A method for identifying and extracting geometric features of a complex thin-walled part

Abstract

The application discloses a kind of complex thin-walled part geometric feature identification and extraction method, comprising: the voxelization processing after pre-processing part model and respectively extracting the multi-view and three-dimensional point cloud modal of part geometric feature model;Extract the feature vector of the voxelization modal, multi-view modal and three-dimensional point cloud modal of complex thin-walled part model, and the feature vector of the three modal extracted is fused;Based on part geometric feature library, the input complex thin-walled part model is identified using navigable small world network algorithm similar geometric feature;In all part geometric feature models in part geometric feature library, obtain the geometric feature model similar to the feature possessed by the input complex thin-walled part model, output and extract the ID number of the highest similarity part geometric feature.The application can identify the similar part geometric feature of the input complex thin-walled part model, improve the reuse rate of existing part geometric feature for operator.

Description

Technical Field

[0001] This invention relates to the field of complex thin-walled parts and the field of geometric feature recognition of parts, specifically to a method for geometric feature recognition and extraction of complex thin-walled parts. Background Technology

[0002] In modern manufacturing, particularly in aerospace and precision instrumentation, complex thin-walled parts have attracted significant attention due to their unique structural characteristics and wide range of applications. These parts typically exhibit diverse product categories and structural variations, with complex and variable geometries that place extremely high demands on machining accuracy and manufacturing processes. Therefore, efficiently and accurately identifying and extracting the geometric features of these complex thin-walled parts has become a key technical challenge in design, manufacturing, and quality control.

[0003] Traditional geometric feature recognition methods are mostly used for simple, regular shapes, such as cylinders and planes, and have limited ability to handle complex free-form surfaces and thin-walled structures. With the rapid development of computer-aided design (CAD) and computer graphics, model feature recognition methods have gradually emerged, but they still suffer from problems such as insufficient recognition accuracy and low efficiency when faced with highly complex and irregular geometric shapes.

[0004] In recent years, with the innovative design and precision manufacturing of numerous high-precision aerospace vehicles, effectively reusing the geometric features of diverse parts and ensuring that design engineers can quickly and accurately locate the required features have become key to improving R&D efficiency and innovation capabilities. Against this backdrop, the revolutionary leap in artificial intelligence and machine learning technologies has opened up new avenues for solving the challenges of geometric feature recognition and extraction for complex thin-walled parts.

[0005] The invention disclosed in CN118351127A presents a 3D point cloud segmentation method for production line part identification, comprising the following steps: For a specific production line to be identified, a VFH sample library containing point clouds of all parts in the production line scene is pre-constructed; 2D image data of the production line is collected, and deep learning-based target detection is performed to obtain the region of interest, which is then reconstructed to obtain a 3D point cloud and preprocessed; the preprocessed production line point cloud is input into a supervoxel clustering algorithm incorporating edge information to generate supervoxels; the supervoxels are input into a region growing segmentation algorithm based on point cloud color and concavity / convexity features to segment the production line part point cloud; the segmented production line part point cloud is matched with the production line part point cloud in the VFH sample library to achieve production line part point cloud identification. This method can generate supervoxels with better boundaries attached to the object boundaries, and the region growing algorithm performs better, proposing a technical route for production line part identification. Summary of the Invention

[0006] The purpose of this invention is to disclose a method for identifying and extracting geometric features of complex thin-walled parts. This method enables the reuse of existing part geometric feature models and the rapid identification and extraction of similar part geometric feature models. It can identify the geometric features of similar parts to the input complex thin-walled part model and push the geometric feature models of parts with high similarity to the operator. This improves the operator's reuse rate of existing part geometric features and provides important data support for subsequent process optimization design.

[0007] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for identifying and extracting geometric features of complex thin-walled parts, comprising the following steps:

[0009] S1, preprocess the input complex thin-walled part model and the existing part geometric feature model; specifically, classify the part geometric features according to their function and structure in the part, encode all the classified part geometric features to form their own unique ID number, and integrate the encoded part geometric features to form a part geometric feature library.

[0010] S2, the preprocessed part model is voxelized and the multi-view and 3D point cloud modes of the part's geometric feature model are extracted respectively; specifically, the preprocessed complex thin-walled part model is voxelized to obtain the voxelized mode of the complex thin-walled part model; projection viewpoints are set from different angles, and the multi-view mode of the complex thin-walled part model is obtained by using the projection method; then the part model format is converted by 3D modeling software, and the 3D point cloud mode of the complex thin-walled part model is obtained by using Open3D.

[0011] S3: Extract feature vectors from the voxelized mode, multi-view mode, and 3D point cloud mode of the complex thin-walled part model, and fuse the feature vectors of the three modes.

