Point cloud data processing method, system and equipment for industrial three-dimensional model construction

Abstract

The invention relates to the technical field of point cloud processing, and discloses a point cloud data processing method, system and device for industrial three-dimensional model construction. The method comprises the following steps: carrying out denoising and down-sampling processing on original point cloud data of an industrial part to obtain preprocessed point cloud data, calculating a local geometric descriptor, converting the local geometric descriptor into an industrial feature recurrent plot through a recurrent plot algorithm, inputting the recurrent plot into a convolutional neural network to extract a high-dimensional geometric feature vector, and mapping the feature vectors into B-spline curved surface control parameters through a deep belief network, and performing point cloud reconstruction quality evaluation according to the control parameters to obtain geometric accuracy error data and curved surface reconstruction integrity data. According to the method, the problems that in existing industrial three-dimensional model construction, a point cloud data processing method lacks intelligent feature recognition, the reconstruction precision is insufficient, and manufacturing constraints are not fully considered are solved, and the recognition accuracy of geometric features of complex industrial parts and the precision of three-dimensional model reconstruction are improved.

Description

Technical Field

[0001] The present application relates to the field of point cloud processing technology, and in particular to a point cloud data processing method, system and device for constructing industrial three-dimensional models. Background Art

[0002] Point cloud data processing technology for building industrial 3D models is a crucial technology in modern manufacturing. It primarily acquires 3D point cloud data of industrial parts through methods such as laser scanning and structured light scanning. Algorithms such as point cloud registration, filtering, segmentation, and surface reconstruction are then used to convert the discrete point cloud data into a complete 3D digital model. Existing technologies typically use the Iterative Closest Point (ICP) algorithm for registration, Poisson reconstruction or marching cubes for surface reconstruction, and traditional geometric algorithms to identify and extract the geometric features of industrial parts. These methods achieve good results when processing simple geometric shapes.

[0003] However, existing technologies have obvious shortcomings when processing complex industrial parts: traditional algorithms are prone to distortion or loss of details when identifying and reconstructing complex geometric features such as thin-walled structures, sharp edges, and internal cavities. Especially for industrial parts with multi-scale geometric features, existing methods have difficulty adaptively identifying and reconstructing different types of geometric features; the point cloud processing process relies heavily on manual parameter adjustment, the degree of automation is not high, and the processing efficiency is low; existing algorithms lack consideration of industrial manufacturing constraints, and the reconstruction results often do not meet actual processing requirements.

[0004] Based on the above analysis, the core problem of existing technologies lies in the lack of intelligent feature recognition and reconstruction methods. Because traditional geometric algorithms cannot automatically learn and adapt to the characteristic patterns of different industrial parts, they suffer from inaccurate feature recognition when processing complex geometric structures. This inaccuracy in feature recognition further affects the quality of subsequent surface reconstruction, making it impossible for the reconstructed model to accurately preserve the geometric features of the original part. Furthermore, the lack of consideration of industrial manufacturing constraints makes the reconstruction results manufacturability issues in practical applications. Therefore, a point cloud data processing method is needed that can intelligently identify industrial features and incorporate manufacturing constraints for high-precision reconstruction. Summary of the Invention

[0005] The present application provides a point cloud data processing method, system and equipment for industrial three-dimensional model construction, which solves the problems of lack of intelligent feature recognition, insufficient reconstruction accuracy and insufficient consideration of manufacturing constraints in existing point cloud data processing methods for industrial three-dimensional model construction, and improves the recognition accuracy of geometric features of complex industrial parts and the accuracy of three-dimensional model reconstruction.

[0006] In the first aspect, the present application provides a point cloud data processing method for industrial three-dimensional model construction, which includes: denoising and downsampling the original point cloud data of industrial parts to obtain preprocessed point cloud data; calculating local geometric descriptors based on the preprocessed point cloud data, and performing two-dimensional image conversion processing on the local geometric descriptors through a recursive graph algorithm to obtain an industrial feature recursive graph; inputting the industrial feature recursive graph into a convolutional neural network for feature extraction processing to obtain a high-dimensional geometric feature vector; performing nonlinear mapping processing on the high-dimensional geometric feature vector through a deep belief network to obtain B-spline surface control parameters; performing point cloud reconstruction quality assessment processing based on the B-spline surface control parameters to obtain geometric accuracy error data and surface reconstruction integrity data.

[0007] In a second aspect, the present application provides a point cloud data processing system for industrial 3D model construction, the point cloud data processing system for industrial 3D model construction comprising: Denoising module, used to perform denoising and downsampling on the original point cloud data of industrial parts to obtain preprocessed point cloud data; a calculation module, configured to calculate a local geometric descriptor based on the pre-processed point cloud data, and perform two-dimensional image conversion processing on the local geometric descriptor using a recursive graph algorithm to obtain an industrial feature recursive graph; An extraction module, configured to input the industrial feature recursive graph into a convolutional neural network for feature extraction processing to obtain a high-dimensional geometric feature vector; A mapping module, configured to perform nonlinear mapping processing on the high-dimensional geometric feature vector through a deep belief network to obtain B-spline surface control parameters; The evaluation module is used to perform point cloud reconstruction quality evaluation processing according to the B-spline surface control parameters to obtain geometric accuracy error data and surface reconstruction completeness data.

[0008] In a third aspect, a point cloud data processing device for industrial three-dimensional model construction is provided, comprising: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor calls the instructions in the memory so that the point cloud data processing device for industrial three-dimensional model construction executes the above-mentioned point cloud data processing method for industrial three-dimensional model construction.

[0009] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned point cloud data processing method for constructing an industrial three-dimensional model.

[0010] The technical solution provided in this application effectively addresses the noise interference and data redundancy issues in traditional point cloud processing by denoising and downsampling the raw point cloud data of industrial parts. Statistical filtering based on manufacturing tolerance constraints ensures that the retained point cloud data meets industrial standards. Industrial feature scale-adaptive voxel grid downsampling dynamically adjusts processing accuracy based on different geometric features, preserving key details while controlling computational complexity. Local geometric descriptors are calculated based on the preprocessed point cloud data and converted into a two-dimensional image using a recursive graph algorithm. This innovatively converts three-dimensional geometric features into a two-dimensional image representation, overcoming the limitations of traditional geometric algorithms in complex feature recognition. The recursive graph algorithm can effectively capture the local geometric relationships and spatial distribution patterns of point clouds. The industrial feature recursive graph is then input into a convolutional neural network for feature extraction, generating a high-dimensional geometric feature vector. This fully leverages the advantages of deep learning in image feature extraction. The multi-layer convolutional structure automatically learns industrial geometric features at different scales, forming a hierarchical feature representation from local details to overall structure. Compared to traditional geometric feature extraction methods, this method offers greater adaptability and robustness.

