A voice driven mechanical component modeling method and system

Abstract

This invention discloses a voice-driven method and system for modeling mechanical components. The user inputs natural language commands containing design intent into a CAD system via voice, which are then parsed into structured operation commands. Simultaneously, a baseline model is selected, and geometric codes are generated for each face based on its local geometric attributes, constructing a geometric structure coding matrix. Next, a machining face prediction network is used to predict each machining face in the machining features, resulting in a sequence composed of all machining face IDs of the machining features, which is denoted as the machining face ID sequence. Finally, the CAD platform API is called to obtain the model of the target mechanical component. This method connects the designer's high-level design intent with low-level geometric operations, avoiding the cumbersome interaction and cognitive burden of manually selecting and manipulating low-level geometric elements (points, edges, faces) in traditional direct modeling, thus significantly improving design efficiency.

Description

Technical Field

[0001] This invention belongs to the field of mechanical component modeling, and specifically relates to a voice-driven mechanical component modeling method and system. Background Technology

[0002] When modeling mechanical components, designers often need to quickly adjust existing models without tracing back to their original design steps. For example, when adjusting the position of a heat sink on an inlet housing, traditional parametric modeling requires locating the corresponding operation in the feature tree, while direct modeling technology allows designers to directly select the surface of the sink and move it by dragging.

[0003] Direct modeling (DMM) technology allows users to intuitively manipulate geometric elements in boundary representation (B-rep) models. Because it supports free geometric editing without a history record, it significantly improves design flexibility and has been widely integrated into mainstream commercial software such as Siemens NX, PTC Creo, and Autodesk Inventor. Unlike traditional parametric modeling, which relies on feature trees and constraints, DMM allows users to directly perform intuitive operations such as pushing, pulling, moving, and rotating geometric elements (e.g., faces, edges, vertices) in boundary representation (B-rep) models, without needing to consider the model's creation history or parent-child dependencies. This results in higher responsiveness and greater operational freedom in scenarios such as conceptual design, reverse engineering, and model modification. For these reasons, it is widely recognized as one of the most important technological advancements in the CAD field over the past decade.

[0004] While current mainstream direct modeling technologies have enabled free editing of CAD models without historical records, significantly improving design flexibility, their interaction methods remain stuck in the explicit selection and manipulation of low-level geometric elements—vertices, edges, and faces—in the boundary representation model. This low-level interaction mode forces designers to manually break down high-level design intentions (such as "move this slot forward 5mm") into a series of tedious and precise geometric selection and parameter setting steps when making any modifications. This not only significantly increases the cognitive load on users but also reduces design efficiency.

[0005] Especially when dealing with complex models, users often need to spend a lot of time selecting dozens or even hundreds of scattered faces and checking the operation parameters one by one, which can easily lead to rework due to misselection or incorrect parameters. Although deep learning technology has made progress in the field of CAD feature recognition in recent years (such as models like UV-Net), these methods generally use discriminative AI to assign a static, context-independent feature label to each face in the model. This "one-size-fits-all" recognition method cannot adapt to the dynamic context of the design process—the same face may belong to different features under different design instructions. Therefore, existing technologies cannot establish a reliable and dynamic mapping from "semantic features" described in natural language to "geometric face sets" in boundary representation models, causing "semantic-driven" direct modeling to remain at the conceptual level and unable to be implemented.

[0006] In summary, current direct modeling still relies on manual manipulation of underlying geometric elements such as points, lines, and surfaces, which is cumbersome, error-prone, and inefficient. Existing AI recognition methods lack user semantic understanding and cannot accurately and efficiently achieve intelligent mapping from semantics to geometry. Summary of the Invention

[0007] To address the shortcomings of the prior art, this invention provides a voice-driven method and system for modeling mechanical components. This system can directly drive the modification of CAD models through natural language commands, overcoming the problems of cumbersome human-computer interaction, high cognitive load, and the inability of the system to understand and respond to the designer's high-level semantic intent in the traditional modeling process.