[0012] S4, based on the part geometric feature library, uses the navigable small-world network algorithm to identify similar geometric features of the input complex thin-walled part model;

[0013] S5 iterates through all the geometric feature models of the parts in the geometric feature library, obtains the geometric feature models that are similar to the features of the input complex thin-walled part model, outputs and extracts the ID number of the geometric feature with the highest similarity.

[0014] Step S1 further includes:

[0015] Step S11 involves standardizing the preprocessed complex thin-walled part model, including the following sub-steps:

[0016] Translating a complex thin-walled part model to place the center of gravity of the complex thin-walled part model at the origin of an absolute coordinate system; rotating and flipping the complex thin-walled part model at different angles centered on the center of gravity of the part model to align the complex thin-walled part model in a standard coordinate plane; scaling the complex thin-walled part model in equal proportion and normalizing the complex thin-walled part model so that the dimensions of the complex thin-walled part model are all in standard unit sizes; the normalization process of the complex thin-walled part model is defined as:

[0017] ω(x, y, z)=a -1 ·D·R·(I-m p )

[0018] wherein, a is the size of the scaling coefficient of the complex thin-walled part model, D represents a flip-invariant diagonal matrix, R represents a rotation matrix after rotation transformation, I represents an identity matrix, m p is the center of gravity of the complex thin-walled part model;

[0019] Step S12: For the characteristic structures of the geometric features of complex thin-walled parts in the part, construct classification rules for part geometric features based on the part geometric features carried in the processing of solid parts, encode and classify the part geometric features, and combine them to form a part geometric feature library.

[0020] Further, in step S2, the process of performing voxelization processing on the preprocessed complex thin-walled part model comprises:

[0021] Converting a continuous three-dimensional space R 3 into a discrete space D in the form of partitioning 3 , in the discrete space, each voxel is a cube, and a voxel point (x, y,z) corresponds to all points in a continuous space (u, v, w) in the original three-dimensional coordinate system, wherein x-1<u≤x, y-1<v≤y, z-1<w≤z, so as to obtain the spatial distribution information of a three-dimensional model, each voxel is assigned a binary value, and 0 or 1 is used to mark whether the voxel is occupied by the model;

[0022] Presetting the width of a single voxel, taking the size R of the voxel space as the voxelization granularity, a larger R indicates a finer and more realistic model, and a smaller R indicates a coarser model;

[0023] Using the selected voxelization granularity R value to amplify the model in the original unit space;

[0024] Mapping continuous points (x, y, z) on the three-dimensional mesh model to voxels (x′, y′, z′) according to the size of each voxel;

[0025] For edges of different lengths, a case-by-case strategy is adopted: If the length l of the edge is less than the preset voxel width, when both endpoints of the edge fall within the same voxel, only that voxel is added to the edge voxel set; if the two endpoints are distributed in two adjacent voxels, both voxels are added to the set; if the length l of the edge exceeds the preset voxel width, the edge is divided into n equal parts. l Segment, ensuring that the length of each segment is no greater than the voxel width, then perform voxelization on the subdivided segments, and summarize all relevant voxel points into the edge voxel set;

[0026] For each triangular facet, two strategies are adopted based on the relationship between the facet's edge length and the voxel width: If the lengths of all three edges of the facet do not exceed the voxel width, all vertices constituting the facet are directly voxelized, the resulting voxel points are collected and added to the facet voxel set; If at least one edge of the facet exceeds the voxel width, starting from the longest edge, the scanline algorithm is used to advance along the longest edge and gradually penetrate into the facet along a direction parallel to another set of edges, and then the edge voxelization method is used for voxelization. All the points obtained in the end are added to the facet voxel set and deduplication is performed.

[0027] Furthermore, in step S2, the process of obtaining the multi-view modalities of a complex thin-walled part model using the projection method includes:

[0028] With the center of gravity of the complex thin-walled part model as the origin, multiple viewpoints are set. The viewpoints are located at the center of the six faces of the complex thin-walled part model: top, bottom, left, right, front, and back. Two auxiliary viewpoints are set on both sides of the diagonal.

[0029] Multiple viewpoints are set for the model of a complex thin-walled part. The projection view method is used to project the model of the complex thin-walled part to obtain a multi-view two-dimensional view of the model of the complex thin-walled part.

[0030] Furthermore, in step S2, the process of obtaining the 3D point cloud mode of the complex thin-walled part model using Open3D includes the following steps:

[0031] The input complex thin-walled part model was converted into an .obj file format using the 3D modeling software UG. Open3D was used to read the .obj file of the mesh model. The 3D coordinate position information of all mesh vertices in the mesh 3D model was obtained through the vertices array. Point cloud data was obtained by downsampling.

[0032] Create a new PointCloud object, fill it with the 3D coordinates of all mesh vertices, save the converted point cloud data to a file, and output it in pcd file format.

