A method and device for reconstructing a point cloud of a curved cylindrical structure, and a storage medium

Abstract

The application provides a curved cylindrical structure point cloud reconstruction method and device and a storage medium, and belongs to the technical field of point cloud reconstruction.The method comprises the following steps: obtaining original curved cylindrical three-dimensional point cloud data and to-be-reconstructed curved cylindrical three-dimensional point cloud data from a three-dimensional scanning device; and pre-processing the original curved cylindrical three-dimensional point cloud data and the to-be-reconstructed curved cylindrical three-dimensional point cloud data to obtain a first pre-processed curved cylindrical three-dimensional point cloud matrix, a second pre-processed curved cylindrical three-dimensional point cloud matrix, to-be-reconstructed scale parameters and to-be-reconstructed centroid coordinates.The application effectively overcomes the problems of topological information loss and mean collapse caused by traditional pooling, completely avoids the mean collapse problem of the prediction control point gathering to the centroid, greatly reduces the training difficulty of the network, and improves the generalization ability to different scanning resolutions, noises and workpiece specifications in an industrial field.

Description

Technical Field

[0001] This invention relates to the field of point cloud reconstruction technology, specifically to a method, apparatus, and storage medium for reconstructing point clouds of curved cylindrical structures. Background Technology

[0002] Curved cylindrical structures (such as aerospace ducts, industrial oil and gas pipelines, and complex curved mechanical shafts) are core geometric components in industrial manufacturing and engineering construction. Their high-precision 3D reverse reconstruction has extremely important application value in digital twins, quality inspection, and intelligent robot grasping. With the popularization of 3D laser scanning and lidar technologies, it has become more convenient to acquire high-density 3D point cloud data of object surfaces. However, how to efficiently and robustly extract the accurate geometric parameters (such as the central axis parameter and cross-sectional radius) of curved cylinders directly from discrete, disordered, and often noisy scanned point clouds remains a current challenge.

[0003] In existing technologies, point cloud reconstruction methods for curved cylindrical structures based on deep learning often employ a phased strategy of "component detection + model fitting." These methods first require complex point cloud segmentation and component identification, either manually or according to rules, followed by piecewise fitting of the cylindrical surface using the least squares method. However, traditional geometric fitting schemes are extremely sensitive to point cloud noise, local missing values, and initial pose. Furthermore, when dealing with complex curved cylinders with large areas and continuously varying curvature, the cumulative error makes it difficult to guarantee global reconstruction accuracy, thus failing to achieve truly efficient end-to-end regression.

[0004] In recent years, end-to-end point cloud reconstruction networks for curved cylindrical structures have been introduced into geometric parameter regression tasks. However, practice has shown that such classic backbone networks have significant inherent defects when applied to curved cylindrical reconstruction tasks: First, these networks typically use a point-by-point multilayer perceptron combined with a single global max pooling to extract point cloud features. Global max pooling directly compresses rich spatial topological information and local morphological features into a one-dimensional global vector, resulting in the obliteration of crucial geometric information at endpoints and boundaries. Second, under the optimization pressure of the mean squared error loss function, due to the lack of local morphological information, the network often exhibits a severe "mean collapse" phenomenon, that is, it tends to learn a "safe mean" close to the centroid of the point cloud, causing the predicted curve control points to incorrectly cluster near the origin and fail to extend towards the ends of the curved cylinder and the bending direction. Third, commonly used point cloud distance metrics algorithms (such as chamfer distance or Hungarian matching) only focus on the spatial proximity of the point set, ignoring the strict order of the control points. For the central axis of a curved cylinder represented by a parametric curve (such as a B-spline), the order of the control points directly determines the tangent direction and topological trend of the curve. If a non-order-preserving distance function is used, the network will fail to converge effectively, thus preventing high-precision parametric reconstruction. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method, apparatus and storage medium for reconstructing point clouds of curved cylindrical structures, which addresses the shortcomings of the prior art.

[0006] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A method for reconstructing point clouds of curved cylindrical structures, comprising the following steps: Obtain the original 3D point cloud data of the curved cylinder and the 3D point cloud data of the curved cylinder to be reconstructed from the 3D scanning equipment; The original three-dimensional point cloud data of the curved cylinder and the three-dimensional point cloud data of the curved cylinder to be reconstructed are preprocessed to obtain a first preprocessed three-dimensional point cloud matrix of the curved cylinder corresponding to the original three-dimensional point cloud data, a second preprocessed three-dimensional point cloud matrix of the curved cylinder corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, a scale parameter to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and centroid coordinates to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed. A training model is constructed, and the training model is analyzed using the first preprocessed three-dimensional point cloud matrix of the curved cylinder to obtain a cross-section prediction model. The cross-section prediction model is used to predict the second preprocessed three-dimensional point cloud matrix of the curved cylinder to obtain the first predicted cross-section radius corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed. The first predicted cross-sectional radius, the scale parameters to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, the centroid coordinates to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed are reconstructed and analyzed. The reconstruction result of the curved cylinder structure point cloud is obtained based on the reconstruction analysis result.

