Natural language representation method and system for three-dimensional model of mechanical part

Abstract

The invention discloses a natural language representation method and system for a three-dimensional model of a mechanical part, and belongs to the technical field of intelligent industrial design. The method comprises the following steps: acquiring geometric features of a historical mechanical part three-dimensional model and multi-dimensional data of engineering attributes; generating structural features of the mechanical part through the constructed segmented intelligent generation model, and performing vectorization processing to generate a semantic expression result of the mechanical part; performing natural language description on the structured features by utilizing a constructed mechanical field corpus; establishing a mapping relation database between the structured features and the natural language description, and performing fine tuning on the pre-trained segmented intelligent generation model; importing multi-dimensional data of the new mechanical part three-dimensional model into the fine-tuned segmented intelligent generation model, and outputting a semantic expression result of the new mechanical part; the problems that an understanding gap and an application barrier exist in a mechanical three-dimensional model, and manual description is low in efficiency and prone to errors are solved.

Description

Technical Field

[0001] This application belongs to the field of intelligent industrial design technology, and in particular relates to a natural language representation method and system for three-dimensional models of mechanical parts. Background Technology

[0002] With the rapid development of artificial intelligence technologies such as Large Language Modeling (LLM), the integration of natural language processing and 3D geometric understanding (multimodal learning) is becoming increasingly mature, making it possible to achieve this representation efficiently, accurately, and automatically. 3D models of mechanical equipment are often stored in specific formats, such as CAD, SLDPRT, and STEP, which are themselves highly accurate carriers of geometric and topological information. However, key engineering knowledge such as implied functional semantics, manufacturing requirements, material properties, and performance constraints is usually implicit in the model structure, annotations, metadata, or the designer's mind. Due to the inherent data characteristics of 3D models of mechanical parts, the following problems often arise: 1. The "understanding gap" in 3D models of mechanical parts: There are huge differences in the understanding and use of 3D models among different fields (design, process, manufacturing, quality inspection, operation and maintenance, procurement, sales), different roles (engineers, technicians, managers, customers), and different software systems (CAD, CAE, CAM, PLM, MES, ERP). Often, some professional terms, engineering drawings, and specific file formats constitute communication barriers.

[0003] 2. "Application Barriers" of 3D Model Data for Mechanical Parts: 3D models of mechanical equipment (such as CAD files, CAE files, and STEP files) are highly accurate carriers of geometric and topological information. However, the key engineering knowledge they contain, such as functional semantics, manufacturing requirements, material properties, and performance constraints, is usually hidden in the model structure, annotations, metadata, or the designer's mind. This makes it difficult for machines to directly understand and process them automatically, resulting in the inability to apply them directly.

[0004] 3. Manually writing natural language descriptions is inefficient and prone to omissions or inaccuracies due to differences in the professional level of personnel. This is especially true for complex parts, where a complete description requires a significant amount of time.

[0005] Therefore, how to transform the three-dimensional model of mechanical parts into a structured natural language description that is both readable and intuitive for human engineers, and provides a semantic basis for computer processing (such as retrieval, reasoning, generation, and verification) has become an urgent problem to be solved in the field of mechanical equipment. Summary of the Invention

[0006] To address the shortcomings of the existing technology, this application provides a natural language representation method and system for three-dimensional models of mechanical parts.

[0007] The technical solution is as follows: On the one hand, a natural language representation method for three-dimensional models of mechanical parts is provided, including the following steps: Acquire multi-dimensional data of historical mechanical parts 3D models, including the geometric features and engineering properties of the mechanical parts 3D models; Based on the multi-dimensional data of the mechanical parts, the segmented intelligent generation model is a three-segment serial model, which is used to generate the structured features of the mechanical parts, vectorize the structured features and generate the semantic expression results of the mechanical parts. The structured features are graph structures composed of nodes and edges. Nodes represent the mechanical parts entities and their constituent features, and edges represent the topological and engineering relationships between geometric features. Construct a corpus of mechanical engineering, and use the mechanical engineering corpus to perform natural language description of the structured features; A mapping relationship database between the structured features and the natural language description is established, and the segmented intelligent generation model is fine-tuned based on the mechanical domain corpus and the mapping relationship database; The multi-dimensional data of the new mechanical part's 3D model is imported into the fine-tuned segmented intelligent generation model, and the semantic representation of the new mechanical part is output.

[0008] In one embodiment, the acquisition of multi-dimensional source data of historical mechanical parts three-dimensional models is achieved by extracting the multi-dimensional source data from CAD software using a data extraction program. The geometric features in the multi-dimensional source data include shape, size, tolerance, and positional relationships between geometric features; The engineering attributes in the multi-dimensional source data include material, weight, and strength.

[0009] In one embodiment, the structured features adopt a hierarchical mesh architecture, distributing the mechanical part entities and the nodes constituting the mechanical part feature information nodes at different levels, with nodes at the same level connected in a mesh.

[0010] In one embodiment, the mechanical part feature information nodes are distributed at different levels by layering the nodes from top to bottom as part nodes, geometric feature combination nodes, and geometric feature nodes, with two or three layers; the mechanical part feature information nodes are distributed at different levels and / or connected by edges between the same level, using a combination of directed and undirected edges, including geometric topological relationships and engineering relationships. The engineering relationships are represented using undirected edges; The geometric topological relationships and process relationships are represented using directed edges; The geometric topological relationships are used to characterize the positional relationships between geometric features and the subordinate relationships between geometric features and engineering features; The engineering relationship characterizes the positional relationship between the geometric features of the combination.

[0011] In one embodiment, the entity object represented by the mechanical part feature information node includes: part nodes, geometric feature nodes, and engineering feature nodes. The component node is the root node of the entire component; The geometric feature nodes are the basic geometric units that make up the parts, including volumes, surfaces, edges, and vertices; The engineering feature nodes are combinations of geometric features, including holes, slots, bosses, chamfers, and fillets.

[0012] In one embodiment, the segmented intelligent generation model is a three-segment serial model used to generate the structural features of the part, vectorize the structural features, and generate the semantic representation of the mechanical part, wherein: The vectorized processing of structured features involves adding a directional encoder and a similarity calculation module to the original image attention network to enhance the model's semantic understanding of nodes and edges. The directional encoder embeds a directional vector into the edge vector of the structured feature and concatenates the node vector with the embedded directional vector. The similarity calculation module compares the similarity values ​​between general word vectors and domain-specific word vectors, and replaces the general word vectors with domain-specific word vectors. The semantic representation of the generated parts is used to guide the segmented intelligent generation model, after fine-tuning, to generate natural language descriptions of the mechanical parts.

