Metal product modeling method and system based on three-dimensional point cloud

Abstract

The embodiment of the application discloses a kind of metal product modeling method and system based on three-dimensional point cloud, method includes: obtaining the original discrete point cloud of target metal product and carries out point cloud regularization processing;Through global scanning, identify density mutation area and normal vector jump area, carry out regional correlation analysis to generate key structural feature boundary;Local point cloud is divided and structure dependent relationship is mined to determine structure correlation rule;Structure section connection relationship network is constructed and topological structure is created;After topological structure and corresponding discrete point cloud segment are aligned, input graph neural network executes inference analysis, realizes structure section correlation reconstruction based on result and obtains three-dimensional model.So, through the coordination of multi-link technical means, the geometric shape of metal product is accurately reproduced while the internal correlation and topological logic of structure are retained, the accuracy and structural expression integrity of three-dimensional model are improved.

Description

Technical Field

[0001] This application relates to the field of 3D reconstruction technology, and in particular to a method and system for modeling metal products based on 3D point clouds. Background Technology

[0002] Against the backdrop of rapid development in digital transformation and reverse engineering technology in the manufacturing industry, 3D reconstruction technology for metal products has become one of the core technologies driving the development of fields such as intelligent manufacturing, digital preservation of cultural relics, and industrial component inspection. Currently, 3D reconstruction technology based on discrete point clouds is the mainstream research direction in this field. Its core is to acquire discrete point cloud data of the target object through various 3D scanning devices, and then reconstruct the 3D model through a series of algorithm processing methods.

[0003] With the development of artificial intelligence technology, existing technologies have attempted to apply AI neural networks to feature learning of point clouds and 3D model reconstruction. Topology construction technology also has certain applications in the field of 3D reconstruction. It is usually based on extracting key feature points to construct the topological relationship of objects, which helps to realize the structured expression of the model.

[0004] However, the existing technologies mentioned above are difficult to accurately capture the inherent logical relationships between structures during the 3D reconstruction of metal products, which makes it easy for the reconstructed 3D model to lose structural topology information. Furthermore, it is difficult to balance reconstruction accuracy and structural integrity for complex metal products. Summary of the Invention

[0005] This application provides a method and system for modeling metal products based on three-dimensional point clouds.

[0006] This application provides a method for modeling metal products based on 3D point clouds, applied to a metal product modeling system. The method includes:

[0007] The original discrete point cloud of the target metal product is obtained, and point cloud regularization is performed based on the spatial distribution characteristics of the original discrete point cloud to obtain a preprocessed discrete point cloud.

[0008] A global scanning analysis is performed on the preprocessed discrete point cloud to identify target regions with spatial distribution differences in the preprocessed discrete point cloud, and density abrupt change regions and normal vector jump regions are determined through the target regions;

[0009] Based on the density abrupt change region and the normal vector jump region, a regional correlation analysis is performed to extract the first edge point set of the density abrupt change region and the second edge point set of the normal vector jump region. The key structural feature boundary of the target metal product is generated by combining the first edge point set and the second edge point set.

[0010] Based on the key structural feature boundaries, the preprocessed discrete point cloud is divided into multiple local point clouds. By comparing the similarity between the structural features of each local point cloud and analyzing the spatial location correlation, the structural dependency relationship between different structural segments is mined. Based on the structural dependency relationship, the structural association rules of each structural segment in the discrete point cloud are determined.

[0011] Based on the structural association rules and the connection relationships between each structural segment, a structural segment connection relationship network is constructed. The structural features of each local point cloud are mapped to the nodes of the structural segment connection relationship network. The structural association rules are transformed into the association edge attributes between nodes. The topology of the target metal product is created through the collaborative mapping of nodes and association edges.

[0012] The topology and its corresponding discrete point cloud segments are aligned and input into a graph neural network to perform inference analysis based on node features and associated edge attributes. Based on the inference analysis results, the association reconstruction of structural segments is performed to obtain a three-dimensional model of the target metal product.

[0013] One embodiment of this application provides a metal product modeling system, including:

[0014] A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement any of the aforementioned methods for modeling metal products based on three-dimensional point clouds.

[0015] One embodiment of this application provides a readable storage medium storing a program or instructions, which, when executed by a processor, implements the steps of the metal product modeling method based on three-dimensional point clouds.

[0016] Therefore, the embodiments of this application have the following beneficial effects: Starting from the full life cycle processing logic of the discrete point cloud of the target metal product, the embodiments of this application form a complete technical link from point cloud preprocessing, feature boundary extraction, structural association rule mining to topological structure construction and three-dimensional model reconstruction, realizing deep collaboration and global optimization of multiple technical means. By identifying density abrupt change regions and normal vector jump regions through full-domain scanning and conducting regional correlation analysis, the key structural feature boundaries of target metal products can be accurately captured, overcoming the limitations of traditional single feature recognition. By dividing local point clouds by key structural feature boundaries and mining structural dependencies to determine structural correlation rules, a precise analysis of the internal logic of metal product structures is achieved. By constructing a structural segment connection relationship network and creating a topological structure, discrete point cloud features are transformed into topological expressions with structured correlations. Finally, combined with graph neural network inference analysis, structural segment correlation reconstruction is achieved. The generated 3D model can not only accurately reproduce the geometric shape of the target metal product, but also completely preserve the internal correlations and topological logic between its structures, greatly improving the accuracy of the 3D model and the integrity of the structural expression, reducing the information loss rate in the 3D reconstruction process of complex metal products, thus serving applications such as reverse engineering, quality inspection, and digital archiving of metal products. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for modeling metal products based on three-dimensional point clouds, provided as an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of the basic structure of a metal product modeling system provided in an embodiment of this application.

[0020] Figure 3 This is a functional block diagram of a metal product modeling device provided in an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] Please see Figure 1 , Figure 1 This is a flowchart of a metal product modeling method based on 3D point cloud provided in an embodiment of this application. The method can be executed by a metal product modeling system or jointly executed by a metal product modeling system and a server. The method may include steps 110-160.

[0023] In the field of 3D point cloud digital modeling of metal products, technical solutions need to handle various parameters with different physical meanings and dimensions, such as spatial coordinates (length units), point cloud density (number of points per unit volume), unitless normal vectors, and unitless similarity indices generated by algorithms. When these parameters are directly used for calculation, comparison, or fusion, dimensional mismatch problems are prone to occur, resulting in calculation results lacking clear physical meaning and affecting the logical rigor and practical feasibility of the modeling process. Therefore, those skilled in the art can accurately identify such potential dimensional errors and overcome them through adaptive normalization techniques. The core of adaptive normalization is to transform parameters with different dimensions and ranges to a unified scale or dimensionless space, eliminating computational obstacles caused by dimensional differences while preserving the relative relationships and distribution characteristics of the original parameters. For example, through Z-score normalization, minimum-maximum scaling, and other methods, different property indices such as point cloud density and normal vector change rate can be mapped to the same numerical range or standard distribution, ensuring the smooth progress of subsequent similarity comparison, correlation analysis, and neural network inference operations, and guaranteeing parameter comparability and the mathematical validity of calculation steps.

[0024] Based on the technical details of the claims, the issue of dimensions is particularly prominent in key stages, and adaptive normalization can address this issue specifically. When extracting key structural feature boundaries, the quantization methods for density abrupt changes (local point cloud count gradient representation) and normal vector jumps (angle change threshold determination) differ significantly. Direct application can easily lead to inaccurate boundary identification. Technicians can normalize these two into dimensionless intensity coefficients, which are then weighted or logically combined for calculation, achieving robust feature extraction. In structural feature similarity comparisons, features such as spatial arrangement, density distribution, and normal vector patterns have different sources and varying numerical ranges. Directly calculating the matching degree can easily lead to feature amplification or neglect. Normalization processes such as distance value scaling, density statistical value relativization, and normal vector difference cosine similarity conversion can standardize all features to a comparable benchmark, ensuring reliable comparison results. In the graph neural network inference stage, node features and associated edge attributes integrate multi-source heterogeneous data with significant differences in dimensions and distribution. Without preprocessing, model training and inference will be affected. Technicians use feature standardization or embedding layer techniques to project these features into a unified vector space, helping the model learn complex relationships and supporting association reconstruction and model generation.

[0025] Adaptive normalization, which handles parameters with different dimensions to overcome dimensional errors, is a standard practice in this field based on existing technologies. It is a standard data preprocessing step in computer vision, point cloud processing, and machine learning, widely used to eliminate feature scale differences and improve model performance. Numerous academic papers, textbooks, and open-source libraries (such as PointNet's coordinate normalization module and TensorFlow's BatchNorm layer) provide technical support. Technicians can flexibly choose strategies according to the scenario without creative effort. This process not only solves the immediate dimensionality problem but also improves the scalability and maintainability of the modeling system. For example, in structural segment association rule mining and scene-based reconstruction, normalization or its extended cross-dimensional mapping calibration techniques achieve fair and accurate fusion of multi-source data. In summary, technicians can seamlessly embed the aforementioned preprocessing steps to ensure consistent computational dimensions, guarantee the mathematical rationality and physical meaning of the technical solution, improve modeling robustness and accuracy, and achieve efficient and reliable point cloud to 3D model conversion based on existing mature technologies, meeting various needs for digital modeling of metal products.

[0026] The target metal product in this embodiment is a metal flange of a certain model. The original discrete point cloud of the flange was acquired by a three-dimensional laser scanner, which includes point cloud data of the flange body, bolt holes, sealing grooves and other structures.

[0027] Step 110: Obtain the original discrete point cloud of the target metal product, and perform point cloud regularization based on the spatial distribution characteristics of the original discrete point cloud to obtain the preprocessed discrete point cloud.

[0028] In this embodiment, the spatial distribution characteristics of the original discrete point cloud include the density distribution, spatial coordinate range, and discreteness of the point cloud. The point cloud normalization process includes the following operations: First, the original discrete point cloud is denoised using a statistical filtering algorithm to calculate the average distance to the neighboring point clouds of each point cloud. Point clouds with a distance greater than a set threshold are identified as noise points and removed. Second, the denoised point cloud is downsampled using a voxel grid downsampling algorithm to divide the point cloud space into several voxel grids, retaining one representative point cloud within each voxel grid to reduce the number of point clouds. Finally, the downsampled point cloud undergoes coordinate normalization, transforming the spatial coordinates of the point cloud to a unified coordinate system, ensuring that the coordinate range of the point cloud is within a set interval. After the above processing, a preprocessed discrete point cloud is obtained. The number of points in this preprocessed discrete point cloud is reduced to approximately 30% of the original discrete point cloud, and the density distribution and coordinate range of the point cloud are uniform.

