Method and system for recognizing machining features of a cad model based on multi-modal fusion

Abstract

The method comprises the following steps: extracting the topological features of the CAD model, including face features and edge features; constructing a face adjacency graph structure according to the face features and the edge features; wherein the nodes of the graph structure represent the faces in the CAD model, and the edges represent the adjacency relationship between the faces; performing feature propagation on the topological information of the CAD model based on a graph neural network to obtain topological features; obtaining multi-view images according to the CAD model; extracting visual features of the multi-view images; performing region feature extraction on the visual features according to the region correspondence relationship of each face in the CAD model in the multi-view, and fusing the corresponding topological features to obtain multi-modal fusion features; and outputting a feature recognition result according to the multi-modal fusion features.

Description

Technical Field

[0001] This application relates to the fields of computer-aided design and intelligent manufacturing technology, specifically to a method and system for identifying machining features of CAD models based on multimodal fusion. Background Technology

[0002] Machining feature recognition is a crucial link between computer-aided design and computer-aided process planning. By automatically identifying machining features such as holes, slots, and bosses from CAD models, it can provide fundamental data support for downstream tasks such as CNC machining, process planning, and manufacturing analysis.

[0003] Existing methods for identifying processing features mainly include geometric rule-based methods, graph matching-based methods, and deep learning-based methods. Geometric rule-based methods achieve feature recognition through predefined geometric constraints; graph matching-based methods typically construct the boundary representation (B-rep) model as a face adjacency graph and perform feature matching based on topological relationships. These methods achieve high recognition accuracy in scenarios with simple, independent features, but when the model contains complex overlapping features, shared boundaries, or non-standard structures, problems such as rule failure and topological ambiguity easily arise, leading to a decrease in recognition accuracy.

[0004] With the development of deep learning technology, researchers have begun to use various data modalities such as point clouds, voxels, meshes, multi-view images, and B-rep for feature recognition. Among them, the multi-view method can use the contour and appearance information in two-dimensional images for feature learning, but it lacks three-dimensional topological constraints and is difficult to accurately represent the connection relationships in complex parts. Although the B-rep-based method can preserve accurate geometric and topological information, most existing methods rely only on graph neural networks to encode the topological structure, which can easily lead to semantic ambiguity in scenarios with overlapping complex features, making it difficult to achieve high-precision feature instance segmentation. Summary of the Invention

[0005] To overcome the problems existing in related technologies, this specification provides a method and system for identifying machining features of CAD models based on multimodal fusion. This method can effectively integrate topological and visual information in CAD models, thereby improving the recognition accuracy and instance differentiation ability of machining features in complex industrial parts.

[0006] The first aspect is a CAD model processing feature recognition method based on multimodal fusion, which includes the following steps:

[0007] S1. Extract the topological features of the CAD model, including face features and edge features;

[0008] S2. Construct a face adjacency graph structure based on face and edge features; where nodes in the graph structure represent faces in the CAD model, and edges represent the adjacency relationships between faces; use a graph neural network to transfer topological information from the CAD model to obtain topological features;

[0009] S3. Obtain multi-view images from the CAD model; extract the visual features of the multi-view images;

[0010] S4. Based on the region correspondence of each face in the CAD model in the multi-view, extract the region features of the visual features and fuse them with the corresponding topological features to obtain the multimodal fusion features.

[0011] S5. Output the feature recognition results based on the multimodal fusion features.

[0012] Preferably, S1 includes:

[0013] S11. Use a geometric modeling kernel to parse the STEP format file of the CAD model to obtain the geometric and topological information in the CAD model;

[0014] S12. Extract surface attributes, edge attributes, and mesh attributes based on topological information;

[0015] S13. Encode the edge attributes to generate edge features; encode the face attributes and grid attributes separately, and concatenate the encoded features to obtain face features.

[0016] Preferably, in S1, the surface attributes include the geometric type, area, centroid coordinates, and NURBS rationality of the surface; the edge attributes include the edge length, whether the edge is a loop, the concavity and convexity of the edge, and the B-spline rationality of the edge; the mesh attributes include the information after surface discretization sampling, the coordinate information of the obtained sampling points, the normal vector information, and the marking information inside and outside the surface.

[0017] Preferably, S4 includes:

[0018] S41. Pre-establish the region correspondence between the surfaces in the CAD model and the regions in the multi-view image;

[0019] S42. Based on the region correspondence, extract the region features of the corresponding regions in the visual features to obtain the accurate region visual features of the corresponding surfaces.

