Deep learning-based machining feature classification and recognition method

Abstract

The present application relates to a kind of processing feature classification identification method based on deep learning, including the following steps: extracting the geometry, topological information of model from the STEP neutral file of part three-dimensional model, and store as memory work format;The intersection edge of two adjacent surfaces in model is concave-convex determination, establishes attribute adjacency graph, and store as extended attribute adjacency matrix;Processing feature subgraph is extracted from attribute adjacency graph, generates processing feature adjacency submatrix, processing feature adjacency submatrix is matched with the extended attribute adjacency matrix in predefined feature library, and processing feature is obtained and highlighted display, obtains two-dimensional color processing feature image, generates processing feature dataset;Processing feature classification identification model based on deep network learning is constructed, and the model is used for identifying classification to the processing feature of the part to be identified after training.The method based on deep network learning effectively solves the problem of complex processing feature recognition difficulty, improves the accuracy of identification result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machining, in particular to a machining feature classification and recognition method based on deep learning. BACKGROUND

[0002] Part feature recognition is a basic technology for realizing automatic design of machining process and automatic programming of numerical control machining. Machining feature is the link between CAD and CAPP (Computer Aided Process Planning) systems, but when CAD and CAPP systems run independently, it leads to the fact that design information and manufacturing information cannot be communicated, and it is necessary to convert design information into manufacturing information through manual means. This process not only affects the standardization of process, but also is difficult to guarantee the machining quality, and reduces the product manufacturing efficiency. At present, one of the effective ways to realize the communication between design information and manufacturing information is machining feature recognition technology.

[0003] The machining feature recognition method mainly includes graph-based, rule-based, volume decomposition-based and trace-based methods, but these machining feature recognition methods have the following problems: (1) features have diversity, different application systems only seek the recognition algorithm of predetermined features according to needs, so that the algorithm cannot adapt to the requirements of feature diversity; (2) without learning ability, the noise resistance of CAD three-dimensional model is poor, the ability to process complex and intersecting features is weak, and the recognition efficiency is low, which cannot guarantee the accuracy of the recognition result. SUMMARY

[0004] In view of the above problems, the present application provides a machining feature classification and recognition method based on deep learning, which combines the traditional machining feature recognition method with the deep learning network, can effectively solve the problem of complex machining feature recognition difficulty, and improve the accuracy of the recognition result.

[0005] In order to solve the above problems, the present application adopts the following technical scheme:

[0006] The machining feature classification and recognition method based on deep learning comprises the following steps:

[0007] Step 1: export a STEP neutral file from a part three-dimensional model, extract the geometry and topology information of the part model from the STEP neutral file, and store the geometry and topology information as a memory work format file;

[0008] Step 2: determine the concave-convexity of the adjacent two face intersecting edges in the part model based on the memory work format file;

[0009] Step 3: According to the concave-convex determination result of the intersecting edge of the two adjacent surfaces, an attribute adjacency graph of the part model is constructed, and is stored in the form of an extended attribute adjacency matrix, the extended attribute adjacency matrix is a symmetric matrix, different values of the diagonal elements represent different surfaces of the corresponding part model, and the non-diagonal elements represent the adjacency attributes between two surfaces, if the intersecting edge of the two surfaces is a concave edge, the unit digit of the element is 0; if the intersecting edge of the two surfaces is a convex edge, the unit digit of the element is 1; if the two surfaces do not intersect, the element is 0;

[0010] Step 4: The processing feature subgraph extraction algorithm is used to decompose the attribute adjacency graph, extract the processing feature subgraph, and generate the corresponding processing feature adjacency submatrix;

[0011] Step 5: The processing feature adjacency submatrix is matched with the extended attribute adjacency matrix of the processing feature in the pre-defined feature library, and the processing feature matching result is obtained;

[0012] Step 6: According to the processing feature matching result, the corresponding processing feature in the part model is highlighted;