[0029] Step 120: Perform a global scan analysis on the preprocessed discrete point cloud to identify target regions with spatial distribution differences in the preprocessed discrete point cloud, and determine the density abrupt change regions and normal vector jump regions through the target regions.

[0030] In this embodiment, the global scanning analysis employs a sliding window scanning algorithm, with the size of the sliding window set according to the point cloud density of the preprocessed discrete point cloud. For each point cloud within a sliding window, its density value is calculated as the ratio of the number of points within the window to the window volume. By comparing the density values ​​within adjacent sliding windows, regions where the density value change exceeds a set threshold are identified as density abrupt change regions. Simultaneously, the normal vector of each point cloud is calculated using a principal component analysis algorithm. By calculating the covariance matrix of the point cloud's neighborhood, the eigenvectors of the covariance matrix are obtained, and the eigenvector corresponding to the smallest eigenvalue is the normal vector of that point cloud. By comparing the angle between the normal vectors of adjacent point clouds, regions where the angle between the normal vectors exceeds a set threshold are identified as normal vector jump regions. For example, at the junction of the flange body and the sealing groove, the density value of the point cloud suddenly drops from high density in the body area to low density in the sealing groove area, and this area is identified as a density abrupt change area; at the edge of the bolt hole of the flange, the normal vector of the point cloud suddenly changes from perpendicular to the surface of the body to perpendicular to the inner wall surface of the bolt hole, and this area is identified as a normal vector jump area.

[0031] Step 130: Perform regional correlation analysis based on density abrupt change regions and normal vector jump regions, extract the first edge point set of the density abrupt change regions and the second edge point set of the normal vector jump regions, and combine the first edge point set and the second edge point set to generate the key structural feature boundary of the target metal product.

[0032] In this embodiment of the application, the regional association analysis includes the following sub-steps:

[0033] Step 131: Perform spatial overlap analysis on the density abrupt change regions and normal vector jump regions to identify overlapping regions and independent regions that do not overlap. The spatial overlap analysis uses a spatial intersection calculation method, comparing the spatial coordinate ranges of the density abrupt change regions and the normal vector jump regions. If the spatial coordinate ranges of the two regions intersect, they are determined to be overlapping regions; otherwise, they are determined to be independent regions. For example, the bolt hole edge region of a flange exhibits both density abrupt changes and normal vector jumps, and this region is determined to be an overlapping region; the flange body edge region only exhibits density abrupt changes and is determined to be an independent density abrupt change region; the bottom region of the flange sealing groove only exhibits normal vector jumps and is determined to be an independent normal vector jump region.

[0034] Step 132: For the overlapping region, extract the point cloud that simultaneously satisfies the density abrupt change characteristic and the normal vector jump characteristic within the overlapping region as the core edge point. Traverse the neighboring point cloud around the core edge point, retain the point cloud with continuous abrupt change trend in the neighborhood, and form the fused edge point set corresponding to the overlapping region.

[0035] In this embodiment, the core edge points are extracted as follows: within the overlapping area, point clouds with both density value change rate and normal vector angle change rate greater than a set threshold are selected as core edge points. The traversal of the neighboring point clouds uses the K-nearest neighbor algorithm, with the K value set according to the point cloud density. The method for determining continuous abrupt change trends is: the density value change rate and normal vector angle change rate of the neighboring point clouds remain within a set range, and the change directions are consistent. For example, in the overlapping area of ​​the bolt hole edges of a flange, point clouds with a density value change rate greater than 0.5 and a normal vector angle change rate greater than 30 degrees are selected as core edge points. Around each core edge point, its 50 neighboring point clouds are traversed, and point clouds with a density value change rate between 0.4 and 0.6, a normal vector angle change rate between 25 and 35 degrees, and consistent change directions are retained to form a fused edge point set.

[0036] Step 133: For the density abrupt change region in the independent region, extract the boundary point cloud along the gradient direction of the density abrupt change, and remove the discretely distributed isolated point cloud in the boundary point cloud to obtain the first edge point set; for the normal vector jump region in the independent region, extract the boundary point cloud along the turning direction of the normal vector change, and remove the discretely distributed isolated point cloud in the boundary point cloud to obtain the second edge point set.

[0037] In this embodiment, the gradient direction of density abrupt changes is determined by calculating the gradient vector of the density value, and the direction of the gradient vector points in the direction of decreasing density value. The boundary point cloud is extracted by selecting the point cloud with the largest rate of change of density value along the gradient direction as the boundary point cloud. Isolated point clouds are removed using a connected component analysis algorithm, calculating the number of neighboring point clouds for each point cloud, and identifying and removing point clouds with fewer than a set threshold of neighboring point clouds as isolated point clouds. For example, in the independent density abrupt change region at the edge of the flange, boundary point clouds are extracted along the gradient direction of decreasing density value, and isolated point clouds with fewer than 10 neighboring point clouds are removed to obtain the first edge point set; in the independent normal vector jump region at the bottom of the flange's sealing groove, boundary point clouds are extracted along the turning direction of the normal vector change, and isolated point clouds with fewer than 8 neighboring point clouds are removed to obtain the second edge point set.

[0038] Step 134: The merged edge point set, the first edge point set, and the second edge point set are spliced ​​together in spatial order to obtain the spliced ​​point cloud.

[0039] In this embodiment, the spatial order is determined as follows: taking the center of the flange as the origin, the edge point sets are sorted clockwise, based on the polar angle of the center coordinates of the edge point sets. The splicing operation uses a point cloud splicing algorithm to connect the sorted edge point sets sequentially, forming a continuous point cloud sequence. For example, the fused edge point set (bolt hole edge), the first edge point set (panel edge), and the second edge point set (bottom of the sealing groove) are sorted clockwise and then spliced ​​to obtain a spliced ​​point cloud.

[0040] Step 135: Based on the spatial distribution trajectory fitting of the splicing point cloud, a continuous curved boundary is generated. The key structural feature boundary of the target metal product is formed by the enclosing of the curved boundary. The spatial position information of the edge point set, the first edge point set and the second edge point set are associated and fused.

[0041] In this embodiment, the spatial distribution trajectory fitting employs a B-spline curve fitting algorithm, using the splicing point agglomeration as control points to generate a continuous B-spline curve. The curve boundary enclosure operation uses a polygon approximation algorithm, discretizing the fitted B-spline curve into several line segments to form a closed polygon boundary. For example, B-spline curve fitting is performed on the splicing point agglomeration to generate continuous curve boundaries for the flange body edge, bolt hole edge, and sealing groove bottom. These curve boundaries enclose the key structural feature boundaries of the flange, which are associated with the spatial location information of each edge point set, such as the center coordinates of the bolt holes and the depth of the sealing groove.

[0042] Step 140: Divide the preprocessed discrete point cloud into multiple local point clouds based on the boundaries of key structural features. By comparing the similarity between the structural features of each local point cloud and analyzing the spatial location correlation, the structural dependency relationship between different structural segments is mined. Based on the structural dependency relationship, the structural association rules of each structural segment in the discrete point cloud are determined.

[0043] In this embodiment, the key structural feature boundaries divide the preprocessed discrete point cloud into multiple local point clouds, such as the disk local point cloud, bolt hole local point cloud, and sealing groove local point cloud. The following is a detailed description of each sub-step of this step:

[0044] Step 141: Using the key structural feature boundary as the segmentation benchmark, the preprocessed discrete point cloud is segmented into multiple non-overlapping point cloud segments. Each point cloud segment corresponds to a local point cloud, and the spatial location range of each local point cloud is recorded.

[0045] In this embodiment, the segmentation operation employs a region growing algorithm. Using point clouds on the boundaries of key structural features as seed points, the algorithm grows outwards, grouping point clouds belonging to the same structural segment together to form point cloud fragments. The spatial location range is recorded by recording the minimum and maximum spatial coordinate values ​​of each local point cloud. For example, the spatial location range of the local point cloud on the disk is X-coordinate 0-100mm, Y-coordinate 0-100mm, and Z-coordinate 0-20mm; the spatial location range of the local point cloud on the bolt hole is X-coordinate 10-20mm, Y-coordinate 10-20mm, and Z-coordinate 0-20mm.

[0046] Step 142: Extract the structural features of each local point cloud. The structural features include the spatial arrangement of the point cloud, the density distribution trend of the local region, and the distribution law of the normal vector.

[0047] In this embodiment, the spatial arrangement morphology is extracted using principal component analysis (PCA) to calculate the covariance matrix of the local point cloud, obtaining the eigenvalues ​​and eigenvectors of the covariance matrix. The direction of the eigenvectors represents the spatial arrangement direction of the local point cloud, and the magnitude of the eigenvalues ​​represents the distribution range of the local point cloud in that direction. The density distribution trend is extracted using kernel density estimation to calculate the density distribution function of the local point cloud, obtaining the peak position and distribution range of the density distribution. The normal vector distribution pattern is extracted using statistical analysis to calculate the histogram of the direction distribution of the normal vectors of the local point cloud, obtaining the main distribution direction and distribution range of the normal vectors. For example, the spatial arrangement of the local point cloud of the disk is planar, the density distribution trend is uniform, and the normal vector distribution pattern is perpendicular to the disk surface; the spatial arrangement of the local point cloud of the bolt hole is cylindrical, the density distribution trend is uniformly distributed along the circumference of the cylinder, and the normal vector distribution pattern is perpendicular to the inner wall surface of the bolt hole.

[0048] Step 143: Perform similarity comparison on the structural features of any two local point clouds, calculate the matching degree of the structural features, and select local point cloud pairs that meet the set matching degree and mark them as similar structural point cloud pairs.

[0049] In this embodiment, the matching degree of structural features is calculated using a weighted summation method, with the weight coefficients set according to the importance of the structural features. Specifically, the calculation method is as follows: First, the matching degree of the spatial arrangement is calculated using the cosine of the angle between the feature vectors; second, the matching degree of the density distribution trend is calculated using the cross-entropy of the density distribution function; finally, the matching degree of the normal vector distribution pattern is calculated using the Bartholomew's distance from the histogram of the normal vector direction distribution. These three matching degrees are then weighted and summed according to the set weight coefficients to obtain the total matching degree of the structural features. A matching degree threshold of 0.8 is set, and local point cloud pairs with a total matching degree greater than 0.8 are marked as similar structural point cloud pairs. For example, the structural feature matching degrees of multiple bolt hole local point clouds of a flange are all greater than 0.8, and they are marked as similar structural point cloud pairs.

[0050] Step 144: Based on the spatial location correlation of all local point clouds, determine the spatial adjacency, inclusion, and parallel relationships of each local point cloud, and generate a spatial location correlation diagram.