[0020] S43. For the precise regional visual features of the same face in multiple views, based on the visibility information of the face in the corresponding views, the precise regional visual features in multiple views are aggregated to generate the visual semantic features of the corresponding face.

[0021] S44. The visual semantic features are concatenated with the corresponding topological features, and the concatenated feature dimensions are mapped through a feature mapping network to obtain the corresponding multimodal fusion features.

[0022] As a preferred option, the aggregation process in S43 is an averaging operation; for surfaces that are not visible in all views, learnable feature vectors are used to initialize and complete the corresponding visual features.

[0023] As a preferred embodiment, S44 also includes:

[0024] Graph neural networks are used to perform information interaction on multimodal fusion features;

[0025] For the multimodal fusion features after transmission, summation pooling is used to aggregate the node information of the whole image. The aggregated features are then mapped to the global latent space through a linear projection layer. The global features are then concatenated with the local features to obtain composite features. The composite features are used to output the feature recognition results.

[0026] Preferably, in S3, the multi-view image is generated by projecting the CAD model under multiple preset perspectives. The preset perspectives include the front view, top view, left view, right view, rear view, bottom view, and eight predefined isometric perspectives.

[0027] Secondly, a CAD model processing feature recognition system based on multimodal fusion includes:

[0028] The topology feature extraction module is used to extract the topology features of the CAD model, including face features and edge features;

[0029] The graph structure construction module is used to construct a face adjacency graph structure based on face and edge features; where nodes in the graph structure represent faces in the CAD model, and edges represent the adjacency relationships between faces; and feature transfer of topological information of the CAD model is performed based on a graph neural network.

[0030] The visual feature extraction module is used to obtain multi-view images from CAD models and extract the visual features of the multi-view images.

[0031] The multimodal fusion module is used to extract regional features from visual features based on the regional correspondence of each face in the CAD model in multiple views, and then fuse them with the corresponding topological features to obtain multimodal fusion features.

[0032] The recognition output module is used to output the corresponding processing feature classification results and feature instance segmentation results based on the multimodal fusion features;

[0033] The CAD model machining feature recognition system based on multimodal fusion is used to implement the CAD model machining feature recognition method and steps based on multimodal fusion as described in the first aspect.

[0034] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon, which, when run on a processor, performs the steps of the CAD model machining feature recognition method based on multimodal fusion as described in the first aspect.

[0035] The technical solutions provided in the embodiments of this specification may have the following beneficial effects:

[0036] This specification's embodiments construct a face adjacency graph structure corresponding to the CAD model and combine it with multi-view visual information to achieve joint representation of topological and visual information, thereby effectively improving the recognition accuracy of machining features in complex industrial parts. By establishing the correspondence between faces in the CAD model and multi-view regions, visual and topological features are accurately fused, effectively alleviating the semantic ambiguity caused by relying solely on topological structures in complex feature overlap scenarios and improving the accuracy of machining feature instance segmentation. By jointly outputting machining feature classification results and feature instance segmentation results, synchronous recognition of machining feature categories and corresponding face sets can be achieved, thereby improving the completeness and practicality of machining feature recognition results. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the information flow processing flow for CAD model processing feature recognition based on multimodal fusion according to Embodiment 1 of the present invention;

[0038] Figure 2 This is a schematic diagram of a preset multi-view CAD model view of Embodiment 1 of the present invention;

[0039] Figure 3 This is a schematic diagram illustrating the processing feature recognition principle of Embodiment 1 of the present invention;

[0040] Figure 4 This is a comparison example of the processing feature recognition effect of a single CAD model in Embodiment 1 of the present invention;

[0041] Figure 5 This is an example diagram of the processing feature classification in Embodiment 1 of the present invention. Detailed Implementation

[0042] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this specification as detailed in the appended claims.

[0043] The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of this specification. The singular forms “a,” “and,” and “the” as used in this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0044] It should be understood that although the terms first, second, third, etc., may be used in this specification to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this specification, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0045] The embodiments described in this specification will now be described in detail.

[0046] Example 1:

[0047] like Figure 1 As shown, Figure 1 This specification illustrates an information flow processing diagram of CAD model machining feature recognition based on multimodal fusion according to an exemplary embodiment, including the following steps:

[0048] S1. Extract face attributes, edge attributes, and mesh attributes from the STEP format file of the CAD model, and perform feature encoding on the face attributes, edge attributes, and mesh attributes to obtain the corresponding face features and edge features; specifically including:

[0049] S11. Use a geometric modeling kernel to parse the STEP format file to obtain geometric and topological information from the CAD model.

