Plastic part feature grid generation method based on machine learning

Abstract

The invention discloses a plastic part feature grid generation method based on machine learning, and relates to the technical field of grid generation, and the method comprises the steps: importing existing different types of plastic part 3D model data into a Cloud Compare, and obtaining corresponding point cloud data; performing feature extraction and classification identification on the obtained point cloud data to obtain corresponding point cloud implicit geometric features and classification results; performing similarity matching on the obtained implicit geometric features and classification results of the point cloud and data samples in a retrieval database, and outputting a corresponding plastic part feature grid according to a similarity matching result; according to the method, the point cloud data is obtained from the CAD file, the point cloud implicit geometric features and the classification result are output in combination with the deep learning network, searching is carried out in the retrieval database according to the point cloud implicit geometric features and the classification result so as to output the grid, generation of the plastic part grid based on machine learning is achieved, and the efficiency of generation of the plastic part grid is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mesh generation, and in particular to a method for generating characteristic meshes of plastic parts based on machine learning. Background Art

[0002] Plastic part simulation plays an important role in automobile R&D. It provides strong technical support for automobile lightweighting, material selection, structural design, molding process optimization and performance testing. Plastic part simulation is a complex and time-consuming process, which includes multiple steps: geometric preprocessing, mid-surface extraction, mesh generation, mesh optimization, material property assignment, boundary conditions and load application, and final simulation analysis and result evaluation. Among them, mesh generation is the most basic and critical preparatory work. This link occupies a large amount of R&D workload, and the quality of the mesh has a significant impact on the simulation results. Therefore, improving the efficiency of plastic part mesh generation has become an urgent need. For a long time, plastic part mesh generation has mainly relied on manual operation. Although the mid-surface mesh generation function of current mainstream commercial software can realize the preliminary division of the mesh, the generated mesh often cannot fully meet the needs of analysis and calculation, and still requires a lot of manual operation for optimization and adjustment. How to improve the output efficiency of plastic part mesh features is a problem we need to solve. To this end, a plastic part feature mesh generation method based on machine learning is now provided. Summary of the invention

[0003] The object of the present invention is to provide a method for generating feature meshes of plastic parts based on machine learning.

[0004] The object of the present invention can be achieved by the following technical solution: A method for generating a feature grid of a plastic part based on machine learning, comprising the following steps:

[0005] Step S1: Import the existing 3D model data of plastic parts of different categories into CloudCompare to obtain the corresponding point cloud data;

[0006] Step S2: extracting features and classifying the obtained point cloud data to obtain corresponding point cloud implicit geometric features and classification results;

[0007] Step S3: Perform similarity matching on the obtained implicit geometric features and classification results of the point cloud and the data samples in the retrieval database, and output the corresponding plastic part feature grid according to the similarity matching results.

[0008] Furthermore, the plastic part 3D model data is a CAD file, which contains the 3D model of the plastic part, and each CAD file is associated with a corresponding plastic part category and mesh data, wherein the mesh data includes a mesh type and a mesh density;

[0009] After the CAD file is input into CloudCompare, the point density and point quantity of the point cloud data are set in CloudCompare to obtain point cloud data corresponding to different categories of plastic parts.

[0010] Furthermore, the process of performing feature extraction and classification recognition on the obtained point cloud data to obtain the corresponding point cloud implicit geometric features and classification results includes:

[0011] The point cloud data output by CloudCompare is classified and summarized according to the categories of plastic parts to obtain the corresponding point cloud data set;

[0012] The obtained point cloud dataset is divided into a base dataset and an incremental dataset;

[0013] Build a classification network model, set the base class task, train the classification network model with the base class data set, obtain the corresponding base class classification network model, input the test set into the trained base class classification network model, obtain the corresponding point cloud implicit geometric features and classification results, and complete the base class task;

[0014] A few-sample classification incremental learning task is set, and according to the set few-sample classification incremental learning task, sample incremental learning is performed on the set base class classification network model through the incremental data set.

