B-Rep feature recognition method and computer equipment

Abstract

The invention discloses a B-Rep feature recognition method and computer equipment. The B-Rep feature recognition method comprises the following steps: S1, acquiring topological features of B-Rep data by using a set airfoil kernel; s2, inputting the topological features and geometric features of the B-Rep data into a pre-constructed classification module to obtain a feature recognition result of the B-Rep data; wherein the topological features comprise neighborhood surface information, neighborhood edge information and neighborhood directed edge information of each directed edge in the B-Rep data. According to the method, accurate identification of the features in the CAD model is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer-aided design / process planning / manufacturing, in particular to a B-Rep feature recognition method and a computer device. BACKGROUND

[0002] Feature recognition (FR) is a key technology for the intelligent integration of computer-aided design (CAD), computer-aided process planning (CAPP), and computer-aided manufacturing (CAM) systems. Traditional feature recognition methods include rule-based, graph-based, and hint-based methods, while deep learning has been gradually applied in this field in recent years. BRepNet is a current deep learning architecture for CAD model feature recognition.

[0003] Current research on CAD model feature recognition mainly relies on deep learning, which identifies features by processing B-Rep data or point cloud data of CAD models. However, these methods often have insufficient recognition accuracy when faced with complex geometric shapes and topological structures in actual industrial scenarios. SUMMARY

[0004] The purpose of the present application is to provide a B-Rep feature recognition method and a computer device to accurately recognize model features.

[0005] To solve the above technical problems, the present application provides a B-Rep feature recognition method, comprising: S1, obtaining the topological features of B-Rep data using a set of wing surface kernels; S2, inputting the topological features and geometric features of B-Rep data into a pre-constructed classification module to obtain the feature recognition result of B-Rep data; Wherein, the topological features include neighborhood face information, neighborhood edge information, and neighborhood directed edge information of each directed edge in the B-Rep data.

[0006] According to the above scheme, the wing surface kernel includes basic topological operators n, p, m, e, and f. For the selected directed edge h, the topological operator n obtains the next directed edge of the directed edge h, the topological operator p obtains the previous directed edge of the directed edge h, the topological operator m obtains the paired directed edge of the directed edge h, the topological operator e obtains the corresponding geometric edge of the directed edge h, and the topological operator f obtains the associated face of the directed edge h. The associated face, geometric edge, and directed edge within the neighborhood of the directed edge constitute the neighborhood face information, neighborhood edge information, and neighborhood directed edge information.

[0007] According to the above scheme, the method for obtaining the geometric features of B-Rep data includes: S201, obtaining UV lattice data of B-Rep data, inputting the UV lattice data into a pre-built multi-key self-attention sub-network to obtain processed lattice information; S202 , obtaining a feature matrix of the B-Rep data, and inputting the processed lattice information and the feature matrix into a pre-built data fusion layer for data fusion to obtain geometric features.

[0008] According to the above scheme, the multi-key self-attention sub-network includes multiple self-attention encoders, and different self-attention encoders process data of different dimensions of UV dot matrix data; after a certain dimension of UV dot matrix data is input into the self-attention encoder, it undergoes linear transformation to obtain Query, Value and multiple groups of different Keys. Each group of Key interacts with Query to generate an independent attention weight distribution. Different attention weight distributions interact with Value to obtain different context vectors. Multiple groups of context vectors are fused through the averaging strategy and output as feature representation.

[0009] According to the above scheme, the method for obtaining geometric features is: aligning the point matrix information on the surface, edge and directed edge with the corresponding feature matrix in the data dimension.

[0010] According to the above scheme, the classification module includes a neural network, multiple auxiliary classifiers, a main classifier, and a weighted fusion layer; the neural network includes an initial layer, multiple intermediate layers, and a final layer. The number of auxiliary classifiers is the same as that of the intermediate layers and they correspond one to one. The multiple auxiliary classifiers collect the features of each intermediate layer respectively and obtain different auxiliary classification results. The main classifier collects the features of the final layer and obtains the main classification result. The weighted fusion layer performs weighted fusion on the main classification result and each auxiliary classification result based on the set weights to obtain the feature recognition result.