[0050] S12. Extract the corresponding face attributes, edge attributes, and mesh attributes from the CAD model. Among them, face attributes include at least one of the following: face geometry type, area, centroid coordinates, and NURBS rational properties; edge attributes include at least one of the following: edge length, whether the edge is a cycle, edge concavity / convexity, and edge B-spline rationality; mesh attributes include at least one of the following: coordinates of the sampled points after discretization, normal vector, and in-plane / out-of-plane marking information.

[0051] S13. Normalize and encode the extracted face, edge, and mesh attributes to generate corresponding face and edge features. A two-layer MLP encoding is used for both face and edge attributes, while a three-layer 2D convolutional neural network is used for mesh attributes. The face and mesh encoded features are then concatenated to obtain the face features. Edge features are obtained directly after edge attribute encoding. ,in Indicates the first One side, Indicates the relationship with the first The first face adjacent to the first face Each side.

[0052] S2. Construct the corresponding face adjacency graph structure data based on face features and edge features. In the graph structure, the nodes represent faces in the CAD model, and the edges represent the adjacency relationships between faces. Based on the graph neural network, the topological information of the CAD model is transferred to obtain the topological features, that is, the feature information contained in each node after feature transfer.

[0053] Specifically, for each face in the CAD model, a corresponding graph node is constructed; for two faces with a common edge, a connecting edge is established between the corresponding nodes, thus forming the face adjacency graph structure corresponding to the CAD model.

[0054] S3. Obtain the corresponding multi-view images based on the CAD model, and extract the corresponding visual features based on the convolutional neural network;

[0055] Specifically, a convolutional neural network is used to extract visual features from the generated multi-view images to obtain corresponding visual feature maps. Preferably, the convolutional neural network includes a residual network structure and further includes a multi-scale feature fusion structure to enhance the visual semantic expression capability at different scales.

[0056] The convolutional neural network architecture consists of two parts. The first part is the first three layers of the ResNet18 residual network, which uses and freezes its pre-trained parameters obtained on the large-scale image dataset ImageNet. The second part is a custom multi-scale feature fusion network.

[0057]

[0058]

[0059] in This indicates that ResNet18 is for the th The first face The layer's output, before upsampling, uses The convolution operation makes and The dimensions are the same. Finally, through multi-scale feature fusion, the result is obtained. ,in Indicates the total number of views. These represent the original dimensions of the view: height and width, both 512. Indicates the first Visual features extracted from each facet.

[0060] In this embodiment, the S1 and S3 data originate from a pre-constructed multimodal machining feature dataset. This dataset contains 34,049 CAD models. Each CAD model sample includes a corresponding STEP format file, multi-view images, multimodal alignment information, and machining feature annotation information. Examples of some CAD models and all machining feature categories are shown below. Figure 5 As shown. Preferably, the training set accounts for 70%, the validation set accounts for 15%, and the test set accounts for 15%.

[0061] The multi-view images and multimodal alignment information are constructed as follows: They are generated using uniformly preset viewpoint parameters to ensure consistent geometric representation across different samples. For example, CAD modeling software combined with scripting tools is used to automate off-screen rendering of each CAD model. Figure 2 As shown in the left image, the preset viewpoints include the front view, top view, left view, right view, rear view, bottom view, and eight predefined isometric viewpoints. To obtain multimodal alignment information, such as... Figure 2 As shown in the right-hand image, in addition to generating a grayscale view for training, a pure color segmentation image was also generated. A unique color value was generated for each face based on its label value, thus obtaining the mapping relationship between faces in the CAD model and the multi-view image regions. The generation methods for different face color values ​​are as follows:

[0062]

[0063]

[0064]

[0065]

[0066] in For predefined large prime numbers, Total capacity of the color space Excluding pure black and pure white, the faces are rendered and colored based on the values ​​of the three RGB channels to obtain a color segmentation map. After obtaining the color segmentation map, the mapping relationship between the current multi-view image region and the faces in the CAD model can be deduced using the above calculation formula, thus allowing the calculation of the first... The face on the 1st The corresponding region position in each view is denoted as . This includes the coordinates of the top-left and bottom-right corners of the corresponding region, and calculates the first... The face on the 1st Visibility in a view :

[0067]

[0068] in This indicates the number of visible pixels on that surface. This indicates the total number of pixels in the view.