[0011] By performing nonlinear mapping on high-dimensional geometric feature vectors using a deep belief network (DBN), the control parameters of the B-spline surface are obtained, enabling intelligent conversion from abstract features to specific geometric parameters. The multi-layered structure of the DNB is capable of learning complex nonlinear mapping relationships. Combined with industrial manufacturing constraints, the generated parameters are ensured to be practically manufacturable. The unsupervised learning properties of the restricted Boltzmann machine (RBM) enable the network to automatically discover implicit patterns and regularities in the data. The point cloud reconstruction quality is assessed based on the B-spline surface control parameters, generating geometric accuracy error data and surface reconstruction integrity data. A comprehensive quality assessment system is established, comprehensively evaluating the reconstruction results through multi-dimensional indicators such as geometric accuracy, shape integrity, and feature retention, ensuring that the reconstructed model meets the accuracy requirements and quality standards of industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0013] Figure 1 A schematic diagram of an embodiment of a point cloud data processing method for constructing an industrial three-dimensional model in an embodiment of the present application; Figure 2 A schematic diagram of an embodiment of a point cloud data processing system for constructing an industrial three-dimensional model in an embodiment of the present application; Figure 3 It is a schematic block diagram of the structure of a point cloud data processing device for constructing an industrial three-dimensional model in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The embodiments of the present application provide a point cloud data processing method, system and device for industrial three-dimensional model construction. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0015] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In one embodiment of the present application, a method for processing point cloud data for constructing an industrial three-dimensional model includes: Step S101: De-noising and down-sampling the original point cloud data of the industrial parts to obtain pre-processed point cloud data; Step S102: Calculate local geometric descriptors based on the pre-processed point cloud data, and perform two-dimensional image conversion processing on the local geometric descriptors using a recursive graph algorithm to obtain an industrial feature recursive graph; Step S103: inputting the industrial feature recursive graph into a convolutional neural network for feature extraction to obtain a high-dimensional geometric feature vector; Step S104: performing nonlinear mapping processing on the high-dimensional geometric feature vector through a deep belief network to obtain B-spline surface control parameters; Step S105 : performing point cloud reconstruction quality assessment processing according to the B-spline surface control parameters to obtain geometric accuracy error data and surface reconstruction integrity data.

[0016] It is understood that the execution subject of this application can be a point cloud data processing system for industrial 3D model construction, or a terminal or server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0017] Specifically, the raw point cloud data of industrial parts is denoised and downsampled. Manufacturing tolerance constraint statistical filtering identifies and removes noise points that exceed the manufacturing accuracy range by calculating the statistical distribution relationship between each point and its neighboring points. Manufacturing tolerance constraints refer to the allowable deviation range set by industrial part processing accuracy standards. Industrial feature scale adaptive voxel grid downsampling dynamically adjusts the voxel size based on the geometric dimensions of different industrial features. A voxel grid is a data structure that divides three-dimensional space into a regular grid of cubes. This reduces data volume by retaining representative points within each voxel. Calculating the industrial processing feature covariance matrix involves statistical analysis of the spatial distribution of point cloud data. The covariance matrix describes the degree of variation in point cloud data across different directions, and its eigenvalue ratios are used to identify different geometric feature types, such as edges, planes, and corners. Industrial semantic block segmentation combines point cloud data with the same manufacturing attributes. Manufacturing semantics refers to geometric feature attributes related to industrial processing. K-nearest neighbor index construction establishes a connection relationship between each point and its nearest neighbors. The k value represents the number of neighboring points and is used for subsequent topological analysis.

[0018] Multidimensional local geometric descriptors contain information in four dimensions: position coordinates, normal vectors, curvature values, and local density. The normal vector represents the vertical direction of the point cloud surface, the curvature value describes the degree of surface curvature, and the local density reflects the distribution density of the point cloud around the point. The industrial processing feature coordinate system is a spatial reference system established based on the primary processing direction of industrial parts. Euclidean distance calculation is used to measure the similarity between descriptor vectors. The manufacturing tolerance recursive threshold judgment process converts the distance matrix into a binary matrix by setting a distance threshold related to manufacturing accuracy. When the distance is less than the threshold, the value is assigned to 1, otherwise it is 0. Specialized coordinate system transformation processing uses corresponding coordinate transformation methods for different industrial features such as fillets, threads, and chamfers. Fillets are transformed using polar coordinates, threads using spiral coordinates, and chamfers using bevel coordinates. Industrial geometric invariance constraints ensure that the recursive graph maintains the geometric properties of industrial features after rotation, scaling, and translation.

[0019] The industrial feature recurrence graph is input into a convolutional neural network for feature extraction. Multiple convolutional layers perform convolution operations on the recurrence graph using convolution kernels of different sizes to extract geometric features at different levels. Convolution is the element-wise multiplication of the convolution kernel with a local area of ​​the image, followed by the summation. Industrial geometric constraint verification screens valid features by checking whether the extracted features conform to industrial geometric laws. Manufacturing constraints refer to geometric restrictions that must be met during industrial processing. The batch normalization layer standardizes the feature graph to eliminate differences between data batches. Standardization adjusts the data to a distribution with a mean of 0 and a variance of 1. Multi-scale feature fusion combines features extracted by different convolutional layers to form a comprehensive feature representation containing multi-level information. The fully connected layer performs high-dimensional mapping, converting the two-dimensional feature graph into a one-dimensional feature vector.

[0020] High-dimensional geometric feature vectors are nonlinearly mapped using a deep belief network. A restricted Boltzmann machine is a two-layer neural network structure that calculates the activation probability of hidden layer units using an energy function that describes the stability of the network state. Manufacturing constraint functions perform deviation checks on geometric parameters such as radius, thickness, and angle to ensure that the parameters meet industrial standards. Contrastive divergence learning is an unsupervised learning algorithm that iteratively updates network weights until convergence. The feature encoder decodes semantic features into control point coordinates, weight coefficients, and node vectors for B-spline surfaces, a mathematical representation used to describe complex surfaces. CAD geometry verification processing checks whether the generated control parameters meet geometric continuity and smoothness requirements.

[0021] Quality assessment is performed based on the control parameters of the B-spline surface. The surface reconstruction process generates a three-dimensional surface model based on the control points, weight coefficients, and node vectors. When the weight coefficients meet the convergence conditions, it indicates that the reconstruction process is stable. The distance calculation algorithm evaluates the geometric accuracy by calculating the point-to-surface distance between the reconstructed surface and the original point cloud. The manufacturing accuracy threshold is set according to industrial standards. The curvature continuity detection algorithm calculates the continuity of the first-order and second-order derivatives of the surface. The smoothness requirement ensures that the surface has no sudden changes. The key feature point positioning algorithm identifies the positions of important geometric features such as hole centers, edges, and corners. The tolerance standards are set according to the requirements of industrial drawings. The comprehensive evaluation algorithm performs weighted calculations on the results of various inspections to generate geometric accuracy error data and surface reconstruction completeness data.

[0022] In a specific embodiment, the process of executing step S101 may specifically include the following steps: Perform statistical filtering based on manufacturing tolerance constraints on the original point cloud data of industrial parts to obtain denoised point cloud data that meets industrial standards; Based on the industrial characteristic scale adaptive voxel grid, the denoised point cloud data is subjected to density gradient downsampling processing to obtain multi-level density distribution point cloud data; Calculate the industrial processing feature covariance matrix based on multi-level density distribution point cloud data to obtain geometric feature classification identification including edges, planes, and corners; The geometric features are classified and identified, and then industrial semantic block processing is performed to obtain a set of point cloud data blocks with manufacturing semantics. Based on the industrial topology continuity constraint, the point cloud data block set is processed by k-nearest neighbor index construction to obtain preprocessed point cloud data.