[0011] According to the above scheme, the wing core, the multi-key self-attention sub-network, and the data fusion layer constitute the encoding module. The encoding module and the classification module constitute the BRepNet++ network. The BRepNet++ network is trained using a self-built dataset.

[0012] According to the above scheme, the self-built data set comes from multiple industrial manufacturing fields, and the B-Rep data in the self-built data set is labeled based on the set rib position recognition standards and labeling categories.

[0013] According to the above scheme, in the rib position identification standard, rib positions are defined as thin-walled reinforcement structures that enhance structural strength and avoid deformation; for each surface in the B-Rep data, the annotation categories include rib position, other, and fuzzy. The fuzzy annotation category indicates that the surface has a topological structure similar to a rib position but does not actually belong to a rib position.

[0014] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the B-Rep feature recognition method described above when executing the computer program.

[0015] Beneficial effects The present invention obtains the topological information of the B-Rep data of the CAD model by introducing the wing surface kernel, obtains more comprehensive surface domain information of the model, and significantly improves the feature recognition accuracy of the subsequent classification module for complex geometric shapes.

[0016] Furthermore, the present invention improves the existing self-attention mechanism, and the obtained multi-key self-attention sub-network generates multiple groups of keys to obtain feature representations from different perspectives, thereby better capturing and integrating fine-grained geometric information and enhancing the ability to understand complex topological relationships.

[0017] Furthermore, the classification module of the present invention can provide additional supervisory signals by adding main classifiers and auxiliary classifiers at different levels of the neural network, thereby significantly improving the accuracy of the classification module in the recognition and reasoning stage.

[0018] Furthermore, the present invention trains the BRepNet++ network through a self-built data set, so that the trained BRepNet++ network can cope with the complex and diverse designs of real industries and meet the feature recognition requirements of real industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the B-Rep feature recognition method according to the first embodiment of the present invention; Figure 2 1 is a schematic diagram of the architecture of the BRepNet++ network according to the first embodiment of the present invention; Figure 3 Schematic diagram of the topological structure of the airfoil core according to the first embodiment of the present invention; Figure 4 Schematic diagram of the structure of the classification module of the first embodiment of the present invention; Figure 5 Schematic diagram of the structure of the multi-key self-attention sub-network of embodiment 1 of the present invention. DETAILED DESCRIPTION

[0020] To make the purpose, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0021] Example 1: Existing feature recognition networks fail to fully consider facet neighborhood information when processing complex CAD models, resulting in limited accuracy in recognizing complex geometric shapes. Furthermore, existing technologies often use synthetic data for algorithm training, rarely considering CAD model data from real industrial scenarios. This results in insufficient adaptability to real-world scenarios and makes it difficult to meet the practical needs of high-precision feature recognition.

[0022] In view of the above situation, this embodiment discloses a B-Rep feature recognition method, which aims to solve the following technical problems: 1) Improve the accuracy of feature recognition of complex geometric shapes and topological structures in CAD models; 2) Enhance the supervisory signal of the neural network during feature recognition and improve the learning effect of the model; 3) Improve the ability to extract fine-grained geometric information from UV-Grid data and enhance the understanding of complex topological relationships; 4) Enhance the applicability of the model in actual industrial scenarios to meet the actual needs of high-precision feature recognition.

[0023] See also Figure 1 The B-Rep feature recognition method of this embodiment includes: S1. Use the set Winged Face Kernel to obtain the topological features of B-Rep data; S2. Input the topological features and geometric features of the B-Rep data into the pre-built classification module (BRepNetClassifyLayers) to obtain the feature recognition results of the B-Rep data; The topological features include the neighborhood face information, neighborhood edge information, and neighborhood directed edge information of each directed edge in the B-Rep data.