[0051] In this embodiment, the spatial position correlation is determined using a spatial distance calculation method. The distance between the center coordinates of any two local point clouds is calculated, and the adjacency relationship is determined based on the distance. The inclusion relationship of the spatial coordinate ranges of the local point clouds is also calculated to determine the inclusion relationship. Furthermore, the parallel relationship is determined based on the structural type and spatial position of the local point clouds. The spatial position correlation graph is generated using a graph structure representation method, with local point clouds as nodes and spatial position relationships as edges to construct the graph structure. For example, if the center distance between the disk's local point cloud and the bolt hole's local point cloud is small, they are determined to be adjacent; if the spatial coordinate range of the bolt hole's local point cloud is contained within the spatial coordinate range of the disk's local point cloud, they are determined to be inclusive; and if multiple bolt hole local point clouds have the same structural type and symmetrical spatial positions, they are determined to be parallel.

[0052] Step 145: Combine the matching results of similar structure point cloud pairs and the spatial location association map to mine the structural dependencies between different structural segments. The structural dependencies include substitution dependencies between similar structures, connection dependencies between adjacent structures, and support dependencies between containing structures.

[0053] In this embodiment, the structural dependency mining employs an association rule mining algorithm. The matching results of similar structural point cloud pairs and the spatial location association map are used as input data to mine the association rules between different structural segments. For example, there is a substitution dependency between the local point clouds of bolt holes in similar structures, meaning the structure of one bolt hole can substitute for the structure of another; there is a connection dependency between the local point clouds of the disk body and the local point clouds of the sealing groove in adjacent structures, meaning the structure of the sealing groove needs to be connected to the structure of the disk body; and there is a support dependency between the local point clouds of the disk body and the local point clouds of the bolt holes containing the structure, meaning the structure of the bolt holes needs the structural support of the disk body.

[0054] Step 146: Based on the structural dependencies obtained from mining, determine the association logic between each structural segment, and transform the association logic into structural association rules. The structural association rules include the association requirements of each structural segment in terms of spatial arrangement and structural form.

[0055] In this embodiment, the determination of association logic employs logical reasoning, deriving the association conditions and results between structural segments based on structural dependencies. The transformation of structural association rules uses a rule representation method, expressing the association logic as an "if-then" rule. For example, the association logic of substitution dependency, "If two bolt holes have similar structures and symmetrical spatial positions, then they can substitute for each other," is transformed into the structural association rule "Bolt holes with similar structures and spatial symmetry can substitute for each other"; the association logic of connection dependency, "If the structure of the sealing groove is connected to the structure of the disc, then the position and size of the sealing groove need to match the structure of the disc," is transformed into the structural association rule "The position and size of the sealing groove need to match the structure of the disc"; the association logic of support dependency, "If the structure of the bolt hole needs the structural support of the disc, then the position and depth of the bolt hole need to be within the structural range of the disc," is transformed into the structural association rule "The position and depth of the bolt hole need to be within the structural range of the disc."

[0056] Step 150: Based on the structural association rules and the connection relationships between each structural segment, construct a structural segment connection relationship network, map the structural features of each local point cloud to the nodes of the structural segment connection relationship network, transform the structural association rules into the association edge attributes between nodes, and create the topology of the target metal product through the collaborative mapping of nodes and association edges.

[0057] In this embodiment, the connection relationships between the structural segments include the connection relationship between the bolt holes and the disc body, the connection relationship between the sealing groove and the disc body, etc. The following is a detailed description of each sub-step of this step:

[0058] Step 151: Determine the connection objects and connection methods of each structural segment based on the connection relationships between each structural segment, and obtain a list of connection relationships.

[0059] In this embodiment, the connection relationships are determined using a structural analysis method. Based on the structural design drawings of the target metal product, the connection objects and connection methods between each structural segment are determined. The connection relationship list is generated in tabular form, recording the name of each structural segment, the name of the connection object, and the connection method. For example, the connection object of the bolt hole is the disc body, and the connection method is embedding; the connection object of the sealing groove is the disc body, and the connection method is integral molding.

[0060] Step 152: Construct network nodes based on each structural segment, and map the structural features of the local point cloud corresponding to each structural segment to the corresponding network node, so that each network node carries its corresponding structural feature information.

[0061] In this embodiment, the network nodes are constructed using node representation methods in graph theory, with each node possessing a unique identifier and attributes. The mapping of structural features utilizes feature vectors, transforming the structural features of the local point cloud into feature vectors, which are then stored in the attributes of the network nodes. For example, the attributes of the network node corresponding to the disk structure segment store feature vectors representing the spatial arrangement, density distribution trend, and normal vector distribution of the local point cloud of the disk; the attributes of the network node corresponding to the bolt hole structure segment store structural feature vectors of the local point cloud of the bolt holes.

[0062] Step 153: Based on the list of connections, construct associated edges between network nodes that have connections, with each associated edge corresponding to a set of connection relationships between structural segments.

[0063] In this embodiment, the construction of associated edges adopts the edge representation method in graph theory, with each edge having a unique identifier and attributes. The correspondence of connection relationships is recorded using the attribute records of the edges, storing the connection objects and connection methods in the attribute records of the associated edges. For example, an associated edge is constructed between a disk network node and a bolt hole network node, with the attribute records that the connection objects are the disk and the bolt hole, and the connection method is "embedded"; an associated edge is constructed between a disk network node and a sealing groove network node, with the attribute records that the connection objects are the disk and the sealing groove, and the connection method is "integral molding".

[0064] Step 154: Convert the structural association rules into corresponding association edge attributes. Different types of structural association rules correspond to different attribute parameters, so that the association edges carry their corresponding structural association rule information.

[0065] In this embodiment, the structural association rules are transformed using a rule encoding method, converting them into numerical attribute parameters and storing them in the attributes of the associated edges. For example, the structural association rule for substitution dependency is converted into the attribute parameter "substitution coefficient," with a value ranging from 0 to 1; a larger value indicates a higher degree of substitution. The structural association rule for connection dependency is converted into the attribute parameter "connection coefficient," with a value ranging from 0 to 1; a larger value indicates a higher degree of connection. The structural association rule for support dependency is converted into the attribute parameter "support coefficient," with a value ranging from 0 to 1; a larger value indicates a higher degree of support.

[0066] Step 155: Construct a structural segment connection relationship network based on network nodes carrying structural feature information and associated edges carrying structural association rule information.

[0067] In this embodiment, the structural segment connection network is constructed using graph representation methods in graph theory, combining network nodes and associated edges into a graph structure. The node set of the graph structure consists of all structural segment network nodes, and the edge set consists of all associated edges. For example, the structural segment connection network of a flange includes network nodes such as disc body nodes, bolt hole nodes, and sealing groove nodes, as well as associated edges such as the associated edges between the disc body and bolt holes, and the associated edges between the disc body and sealing groove.

[0068] Step 156: Through the collaborative mapping of nodes and associated edges in the structural segment connection relationship network, the spatial positional relationship, structural feature association and connection relationship of each structural segment are integrated into a unified spatial structural relationship network as the topology of the target metal product. The topology retains the structural feature information of each network node and the structural association rule information of each associated edge.

[0069] In this embodiment, the collaborative mapping is implemented using a graph embedding algorithm, which maps the features of network nodes and associated edges into a low-dimensional space to form a spatial structural relationship network. The spatial structural relationship network is generated using a visualization method, displaying the mapped network nodes and associated edges graphically to intuitively reflect the spatial positional relationships, structural feature associations, and connection relationships between each structural segment. For example, in the topology of a flange, the disc body node is located at the center, bolt hole nodes are distributed around the disc body node, sealing groove nodes are located at the edge of the disc body node, and associated edges connect each node, reflecting the connection relationships and structural association rules between each structural segment.

[0070] Step 1561: Extract the structural feature information carried by each network node in the structural segment connection relationship network, analyze the spatial arrangement of the local point cloud corresponding to each structural feature information, the density distribution trend of the local region and the normal vector distribution law, and extract the structural association rule information carried by each associated edge, and decompose the association requirements between each structural segment in terms of spatial arrangement and structural form.

[0071] In this embodiment, the extraction of structural feature information employs an attribute reading method. This method reads the structural feature vectors stored in the network node attributes and analyzes them to obtain the spatial arrangement, density distribution trend, and normal vector distribution pattern. The extraction of structural association rule information also employs an attribute reading method. This method reads the rule encoding parameters stored in the associated edge attributes and decomposes them to obtain the association requirements in terms of spatial arrangement and structural form. For example, analyzing the structural feature information of the disk node reveals that the spatial arrangement of the local point cloud of the disk is planar, the density distribution trend is uniform, and the normal vector distribution pattern is perpendicular to the disk surface. Analyzing the structural association rule information of the associated edge between the disk and the bolt hole reveals that the position and depth of the bolt hole must be within the disk's structural range.

[0072] Step 1562: Based on the structural feature information obtained from parsing and the structural association rule information after decomposition, create a node-edge collaborative mapping matrix, bind the identification information of each network node with the corresponding associated edge identification information, so that the network node and the associated edge form a one-to-one mapping relationship.

[0073] In this embodiment, the node-edge collaborative mapping matrix is ​​created using a matrix representation method. Rows represent the identification information of network nodes, and columns represent the identification information of associated edges. A value of 1 indicates that a network node and its associated edge have a mapping relationship, while a value of 0 indicates that no mapping relationship exists. For example, if the identification information of a disk node is 1, and the identification information of the associated edge between the disk and the bolt hole is 101, then the element in the first row and 101st column of the matrix has a value of 1, indicating that the disk node and this associated edge have a mapping relationship.

[0074] Step 1563: Retrieve the spatial location range data of the structural segments corresponding to each network node, and input the spatial location range data into the node-edge collaborative mapping matrix to realize the associated storage of structural feature information, structural association rule information and spatial location range data in the matrix.

[0075] In this embodiment, the spatial location range data is retrieved using an attribute reading method, which reads the spatial location range data stored in the network node attributes. The associated storage is implemented using a matrix expansion method, adding columns to the node-edge collaborative mapping matrix to store the spatial location range data. For example, X-coordinate range columns, Y-coordinate range columns, and Z-coordinate range columns are added to the node-edge collaborative mapping matrix to store the spatial location range data of each structural segment.

[0076] Step 1564: Based on the node-edge collaborative mapping matrix, and combined with the spatial arrangement pattern, density distribution trend, and normal vector distribution pattern obtained from the analysis, cross-compare the structural feature information of each network node, calculate the similarity of spatial arrangement pattern, the consistency of density distribution trend, and the matching degree of normal vector distribution pattern among different nodes, and screen out network node groups whose similarity of spatial arrangement pattern, the consistency of density distribution trend, and the matching degree of normal vector distribution pattern all meet the corresponding preset index thresholds and mark them as similar structural feature node groups; at the same time, classify and organize the structural association rule information of each associated edge to obtain different types of association rule groups.