[0008] The technical solution of the present invention is as follows:

[0009] I. A voice-driven modeling method for mechanical components

[0010] 1) Based on the structure of the target mechanical component, find and obtain a boundary representation model that is most similar to the target mechanical component in terms of structure in the model database, and use this boundary representation model as the benchmark model; then, record the structures in the benchmark model that are different from the target mechanical component as processing features;

[0011] 2) Determine one surface in the processing feature and record it as the starting processing surface; and obtain the voice processing instruction, then parse the voice processing instruction to obtain the structured operation instruction;

[0012] 3) Generate the geometric code corresponding to each face based on the local geometric attributes of each face in the baseline model, and then construct the geometric structure coding matrix of the baseline model by combining the adjacency topology.

[0013] 4) Based on the initial machining surface, structured operation instructions, and geometric structure encoding matrix, a Transformer-based machining surface prediction network is used to predict each machining surface in the machining feature, thereby obtaining a sequence composed of all machining surface IDs of the machining feature in sequence and denoted as the machining surface ID sequence.

[0014] 5) Input the structured operation instructions and the machining surface ID sequence into the direct modeling system, automatically execute the CAD platform API call, and finally obtain the model of the target mechanical component.

[0015] 6) If there is more than one processing feature, repeat steps 2)-5) to adjust the different processing features sequentially based on the baseline model after the previous adjustment until the final model is obtained.

[0016] The process of parsing the speech processing instructions to obtain structured operation instructions specifically includes:

[0017] The speech processing instruction is converted into text format using a speech recognition model. Then, the text format speech processing instruction is semantically parsed by a large model to obtain JSON data containing the target feature name, operation type, and operation parameters, which is recorded as a structured operation instruction.

[0018] Specifically, 3) refers to:

[0019] 3.1) Decompose each face of the baseline model so that each face is composed of multiple triangular facets, each edge on each face is composed of multiple line segments, and each loop is a polygon composed of multiple line segments; then extract the direction, coordinate points, center point, corner point, normal and adjacent triangular facet number of each polygon;

[0020] 3.2) Extracting local shape features: Use a MeshNet neural network to encode all triangular facets of each face to obtain the initial encoding of that face; and use a convolutional neural network to encode all polygons of each face to obtain the current ring encoding; add the current ring encoding to the corresponding initial encoding to obtain the current geometric encoding;

[0021] 3.3) Repeat steps 3.1)-3.2) to encode the remaining surfaces of the baseline model and obtain the geometric codes corresponding to the remaining surfaces;

[0022] 3.4) Based on the adjacency topology, the geometric codes of all faces of the baseline model are aggregated to obtain the geometric structure coding matrix of the baseline model.

[0023] Specifically, 4) refers to:

[0024] 4.1) Convert the target feature names in the structured operation instructions into text encoding; and construct labels based on the currently processed features;

[0025] 4.2) Train the Transformer-based processing surface prediction network based on text encoding, geometric structure encoding matrix and labels to obtain the trained processing surface prediction network;

[0026] 4.3) Input the geometric code of the starting processing surface into the trained processing surface prediction network. The network regressively generates the geometric code of the next processing surface until the prediction ends. The sequence composed of all processing surface IDs of the processing features is recorded as the processing surface ID sequence.

[0027] Specifically, 4.2) refers to:

[0028] The Transformer-based processing surface prediction network includes a connected Transformer-based encoder and a Transformer-based decoder; the text encoding and geometric structure encoding matrix are input together into the Transformer-based encoder, and the encoder outputs a geometric structure encoding matrix with semantic information.

[0029] A geometric code from a semantically information-rich geometric structure encoding matrix is ​​randomly selected and used as input to a Transformer-based decoder. The decoder then generates the geometric code for the next face autoregressively. Based on the geometric code of the next face, the face in the semantically information-rich geometric structure encoding matrix that is most similar to the geometric code of the next face is identified and its ID is recorded. This ID is then added to the face ID sequence. The decoder predicts the geometric code of the next face based on the current face ID sequence and its input. The loss function is calculated based on the final face ID sequence and labels, and backpropagation and gradient descent are performed to update the parameters of the Transformer-based processing face prediction network. After several iterations of training and convergence of the loss function, the final network weights are obtained, thus yielding the trained processing face prediction network.