[0024] Further, see Figure 3 ,The wing kernel includes basic topological operators n, p, m, e, f; For the selected directed edge h, the topological operator n obtains the next directed edge of the directed edge h, the topological operator p obtains the previous directed edge of the directed edge h, the topological operator m obtains the paired directed edge of the directed edge h, the topological operator e obtains the corresponding geometric edge of the directed edge h, and the topological operator f obtains the associated face of the directed edge h; The results of obtaining the associated faces, geometric edges, and directed edges in the directed edge neighborhood constitute neighborhood face information, neighborhood edge information, and neighborhood directed edge information; Specifically, the wing core B-Rep ring is composed of directed edges, and the complex relationships between faces in the CAD model are captured through systematic topological operators. The operation rules of each topological operator are as follows: n(h) operator: Get the next directed edge of the current directed edge h (next coedge) p(h) operator: Get the previous directed edge of the current directed edge h (previous coedge) m(h) operator: Get the paired directed edge of the current directed edge h (mate coedge) e(h) operator: Get the geometric edge (edge ​​of coedge) corresponding to the current directed edge h f(h) operator: Get the face of coedge associated with the current directed edge h; For each directed edge h, the wing surface kernel obtains the information of the face, edge, and directed edge in the neighborhood of the directed edge through a set of operators; The information of the neighboring faces includes F={"f", "mf", "nmf", "pmf", "mnmf", "mpmf"}; The information of the neighborhood edges includes E={"e", "ne", "pe", "mne", "mpe"}; The information of the neighborhood directed edges includes C={"i", "m", "n", "p", "mn", "mp"}; Define the winged surface topology kernel Kernel={F, E, C}.

[0025] Each element in F, E, and C represents the combination of the results obtained by each operator, as well as the directed edge h itself "i".

[0026] Further, see Figure 2 , the method for obtaining the geometric features of B-Rep data includes: S201, obtaining UV lattice data of B-Rep data, inputting the UV lattice data into a pre-built multi-key self-attention sub-network to obtain processed lattice information; S202 , obtaining a feature matrix of the B-Rep data, and inputting the processed lattice information and the feature matrix into a pre-built data fusion layer for data fusion to obtain geometric features.

[0027] Further, see Figure 5The multi-key self-attention subnetwork includes multiple self-attention encoders, and different self-attention encoders process different dimensions of UV dot matrix data. After a certain dimension of UV dot matrix data is input into the self-attention encoder, it undergoes linear transformation to obtain Query, Value, and multiple groups of different Keys. Each group of Key interacts with Query to generate an independent attention weight distribution. Different attention weight distributions interact with Value to obtain different context vectors. Multiple groups of context vectors are fused through an averaging strategy and output as feature representations. In this embodiment, the implementation of the multi-key self-attention sub-network includes the following steps: 1) Building a basic architecture for multi-key self-attention: adding an independent key projection layer to the traditional self-attention module to generate feature representations from different perspectives. UV-grid data undergoes a linear transformation to generate queries, values, and multiple sets of keys. Each key interacts with the query to generate an independent attention weight distribution, which then interacts with the value to generate a corresponding context vector. Finally, an averaging strategy is used to fuse multiple context vectors to form a feature representation. 2) For UV-Grid data of different dimensions, a self-attention mechanism is packaged to implement a feature encoder. A multi-key self-attention module is integrated to replace the traditional convolutional layer for feature extraction. Layer normalization is added to enhance training stability. Residual connections are used to ensure smooth information transmission. A feedforward network with LeakyReLU activation is used to enhance nonlinear expression capabilities. 3) Feature encoding modules are implemented for curves and surfaces separately. First, a linear layer is used to map the input channels to a fixed-dimensional feature space and a LayerNorm normalization layer is added. At the same time, a learnable position encoding parameter is introduced to capture the sequential position information of the curve data. Four layers of attention encoding layers are stacked in sequence to extract more abstract feature representations layer by layer. Finally, the final output dimension is adjusted through a fully connected layer.

[0028] Furthermore, the method for obtaining geometric features is to align the point lattice information on the surface, edge and directed edge with the corresponding feature matrix in the data dimension.