[0013] Step 7: The three-dimensional processing feature in the part model is rotated and view-captured at multiple angles based on the multi-angle dimension reduction capturing method, two-dimensional color processing feature images are captured at different angles, a processing feature dataset is generated and the dataset is preprocessed;

[0014] Step 8: The processing features contained in the two-dimensional color processing feature images in the dataset are framed and the type of each processing feature is labeled, the information labeling of the dataset is completed, and the labeled dataset is divided into a training set and a test set;

[0015] Step 9: A processing feature classification and recognition model based on deep network learning is constructed, and the training set and the test set are used to train and test the processing feature classification and recognition model, so as to obtain a trained processing feature classification and recognition model, the processing feature classification and recognition model includes a backbone feature extraction network, a reinforced feature fusion network and a prediction output network, wherein the backbone feature extraction network adopts a Backbone structure to extract three feature maps, the reinforced feature fusion network includes two feature pyramid structures of FPN and PAN, and the three input feature maps are fused, three feature maps of different sizes are obtained after fusion, and the prediction output network performs convolution operation on the three feature maps of different sizes respectively, and outputs a prediction result;

[0016] Step 10: The trained processing feature classification and recognition model is used to recognize the two-dimensional color processing feature image corresponding to the part model to be recognized, and a processing feature classification and recognition result is obtained.

[0017] The present application has the following beneficial effects:

[0018] The machining feature classification and recognition method based on deep learning provided by the application firstly derives the corresponding STEP neutral file of the part three-dimensional model, and extracts the geometric and topological information of the model in the form of memory work format; then the concave-convex property of the intersecting edge of two adjacent surfaces in the model is determined based on the memory work format of STEP, so as to establish the attribute adjacency graph of the model, and store the attribute adjacency graph in the form of extended attribute adjacency matrix; the machining feature subgraph extraction algorithm is used to extract the machining feature subgraph from the attribute adjacency graph, and generate the machining feature adjacency submatrix; the machining feature adjacency submatrix is matched with the extended attribute adjacency matrix in the predefined feature library, so as to obtain the machining feature corresponding to the machining feature subgraph; the machining feature obtained by matching is highlighted in the part model, a two-dimensional color machining feature image is obtained based on the multi-angle dimension reduction capturing method, a machining feature dataset is generated, and the dataset is preprocessed to form a training set and a test set; finally, a machining feature classification and recognition model based on deep network learning is constructed, and the training set and the test set are used to complete the training and testing of the model, so that the trained machining feature classification and recognition model can realize the recognition and classification of the machining feature of the part to be recognized. The machining feature recognition technology based on the attribute adjacency graph and the deep neural network technology are combined, the advantages of the two are combined, and the machining feature diversity can be adapted. Meanwhile, the machining feature classification and recognition model based on deep learning not only has learning ability, but also can effectively solve the problem of complex machining feature recognition, so that the recognition rate is higher. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The flowchart of the machining feature classification and recognition method based on deep learning described in the embodiments of the application;

[0020] Figure 2 The concave-convex property determination diagram for the straight line edge intersecting between two adjacent surfaces;

[0021] Figure 3 The concave-convex property determination diagram for the circular edge intersecting between two adjacent surfaces;

[0022] Figure 4 The flowchart of the process of constructing the attribute adjacency graph;

[0023] Figure 5 The machining feature model diagram corresponding to the U-shaped groove;

[0024] Figure 6 The attribute adjacency graph corresponding to the U-shaped groove;

[0025] Figure 7 The machining feature model diagram corresponding to the step;

[0026] Figure 8 The attribute adjacency graph corresponding to the step-shaped groove;

[0027] Figure 9 A machining feature model schematic diagram corresponding to the stepped hole is shown in Figure 1;

[0028] Figure 10 An attribute adjacency graph corresponding to the stepped hole is shown in Figure 2;

[0029] Figure 11 A machining feature model schematic diagram corresponding to the through slot is shown in Figure 3;

[0030] Figure 12 An attribute adjacency graph corresponding to the through slot is shown in Figure 4;