[0023] Specifically, the statistical filtering process based on manufacturing tolerance constraints first calculates the statistical distribution of the distance between each point in the original point cloud and its neighboring points. The manufacturing tolerance constraint refers to the allowable range of geometric deviations set according to the industrial part processing accuracy standards. The statistical filtering algorithm identifies outliers whose distance distribution exceeds the normal range by calculating the average distance and standard deviation from each point in the point cloud to its neighboring points. These outliers are usually noise generated by reflection errors or environmental interference during the laser scanning process. The algorithm traverses each point in the point cloud, searches for the nearest points in its neighborhood, and calculates the average distance from the point to the neighboring points. When the average distance of a point exceeds the threshold of the global average distance plus a multiple of the standard deviation, the point is marked as a noise point and removed from the point cloud. The manufacturing tolerance constraint ensures that the retained point cloud data meets the geometric accuracy standards of actual manufacturing by associating the statistical threshold with the industrial processing accuracy requirements. The denoised point cloud data retains the true geometric characteristics of industrial parts.

[0024] The industrial feature scale-adaptive voxel grid downsampling process dynamically adjusts the voxel size based on the geometric dimensions of different industrial features. A voxel grid is a data structure that divides three-dimensional space into regular cubic units. Density gradient downsampling first analyzes the density distribution of the point cloud in space, calculates the point cloud density in each local area, and then adjusts the resolution of the voxel grid based on the density gradient. Smaller voxel sizes are used in areas of high point cloud density to preserve detailed features, while larger voxel sizes are used in areas of low density to reduce data redundancy. The algorithm calculates the local density function of the point cloud to identify boundary regions with drastic density changes and planar regions with relatively uniform density, and samples these regions with appropriate voxel sizes. The industrial feature scale-adaptive mechanism automatically selects the appropriate voxel resolution based on the typical size range of different industrial features, such as circular holes, threads, and chamfers, ensuring that the downsampled point cloud maintains key geometric features while controlling data size.

[0025] Industrial processing feature covariance matrix calculation identifies different types of geometric features by analyzing the spatial distribution characteristics of multi-level density-distributed point cloud data. The covariance matrix is ​​a statistic that describes the degree of data variation in different directions. For each point and its neighborhood, the covariance values ​​in three directions are calculated to form a 3×3 matrix. The algorithm calculates the three eigenvalues ​​of the covariance matrix and determines the geometric feature type based on the ratio between them: when the first eigenvalue is 0.8, the second eigenvalue is 0.08, and the third eigenvalue is 0.008, the eigenvalue ratio is 100:10:1, indicating that the point cloud varies dramatically in a certain direction, corresponding to an edge feature. When the first eigenvalue is 0.5, the second eigenvalue is 0.48, and the third eigenvalue is 0.05, the eigenvalue ratio is 10:9.6:1. The first two eigenvalues ​​are close and much larger than the third eigenvalue, indicating that the point cloud is distributed in a plane, corresponding to a plane feature. When the first eigenvalue is 0.35, the second eigenvalue is 0.32, and the third eigenvalue is 0.3, the eigenvalue ratio is 1.17:1.07:1. The three eigenvalues ​​are close, indicating that the point cloud varies evenly in all directions, corresponding to a corner feature. Geometric feature classification identifiers are generated by quantizing the eigenvalue ratios. Each point is assigned a corresponding feature type label, forming a classified data structure containing edge, plane, and corner information.

[0026] Industrial semantic block processing merges point cloud data with the same geometric feature classification identifier into data blocks. Manufacturing semantics refers to attribute information related to industrial processing technology and geometric functions. The algorithm aggregates adjacent point cloud data with the same feature type into connected regions based on the geometric feature classification identifier. Each connected region represents a geometric feature with specific manufacturing semantics. Block processing uses a region growing algorithm, starting from a seed point and expanding to the surrounding area, gradually adding points with the same feature type and spatial proximity to the same data block until further expansion is impossible. Manufacturing semantic information includes the feature's geometric type, dimensional parameters, position information, and process attributes. Each data block carries a corresponding semantic label to form a collection of point cloud data blocks with industrial significance.

[0027] The k-nearest neighbor index construction process for industrial topological continuity constraints establishes a connection relationship between each point in a set of point cloud data blocks and its neighboring points. The k-nearest neighbor index is a spatial data structure that records the connection information of each point and its k nearest neighbor points. The k value is determined based on the geometric complexity and density distribution of the industrial features. The topological continuity constraint ensures that the connection relationship between adjacent points meets the geometric continuity requirements of industrial parts and avoids unreasonable jump connections. The algorithm first calculates the Euclidean distance from each point to all other points, and then selects the k points with the smallest distance as neighbors to establish an adjacency table. Industrial topological continuity is verified by checking the geometric rationality of the adjacency relationship, including distance constraints, angle constraints, and curvature continuity constraints. Only adjacency relationships that meet these constraints are retained in the index structure.

[0028] In a specific embodiment, the process of executing step S102 may specifically include the following steps: Construct a multi-dimensional local geometric descriptor containing position coordinates, normal vectors, curvature values ​​and local density based on the preprocessed point cloud data; Based on the industrial processing feature coordinate system, the Euclidean distance calculation is performed on the local geometric descriptor to obtain the industrial feature distance matrix; The industrial characteristic distance matrix is ​​processed by manufacturing tolerance recursive threshold judgment to obtain an industrial standard recursive binary matrix; According to different industrial feature types such as fillet, thread and chamfer, the recursive binary matrix is ​​transformed into a special coordinate system to obtain a recursive image of multiple types of industrial features; Based on the industrial geometric invariance constraint, the industrial feature recursive image is rotated, scaled and translated to obtain the industrial feature recursive graph.

[0029] Specifically, the multidimensional local geometric descriptor construction process extracts four key geometric attributes of each point from the preprocessed point cloud data. The position coordinates include the x, y, and z coordinate values ​​of the point in three-dimensional space. The normal vector is obtained by calculating the perpendicular direction of the plane fitted by the neighborhood point cloud of the point, indicating the orientation information of the surface at that point. The curvature value calculation involves analyzing the neighborhood geometric changes of the point. By fitting the local surface and calculating its second-order derivative, a quantitative value of the degree of surface curvature is obtained. A high curvature value indicates that the surface is severely curved, and a low curvature value indicates that the surface is relatively flat. The local density is calculated by counting the number of other points in the neighborhood of the point, reflecting the density of the point cloud distribution in the area. The algorithm traverses each point in the preprocessed point cloud, searches for neighboring points within a set radius, uses the least squares method to fit the local plane to calculate the normal vector, and simultaneously fits the quadratic surface to calculate the principal curvature and mean curvature. The local density is calculated by counting the number of neighborhood points. The four attribute values ​​are combined to form a 12-dimensional feature vector as the local geometric descriptor of the point.