[0013] In one embodiment, establishing the mapping relationship database between the structured features and the natural language description is a mapping relationship between the feature information of nodes and edges and the structured natural language, formed by expressing the feature information of nodes and edges in structured natural language.

[0014] In one embodiment, importing the multi-dimensional data of the new mechanical part's three-dimensional model into the fine-tuned segmented intelligent generation model includes: First, incremental pre-training of the large language model is performed using a corpus in the mechanical field. Then, the model, which has been incrementally pre-trained, is fully trained using a mapping relation database to make the model deeply adaptable to the natural language description in the mechanical field. Then, by using preset prompt word templates, a segmented model for intelligently generating natural language descriptions of mechanical parts is obtained after training.

[0015] On the other hand, a natural language representation system for a three-dimensional model of a mechanical part is provided, for implementing the natural language representation method for a three-dimensional model of a mechanical part as described above, including: The feature extraction module is used to extract the geometric features, engineering attributes, and process information of the 3D model of mechanical parts; The segmented intelligent generation module is used to extract multi-dimensional source data of mechanical parts, generate structured features of mechanical parts, and perform vectorization processing to intelligently generate semantic representation results of mechanical parts. The storage module is used to store corpora and mapping relationship databases in the field of mechanics; The output module is used to output and display the semantic representation of mechanical parts.

[0016] In a preferred embodiment, the segmented intelligent generation module includes: The structured feature generation submodule is used to generate structured features for mechanical parts; A structured feature vectorization processing submodule is used to perform vectorization processing on the structured features; The semantic representation result generation submodule is used to generate semantic representation results for mechanical parts.

[0017] The technical solution includes at least the following technical effects: This invention constructs a segmented intelligent generation model, which transforms the geometric, engineering, and technological multi-dimensional data of a 3D model of a mechanical part into a hierarchical, network-like structured feature, and establishes a mapping relationship between this structured feature and natural language. Ultimately, it drives a large language model (i.e., the segmented intelligent generation model) trained with domain knowledge to generate accurate and structured semantic expressions, thus solving the problems of "understanding gap" and "application barrier" in mechanical 3D models, as well as the low efficiency and error-prone nature of manual description.

[0018] This invention parses the 3D model of a mechanical part into structured features rich in domain semantics, and then performs professional fine-tuning on a pre-trained large language model based on these features. This enables end-to-end automated generation from model to text, ultimately achieving a comprehensive effect of high accuracy, strong professionalism, and significantly improved efficiency in the generation and description technology. It effectively solves the core pain points of traditional methods, such as low efficiency and unprofessional and unreliable general model representations.

[0019] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit this application. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] Figure 1 A schematic diagram of the structure of a natural language representation system for a three-dimensional model of a mechanical part, provided as a preferred embodiment of this application; Figure 2 A flowchart of a natural language representation method for a three-dimensional model of a mechanical part, provided as a preferred embodiment of this application; Figure 3 A layered network diagram of "part-geometric feature combination-geometric feature" for a stepped shaft provided in a preferred embodiment of this application; Figure 4 An overall architecture diagram of the improved BERT-Encoder and Transformer-Decoder provided in a preferred embodiment of this application; Figure 5 This is a network structure diagram of an improved BERT-Base provided in a preferred embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] This embodiment uses a typical mechanical part—a "stepped shaft"—as an example to fully illustrate the entire process from data extraction to natural language description generation. The system architecture diagram for this case is as follows: Figure 1 As shown: The natural language representation system for the 3D model of the mechanical part includes: The feature extraction module is used to extract the geometric features, engineering attributes, and process information of the 3D model of mechanical parts.

[0025] The segmented intelligent generation module is used to extract multi-dimensional source data of mechanical parts, generate structured features of mechanical parts, and perform vectorization processing to intelligently generate semantic representation results of mechanical parts.

[0026] The storage module is used to store corpora and mapping relationship databases in the field of mechanics.

[0027] The output module is used to output and display the semantic representation of mechanical parts.

[0028] In a preferred embodiment, the segmented intelligent generation module includes: The structured feature generation submodule is used to generate structured features for mechanical parts; A structured feature vectorization processing submodule is used to perform vectorization processing on the structured features; The semantic representation result generation submodule is used to generate semantic representation results for mechanical parts.

[0029] This stepped shaft part contains various geometric features, engineering properties, and process requirements, which fully demonstrate the effectiveness and advancement of this method. The specific implementation steps are as follows: Figure 2 As shown: Step S1: Obtain multi-dimensional data of the three-dimensional model of the mechanical part, including the geometric features and engineering attributes of the three-dimensional model of the mechanical part; In a preferred embodiment, a data extraction program developed based on Python and CAD secondary development interfaces (such as SolidWorks, API, AutoCAD, .NET API) is used to extract multi-dimensional data from SolidWorks software, which is divided into two main categories: geometric features and engineering attributes.

[0030] It should be noted that the data extraction program, developed using Python and CAD secondary development interfaces (such as SolidWorks, API, AutoCAD, and .NET API), provides separate secondary development interfaces for each mainstream CAD software, normalizing the data according to the format shown in Table 1. Although different CAD software (UG / NX, CATIA, Pro / E) have different internal data structures and APIs, they all provide interfaces for accessing geometric data, attributes, and annotations. This embodiment primarily uses SolidWorks.

[0031] The key to this method lies in the subsequent construction and semantic understanding of "structured features". Regardless of the CAD software used to extract the information such as "cylinder diameter 50mm" and "heat treatment tempering", once it enters the processing flow of this application, it will be uniformly represented as nodes (geometric feature nodes, process information nodes) and edges (process relationships) in a graph structure. Data normalization is performed at the semantic level through the structured feature graph model defined in this application, rather than forcibly unifying the original data format. Therefore, this method framework has the potential and adaptability for cross-platform applications.

[0032] Regarding the representation capability of complex geometric features: This application defines "geometric feature node" as "the basic geometric unit constituting a part, including volume, surface, edge, and vertex". A complex freeform surface is composed of a series of "surfaces" within a CAD system; an irregular cavity is also a "volume" surrounded by multiple "surfaces".

[0033] This application addresses complexity through a hierarchical mesh structure. In the "geometric feature layer," complex surfaces can be represented as one or more "face" nodes. Simultaneously, in the higher-level "geometric feature combination layer," a node such as a "combination of irregular cavities" can be defined, connected to multiple lower-level "face" nodes via "directed edges." This structure is fully capable of representing complex features and describing the connections and orientations between faces through the "geometric topological relationships" of the edges. This application provides a flexible representation framework, rather than being limited to handling simple, regular features.