[0031] Figure 13 A machining feature model schematic diagram corresponding to the through hole is shown in Figure 5;

[0032] Figure 14 An attribute adjacency graph corresponding to the through hole is shown in Figure 6;

[0033] Figure 15 A machining feature model schematic diagram corresponding to the protrusion is shown in Figure 7;

[0034] Figure 16 An attribute adjacency graph corresponding to the protrusion is shown in Figure 8;

[0035] Figure 17 A flowchart of matching between a machining feature adjacency sub-matrix and an extended attribute adjacency matrix in a predefined feature library is shown in Figure 9;

[0036] Figure 18 A network structure schematic diagram of a machining feature classification recognition model in an embodiment of the present application is shown in Figure 10;

[0037] Figure 19 Two CSP module structure schematic diagrams are shown in Figures 11 and 12. DETAILED DESCRIPTION

[0038] In order to more clearly express the application purpose, technical scheme and advantages of the present application, the present application is described in detail below with reference to the accompanying drawings and embodiments.

[0039] As shown in Figure 1, Figure 1 the present application embodiment provides a machining feature classification recognition method based on deep learning, which specifically includes the following steps:

[0040] Step 1: export the corresponding STEP neutral file from the part three-dimensional model, extract the geometry and topology information of the part model from the STEP neutral file, and store the extracted geometry and topology information as a memory work format file.

[0041] In the embodiment, the STEP neutral file is a product data exchange method of the STEP standard, which mainly includes two parts: a header section and a data section. The header section describes the basic information of the entire file. The entity instances in the data section are nested and called each other, forming geometric entities and features in the entities. Moreover, the same entity instance can be repeatedly called by multiple instances. In order to conveniently obtain the data information in the STEP, the entity information contained therein is accessed through a pointer, so as to extract the geometric and topological information, and store the extracted information in a memory working format, thereby providing an information source for the subsequent concave-convex judgment of the edges and the construction of the attribute adjacency graph.

[0042] Step 2: Based on the memory working format file, the concave-convexity of the intersecting edge between the two adjacent faces in the part model is judged. If the intersecting edge is a convex edge, the corresponding attribute value is 1; if the intersecting edge is a concave edge, the corresponding attribute value is 0.

[0043] In the embodiment, the type of the intersecting edge between the two adjacent faces is mainly a straight line edge and a circular edge, Figure 2 and Figure 3 are schematic diagrams for judging the concave-convexity when the intersecting edge between the two adjacent faces is a straight line edge and a circular edge, respectively. In Figure 2 and Figure 3 , it is assumed that the two intersecting faces are face F1 and face F2.

[0044] As shown in Figure 2 , when the intersecting edge is a straight line edge, the concave-convexity judgment method is as follows: the memory working format file is traversed to find the ADVANCED_FACE corresponding to face F1 and face F2, the direction vector given by the attribute direction DIRECTION of the second attribute direction AXIS2_PLACEMENT_3D of the three-dimensional space of face F1 and face F2 is obtained, the outer normal vector of the face is determined according to the direction identification Boolean value of face F1 and face F2, if the Boolean value is true, the direction vector given by the attribute direction DIRECTION is the outer normal vector of the face, otherwise the outer normal vector of the face is reversed, the outer normal vector N1 of face F1 and the outer normal vector N2 of face F2 are obtained. Taking face F1 as the reference face, the FACE_OUTER_BOUND or FACE_BOUND constituting face F1 is found, the EDGE_LOOP thereof is obtained, the ORIENTED_EDGE corresponding to the edge e and the EDGE_CURVE thereof are obtained, and finally the direction vector corresponding to the attribute direction DIRECTION of the LINE of the geometric definition curve of the edge e is obtained. When the direction identification Boolean values of the edges are all true, the direction vector corresponding to the attribute direction DIRECTION is the direction vector N e of the edge e, otherwise the direction vector N e is reversed. According to the obtained outer normal vector N1, outer normal vector N2 and direction vector N eN can be calculated e The vector product of N and N2 yields the direction vector N, i.e., N = N2. e ×N2, the concavity or convexity of edge e is determined by judging the angle θ between N and N1: when |θ|<π, edge e is concave; when |θ|>π, edge e is convex.

