Multi-information fusion processing feature recognition method and device based on graph neural network

Abstract

The invention discloses a multi-information fusion processing feature recognition method and device based on a graph neural network, and the method comprises the steps: representing a three-dimensional model through an adjacent graph, constructing a connection relation matrix between surfaces, carrying out the sampling and coding of the surfaces and edges of the three-dimensional model, and enabling the surfaces and edges to serve as the corresponding feature tensors of the nodes and edges of the adjacent graph, constructing an attribute adjacency graph data set of the part; constructing a graph neural network, a processing feature semantic classifier and a bottom surface classifier, and training to enable the network to perform processing feature semantic classification and bottom surface classification on each surface; and in combination with the output results of the two classifiers and the connection relation matrix, displaying the identified feature instances by using a depth map library, realizing the identification of the processing feature instances of the part, and obtaining the surfaces forming the feature instances. According to the method, the problems of realizing machining feature identification and obtaining the surface forming the machining feature instance in the part with the complex machining feature are solved.

Description

Technical Field

[0001] The present invention relates to a method for processing feature recognition, and in particular to a method and device for processing feature recognition based on multi-information fusion of graph neural network. Background Art

[0002] With the rapid development of the economy and the intensification of market competition, various advanced design technologies and advanced manufacturing technologies have been popularized and applied in the process of product development and production. Among them, the popularization of 3D CAD and CAPP systems has promoted the efficiency of product design and production in enterprises. The design department of the enterprise uses 3D design software such as UG, CATIA and Pro / E to realize the parametric modeling of parts; the process department uses 3D CAPP system to improve the efficiency of process design. However, the independent operation of CAD and CAPP systems has led to the inability of design information and manufacturing information to communicate with each other. The traditional way is to convert design information into manufacturing information by manual means, which increases product development time and reduces the competitiveness of enterprises. Therefore, studying processing feature recognition technology and realizing the communication between design information and manufacturing information is one of the effective ways to improve the efficiency of product design and process planning, which is of great significance to improving enterprise production efficiency and reducing product production costs.

[0003] In this field, researchers have proposed a variety of feature recognition methods based on traces, graphs, and volume decomposition. In traditional feature recognition methods, researchers must develop reliable recognition rules based on a full understanding of the various processing feature structures. With the continuous development of deep learning technology, the application of multi-layer neural network technology to identify processing features is deepening, and the representation of part models is constantly updated. Using the attribute adjacency graph and graph neural network model, the original model can be converted into data that can be processed by the neural network to maintain the geometric topological characteristics of the part model, and automatically learn complex feature representations, which can efficiently parse the processing features. However, the existing research on processing feature recognition using attribute adjacency graphs and graph neural network models mainly realizes the semantic classification of processing features of part faces, rather than processing feature instance recognition, that is, it cannot determine which faces the processing feature instance is composed of, so that it has no practical guiding significance for subsequent process design.

[0004] In view of this, the design and implementation of a multi-information fusion processing feature recognition method based on graph neural network is of great significance for the integration of CAD and CAPP systems, automatic generation of part process procedures and automatic decision-making of process parameters. Summary of the invention

[0005] Purpose of the invention: The present invention provides a multi-information fusion processing feature recognition method and device based on graph neural network, which not only solves the problem of distortion in part model representation, but also solves the problem of difficulty in identifying complex features and cross-features.

[0006] Technical solution: The multi-information fusion processing feature recognition method based on graph neural network described in the present invention comprises the following steps:

[0007] (1) Select a series of basic processing features and combine them to form a multi-featured part digital model and label them to form a processing feature data set;

[0008] (2) Preprocessing the processing feature data set, including sampling the faces and edges of the parts and constructing the face adjacency graph, constructing the face connection relationship matrix, constructing the geometric attributes of the faces and edges, and constructing the attribute table of the faces and edges;

[0009] (3) encoding the data processed in step (2), including encoding the face edge data and encoding the geometric attributes of the face and edge; the encoded data is used as the feature vectors of the nodes and edges in the adjacency graph, thereby forming an attribute adjacency graph data set;

[0010] (4) Construct a graph neural network with a processing feature semantic classifier and a bottom surface classifier to extract the processing feature information composed of the part face and edge combination, and perform processing feature semantic classification and bottom surface classification on the node surface;

[0011] (5) The outputs of the two classifiers are integrated with the connection relationship matrix, and the recognition effect is demonstrated using the deep image library.