[0029] Further, see Figure 2 、 Figure 5 The classification module includes a neural network, multiple auxiliary classifiers, a main classifier, and a weighted fusion layer. The neural network includes an initial layer, multiple intermediate layers, and a final layer. The number of auxiliary classifiers is the same as the number of intermediate layers and they correspond one to one. The multiple auxiliary classifiers collect the features of each intermediate layer and obtain different auxiliary classification results. The main classifier collects the features of the final layer and obtains the main classification result. The weighted fusion layer performs weighted fusion on the main classification result and the auxiliary classification results based on the set weights to obtain the feature recognition result. Specifically, the implementation of the classification module in this embodiment includes the following steps: 1) Collecting intermediate layer features at different depth positions of the neural network, and finally collecting the final layer features; from the 2nd layer to the N-1th layer, each layer has an auxiliary classifier to process and convert the input geometric features; the last layer is provided with a main classifier to process the final output and generate the final surface feature; each classifier is composed of a linear layer to map the surface feature generated by the layer to a feature recognition classification space; 2) Weighted calculation of the prediction results of each layer by the fusion mechanism; in the training stage, the weighted fusion method is used to integrate the classification results at different depths; the final classification score is composed of the output of the main classifier and the weighted sum of the outputs of the auxiliary classifiers; in this embodiment, the contribution of each auxiliary classifier is given a weight of 0.1, and the weight of the main classifier is set to 1 to balance the importance of the output results of each layer; 3) In the inference stage, the consistency with the training stage is maintained, and the rich feature information extracted from different depth layers is used by the auxiliary classifiers to ensure that the model behaves consistently in the two stages and fully utilizes the information of the intermediate layers for final decision-making.

[0030] Further, referring to Figure 2 , the wing surface kernel, the multi-key self-attention subnetwork, and the data fusion layer constitute an encoding module (UVNetAttentionEncoder), the encoding module and the classification module constitute a BRepNet++ network, and the BRepNet++ network is trained using a self-built dataset; The encoding module replaces the traditional convolution layer with an improved self-attention mechanism to effectively extract the geometric information of the UV-grid data; the classification module introduces rich topological relationships, deeply fuses geometric features and topological features, and then processes the fusion features to obtain feature recognition results; In the encoding module, multiple consecutive attention encoders are used to process parameterized data such as faces, edges, and vertices in UV-grid format, align the point array information with the attribute feature matrix of each face, and generate fusion data representation with geometric explicitness; In the classification module, the wing surface kernel is constructed by a systemically defined topological operator to capture the topological structure relationship in the CAD model, deeply fuse the geometric features and topological features, and perform hierarchical processing on the outputs of each layer, assign different weights according to the network depth, and fuse the features of multiple layers to fully utilize the hidden feature information in the intermediate layers of the network, and finally output the inference results through the classification layer to realize high-precision feature recognition and segmentation of the CAD model.

[0031] Further, the self-built dataset is derived from multiple industrial manufacturing fields and is annotated based on the set rib position recognition standard and annotation categories.

[0032] Further, in the rib position recognition standard, the rib position is defined as a thin-walled reinforcing structure that strengthens the structure strength and avoids deformation; for each face in the B-Rep data, the annotation categories include rib position, other, and ambiguous, and the annotation category of ambiguous indicates that the face has a rib position similar topological structure but does not actually belong to the rib position.