[0034] The specific data is presented in tabular form as follows: Table 1: Multi-dimensional Data (Parameter) Extraction Table Step S2: Based on the multi-dimensional data of the mechanical parts, construct a segmented intelligent generation model. The segmented intelligent generation model is a three-segment serial model, which is used to generate the structured features of the mechanical parts, vectorize the structured features, and generate the semantic expression results of the mechanical parts.

[0035] The structured features are graph structures composed of nodes and edges, where nodes represent part entities and their constituent features, and edges represent the topological and engineering relationships between geometric features.

[0036] The segmented intelligent generative model constructed in this step includes three core sub-modules: a structured feature generation sub-module, a structured feature vectorization processing sub-module, and a semantic expression result generation sub-module.

[0037] The overall architecture adopts a segmented design of "feature modeling → vector encoding → semantic generation" to ensure the independence and synergy of each link.

[0038] Step S20: Generate structured features; The structured features of the stepped axis adopt a three-layer hierarchical network architecture of "part-geometric feature combination-geometric feature". Nodes are distributed hierarchically, and nodes within the same level are connected in a network. Edges represent different types of dependencies.

[0039] Step S201: Node definition and hierarchy division; Based on the multi-dimensional data in Table 1, four types of nodes are defined (part nodes, geometric feature combination nodes, geometric feature nodes, and engineering feature nodes), and the hierarchy is divided according to a three-layer architecture, as follows: First layer (root node layer): component nodes; Node Name: Stepped Shaft_φ50-φ30-φ20; Node Attributes: Material = 45 steel, Weight = 3.25kg, Tensile Strength ≥ 600MPa; Node Type: Root Node (Unique).

[0040] Second layer (geometric feature combination layer): Engineering feature nodes (geometric feature combination nodes); It contains 3 nodes, namely: Shaft assembly node: Its attributes are composed of shaft segment 1, shaft segment 2, and shaft segment 3, and its function is to realize transmission and support; Key connection combination node: The attribute is that it is composed of flat keyway 1, and the function is to realize the circumferential fixation of the shaft and the hub; Transitional combination node: Its attributes consist of a relief groove and 4 chamfers, and its function is to facilitate assembly and reduce stress concentration.

[0041] Third layer (geometric feature layer): Geometric feature nodes; It contains 9 nodes, namely: Shaft segment 1 node: Attribute = Cylindrical, φ50h7×120mm; Shaft segment 2 node: Attribute = Cylindrical, φ30k6×80mm; Shaft segment 3 nodes: Attribute = Cylindrical, φ20×50mm; Flat keyway 1 node: Attribute = rectangular slot, 10×5×30mm; Unsink groove node: Attribute = Annular groove, 5×2mm; Chamfer 1 node: property = C2 × 45°; Chamfer 2 node: property = C2 × 45°; Chamfer 3-node: property = C2 × 45°; Chamfer 4-node: property = C2×45°.

[0042] Supplementary nodes (process / engineering attribute nodes): Heat treatment details: Properties = tempering, hardness 220-250 HBW; Surface treatment node: Attribute = chrome plating for shaft segment 1, sandblasting for the rest; Step S202: Define edges and connection methods; A combination of directed and undirected edges is used to connect nodes at different and the same level. The types and meanings of the edges are as follows: Cross-level edges: Level 1 - Level 2, Level 2 - Level 3: Part node → Shaft assembly node: Directed edge, representing geometric topological dependency relationship: Shaft assembly is the core component of part). Part node → Key connection composite node: Directed edge, geometric topological dependency relationship; Component node → Transitional composite node: Directed edge, geometric and topological dependency relationship; Part node ↔ Process combination node: Undirected edge, engineering relationship: Process combination serves to improve the overall performance of the part; Axis body composite node → Axis segment 1 node: Directed edge, geometric and topological dependency relationship; Axis body combined node → Axis segment 2 node: Directed edge, geometric and topological dependency relationship; Shaft body combined node → Shaft segment 3 node: Directed edge, geometric and topological dependency relationship; Key connection combination node → Flat keyway 1 node: Directed edge, geometric topological dependency relationship; Transitional composite node → tool relief groove node: directed edge, geometric and topological dependency relationship; Transitional combination node → Beveled nodes 1-4: Directed edge, geometric topological dependency relationship.

[0043] Edges at the same level (within the third level, within supplementary nodes): Node 1 of axis segment → Node 2 of axis segment: Directed edge, geometric and topological relationship: Axis segment 1 and axis segment 2 are coaxial and their end faces fit together; Node 2 of axis segment → Node 3 of axis segment: Directed edge, geometric and topological relationship: coaxial and end faces fit together; Shaft segment 1 node ↔ Keyway 1 node: Undirected edge, engineering relationship: The keyway is located in the middle of shaft segment 1, and the two are related in position; Shaft segment 1 node → tool relief groove node: directed edge, geometric topological relationship: the tool relief groove is located at the transition between shaft segment 1 and shaft segment 2; Surface treatment node → mating node: undirected edge, engineering relationship: surface treatment quality affects mating accuracy.

[0044] Step S203: Draw a layered network diagram; Based on the nodes and edges defined above, a layered mesh diagram of the stepped axis "part-geometric feature combination-geometric feature" is drawn as shown in Figure 3. Note: Figure 3 Only the properties of node 1 of axis segment are displayed.

[0045] Step S21: Structured feature vectorization processing.

[0046] This embodiment adds a directional encoder and a similarity calculation module to the original Graph Attention Network (GAT) to vectorize the constructed graph structure (structured features), enhancing the model's semantic understanding of nodes and edges. The specific process is as follows: Step S211: Node vectorization processing; The attribute descriptions of each node are pre-encoded using the Word2Vec model to generate initial node vectors. For example, the attribute description of node segment 1 is "cylindrical, φ50h7×120mm", which, after Word2Vec encoding, yields an initial vector V1 with a dimension of 256; similarly, initial vectors (all with a dimension of 256) are generated for all nodes.

[0047] Step S212: The orientation encoder embeds the orientation vector; The direction encoder embeds different direction vectors for directed and undirected edges in the graph structure: for directed edges (e.g., A→B), direction vector D1 (dimension 128) is embedded, representing the subordinate / order relationship "from A to B"; for undirected edges (e.g., A↔B), direction vector D2 (dimension 128) is embedded, representing the "bidirectional association" relationship. The initial node vector is concatenated with the corresponding direction vector to obtain the node vector with fused direction information (dimension 256+128=384). For example, the fused vector for node 1 corresponding to the directed edge from node 1 to node 2 is V1+D1, and the fused vector for node 1 corresponding to the undirected edge from node 1 to node 1 is V1+D2. Because of the embedded direction vectors, the original attention mechanism formula needs to be modified, as follows: ; in, The weight matrix to be learned. and Let and represent the original feature vectors of node and node , respectively. Encode the embedding vector for the direction of the edge. This represents the vector concatenation operation.