[0012] Furthermore, the labeling implementation process in step (1) is as follows:

[0013] Construct feature classification label vector T for each face m , the dimension m is determined by the total category of the machining feature. If the surface belongs to a certain feature, the dimension value is 1, and the other dimensions are 0;

[0014] Construct a bottom surface classification label a for each face i , if the face is a bottom face, the value is 1, otherwise it is 0.

[0015] Furthermore, the data preprocessing process in step (2) is as follows:

[0016] UV-Net is used to characterize and sample the faces and edges of the parts, and the input tensor of the faces and edges is constructed. The input tensor of the faces contains the absolute coordinates xyz of the sampling points and the derivative d of the sampling points in the UV direction. U and d V ; The input tensor of the edge contains the absolute coordinates xyz of the sampling point, and the sampling point is on the surface f iThe tangent vector T on xi , T yi , T zi , the sampling point is on the surface f j The tangent vector T on xj , T yj , T zj ;

[0017] Use Python's OCC module to read the part model step file, identify the faces and edges of the model, and assign a unique identification number f i , l j , build an adjacency graph with faces as nodes and edges as relationships; build a connection relationship matrix L for all faces of the entire part n×n , n represents the number of part faces, and the matrix value l ij Indicates the connection relationship between face i and face j. When identifying faces and edges, if they are connected, then l ij is 1, otherwise it is 0, and l ii The value is also set to 1;

[0018] Use Python's OCC module to read the attribute information of the face and edge in the part model; the attributes of the face include the face number, type, area, and concavity; the attributes of the edge include the edge number, type, length, concavity, adjacent face f i Number and adjacent faces f j serial number.

[0019] Furthermore, the implementation process of step (3) is as follows:

[0020] Data encoding is used for face edges to extract spatial feature information of face edges. The face encoder contains three layers of two-dimensional convolutional neural networks, each layer contains a 2DCNN layer, a regularization layer, and an activation layer. The sampled data of the face is processed by the encoder to obtain the corresponding vector. The edge encoder contains three layers of one-dimensional convolutional neural networks, each layer contains a 1DCNN layer, a regularization layer, and an activation layer. The sampled data of the edge is processed by the encoder to obtain the corresponding vector.

[0021] The face and edge attributes are encoded. The geometric attribute encoders of the face and edge contain a multi-layer perceptron neural network, each layer of which contains a linear layer, a regularization layer, and an activation layer. The geometric attributes of the face and edge are input into their respective encoders to obtain corresponding vectors.

[0022] Furthermore, the implementation process of step (4) is as follows:

[0023] The face encoding vector and face attribute encoding vector are fused as the feature vector of the graph node; the edge encoding vector and edge attribute encoding vector are fused as the feature vector of the graph edge; the graph neural network contains s graph neural sub-networks, each of which contains normalization, information transfer neural network MPNN, normalization, and multi-layer perceptron;

[0024] Construct processing feature semantic classifiers and bottom surface classifiers to extract more information about parts. Both semantic recognition and bottom surface recognition classifiers contain multi-layer perceptron neural networks, each of which includes a linear layer, a normalization layer, and an activation layer. Inputting the data obtained through graph neural network processing into the two tasks can determine whether the node surface belongs to a certain processing feature and whether it belongs to the bottom surface of the processing feature; the loss functions of semantic recognition and bottom surface recognition classifiers are as follows:

[0025]

[0026] L=αL s +βL b

[0027] Among them, L s , L b , L represent semantic recognition loss, bottom surface recognition loss and comprehensive task loss function respectively, N represents the number of node faces, K represents the number of processing feature categories, t i represents the label vector of node i, p i represents the predicted probability vector of node i, and α and β are hyperparameters that weigh the semantic recognition loss and the bottom surface recognition loss against the comprehensive task loss.