[0030] The industrial processing feature coordinate system is a specialized spatial reference system established based on the primary processing directions and geometric features of industrial parts, distinct from the universal Cartesian coordinate system. The Euclidean distance calculation process calculates the straight-line distance between two points in a 12-dimensional feature vector space by comparing the similarities between the local geometric descriptors of different points. The algorithm traverses all point pairs, calculates the sum of the squares of the differences in the geometric descriptor vectors for each pair of points, and then takes the square root to obtain the Euclidean distance value. The industrial feature distance matrix is ​​a symmetric matrix with rows and columns corresponding to points in the point cloud. The matrix element values ​​represent the geometric similarity between corresponding point pairs. Smaller distance values ​​indicate more similar geometric features between two points, while larger distance values ​​indicate greater geometric differences. The algorithm then enters the Euclidean distance values ​​for all point pairs into the corresponding positions in the matrix to form a description of the distance relationship.

[0031] The manufacturing tolerance recursive thresholding process converts continuous distance values ​​into discrete binary data. Manufacturing tolerance is the allowable range of geometric deviations set based on industrial machining precision requirements. The recursive threshold is determined by analyzing the machining precision level of industrial parts. When the value of an element in the distance matrix is ​​less than or equal to the threshold, the corresponding position is assigned a value of 1, indicating similar geometric features between the two points. When the value of an element is greater than the threshold, it is assigned a value of 0, indicating significant geometric differences. This thresholding process generates an industrial standard recursive binary matrix containing only the values ​​0 and 1, forming a digital representation of the black and white pixel image. The algorithm iterates through each element of the distance matrix, binarizing it according to the threshold conditions, and generates a recursive binary matrix of the same size.

[0032] Dedicated coordinate system transformation processing uses corresponding mathematical transformation methods for different industrial feature types such as fillets, threads, and chamfers. Fillet features use polar coordinate system transformation to convert rectangular coordinates into radius and angle representations, highlighting the geometric characteristics of the arc; thread features use spiral coordinate system transformation, introducing spiral angle and pitch parameters to strengthen the geometric regularity of the spiral line; chamfer features use oblique coordinate system transformation, adjusting the coordinate axis direction according to the chamfer angle to optimize the edge feature representation. The algorithm selects the corresponding transformation matrix based on the geometric feature classification identifier and performs coordinate transformation operations on the recursive binary matrix. Multi-type industrial feature recursive images are generated by converting the transformed binary matrix into a grayscale image. Each industrial feature type corresponds to different image patterns and texture features.

[0033] Industrial geometry invariance constraints ensure that the recursive image maintains the essential geometric properties of the industrial feature after spatial transformation. Rotation transforms the image's angle around its center point through matrix multiplication, scaling adjusts the image's scale, and translation changes the image's position within the coordinate system. The algorithm performs multiple transformations on each industrial feature recursive image to generate a set of transformed images. Invariance constraints are verified by examining the consistency of the geometric features of the images before and after the transformation, including the degree of preservation of feature point positional relationships, geometric shape proportions, and topological connectivity. The industrial feature recursive graph synthesizes the results of these multiple transformations to generate a standardized image representation.

[0034] In a specific embodiment, the process of executing step S103 may specifically include the following steps: The industrial feature recursive graph is input into the multi-layer convolution layer for industrial geometric feature extraction to obtain a multi-scale industrial feature graph set; Based on the industrial geometric constraint module, the geometric consistency verification process of the multi-scale industrial feature map set is carried out to obtain the geometric feature map that meets the manufacturing constraints; Performing industrial normalization on the geometric feature map according to the batch normalization layer to obtain a standardized industrial feature map; The standardized industrial feature map is input into the multi-scale feature fusion module for industrial semantic fusion processing to obtain a fused industrial feature map; The fused industrial feature map is processed into high-dimensionality based on the fully connected layer to obtain a high-dimensional geometric feature vector.

[0035] Specifically, the multi-layer convolutional layer processing takes the industrial feature recursive graph as input and performs layer-by-layer feature extraction. The convolutional layer is the basic unit for image feature detection in neural networks. It extracts local features by performing a sliding window operation between the convolution kernel and the input image. During the industrial geometric feature extraction process, convolution kernels of different sizes detect geometric patterns at different scales. Small convolution kernels identify detailed features such as thread texture and surface roughness, medium-sized convolution kernels capture local geometric shapes such as fillets and chamfers, and large-sized convolution kernels extract global structural features such as the distribution of planar and curved surfaces. Each convolution layer contains multiple feature channels, each corresponding to a different type of industrial feature detector. Convolution operations generate feature maps. The multi-scale industrial feature map collection is composed of the outputs of convolution layers at different levels. Shallow convolution layers generate high-resolution detail feature maps, while deep convolution layers generate low-resolution semantic feature maps, forming a feature hierarchy from details to the whole. The algorithm passes the recursive graph through each convolution layer, applying an activation function after each convolution operation to enhance feature representation, thereby generating a collection of feature maps containing information at multiple scales.

[0036] The industrial geometric constraint verification process selects valid features by checking whether the multi-scale feature map set conforms to the geometric laws of industrial manufacturing. Geometric consistency verification includes three aspects: dimensional continuity check, angle rationality check, and surface smoothness check. Dimensional continuity ensures that the dimensional changes of adjacent features meet the processing requirements, angle rationality verifies that the angular relationship between features meets the industrial design standards, and surface smoothness checks the smooth transition of feature boundaries. Manufacturing constraints are verified using preset industrial standard parameters, including minimum processing radius, maximum tilt angle, surface roughness range and other restrictions. The algorithm traverses each pixel area in the feature map, calculates its geometric attribute parameters, and compares them with the manufacturing constraints. Feature areas that do not meet the constraints are marked as invalid and removed from the feature map. Geometric feature maps that meet manufacturing constraints only retain feature areas that have passed verification, ensuring that the feature data processed subsequently meets the feasibility requirements of actual industrial manufacturing.

[0037] The batch normalization layer performs industrial normalization on the geometric feature maps. Batch normalization is a technique used in deep learning to stabilize the training process by adjusting the distribution of feature data to eliminate differences between batches. Industrial normalization sets normalization parameters, including target values ​​for mean and variance, based on the precision level and quality standards of industrial manufacturing. The algorithm calculates the statistical characteristics of the feature maps for the current batch, including the mean and standard deviation for each channel, and then adjusts the feature values ​​to a standard distribution. The normalization process achieves a distribution with zero mean and unit variance by subtracting the mean and dividing by the standard deviation. Learnable scaling and offset parameters are then applied to adapt to the specific industrial application requirements. The standardized industrial feature maps have a uniform value range and distribution characteristics, eliminating scale differences between different industrial parts and making the feature representation more stable and consistent.