[0048] Step S213: The similarity calculation module optimizes word vectors; A general-purpose word vector library and a mechanical-domain-specific word vector library are constructed. The similarity calculation module calculates the similarity value of corresponding words in the two word vector libraries using the cosine similarity formula. Preferably, the general-purpose word vector library uses GloVe pre-trained word vectors, while the mechanical-domain-specific word vector library is built based on mechanical design manuals and process specification documents.

[0049] If the similarity value is lower than a preset threshold (set to 0.6 in this embodiment), the general word vector is replaced with a domain-specific word vector to ensure the domain adaptability of the word vectors. For example, "quality adjustment" has no corresponding vector in the general word vector library, so it is directly replaced with a domain-specific word vector. For example, the similarity value between the vector of "cylindrical" in the general word vector library and the domain-specific word vector is 0.58 (lower than 0.6), so it is replaced and optimized. Finally, the optimized structured feature vector matrix is ​​obtained (dimension is the number of nodes × 384).

[0050] Step S23: Construction of the semantic representation result generation submodule; A large language model (based on an improved BERT-Base model) is used as the core for semantic generation, guiding the pre-trained model to fine-tune and generate natural language descriptions of the stepped axis. The input of the submodule is the optimized structured feature vector matrix, and the output is natural language text conforming to mechanical field standards. The output format is preset to a structured text framework of "overall part description → core feature description → engineering attribute description → process information description".

[0051] The specific improvements to the BERT-Base model are as follows: Construct a hybrid architecture of BERT-Encoder + Transformer-Decoder.

[0052] The original BERT is a bidirectional autoencoder model, adept at semantic understanding but lacking native text generation capabilities. This application extends its Encoder-Decoder architecture, retaining BERT's semantic understanding advantages and introducing a Transformer-Decoder to achieve native text generation capabilities. The modified model architecture is "BERT-Encoder + Sequence Generation Transformer-Decoder," with the specific design as follows: 1. Improvements to BERT-Encoder: Based on the original BERT-Base, a coding layer specific to the mechanical field is incorporated, including: 1) Node feature encoding layer: embeds the node vectors of structured features (including geometric, engineering, and process attributes) into the word vector layer of BERT; 2) Edge relation encoding layer, which integrates the direction vectors and relation type vectors of directed / undirected edges into the BERT position encoding layer; 3) Domain semantic encoding layer: Replace general word vectors with domain-specific word vectors to enhance the model's ability to understand the semantics of domain-specific terms.

[0053] The core function of this part is to perform deep semantic encoding on the structured feature vector matrix and output a semantically encoded vector containing the feature relationships in the mechanical field, which provides a foundation for subsequent text generation.

[0054] 2. Transformer-Decoder Design: The classic Transformer-Decoder structure is introduced (using a 6-layer decoder and a multi-head attention mechanism of 12 heads). An autoregressive generation method is adopted, taking the semantic encoding vector output by BERT-Encoder as input and combining it with a mechanical domain sentence template library to realize the generation of natural language text sequences from semantic encoding.

[0055] The "BERT-Encoder+Transformer-Decoder" hybrid architecture described in this application is not a simple concatenation, but rather an end-to-end sequence generation task achieved by using the encoder output as the decoder input. The integrated architecture diagram is shown below. Figure 4 As shown, the specific integration method is as follows: The BERT-Encoder part uses a pre-trained BERT-Base model (12-layer Transformer Encoder) as the encoder. The input is a structured feature vector matrix (dimension: nodes x 384), which is mapped to a sequence of word vectors through the embedding layer. The BERT-Encoder output is a context-aware sequence of hidden states with dimensions [sequence length, 768].

[0056] Bridge Layer: A fully connected projection layer is introduced between the BERT Encoder and Transformer Decoder, mapping the 768-dimensional vector output by the Encoder to a 512-dimensional vector input by the Decoder, thus adapting to the input dimension of the Decoder. This layer simultaneously achieves dimension alignment and feature fusion, ensuring efficient information transfer from the Encoder to the Decoder.

[0057] The Transformer-Decoder section employs a standard Transformer-Decoder architecture, comprising six Decoder Layers, each with a 12-head multi-head attention mechanism. The Decoder's input consists of: the encoder output (projected through the Bridge Layer) serving as the Key and Value for the "Encoder-Decoder Attention"; and the word vector sequence of the target text (during training) or the previous word generated by autoregression (during inference) serving as the Query. The Decoder output is a probability distribution over the vocabulary, which, after passing through a Softmax layer, generates a natural language description text.

[0058] Pre-training task adjustment: To adapt to text generation tasks, this architecture adopts a sequence-to-sequence (Seq2Seq) training objective in the pre-training stage, specifically combining masked language modeling (MLM) and causal language modeling (CLM): the Encoder part is still trained using the MLM task; the Decoder part is trained using the CLM task (autoregressive generation); During the end-to-end fine-tuning stage, an "encoder-decoder joint training" strategy is adopted to optimize the overall generation quality.

[0059] The improved BERT-Base is a bidirectional pre-trained language model based on the Transformer Encoder. The detailed network structure diagram of the improved model is shown below. Figure 5 As shown, its network structure is as follows: Embedding layer: Converts the input words into 768-dimensional vectors, including Token Embedding, Segment Embedding and Position Embedding. Token Embedding maps each node vector (384-dimensional) to 768-dimensional vectors through a linear layer. Segment Embedding is set to 0 here. Position Embedding: Uses learnable positional encoding, supports a maximum of 512 nodes, and outputs: [sequence length, 768].

[0060] BERT-Encoder layer: 12 Transformer Encoder layers, each layer specifically includes: Multi-Head Self-Attention (12 heads), each head has a dimension of 64, Feed-Forward Network (the first layer has a linear transformation of 768→3072, the activation function is GELU, the intermediate layer has a dimension of 3072, the second layer has a linear transformation of 3072→768), Layer Normalization, and Residual Connection.

[0061] Output layer: The output of the last Encoder layer serves as the semantic representation of the entire sentence.

[0062] This application retains the entire Encoder structure of BERT-Base, and only connects a newly designed Transformer-Decoder after it, thereby realizing a complete process from semantic understanding to text generation.

[0063] Step S3: Construct a corpus in the mechanical field, and use the corpus in the mechanical field to perform natural language description of the structured features of the mechanical parts.

[0064] Step S31: Construct a corpus for the mechanical field.