[0033] Furthermore, in step S3, the process of extracting the feature vectors of the voxelized modes of the complex thin-walled part model includes the following steps:

[0034] The voxelized modalities are input into the VoxNet network. After passing through the first 3D convolutional layer, voxel features are extracted and a feature map is output. Then, the voxel features are reduced in dimensionality and key information is preserved by the pooling layer. By repeatedly stacking convolutional and pooling layers, the voxel feature extraction is deepened and the feature dimensionality is reduced. Finally, the voxel features are integrated and nonlinearly transformed by a fully connected layer with ReLU as the activation function, and the final voxel feature vector is output through the output layer.

[0035] Furthermore, in step S3, the process of extracting the feature vectors of the multi-view modalities of the complex thin-walled part model includes the following steps:

[0036] Construct a multi-channel input convolutional neural network to extract multiple 2D images. Figure 1 The image is input into multiple channels in a corresponding manner. The image enters the convolutional layer from the input layer and passes through an activation function to obtain the feature map output value. The formula for the obtained feature map is:

[0037] x l =f(w l x l +b l )

[0038] Where l represents the layer number, ω represents the weight, b represents the offset, and f represents the activation function;

[0039] Pooling layers and convolutional layers are connected adjacently to reduce the dimensionality of the feature map. The pooling layer is calculated as follows:

[0040]

[0041] In the formula, x j l Let f represent the output value of the j-th feature map in the l-th layer, and let f represent the activation function, β. j l `down` represents the downsampling function, `down` represents the downsampling operation, and `x` represents the downsampling function. j l-1 b represents the output value of the j-th feature map in the (l-1)-th layer. j l This represents the bias value of the j-th feature map in the l-th layer;

[0042] By stacking multiple convolutional and pooling layers, multi-view feature vectors of the part model are obtained.

[0043] Furthermore, in step S3, the process of extracting the feature vectors of the three-dimensional point cloud modes of the complex thin-walled part model includes the following steps:

[0044] The Pointconv network is used to extract feature vectors from the point cloud of the model. The input set of 3D point cloud data of the part is {p i |i=1,...n}, where each point contains a position vector (x, y, z) and its features, and n represents the number of points in the point cloud data. The calculation formula for the PointConv network is as follows:

[0045]

[0046] In the formula, F(x+δ) x y+δ y ,z+δ z ) is a characteristic of a point centered at point p = (x, y, z) in a local region, (δ) x δ y δ z W(δ) represents the coordinates of any location within a local region. x δ y δ z ) is the coordinate (δ) x δ y δ z The approximate weight function of ), S(δ) x δ y δ z ) is a point (δ) x δ y δ z The inverse density of )

[0047] By stacking multiple layers of PointConv networks, the feature vectors of the 3D point cloud of the part are extracted.

[0048] Furthermore, in step S3, the Concat fusion method is used to connect the feature vectors of the three extracted modalities along the same dimension to form a feature vector with a higher dimension. The calculation formula is as follows:

[0049]

[0050] In the formula, the number of feature channels used for fusion is X voxel features. i ={X1, X2, ..., X s}, Multi-view feature Y i ={Y1, Y2, ..., Y s} and 3D point cloud features Z i ={Z1, Z2, ..., Z s}, where s represents the arbitrary number of feature channels, and C i This represents the convolution operation on the i-th voxel.

[0051] Step S4 further includes:

[0052] Given a feature element q, a multi-layer HNSW graph structure, and the number of nearest neighbors to be found k, set a dynamic candidate list ef to control the accuracy of the search; set an empty list W to store the closest feature element found so far, and set the entry point ep for the start of the search.

[0053] During the search, the search proceeds from the highest level downwards. A search is performed at each level using a specified level query to find the feature element closest to the query point q and update the entry point ep to the feature element in the search results that is closest to q. When the search reaches the bottom level, the search stops and the K closest feature elements in the bottom search results are returned.

[0054] Returns a list W containing the K nearest neighbor features of the input feature element q.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] First, the method for identifying and extracting geometric features of complex thin-walled parts according to the present invention achieves the identification and extraction of geometric feature models of complex thin-walled parts through a practical and effective method. The retrieval method mainly targets the geometric features of complex thin-walled parts, which have a rich variety of geometric features that are easy to classify. The present invention provides more accurate identification of geometric feature models of complex thin-walled parts and also allows designers to reuse existing geometric feature models of parts.

[0057] Secondly, the method for identifying and extracting geometric features of complex thin-walled parts in this invention preprocesses the input complex thin-walled part model to ensure that the relative positions of all models in space are consistent, which also shortens the time for subsequent model voxelization and extraction of multiple views and 3D point clouds of the part model.