[0038] Multi-scale feature fusion integrates information at different scales within standardized industrial feature maps. Feature fusion is the process of combining multiple feature representations into a unified representation. Industrial semantic fusion organically combines detailed features with overall features by analyzing the semantic relevance of feature maps at different scales. The algorithm employs two fusion strategies: feature concatenation and weighted summation. Feature concatenation connects feature maps at different scales along the channel dimension to form a comprehensive feature representation containing multi-scale information. Weighted summation linearly combines features at different scales by assigning weight coefficients based on their importance. The fusion process addresses the differences in resolution between feature maps at different scales, adjusting the feature maps to a uniform size through upsampling or downsampling. The fused industrial feature map contains complete geometric information from local details to global structure, forming a hierarchical industrial feature representation.

[0039] The fully connected layer performs high-dimensional mapping, converting the two-dimensional fused industrial feature map into a one-dimensional, high-dimensional feature vector. A fully connected layer is a layer in a neural network where every input node is connected to every output node, achieving a linear transformation from input to output through matrix multiplication. The high-dimensional mapping process flattens the fused feature map, expanding the two-dimensional matrix into a one-dimensional vector, which is then linearly transformed using the fully connected layer's weight matrix. The algorithm calculates the product of the input vector and the weight matrix, adds a bias vector, and then applies an activation function to produce the output vector. The dimensionality of the high-dimensional geometric feature vector is determined by the complexity of the industrial application and contains an abstract representation of all the geometric information extracted and processed from the original recursive graph.

[0040] In a specific embodiment, the process of performing high-dimensional mapping processing on the fused industrial feature graph based on the fully connected layer may specifically include the following steps: The fused industrial feature map is flattened and expanded to obtain a one-dimensional feature sequence containing industrial processing feature information; Based on the first fully connected layer, the one-dimensional feature sequence is processed into industrial part geometric semantic encoding to obtain an industrial semantic encoding vector containing cylindrical surface, plane, and thread features; The industrial semantic coding vector is processed with manufacturing feature nonlinear mapping according to the ReLU activation function to obtain the activated industrial geometric feature vector; The activated industrial geometric feature vector is input into the second fully connected layer to perform industrial CAD feature dimension compression processing to obtain a compressed feature vector suitable for B-spline reconstruction; The compressed feature vector is normalized according to the industrial standard based on the L2 norm constraint to obtain a high-dimensional geometric feature vector.

[0041] Specifically, the flattening process converts the two-dimensional fused industrial feature map into a one-dimensional data structure. Flattening is the process of rearranging a multidimensional array into a one-dimensional vector. A fused industrial feature map typically has three dimensions: height, width, and number of channels. The flattening operation arranges all pixel values ​​into a continuous one-dimensional array in row-first order. The algorithm traverses each pixel position in the feature map, scanning row by row from the first row and first column, and sequentially adds all channel values ​​of each pixel to the one-dimensional sequence until the entire feature map is processed. The industrial processing feature information maintains its original numerical relationship during the flattening process, but only changes the organization of the data from a two-dimensional spatial representation to a linear sequence representation. The one-dimensional feature sequence contains all the geometric information in the original feature map, including geometric properties of industrial parts such as surface texture, edge contours, and local curvature. Each element in the sequence corresponds to the feature value of a specific position and channel in the feature map.

[0042] The geometric semantic encoding process for industrial parts converts a one-dimensional feature sequence into a coded representation with industrial semantic meaning through the first fully connected layer. The fully connected layer linearly transforms the input using a weight matrix and a bias vector. The weight parameters learn the correlations between different industrial features during training. The algorithm calculates the matrix multiplication of the input feature sequence and the weight matrix, with each output neuron corresponding to the weighted sum of all elements of the input sequence. The industrial semantic encoding process converts raw pixel-level features into semantic features with industrial significance. The different dimensions of the encoding vector correspond to specific industrial geometric feature types, such as cylinders, planes, and threads. The cylindrical feature encoding includes radius and axial information, the plane feature encoding includes the normal vector and plane equation parameters, and the thread feature encoding includes pitch, thread angle, and thread profile parameters. The industrial semantic encoding vector uses the weight parameters learned during training to map low-level pixel features to high-level semantic features.

[0043] The ReLU activation function performs a nonlinear transformation on the industrial semantic encoding vector. The ReLU function is a rectified linear unit activation function that performs piecewise linear processing on the input value. The manufacturing feature nonlinear mapping converts the linear encoding result into a nonlinear representation through the ReLU function. The ReLU function maintains the positive value unchanged and sets the negative value to zero. The mathematical expression is that the output is equal to the maximum value of the input and zero. The algorithm traverses each element of the encoding vector, retaining the original value when the element value is greater than zero and setting it to zero when the element value is less than or equal to zero. The nonlinear mapping introduces sparsity in the feature representation. Only the activated feature dimensions participate in subsequent calculations, suppressing the influence of negative features. The activated industrial geometric feature vector has stronger expressive power. The nonlinear transformation enables the network to learn complex industrial feature patterns and geometric relationships.

[0044] Industrial CAD feature dimension compression processing compresses the activated feature vector to a dimensional scale suitable for B-spline reconstruction through the second fully connected layer. Dimension compression is a dimensionality reduction technique that maps high-dimensional features to a low-dimensional space through linear transformation while maintaining important geometric information. The algorithm calculates the product of the activated feature vector and the compression weight matrix, where the number of rows in the compression weight matrix is ​​equal to the input dimension and the number of columns is equal to the target compression dimension. B-spline reconstruction requires a specific number of control parameters, including control point coordinates, weight coefficients, and node vectors. The dimension of the compressed feature vector matches the total number of these parameters. The compression process selects the most important feature dimensions through learned weight parameters and removes redundant information and noise interference. The compressed feature vector suitable for B-spline reconstruction contains the core geometric information required to build a three-dimensional model. Dimension compression balances the requirements of information preservation and computational efficiency.

[0045] L2-norm constrained normalization ensures that the compressed feature vectors meet industry-standard numerical range requirements. The L2 norm is the square root of the sum of the squares of all elements in a vector and is used to measure the magnitude of the vector. Industrial-standard normalization uses the L2-norm constraint to adjust the feature vectors to a standardized numerical range, preventing excessively large or small values ​​from affecting the accuracy of subsequent calculations. The algorithm first calculates the L2-norm value of the compressed feature vector and then divides each element of the vector by this norm value to obtain a normalized vector of unit length. Normalization eliminates scale differences between different industrial parts, ensuring that the feature representation has a unified numerical range and a benchmark for comparison. After L2-norm constraints, high-dimensional geometric feature vectors have stable numerical characteristics that meet the input requirements of deep belief networks.

[0046] In a specific embodiment, the process of executing step S104 may specifically include the following steps: The high-dimensional geometric feature vector is input into the restricted Boltzmann machine for energy function calculation. When the activation probability of the hidden layer unit is greater than the threshold, the activated hidden feature is obtained. Based on the manufacturing constraint function, the radius, thickness and angle deviation of the activated hidden features are judged and processed. When the deviation value is less than the standard tolerance, the qualified constraint feature is obtained; The qualified constraint features are input into the restricted Boltzmann machine for contrastive divergence learning, and stable semantic features are obtained when the weight update converges; According to the feature encoder, the control points, weights, and node vectors of the stable semantic features are decoded, and when the decoding error meets the accuracy requirements, the candidate B-spline parameters are obtained; The candidate B-spline parameters are subjected to CAD geometric verification based on the quality assessment module, and the B-spline surface control parameters are obtained when the geometric continuity and smoothness meet the preset standards.