[0065] The corpus was constructed using a combination of publicly available data collection and domain-specific data annotation, and comprises three sub-corpora: (1) Basic language material sub - library: Collect public materials such as mechanical design manuals (such as "Mechanical Design Manual (Sixth Edition)"), GB / T standard documents (such as GB / T 1800.2 - 2009, GB / T 307.1 - 2015), and mechanical part atlases, and extract texts such as part descriptions, feature terms, and process expressions.

[0066] (2) Special language material sub - library for shaft parts: For shaft parts (such as stepped shafts, smooth shafts, spline shafts, etc.), collect texts such as shaft part design specifications, process cards, and technical manuals, extract relevant descriptions such as shaft segments, keyways, chamfers, heat treatment, etc., and perform manual annotation (annotation dimensions: feature type, parameter accuracy, relationship expression).

[0067] (3) Sentence pattern template sub - library: Summarize common sentence patterns for part descriptions in the mechanical field, and construct sentence patterns, such as "[Part name] consists of [Feature combination 1], [Feature combination 2], and the main structure is [Core feature]", "[Feature A] is located at the [Position relationship] of [Feature B], and the dimensional parameter is [Specific parameter]", "[Process name] process parameters are [Process parameters], and the purpose is [Process purpose]", etc. There are a total of 100 sentence pattern templates.

[0068] Pre - process the collected language materials: Remove duplicate texts, correct typos, perform word segmentation (using the jieba word - segmentation tool and adding a mechanical - field exclusive dictionary, such as "quenching and tempering treatment", "clearance fit", etc.), and remove stop words (such as "of", "already", "and", etc.), and finally form a standardized mechanical - field language corpus.

[0069] It should be noted that: Taking the stepped shaft as a single example to disclose the training process, it is not a training sample and only contains this part. In fact, the training language corpus of this application covers multiple types of mechanical parts and a large amount of industrial actual data, and a data enhancement mechanism + regularization strategy is designed to effectively suppress over - fitting, and can handle data changes and different types of mechanical parts. The specific description is as follows: (1) The actual scale and composition of the incremental pre - training language corpus.

[0070] 1. Basic language material sub - library: In addition to the "Mechanical Design Manual" and GB / T standard documents, it also integrates part technical manuals and design specifications of mainstream domestic mechanical manufacturing enterprises, and the amount of language materials reaches more than 50,000 pieces; 2. Special language material sub - libraries for multiple types of parts: It does not only contain shaft parts, but also covers more than 10 types of typical mechanical parts such as disk parts, bracket parts, box parts, gear parts, and connecting parts. For each type of part, industrial actual design and process description language materials are collected.

[0071] 3. Sentence Template Sub-library: In addition to the original 100 basic sentence templates, 500 exclusive sentence templates have been added based on different part types, covering the description logic of various parts.

[0072] During incremental pre-training, the corpus of 100,000+ words is used as training data to enable the model to fully learn the terminology, sentence structure and expression habits of the mechanical field, rather than just learning knowledge related to shaft parts.

[0073] (ii) The actual sample range of full training and prompt word fine-tuning.

[0074] Taking a single step axis as an example, the process of constructing mapping relationships and generating natural language text is disclosed. It is not that the full training and fine-tuning of prompt words use only this sample; the actual training samples are diverse and comprehensive. 1. Fully trained mapping database: It does not only contain node / edge-text fragment mappings of a single step shaft, but covers multiple types of parts such as shafts, discs, and flanges, and has constructed tens of thousands of node / edge-text fragment mapping pairs. Each type of part contains mapping relationships with different structures, parameters, and processes, enabling the model to learn the structured features of different parts and the mapping rules of natural language. 2. Cue Word Fine-tuning (20 Text Templates): These 20 text templates are descriptions of the same part with varying levels of detail and emphasis. The purpose is not to train the model to memorize the description of a specific part, but rather to guide the model to learn how to organize the text description structure as "overall part description → core feature description → …", and how to adjust the output under different levels of detail requirements.

[0075] Step 32: Structured feature natural language description based on corpus.

[0076] By combining the corpus and sentence templates from the mechanical engineering corpus, the hierarchical axis structured features (graph structure) constructed in step S2 are described using natural language to form an initial natural language text set. For example: (1) Overall description of the part: "This part is a stepped shaft, model number stepped shaft_φ50-φ30-φ20, made of 45 steel, weighing 3.25kg, with a tensile strength of not less than 600MPa and a yield strength of not less than 355MPa. (2) Description of core features: "The stepped shaft body is composed of three coaxial shaft segments, namely shaft segment 1, shaft segment 2 and shaft segment 3; shaft segment 1 is cylindrical with a diameter of φ50h7mm and a length of 120mm; shaft segment 2 is cylindrical with a diameter of φ30k6mm and a length of 80mm; shaft segment 3 is cylindrical with a diameter of φ20mm and a length of 50mm; a rectangular flat keyway is provided in the middle of shaft segment 1 with a width of 10mm, a depth of 5mm and a length of 30mm; an annular relief groove is provided at the transition between shaft segment 1 and shaft segment 2 with a width of 5mm and a depth of 2mm; a total of 45° chamfers are provided at both ends of the part and at the transition between shaft segments, with a chamfer size of C2mm." (3) Engineering attribute description: "Shaft section 1 and the inner hole of the hub adopt H7 / h7 clearance fit, and shaft section 2 and the inner ring of the bearing adopt K6 / h5 transition fit." The above natural language descriptions strictly correspond to the relationships between node attributes and edges in the structured features, ensuring the accuracy and completeness of the descriptions. A total of 20 initial natural language texts were generated (covering descriptions of varying levels of detail).

[0077] Step S4: Establish a mapping relationship database between the structured features of mechanical parts and natural language descriptions, and use the mechanical domain corpus and mapping relationship database to train and fine-tune the segmented intelligent generation model in stages.

[0078] Step S41: Based on the structured features (nodes, edges) of step S2 and the natural language description of step S3, establish a mapping relationship database between the two.

[0079] The mapping relationships are constructed in two categories: "node-text fragment" and "edge-text fragment," and stored in key-value pair format, as follows: Node-text fragment mapping: Key is the unique identifier of the node (e.g., “part node_stepped shaft”, “shaft segment 1 node”), and Value is the natural language description fragment of the corresponding node (e.g., “This part is a stepped shaft, model is stepped shaft_φ50-φ30-φ20”, “shaft segment 1 is cylindrical, diameter φ50h7mm, length 120mm”).

[0080] Edge-text fragment mapping: Key is the unique identifier of the edge (e.g., “part node → shaft body combination node”, “shaft segment 1 node → shaft segment 2 node”), and Value is the natural language description fragment of the relationship represented by the corresponding edge (e.g., “the stepped shaft body is composed of three coaxial shaft segments”, “shaft segment 1 and shaft segment 2 are coaxial and their end faces fit together”).