[0028] Furthermore, the implementation process of step (5) is as follows:

[0029] The processing feature semantic classifier can output the node surface f i The processing feature classification vector f i,m , m represents the number of machining feature types, and the machining feature classification vectors of all faces of the part constitute the machining feature classification matrix F of the part n×m , n represents the number of part faces; the bottom surface classifier can output the node face f i The bottom classification vector b i,1 , the bottom surface classification vectors of all parts constitute the bottom surface classification matrix B of the part n×1 ; Classify the processing feature matrix F n×m , bottom classification matrix B n×1 and the connection relationship matrix L constructed in claim 3 n×n Splice to get the part processing feature recognition matrix Z n×(m+1+n) ; Integrate the information of the matrix and use the deep library tool to display the processing feature recognition results.

[0030] A device according to the present invention includes a memory and a processor, wherein:

[0031] A memory for storing computer programs that can be run on the processor;

[0032] A processor is used to execute the steps of the multi-information fusion processing feature recognition method based on graph neural network as described above when running the computer program.

[0033] A storage medium described in the present invention stores a computer program on the storage medium, and when the computer program is executed by at least one processor, the steps of the multi-information fusion processing feature recognition method based on graph neural network as described above are implemented.

[0034] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. The present invention uses attribute adjacency graph and graph neural network to identify processing features, thereby improving the adaptability of the method;

[0036] 2. Compared with the existing methods, the present invention combines the connection relationship between the bottom surface and the surface of the processing feature, which can not only identify that a single surface belongs to a certain type of feature, but also determine which surfaces together constitute a processing feature, thereby providing practical technical support for subsequent process planning;

[0037] 3. The present invention uses UV-Net to characterize the faces and edges of parts, can characterize small processing features, and normalize the sampling points, which is conducive to processing new parts. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the multi-information fusion processing feature recognition method based on graph neural network;

[0039] Figure 2 It is a schematic diagram of the processing feature recognition network structure proposed by the present invention. DETAILED DESCRIPTION

[0040] The present invention is further described in detail below with reference to the accompanying drawings.

[0041] Aiming at the information connectivity between CAD and CAPP, and the effective identification and expression of part processing features, this paper proposes a multi-information fusion processing feature recognition method based on graph neural network, such as Figure 1 As shown, the following steps are included:

[0042] Step 1: Select a series of basic processing features and combine them to form a part digital model with multiple features and label them to form a processing feature data set.

[0043] Construct feature classification label vector T for each face m , the total category of the processing feature of dimension m is determined. If the surface belongs to a certain feature, the value of this dimension is 1, and the other dimensions are 0; construct the bottom surface classification label a for each surface i , if the face is a bottom face, the value is 1, otherwise it is 0.

[0044] Step 2: Preprocess the processing feature data set, including sampling the faces and edges of the part and constructing the face adjacency graph, constructing the face connection relationship matrix, constructing the geometric properties of the faces and edges, and constructing the attribute table of the faces and edges.

[0045] UV-Net is used to characterize and sample the faces and edges of the parts, and the input tensor of the faces and edges is constructed. The input tensor of the faces contains the absolute coordinates xyz of the sampling points and the derivative d of the sampling points in the UV direction. U and d V ; The input tensor of the edge contains the absolute coordinates xyz of the sampling point, and the sampling point is on the surface f i The tangent vector T on xi , T yi , T zi , the sampling point is on the surface f j The tangent vector T on xj , T yj , T zj .

[0046] Use Python's OCC module to read the part model step file, identify the faces and edges of the model, and assign a unique identification number f i , l j , build an adjacency graph with faces as nodes and edges as relationships; build a connection relationship matrix L for all faces of the entire part n×n , n represents the number of part faces, and the matrix value l ij Indicates the connection relationship between face i and face j. When identifying faces and edges, if they are connected, then l ij is 1, otherwise it is 0, and l ii The value is also set to 1.

[0047] Use Python's OCC module to read the attribute information of the face and edge in the part model. The attributes of the face include the face number, type, area, and concavity; the attributes of the edge include the edge number, type, length, concavity, and adjacent face f i Number, adjacent face f j serial number.

[0048] Step 3: Encode the data processed in step 2, including encoding the face edges and the geometric attributes of the faces and edges; the encoded data is used as the feature vectors of the nodes and edges in the adjacency graph to form an attribute adjacency graph data set.