[0047] Specifically, the energy function calculation process of the restricted Boltzmann machine uses high-dimensional geometric feature vectors as visible layer unit input. The restricted Boltzmann machine is a two-layer neural network structure consisting of a visible layer and a hidden layer. There are connection weights between the two layers, but no connections within the same layer. The energy function describes the stability of the network's current state and determines the system state by calculating the interaction energy between visible and hidden layer units. The algorithm assigns the high-dimensional geometric feature vectors to the visible layer units and then calculates the activation probability of each hidden layer unit. The activation probability is calculated as the weighted sum of the visible layer units using the sigmoid function. When the activation probability of a hidden layer unit exceeds a set threshold, the unit is activated and outputs 1; otherwise, it outputs 0. The threshold setting is determined based on the complexity and feature density of the industrial part. Complex parts use a lower threshold to activate more feature units, while simple parts use a higher threshold to reduce redundant activations. The activated hidden feature is composed of all activated hidden layer units, forming a sparse binary vector representation that contains the key geometric information extracted from the input features.

[0048] The manufacturing constraint function verifies the industrial manufacturing feasibility of activated hidden features. Manufacturing constraints include industrial standard requirements such as geometric dimensions, angular range, and surface quality. Radius deviation judgment checks whether the radius values ​​of arc and cylindrical features are within the allowable range. Thickness deviation judgment verifies whether the wall thickness and plate thickness meet the structural strength requirements. Angle deviation judgment confirms whether the chamfer and taper angles meet the processing technology restrictions. The algorithm decodes the activated hidden features into corresponding geometric parameters, obtains the standard tolerance range through table lookup or calculation, and then calculates the deviation between the actual parameters and the standard values. When the radius deviation is less than the radius allowable value of the standard tolerance, the thickness deviation is less than the thickness allowable value, and the angle deviation is less than the angle allowable value, the feature is marked as qualified. The qualified constraint feature only retains hidden units that pass all manufacturing constraint checks, filters out feature representations that do not meet industrial manufacturing requirements, and ensures that the subsequently generated B-spline parameters are practically feasible for processing.

[0049] Contrastive divergence learning optimizes network parameters by iteratively updating the connection weights of a restricted Boltzmann machine. Contrastive divergence is an unsupervised learning algorithm that adjusts weights by comparing the data distribution with the model distribution. The algorithm first reconstructs the visible layer units based on the qualified constraint features. Then, based on the reconstruction, it recalculates the hidden layer activations. This process is repeated multiple times to form a Gibbs sampling chain. Weights are updated by calculating the statistical difference between the original and reconstructed data. The update amount is equal to the learning rate multiplied by the gradient value. Learning stops when the change in weights over multiple consecutive iterations is less than a convergence threshold. Convergence of weight updates indicates that the network has learned the stable statistical characteristics of the data and that network parameters no longer change significantly. The stable semantic features represented by the converged hidden layer units contain the learned and optimized industrial geometric feature encoding, which has stronger expressive power and generalization performance.

[0050] The feature encoder decoding process converts stable semantic features into mathematical parameter representations of B-spline surfaces. The feature encoder is a trained neural network specifically designed to map abstract features into specific geometric parameters. Control point decoding extracts the three-dimensional coordinates of the control points of the B-spline surface from the semantic features. Weight decoding calculates the coefficient of influence of each control point on the surface shape. Node vector decoding determines the parameter sequence of the B-spline basis functions. The algorithm inputs the stable semantic features into the encoder network and generates the corresponding B-spline parameters through multiple layers of nonlinear transformations. The decoding error is calculated by comparing the difference between the decoded output and the target parameters. When the root mean square value of the decoding error is less than the required accuracy threshold, the decoding result is considered to meet the quality standard. The candidate B-spline parameters contain all the mathematical information required to construct the three-dimensional surface. The accuracy of the parameters directly affects the geometric accuracy of the final reconstructed model.

[0051] The quality assessment process verifies the geometric rationality and mathematical correctness of the candidate B-spline parameters. The quality assessment includes two aspects: geometric continuity check and surface smoothness check. Geometric continuity verifies whether the boundaries between adjacent surface patches are smoothly connected, and smoothness checks whether the first-order and second-order derivatives of the surface are continuous. The algorithm reconstructs the surface based on the B-spline parameters, calculates the curvature distribution and normal vector changes of the surface, and checks whether there are sudden changes or discontinuities. CAD geometry verification includes topological consistency check, dimensional accuracy check and shape rationality check to ensure that the generated surface meets industrial design requirements. When the geometric continuity index meets the continuity standard and the smoothness index meets the smoothness standard, the candidate parameters pass the quality assessment. The B-spline surface control parameters are composed of candidate parameters that have passed all verifications and contain the complete control information required to reconstruct the industrial 3D model.

[0052] In a specific embodiment, the process of executing step S105 may specifically include the following steps: Surface reconstruction is performed according to the B-spline surface control parameters, and a reconstructed surface model is obtained when the control point weight coefficients meet the convergence conditions; The point cloud comparison analysis of the reconstructed surface model is performed based on the distance calculation algorithm. When the average distance error is less than the manufacturing accuracy threshold, the geometric accuracy error data is obtained. The reconstructed surface model is input into the curvature continuity detection algorithm for shape accuracy verification. When the curvature change rate meets the smoothness requirement, the shape integrity evaluation result is obtained. The position deviation of the shape integrity evaluation results is calculated based on the key feature point positioning algorithm. When the position deviation of the hole center, edge, and corner point meets the tolerance standard, the feature retention rate data is obtained. The feature retention rate data is quantitatively processed based on the comprehensive evaluation algorithm to measure the model integrity. When both the feature retention rate and the model continuity meet the standards, the surface reconstruction integrity data is obtained.

[0053] Specifically, the B-spline surface reconstruction process generates a three-dimensional geometric model based on the control parameters. The B-spline surface is a parametric surface representation method defined by control points, weight coefficients and node vectors. The reconstruction algorithm first reads the three-dimensional coordinate data of the control points, then applies the corresponding weight coefficients to adjust the degree of influence of each control point on the surface shape, and finally determines the distribution of the basis function according to the node vectors. The convergence condition of the control point weight coefficient is judged by iteratively calculating the weight change. When the weight change of multiple consecutive iterations is less than the preset threshold, it is considered to have converged. The algorithm recalculates the position of the sampling point on the surface in each iteration and adjusts the weight coefficient by minimizing the distance error between the sampling point and the original point cloud. During the convergence process, the weight coefficient is gradually optimized until the surface shape is stable. At this time, the reconstructed surface model generated contains the main geometric features of the original point cloud. The reconstructed surface model is represented in the form of a triangular mesh or NURBS surface, and each surface point has accurate three-dimensional coordinates and normal vector information.