[0081] Step S42: Fine-tune the pre-trained segmented intelligent generative model in stages; This case study employs a three-stage training strategy: incremental pre-training → full training → prompt word fine-tuning. The segmented intelligent generation model constructed in step S2 is trained and fine-tuned using this strategy. The specific process is as follows: Phase 1: Incremental Pre-training (Domain Adaptation) Data: Corpus of mechanical engineering (100,000+ entries) Task: Masked Language Modeling (MLM) task, randomized masking mechanics terminology. Objective: To enable the model to learn mechanical terminology and sentence structures. Training method: Continue pre-training on the basis of BERT-Base to learn mechanical domain terminology and sentence structures. Training parameters: Batch Size=32, Learning Rate=2e-5, Epochs=5, Optimizer=AdamW.

[0082] Using the mechanical domain corpus constructed in step S3 as training data, the basic BERT-Base model was incrementally pre-trained to allow the model to learn mechanical domain terminology, sentence structures, and expression habits. Training parameters were set as follows: Batch Size = 32, Learning Rate = 2e-5, Epochs = 5, and the AdamW optimizer was used. During training, a Masked Language Modeling (MLM) task was employed, randomly masking domain terms (such as "conditioning" and "gap fit") from the corpus, and having the model predict the correct term at the masked location. After incremental pre-training, the model gained the ability to recognize mechanical domain terminology and initially possessed the ability to understand and generate mechanical domain text.

[0083] Phase Two: Full Training (Mapping Adaptation) Data: Mapping relation database (nodes / edges → text fragments) Task: Sequence generation task, input is structured feature map vector, output is corresponding text fragment. Model: Using the Encoder trained in the first stage, connect it to the newly initialized Decoder. Training method: The input of the Encoder is a sequence of structured feature vectors, the input of the Decoder is the first n-1 words of the natural language description, and the output of the Decoder is the nth word of the natural language description.

[0084] Training parameters: Batch Size=16, LearningRate=1e-5, Epochs=8, Loss=Cross-Entropy Loss.

[0085] Using "node / edge-text fragment" mapping pairs from a mapping relation database as training data, the incrementally pre-trained model was fully trained to learn the precise mapping relationship between structured features and natural language descriptions. The training task was a "structured feature → text fragment" generation task, with the input being the structured features obtained in step S2 and the output being the corresponding natural language text fragments. Training parameters were set as follows: batch size = 16, learning rate = 1e-5, number of training epochs = 8, and the cross-entropy loss function was used. During training, an early stopping strategy was adopted, stopping training when the validation set loss did not decrease for three consecutive epochs. After full training, the matching accuracy of the text fragments generated by the model with the structured features was significantly improved.

[0086] Phase 3: Fine-tuning of prompt words (generation optimization) Data: Prompt word templates + structured features → complete descriptive text Task: Use prompt word templates to guide the model to generate a structurally complete and logically coherent technical description. Training parameters: Batch Size=8, Learning Rate=5e-6, Epochs=3.

[0087] Design a dedicated prompt word template for describing mechanical parts, and fine-tune the fully trained model to optimize the logic and standardization of the generated text. This case study provides sample prompt word templates: "Please generate a standardized natural language description based on the following structural features of the mechanical parts, including overall part information, core geometric features, engineering attributes, and process information: [Input structured feature vector summary]" and "Please describe in detail the positional relationships and parameters of the following geometric features: [Input feature node vector]". The fine-tuning data uses 20 initial natural language texts generated in step S3, constructed in the form of "prompt word + structured feature → target text". Training parameters are set as follows: batch size = 8, learning rate = 5e-6, training epochs = 3. After prompt word fine-tuning, the logical coherence of the model-generated text is significantly improved, and the standardization of expression meets the requirements of technical documents in the mechanical field.

[0088] Step S5: Import the multi-dimensional data of the new mechanical part's 3D model into the segmented intelligent generation model and output the semantic representation of the mechanical part.

[0089] A new 3D model of a stepped shaft (model: stepped shaft_φ60-φ40-φ30, with a structure similar to the core stepped shaft in the embodiment, but different in size and process parameters) was selected as the test object. Its multi-dimensional data (extracted from CAD software using the self-developed program in step S1) was imported into the trained and fine-tuned segmented intelligent generation model. The model automatically completed the generation of structured features, vectorization processing, and semantic expression results. The final output natural language expression results are as follows: This part is a stepped shaft, model number stepped shaft_φ60-φ40-φ30, made of 45# steel, weighing 5.12kg, with a tensile strength of not less than 600MPa and a yield strength of not less than 355MPa. The stepped shaft body consists of three coaxial shaft segments: shaft segment 1, shaft segment 2, and shaft segment 3. Shaft segment 1 is cylindrical, with a diameter of φ60h7mm (upper deviation 0mm, lower deviation -0.03mm) and a length of 150mm; shaft segment 2 is cylindrical, with a diameter of φ40k6mm (upper deviation +0.022mm, lower deviation +0.003mm) and a length of 100mm; shaft segment 3 is cylindrical, with a diameter of φ30mm and a length of 60mm. Shaft segment 2 has a rectangular flat keyway in the middle, with a groove width of 12mm. The groove is 6mm deep and 40mm long. An annular relief groove, 6mm wide and 2.5mm deep, is provided at the transition points between shaft segment 1 and shaft segment 2, and between shaft segment 2 and shaft segment 3. Six 45° chamfers, each with a C3mm chamfer size, are provided at both ends of the part and at the shaft segment transition points. In terms of engineering properties, shaft segment 1 and the inner bore of the hub use an H7 / h7 clearance fit, while shaft segment 2 and the inner ring of the bearing use a K6 / h5 transition fit. Regarding process information, the parts require heat treatment, achieving a hardness of 220-250 HBW after heat treatment. Surface treatment includes chrome plating (0.01-0.02mm thickness) on shaft segments 1 and 2, and sandblasting on other surfaces. Machining methods include turning (outer diameter of shaft segments), milling (flat keyway), and grinding (outer diameter of shaft segments 1 and 2). This application provides a systematic solution to address the long-standing semantic gap in 3D models within the mechanical field. Its innovation lies primarily in: 1. A hierarchical mesh structured feature that integrates geometric, engineering, and technological semantics is proposed as a general and computable intermediate semantic representation connecting 3D data and natural language.

[0090] 2. A segmented intelligent generation model architecture of "feature modeling → vector encoding → semantic generation" was designed. In particular, the vectorization processing steps of directional encoding and domain word vector optimization were added to enhance the model's understanding of engineering semantics.