[0077] In this embodiment, the similarity of spatial arrangement is calculated using the cosine of the angle between feature vectors; the consistency of density distribution trends is calculated using the cross-entropy of the density distribution function; and the matching degree of normal vector distribution patterns is calculated using the Bach distance of the histogram of normal vector direction distributions. Preset threshold values ​​are set according to the structural design requirements of the target metal product. For example, the similarity threshold for spatial arrangement is 0.8, the consistency threshold for density distribution trends is 0.7, and the matching degree threshold for normal vector distribution patterns is 0.8. Similar structural feature node groups are selected using a logical AND operation, marking network node groups that meet all three thresholds as similar structural feature node groups. Clustering algorithms are used to classify and organize association rule groups, grouping association edges with similar structural association rule information into the same group. For example, multiple bolt hole nodes are grouped into similar structural feature node groups; and association edges with substitution dependencies are grouped into the same association rule group.

[0078] Step 1565: Match similar structural feature node groups with corresponding association rule groups, and combine the spatial location range data of each group with the spatial arrangement pattern and density distribution trend information obtained by parsing to construct an initial spatial structure relationship network, so that the node arrangement in the initial network is adapted to the spatial arrangement pattern of the structural segment, and the connection relationship between nodes is matched with the density distribution trend.

[0079] In this embodiment, the matching operation employs a rule-based matching method, matching the rule information of similar structural feature node groups with the association rule groups to determine the connection relationships between nodes and associated edges. The initial spatial structure relationship network is constructed using a graph generation algorithm, generating the initial spatial structure relationship network based on the matched node and associated edge information. For example, similar structural feature node groups (bolt hole nodes) are matched with association rule groups representing substitution dependencies to construct the initial spatial structure relationship network, ensuring that bolt hole nodes are distributed around disk nodes, and that the connection relationships between nodes match the density distribution trend of bolt holes.

[0080] Step 1566: Perform network topology optimization on the initial spatial structure network. Based on the distribution pattern of normal vectors obtained from the analysis, adjust the arrangement order of network nodes to make the node arrangement consistent with the actual spatial position distribution of the corresponding structural segments. Optimize the connection method of the associated edges to make the actual connection relationship between the associated edges and the corresponding structural segments, and make the distribution pattern of normal vectors of the optimized nodes match the connection logic of adjacent nodes.

[0081] In this embodiment, network topology optimization employs a genetic algorithm, using node arrangement order and associated edge connection method as optimization variables, and using the matching degree between node arrangement and actual spatial location distribution, the matching degree between associated edge connection method and actual connection relationship, and the matching degree between normal vector distribution pattern and adjacent node connection logic as optimization objectives for multi-objective optimization. For example, the arrangement order of bolt hole nodes is adjusted to match the actual spatial location distribution of bolt holes on the flange; the associated edge connection method between bolt hole nodes and disk body nodes is optimized to match the actual connection relationship of bolt holes embedded in the disk body; and the normal vector distribution pattern of nodes is adjusted to match the connection logic of adjacent nodes.

[0082] Step 1567: Use the optimized spatial structure relationship network as the topology of the target metal product. The topology stores the structural feature information of each network node, the structural association rule information of each associated edge, and the spatial location range data of each structural segment. The information forms a coherent association system through the node-edge collaborative mapping relationship.

[0083] In this embodiment, the topology is stored in a graph database, where network nodes, associated edges, and various information are stored. The association and access between these information are achieved through node-edge collaborative mapping. For example, the topology of a flange is stored in the graph database. By querying the flange's nodes, one can obtain the flange's structural feature information, the structural association rules of associated edges, and the spatial location range data of the flange.

[0084] Step 160: Align the topology and its corresponding discrete point cloud segments and input them into a graph neural network to perform inference analysis based on node features and associated edge attributes. Based on the inference analysis results, perform associated reconstruction of structural segments to obtain a three-dimensional model of the target metal product.

[0085] In this embodiment, the graph neural network employs the GraphSAGE algorithm, which can effectively process graph structure data and extract features of nodes and edges. The sub-steps of this step are described in detail below:

[0086] Step 161: Extract the discrete point cloud segments corresponding to each network node in the topology and determine the spatial coordinate information of each discrete point cloud segment.

[0087] In this embodiment, the extraction of discrete point cloud segments employs an attribute reading method. This method reads the identifier information of the discrete point cloud segments stored in the network node attributes and extracts the corresponding discrete point cloud segments from the point cloud database based on the identifier information. The determination of spatial coordinate information uses a coordinate reading method, which reads the spatial coordinate values ​​of each point cloud in the discrete point cloud segment. For example, extracting the discrete point cloud segment corresponding to a disk node and determining the X, Y, and Z coordinate values ​​of each point cloud in that segment.

[0088] Step 162: Using the spatial structure relationship network of the topology as a reference, align the spatial coordinate information of each discrete point cloud fragment with the spatial position of the corresponding network node in the topology to achieve spatial position matching between the discrete point cloud fragment and the topology.

[0089] In this embodiment, spatial alignment is achieved using a coordinate transformation method. Based on the spatial location information of network nodes in the topology, a coordinate transformation matrix for discrete point cloud segments is calculated, transforming the spatial coordinate information of the discrete point cloud segments to the coordinate system of the topology. For example, the spatial coordinate information of the bolt hole discrete point cloud segments is transformed to the coordinate system of the disk nodes, so that the position of the bolt holes matches the position of the disk.

[0090] Step 163: Input the aligned topology and discrete point cloud fragments into the graph neural network, and extract the node features of each network node and the associated edge attributes of each associated edge through the graph neural network.

[0091] In this embodiment, the input to the graph neural network is graph-structured data. The network nodes and associated edges of the topological structure are used as the nodes and edges of the graph, and the spatial coordinate information of discrete point cloud fragments is used as the feature input to the nodes. Node feature extraction uses the aggregation function of the GraphSAGE algorithm to aggregate the features of the node's neighboring nodes, obtaining the node's feature vector. Associated edge attribute extraction uses an edge feature extraction function to extract the attribute information of the associated edges, obtaining the associated edge's feature vector. For example, the feature vector extracted from the disk nodes using the GraphSAGE algorithm includes information on the disk's spatial arrangement, density distribution trend, and normal vector distribution pattern; the feature vector extracted from the associated edges between the disk and bolt holes includes information on the connection method and structural association rules.

[0092] Step 164: Perform inference analysis based on node features and associated edge attributes to mine potential structural association information between structural segments and label undefined associations in the topology. Combine the structural association information and undefined associations to generate structural association inference results.

[0093] In this embodiment, the inference analysis is implemented using a graph neural network classification function. Node features and associated edge attributes are input into the classification function to obtain structural association information between structural segments. The mining of potential structural association information employs an association rule mining algorithm to extract undefined structural association information from the classification results. The labeling of undefined associations uses an attribute addition method, adding identifiers for undefined associations to the associated edge attributes of the topology. The generation of structural association inference results uses a report generation method, compiling the mined structural association information and labeled undefined associations into a report. For example, it mines potential structural association information between bolt holes and sealing grooves (i.e., the position of the bolt hole needs to match the position of the sealing groove); labels the undefined associations between bolt holes and sealing grooves in the topology; and generates a structural association inference result report.

[0094] Step 1641: Use a standardization algorithm to process node features and associated edge attributes, eliminate the dimensional differences between node features of different dimensions and the numerical deviations between associated edge attributes of different types, and obtain standardized node features and standardized associated edge attributes.

[0095] In this embodiment, the standardization algorithm employs the Z-score standardization method to calculate the mean and standard deviation of node features and associated edge attributes, converting them into standardized data with a mean of 0 and a standard deviation of 1. For example, the feature vectors of disk nodes are Z-score standardized to obtain standardized node features; the feature vectors of the associated edges between the disk and bolt holes are Z-score standardized to obtain standardized associated edge attributes.

[0096] Step 1642: Input the standardized node features and standardized associated edge attributes into the feature deepening layer of the graph neural network. Mine the standardized node features through the convolution operation of the feature deepening layer to extract deep node features containing structural segment details. Then, process the standardized associated edge attributes through feature enhancement operation to obtain enhanced associated edge attributes.

[0097] In this embodiment, the feature enhancement layer employs a convolutional neural network layer, with the kernel size set according to the dimension of the node features. The convolution operation is implemented using a sliding window approach, performing convolution operations on standardized node features to extract deep node features. The feature enhancement operation employs a multilayer perceptron, performing nonlinear transformations on standardized associated edge attributes to obtain enhanced associated edge attributes. For example, by performing convolution operations on standardized disk node features through a convolutional neural network layer, deep node features containing detailed features of the disk sealing groove are extracted; the associated edge attributes between the standardized disk and bolt holes are processed using a multilayer perceptron to obtain enhanced associated edge attributes.

[0098] Step 1643: Input the deep node features and reinforced association edge attributes into the inference operation layer of the graph neural network to construct the structural association inference path, and perform calculation and analysis on the association relationship between each network node based on the preset inference logic.

[0099] In this embodiment, the inference operation layer employs a recurrent neural network (RNN) layer, with the hidden layer size determined based on the dimensionality of the deep node features. The construction of the structural association inference path utilizes a sequence generation method, generating a sequence path for structural association inference based on the deep node features and reinforced association edge attributes. The preset inference logic is set according to the structural design rules of the target metal product; for example, "if two structural segments are spatially adjacent and have similar structural features, then there is a structural association between them." The computational analysis is implemented using a logical inference method, judging and analyzing the association relationships between each network node according to the preset inference logic. For example, a structural association inference path is constructed between the disc node, bolt hole node, and sealing groove node, and their association relationships are analyzed based on the preset inference logic.

[0100] Step 1644: Mine potential structural association information that is not clearly defined between structural segments through reasoning operations. Potential structural association information includes indirect connection relationships between non-directly adjacent structural segments, substitution relationships between different similar structural segments, and collaborative support relationships between cross-regional structural segments.

[0101] In this embodiment, the mining of potential structural association information employs an association rule mining algorithm to extract structural association information that is not explicitly defined from the results of inference operations. For example, it mines the indirect connection relationship between bolt holes and sealing grooves, i.e., the position of the bolt hole needs to be matched with the position of the sealing groove through the structure of the disc; it mines the substitution relationship between different bolt holes, i.e., the structure of one bolt hole can substitute for the structure of another bolt hole; and it mines the cooperative support relationship between the disc and the sealing groove, i.e., the structure of the disc needs to support the structure of the sealing groove.

[0102] Step 1645: Based on the potential structural association information obtained from mining, traverse all network nodes and associated edges in the topology, locate the unlabeled associations in the topology, classify the unlabeled associations by type and assign them unique identifiers.