[0015] Furthermore, the classification network model is trained by the base class data set to obtain the corresponding base class classification network model, including:

[0016] The obtained point cloud data is projected from multiple perspectives to obtain point cloud view images from different perspectives;

[0017] Input the multi-view point cloud view image into the visual encoder of the pre-trained model CLIP and output the initial encoding features;

[0018] The initial coding features corresponding to the point cloud view images obtained from different perspectives are spliced, and the spliced ​​initial coding features are input into a two-layer MLP network to obtain the corresponding global image features;

[0019] Input the point cloud data into the 3D network, map the output of the 3D network through a two-layer MLP network, and obtain the corresponding 3D features;

[0020] The obtained 3D features are fused with the global image features to obtain implicit geometric features of the point cloud, and at the same time, text features of the point cloud data are obtained;

[0021] Generate corresponding text prompts according to the plastic part category corresponding to the point cloud data, and extract features of the generated text prompts through the text encoder of the CLIP model to obtain corresponding text features;

[0022] Calculating the similarity between the obtained implicit geometric features of the point cloud and the text features to obtain similarity values ​​between the point cloud data and each plastic part category, and outputting the plastic part category corresponding to the maximum similarity value as a prediction classification result;

[0023] The corresponding loss function is obtained according to the predicted classification result and the actual plastic part category, the base class classification network model is updated according to the obtained loss function, and the above operation is repeated iteratively.

[0024] Furthermore, the process of sample incremental learning of the classification network model includes:

[0025] Setting a few-sample classification incremental learning task, wherein the few-sample classification incremental learning task includes n learning stages, and each learning stage is numbered as i, where i=1, 2, ..., n;

[0026] Set corresponding training sets and test sets for each learning stage;

[0027] According to the learning stages corresponding to the label order, the classification network model is trained in sequence through the training sets corresponding to the learning stages;

[0028] The training set of the learning stage corresponding to label i=1 is input into the trained base class classification network model, and then the training result of the base class classification network model is evaluated by the test set corresponding to the learning stage, thus completing the learning process of the learning stage of label i=1;

[0029] Input the training set corresponding to the learning stage corresponding to label i=2 into the base class classification network model that completes the learning process of the learning stage labeled i=1, evaluate the training results of the base class classification network model through the test sets corresponding to the learning stages labeled i=1 and label i=2, and complete the learning process of the learning stage labeled i=2;

[0030] The training set corresponding to the learning stage corresponding to label i=3 is input into the base class classification network model that completes the learning process of the learning stage labeled i=2, and the training results of the base class classification network model are evaluated through the test sets corresponding to the learning stages labeled i=1, i=2 and i=3, completing the learning process of the learning stage labeled i=3, and so on, until all learning stages are completed.

[0031] Furthermore, the obtained implicit geometric features and classification results of the point cloud are matched with the data samples in the retrieval database for similarity, and the process of outputting the corresponding plastic part feature grid according to the similarity matching results includes:

[0032] Building a retrieval database, and importing sample data of different plastic parts categories into the retrieval database, wherein the sample data includes candidate categories, candidate features, and grid data;

[0033] Input the obtained implicit geometric features of the point cloud and the classification results into the retrieval database, match the classification results with the candidate categories, determine the corresponding retrieval range according to the matching results, and filter the sample data corresponding to the classification results according to the retrieval range;

[0034] The implicit geometric features of the point cloud are then similarly calculated with the candidate features in the screened sample data to obtain the corresponding cosine similarity. The candidate feature corresponding to the maximum value of the obtained cosine similarity is used as the matching result, and the grid data corresponding to the matched candidate feature is output as the plastic part feature grid.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] By obtaining point cloud data from CAD files and combining it with the deep learning network to output the implicit geometric features and classification results of the point cloud, the mesh is output by searching in the retrieval database based on the implicit geometric features and classification results of the point cloud. This realizes a plastic part mesh generation solution based on machine learning and improves the efficiency of plastic part mesh generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0038] Figure 1 It is a schematic diagram of the present invention. DETAILED DESCRIPTION

[0039] like Figure 1 As shown, a method for generating a feature mesh of a plastic part based on machine learning comprises the following steps:

[0040] Step S1: Import the existing 3D model data of plastic parts of different categories into CloudCompare to obtain the corresponding point cloud data;

[0041] Step S2: extracting features and classifying the obtained point cloud data to obtain corresponding point cloud implicit geometric features and classification results;

[0042] Step S3: Perform similarity matching on the obtained implicit geometric features and classification results of the point cloud and the data samples in the retrieval database, and output the corresponding plastic part feature grid according to the similarity matching results.