[0091] 3. An innovative three-stage training strategy of "incremental pre-training → full training → prompt word fine-tuning" was adopted, which enabled the general-purpose large language model to efficiently and deeply adapt to professional description tasks in the mechanical field and reduce the dependence on massive amounts of accurate matching data.

[0092] This application addresses the "understanding gap" and "application barrier" existing in mechanical 3D models, as well as the problems of low efficiency and error susceptibility of manual description, by proposing an innovative natural language representation method and system. Its core innovation lies in constructing a segmented intelligent generation model. This model transforms the multi-dimensional geometric, engineering, and technological data of the 3D model of mechanical parts into a hierarchical, network-like structured feature, establishes a mapping relationship between this structured feature and natural language, and ultimately drives a segmented intelligent generation model (large language model) deeply trained with domain knowledge to generate accurate and structured semantic expressions.

[0093] Specifically, the beneficial effects of this application are reflected in: 1. From Multi-Source Data to Structured Data Representation: A self-developed program extracts multi-dimensional data (geometric, engineering, and process data) from CAD software and constructs a hierarchical network structure of "part-feature combination-feature" for unified representation. This structure uses nodes to represent entities and features, and directed / undirected edges to accurately depict geometric topology, engineering dependencies, and process sequence relationships, providing a semantically rich and relationally structured foundation for subsequent natural language generation.

[0094] 2. Domain-Knowledge-Based Phased Model Training: An innovative phased strategy of "incremental pre-training + full training + prompting fine-tuning" is adopted to train the model. First, the professional knowledge of the model is enhanced using a corpus of mechanical engineering. Then, deep adaptation is performed through a mapping database of structured features and natural language, enabling the large language model to deeply understand the professional terminology and descriptive logic of the mechanical engineering field, thereby generating professional and accurate descriptive text.

[0095] 3. Achieving efficient, accurate, and computable natural language output: This method can automatically transform new 3D models into structured natural language descriptions. This not only greatly improves the efficiency and consistency of description generation, avoiding human omissions and biases, but also produces semantic results that are both human-readable and machine-processable, providing a crucial semantic foundation for downstream intelligent applications such as model retrieval, knowledge reasoning, and process planning.

[0096] 4. Forming a complete system solution: The supporting system integrates feature extraction, intelligent generation, storage and output modules, productizes the above methods and processes, and realizes end-to-end automation from 3D model input to natural language report output, effectively promoting knowledge sharing and collaboration among multiple fields such as design, manufacturing, and operation and maintenance.

[0097] Starting from the technical problem, existing technology, publication number CN120783362A, discloses a geometric feature recognition method based on a large language model, which mainly solves the problems of automation and semanticization of feature recognition in three-dimensional geometric models. Its focus is on extracting geometric elements from STL, STEP, and other format files and identifying their topological relationships, but it is still limited to the geometric level. Existing technology publication number CN115034222A discloses a multi-dimensional named entity recognition method based on mechanical parts processing literature data, which is aimed at textual materials in the field of mechanical processing. It aims to extract named entities such as cutting tools, machine tools, and materials from unstructured documents, falling under the category of pure text information extraction.

[0098] This application aims to address the "understanding gap" and "application barriers" existing in mechanical 3D models. This deeper issue not only focuses on geometry, but also integrates engineering attributes (such as material, weight, and strength) and process information (such as processing methods, heat treatment, and tolerance fits). It aims to build a complete knowledge representation that integrates geometric, engineering, and process semantics, thereby supporting semantic integration and automated processing across multiple stages from design to manufacturing.

[0099] Regarding data sources and representation methods, publication number CN120783362A proposes a geometric feature recognition method based on a large language model. Its data source is the mesh or boundary representation file of a 3D model, which is transformed into a structured natural language description (B-rep) through rule templates. Publication number CN115034222A discloses a multi-dimensional named entity recognition method based on mechanical parts processing literature data. Based on text corpora such as mechanical processing manuals and papers, it is transformed into a word vector sequence after annotation. This application employs a multi-source heterogeneous data fusion approach, directly extracting geometric features, engineering attributes, and process information from CAD systems (such as SolidWorks) and constructing them into a hierarchical network graph structure. In this structure, nodes represent parts, geometric features, engineering features, and process steps, while edges precisely represent various relationships such as geometric topology, engineering dependencies, and process sequences. This representation method not only includes richer semantic dimensions but also explicitly models the complex relationships between features through a graph structure, laying a more solid and structured foundation for subsequent semantic generation.

[0100] Regarding model architecture and processing flow, publication number CN120783362A proposes a geometric feature recognition method based on a large language model. It relies on the reasoning ability of a pre-trained large language model to understand geometric description text and output feature topology maps through an attention mechanism. Publication number CN115034222A discloses a multi-dimensional named entity recognition method based on mechanical parts processing literature data. It adopts a classic NLP pipeline and combines BERT, BiLSTM, Attention, MOEE and CRF models for entity recognition.

[0101] This application innovatively proposes a segmented intelligent generation model. The model first constructs a graph structure from multi-dimensional data, then vectorizes it using an improved Graph Attention Network (GAT). It innovatively incorporates a directional encoder to distinguish the different relationships represented by directed / undirected edges, and introduces a similarity calculation module to replace general word vectors with mechanical domain-specific word vectors to enhance the accuracy of domain semantics. Finally, a large language model adapted for deep domain adaptation generates the final natural language description. This segmented process of "feature modeling → vector encoding → semantic generation" ensures independent optimization and overall synergy at each stage.

[0102] Regarding training strategies and adaptive capabilities, CN120783362A proposes a geometric feature recognition method based on a large language model, which uses incremental learning and prompt word optimization to adapt to new geometric features; CN115034222A discloses a multi-dimensional named entity recognition method based on mechanical parts processing literature data, which pre-trains BERT using domain corpus and then uses labeled data for incremental training.

[0103] This application, however, designs a more systematic and in-depth three-stage training strategy: the first stage uses a constructed corpus of mechanical engineering linguistics for incremental pre-training, enabling the model to master professional terminology and sentence structures; the second stage uses a database of mapping relationships between structured features and natural language fragments for full training, enabling the model to learn accurate conversion from structured features to text; the third stage uses carefully designed prompt word templates for fine-tuning, optimizing the logic and standardization of the generated text. This tiered training strategy ensures that the model not only possesses domain knowledge but also masters the professional expression logic of mechanical engineering, thereby generating high-quality, structured descriptive text.

[0104] Regarding output format and application scenarios, CN120783362A proposes a geometric feature recognition method based on a large language model, which outputs a JSON / XML file containing feature semantic tags and topological relationships, mainly used for integration into CAD or PLM systems; CN115034222A discloses a multi-dimensional named entity recognition method based on mechanical parts processing literature data, which outputs a named entity tag sequence, serving knowledge graph construction or literature mining.