[0103] In this embodiment, the location of unlabeled relationships employs a traversal search method, traversing all network nodes and associated edges in the topology to find node pairs and edges with unlabeled relationships. Type classification is implemented using a clustering algorithm, dividing unlabeled relationships according to structural association types. Unique identifiers are assigned using a numbering method, assigning a unique number to each unlabeled relationship. For example, the unlabeled relationship between a bolt hole and a sealing groove is located, classified as an indirect connection relationship, and assigned the unique identifier R001.

[0104] Step 1646: Integrate the potential structural association information mined and the undefined association relationships of the location annotations, sort the integrated information according to the arrangement order of network nodes, and generate a structured information set as the structural association inference result.

[0105] In this embodiment, the integration operation employs a data fusion method to combine potential structural association information and information with undefined relationships. The sorting operation uses a node arrangement order method to sort the integrated information according to the arrangement order of network nodes in the topology. The structured information set is generated in tabular form, recording the type of integrated information, associated nodes, association rules, etc. For example, a structural association reasoning result table is generated, recording the indirect connection relationship between bolt holes and sealing grooves, the associated nodes being bolt hole nodes and sealing groove nodes, and the association rule that the position of the bolt hole must match the position of the sealing groove.

[0106] Step 165: Based on the structural association reasoning results, perform association reconstruction on the discrete point cloud segments corresponding to each structural segment to obtain the target point cloud model.

[0107] In this embodiment, the correlation reconstruction is achieved using a point cloud stitching algorithm, which stitches and fuses the discrete point cloud fragments corresponding to each structural segment according to the requirements of the structural correlation reasoning results. The following is a detailed description of each sub-step of this step:

[0108] Step 1651: Analyze the structural association reasoning results, extract the association strength parameters and association position coordinate information between each structural segment, set the reconstruction priority threshold based on the association strength parameters, divide the structural segment combination with the association strength parameter higher than the reconstruction priority threshold into the first priority reconstruction unit, and divide the structural segment combination with the association strength parameter lower than or equal to the reconstruction priority threshold into the second priority reconstruction unit.

[0109] In this embodiment, the correlation strength parameters are extracted using an attribute reading method, which reads the correlation strength parameters stored in the structural correlation inference results. The reconstruction priority threshold is set according to the structural importance of the target metal product. For example, if the reconstruction priority threshold is set to 0.8, structural segments with correlation strength parameters higher than 0.8 are classified as first-priority reconstruction units, such as the combination of a disc and bolt holes; structural segments with correlation strength parameters lower than or equal to 0.8 are classified as second-priority reconstruction units, such as the combination of a disc and a sealing groove.

[0110] Step 1652: Retrieve the discrete point cloud segments corresponding to each structural segment in the first priority reconstruction unit, extract the spatial coordinate data of each discrete point cloud segment, and establish a coordinate mapping coordinate system based on the spatial structural relationship network in the topology.

[0111] In this embodiment, the discrete point cloud fragments are retrieved using an attribute reading method. The identification information of the discrete point cloud fragments corresponding to each structural segment in the first priority reconstruction unit is read, and the corresponding discrete point cloud fragments are extracted from the point cloud database based on the identification information. Spatial coordinate data is extracted using a coordinate reading method, reading the spatial coordinate values ​​of each point cloud in the discrete point cloud fragment. The coordinate mapping coordinate system is established using a coordinate system transformation method, with the center of the spatial structural network in the topology as the origin and the main arrangement direction of the structural segments as the coordinate axis direction. For example, discrete point cloud fragments corresponding to the disk body and bolt holes are retrieved, their spatial coordinate data is extracted, and a coordinate mapping coordinate system is established with the center of the disk body as the origin.

[0112] Step 1653: Transform the spatial coordinate data of each discrete point cloud segment to the coordinate mapping coordinate system, and adjust the spatial attitude of each discrete point cloud segment so that the coordinate data of the connecting parts of the corresponding structural segments of each discrete point cloud segment are matched.

[0113] In this embodiment, coordinate transformation is achieved using a coordinate transformation matrix. The coordinate transformation matrix of the discrete point cloud segment is calculated based on the parameters of the coordinate mapping coordinate system, transforming the spatial coordinate data of the discrete point cloud segment to the coordinate mapping coordinate system. Spatial attitude adjustment employs rotation and translation operations. Based on the coordinate data requirements of the structural segment connection points, the rotation angle and translation distance of the discrete point cloud segment are adjusted to match the coordinate data of the connection points. For example, the spatial coordinate data of the bolt hole discrete point cloud segment is transformed to the coordinate mapping coordinate system, and the rotation angle and translation distance of the bolt hole are adjusted to match the edge coordinate data of the bolt hole with the bolt hole installation position coordinate data of the disk.

[0114] Step 1654: Perform point cloud fusion processing on the connecting parts of the discrete point cloud segments after coordinate matching, calculate the spatial distance between adjacent point clouds at the connecting parts, and resample the point clouds at the connecting parts based on the spatial distance to make the point cloud distribution density at the connecting parts consistent with that of the adjacent areas.

[0115] In this embodiment, the point cloud fusion processing employs a point cloud stitching algorithm to stitch together the point clouds at the connecting parts of discrete point cloud segments after coordinate matching. Spatial distance is calculated using the Euclidean distance formula to determine the spatial distance between adjacent point clouds at the connecting parts. Resampling is implemented using a voxel grid downsampling algorithm, setting the voxel grid size based on the spatial distance and resampling the point clouds at the connecting parts. For example, when fusing the point clouds at the connection between the disk and bolt holes, the spatial distance between adjacent point clouds is calculated, the voxel grid size is set to 0.5 mm, and the point clouds at the connecting parts are resampled to ensure that the point cloud distribution density at the connecting parts is consistent with the point cloud distribution density of the disk.

[0116] Step 1655: After completing the association reconstruction of the first priority reconstruction unit, retrieve the discrete point cloud fragment corresponding to the second priority reconstruction unit, convert its spatial coordinate data to the coordinate mapping coordinate system using the same coordinate mapping method, and combine the association position coordinate information in the structural association reasoning result to dock the discrete point cloud fragment corresponding to the second priority reconstruction unit with the first priority reconstruction unit that has been reconstructed.

[0117] In this embodiment, the discrete point cloud fragments are retrieved using an attribute reading method. The identifier information of the discrete point cloud fragment corresponding to the second priority reconstruction unit is read, and the corresponding discrete point cloud fragment is extracted from the point cloud database based on the identifier information. The coordinate mapping method is the same as that used for the first priority reconstruction unit, converting the spatial coordinate data of the discrete point cloud fragments to a coordinate mapping coordinate system. The docking operation is implemented using a point cloud stitching algorithm. Based on the associated position coordinate information in the structural association inference results, the discrete point cloud fragments corresponding to the second priority reconstruction unit are stitched together with the already reconstructed first priority reconstruction unit. For example, the discrete point cloud fragment corresponding to the sealing groove is retrieved, its spatial coordinate data is converted to a coordinate mapping coordinate system, and based on the sealing groove position coordinate information in the structural association inference results, the discrete point cloud fragment of the sealing groove is docked with the already reconstructed disk and bolt hole combination.

[0118] Step 1656: Based on the overall point cloud structure after docking, perform a traversal check to locate the point cloud regions with coordinate deviations, and correct the point cloud coordinates of the deviation regions based on the spatial structure relationship network in the topology.

[0119] In this embodiment, the traversal check is implemented using a point cloud traversal algorithm. This algorithm traverses the overall point cloud structure after docking and calculates the deviation between the spatial coordinates of each point cloud and the spatial position information in the topology. The coordinate deviation is located using a threshold judgment method, identifying point cloud regions with deviations greater than a set threshold as areas with coordinate deviations. Coordinate correction is implemented using a coordinate transformation method. Based on the coordinate information of the spatial structure network in the topology, the correction value of the point cloud coordinates in the deviation area is calculated, and the point cloud coordinates of the deviation area are corrected. For example, by traversing the overall point cloud structure after docking, point cloud regions with coordinate deviations at the edge of the sealing groove are located, and the point cloud coordinates of these regions are corrected based on the sealing groove position information in the topology.

[0120] Step 1657: Integrate all the discrete point cloud fragments that have been reconstructed and docked to generate an initial point cloud set. Screen out the duplicate point clouds generated during the reconstruction process in the initial point cloud set to obtain the target point cloud model.

[0121] In this embodiment, the initial point cloud set is generated using a point cloud stitching algorithm, which stitches together all the discrete point cloud segments that have completed reconstruction and docking. Duplicate point clouds are removed using a statistical filtering algorithm, which calculates the number of neighboring point clouds for each point cloud and identifies point clouds with a number of neighboring point clouds exceeding a set threshold as duplicate point clouds and removes them. For example, the initial point cloud set is generated by integrating discrete point cloud segments of the disc body, bolt holes, and sealing grooves; duplicate point clouds are then removed to obtain the target point cloud model of the flange.

[0122] Step 166: Optimize the structural coherence of the target point cloud model, delete the structurally conflicting point clouds in the target point cloud model, and obtain the three-dimensional model of the target metal product. The three-dimensional model retains the spatial relationship of the topological structure and the detailed features of the discrete point cloud.

[0123] In this embodiment, structural coherence optimization employs a point cloud smoothing algorithm to smooth the target point cloud model, making the point cloud structure more coherent. The deletion of structural conflict point clouds utilizes a conflict detection algorithm to detect and delete structurally conflicting point clouds, such as overlapping and intersecting point clouds, present in the target point cloud model. For example, smoothing the target point cloud model of a flange and deleting overlapping point clouds yields a three-dimensional model of the flange that retains the spatial relationships between the disc body, bolt holes, and sealing grooves, as well as the detailed features of the point cloud.

[0124] In an optional technical solution, the method further includes:

[0125] Step 210: Based on the three-dimensional model of the target metal product, perform hierarchical decomposition of features to obtain a hierarchical feature set containing a basic feature layer, a core feature layer, and a related feature layer.

[0126] In this embodiment, the hierarchical decomposition of features employs a feature decomposition algorithm, dividing the features of the 3D model into different levels according to their importance and degree of abstraction. The basic feature layer contains the basic geometric features of the 3D model, such as the spatial coordinates and normal vectors of the point cloud; the core feature layer contains the key structural features of the 3D model, such as the planar features of the disc and the cylindrical features of the bolt holes; the associated feature layer contains the structural associated features of the 3D model, such as the connection features between the bolt holes and the disc, and the connection features between the sealing groove and the disc. For example, the hierarchical feature set of the 3D model of the flange includes the basic feature layer (spatial coordinates and normal vectors of the point cloud), the core feature layer (planar features of the disc, cylindrical features of the bolt holes, and annular features of the sealing groove), and the associated feature layer (embedded connection features of the bolt holes and the disc, and integral forming connection features of the sealing groove and the disc).