[0043] It should be further explained that, in the specific implementation process, the plastic part 3D model data is a CAD file, which contains the 3D model of the plastic part, and each CAD file is associated with a corresponding plastic part category. For CAD files of different plastic part categories, corresponding mesh types and mesh densities are set according to requirements to obtain corresponding mesh data;

[0044] After setting the point density and number of points in CloudCompare, the point cloud data corresponding to different categories of plastic parts are output.

[0045] It should be further explained that, in the specific implementation process, the process of extracting features and classifying the obtained point cloud data to obtain the corresponding point cloud implicit geometric features and classification results includes:

[0046] The point cloud data output by CloudCompare is classified and summarized according to the categories of plastic parts to obtain the corresponding point cloud data set;

[0047] The obtained point cloud data set is divided into a base class data set and an incremental data set; it should be further explained that, in the specific implementation process, a data volume threshold is set, if the number of point cloud data in the point cloud data set reaches the data volume threshold, then the point cloud data set is recorded as the base class data set, otherwise, the point cloud data set is recorded as the incremental data set;

[0048] Build a classification network model, set base class tasks and few-sample classification incremental learning tasks for the classification network model, and use the incremental data set to perform sample incremental learning on the classification network model through the set few-sample classification incremental learning tasks;

[0049] The classification network model is trained by the set base class task and the base class data set to obtain the corresponding base class classification network model. The test set is input into the trained base class classification network model to obtain the corresponding point cloud implicit geometric features and classification results, thus completing the base class task.

[0050] It should be further explained that, in the specific implementation process, the classification network model that completes the sample incremental learning task is trained through the base class data set, and the process of obtaining the corresponding base class classification network model includes:

[0051] The obtained point cloud data is projected from multiple perspectives to obtain point cloud view images from different perspectives. The point cloud view images from different perspectives are aggregated to obtain the corresponding global image features. The specific process is as follows:

[0052] The multi-view point cloud view image is input into the visual encoder of the pre-trained model CLIP, and the initial encoding feature is output, denoted as F 2D ;

[0053] The initial coding features corresponding to the point cloud view images obtained from different perspectives are spliced, and the spliced ​​initial coding features are input into the two-layer MLP network to obtain the corresponding global image features, denoted as f d ;

[0054]

[0055] in, Represent the first layer MLP network and the second layer MLP network respectively;

[0056] The point cloud data is input into the 3D network and mapped through the two-layer MLP network to obtain the corresponding 3D features. The specific process includes:

[0057] Input the point cloud data into the 3D network and record the output of the 3D network as f 3D ;

[0058] The output of the 3D network is mapped through a two-layer MLP network to obtain the corresponding 3D features. The 3D features obtained through the MLP network mapping are recorded as f p ;

[0059] in,

[0060] in Represent the first layer MLP network and the second layer MLP network respectively;

[0061] The obtained 3D features are fused with the global image features to obtain the implicit geometric features of the point cloud, and the text features of the point cloud data are obtained at the same time. The obtained implicit geometric features of the point cloud are recorded as f g ;

[0062] in,

[0063] Among them, max() means finding the maximum value, and avg() means finding the average value;

[0064] Generate corresponding text prompts according to the plastic part category corresponding to the point cloud data, and extract features of the generated text prompts through the text encoder of the CLIP model to obtain the corresponding text features, which are recorded as F t ;

[0065] The obtained implicit geometric features of the point cloud are calculated similarly with the text features to obtain similarity values ​​between the point cloud data and each plastic part category, and the plastic part category corresponding to the maximum similarity value is output as a predicted classification result, and the similarity value is recorded as L;

[0066] Where, L = cos(fg , F t );

[0067] The maximum value among the obtained similarity values ​​is taken as the output of the plastic part category, and the output result is recorded as p;

[0068] Where p = softmax([L]);

[0069] The corresponding loss function is obtained according to the predicted classification result and the actual plastic part category, and the base class classification network model is updated according to the obtained loss function. The above operation is repeated iteratively, where the number of iterations is 100 times;

[0070] The loss function is recorded as Loss cls =L ce (p, y);

[0071] Among them, L ce () represents the cross entropy loss function, and y is the true category corresponding to the point cloud data.