[0069] S4. Based on the region correspondence of each face in the CAD model in multiple views, extract region features from the visual features and fuse them with the corresponding topological features to obtain multimodal fusion features, thereby achieving joint representation of visual and topological information and transferring the fusion features; specifically including:

[0070] S41. Establish a mapping relationship between the faces in the CAD model and the multi-view image areas in advance, so as to obtain the corresponding area position of each face in different views.

[0071] S42. Based on the region correspondence, region feature extraction is performed on the corresponding regions in the visual features to obtain the accurate region visual features of the corresponding surfaces.

[0072] S43. For the precise regional visual features of the same surface in multiple views, based on the visibility information of the surface in the corresponding views... The system aggregates the visual features of precise regions in multiple views to generate the visual semantic features of the corresponding surfaces.

[0073] Preferably, the aggregation process is an averaging operation. For faces that are not visible in all views, learnable feature vectors are used to initialize and complete the corresponding visual features.

[0074] S44. Subsequently, the obtained visual semantic features are concatenated with the corresponding topological features, and the dimensions of the fused features are mapped through a feature mapping network to obtain the corresponding multimodal fused features. A graph neural network is then used to further interact with the fused multimodal features to enhance the contextual semantic association between different features.

[0075] like Figure 3 As shown, in this embodiment, based on the pre-obtained area boundary information... Using ROIAlign operations to achieve from Extracting precise region features from visual features :

[0076]

[0077] In obtaining Then, a two-layer MLP is used to flatten it to obtain accurate region visual encoding. , indicating the first The first face Precise visual encoding of the region of each view.

[0078] Based on the pre-obtained surface visibility information, the visual encoding of the precise region is aggregated:

[0079]

[0080]

[0081] The final aggregation yields the first Precise visual features of each surface region This is the visibility matrix. For faces that are not visible in all views, learnable feature vectors are used to initialize and complete the corresponding visual features.

[0082] For obtaining the first Precise regional visual features of each facet and topological features First, a concatenation operation is performed, and then a two-layer MLP is used to map its features back to 128 dimensions to obtain the multimodal fusion features of each facet:

[0083]

[0084] In this embodiment, for the fused features after transmission, summation pooling is used to aggregate the node information of the entire image. The aggregated features are mapped to the global latent space through a linear projection layer, and the global features are concatenated with the local features to obtain composite features. The feature instance segmentation result is output based on the composite features.

[0085] In S2 and S4, the feature transfer method of the graph neural network used is defined as follows:

[0086]

[0087] in express The first in An aggregation operation, The operations in the code include summation and finding the maximum value. Indicates the relationship between in the graph structure and All nodes that are adjacent to a given node. Representing node characteristics, Represents a node With nodes The edge characteristics of the connecting edges, This indicates a splicing operation. This indicates that the features obtained from the aggregation operation are concatenated.

[0088] S5. Based on the transmitted multimodal features, output the feature recognition results;

[0089] Specifically, the corresponding processing feature classification result (semantic classification output) is output through the classification output head, and the corresponding feature instance segmentation result is output through the instance segmentation output head.

[0090] The semantic classification output is a softmax classifier that maps composite features to... The probability matrix represents all the CAD models. Each side belongs to The confidence level of the processing feature is constrained using cross-entropy loss:

[0091]

[0092] in , For category tags, To predict probabilities.

[0093] The instance segmentation output is processed by two independent two-layer MLPs to operate on the composite features, and a connection score matrix is ​​obtained by matrix multiplication to predict the connection relationships between surfaces. The loss function is binary cross-entropy loss.

[0094]

[0095] in For the adjacency matrix labels, This is the connection score matrix for the predicted output.

[0096] To balance the convergence speed across different tasks and prevent the model from getting stuck in local optima during the early stages of training, a dynamic weight warm-up mechanism is introduced. The total loss function is defined as:

[0097]

[0098] in It's about training rounds. Dynamic weighting function:

[0099]

[0100] in Indicates the start of the warm-up cycle. This indicates the round in which the preheating process has ended.

[0101] After training in this step, a multimodal fusion CAD model for processing feature recognition with fixed parameters can be obtained. Furthermore, in this embodiment, after training the multimodal semantic fusion model, the performance of processing feature recognition can be evaluated during the inference phase. Specifically, the obtained processing feature recognition results are quantitatively analyzed and evaluated on a pre-divided test dataset.

[0102] In the model training process of this embodiment, a preset training configuration is used to ensure model convergence stability and training efficiency. This training configuration includes batch size, number of training epochs, optimizer type, learning rate scheduling strategy, and gradient pruning parameters.