[0127] Step 220: Conduct multi-dimensional adaptation requirement analysis of application scenarios based on the hierarchical feature set, and output a scenario adaptation requirement feature mapping table; dynamically adapt and adjust each feature layer in the hierarchical feature set based on the scenario adaptation requirement feature mapping table to obtain the initial adjusted hierarchical feature set.

[0128] In this embodiment, the multi-dimensional adaptation requirement analysis for application scenarios employs a requirement survey method. This method surveys the feature requirements of 3D models for metal products under different application scenarios. For example, machining scenarios require precision features of the 3D model, while assembly scenarios require assembly features. The output of the scenario adaptation requirement feature mapping table is in tabular form, recording the degree of demand for each feature layer under different application scenarios. Dynamic adaptation adjustment is achieved using a feature weight adjustment method. Based on the degree of demand in the scenario adaptation requirement feature mapping table, the weight coefficients of each feature layer are adjusted to obtain an initial hierarchical feature set. For example, in the machining scenario, the weight coefficient of the core feature layer is adjusted to 0.6, the weight coefficient of the basic feature layer is adjusted to 0.3, and the weight coefficient of the related feature layer is adjusted to 0.1; in the assembly scenario, the weight coefficient of the related feature layer is adjusted to 0.5, the weight coefficient of the core feature layer is adjusted to 0.3, and the weight coefficient of the basic feature layer is adjusted to 0.2.

[0129] Step 230: Perform cross-scene adaptation compatibility verification on the initial hierarchical feature set and generate compatibility verification results; perform multiple rounds of iterative adjustment on the initial hierarchical feature set based on the compatibility verification results to obtain the final hierarchical feature set.

[0130] In this embodiment, cross-scenario adaptation compatibility verification employs a feature matching algorithm to match the initial hierarchical feature set with the required features of different application scenarios and calculate the matching degree. The compatibility verification results are generated in report form, recording the matching degree results and any existing compatibility issues. Multi-round iterative adjustments are implemented using a feedback adjustment method. Based on the feedback information from the compatibility verification results, the weight coefficients of the initial hierarchical feature set are adjusted, and cross-scenario adaptation compatibility verification is repeated until a final hierarchical feature set that meets compatibility requirements is obtained. For example, compatibility verification of the initial hierarchical feature set is performed for machining and assembly scenarios, generating a compatibility verification report. Based on the feedback information in the report, the weight coefficients are adjusted, and multiple rounds of iterative adjustments are performed to obtain the final hierarchical feature set.

[0131] Step 240: Based on the final hierarchical feature set, perform scene-based reconstruction of the target metal product's 3D model to obtain a scene-customized 3D model; perform feature adaptation effectiveness analysis on the scene-customized 3D model to obtain adaptation effectiveness results.

[0132] In this embodiment, the scene-based reconstruction is achieved using a model reconstruction algorithm. Based on the feature information of the final hierarchical feature set, the 3D model of the target metal product is reconstructed to generate a scene-customized 3D model. The feature adaptation effectiveness analysis is implemented using an effect evaluation method to assess the adaptation effect of the scene-customized 3D model in the corresponding application scenario, obtaining an adaptation effectiveness result. For example, reconstructing the 3D model of a flange based on the final hierarchical feature set yields a machining scene-customized 3D model. Evaluating the adaptation effect of this model in the machining scene yields a good adaptation effectiveness result.

[0133] Step 250: Generate scenario application execution guidelines based on the scenario-customized 3D model and the adaptation validity results. The scenario application execution guidelines are used to associate the model calling process and feature matching logic under different subsequent application scenarios.

[0134] In this embodiment, the scene application execution guide is generated using a document writing method. Based on the feature information and adaptation effectiveness results of the scene-customized 3D model, the scene application execution guide is written, recording the model calling process and feature matching logic under different application scenarios. For example, the execution guide for the machining scene is generated, recording the steps of model calling, feature matching rules, and other content.

[0135] Step 260: Based on the scene application execution guide, associate and bind the scene-customized 3D model with the application scene, and output the scene binding results and the corresponding model application startup parameters.

[0136] In this embodiment, the application scenario association and binding adopts a rule-based binding method. Based on the rules of the scenario application execution guidelines, the customized 3D model of the scenario is bound to the corresponding application scenario. The scenario binding result is output in report form, recording the bound application scenario and model information. The model application startup parameters are output in the form of a parameter configuration file, recording the startup parameters of the model application, such as model path and feature weights. For example, binding a customized 3D model of a machining scenario to a machining application scenario outputs a scenario binding result report and a model application startup parameter configuration file.

[0137] In an optional technical solution, the method further includes:

[0138] Step 310: Obtain multi-source correlation data related to the target metal product; perform hierarchical filtering of the multi-source correlation data based on the feature dimensions of the three-dimensional model of the target metal product to obtain a preliminary filtered data set; perform correlation verification between the data dimensions and model feature dimensions of the preliminary filtered data set and output the correlation verification results; perform secondary filtering of the preliminary filtered data set based on the correlation verification results to obtain a correlation data set.

[0139] In this embodiment, the multi-source associated data includes design drawings, processing technology data, and inspection data of the target metal product. Feature dimensions are determined using a feature extraction algorithm to extract the feature dimensions of the 3D model, such as geometric and structural feature dimensions. Hierarchical filtering is implemented using a data filtering algorithm to filter the multi-source associated data based on the feature dimensions, obtaining a preliminary filtered data set. Association verification is implemented using a correlation analysis algorithm to calculate the correlation between the data dimensions of the preliminary filtered data set and the model feature dimensions, outputting the association verification result. Secondary filtering is implemented using a data filtering algorithm to filter the preliminary filtered data set based on the association verification result, obtaining an associated data set. For example, design drawings, processing technology data, and inspection data of a flange are acquired; hierarchical filtering of the multi-source associated data is performed based on the geometric and structural feature dimensions of the flange's 3D model to obtain a preliminary filtered data set; association verification is performed on the preliminary filtered data set, outputting an association verification report; secondary filtering is performed on the preliminary filtered data set based on the report, obtaining the associated data set.

[0140] Step 320: Establish the cross-dimensional mapping relationship between the feature dimensions of the 3D model of the target metal product and the attribute dimensions of the associated data set, and generate an initial dimension mapping matrix; perform dimension adaptation calibration on the initial dimension mapping matrix to obtain the calibrated dimension mapping matrix.

[0141] In this embodiment, the establishment of cross-dimensional mapping relationships employs a mapping algorithm to map the feature dimensions of the 3D model to the attribute dimensions of the associated data set, generating an initial dimension mapping matrix. Dimensional adaptability calibration is implemented using a matrix calibration algorithm to calibrate the initial dimension mapping matrix, adjusting the weight coefficients of the mapping relationships to obtain a calibrated dimension mapping matrix. For example, a cross-dimensional mapping relationship is established between the geometric feature dimensions (such as diameter and thickness) of the flange 3D model and the attribute dimensions (such as design diameter and processing diameter) of the associated data set, generating an initial dimension mapping matrix; the initial dimension mapping matrix is ​​then calibrated to obtain a calibrated dimension mapping matrix.

[0142] Step 330: Based on the calibrated dimension mapping matrix, the 3D model of the target metal product and the associated data set are fused to obtain an initial model-data fusion body; feature enhancement and noise filtering are performed on the initial model-data fusion body to obtain an optimized model-data fusion body; based on the optimized model-data fusion body, association rule mining and analysis are carried out to obtain a multi-level association rule set; the 3D model part in the optimized model-data fusion body is labeled with association features using the multi-level association rule set to obtain the labeled 3D model.

[0143] In this embodiment, the fusion processing employs a data fusion algorithm to fuse the feature information of the 3D model with the attribute information of the associated data set, resulting in an initial model-data fusion body. Feature enhancement and noise filtering are implemented using feature enhancement and noise filtering algorithms to process the initial model-data fusion body, yielding an optimized model-data fusion body. Association rule mining analysis utilizes an association rule mining algorithm to mine multi-level association rules from the optimized model-data fusion body, resulting in a multi-level association rule set. Association feature annotation employs a feature annotation algorithm to annotate the 3D model portion of the optimized model-data fusion body using the multi-level association rule set, resulting in an annotated 3D model. For example, the 3D model of a flange is fused with the associated data set to obtain an initial model-data fusion body; feature enhancement and noise filtering are then applied to it to obtain an optimized model-data fusion body; association rules between the design diameter and the processing diameter are mined from the optimized model-data fusion body, resulting in a multi-level association rule set; this rule set is then used to annotate the 3D model portion with association features, resulting in an annotated 3D model of the flange.

[0144] Step 340: Generate a guide for retrieving associated data and analyzing features based on the 3D model after association annotation and the multi-level association rule set. The guide is used to standardize the order of retrieving associated data and the analysis logic of model features.

[0145] In this embodiment, the generation of associated data retrieval and feature analysis guidelines adopts a document writing method. Based on the feature information of the annotated 3D model and the multi-level association rule set, the associated data retrieval and feature analysis guidelines are written, recording the retrieval order of associated data and the analysis logic of model features. For example, generating guidelines for flange associated data retrieval and feature analysis records the steps for retrieving associated data, the analysis rules of model features, and other content.

[0146] In an optional technical solution, the method further includes:

[0147] Step 410: Collect application feedback data of the 3D model of the target metal product in various subsequent application scenarios, and perform multi-dimensional analysis on the application feedback data to obtain analysis results including validity dimension, relevance dimension and credibility dimension.

[0148] In this embodiment, the application feedback data is collected using a data acquisition method, gathering feedback data such as usage data and user evaluations of the 3D model in various application scenarios. Multi-dimensional analysis is achieved using data analysis algorithms to analyze the validity, relevance, and credibility of the application feedback data, yielding analytical results. For example, by collecting application feedback data on the flange 3D model in machining and assembly scenarios and performing multi-dimensional analysis, the results show good validity, high relevance, and good credibility.

[0149] Step 420: Extract effective feedback features based on the analysis results and generate an effective feedback feature set; compare and correlate the effective feedback feature set with the feature set of the three-dimensional model of the target metal product to identify feature deviation information; perform source tracing analysis on the feature deviation information to determine the related factors and scope of influence of the deviation, and output the deviation source tracing results.

[0150] In this embodiment, the extraction of effective feedback features employs a feature extraction algorithm to extract effective feedback features from the analysis results, generating an effective feedback feature set. The correlation comparison is implemented using a feature matching algorithm, matching the effective feedback feature set with the feature set of the 3D model to identify feature deviation information. The source tracing analysis is implemented using a causal analysis algorithm to analyze the causes and scope of influence of the feature deviation information, outputting the deviation source tracing results. For example, the effective feedback feature set is extracted and compared with the feature set of the flange 3D model to identify the feature deviation information of the sealing groove depth; the source tracing analysis of this deviation information determines that the related factor causing the deviation is processing error, and the scope of influence is sealing performance; a deviation source tracing result report is output.