[0105] The final output of this invention is a well-structured, hierarchical, and highly readable natural language description report, covering overall part information, core geometric features, engineering attributes, and process information. This output is easy for engineers to read and understand directly, and its structured and semantic characteristics allow it to be directly parsed and processed by computer systems, thus supporting a wider range of downstream applications such as intelligent process planning, automated review, cross-system semantic exchange, and digital delivery.

[0106] In summary, the method and system employed in this application achieve the following: First, it realizes a unified representation and deep fusion of multi-dimensional information (geometry, engineering, and process) of mechanical parts, breaking through the limitations of single data modalities. Second, it innovatively adopts a hierarchical network graph structure and an improved graph neural network encoding method to achieve explicit modeling and accurate vectorization of complex engineering relationships. Third, its systematic three-stage training framework ensures a deep integration of professionalism and accuracy in the generated model. Finally, this application provides an end-to-end automated solution from 3D models to structured text, generating semantic descriptions that are both human-readable and machine-processable, effectively bridging the "understanding gap" between different fields, roles, and systems, and providing a crucial semantic foundation for knowledge transfer and collaborative applications in the intelligent manufacturing environment.

[0107] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

Claims

1. A natural language representation method for a three-dimensional model of a mechanical part, characterized in that, include: Acquire multi-dimensional data of historical mechanical parts 3D models, including the geometric features and engineering properties of the mechanical parts 3D models; Based on the multi-dimensional data of the mechanical parts, a segmented intelligent generation model is constructed. The segmented intelligent generation model is a three-segment serial model, which is used to generate the structured features of the mechanical parts, vectorize the structured features, and generate the semantic expression results of the mechanical parts. The structured features are graph structures composed of nodes and edges. Nodes represent the mechanical parts entities and their constituent features, and edges represent the topological and engineering relationships between geometric features. Construct a corpus of mechanical engineering, and use the mechanical engineering corpus to perform natural language description of the structured features; A mapping relationship database between the structured features and the natural language description is established, and a segmented intelligent generation model is trained and fine-tuned in stages based on the mechanical domain corpus and the mapping relationship database; The multi-dimensional data of the new mechanical part's 3D model is imported into the fine-tuned segmented intelligent generation model, and the semantic representation of the new mechanical part is output.

2. The natural language representation method for three-dimensional models of mechanical parts according to claim 1, characterized in that, The process involves acquiring multi-dimensional data of historical mechanical parts' three-dimensional models and extracting this multi-dimensional data from CAD software using a data extraction program. The geometric features in the multidimensional data include shape, size, tolerance, and positional relationships between geometric features; The engineering attributes in the multi-dimensional data include material, weight, and strength.

3. The natural language representation method for three-dimensional models of mechanical parts according to claim 1, characterized in that, The structured features adopt a hierarchical network architecture, distributing the mechanical part entities and the nodes that constitute the mechanical part feature information at different levels, with nodes at the same level connected in a network.

4. The natural language representation method for three-dimensional models of mechanical parts according to claim 3, characterized in that, The mechanical part feature information nodes are distributed at different levels by layering the nodes from top to bottom as part nodes, geometric feature combination nodes, and geometric feature nodes, with two or three layers; the mechanical part feature information nodes are distributed at different levels and connected at the same level by edges, which is a combination of directed and undirected edges, including geometric topological relationships and engineering relationships. The engineering relationships are represented using undirected edges; The geometric topological relationships and process relationships are represented using directed edges; The geometric topological relationships are used to characterize the positional relationships between geometric features and the subordinate relationships between geometric features and engineering features; The engineering relationship characterizes the positional relationship between the geometric features of the combination.

5. The natural language representation method for three-dimensional models of mechanical parts according to claim 4, characterized in that, The entity objects represented by the mechanical part feature information nodes include: part nodes, geometric feature nodes, and engineering feature nodes. The component node is the root node of the entire component; The geometric feature nodes are the basic geometric units that make up the parts, including volumes, surfaces, edges, and vertices; The engineering feature nodes are combinations of geometric features, including holes, slots, bosses, chamfers, and fillets.

6. The natural language representation method for three-dimensional models of mechanical parts according to claim 1, characterized in that, The segmented intelligent generation model is a three-segment serial model used to generate the structural features of parts, vectorize the structural features, and generate the semantic representation of mechanical parts, wherein: The vectorized processing of structured features involves adding a directional encoder and a similarity calculation module to the original image attention network to enhance the model's semantic understanding of nodes and edges. The directional encoder embeds a directional vector into the edge vector of the structured feature and concatenates the node vector with the embedded directional vector. The similarity calculation module compares the similarity values ​​between general word vectors and domain-specific word vectors, and replaces the general word vectors with domain-specific word vectors. The semantic representation of the generated parts is used to guide the segmented intelligent generation model, after fine-tuning, to generate natural language descriptions of the mechanical parts.

7. The natural language representation method for three-dimensional models of mechanical parts according to claim 1, characterized in that, The establishment of the mapping relationship database between the structured features and natural language descriptions is a mapping relationship between the feature information of nodes and edges and the structured natural language, formed by expressing the feature information of nodes and edges in structured natural language.

8. The natural language representation method for three-dimensional models of mechanical parts according to claim 1, characterized in that, The process of importing multi-dimensional data from the new 3D model of the mechanical part into the fine-tuned segmented intelligent generation model includes: First, incremental pre-training of the large language model is performed using a corpus in the mechanical field. Then, the mapping relationship database is used to perform full training on the model after incremental pre-training; Then, by using preset prompt word templates, a segmented model for intelligently generating natural language descriptions of mechanical parts is obtained after training.

9. A natural language representation system for a three-dimensional model of a mechanical part, used to implement the natural language representation method for a three-dimensional model of a mechanical part as described in any one of claims 1 to 8, characterized in that, include: The feature extraction module is used to extract the geometric features, engineering attributes, and process information of the 3D model of mechanical parts; The segmented intelligent generation module is used to extract multi-dimensional source data of mechanical parts, generate structured features of mechanical parts, and perform vectorization processing to intelligently generate semantic representation results of mechanical parts. The storage module is used to store corpora and mapping relationship databases in the field of mechanics; The output module is used to output and display the semantic representation of mechanical parts.

10. The natural language representation system for a three-dimensional model of a mechanical part as described in claim 9, characterized in that, The segmented intelligent generation module includes: The structured feature generation submodule is used to generate structured features for mechanical parts; A structured feature vectorization processing submodule is used to perform vectorization processing on the structured features; The semantic representation result generation submodule is used to generate semantic representation results for mechanical parts.

Images