple fields.

The design flexibility is improved: the invention can automatically deduce the 3D model structure meeting the requirements according to the input requirements of the user. The user can realize flexible design and optimization by adjusting the input parameters and conditions.

The product quality and the competitiveness are improved: the intelligent mechanical design can improve the design efficiency, greatly reduce the design period, thereby improving the product quality and the line-down speed, and further improving the production efficiency and the competitiveness of enterprises.

While the foregoing enumerated embodiments describe the design methods, technical routes, system flows, and system modules of the present invention, it should be appreciated that many variations and modifications may be made without departing from the scope of the present invention. It is therefore intended that the foregoing detailed description be regarded as illustrative rather than limiting, and that it be understood that it is the following claims, including all equivalents, that are intended to define the methods and systems of this invention.

Claims

1. The three-dimensional CAD generation type design method based on the knowledge graph and the machine learning is characterized by comprising the following steps:

step 4: according to the processing information of the part model to be matched, the 3D model structure and parameters, adopting cloud containerization deployment to generate a model, and outputting a 3D CAD model;

Said step 1) comprises the steps of:

1.4 According to the trained neural network model, calculating the feature vector of each part three-dimensional model, and completing the construction of a database and the training of the multi-view convolutional neural network;

Said step 2) comprises the steps of:

2.5 According to the extracted processing characteristics, further matching in the similar model obtained by rough matching to obtain a model with the type, the number and the size of the processing characteristics closest to the model to be matched, completing matching, and obtaining corresponding processing data from a model library;

(1) Acquiring expert rules, and converting the expert rules into triple structured data by using a mode of combining natural language processing and automatic extraction with manual correction;

(2) Storing the triad structured data into logic atlas and visualizing it;

(3) Reasoning is carried out based on the established logic map;

(4) Inputting the requirements and conditions of the mechanical structure to be designed, and obtaining the 3D model structure and parameters by reasoning and visualizing the results.

2. The three-dimensional CAD-generated design method based on knowledge-graph and machine learning as claimed in claim 1, wherein the CAD model matching method based on multi-view and process feature recognition comprises the steps of:

3. The three-dimensional CAD-generated design method based on knowledge graph and machine learning as claimed in claim 1, wherein said step (1) comprises the steps of:

(1.1) obtaining text data in a mechanical manual, and storing the text data in a readable electronic document format, the text data comprising: design parameters, material characteristics, and process requirements;

(1.2) sequentially cleaning the text data;

(1.3) analyzing the cleaned text data by a natural language processing method, and identifying and extracting keywords, phrases and grammar structures in the text data as rules;

(1.4) converting the extracted rule into triple structured data and storing the triple structured data into an SQL database;

(1.5) validating and revising rules in the SQL database using actual mechanical design cases.

4. The three-dimensional CAD-generated design method based on knowledge graph and machine learning according to claim 1, wherein said step 4 comprises the steps of:

① Packaging the generated model into a container, and deploying on a cloud server;

② The design requirement of the user is converted into design parameters, an optimization target and processing information of the part model to be matched after natural language processing, and the processing information is used as input of a generated model to generate and output a 3D CAD model; the design requirements of the user include text language, excel documents and 3D auxiliary models.

5. A three-dimensional CAD-generated design system based on knowledge graph and machine learning for implementing the three-dimensional CAD-generated design method based on knowledge graph and machine learning as set forth in claim 1, comprising:

6. The three-dimensional CAD generation type design device based on the knowledge graph and the machine learning is characterized by comprising a memory and a processor; the memory is used for storing a computer program; the processor, when executing the computer program, is configured to implement the three-dimensional CAD-generated design method based on knowledge graph and machine learning as claimed in any one of claims 1-4.

7. A computer-readable storage medium, wherein a computer program is stored on the storage medium, which when executed by a processor, implements the three-dimensional CAD-generated design method based on knowledge-graph and machine learning as claimed in any one of claims 1 to 4.