[0030] Calculate the cosine similarity between the geometric code of the next face and other faces in the geometric structure encoding matrix with semantic information, and denote the face with the largest cosine similarity as the face most similar to the geometric code of the next face.

[0031] II. A Voice-Driven Mechanical Component Modeling System

[0032] Storage units are used to store boundary representation models of different mechanical components;

[0033] The model selection unit is used to find and obtain a boundary representation model that is most similar to the target mechanical component in terms of structure in the model database, and use the boundary representation model as the reference model; and to record the structures in the reference model that are different from the target mechanical component as processing features.

[0034] The user action acquisition unit is used to acquire the starting processing surface of the processing features and to acquire the voice processing instructions.

[0035] The speech parsing unit is used to parse the speech processing instructions to obtain structured operation instructions;

[0036] The geometric coding unit is used to generate the geometric code corresponding to each face based on the local geometric attributes of each face in the benchmark model, and then combine the adjacency topology to construct the geometric structure coding matrix of the benchmark model.

[0037] The surface sequence generation unit is used to predict each processing surface in the processing feature based on the starting processing surface, structured operation instructions and geometric structure encoding matrix, using a Transformer-based processing surface prediction network, thereby obtaining a sequence composed of all processing surface IDs of the processing feature in sequence and denoted as the processing surface ID sequence.

[0038] The direct modeling API call unit is used to automatically and execute CAD platform API calls based on structured operation instructions and machining surface ID sequences, ultimately obtaining the model of the target mechanical component.

[0039] III. A computer device

[0040] The computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the voice-driven mechanical component modeling method.

[0041] IV. A computer-readable storage medium

[0042] The medium stores a computer program, which, when executed by a processor, implements the steps of the voice-driven mechanical component modeling method.

[0043] The beneficial effects of this invention are:

[0044] This invention directly drives the modification of CAD models in the database through natural language commands, eliminating the need for designers to manually select dozens of scattered faces or memorize complex menus, thus greatly improving operational efficiency. At the same time, with the help of generative feature recognition technology based on deep learning, the system can accurately locate the operation object according to the user's semantic context, completely avoiding misselection and rework caused by static recognition in traditional methods, and significantly reducing cognitive load. Attached Figure Description

[0045] Figure 1 This is the overall process framework diagram of the present invention.

[0046] Figure 2 This is a schematic diagram of geometric coding calculation.

[0047] Figure 3 This is a network architecture diagram.

[0048] Figure 4 This is a schematic diagram of the process of generating the face ID sequence.

[0049] Figure 5 This is a flowchart of the intelligent direct modeling process.

[0050] Figure 6 This is a flowchart of the method of the present invention. Detailed Implementation

[0051] The present invention will now be described in further detail with reference to the accompanying drawings.

[0052] like Figure 6 As shown, the voice-driven mechanical component modeling method proposed in this invention includes the following steps:

[0053] 1) Based on the structure of the target mechanical component, find and obtain the boundary representation model that is most similar to the target mechanical component in terms of structure in the model database of the CAD system. Use this boundary representation model as the reference model. Then, record the structures in the reference model that are different from the target mechanical component as machining features. Boundary representation model is a commonly used representation form in CAD models. Each machining feature consists of multiple faces of a structure in the boundary representation model and is a group of faces with actual machining significance.

[0054] The target mechanical components can be stepped shafts and splined shafts among shaft-type parts, and flanges (with bolt holes) and end caps (possibly with bosses and mounting holes) among disc-type parts. The model database includes various standard part families and common blank models. Standard part families include smooth shafts, non-perforated flanges, standard shapes, etc., and common blank models include rectangular blocks, stepped shaft blanks, ring blanks, etc.

[0055] 2) Identify a face within a machining feature and designate it as the starting machining face. In one feasible implementation, the user clicks on a face within the machining feature they wish to modify using the mouse, and the clicked face is designated as the starting machining face. Additionally, the user issues a voice machining command to the CAD system, which retrieves the corresponding voice machining command for that feature through a voice interaction plugin. The voice machining command is then parsed to obtain a structured operation instruction.