[0058] Third, the method for identifying and extracting geometric features of complex thin-walled parts of the present invention establishes classification rules for adjusting the geometric structure of parts by utilizing the structural position of the geometric features of parts in the part model, clarifies the classification criteria for the geometric features of parts, constructs a geometric feature library of parts, improves the utilization rate of existing geometric features of parts, shortens the design time and design cost for designers, and improves the accuracy of the identification of geometric features of parts and the efficiency of subsequent part processing and production.

[0059] Fourth, the method for identifying and extracting geometric features of complex thin-walled parts in this invention employs the advanced Navigable Small World (HNSW) algorithm during part geometric feature identification. Due to its multi-layered navigation graph structure, it enables rapid identification of part geometric features. During the algorithm's identification process, a heuristic search strategy is used to achieve efficient searching and identification of part geometric features.

[0060] Fifth, the method for identifying and extracting geometric features of complex thin-walled parts of the present invention identifies the geometric features of the parts through the advanced HNSW algorithm, and finally outputs and extracts the geometric feature code ID number of the part with the highest similarity, providing operators with a more accurate geometric feature model of similar parts, which not only improves the utilization rate of the geometric features of the parts, but also reduces the cost of designing new parts. Attached Figure Description

[0061] Figure 1 This is a flowchart of the method for identifying and extracting geometric features of complex thin-walled parts according to an embodiment of the present invention;

[0062] Figure 2 This is a schematic diagram illustrating the classification of the part geometric feature library provided in an embodiment of the present invention. Detailed Implementation

[0063] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0064] like Figure 1 and Figure 2 As shown, the method for identifying and extracting geometric features of parts according to the present invention includes:

[0065] Step 1: Preprocess the input complex thin-walled part model and existing part geometric feature models, understand the function and structure of the complex thin-walled part geometric features in the part, construct a part geometric feature library, and encode all part geometric features.

[0066] First, the complex thin-walled part model is preprocessed. Due to differences in designers, even the same part model can have multiple output forms in different coordinate systems due to variations in spatial location and angles. The specific preprocessing steps are as follows:

[0067] 1. Translation Process: To translate the part model, the position of the origin of the part model needs to be normalized. In the coordinate system (x, y, z), let the new part model be A(x, y, z), and the translated model be I(x, y, z), whose expression is:

[0068] I(x, y, z) = A(xm) p ym p zm p )

[0069] In the formula, m p This is the center of gravity of the part's 3D model.

[0070] 2. Rotation Processing: Rotation processing typically involves rotating the part model around its center of gravity at different angles to ensure that the model maintains its inherent geometric properties and structural integrity at all angles. Its main purpose is to correct deviations in the model at different angles. During the rotation process, the covariance matrix C of the point set needs to be calculated. p This is to achieve the final rotation processing of the part model. Its expression is:

[0071]

[0072] In the formula, p i p j s represents the centroid of each triangular mesh. i s j These represent the area of ​​each triangular grid.

[0073] 3. Scaling Processing: Scaling processing refers to scaling the 3D model of the part proportionally in the 3D coordinate system to ensure that the part model can be scaled to a uniform scale without any distortion. Let the part model A(x, y, z) of size L×W×H be transformed into I(x, y, z) as shown in formula (4):

[0074] I(x,y,z)=A(int(a×x), int(a×y), int(a×z))

[0075] In the formula, int represents the floor function, and the scaling factor a = 1 / n. When the scaling factor n > 1, the three-dimensional model of the part is reduced in size; when the scaling factor n < 1, the three-dimensional model of the part is enlarged. This allows us to obtain a three-dimensional model with rotation invariance.

[0076] After the above three processes, the preprocessing of the part model is completed. This process is defined as follows:

[0077] ω(x, y, z) = a -1 ·D·R·(Im p )

[0078] Where a is the scaling factor of the 3D model, D represents the diagonal matrix with flip invariance, R represents the rotation matrix after rotation transformation, and m p This is the center of gravity of the part's 3D model.

[0079] Secondly, the functionality of the geometric features of the parts involves classifying and encoding the existing geometric feature models of complex thin-walled parts to form a geometric feature library for the parts.

[0080] Part geometric feature classification such as Figure 2As shown in the present invention, taking frame segments as an example, they are divided into shape units and transition units, and are further classified according to the features carried in the functional features. The shape units are divided into profile surfaces and end surfaces, and the transition units are divided into chamfers and transition fillets. After completing the model classification, all geometric feature models of complex thin-walled parts are encoded. Taking the geometric features of frame segment parts as an example, such as "frame segment-front end face-hole", the feature coding consists of three parts: "part name-feature position-feature name", and then each exclusive ID number is formed, and all coded part geometric feature models are integrated to form a part geometric feature library.