[0045] like Figure 3 As shown, when the intersecting edges are rounded, the method for determining concavity / convexity is as follows: Traverse the memory working format file to find the high-level face ADVANCED_FACE corresponding to face F1 and face F2. The external normal vector of face F1 is N1, which is obtained in the same way as the method for obtaining the external normal vector of face F1 when the intersecting edges are straight lines. The external normal vector of face F2, N2, needs to be calculated through the following steps: First, obtain the coordinates of the first attribute vertex VERTEX_POINT in the edge curve EDGE_CURVE of the geometric definition curve CIRCLE of face F1, denoted as point P. Then, find the coordinates of the center of the circle corresponding to the first attribute Cartesian coordinate point CARTESIAN_POINT in the 3D orientation AXIS2_PLACEMENT_3D of the geometric definition curve CIRCLE, denoted as point P0. The vector P is formed by points P and P0. V Its direction is from point P to point P0; obtain the direction vector V corresponding to the second attribute DIRECTION in the geometric definition curve CIRCLE of face F2; calculate vector P V The vector product of the direction vector V and the vector R is obtained. V R V =P V ×V; then calculate the outer normal vector N2 of surface F2 based on whether it is the outer surface of the cylindrical surface that constitutes the solid surface of the part. If it is the outer surface, then the outer normal vector N2 = R. V ×V, otherwise the outer normal vector N2 = V×R V ; Calculate the direction vector N of edge e e =P V ×N1, then calculate the direction vector N e The vector product of the outward normal vector N2 and the outward normal vector N2 yields the direction vector N = N. e ×N2, and finally the concavity or convexity of edge e is determined by the angle θ between the direction vector N and the external normal vector N1: when |θ|<π, edge e is concave; when |θ|>π, edge e is convex.

[0046] Step 3: According to the concave-convex determination result of the intersecting edge of the adjacent two surfaces, the attribute adjacency graph of the part model is constructed and stored in the form of an extended attribute adjacency matrix. The extended attribute adjacency matrix is a symmetric matrix, and different values of the diagonal elements represent different surfaces of the corresponding part model, and the non-diagonal elements represent the adjacency attributes between two surfaces. If the intersecting edge of the two surfaces is a concave edge, the unit digit of the element is 0; if the intersecting edge of the two surfaces is a convex edge, the unit digit of the element is 1; if the two surfaces do not intersect, the element is 0.

[0047] In this step, the process of constructing the attribute adjacency graph of the part model specifically includes the following steps, as shown in Figure 4

[0048]

[0049] The extended attribute adjacency matrix is expressed by a two-dimensional array of n x n, and different values are assigned to the elements, which are defined as follows:

[0050] (1) In the extended attribute adjacency matrix A[n, n], when i = j (i≤n, j≤n), the element A[i, j] in the ith row and jth column represents the type of the surface. The value corresponding to A[i, j] is shown in Table 1.

[0051] Table 1 Definition of node attribute values of extended attribute adjacency matrix.

[0052]

[0053]

[0054] (2) In the extended attribute adjacency matrix A[n, n], when i≠j, the ith surface and the jth surface do not intersect, then A[i, j]=0; the ith surface and the jth surface intersect, and the value corresponding to A[i, j] is shown in Table 2.

[0055] Table 2 Definition of concave-convex attribute values of the edge of the extended attribute adjacency matrix.

[0056] ​​

[0057] Step 4: The attribute adjacency graph is decomposed by using the machining feature subgraph extraction algorithm to extract the machining feature subgraph and generate the corresponding machining feature adjacency submatrix.