[0056] In one feasible implementation, the structured operation instructions are obtained after parsing the speech processing instructions, specifically including:

[0057] The speech processing instruction is converted into text format using a speech recognition model. Then, the text format speech processing instruction is semantically parsed by a large model to obtain JSON data containing the target feature name, operation type, and operation parameters, which is recorded as a structured operation instruction. The JSON data is structured data.

[0058] Optionally, a commonly used speech recognition model, such as Whisper, can be selected. For example, the speech processing instruction output by the speech recognition model in text format is: "Help me rotate this waist-shaped slot 90 degrees clockwise around the Z-axis."

[0059] The process of semantic parsing of text-formatted speech processing instructions by the large model consists of the following three steps:

[0060] Feature recognition and localization: The Large Model (LLM) first identifies the target CAD feature name explicitly mentioned in the voice processing instructions, such as "waist-shaped through slot", "circular blind hole" or "hexagonal channel", to ensure that the system operates on the correct target;

[0061] Operation and parameter extraction: The large model (LLM) sequentially extracts one or more operation types (such as "move", "rotate", "offset", "bevel") and their corresponding parameters (such as distance, angle, direction axis, radius, etc.) contained in the instructions to ensure the integrity of the geometric operation intent;

[0062] Structured Output: Large Models (LLMs) organize the extracted features, operations, and parameters into a predefined JSON format for output. For example, the output JSON data might look like this:

[0063] {"Features": {

[0064] Name: Waist-shaped through groove

[0065] "operate": [{

[0066] Type: Rotation

[0067] "Parameters": {

[0068] Angle: 90

[0069] Direction: Clockwise

[0070] "Coordinate axis": "Z-axis"}}]}

[0071] 3) Generate the geometric code corresponding to each face based on the local geometric attributes of each face in the baseline model, and then construct the geometric structure coding matrix of the baseline model by combining the adjacency topology; the geometric structure coding matrix is ​​the set of geometric codes of all faces in the baseline model.

[0072] 3) Specifically:

[0073] 3.1) Using the OpenCascade library, each face of the baseline model is split into multiple triangular facets, each edge of each face is composed of multiple line segments, and each loop (the closed boundary of the face) is a polygon composed of multiple line segments, such as... Figure 2 As shown in (a); next, extract the orientation, coordinates (data dimensions are M×3, M×3 respectively, where M is the number of line segments in the polygon), center point, corner point, normal, and adjacent triangle numbers of each polygon (data dimensions are N×3, N×9, N×3, N×3 respectively, where N is the number of triangles), as shown in (a). Figure 2 As shown in (b).

[0074] 3.2) Extracting local shape features: such as Figure 3 As shown, a MeshNet neural network is used to encode all triangular facets of each face to obtain the initial code (N×256) for that face; and a convolutional neural network (CNN) is used to encode all polygons (i.e., loops) of each face to obtain the current loop code (M×256); the current loop code is added to the corresponding initial code to obtain the current geometric code.

[0075] 3.3) Repeat steps 3.1)-3.2) to encode the remaining surfaces of the baseline model to obtain the geometric codes corresponding to the remaining surfaces.

[0076] 3.4) Based on the adjacency topology, the geometric codes of all faces of the baseline model are aggregated to obtain the geometric structure coding matrix of the baseline model. The aggregation operation can be either concatenation or addition.

[0077] 4) Based on the initial machining surface, structured operation instructions, and geometric structure encoding matrix, a Transformer-based machining surface prediction network is used to predict each machining surface in the machining feature, thereby obtaining a sequence composed of all machining surface IDs of the machining feature in sequence and denoted as the machining surface ID sequence.

[0078] 4) Specifically:

[0079] 4.1) The target feature name (e.g., "waist-shaped through slot") in the structured operation instructions is converted into a 256-dimensional text encoding using a text embedding model; and a label is constructed based on the current processing feature, i.e., an unordered sequence of face IDs consisting of all face IDs of each processing feature, such as [1, 2, 3, ..., ...]. <eos>],in <eos>The terminator indicates the end of the calculation and can be represented by the number "0". This label is used to calculate the loss and gradient descent.