[0081] Step 2: Voxelize the preprocessed part model and extract the multi-view and three-dimensional point cloud modalities of the part geometric feature model respectively

[0082] 1. For the preprocessed part model in step 1, voxelize the part model: convert a continuous three-dimensional space R<3> into a discrete space D in the form of partitioning 3 , assuming that a voxel is a 1×1×1 cube in the discrete space, then the voxel point (x, y,z) corresponds to all points in a continuous space (u, v, w) in the original three-dimensional coordinate system, where x-1<u≤x, y-1<v≤y, z-1<w≤z. After such processing, the spatial distribution information of a three-dimensional model can be obtained. Each voxel is assigned a binary value, and 0 or 1 is used to mark whether the voxel is occupied by the model.

[0083] (1) Voxelization granularity

[0084] SEWR is used to better describe the fineness of the model after voxel processing. In order to prevent the loss of too many three-dimensional spatial features after voxelization of the model, the width of a single voxel is fixed as l=1, so that the size of the voxel space can be changed. In this case, the size R of the voxel space is the voxelization granularity. A larger R indicates that the model is finer and more realistic, while a smaller R indicates that the model is coarser. In order to make the measurement length of the voxel coordinate system exactly correspond to that of the three-dimensional vertex coordinate system, so as to facilitate subsequent voxelization calculation, we amplify the model in the original unit space. Assuming that the granularity R value has been selected, for each vertex v=(x, y, z), we have:

[0085] (2) Vertex voxelization

[0086] It is assumed that the three-dimensional model after coordinate scale normalization can just be enclosed by an L×L×L cubic bounding box. If the cubic bounding box is divided into n×n×n voxels, then the size of each voxel is w×w×w, where w=L / n.

[0087] Therefore, for a continuous point (x, y, z) on the three-dimensional mesh model, the process of mapping it to the voxel (x′, y′, z′) is shown in the following formula.

[0088]

[0089] (3) Edge voxelization

[0090] In the edge voxelization process of the 3D mesh model, a case-by-case handling strategy is adopted for edges of different lengths: If the edge length l is less than the preset voxel width (set to 1), there are two cases: when the two endpoints of the edge fall within the same voxel, only that voxel point is included in the edge voxel set; if the two endpoints are distributed in two adjacent voxels, then both voxel points are added to the set. For the edge length exceeding the voxel width, i.e., l > 1, the edge is divided into n equal parts using the equal division method. l Each segment is segmented to ensure that the length of each segment is no greater than the voxel width. Then, these sub-segments are voxelized according to the first two rules, and all relevant voxel points are aggregated into the edge voxel set.

[0091] (4) Surface voxelization

[0092] When processing the voxelization of each triangular facet in the 3D mesh model, two strategies are mainly adopted based on the relationship between the facet's edge length and the voxel width: If the length of all three edges of the facet does not exceed the voxel width, the facet may completely or partially cover one or several adjacent voxels. In this case, all vertices constituting the facet are directly voxelized, and the resulting voxel points are collected and added to the facet voxel set. Conversely, if at least one edge of the facet exceeds the voxel width, a more refined method is adopted—starting from the longest edge, a scanline algorithm is used to advance along this edge, and then gradually penetrates into the facet along a direction parallel to another set of edges. Then, edge voxelization is used for voxelization. All the points obtained are finally added to the facet voxel set, and deduplication is performed.

[0093] 2. Extract multi-view 2D views of the part model

[0094] Taking a frame segment part as an example, when extracting the frame segment into the 2D view of the part model, eight viewpoints are set with the center of gravity of the part model as the origin. These viewpoints are located at the centers of the top, bottom, left, right, front, and back faces of the part model. To increase the diversity of the view data, two auxiliary viewpoints are set on both sides of the diagonal. Using the eight viewpoints of the part model, the part model is projected using the projection view method to obtain a multi-view 2D view of the part model.

[0095] 3. Extract the 3D point cloud from the part model.

[0096] The input complex thin-walled part model file format is converted into an .obj file format suitable for point cloud processing using the 3D modeling software UG. Open3D is then used to read the .obj file of the mesh model. The 3D coordinates of all mesh vertices are then obtained from the mesh 3D model using the `vertices` array. Due to the complex structure of the thin-walled part and the large amount of mesh vertex data, downsampling is used to reduce data complexity and improve subsequent processing speed when acquiring point cloud data. Next, a new `PointCloud` object is created and filled with the coordinates of all vertices. Finally, the converted point cloud data is saved to a file and output in .pcd file format.

[0097] Step 3: Extract feature vectors from the three modes of the part's geometric feature model, and then fuse the feature vectors of the three modes.