[0151] Step 430: Based on the deviation tracing results, formulate a multi-dimensional optimization strategy for the 3D model of the target metal product; perform the first round of feature adjustment on the 3D model of the target metal product according to the multi-dimensional optimization strategy to obtain a preliminary optimized 3D model; verify the application effect of the preliminary optimized 3D model in multiple scenarios and generate the effect verification results; iteratively optimize and adjust the preliminary optimized 3D model according to the effect verification results until the final optimized 3D model that meets the preset application effect requirements is obtained.

[0152] In this embodiment, the multi-dimensional optimization strategy is formulated using a strategy planning method. Optimization strategies, such as adjusting feature parameters and optimizing model structure, are developed based on deviation tracing results. The initial feature adjustment employs a model modification algorithm, modifying the 3D model according to the optimization strategy to obtain a preliminary optimized 3D model. Multi-scenario application effect verification utilizes an effect evaluation method, assessing the effectiveness of the preliminary optimized 3D model in various application scenarios and generating effect verification results. Iterative optimization adjustment employs a feedback adjustment method, adjusting the preliminary optimized 3D model based on feedback information from the effect verification results, and repeatedly performing multi-scenario application effect verification until the final optimized 3D model is obtained. For example, the optimization strategy formulated based on deviation tracing results is to adjust the feature parameter of the sealing groove depth; the first round of feature adjustment is performed on the flange 3D model according to this strategy to obtain a preliminary optimized 3D model; the application effect verification of the preliminary optimized 3D model is performed in machining and assembly scenarios, generating an effect verification report; iterative optimization adjustment is performed based on the feedback information in the report to obtain the final optimized 3D model.

[0153] Step 440: Update the original 3D model of the target metal product using the final optimized 3D model, and generate a model optimization report simultaneously. The model optimization report is used to record the deviation tracing process, the content of the optimization strategy, and the logic of the effect changes of each round of optimization.

[0154] In this embodiment, the model update is implemented using a model replacement method, replacing the original 3D model of the target metal product with the final optimized 3D model. The model optimization report is generated using a document writing method, recording the deviation tracing process, optimization strategy content, and the logic of the effect changes in each round of optimization. For example, replacing the original flange 3D model with the final optimized 3D model generates a model optimization report, recording the deviation tracing steps, optimization strategy content, and the effect changes in each round of optimization.

[0155] Through the above steps, the metal product modeling system completes the three-dimensional modeling of the flange, obtaining a three-dimensional model of the flange. This model can be applied to multiple application scenarios such as machining and assembly, and has high accuracy and adaptability.

[0156] This application's embodiments start from the full lifecycle processing logic of discrete point clouds of target metal products, forming a complete technical chain from point cloud preprocessing, feature boundary extraction, structural association rule mining to topological structure construction and 3D model reconstruction, realizing deep collaboration and global optimization of multiple technical means. By identifying density abrupt change regions and normal vector jump regions through full-domain scanning and conducting regional association analysis, the key structural feature boundaries of the target metal product can be accurately captured, breaking through the limitations of traditional single feature recognition. By dividing the local point cloud by key structural feature boundaries and mining structural dependencies to determine structural association rules, a precise analysis of the internal logic of the metal product structure is achieved. By constructing a structural segment connection relationship network and creating a topological structure, discrete point cloud features are transformed into topological expressions with structured associations. Finally, combined with graph neural network inference analysis, structural segment association reconstruction is achieved. The generated 3D model can not only accurately reproduce the geometric shape of the target metal product, but also completely preserve the internal associations and topological logic between its structures, greatly improving the accuracy of the 3D model and the integrity of the structural expression, reducing the information loss rate in the 3D reconstruction process of complex metal products, thus serving application scenarios such as reverse engineering, quality inspection, and digital archiving of metal products.

[0157] Please see Figure 2 The figure is a schematic diagram of the basic structure of a metal product modeling system 200 provided in an embodiment of this application. The metal product modeling system 200 includes: a processor 201; a storage device 202 on which a computer program 2020 is stored; and a network interface 203 for providing network communication functions. When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the metal product modeling methods based on three-dimensional point clouds.

[0158] Please see Figure 3 This application provides a functional block diagram of a metal product modeling device, which includes:

[0159] The point cloud normalization module is used to acquire the original discrete point cloud of the target metal product, and perform point cloud normalization based on the spatial distribution characteristics of the original discrete point cloud to obtain a preprocessed discrete point cloud.

[0160] The point cloud global scanning module is used to perform global scanning analysis on the preprocessed discrete point cloud, identify target regions with spatial distribution differences in the preprocessed discrete point cloud, and determine density abrupt regions and normal vector jump regions through the target regions.

[0161] The feature boundary generation module is used to perform regional correlation analysis based on the density abrupt change region and the normal vector jump region, extract the first edge point set of the density abrupt change region and the second edge point set of the normal vector jump region, and combine the first edge point set and the second edge point set to generate the key structural feature boundary of the target metal product.

[0162] The structural association determination module is used to divide the preprocessed discrete point cloud into multiple local point clouds based on the key structural feature boundaries, and to mine the structural dependency relationship between different structural segments by comparing the similarity between the structural features of each local point cloud and analyzing the spatial location correlation. Based on the structural dependency relationship, the structural association rules of each structural segment in the discrete point cloud are determined.

[0163] The topology creation module is used to construct a structural segment connection relationship network based on the structural association rules and the connection relationship between each structural segment, map the structural features of each local point cloud to the nodes of the structural segment connection relationship network, convert the structural association rules into the association edge attributes between nodes, and create the topology of the target metal product through the collaborative mapping of nodes and association edges.

[0164] The 3D reasoning and reconstruction module is used to align the topological structure and its corresponding discrete point cloud segments and input them into a graph neural network to perform reasoning analysis based on node features and associated edge attributes. Based on the reasoning analysis results, the module performs associated reconstruction of structural segments to obtain a 3D model of the target metal product.

[0165] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0166] Furthermore, it should be noted that this application also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. The processor of the metal product modeling system reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the metal product modeling system to perform the aforementioned... Figure 1 The methods described in the corresponding embodiments are already known, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program product embodiments related to this application, please refer to the description of the method embodiments of this application.

[0167] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. A method of modeling a metal product based on a three-dimensional point cloud, the method comprising: The method comprises: obtaining an original discrete point cloud of a target metal product, performing point cloud regularization processing based on the spatial distribution characteristics of the original discrete point cloud to obtain a preprocessed discrete point cloud; performing global scanning analysis on the preprocessed discrete point cloud to identify a target region with spatial distribution difference in the preprocessed discrete point cloud, and determining a density mutation region and a normal vector jump region through the target region; the global scanning analysis adopts a sliding window scanning algorithm, and the size of the sliding window is set according to the point cloud density of the preprocessed discrete point cloud; for the point cloud in each sliding window, the density value thereof is calculated, and the calculation method of the density value is the ratio of the number of point clouds in the window to the volume of the window; by comparing the density values in adjacent sliding windows, a region with a density value change greater than a set threshold is identified as a density mutation region; at the same time, the normal vector of each point cloud is calculated, and the calculation of the normal vector adopts a principal component analysis algorithm, and the feature vector of the covariance matrix of the point cloud neighborhood is obtained by calculating the covariance matrix of the point cloud neighborhood, wherein the feature vector corresponding to the smallest eigenvalue is the normal vector of the point cloud; by comparing the included angles of the normal vectors of adjacent point clouds, a region with an included angle of the normal vector greater than a set threshold is identified as a normal vector jump region; performing region correlation analysis based on the density mutation region and the normal vector jump region, extracting a first edge point set of the density mutation region and a second edge point set of the normal vector jump region, and combining the first edge point set and the second edge point set to generate a key structural feature boundary of the target metal product; dividing the preprocessed discrete point cloud into a plurality of local point clouds according to the key structural feature boundary, comparing the structural features of the local point clouds and analyzing the spatial position correlation, mining the structural dependency relationship between different structural segments, and determining the structural correlation rule of each structural segment in the discrete point cloud according to the structural dependency relationship; constructing a structural segment connection relationship network according to the structural correlation rule and the connection relationship between the structural segments, mapping the structural features of the local point clouds to the nodes of the structural segment connection relationship network, converting the structural correlation rule to the correlation edge attribute between the nodes, and creating a topology structure of the target metal product through the collaborative mapping of the nodes and the correlation edges; aligning the topology structure and the corresponding discrete point cloud segment and inputting them into a graph neural network to perform reasoning analysis based on the node features and the correlation edge attributes, performing correlation reconstruction of the structural segments based on the reasoning analysis result, and obtaining a three-dimensional model of the target metal product.

2. The method of claim 1, wherein, The region correlation analysis based on the density mutation region and the normal vector jump region, the extraction of the first edge point set of the density mutation region and the second edge point set of the normal vector jump region, and the combination of the first edge point set and the second edge point set to generate the key structural feature boundary of the target metal product comprise: performing spatial position overlap analysis on the density mutation region and the normal vector jump region, and screening out overlapping regions with spatial overlap and independent regions without spatial overlap; For the overlapping region, the point cloud in the overlapping region that meets both the density mutation characteristic and the normal vector jump characteristic is extracted as a core edge point, the neighborhood point cloud around the core edge point is traversed, the point cloud with a continuous mutation trend in the neighborhood is retained to form a fusion edge point set corresponding to the overlapping region; For the density mutation region in the independent region, boundary point cloud is extracted along the gradient direction of the density mutation, and isolated point cloud that is discretely distributed in the boundary point cloud is removed to obtain the first edge point set; for the normal vector jump region in the independent region, boundary point cloud is extracted along the turning direction of the normal vector change, and isolated point cloud that is discretely distributed in the boundary point cloud is removed to obtain the second edge point set; The fusion edge point set, the first edge point set and the second edge point set are spliced in order of spatial position to obtain a spliced point cloud set; A continuous curve boundary is generated based on the spatial distribution trajectory fitting of the spliced point cloud set, and a key structural feature boundary of the target metal product is formed by the curve boundary, and the key structural feature boundary is associated with the spatial position information of the fusion edge point set, the first edge point set and the second edge point set.