[0007] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: A point cloud reconstruction device for a curved cylindrical structure, comprising: The point cloud data acquisition module is used to obtain the original 3D point cloud data of the curved cylinder and the 3D point cloud data of the curved cylinder to be reconstructed from the 3D scanning device. The preprocessing module is used to preprocess the original curved cylindrical 3D point cloud data and the curved cylindrical 3D point cloud data to be reconstructed to obtain a first preprocessed curved cylindrical 3D point cloud matrix corresponding to the original curved cylindrical 3D point cloud data, a second preprocessed curved cylindrical 3D point cloud matrix corresponding to the curved cylindrical 3D point cloud data to be reconstructed, a scale parameter to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed, and centroid coordinates to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed. The model analysis module is used to construct a training model and perform model analysis on the training model using the first preprocessed curved cylindrical 3D point cloud matrix to obtain a cross-section prediction model. The prediction module is used to predict the second preprocessed curved cylinder three-dimensional point cloud matrix through the cross-section prediction model to obtain the first predicted cross-section radius corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed. The reconstruction result module is used to perform reconstruction analysis on the first predicted cross-sectional radius, the scale parameters to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, the centroid coordinates to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and to obtain the point cloud reconstruction result of the curved cylinder structure based on the reconstruction analysis result.

[0008] Based on the above-mentioned method for reconstructing point clouds of curved cylindrical structures, the present invention also provides a system for reconstructing point clouds of curved cylindrical structures.

[0009] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a curved cylindrical structure point cloud reconstruction system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the curved cylindrical structure point cloud reconstruction method as described above.

[0010] Based on the above-mentioned method for reconstructing point clouds of curved cylindrical structures, the present invention also provides a computer-readable storage medium.

[0011] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the point cloud reconstruction method for curved cylindrical structures as described above.

[0012] The beneficial effects of this invention are as follows: By obtaining original 3D point cloud data of a curved cylinder and 3D point cloud data of a curved cylinder to be reconstructed from a 3D scanning device, preprocessing the original and reconstructed 3D point cloud data of the curved cylinder yields a first preprocessed 3D point cloud matrix, a second preprocessed 3D point cloud matrix of the curved cylinder, the scale parameters to be reconstructed, and the centroid coordinates to be reconstructed. Model analysis of the training model using the first preprocessed 3D point cloud matrix of the curved cylinder yields a cross-section prediction model. Prediction of the second preprocessed 3D point cloud matrix of the curved cylinder using the cross-section prediction model yields a first predicted cross-section radius and a first predicted control point matrix. The cross-sectional radius, the scale parameters to be reconstructed, the coordinates of the centroid to be reconstructed, and the matrix of the first predicted control point are used for reconstruction analysis. Based on the reconstruction analysis results, the point cloud reconstruction results of the curved cylindrical structure are obtained. This effectively overcomes the problems of topological information loss and mean collapse caused by traditional pooling. It realizes efficient and high-precision direct regression from disordered point cloud to the geometric parameters of the curved cylinder, greatly improves the robustness of inverse reconstruction, overcomes the information loss caused by single pooling, completely avoids the mean collapse problem of predicted control points gathering towards the centroid, greatly reduces the training difficulty of the network, and improves the generalization ability to different scanning resolutions, noise, and workpiece specifications in industrial sites. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the point cloud reconstruction method for curved cylindrical structures provided in an embodiment of the present invention. Figure 2 A schematic diagram of the training model for the point cloud reconstruction method of curved cylindrical structure provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the dynamic graph convolutional neural network encoder for the point cloud reconstruction method of curved cylindrical structure provided in an embodiment of the present invention. Figure 4 A schematic diagram of the reconstruction analysis of the point cloud reconstruction method for curved cylindrical structures provided in an embodiment of the present invention; Figure 5 A flowchart of a point cloud reconstruction method for a curved cylindrical structure provided in an embodiment of the present invention; Figure 6 This is a structural block diagram of the point cloud reconstruction method for curved cylindrical structures provided in an embodiment of the present invention; Figure 7 The convergence curve of the model training loss for the point cloud reconstruction method of the curved cylindrical structure provided in this embodiment of the invention; Figure 8 A schematic diagram comparing the three-dimensional reconstruction effect of curved cylindrical point cloud structure point cloud provided in the embodiment of the present invention with that of the prior art. Figure 9 This is a block diagram of a point cloud reconstruction device for a curved cylindrical structure provided in an embodiment of the present invention. Detailed Implementation

[0014] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0015] Figure 1 This is a flowchart illustrating a point cloud reconstruction method for a curved cylindrical structure provided in an embodiment of the present invention.