[0058] In this embodiment, the machining feature subgraph extraction algorithm based on the feature base surface and the main shaft rotary surface is used to extract the machining feature subgraph from the attribute adjacency graph and generate the machining feature adjacency submatrix, and the specific algorithm is as follows:

[0059] Firstly, all nodes in the attribute adjacency graph are traversed to find the feature base surface nodes and the nodes adjacent to the feature base surface nodes and having the same characteristics, the machining feature subgraph containing the feature base surface and the corresponding machining feature adjacency submatrix are extracted, and the simplified attribute adjacency graph is obtained; continue to traverse to find all the main shaft rotary surface nodes, delete the nodes, and obtain the machining feature subgraph without the feature base surface and the corresponding machining feature adjacency submatrix.

[0060] Step 5: The machining feature adjacency submatrix is matched with the extended attribute adjacency matrix of the machining feature in the predefined feature library to obtain the machining feature matching result, wherein the predefined feature library includes the machining feature and the corresponding machining feature model, attribute adjacency graph and extended attribute adjacency matrix.

[0061] The predefined feature library uses the attribute adjacency graph and the extended attribute adjacency matrix to predefine the machining feature model, which mainly includes the machining features such as U-shaped groove, step, stepped hole, through groove, through hole and protrusion and the corresponding extended attribute adjacency matrix, as shown in Table 3; the machining feature model and the attribute adjacency graph corresponding to each machining feature are shown in Table 4. Figures 5-16

[0062] Table 3: Predefined feature library.

[0063]

[0064] The machining feature adjacency submatrix to be matched is matched with the extended attribute adjacency matrix of the machining feature in the predefined feature library, and finally the machining feature matching result is output, and the specific matching process is shown in Table 5, including the following steps: Figure 17

[0065] ​​The dimension of the machining feature adjacency sub-matrix to be matched is determined; a set of extended attribute adjacency matrices with the same dimension is found in the pre-defined feature library, and the search range in the pre-defined feature library can be narrowed according to the dimension; one of the extended attribute adjacency matrices in the set is selected, and it is judged whether the diagonal line and each row element of the extended attribute adjacency matrix are similar to those of the machining feature adjacency sub-matrix to be matched, if not, the next extended attribute adjacency matrix in the set is selected for judgment, if yes, the row and column elements of the machining feature adjacency sub-matrix to be matched are transformed according to the extended attribute adjacency matrix; it is judged whether the corresponding elements of the two matrices after transformation are the same, if yes, the matching is successful, and the machining feature matching result is obtained, otherwise, the next extended attribute adjacency matrix in the set is selected for judgment.

[0066] Step 6: According to the machining feature matching result, the corresponding machining feature in the part model is highlighted. This step highlights the matched machining feature in the part model, making it easier to capture the two-dimensional image of different machining features subsequently.

[0067] Step 7: The three-dimensional machining feature in the part model is subjected to multi-angle rotation and view capture operation based on the multi-angle dimension reduction capture method, and two-dimensional color machining feature images are captured at different angles to generate a machining feature dataset.

[0068] Step 8: The machining feature dataset is preprocessed, specifically, the machining features contained in the two-dimensional color machining feature images in the dataset are framed out, and the types of each machining feature are labeled, the information labeling of the dataset is completed, and then the labeled dataset is divided into a training set and a test set, which are used as training data and test data of the machining feature classification recognition model respectively.

[0069] Step 9: A machining feature classification recognition model based on deep network learning is constructed, and the training set and the test set are used to train and test the machining feature classification recognition model, and a trained machining feature classification recognition model is obtained. The machining feature classification recognition model includes four parts of inputs, backbone feature extraction network, enhanced feature fusion network and output, as shown in Figure 18 .

[0070] The machining feature classification recognition model adopts two different CSP modules, and the structure is as shown in Figure 19As shown. Among them, the CSP1_X module is used in the backbone feature extraction network, and the CSP2_X module is used in the enhanced feature fusion network. In the CSP1_X module, X represents the number of residual components used; when X = 1, that is, the CSP1_1 module, it means that one residual component is used; when X = 3, that is, the CSP1_3 module, it means that three residual components are used. In the CSP2_X module, X represents the number of convolution groups used; when X = 1, that is, the CSP2_1 module, it means that one convolution group is used.