[0054] The distance calculation algorithm evaluates the reconstruction accuracy by comparing the spatial position difference between the reconstructed surface model and the original point cloud data. The point cloud comparative analysis process first establishes the correspondence between each point in the original point cloud and the nearest point on the reconstructed surface, and then calculates the perpendicular distance from the point to the surface. The algorithm traverses each point in the original point cloud, searches for the surface position closest to the point on the reconstructed surface, and calculates the Euclidean distance between the two. The average distance error is obtained by taking the arithmetic average of the distance values ​​of all points, reflecting the overall degree of deviation between the reconstructed model and the original data. The manufacturing accuracy threshold is set according to the processing accuracy level of industrial parts. Precision parts use a smaller threshold, and ordinary parts use a larger threshold. When the average distance error is less than the manufacturing accuracy threshold, the reconstructed model is considered to meet the geometric accuracy requirements. The geometric accuracy error data includes statistical indicators such as average error, maximum error, and standard deviation, which quantitatively describe the geometric accuracy of the reconstructed model.

[0055] The curvature continuity detection algorithm analyzes the local geometric characteristics of the reconstructed surface model. Curvature is a geometric quantity that describes the degree of curvature of a surface. Continuity detection verifies whether there are sudden changes or rough areas on the surface. Shape accuracy verification processing evaluates surface quality by calculating the principal curvature and average curvature distribution of the surface. The algorithm evenly samples detection points on the reconstructed surface, calculates the curvature value of each detection point, and then analyzes the curvature change rate between adjacent points. The curvature change rate is obtained by calculating the difference in the curvature values ​​of adjacent sampling points. Excessive change rate indicates that there are rough areas on the surface. Smoothness requirements are set according to surface quality standards for industrial applications. Aviation parts require higher smoothness, and mechanical parts require medium smoothness. The shape integrity evaluation results comprehensively consider the uniformity, continuity and smoothness indicators of the curvature distribution to generate evaluation data describing the surface shape quality.

[0056] The key feature point location algorithm identifies functionally important geometric locations within industrial parts. Key feature points include hole centers, edges, corners, and other locations that are crucial to part function and assembly. Position deviation calculation evaluates feature preservation accuracy by comparing the locations of feature points in the reconstructed model with the design reference locations. The algorithm automatically identifies hole centers within the reconstructed surface model and determines their coordinates by fitting a circular boundary. It then detects edge locations, identifying surface boundary lines based on curvature variations. Finally, it locates corner points, determining their coordinates by analyzing the intersections of multiple edges. Hole center position deviation is calculated by calculating the distance between the reconstructed and designed hole centers. Edge position deviation is calculated by calculating the average offset of edge lines. Corner position deviation is calculated by calculating the difference in corner coordinates. Tolerance standards are set based on the dimensional and positional tolerance requirements of the industrial drawing, with different feature types assigned different tolerance grades. Feature preservation rate data is calculated by statistically analyzing the proportion of feature points that meet tolerance requirements, reflecting the reconstructed model's ability to preserve key features.

[0057] The comprehensive evaluation algorithm combines the feature retention rate data with the model continuity index through weighted synthesis, and the comprehensive evaluation takes into account performance indicators in multiple dimensions such as geometric accuracy, shape quality, and feature integrity. The quantitative processing of model integrity converts each indicator into a unified numerical score by establishing an evaluation function. The algorithm sets weight coefficients for indicators such as feature retention rate, model continuity, and geometric accuracy, and adjusts the weight distribution according to specific application requirements. A feature retention rate that meets the standard indicates that the reconstructed model retains the key functional characteristics of the original part, and a model continuity that meets the standard indicates that the reconstructed surface has good mathematical properties. The comprehensive evaluation function calculates the overall score through weighted summation. When the overall score exceeds the preset standard, the reconstruction quality is considered qualified. The surface reconstruction integrity data contains the scores of each sub-indicator and the comprehensive score to form a quality assessment report.

[0058] The above describes the point cloud data processing method for industrial 3D model construction in the embodiment of the present application. The following describes the point cloud data processing system for industrial 3D model construction in the embodiment of the present application. Figure 2 In one embodiment of the present application, a point cloud data processing system for constructing an industrial three-dimensional model includes: Denoising module, used to perform denoising and downsampling on the original point cloud data of industrial parts to obtain preprocessed point cloud data; a calculation module, configured to calculate a local geometric descriptor based on the pre-processed point cloud data, and perform two-dimensional image conversion processing on the local geometric descriptor using a recursive graph algorithm to obtain an industrial feature recursive graph; An extraction module, configured to input the industrial feature recursive graph into a convolutional neural network for feature extraction processing to obtain a high-dimensional geometric feature vector; A mapping module, configured to perform nonlinear mapping processing on the high-dimensional geometric feature vector through a deep belief network to obtain B-spline surface control parameters; The evaluation module is used to perform point cloud reconstruction quality evaluation processing according to the B-spline surface control parameters to obtain geometric accuracy error data and surface reconstruction completeness data.

[0059] above Figure 2 The point cloud data processing system for industrial three-dimensional model construction in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The point cloud data processing device for industrial three-dimensional model construction in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0060] Reference Figure 3 In an embodiment of the present invention, a point cloud data processing device for industrial 3D model construction is further provided. The point cloud data processing device for industrial 3D model construction can be a server, and its internal structure can be as follows: Figure 3 As shown. The point cloud data processing device for industrial three-dimensional model construction includes a processor, a memory, a display screen, an input device, a network interface and a database connected via a system bus. Among them, the computer-designed processor is used to provide computing and control capabilities. The memory of the point cloud data processing device for industrial three-dimensional model construction includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the point cloud data processing device for industrial three-dimensional model construction is used to store the corresponding data in this embodiment. The network interface of the point cloud data processing device for industrial three-dimensional model construction is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0061] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the point cloud data processing device for industrial three-dimensional model construction to which the solution of the present invention is applied.

[0062] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the point cloud data processing method for constructing an industrial three-dimensional model.

[0063] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0064] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a point cloud data processing device for industrial three-dimensional model construction (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0065] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A point cloud data processing method for industrial 3D model construction, characterized in that: The method comprises: De-noising and down-sampling are performed on the original point cloud data of industrial parts to obtain pre-processed point cloud data; Calculating a local geometric descriptor based on the pre-processed point cloud data, and performing two-dimensional image conversion processing on the local geometric descriptor using a recursive graph algorithm to obtain an industrial feature recursive graph; Inputting the industrial feature recursive graph into a convolutional neural network for feature extraction processing to obtain a high-dimensional geometric feature vector; Performing nonlinear mapping processing on the high-dimensional geometric feature vector through a deep belief network to obtain B-spline surface control parameters; The point cloud reconstruction quality assessment process is performed according to the B-spline surface control parameters to obtain geometric accuracy error data and surface reconstruction completeness data.