[0072] It should be further explained that, in the specific implementation process, the process of sample incremental learning of the classification network model includes:

[0073] Setting a few-sample classification incremental learning task, wherein the few-sample classification incremental learning task includes n learning stages, and each learning stage is numbered as i, where i=1, 2, ..., n;

[0074] Set the corresponding training set and test set for each learning stage, and the training set and test set corresponding to the learning stage labeled i are respectively recorded as and

[0075] According to the learning stages corresponding to the label order, the classification network model is trained in sequence through the training sets corresponding to the learning stages;

[0076] The training set of the learning stage corresponding to label i=1 is input into the trained base class classification network model, and then the training result of the base class classification network model is evaluated by the test set corresponding to the learning stage, thus completing the learning process of the learning stage of label i=1;

[0077] Input the training set corresponding to the learning stage corresponding to label i=2 into the classification network model that completes the learning process of the learning stage labeled i=1, evaluate the training results of the classification network model through the test sets corresponding to the learning stages labeled i=1 and label i=2, and complete the learning process of the learning stage labeled i=2;

[0078] The training set corresponding to the learning stage corresponding to label i=3 is input into the classification network model that completes the learning process of the learning stage labeled i=2, and the training results of the classification network model are evaluated through the test sets corresponding to the learning stages labeled i=1, i=2 and i=3, completing the learning process of the learning stage labeled i=3, and so on, until all learning stages are completed.

[0079] It should be further explained that, in the specific implementation process, the obtained implicit geometric features and classification results of the point cloud are matched with the data samples in the retrieval database for similarity, and the process of outputting the corresponding plastic part feature grid according to the similarity matching results includes:

[0080] Constructing a retrieval database, and importing sample data of different plastic parts categories into the retrieval database, wherein the sample data includes candidate categories and candidate features;

[0081] Input the obtained implicit geometric features of the point cloud and the classification results into the retrieval database, match the classification results with the candidate categories, determine the corresponding retrieval range according to the matching results, and filter the sample data corresponding to the classification results according to the retrieval range;

[0082] The implicit geometric features of the point cloud are then similarly calculated with the candidate features in the screened sample data to obtain the corresponding cosine similarity. The candidate feature corresponding to the maximum value of the obtained cosine similarity is used as the matching result, and the grid data corresponding to the matched candidate feature is output as the plastic part feature grid.

[0083] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments with equivalent changes without departing from the scope of the technical solution of the present invention. However, any modification or equivalent replacement of the above embodiments made according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A method for generating feature mesh of plastic parts based on machine learning, characterized in that: The following steps are involved: Step S1: Import the existing 3D model data of plastic parts of different categories into CloudCompare to obtain the corresponding point cloud data; Step S2: extracting features and classifying the obtained point cloud data to obtain corresponding point cloud implicit geometric features and classification results; Step S3: Perform similarity matching on the obtained implicit geometric features and classification results of the point cloud and the data samples in the retrieval database, and output the corresponding plastic part feature grid according to the similarity matching results.

2. The method for generating feature mesh of plastic parts based on machine learning according to claim 1, characterized in that: The plastic part 3D model data is a CAD file, which contains a 3D model of the plastic part, and each CAD file is associated with a corresponding plastic part category and mesh data, wherein the mesh data includes a mesh type and a mesh density; After the CAD file is input into CloudCompare, the point density and point quantity of the point cloud data are set in CloudCompare to obtain point cloud data corresponding to different categories of plastic parts.