[0071] The backbone feature extraction network adopts the Backbone structure. In the first layer, the Focus module is used to perform slicing and channel splicing operations on the 640x640x3 input image, so that the input image is changed from 3 channels to 12 channels, and then a convolution operation with 32 convolution kernels is performed to obtain a 320x320x32 feature map, effectively reducing the information loss caused by downsampling, reducing the calculation amount and improving the model training speed; then the Conv convolution, BN layer and activation function RELU, that is, the CBL module, are used for convolution operation to obtain a 160x160x64 feature map; then the CSP1_1 module divides the 160x160x64 feature map into two parts, and each part is subjected to a convolution operation to reduce the number of channels by half, one part is first subjected to the CBL module, then subjected to a residual structure and subjected to a convolution to obtain a feature Figure 1 , and the other part is directly subjected to a convolution to obtain a feature Figure 2 , and finally the features Figure 1 and Figure 2 are concatenated and input into the CBL module to obtain an output 160x160x64 feature map.

[0072] Continue to perform convolution operations through the CBL and CSP1_3 modules in Figure 18 to obtain 80x80x128, 40x40x256 and 20x20x512 feature maps, respectively;

[0073] The SPP module is added at the end of the Backbone structure to perform a convolution operation on the 20x20x512 feature map through the CBL module, and then the feature map is divided into 5x5, 9x9 and 13x13 grids for three times of maximum pooling operations to extract different features. The features are fused together using Concat, and finally a 20x20x512 feature map is output after convolution operation through the CBL module.

[0074] The reinforced feature fusion network is a Neck structure, mainly including two feature pyramid structures of FPN and PAN, and adding CSP structure and convolution operation, using the FPN structure to perform up-sampling feature fusion in a top-down manner, and using the PAN structure to fuse each output feature map of the FPN structure with a new feature map in a bottom-up manner, so that the feature maps of different layers are fused, which is more conducive to feature information extraction.

[0075] Finally, the prediction output network, i.e., the output end, performs convolution operation on the three feature maps of different sizes fused by the Neck structure, with a convolution kernel size of 1x1, so as to perform prediction, and after non-maximum suppression (NMS), the prediction result is output.

[0076] Step 10: Using the trained machining feature classification and recognition model to recognize the two-dimensional color machining feature image corresponding to the part model to be recognized, to obtain the machining feature classification and recognition result.

[0077] In this embodiment, the machining feature data set is used as the input of the deep learning network model for training, to obtain the trained machining feature classification and recognition model. After the part model to be recognized is processed by steps 1 to 7, the image information of the part model to be recognized, i.e., the two-dimensional color machining feature image corresponding to the part model to be recognized, is obtained. After the image is input into the trained machining feature classification and recognition model, the model outputs the final machining feature classification and recognition result. The machining feature classification and recognition method of this embodiment is based on deep network learning, effectively solving the problem of difficult recognition of complex machining features, and improving the accuracy of the recognition result.

[0078] The existing algorithm cannot adapt to the requirement of feature diversity, while the machining feature recognition technology based on attribute adjacency graph and the deep neural network technology are combined in the present application, which can adapt to the diversity of machining features. The existing algorithm does not have learning ability, has weak ability to process complex and intersecting features, has low recognition efficiency, and cannot guarantee the accuracy of the recognition result. The machining feature classification and recognition model based on deep learning used in the present application not only has learning ability, but also effectively solves the problem of difficult recognition of complex machining features, so that the recognition rate is higher.

[0079] The technical features of the above-described embodiments can be combined arbitrarily, and to make the description concise, all possible combinations of the technical features in the above-described embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present disclosure.