[0098] Taking a frame segment part as an example, the voxelized part model in step two is first input into the VoxNet network. After passing through the first 3D convolutional layer, voxel features are extracted and a feature map is output. Then, it enters the pooling layer to reduce the dimensionality of the voxel features and retain key information. By repeatedly stacking convolutional and pooling layers, the voxel feature extraction is deepened and the feature dimensionality is reduced. Finally, the voxel features are integrated and nonlinearly transformed through a fully connected layer with ReLU as the activation function, and the final voxel feature vector is output through the output layer.

[0099] Secondly, a multi-channel input convolutional neural network is constructed. The eight 2D images extracted in step two are input into the eight channels respectively. The images enter the convolutional layer from the input layer and obtain feature map output values ​​through the activation function. The formula for obtaining the feature map is as follows:

[0100] x l =f(w l x l +b l )

[0101] Where l represents the layer number, ω represents the weight, b represents the offset, and f represents the activation function.

[0102] Connecting pooling layers and convolutional layers adjacently reduces the dimensionality of the feature map. The pooling layer is calculated as follows:

[0103]

[0104] In the formula, β j l down represents the downsampling function, and down represents the downsampling operation.

[0105] By stacking multiple convolutional and pooling layers, multi-view feature vectors of the part model are obtained.

[0106] Then, the 3D point cloud of the model obtained in step two is used to extract feature vectors from the point cloud using the Pointconv network. The input set of 3D point cloud data of the part is {p i |i=1,...n}, where each point contains a position vector (x, y, z) and its features. In continuous space, point cloud data can be regarded as a non-uniform sample. Therefore, PointConv is defined as follows:

[0107]

[0108] In the formula, F(x+δ) x y+δ y ,z+δ z ) is a characteristic of a point centered at point p = (x, y, z) in a local region, (δ) x δ y δ z W(δ) represents the coordinates of any location within a local region. x δ y δ z ) is the coordinate (δ) x δ y δ z The approximate weight function of ), S(δ) x δ y δ z ) is a point (δ) x δ y δ z The inverse density of ).

[0109] Therefore, by stacking multiple layers of PointConv networks, the feature vector of the three-dimensional point cloud of the part is extracted.

[0110] Finally, the Concat fusion method is used to fuse the three feature vectors of the part, connecting them along the same dimension to form a higher-dimensional feature vector. This essentially merges the feature channel counts of the three modal feature vectors of the part's geometric feature model. The calculation formula is as follows:

[0111]

[0112] In the formula, the number of feature channels used for fusion is X voxel features. i ={X1, X2, ..., X s}, Multi-view feature Y i ={Y1, Y2, ..., Y s} and 3D point cloud features Z i ={Z1,Z2,...,Z} s}, where s represents the arbitrary number of feature channels, and C iThis represents the convolution operation on the i-th voxel.

[0113] Step 4: Part geometric feature library. The HNSW algorithm is used to identify similar geometric features in the input complex thin-walled part geometric feature model.

[0114] Taking the frame segment part model as an example, step four fusion obtains the part model feature vector, and the HNSW algorithm is used to identify the similarity of the part geometric features carried by the part model.

[0115] First, input a feature element q, given a multi-level HNSW graph structure, and a number of nearest neighbors k to be found. Set a dynamic candidate list ef to control the precision of the search. Then, set an empty list W to store the closest feature element found so far, and set the entry point ep for the initial search, usually a feature element from the highest level. Next, during the search, search downwards from the highest level, using a SEARCH-LAYER (specified layer query) to perform a search at each level, finding the feature element closest to the query point q, and updating the entry point ep to the feature element in the search results that is closest to q. When the search reaches level zero (the bottom level), stop the search and return the K closest feature elements from the bottom search results. Finally, return a list W containing the K nearest neighbor feature elements of the input feature element q.

[0116] Step 5: Similar part geometric feature models have been identified. Output and extract the geometric feature code ID number of the part with the highest similarity.

[0117] By traversing all the geometric feature models of the parts in the geometric feature library, we obtain similar geometric feature models that have the same features as the input complex thin-walled part model. Finally, we output and extract the ID number of the similar part geometric model.

[0118] The present invention provides a method for identifying and extracting geometric features of complex thin-walled parts. By precisely extracting multi-dimensional information of the geometric features of the parts and using an advanced nearest neighbor search algorithm, similar geometric shapes are quickly and accurately matched in the geometric feature library. This method not only greatly enhances the accuracy and efficiency of feature recognition but also significantly reduces the processing cycle. It lays a solid data foundation for subsequent manufacturing process planning, process strategy optimization, and even the construction of a quality monitoring system, thus accelerating the intelligent transformation and upgrading of the aerospace manufacturing industry.