[0103] Furthermore, to quantitatively evaluate the obtained processing feature recognition results, this embodiment introduces multiple evaluation metrics for the command layer and geometry layer, including classification accuracy (Acc), mean intersection-over-union ratio (mIoU), segmentation accuracy (Acc), mAP50, localization F-1 (F-1), and mean coverage (mCov), which are defined as follows:

[0104]

[0105]

[0106] in This indicates the number of faces that were correctly predicted in the classification. This indicates the total number of faces. Indicates the label is the first The face set of a class Indicates the prediction is the first The set of faces of a class.

[0107]

[0108]

[0109]

[0110]

[0111]

[0112]

[0113] in, Represents the collection of all instances in the label. Represents the set of all predicted instances. This represents an instance whose IoU with the label instance is equal to 1 and which is correctly classified.

[0114] The method described in this application is compared with other deep learning-based methods.

[0115] Table 1 Comparison of processing feature recognition performance of different methods on the main dataset.

[0116]

[0117] Table 2 Comparison of processing feature recognition performance of different methods across datasets

[0118]

[0119] Experimental results show that the multimodal fusion-based processing feature recognition method proposed in this application achieves higher instance segmentation accuracy in complex overlapping feature scenarios compared to methods relying solely on topological or visual information. This is because this application establishes a correspondence between surfaces and multi-view regions in the CAD model, achieving precise fusion between visual and topological semantic information, thereby effectively alleviating semantic ambiguity issues under complex topological structures.

[0120] Furthermore, the embodiments of this application enhance the contextual association ability between different features by further propagating the fused multimodal features, thereby improving the stability and accuracy of the processed feature classification and instance segmentation results.

[0121] like Figure 4 As shown in Table 3, in order to verify the processing feature recognition capability of the method proposed in this application embodiment in complex industrial parts scenarios, 3D CAD models of different complexities were selected to compare the single-model processing feature recognition accuracy.

[0122] Table 3. Classification and segmentation accuracy of different methods for each 3D CAD model.

[0123]

[0124] Experimental results show that as the complexity of CAD models increases, the recognition performance of existing processing feature recognition methods all decline to varying degrees. In contrast, the embodiments of this application, by fusing multi-view visual information and B-Rep topological information, can utilize the contour boundaries and local geometric semantic information in the visual modality to more accurately distinguish instances of the same category of processing features, thereby effectively improving the accuracy of processing feature recognition and instance segmentation capability in complex scenes.

[0125] Example 2:

[0126] A CAD model machining feature recognition system based on multimodal fusion includes:

[0127] The topology feature extraction module is used to extract the topology features of the CAD model, including face features and edge features;

[0128] The graph structure construction module is used to construct graph structure data based on face and edge features; where nodes in the graph structure represent faces in the CAD model, and edges represent the adjacency relationships between faces; and feature transfer of topological information of the CAD model is performed based on graph neural networks.

[0129] The visual feature extraction module is used to obtain multi-view images from CAD models and extract the visual features of the multi-view images.

[0130] The multimodal fusion module is used to extract regional features from visual features based on the regional correspondence of each face in the CAD model in multiple views, and then fuse them with the corresponding topological features to obtain multimodal fusion features.

[0131] The recognition output module is used to output the corresponding processing feature classification results and feature instance segmentation results based on the multimodal fusion features.

[0132] Example 3:

[0133] A computer-readable storage medium is provided, on which a computer program is stored, which, when run on a processor, executes the steps of a CAD model machining feature recognition method based on multimodal fusion.

[0134] For system embodiments, since they basically correspond to method embodiments, relevant details can be found in the descriptions of the method embodiments. The apparatus embodiments described above are merely illustrative; the modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, meaning they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of the solution described in this specification, depending on actual needs. Those skilled in the art can understand and implement this without any inventive effort.

[0135] It should be understood that this specification is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this specification is limited only by the appended claims.

[0136] The above are merely preferred embodiments of this specification and are not intended to limit this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification shall be included within the scope of protection of this specification.