[0033] To verify the basic performance of the BRepNet++ network proposed in this embodiment, this embodiment verifies it on three open source data sets MFCAD, MFCAD++, and Fusion 360 respectively. MFCAD contains 15,488 CAD models, a total of 16 feature categories; each CAD model contains an average of 22.62 faces; the distribution of each category is relatively balanced, with an average of 21,893 faces per category and a minimum of 4,849 faces; for this data set, the BRepNet++ network achieves an accuracy of more than 99%, indicating that the BRepNet++ network has excellent performance on balanced synthetic data; MFCAD++ contains 57,563 CAD models, a total of 25 categories; each CAD model contains an average of 29.99 faces, and this data set is larger and has more categories; each category has an average of 69,048 faces; for this data set, the BRepNet++ network still maintains a high performance of more than 99%, which verifies the stability of the BRepNet++ network when processing diversified synthetic data; Fusion 360 is a real product CAD data set containing 30,680 CAD models, divided into 8 categories; the average number of faces per CAD model in this data set is 14.74, but the maximum number of faces is 421. At the same time, the category distribution is extremely unbalanced, with an average of 65,759 faces per category, but the maximum category (266,947 faces) and the minimum category (435 faces) have a difference of more than 600 times; on this data set, the BRepNet++ network achieves an accuracy of more than 95%, and significantly outperforms other networks, proving the robustness of the BRepNet++ network when processing unbalanced real data.

[0034] Table 1 Recognition accuracy of each network model on different open source data sets

[0035] In order to better meet the needs of actual industrial applications, this embodiment constructs a self-built data set focusing on plastic part rib position recognition, and performs transfer learning and fine-tuning on this basis. The self-built dataset collects 74 complex plastic product CAD models from real industrial environments, which come from automotive interiors, electronic device housings, etc., representing the complexity of real industrial design; At the same time, clear rib position identification standards are formulated: rib position is defined as a thin-walled reinforcing structure to strengthen the strength and avoid deformation, usually in the form of rib-shaped geometry extending from the base surface in a specific direction; When manually labeling, each face is classified as "rib position", "other" and "ambiguous", where ambiguous is used to represent the face with similar topology structure as rib position but not actually belonging to rib position; The self-built dataset totals 111 CAD models, containing 138,722 faces, of which 16,672 rib position faces account for 12%; Each CAD model contains an average of 1,249.75 faces, which is much higher than the open source dataset, reflecting the complexity of real industrial models; The number of model faces is widely distributed, ranging from a minimum of 64 faces to a maximum of 16,889 faces, showing the diversity of real design; The number distribution of the two categories is uneven, which is consistent with the industrial reality.

[0036] Table 2 Recognition accuracy of each network model on the self-built dataset

[0037] Through the construction of the self-built dataset and the practice of transfer learning, the BRepNet++ network successfully adapts to the feature recognition needs in real industrial scenarios, especially for the key rib feature recognition in mold design, verifying the practicability of the model in professional field applications.

[0038] By applying the method of the application, at least the following beneficial effects are provided: 1) By introducing a winged face kernel, more comprehensive face neighborhood information is provided, significantly improving the feature recognition accuracy of complex geometric shapes; 2) By adding multi-classification modules at different depths of the neural network, additional supervision signals are provided, greatly improving the accuracy of the reasoning stage; 3) By improving the self-attention mechanism, fine-grained geometric information is better captured and integrated, enhancing the understanding ability of complex topological relationships; 4) The model performance is verified on synthetic and real CAD datasets, proving its superior performance and practical application potential in feature recognition tasks.

[0039] Example two: This embodiment shares the same principles as the first embodiment. Building upon the first embodiment, this embodiment provides a computer device, such as a smartphone, tablet computer, laptop computer, desktop computer, rack server, blade server, tower server, or cabinet server (including a standalone server or a server cluster consisting of multiple servers), capable of executing programs. The computer device of this embodiment includes at least, but is not limited to, a memory and a processor that are interconnected via a system bus.

[0040] In this embodiment, the memory (i.e., readable storage medium) includes flash memory, a hard disk, a multimedia card, a card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and programmable read-only memory (PROM). The memory may also be an external storage device of the computer device, such as a plug-in hard disk, a SmartMedia Card (SMC), a Secure Digital (SD) card, a flash memory card, etc. Of course, the memory may also include both the internal storage unit of the computer device and its external storage device. In this embodiment, the memory is generally used to store the operating system and various application software installed on the computer device, such as the program code of the B-Rep feature recognition method in Example 1. Furthermore, the memory may also be used to temporarily store various types of data that have been output or are about to be output.