[0119] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0120] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0121] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0122] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, causing a series of operational steps to be executed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that run on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0123] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0124] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for identifying and extracting geometric features of complex thin-walled parts, characterized in that, The method for identifying and extracting geometric features of complex thin-walled parts includes the following steps: S1, preprocess the input complex thin-walled part model and the existing part geometric feature model; specifically, classify the part geometric features according to their function and structure in the part, encode all the classified part geometric features to form their own unique ID number, and integrate the encoded part geometric features to form a part geometric feature library. S2, the preprocessed part model is voxelized and the multi-view and 3D point cloud modes of the part's geometric feature model are extracted respectively; specifically, the preprocessed complex thin-walled part model is voxelized to obtain the voxelized mode of the complex thin-walled part model; projection viewpoints are set from different angles, and the multi-view mode of the complex thin-walled part model is obtained by using the projection method; then the part model format is converted by 3D modeling software, and the 3D point cloud mode of the complex thin-walled part model is obtained by using Open3D. S3: Extract feature vectors from the voxelized mode, multi-view mode, and 3D point cloud mode of the complex thin-walled part model, and fuse the feature vectors of the three modes. S4, based on the part geometric feature library, uses the navigable small-world network algorithm to identify similar geometric features of the input complex thin-walled part model; S5: Traverse all part geometric feature models in the part geometric feature library, obtain geometric feature models with features similar to the input complex thin-walled part model, output and extract the ID number of the part geometric feature with the highest similarity. Step S4 further includes: Given a feature element q, a multi-layer HNSW graph structure, and the number of nearest neighbors to be found k, set a dynamic candidate list ef to control the accuracy of the search; set an empty list W to store the closest feature element found so far, and set the entry point ep for the start of the search. During the search, the search proceeds from the highest level downwards. A search is performed at each level using a specified level query to find the feature element closest to the query point q and update the entry point ep to the feature element in the search results that is closest to q. When the search reaches the bottom level, the search stops and the K closest feature elements in the bottom search results are returned. Returns a list W containing the K nearest neighbor features of the input feature element q.

2. The method for identifying and extracting geometric features of complex thin-walled parts according to claim 1, characterized in that, Step S1 further includes: Step S11 involves standardizing the preprocessed complex thin-walled part model, including the following sub-steps: The complex thin-walled part model is translated to place its center of gravity at the origin of the absolute coordinate system. Then, the model is rotated and flipped at different angles around its center of gravity to align it in the standard coordinate plane. Finally, the model is scaled proportionally and normalized to ensure all dimensions are in standard units. The standardization process for the complex thin-walled part model is defined as follows: Where 'a' represents the scaling factor of the complex thin-walled part model, 'D' represents the diagonal matrix with flip invariance, 'R' represents the rotation matrix after rotation transformation, 'I' represents the identity matrix, and 'm' represents the model size. p The center of gravity of the complex thin-walled part model; Step S12: For the feature structure of the geometric features of complex thin-walled parts, construct the part geometric feature classification rules based on the part geometric features carried in the machining of solid parts, encode the classified part geometric features and combine them to form a part geometric feature library.

3. The method for identifying and extracting geometric features of complex thin-walled parts according to claim 1, characterized in that, Step S2, the process of voxelizing the preprocessed complex thin-walled part model, includes: Given a continuous three-dimensional space R 3 Transform it into a discrete space D in the form of blocks. 3 In discrete space, each voxel is a cube, and voxel points Corresponding to a continuous space in the original three-dimensional coordinate system All points in the model are used to obtain the spatial distribution information of a three-dimensional model, where xl < u ≤ x, yl < v ≤ y, zl < w ≤ z. Each voxel is assigned a binary value, and 0 or 1 is used to mark whether the voxel is occupied by the model. The width of a single voxel is preset, and the size of the voxel space R is used as the voxelization granularity. The larger the R is, the more refined and realistic the model is, and the smaller the R is, the coarser the model is. The model in the original unit space is magnified using the selected voxelization granularity R value; Based on the size of each voxel, the continuous points (x,y,z) on the 3D mesh model are mapped to voxels (x',y',z'); For edges of different lengths, a case-by-case strategy is adopted: If the length l of the edge is less than the preset voxel width, when both endpoints of the edge fall within the same voxel, only that voxel is added to the edge voxel set; if the two endpoints are distributed in two adjacent voxels, both voxels are added to the set; if the length l of the edge exceeds the preset voxel width, the edge is divided into n equal parts. l Segment, ensuring that the length of each segment is no greater than the voxel width, then perform voxelization on the subdivided segments, and summarize all relevant voxel points into the edge voxel set; For each triangular facet, two strategies are adopted based on the relationship between the facet's edge length and the voxel width: If the lengths of all three edges of the facet do not exceed the voxel width, all vertices constituting the facet are directly voxelized, the resulting voxel points are collected and added to the facet voxel set; If at least one edge of the facet exceeds the voxel width, starting from the longest edge, the scanline algorithm is used to advance along the longest edge and gradually penetrate into the facet along a direction parallel to another set of edges, and then the edge voxelization method is used for voxelization. All the points obtained in the end are added to the facet voxel set and deduplication is performed.