[0080] 4.2) Train the Transformer-based processing surface prediction network based on text encoding, geometric structure encoding matrix and labels to obtain the trained processing surface prediction network;

[0081] 4.2) Specifically:

[0082] The Transformer-based processing surface prediction network consists of a connected Transformer-based encoder and a Transformer-based decoder; the text encoding and geometric structure encoding matrix are input together into the Transformer-based encoder, and the encoder outputs a geometric structure encoding matrix with semantic information.

[0083] A geometric code is randomly selected from the geometric structure encoding matrix containing semantic information and used as input to the Transformer-based decoder. The decoder then generates the geometric code for the next face via autoregression. Figure 4 As shown; autoregression refers to the model iteratively predicting the geometric code of the next face based on the generated historical sequence of face geometric codes. Based on the geometric code of the next face, the face most similar to the geometric code of the next face in the geometric structure encoding matrix with semantic information is identified and its ID is recorded and added to the face ID sequence. The decoder predicts the geometric code of the next face based on the current face ID sequence and the decoder input. The loss function is calculated based on the final face ID sequence and labels, and backpropagation and gradient descent are performed to update the parameters of the Transformer-based processing face prediction network. The loss function is specifically the binary cross loss function (BCE). After several iterations of training and the loss function converges, the final network weights are obtained, thus obtaining the trained processing face prediction network.

[0084] The formula for the binary cross-loss function BCE is as follows:

[0085]

[0086] Where BCELoss is the binary crossover loss; N is the sequence length; The true label is the i-th position in the sequence; This represents the label probability value predicted at position i by a Transformer-based processing surface prediction network.

[0087] In one feasible implementation, the cosine similarity between the geometric code of the next face and other faces in the geometric structure encoding matrix with semantic information is calculated, and the face with the largest cosine similarity is recorded as the face most similar to the geometric code of the next face.

[0088] 4.3) The geometric code of the starting machining surface is input into the trained machining surface prediction network. The network autoregressively generates the geometric code of the next machining surface until the prediction ends, i.e., an end marker is generated. <eos>The sequence consisting of all the processing surface IDs of the processing feature is denoted as the processing surface ID sequence. The tag construction ends with "...". <eos>"The marker ends here. The trained model will also generate this marker during inference."

[0089] 5) Input the operation type and parameters of the structured operation instructions together with the machining surface ID sequence into the direct modeling system, automatically execute the CAD platform API call, complete the automatic selection of features in the benchmark model and the direct modeling operation, and finally obtain the model of the target mechanical component. Based on this simulation model, simulation, machining and manufacturing are carried out.

[0090] The CAD platform's API should be selected from software that offers open secondary development interfaces and direct modeling operations, such as Siemens NX and SolidWorks.

[0091] 5) Specifically:

[0092] 5.1) The structured operation instructions (JSON format, including operation type and parameters) and the machining surface ID sequence are used as input to the direct modeling system. The API of the CAD software converts the machining surface ID sequence into a set of actual geometric surface objects in the system and binds them to the operation target in the API code.

[0093] 5.2) Inject the operation parameters (such as distance, angle, and direction) from the JSON data into the parameter fields of the corresponding API to complete the final configuration before execution;

[0094] 5.3) The system calls the "Submit" or "Execute" interface of the CAD platform, silently executing API code in the background to directly drive the CAD kernel to complete geometric modifications (such as moving, rotating, offsetting, deleting, etc.). Figure 5 As shown; and outputs the final model for further applications, such as simulation, processing, and manufacturing.

[0095] 6) If there is more than one machining feature, repeat steps 2)-5), adjusting different machining features sequentially based on the previously adjusted baseline model until the final model is obtained. For example, the final model can be exported and imported into a CAM system in STEP format to automatically generate machining paths. Taking a bracket part as an example, the user can use voice commands to sequentially complete operations such as "rotate the waist-shaped through groove 90 degrees clockwise around the Z-axis" and "chamfer the countersunk hole by 0.5mm". The system automatically recognizes the machining features and calls the API to complete the modeling update; the exported model is directly used in Mastercam software to generate five-axis toolpaths and drive CNC machine tool machining.