[0041] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of a computer device. In this embodiment, the processor is used to execute program code stored in a memory or process data to implement the B-Rep feature recognition method of Example 1.

[0042] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, or two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0043] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A B-Rep feature recognition method, characterized in that: include: S1. Obtain the topological features of B-Rep data using the set airfoil kernel; S2. Input the topological features and the geometric features of the B-Rep data into a pre-built classification module to obtain the feature recognition results of the B-Rep data; The topological features include the neighborhood face information, neighborhood edge information, and neighborhood directed edge information of each directed edge in the B-Rep data.

2. The B-Rep feature recognition method according to claim 1, characterized in that: The wing kernel includes basic topological operators n, p, m, e, and f; For the selected directed edge h, the topological operator n obtains the next directed edge of the directed edge h, the topological operator p obtains the previous directed edge of the directed edge h, the topological operator m obtains the paired directed edge of the directed edge h, the topological operator e obtains the corresponding geometric edge of the directed edge h, and the topological operator f obtains the associated face of the directed edge h; The results of obtaining associated faces, geometric edges, and directed edges within the directed edge neighborhood constitute neighborhood face information, neighborhood edge information, and neighborhood directed edge information.

3. The B-Rep feature recognition method according to claim 1, characterized in that: Methods for obtaining geometric features of B-Rep data include: S201, obtaining UV lattice data of B-Rep data, inputting the UV lattice data into a pre-built multi-key self-attention sub-network to obtain processed lattice information; S202 , obtaining a feature matrix of the B-Rep data, and inputting the processed lattice information and the feature matrix into a pre-built data fusion layer for data fusion to obtain geometric features.

4. The B-Rep feature recognition method according to claim 3, characterized in that: The multi-key self-attention subnetwork includes multiple self-attention encoders, and different self-attention encoders process data of different dimensions of UV dot matrix data. After a certain dimension of UV dot matrix data is input into the self-attention encoder, it undergoes linear transformation to obtain query, value and multiple groups of different keys. Each group of keys interacts with the query to generate an independent attention weight distribution. Different attention weight distributions interact with the value to obtain different context vectors. Multiple groups of context vectors are fused through the averaging strategy and output as feature representation.

5. The B-Rep feature recognition method according to claim 3, characterized in that: The method for obtaining geometric features is to align the point lattice information on faces, edges and directed edges with the corresponding feature matrix in the data dimension.

6. The B-Rep feature recognition method according to claim 1, characterized in that: The classification module includes a neural network, multiple auxiliary classifiers, a main classifier, and a weighted fusion layer; the neural network includes an initial layer, multiple intermediate layers, and a final layer. The number of auxiliary classifiers is the same as that of the intermediate layers and they correspond one to one. Multiple auxiliary classifiers collect the features of each intermediate layer respectively and obtain different auxiliary classification results. The main classifier collects the features of the final layer and obtains the main classification result. The weighted fusion layer performs weighted fusion on the main classification result and each auxiliary classification result based on the set weights to obtain the feature recognition result.

7. The B-Rep feature recognition method according to claim 3, characterized in that: The wing core, multi-key self-attention sub-network, and data fusion layer constitute the encoding module. The encoding module and the classification module constitute the BRepNet++ network. The BRepNet++ network is trained using a self-built dataset.

8. The B-Rep feature recognition method according to claim 7, characterized in that: The self-built datasets come from multiple industrial manufacturing fields, and the B-Rep data in the self-built datasets are labeled based on the set rib position recognition standards and annotation categories.

9. The B-Rep feature recognition method according to claim 8, wherein: In the rib identification standard, ribs are defined as thin-walled reinforcement structures that enhance structural strength and prevent deformation. For each surface in the B-Rep data, the annotation categories include rib, other, and fuzzy. The fuzzy annotation category indicates that the surface has a topological structure similar to a rib but does not actually belong to a rib.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the B-Rep feature recognition method according to any one of claims 1 to 9 are implemented.

Images