3. The method of claim 1, wherein, The preprocessed discrete point cloud is divided into a plurality of local point clouds according to the key structural feature boundary, the structural dependency relationship between different structural segments is mined through similarity comparison and spatial position correlation analysis between the structural features of each local point cloud, and the structural correlation rule of each structural segment in the discrete point cloud is determined according to the structural dependency relationship, including: The preprocessed discrete point cloud is divided into a plurality of non-overlapping point cloud segments with the key structural feature boundary as a segmentation reference, each point cloud segment corresponds to a local point cloud, and the spatial position range of each local point cloud is recorded; The structural features of each local point cloud are extracted, including the spatial arrangement form of the point cloud, the density distribution trend of the local region, and the normal vector distribution law; The structural features of any two local point clouds are compared for similarity, the matching degree of the structural features is calculated, and the local point cloud pairs whose matching degree meets the set condition are selected and marked as similar structural point cloud pairs; According to the spatial position correlation of all local point clouds, the adjacent relationship, the containing relationship and the parallel relationship of each local point cloud in space are determined, and a spatial position correlation graph is generated; The matching results of the similar structural point cloud pairs and the spatial position correlation graph are combined to mine the structural dependency relationship between different structural segments, including the substitution dependency relationship between similar structures, the connection dependency relationship between adjacent structures, and the support dependency relationship between containing structures; Based on the mined structural dependency relationship, the association logic between each structural segment is determined, and the association logic is converted into a structural correlation rule, including the association requirements of each structural segment in spatial arrangement and structural form.

4. The method of claim 1, wherein, The structure segment connection relationship network is constructed based on the structure association rules and the connection relationship between the structure segments, the structure features of each local point cloud are mapped to the nodes of the structure segment connection relationship network, the structure association rules are converted into the association edge attributes between the nodes, and the topological structure of the target metal product is created through the collaborative mapping of the nodes and the association edges, including: Determine the connection object and the connection mode of each structure segment according to the connection relationship between the structure segments, and obtain a connection relationship list; Construct network nodes based on each structure segment, map the structure features of the local point cloud corresponding to each structure segment to the corresponding network nodes, so that each network node carries the corresponding structure feature information; According to the connection relationship list, construct association edges between network nodes with connection relationship, and each association edge corresponds to the connection relationship between a group of structure segments; The structure association rules are converted into corresponding association edge attributes, different types of structure association rules correspond to different attribute parameters, so that the association edge carries the corresponding structure association rule information; Based on the network nodes carrying structure feature information and the association edges carrying structure association rule information, a structure segment connection relationship network is constructed; Through the collaborative mapping of the nodes and the association edges in the structure segment connection relationship network, the spatial position relationship, the structure feature association and the connection relationship of each structure segment are integrated into a unified spatial structure relationship network as the topological structure of the target metal product, and the topological structure retains the structure feature information of each network node and the structure association rule information of each association edge.

5. The method of claim 4, wherein, The collaborative mapping of the nodes and the association edges in the structure segment connection relationship network, the spatial position relationship, the structure feature association and the connection relationship of each structure segment are integrated into a unified spatial structure relationship network as the topological structure of the target metal product, including: Extract the structure feature information carried by each network node in the structure segment connection relationship network, analyze the spatial arrangement form, the density distribution trend and the normal vector distribution law of the local point cloud corresponding to each structure feature information, and extract the structure association rule information carried by each association edge, and then disassemble to obtain the association requirements between the structure segments in the spatial arrangement and the structure form; Based on the analyzed structure feature information and the disassembled structure association rule information, a node-edge collaborative mapping matrix is created, the identification information of each network node is bound with the identification information of the corresponding association edge, so that the network node and the association edge form a one-to-one mapping relationship; Call the spatial position range data of the structure segment corresponding to each network node, and record the spatial position range data in the node-edge collaborative mapping matrix to realize the associated storage of the structure feature information, the structure association rule information and the spatial position range data in the matrix; Based on the node-edge collaborative mapping matrix, the structural characteristic information of each network node is cross-compared by combining the analytically obtained spatial arrangement form, density distribution trend and normal vector distribution law, the similarity of the spatial arrangement form, the consistency of the density distribution trend and the matching degree of the normal vector distribution law between different nodes are calculated, and the network node group whose similarity of the spatial arrangement form, the consistency of the density distribution trend and the matching degree of the normal vector distribution law all meet the corresponding preset index threshold is marked as a similar structural characteristic node group; meanwhile, the structural association rule information of each associated edge is classified and arranged, and different types of association rule groups are divided; The similar structural characteristic node group and the corresponding type of association rule group are matched, and the initial spatial structure relationship network is constructed by combining the spatial position range data of each group and the analytically obtained spatial arrangement form and density distribution trend information, so that the node arrangement and the spatial arrangement form of the structure segment in the initial network are adapted, and the connection relationship between the nodes and the density distribution trend are matched; The network topology of the initial spatial structure relationship network is optimized, the arrangement order of the network nodes is adjusted to make the node arrangement consistent with the actual spatial position distribution of the corresponding structure segment by combining the analytically obtained normal vector distribution law, and the connection mode of the associated edge is optimized to make the actual connection relationship between the structure segments corresponding to the associated edge, and the normal vector distribution law of the optimized node is matched with the connection logic of the adjacent node; The optimized spatial structure relationship network is taken as the topological structure of the target metal product, the structural characteristic information of each network node, the structural association rule information of each associated edge and the spatial position range data of each structure segment are stored in the topological structure, and the information is connected to form a coherent association system through the node-edge collaborative mapping relationship.

6. The method according to any one of claims 1 to 5, wherein, The topological structure and the corresponding discrete point cloud segment are aligned and input into a graph neural network to perform reasoning analysis based on node features and associated edge attributes, and the associated reconstruction of the structure segment is performed based on the reasoning analysis result to obtain a three-dimensional model of the target metal product, including: Extracting the discrete point cloud segment corresponding to each network node in the topological structure, and determining the spatial coordinate information of each discrete point cloud segment; Aligning the spatial coordinate information of each discrete point cloud segment with the spatial position of the corresponding network node in the topological structure based on the spatial structure relationship network of the topological structure, so as to realize the spatial position matching of the discrete point cloud segment and the topological structure; Inputting the aligned topological structure and discrete point cloud segment into a graph neural network to extract the node features of each network node and the associated edge attributes of each associated edge through the graph neural network; Performing reasoning analysis based on the node features and the associated edge attributes, mining the potential structural association information between each structure segment and labeling the undefined association relationship in the topological structure, and generating a structure association reasoning result combining the structural association information and the undefined association relationship; According to the structure association reasoning result, the discrete point cloud segment corresponding to each structure segment is associated and reconstructed to obtain a target point cloud model; The target point cloud model is subjected to structure consistency optimization, and structure conflict point clouds existing in the target point cloud model are deleted, so as to obtain a three-dimensional model of the target metal product, and the three-dimensional model retains the spatial correlation of the topological structure and the detail characteristics of the discrete point cloud.

7. The method of claim 6, wherein, The inference analysis is performed based on the node features and the associated edge attributes, potential structure correlation information between structure segments is mined, and unbounded correlation relationships in the topological structure are labeled, and a structure correlation inference result is generated in combination with the structure correlation information and the unbounded correlation relationships, including: The node features and the associated edge attributes are processed by using a standardization algorithm, so as to eliminate dimensional differences between different dimension node features and numerical deviations between different types of associated edge attributes, and standardized node features and standardized associated edge attributes are obtained; The standardized node features and the standardized associated edge attributes are input into the feature deepening layer of the graph neural network, the standardized node features are mined by convolution operation of the feature deepening layer, deep layer node features containing structure segment detail features are extracted, and the standardized associated edge attributes are processed by feature strengthening operation to obtain strengthened associated edge attributes; The deep layer node features and the strengthened associated edge attributes are input into the inference operation layer of the graph neural network to construct a structure correlation inference path, and the correlation relationships between network nodes are analyzed based on a preset inference logic; Potential structure correlation information between structure segments that is not explicitly bounded is mined by inference operation, and the potential structure correlation information includes indirect connection relationships between non-directly adjacent structure segments, substitution correlation relationships between different similar structure segments, and cooperative support relationships between cross-region structure segments; Based on the mined potential structure correlation information, all network nodes and associated edges in the topological structure are traversed, and unmarked correlation relationships in the topological structure are located, the unmarked correlation relationships are classified by type and given a unique identifier; The mined potential structure correlation information and the located unbounded correlation relationships are integrated, the integrated information is sorted according to the arrangement order of the network nodes, and a structured information set is generated as a structure correlation inference result.

8. The method of claim 6, wherein, The discrete point cloud segments corresponding to the structure segments are associated and reconstructed according to the structure correlation inference result, and a target point cloud model is obtained, including: The structure correlation inference result is analyzed, the correlation strength parameters and the correlation position coordinate information between the structure segments are extracted, a reconstruction priority threshold is set based on the correlation strength parameters, structure segment combinations with correlation strength parameters higher than the reconstruction priority threshold are classified into a first priority reconstruction unit, and structure segment combinations with correlation strength parameters lower than or equal to the reconstruction priority threshold are classified into a second priority reconstruction unit; The discrete point cloud segments corresponding to the structure segments in the first priority reconstruction unit are called, and the spatial coordinate data of each discrete point cloud segment is extracted, a coordinate mapping coordinate system is established based on the spatial structure relationship network in the topological structure; The spatial coordinate data of each discrete point cloud segment is converted to a coordinate mapping coordinate system, the spatial attitude of each discrete point cloud segment is adjusted, and the coordinate data of the connecting part of each discrete point cloud segment corresponding to the structure segment is matched; The point cloud fusion processing is performed on the connecting part of the discrete point cloud segment after the coordinate matching, the spatial distance of the adjacent point clouds of the connecting part is calculated, the point cloud of the connecting part is resampled based on the spatial distance, and the point cloud distribution density of the connecting part is consistent with the adjacent area; After the association reconstruction of the first priority reconstruction unit is completed, the discrete point cloud segment corresponding to the second priority reconstruction unit is called, the spatial coordinate data thereof is converted to the coordinate mapping coordinate system by using the same coordinate mapping mode, the discrete point cloud segment corresponding to the second priority reconstruction unit is connected to the first priority reconstruction unit which has completed the reconstruction in combination with the association position coordinate information in the structure association inference result; According to the traversal inspection of the whole point cloud structure after the connection, the point cloud area with coordinate deviation is located, and the point cloud coordinate of the deviation area is corrected based on the spatial structure relationship network in the topological structure; The initial point cloud set is generated by integrating all the discrete point cloud segments which have completed the reconstruction and connection, the repeated point clouds in the initial point cloud set due to the reconstruction process are screened out, and the target point cloud model is obtained.

9. A metal article modeling system characterized by, It comprises: A processor, a storage device having a computer program stored thereon, and a network interface for providing network communication function. When the computer program is executed by the processor, the processor implements the three-dimensional point cloud-based metal product modeling method according to any one of claims 1-8.

10. A readable storage medium, characterized by, The program or instruction stored on the readable storage medium is executed by the processor to implement the three-dimensional point cloud-based metal product modeling method according to any one of claims 1-8.

Images