Claims

1. A method for identifying machining features of CAD models based on multimodal fusion, characterized in that, Includes the following steps: S1. Extract the topological features of the CAD model, including face features and edge features; S2. Construct a face adjacency graph structure based on face and edge features; where nodes in the graph structure represent faces in the CAD model, and edges represent the adjacency relationships between faces; use a graph neural network to transfer topological information from the CAD model to obtain topological features; S3. Obtain multi-view images based on the CAD model; extract visual features from the multi-view images; the convolutional neural network used for visual feature extraction consists of two parts: the first part is the first three layers of the residual network ResNet18, using and freezing its pre-trained parameters obtained on the large-scale image dataset ImageNet; the second part is a custom multi-scale feature fusion network. in, This indicates that ResNet18 is for the th The first face The layer's output, before upsampling, uses The convolution operation makes the first Layer output and Layer output The dimensions are the same; after multi-scale feature fusion, the first dimension is obtained. Visual features extracted from individual faces , Indicates the total number of views. These represent the original dimensions of the view: height and width, respectively. S4. Based on the region correspondence of each face in the CAD model in multiple views, extract region features from the visual features and fuse them with the corresponding topological features to obtain multimodal fusion features; including: S41. Pre-establish the region correspondence between the surfaces in the CAD model and the regions in the multi-view image; S42. Based on the region correspondence, extract the region features of the corresponding regions in the visual features to obtain the accurate region visual features of the corresponding surfaces. S43. For the precise regional visual features of the same face in multiple views, based on the visibility information of the face in the corresponding views, the precise regional visual features in multiple views are aggregated to generate the visual semantic features of the corresponding face. S44. The visual semantic features are concatenated with the corresponding topological features, and the concatenated feature dimensions are mapped through a feature mapping network to obtain the corresponding multimodal fusion features. S5. Output the feature recognition results based on the multimodal fusion features.

2. The CAD model machining feature recognition method based on multimodal fusion according to claim 1, characterized in that, S1 includes: S11. Use a geometric modeling kernel to parse the STEP format file of the CAD model to obtain the geometric and topological information in the CAD model; S12. Extract surface attributes, edge attributes, and mesh attributes based on topological information; S13. Encode the edge attributes to generate edge features; The surface attributes and grid attributes are encoded separately, and the encoded features are concatenated to obtain the surface features.

3. The CAD model machining feature recognition method based on multimodal fusion according to claim 2, characterized in that, In S1, surface attributes include the geometric type, area, centroid coordinates, and NURBS rationality of the surface; edge attributes include edge length, whether the edge is a cycle, edge concavity / convexity, and edge B-spline rationality; mesh attributes include information after surface discretization, coordinate information of the obtained sampling points, normal vector information, and surface / inside / outside marking information.

4. The CAD model machining feature recognition method based on multimodal fusion according to claim 1, characterized in that, In S43, the aggregation process is an averaging operation; for faces that are not visible in all views, learnable feature vectors are used to initialize and complete the corresponding visual features.

5. The CAD model machining feature recognition method based on multimodal fusion according to claim 1, characterized in that, S44 also includes: Graph neural networks are used to perform information interaction on multimodal fusion features; For the multimodal fusion features after transmission, summation pooling is used to aggregate the node information of the whole image. The aggregated features are then mapped to the global latent space through a linear projection layer. The global features are then concatenated with the local features to obtain composite features. The composite features are used to output the feature recognition results.

6. The CAD model machining feature recognition method based on multimodal fusion according to claim 1, characterized in that, In S3, multi-view images are generated by projecting the CAD model under multiple preset perspectives. The preset perspectives include the front view, top view, left view, right view, rear view, bottom view, and eight predefined isometric perspectives.

7. A CAD model machining feature recognition system based on multimodal fusion, characterized in that, include: The topology feature extraction module is used to extract the topology features of the CAD model, including face features and edge features; The graph structure construction module is used to construct a face adjacency graph structure based on face and edge features; where nodes in the graph structure represent faces in the CAD model, and edges represent the adjacency relationships between faces; and feature transfer of topological information of the CAD model is performed based on a graph neural network. The visual feature extraction module is used to obtain multi-view images from CAD models and extract the visual features of the multi-view images. The multimodal fusion module is used to extract regional features from visual features based on the regional correspondence of each face in the CAD model in multiple views, and then fuse them with the corresponding topological features to obtain multimodal fusion features. The recognition output module is used to output the corresponding processing feature classification results and feature instance segmentation results based on the multimodal fusion features; The CAD model machining feature recognition system based on multimodal fusion is used to implement the CAD model machining feature recognition method and steps based on multimodal fusion as described in claim 1.

8. A computer-readable storage medium, characterized in that, It stores a computer program, which executes the steps of the CAD model machining feature recognition method based on multimodal fusion as described in claim 1 when the program runs on the processor.

Images