[0096] like Figure 1 As shown, the present invention also proposes a voice-driven mechanical component modeling system, the device comprising:

[0097] Storage units are used to store boundary representation models of different mechanical components;

[0098] The model selection unit is used to find and obtain a boundary representation model that is most similar to the target mechanical component in terms of structure in the model database, and use the boundary representation model as the reference model; and to record the structures in the reference model that are different from the target mechanical component as processing features.

[0099] The user action acquisition unit is used to acquire the starting processing surface of the processing features and to acquire the voice processing instructions.

[0100] The speech parsing unit is used to parse the speech processing instructions to obtain structured operation instructions;

[0101] The speech parsing unit includes a speech recognition module and a large model module.

[0102] The geometric coding unit is used to generate the geometric code corresponding to each face based on the local geometric attributes of each face in the benchmark model, and then combine the adjacency topology to construct the geometric structure coding matrix of the benchmark model.

[0103] The surface sequence generation unit is used to predict each processing surface in the processing feature based on the starting processing surface, structured operation instructions and geometric structure encoding matrix, using a Transformer-based processing surface prediction network, thereby obtaining a sequence composed of all processing surface IDs of the processing feature in sequence and denoted as the processing surface ID sequence.

[0104] The surface sequence generation unit includes a text encoding module and a neural network module.

[0105] The direct modeling API call unit is used to automatically and execute CAD platform API calls based on structured operation instructions and machining surface ID sequences, ultimately obtaining the model of the target mechanical component.

[0106] The present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a voice-driven mechanical component modeling method.

[0107] The present invention also proposes a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of a voice-driven mechanical component modeling method.< / eos> < / eos> < / eos> < / eos>

Claims

1. A voice-driven mechanical component modeling method, characterized by, The method comprises the following steps: 1) According to the structure of the target mechanical element, a boundary representation model most similar in structure to the target mechanical element is found and obtained in a model database, and the boundary representation model is taken as a reference model; then structures different from the target mechanical element in the reference model are recorded as machining features; 2) A face in the machining features is determined and recorded as a starting machining face; and a voice machining instruction is obtained, and then a structured operation instruction is obtained after the voice machining instruction is parsed; 3) A geometry coding matrix of the reference model is constructed according to the local geometry attributes of each face in the reference model, and the geometry coding matrix is combined with an adjacency topological relationship; 4) Based on the starting machining face, the structured operation instruction and the geometry coding matrix, a machining face prediction network based on Transformer is used to predict each machining face in the machining features, so that a sequence composed of all machining face IDs of the machining features in sequence is obtained and recorded as a machining face ID sequence; 5) The structured operation instruction is input into a direct modeling system together with the machining face ID sequence, and a CAD platform API call is automatically executed, and finally a model of the target mechanical element is obtained.

2. A method of voice-driven mechanical component modeling according to claim 1, wherein, The method further comprises the following steps: 6) If the machining features are more than one, steps 2) to 5) are repeated, and different machining features are adjusted in sequence on the basis of the reference model adjusted in the previous time, until the final model is obtained.

3. The method of modeling a mechanical element driven by speech of claim 1, wherein, The structured operation instruction obtained after the voice machining instruction is parsed comprises: The voice machining instruction is converted into a text format by using a voice recognition model; and after the voice machining instruction in the text format is semantically parsed by a large model, JSON data containing a target feature name, an operation type and an operation parameter are obtained and recorded as the structured operation instruction.