[0049] Data encoding is used for face edges to extract spatial feature information of face edges. The face encoder contains three layers of two-dimensional convolutional neural networks, each layer contains a 2DCNN layer, a regularization layer, and an activation layer. The sampled data of the face is processed by the encoder to obtain the corresponding vector. The edge encoder contains three layers of one-dimensional convolutional neural networks, each layer contains a 1DCNN layer, a regularization layer, and an activation layer. The sampled data of the edge is processed by the encoder to obtain the corresponding vector.

[0050] The face and edge attributes are encoded. The geometric attribute encoders of the face and edge contain a multi-layer perceptron neural network, each layer of which contains a linear layer, a regularization layer, and an activation layer. The geometric attributes of the face and edge are input into their respective encoders to obtain corresponding vectors.

[0051] In this embodiment, the points of the adjacency graph represent the faces of the parts, denoted by f i , the edges of the adjacency graph represent the edges of the parts, denoted by l i ; Connection relationship matrix L n×n Indicates the connection between the surfaces, l ij Indicates face i With noodles j Connection relationship between them, if connected, then l ij is 1, otherwise it is 0, where l ii is 1; OCC is used to sample the surface and edge, and the surface sampling data F ic Including the absolute coordinates xyz of the sampling point and the derivative d of the sampling point in the UV direction U and d V , the sampling data size of the surface is [u f ,v f ,3,2], the edge sampling data E ic Including the absolute coordinates xyz of the sampling point, the sampling point on the surface f i The tangent vector T on xi , T yi , T zi , the sampling point is on the surface f j The tangent vector T on xj , T yj , T zj , the sampling data size of the edge is [u e ,3,3]; Use OCC to read the attribute information of the surface and edge, the attribute F of the surface is Including number, type, convexity and area, edge attributes E is Including number, type, convexity, length, adjacent surface f i The number and adjacent faces f j The part attribute adjacency graph dataset composed of all the above features is divided into a training set and a test set in a ratio of 8:2.

[0052] Step 4: Construct a processing feature recognition network based on graph neural network, integrate the geometric information, topological information and attribute information of the face edge, and realize both the semantic classification of face features and the classification of feature instances. The main structure of the processing feature recognition network based on graph neural network is as follows: Figure 2 As shown, the details are as follows:

[0053] The face encoding vector and face attribute encoding vector are fused as the feature vector of the graph node; the edge encoding vector and edge attribute encoding vector are fused as the feature vector of the graph edge; the graph neural network contains s graph neural sub-networks, each of which contains normalization, information transfer neural network MPNN, normalization, and multi-sensor;

[0054] Construct processing feature semantic classifiers and bottom surface classifiers to extract more information about parts. Both semantic recognition and bottom surface recognition classifiers contain multi-layer perceptron neural networks, each of which includes a linear layer, a normalization layer, and an activation layer. Inputting the data obtained through graph neural network processing into the two tasks can determine whether the node surface belongs to a certain processing feature and whether it belongs to the bottom surface of the processing feature; the loss functions of the two classifiers are as follows:

[0055]

[0056]

[0057] L=aL s +βL b

[0058] Among them, L s , L b , L represent semantic recognition loss, bottom surface recognition loss and comprehensive task loss function respectively, N represents the number of node faces, K represents the number of processing feature categories, t i represents the label vector of node i, p i represents the predicted probability vector of node i, and α and β are hyperparameters that weigh the semantic recognition loss and the bottom surface recognition loss against the comprehensive task loss.

[0059] The sampling data of the face and edge are encoded using a two-dimensional convolutional neural network and a one-dimensional convolutional neural network to obtain the corresponding vector F ie and E ie ; Use multi-perceptron to encode the attributes of the face edges and obtain the corresponding vector F im and E im ; Concatenate the sampling vector and the attribute vector to form the feature vector F of the node face and edge in the attribute adjacency graph i and E i , where F i =[F ie |F im ], E i =[Eie |E im ], the face feature matrix F of the part is composed of all face feature vectors, and the edge feature matrix E is composed of all edge feature vectors. A graph neural network is constructed, with F and E as input, and after processing by Norm, MPNN, MLP, etc., the information of nodes and edges is extracted to obtain the information feature Z of the high-dimensional space.