4. The method for identifying and extracting geometric features of complex thin-walled parts according to claim 1, characterized in that, Step S2, the process of obtaining the multi-view modalities of a complex thin-walled part model using the projection method, includes: With the center of gravity of the complex thin-walled part model as the origin, multiple viewpoints are set. The viewpoints are located at the center of the six faces of the complex thin-walled part model: top, bottom, left, right, front, and back. Two auxiliary viewpoints are set on both sides of the diagonal. Multiple viewpoints are set for the model of a complex thin-walled part. The projection view method is used to project the model of the complex thin-walled part to obtain a multi-view two-dimensional view of the model of the complex thin-walled part.

5. The method for identifying and extracting geometric features of complex thin-walled parts according to claim 1, characterized in that, Step S2, the process of obtaining the 3D point cloud mode of a complex thin-walled part model using Open3D includes the following steps: The input complex thin-walled part model was converted into an .obj file format using the 3D modeling software UG. Open3D was used to read the .obj file of the mesh model. The 3D coordinate position information of all mesh vertices in the mesh 3D model was obtained through the vertices array. Point cloud data was obtained by downsampling. Create a new PointCloud object, fill it with the 3D coordinates of all mesh vertices, save the converted point cloud data to a file, and output it in pcd file format.

6. The method for identifying and extracting geometric features of complex thin-walled parts according to claim 1, characterized in that, Step S3, the process of extracting the feature vectors of the voxelized modes of the complex thin-walled part model, includes the following steps: The voxelized modalities are input into the VoxNet network. After passing through the first 3D convolutional layer, voxel features are extracted and a feature map is output. Then, the voxel features are reduced in dimensionality and key information is preserved by the pooling layer. By repeatedly stacking convolutional and pooling layers, the voxel feature extraction is deepened and the feature dimensionality is reduced. Finally, the voxel features are integrated and nonlinearly transformed by a fully connected layer with ReLU as the activation function, and the final voxel feature vector is output through the output layer.

7. The method for identifying and extracting geometric features of complex thin-walled parts according to claim 1, characterized in that, Step S3, the process of extracting the feature vectors of the multi-view modalities of the complex thin-walled part model, includes the following steps: A multi-channel input convolutional neural network is constructed, in which multiple extracted 2D views are input one-to-one into multiple channels. The image enters the convolutional layer from the input layer and obtains the feature map output value through the activation function. The formula for obtaining the feature map is: Where l represents the layer number, ω represents the weight, b represents the offset, and f represents the activation function; Pooling layers and convolutional layers are connected adjacently to reduce the dimensionality of the feature map. The pooling layer is calculated as follows: In the formula, x j l Let f represent the output value of the j-th feature map in the l-th layer, and let f represent the activation function, β. j l `down` represents the downsampling function, `down` represents the downsampling operation, and `x` represents the downsampling function. j l-1 b represents the output value of the j-th feature map in the (l-1)-th layer. j l This represents the bias value of the j-th feature map in the l-th layer; By stacking multiple convolutional and pooling layers, multi-view feature vectors of the part model are obtained.

8. The method for identifying and extracting geometric features of complex thin-walled parts according to claim 1, characterized in that, Step S3, the process of extracting the feature vectors of the 3D point cloud modes of the complex thin-walled part model, includes the following steps: The Pointconv network is used to extract feature vectors from the point cloud of the model. The input set of 3D point cloud data of the part is {p i |i=1,…n}, where each point contains a position vector (x, y, z) and its features, and n represents the number of points in the point cloud data. The calculation formula for the PointConv network is as follows: In the formula, F(x+δ) x , y+δ y , z+δ z ) is a feature of a point centered at point p=(x, y, z) in a local region, (δ) x ,δ y , δ z W(δ) represents the coordinates of any location within a local region. x , δ y , δ z ) is the coordinate (δ) x , δ y , δ z The approximate weight function of ), S(δ) x , δ y , δ z ) is a point (δ) x , δ y , δ z The inverse density of ) By stacking multiple layers of PointConv networks, the feature vectors of the 3D point cloud of the part are extracted.

9. The method for identifying and extracting geometric features of complex thin-walled parts according to claim 1, characterized in that, In step S3, the Concat fusion method is used to connect the feature vectors of the three extracted modalities along the same dimension to form a feature vector with a higher dimension. The calculation formula is as follows: In the formula, the number of feature channels used for fusion is X voxel features. i ={X1,X2,……,X s }, Multi-view feature Y i ={Y1,Y2,……,Y s } and 3D point cloud features Z i ={Z1,Z2,……,Z s }, where s represents the arbitrary number of feature channels, and C i This represents the convolution operation on the i-th voxel.

Images