4. The method of modeling a mechanical element driven by speech of claim 1, wherein, 3) Specifically comprises: 3.1) Each face of the reference model is split, so that each face is composed of multiple triangular facets, each edge on each face is composed of multiple line segments, and each loop is a polygon composed of multiple line segments; then the direction, coordinate point, center point, corner point, normal and adjacent triangular facet number of each polygon are extracted; 3.2) Local shape features are extracted: the MeshNet neural network is used to encode all triangular facets of each face to obtain an initial coding of the face; and the convolutional neural network is used to encode all polygons of each face to obtain a loop coding of the current face; after the loop coding of the current face is added to the corresponding initial coding, the geometry coding of the current face is obtained; 3.3) Steps 3.1) to 3.2) are repeated to encode the remaining faces of the reference model to obtain the geometry coding corresponding to the remaining faces; 3.4) The geometry coding of all faces of the reference model is aggregated according to the adjacency topological relationship to obtain the geometry coding matrix of the reference model.

5. The method of modeling a mechanical element driven by speech of claim 1, wherein, 4) Specifically comprises: 4.1) The target feature name in the structured operation instruction is converted into a text coding; and a label is constructed according to the current machining feature; 4.2) The machining face prediction network based on Transformer is trained according to the text coding, the geometry coding matrix and the label, and a trained machining face prediction network is obtained; 4.3) input the geometry code of the starting machining surface into the trained machining surface prediction network, and the network generates the geometry code of the next machining surface in a self-regressive manner until the end of prediction, and the sequence composed of all machining surface IDs of the machining feature is recorded as the machining surface ID sequence.

6. A method of voice-driven mechanical component modeling according to claim 5, wherein, The 4.2) is specifically: The machining surface prediction network based on the Transformer comprises a connected encoder based on the Transformer and a decoder based on the Transformer; the text code and the geometry code matrix are input into the encoder based on the Transformer, and the encoder outputs the geometry code matrix with semantic information; a geometry code in the geometry code matrix with semantic information is randomly selected and used as the input of the decoder based on the Transformer, the decoder generates the geometry code of the next surface in a self-regressive manner; based on the geometry code of the next surface, the surface in the geometry code matrix with semantic information that is most similar to the geometry code of the next surface is determined and the ID of the most similar surface is recorded, and the surface ID is added to the surface ID sequence; the decoder predicts the geometry code of the next surface according to the current surface ID sequence and in combination with the input of the decoder, the loss function is calculated according to the final surface ID sequence and the label, and the back propagation and gradient descent are performed to update the parameters of the machining surface prediction network based on the Transformer; after the iterative training for several times and the convergence of the loss function, the final network weight is obtained, so that the trained machining surface prediction network is obtained.

7. A method of voice-driven mechanical component modeling according to claim 6, wherein, The cosine similarity between the geometry code of the next surface and other surfaces in the geometry code matrix with semantic information is calculated, and the surface with the maximum cosine similarity is recorded as the surface most similar to the geometry code of the next surface.

8. A voice driven mechanical element modeling system, characterized by, It comprises: a storage unit for storing boundary representation models of different mechanical elements; a model selection unit for finding and obtaining a boundary representation model most similar in structure to a target mechanical element from a model database according to the structure of the target mechanical element, taking the boundary representation model as a reference model, and recording the structure different from the target mechanical element in the reference model as a machining feature; a user action acquisition unit for acquiring a starting machining surface of the machining feature and acquiring a voice machining instruction; a voice analysis unit for analyzing the voice machining instruction to obtain a structured operation instruction; a geometry coding unit for generating a geometry code corresponding to each surface in the reference model according to the local geometry attribute corresponding to each surface in the reference model, and constructing a geometry code matrix of the reference model in combination with the adjacency topological relationship; a surface sequence generation unit for predicting each machining surface in the machining feature based on the starting machining surface, the structured operation instruction and the geometry code matrix by using the machining surface prediction network based on the Transformer, so as to obtain a sequence composed of all machining surface IDs of the machining feature and record the sequence as a machining surface ID sequence; a direct modeling API calling unit for automatically executing CAD platform API calling according to the structured operation instruction and the machining surface ID sequence, and finally obtaining a model of the target mechanical element. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. When the processor executes the computer program, it implements the steps of the voice-driven mechanical component modeling method according to any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the voice-driven mechanical component modeling method according to any one of claims 1 to 7.

Images