[0060] The training process of the processing feature recognition network model based on graph neural network is input as F c 、E c 、F s With E s , the output is F n×24 With B n×1 The amount of training data is 40,000 data in the part data set, and 1,000 data are selected for verification. The number of training rounds is 1,000 rounds, the learning rate is 0.006, the optimizer is Adam, and the training batch is set to 64. The processing feature recognition network based on the graph neural network obtained by the final training can achieve a relatively good effect on the recognition of part processing features.

[0061] Step 5: Fusion the outputs of the two classifiers and the connection relationship matrix, and use the deep image library to demonstrate the recognition effect.

[0062] The processing feature semantic classifier can output the node surface f i The processing feature classification vector f i,m , m represents the number of machining feature types, and the machining feature classification vectors of all faces of the part constitute the machining feature classification matrix F of the part n×m , n represents the number of part faces; the bottom surface classifier can output the node face f i The bottom classification vector b i,1 , the bottom surface classification vectors of all parts constitute the bottom surface classification matrix B of the part n×1 ; Classify the processing feature matrix F n×m , bottom classification matrix B n×1 And the connection relationship matrix L n×n Splice to get the part processing feature recognition matrix Z n×(m+1+n) By integrating the information of this matrix, the processing feature recognition results can be displayed using the Deep Gallery tool.

[0063] The present invention also provides a device, including a memory and a processor, wherein the memory is used to store a computer program that can be run on the processor; the processor is used to execute the steps of the multi-information fusion processing feature recognition method based on graph neural network as described above when running the computer program.

[0064] The present invention also provides a storage medium having a computer program stored thereon. When the computer program is executed by at least one processor, the steps of the multi-information fusion processing feature recognition method based on graph neural network as described above are implemented.

[0065] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-information fusion processing feature recognition method based on graph neural network, characterized in that: The following steps are involved: (1) Select a series of basic processing features and combine them to form a multi-featured part digital model and label them to form a processing feature data set; (2) Preprocessing the processing feature data set, including sampling the faces and edges of the parts and constructing the face adjacency graph, constructing the face connection relationship matrix, constructing the geometric attributes of the faces and edges, and constructing the attribute table of the faces and edges; (3) encoding the data processed in step (2), including encoding the face edge data and encoding the geometric attributes of the face and edge; the encoded data is used as the feature vectors of the nodes and edges in the adjacency graph, thereby forming an attribute adjacency graph data set; (4) Construct a graph neural network with a processing feature semantic classifier and a bottom surface classifier to extract the processing feature information composed of the part face and edge combination, and perform processing feature semantic classification and bottom surface classification on the node surface; (5) The outputs of the two classifiers are integrated with the connection relationship matrix, and the recognition effect is demonstrated using the deep image library.

2. The multi-information fusion processing feature recognition method based on graph neural network according to claim 1 is characterized in that: The labeling process in step (1) is as follows: Construct feature classification label vector T for each face m , the dimension m is determined by the total category of the machining feature. If the surface belongs to a certain feature, the dimension value is 1, and the other dimensions are 0; Construct a bottom surface classification label a for each face i , if the face belongs to a face, the value is 1, otherwise it is 0.

3. The multi-information fusion processing feature recognition method based on graph neural network according to claim 1 is characterized in that: The data preprocessing process described in step (2) is as follows: UV-Net is used to characterize and sample the faces and edges of the parts, and the input tensor of the faces and edges is constructed. The input tensor of the faces contains the absolute coordinates xyz of the sampling points and the derivative d of the sampling points in the UV direction. U and d V ; The input tensor of the edge contains the absolute coordinates xyz of the sampling point, and the sampling point is on the surface f i The tangent vector T on xi 、T yi 、T zi , the sampling point is on the surface f j The tangent vector T on xj 、T yj 、T zj ; Use Python's OCC module to read the part model step file, identify the faces and edges of the model, and assign a unique identification number f i , l j , build an adjacency graph with faces as nodes and edges as relationships; build a connection relationship matrix L for all faces of the entire part n×n , n represents the number of part faces, and the matrix value l ij Indicates the connection relationship between face i and face j. When identifying faces and edges, if they are connected, then l ij is 1, otherwise it is 0, and l ii The value is also set to 1; Use Python's OCC module to read the attribute information of faces and edges in the part model.