[0080] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A deep learning-based machining feature classification and recognition method, characterized in that, The method comprises the following steps: Step 1: exporting a STEP neutral file from a part three-dimensional model, extracting geometry and topology information of the part model from the STEP neutral file, and storing the geometry and topology information as a memory working format file; Step 2: judging the concave-convex property of an intersection edge between two adjacent surfaces in the part model based on the memory working format file; Step 3: constructing an attribute adjacency graph of the part model according to the concave-convex property judgment result of the intersection edge between the two adjacent surfaces, and storing the attribute adjacency graph in the form of an extended attribute adjacency matrix, wherein the extended attribute adjacency matrix is a symmetric matrix, different values of diagonal elements represent different surfaces of the corresponding part model, and non-diagonal elements represent the adjacency attributes between two surfaces; if the intersection edge between the two surfaces is a concave edge, the unit digit of the element is 0; if the intersection edge between the two surfaces is a convex edge, the unit digit of the element is 1; if the two surfaces do not intersect, the element is 0; Step 4: decomposing the attribute adjacency graph by using a machining feature subgraph extraction algorithm, extracting a machining feature subgraph, and generating a corresponding machining feature adjacency submatrix; Step 5: matching the machining feature adjacency submatrix with an extended attribute adjacency matrix of a machining feature in a pre-defined feature library to obtain a machining feature matching result; Step 6: highlighting the corresponding machining feature in the part model according to the machining feature matching result; Step 7: performing multi-angle rotation and view capture operations on the three-dimensional machining feature in the part model based on a multi-angle dimension reduction capture method, capturing two-dimensional color machining feature images under different angles, generating a machining feature dataset, and preprocessing the dataset; Step 8: framing the machining feature contained in the two-dimensional color machining feature images in the dataset, labeling the type of each machining feature, completing the information labeling of the dataset, and then dividing the labeled dataset into a training set and a test set; Step 9: constructing a machining feature classification and recognition model based on deep network learning, training and testing the machining feature classification and recognition model by using the training set and the test set, obtaining a trained machining feature classification and recognition model, and the machining feature classification and recognition model comprises an input end, a backbone feature extraction network, a reinforced feature fusion network and an output end, wherein the backbone feature extraction network adopts a Backbone structure to extract three feature maps, the reinforced feature fusion network comprises two feature pyramid structures of FPN and PAN, and the three input feature maps are fused to obtain three feature maps of different sizes, and the output end performs convolution operations on the three feature maps of different sizes to output a prediction result; Step 10: identifying the two-dimensional color machining feature image corresponding to the part model to be identified by using the trained machining feature classification and recognition model, and obtaining a machining feature classification and recognition result. 2.The deep learning-based machining feature classification and identification method according to claim 1, characterized in that, When the intersection edge between the two adjacent surfaces is a straight line edge, the concave-convex property judgment method is as follows: Suppose the two intersecting faces are face F1 and face F2, traverse the memory working format file to find the advanced face ADVANCED_FACE corresponding to face F1 and face F2, obtain the direction vector given by the second attribute direction DIRECTION of the axis two orientation AXIS2_PLACEMENT_3D of the three-dimensional space of face F1 and face F2, determine the outer normal vector of the face according to the direction identification Boolean value of face F1 and face F2, if the Boolean value is true, the direction vector given by the attribute direction DIRECTION is the outer normal vector of the face, otherwise the outer normal vector of the face is reversed, obtain the outer normal vector N1 of face F1 and the outer normal vector N2 of face F2; Taking the face F1 as a reference face, a face boundary FACE_OUTER_BOUND or FACE_BOUND constituting the face F1 is found, a side loop EDGE_LOOP thereof is acquired, a directed side ORIENTED_EDGE corresponding to a side e thereof is obtained, a side curve EDGE_CURVE is obtained, and finally a direction vector corresponding to a property direction DIRECTION of a geometric definition curve LINE of the side e is obtained. When the direction identification Boolean values of the sides are all true, the direction vector corresponding to the property direction DIRECTION is the direction vector N of the side e e , and vice versa. Calculate the direction vector N e The vector product of the outward normal vector N2 and the outward normal vector N2 yields the direction vector N = N. e ×N2, the concavity / convexity of edge e is determined by the angle θ between the direction vector N and the external normal vector N1: when |θ|<π, edge e is concave; when |θ|>π, edge e is convex. 