2. The point cloud data processing method for industrial three-dimensional model construction according to claim 1, characterized in that: The denoising and downsampling processing of the original point cloud data of the industrial parts to obtain pre-processed point cloud data includes: Performing statistical filtering based on manufacturing tolerance constraints on the original point cloud data of the industrial parts to obtain denoised point cloud data that meets industrial standards; Performing density gradient downsampling processing on the denoised point cloud data based on an industrial characteristic scale-adaptive voxel grid to obtain multi-level density distribution point cloud data; Calculating the industrial processing feature covariance matrix based on the multi-level density distribution point cloud data to obtain geometric feature classification identification including edges, planes, and corner points; The geometric feature classification identifier is subjected to industrial semantic block processing to obtain a set of point cloud data blocks with manufacturing semantics; The point cloud data block set is subjected to k-nearest neighbor index construction processing based on industrial topology continuity constraints to obtain the pre-processed point cloud data.

3. The point cloud data processing method for industrial three-dimensional model construction according to claim 1, characterized in that: The method of calculating a local geometric descriptor based on the pre-processed point cloud data and performing two-dimensional image conversion processing on the local geometric descriptor by a recursive graph algorithm to obtain an industrial feature recursive graph includes: Constructing a multi-dimensional local geometric descriptor including position coordinates, normal vectors, curvature values ​​and local density according to the pre-processed point cloud data; Performing Euclidean distance calculation on the local geometric descriptor based on the industrial processing feature coordinate system to obtain an industrial feature distance matrix; The industrial characteristic distance matrix is ​​processed by manufacturing tolerance recursive threshold judgment to obtain an industrial standard recursive binary matrix; Performing a special coordinate system transformation process on the recursive binary matrix according to different industrial feature types such as fillets, threads, and chamfers to obtain a multi-type industrial feature recursive image; The industrial feature recursive image is rotated, scaled, and translated based on industrial geometric invariance constraints to obtain the industrial feature recursive graph.

4. The point cloud data processing method for industrial three-dimensional model construction according to claim 1, characterized in that: The industrial feature recursive graph is input into a convolutional neural network for feature extraction processing to obtain a high-dimensional geometric feature vector, including: Inputting the industrial feature recursive graph into a multi-layer convolutional layer to perform industrial geometric feature extraction processing to obtain a multi-scale industrial feature graph set; Performing geometric consistency verification on the multi-scale industrial feature graph set based on the industrial geometric constraint module to obtain a geometric feature graph that meets the manufacturing constraints; performing industrial normalization processing on the geometric feature map according to a batch normalization layer to obtain a standardized industrial feature map; Inputting the standardized industrial feature map into a multi-scale feature fusion module for industrial semantic fusion processing to obtain a fused industrial feature map; The fused industrial feature graph is subjected to high-dimensional mapping processing based on a fully connected layer to obtain the high-dimensional geometric feature vector.

5. The point cloud data processing method for industrial three-dimensional model construction according to claim 4, characterized in that: The performing high-dimensional mapping processing on the fused industrial feature graph based on the fully connected layer to obtain the high-dimensional geometric feature vector includes: Flattening and expanding the fused industrial feature graph to obtain a one-dimensional feature sequence containing industrial processing feature information; Performing industrial part geometric semantic coding processing on the one-dimensional feature sequence based on the first fully connected layer to obtain an industrial semantic coding vector containing cylindrical surface, plane, and thread features; Performing manufacturing feature nonlinear mapping processing on the industrial semantic coding vector according to the ReLU activation function to obtain an activated industrial geometric feature vector; Inputting the activated industrial geometric feature vector into the second fully connected layer to perform industrial CAD feature dimension compression processing to obtain a compressed feature vector suitable for B-spline reconstruction; The compressed feature vector is normalized according to an industrial standard based on an L2 norm constraint to obtain the high-dimensional geometric feature vector.

6. The point cloud data processing method for industrial three-dimensional model construction according to claim 1, characterized in that: The nonlinear mapping process is performed on the high-dimensional geometric feature vector by the deep belief network to obtain the B-spline surface control parameters, including: Inputting the high-dimensional geometric feature vector into a restricted Boltzmann machine for energy function calculation processing, and obtaining an activated hidden feature when the activation probability of the hidden layer unit is greater than a threshold; Performing radius, thickness, and angle deviation judgment processing on the activated hidden feature based on the manufacturing constraint function, and obtaining a qualified constraint feature when the deviation value is less than the standard tolerance; Inputting the qualified constraint features into a restricted Boltzmann machine for contrastive divergence learning processing, and obtaining stable semantic features when the weight update converges; Decoding the control points, weights, and node vectors of the stable semantic features according to the feature encoder, and obtaining candidate B-spline parameters when the decoding error meets the accuracy requirement; The candidate B-spline parameters are subjected to CAD geometric verification processing based on a quality assessment module, and the B-spline surface control parameters are obtained when the geometric continuity and smoothness meet preset standards.

7. The point cloud data processing method for industrial three-dimensional model construction according to claim 1, characterized in that: The point cloud reconstruction quality assessment process is performed according to the B-spline surface control parameters to obtain geometric accuracy error data and surface reconstruction integrity data, including: Performing surface reconstruction processing according to the B-spline surface control parameters, and obtaining a reconstructed surface model when the control point weight coefficients meet the convergence conditions; Performing point cloud comparison analysis on the reconstructed surface model based on a distance calculation algorithm, and obtaining the geometric accuracy error data when the average distance error is less than a manufacturing accuracy threshold; Inputting the reconstructed surface model into a curvature continuity detection algorithm for shape accuracy verification processing, and obtaining a shape integrity evaluation result when the curvature change rate meets the smoothness requirement; Performing position deviation calculation on the shape integrity evaluation result according to the key feature point positioning algorithm, and obtaining feature retention rate data when the position deviations of the hole center, edge, and corner point meet the tolerance standards; The feature retention rate data is subjected to model integrity quantification based on a comprehensive evaluation algorithm, and the surface reconstruction integrity data is obtained when both the feature retention rate and the model continuity meet the standards.

8. A point cloud data processing system for industrial 3D model construction, characterized in that: A method for processing point cloud data for constructing an industrial three-dimensional model according to any one of claims 1 to 7, wherein the point cloud data processing system for constructing an industrial three-dimensional model comprises: Denoising module, used to perform denoising and downsampling on the original point cloud data of industrial parts to obtain preprocessed point cloud data; a calculation module, configured to calculate a local geometric descriptor based on the pre-processed point cloud data, and perform two-dimensional image conversion processing on the local geometric descriptor using a recursive graph algorithm to obtain an industrial feature recursive graph; An extraction module, configured to input the industrial feature recursive graph into a convolutional neural network for feature extraction processing to obtain a high-dimensional geometric feature vector; A mapping module, configured to perform nonlinear mapping processing on the high-dimensional geometric feature vector through a deep belief network to obtain B-spline surface control parameters; The evaluation module is used to perform point cloud reconstruction quality evaluation processing according to the B-spline surface control parameters to obtain geometric accuracy error data and surface reconstruction completeness data.

9. A point cloud data processing device for industrial 3D model construction, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the method for processing point cloud data for constructing an industrial three-dimensional model according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is caused to execute the point cloud data processing method for industrial three-dimensional model construction according to any one of claims 1 to 7.

Images