4. The multi-information fusion processing feature recognition method based on graph neural network according to claim 1 is characterized in that: The implementation process of step (3) is as follows: Data encoding is used for face edges to extract spatial feature information of face edges. The face encoder contains three layers of two-dimensional convolutional neural networks, each layer contains a 2DCNN layer, a regularization layer, and an activation layer. The sampled data of the face is processed by the encoder to obtain the corresponding vector. The edge encoder contains three layers of one-dimensional convolutional neural networks, each layer contains a 1DCNN layer, a regularization layer, and an activation layer. The sampled data of the edge is processed by the encoder to obtain the corresponding vector. The face and edge attributes are encoded. The geometric attribute encoders of the face and edge contain a multi-layer perceptron neural network, each layer of which contains a linear layer, a regularization layer, and an activation layer. The geometric attributes of the face and edge are input into their respective encoders to obtain corresponding vectors.

5. The multi-information fusion processing feature recognition method based on graph neural network according to claim 1 is characterized in that: The implementation process of step (4) is as follows: The face encoding vector and face attribute encoding vector are fused as the feature vector of the graph node; the edge encoding vector and edge attribute encoding vector are fused as the feature vector of the graph edge; the graph neural network contains s graph neural sub-networks, each of which contains normalization, information transfer neural network MPNN, normalization, and multi-layer perceptron; Construct processing feature semantic classifiers and bottom surface classifiers to extract more information about parts. Both semantic recognition and bottom surface recognition classifiers contain multi-layer perceptron neural networks, each of which includes a linear layer, a normalization layer, and an activation layer. Inputting the data obtained through graph neural network processing into the two tasks can determine whether the node surface belongs to a certain processing feature and whether it belongs to the bottom surface of the processing feature; The loss functions of the semantic recognition and bottom surface recognition classifiers are as follows: L=αL s +βL b Among them, L s , L b , L represent semantic recognition loss, bottom surface recognition loss and comprehensive task loss function respectively, N represents the number of node faces, K represents the number of processing feature categories, t i represents the label vector of node i, p i represents the predicted probability vector of node i, and α and β are hyperparameters that weigh the semantic recognition loss and the bottom surface recognition loss against the comprehensive task loss.

6. The multi-information fusion processing feature recognition method based on graph neural network according to claim 1 is characterized in that: The implementation process of step (5) is as follows: The processing feature semantic classifier can output the node surface f i The processing feature classification vector f i,m , m represents the number of machining feature types, and the machining feature classification vectors of all faces of the part constitute the machining feature classification matrix F of the part n×m , n represents the number of part faces; the bottom surface classifier can output the node face f i The bottom classification vector b i,1 , the bottom surface classification vectors of all parts constitute the bottom surface classification matrix B of the part n×1 ; Classify the processing feature matrix F n×m , bottom classification matrix B n×1 and the connection relationship matrix L constructed in claim 3 n×n Splice to get the part processing feature recognition matrix Z n×(m+1+n) ; Integrate the information of the matrix and use the deep library tool to display the processing feature recognition results.

7. The multi-information fusion processing feature recognition method based on graph neural network according to claim 3 is characterized in that: The attributes of the surface include the number, type, area, and concavity of the surface.

8. The multi-information fusion processing feature recognition method based on graph neural network according to claim 3 is characterized in that: The attributes of the edge include the edge number, type, length, concavity, adjacent faces, i Number and adjacent faces f j serial number.

9. A device, characterized in that: comprising a memory and a processor, wherein: A memory for storing computer programs that can be run on the processor; A processor is used to execute the steps of the multi-information fusion processing feature recognition method based on graph neural network as described in any one of claims 1 to 8 when running the computer program.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by at least one processor, implements the steps of the multi-information fusion processing feature recognition method based on graph neural network as described in any one of claims 1 to 8.

Images