3.The deep learning-based machining feature classification and recognition method according to claim 2, characterized in that, When the intersecting edge of the two adjacent faces is a circular edge, the concave-convex determination method is as follows: Suppose the two intersecting faces are face F1 and face F2, traverse the memory working format file to find the advanced face ADVANCED_FACE corresponding to face F1, obtain the direction vector given by the second attribute direction DIRECTION of the axis two orientation AXIS2_PLACEMENT_3D of the three-dimensional space of face F1, determine the outer normal vector of the face according to the direction identification Boolean value of face F1, if the Boolean value is true, the direction vector given by the attribute direction DIRECTION is the outer normal vector of the face, otherwise the outer normal vector of the face is reversed, obtain the outer normal vector N1 of face F1; The external normal vector N2 of surface F2 is calculated through the following steps: First, obtain the coordinates of the first attribute vertex VERTEX_POINT in the edge curve EDGE_CURVE of the geometric definition curve CIRCLE of surface F1, denoted as point P. Then, find the center coordinates of the first attribute Cartesian coordinate point CARTESIAN_POINT in the AXIS2_PLACEMENT_3D of the geometric definition curve CIRCLE, denoted as point P0. The vector P is formed by points P and P0. V Its direction is from point P to point P0; obtain the direction vector V corresponding to the second attribute DIRECTION in the geometric definition curve CIRCLE of face F2; calculate vector P V The vector product of the direction vector V and the vector R is obtained. V =P V ×V; then calculate the external normal vector N2 of surface F2 based on whether it is the outer surface of the cylindrical surface that constitutes the solid surface of the part. If it is the outer surface, then the external normal vector N2 = R. V ×V, otherwise the outer normal vector N2 = V×R V ; Calculate the direction vector N of edge e e =P V ×N1, then calculate the direction vector N e The vector product of the outward normal vector N2 and the outward normal vector N2 yields the direction vector N = N. e ×N2, the concavity / convexity of edge e is determined by the angle θ between the direction vector N and the external normal vector N1: when |θ|<π, edge e is concave; when |θ|>π, edge e is convex. 4.The deep learning-based machining feature classification and identification method according to claim 1, characterized in that, The process of constructing the attribute adjacency graph of the part model includes the following steps: Traverse the memory working format file to obtain all advanced face ADVANCED_FACE rows, find the corresponding geometric definition surface, judge whether each two faces intersect, and construct the attribute adjacency graph combined with the concave-convex determination result of the intersecting edge of the two adjacent faces, which is defined as: G={F, E, A}, wherein F represents the node in the attribute adjacency graph G, that is, the corresponding face in the part model; E represents the connection between two nodes in the attribute adjacency graph G, that is, the common edge of the two adjacent faces; A represents the concave-convex property of the edge in the attribute adjacency graph G. 5.The deep learning-based machining feature classification and recognition method according to claim 1, characterized in that, Step 4 includes the following steps: Traverse all face nodes in the attribute adjacency graph, find the feature base face node and the face node adjacent to the feature base face node and having the same characteristic, extract the machining feature subgraph containing the feature base face and the corresponding machining feature adjacency submatrix; Continue to traverse, find all main shaft rotation face nodes, delete such face nodes, obtain the machining feature subgraph without feature base face and the corresponding machining feature adjacency submatrix. 6.The deep learning-based machining feature classification and identification method according to claim 1, characterized in that, The pre-defined feature library includes machining features and corresponding machining feature models, attribute adjacency graphs and extended attribute adjacency matrices. The pre-defined feature library includes machining features and corresponding machining feature models, attribute adjacency graphs and extended attribute adjacency matrices.

Images