3. The method for generating feature mesh of plastic parts based on machine learning according to claim 2, characterized in that: The process of extracting features and classifying the obtained point cloud data to obtain the corresponding point cloud implicit geometric features and classification results includes: The point cloud data output by CloudCompare is classified and summarized according to the categories of plastic parts to obtain the corresponding point cloud data set; The obtained point cloud dataset is divided into a base dataset and an incremental dataset; Build a classification network model, set the base class task, train the classification network model with the base class data set, obtain the corresponding base class classification network model, input the test set into the trained base class classification network model, obtain the corresponding point cloud implicit geometric features and classification results, and complete the base class task; A few-sample classification incremental learning task is set, and according to the set few-sample classification incremental learning task, sample incremental learning is performed on the set base class classification network model through the incremental data set.

4. The method for generating feature mesh of plastic parts based on machine learning according to claim 3, characterized in that: The process of training the classification network model through the base class data set and obtaining the corresponding base class classification network model includes: The obtained point cloud data is projected from multiple perspectives to obtain point cloud view images from different perspectives; Input the multi-view point cloud view image into the visual encoder of the pre-trained model CLIP and output the initial encoding features; The initial coding features corresponding to the point cloud view images obtained from different perspectives are spliced, and the spliced ​​initial coding features are input into a two-layer MLP network to obtain the corresponding global image features; Input the point cloud data into the 3D network, map the output of the 3D network through a two-layer MLP network, and obtain the corresponding 3D features; The obtained 3D features are fused with the global image features to obtain implicit geometric features of the point cloud, and at the same time, text features of the point cloud data are obtained; Generate corresponding text prompts according to the plastic part category corresponding to the point cloud data, and extract features of the generated text prompts through the text encoder of the CLIP model to obtain corresponding text features; Calculating the similarity between the obtained implicit geometric features of the point cloud and the text features to obtain similarity values ​​between the point cloud data and each plastic part category, and outputting the plastic part category corresponding to the maximum similarity value as a prediction classification result; The corresponding loss function is obtained according to the predicted classification result and the actual plastic part category, the base class classification network model is updated according to the obtained loss function, and the above operation is repeated iteratively.

5. The method for generating feature mesh of plastic parts based on machine learning according to claim 4, characterized in that: The process of sample incremental learning of the classification network model includes: Setting a few-sample classification incremental learning task, wherein the few-sample classification incremental learning task includes n learning stages, and each learning stage is numbered as i, where i=1, 2, ..., n; Set corresponding training sets and test sets for each learning stage; According to the learning stages corresponding to the label order, the classification network model is trained in sequence through the training sets corresponding to the learning stages; The training set of the learning stage corresponding to label i=1 is input into the trained base class classification network model, and then the training result of the base class classification network model is evaluated by the test set corresponding to the learning stage, thus completing the learning process of the learning stage of label i=1; Input the training set corresponding to the learning stage corresponding to label i=2 into the base class classification network model that completes the learning process of the learning stage labeled i=1, evaluate the training results of the base class classification network model through the test sets corresponding to the learning stages labeled i=1 and label i=2, and complete the learning process of the learning stage labeled i=2; The training set corresponding to the learning stage corresponding to label i=3 is input into the base class classification network model that completes the learning process of the learning stage labeled i=2, and the training results of the base class classification network model are evaluated through the test sets corresponding to the learning stages labeled i=1, i=2 and i=3, completing the learning process of the learning stage labeled i=3, and so on, until all learning stages are completed.

6. The method for generating feature mesh of plastic parts based on machine learning according to claim 5, characterized in that: The process of performing similarity matching between the obtained implicit geometric features and classification results of the point cloud and the data samples in the retrieval database and outputting the corresponding plastic part feature mesh according to the similarity matching results includes: Building a retrieval database, and importing sample data of different plastic parts categories into the retrieval database, wherein the sample data includes candidate categories, candidate features, and grid data; Input the obtained implicit geometric features of the point cloud and the classification results into the retrieval database, match the classification results with the candidate categories, determine the corresponding retrieval range according to the matching results, and filter the sample data corresponding to the classification results according to the retrieval range; The implicit geometric features of the point cloud are then similarly calculated with the candidate features in the screened sample data to obtain the corresponding cosine similarity. The candidate feature corresponding to the maximum value of the obtained cosine similarity is used as the matching result, and the grid data corresponding to the matched candidate feature is output as the plastic part feature grid.

Images