[0016] like Figure 1 As shown, a method for reconstructing point clouds of curved cylindrical structures includes the following steps: S1: Obtain the original 3D point cloud data of the curved cylinder and the 3D point cloud data of the curved cylinder to be reconstructed from the 3D scanning device; S2: Preprocess the original curved cylindrical 3D point cloud data and the curved cylindrical 3D point cloud data to be reconstructed to obtain a first preprocessed curved cylindrical 3D point cloud matrix corresponding to the original curved cylindrical 3D point cloud data, a second preprocessed curved cylindrical 3D point cloud matrix corresponding to the curved cylindrical 3D point cloud data to be reconstructed, a scale parameter to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed, and centroid coordinates to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed. S3: Construct a training model, and perform model analysis on the training model using the first preprocessed curved cylindrical 3D point cloud matrix to obtain a cross-section prediction model; S4: The second preprocessed curved cylinder three-dimensional point cloud matrix is ​​predicted by the cross-section prediction model to obtain the first predicted cross-section radius corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed. S5: Perform reconstruction analysis on the first predicted cross-sectional radius, the scale parameters to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, the centroid coordinates to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and obtain the point cloud reconstruction result of the curved cylinder structure based on the reconstruction analysis results.

[0017] It should be understood that the original three-dimensional point cloud data of the curved cylinder to be reconstructed (i.e., the original three-dimensional point cloud data of the curved cylinder and the three-dimensional point cloud data of the curved cylinder to be reconstructed) is obtained. This data is usually a coordinate matrix from a three-dimensional scanning device (i.e., a three-dimensional scanning device), which contains M unordered spatial discrete points.

[0018] In the above embodiments, the original three-dimensional point cloud data of the curved cylinder and the three-dimensional point cloud data of the curved cylinder to be reconstructed are obtained from a three-dimensional scanning device. Preprocessing of the original and reconstructed three-dimensional point cloud data yields a first preprocessed three-dimensional point cloud matrix, a second preprocessed three-dimensional point cloud matrix, the scale parameters to be reconstructed, and the centroid coordinates to be reconstructed. Model analysis of the training model using the first preprocessed three-dimensional point cloud matrix yields a cross-section prediction model. Prediction of the second preprocessed three-dimensional point cloud matrix using the cross-section prediction model yields a first predicted cross-section radius and a first predicted control point matrix. The cross-sectional radius, the scale parameters to be reconstructed, the coordinates of the centroid to be reconstructed, and the matrix of the first predicted control point are reconstructed. Based on the reconstruction analysis results, the point cloud reconstruction results of the curved cylindrical structure are obtained. This effectively overcomes the problems of topological information loss and mean collapse caused by traditional pooling. It realizes efficient and high-precision direct regression from disordered point cloud to the geometric parameters of the curved cylinder, greatly improves the robustness of inverse reconstruction, overcomes the information loss caused by single pooling, completely avoids the mean collapse problem of predicted control points gathering towards the centroid, greatly reduces the training difficulty of the network, and improves the generalization ability to different scanning resolutions, noise, and workpiece specifications in industrial sites.

[0019] Optionally, as an embodiment of the present invention, the original three-dimensional point cloud data of the curved cylinder includes multiple discrete points of the original three-dimensional point cloud of the curved cylinder, and the three-dimensional point cloud data of the curved cylinder to be reconstructed includes multiple discrete points of the three-dimensional point cloud of the curved cylinder to be reconstructed. The process of preprocessing the original curved cylindrical 3D point cloud data and the curved cylindrical 3D point cloud data to be reconstructed to obtain a first preprocessed curved cylindrical 3D point cloud matrix corresponding to the original curved cylindrical 3D point cloud data, a second preprocessed curved cylindrical 3D point cloud matrix corresponding to the curved cylindrical 3D point cloud data to be reconstructed, the scale parameters to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed, and the centroid coordinates to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed includes: According to the preset sampling scale, all the original three-dimensional point cloud discrete points of the curved cylinder are sampled and processed to obtain multiple first-sampled three-dimensional point cloud discrete points of the curved cylinder corresponding to the original three-dimensional point cloud data of the curved cylinder. According to the preset sampling scale, all discrete points of the three-dimensional point cloud of the cylindrical body to be reconstructed are sampled to obtain multiple discrete points of the three-dimensional point cloud of the cylindrical body after second sampling, which correspond to the three-dimensional point cloud data of the cylindrical body to be reconstructed. The first formula is used to calculate the original centroid coordinates corresponding to the original 3D point cloud data of the curved cylinder and the original centroid coordinates corresponding to the original 3D point cloud data of the curved cylinder, as well as the original centroid coordinates corresponding to the original 3D point cloud data of the curved cylinder and the original centroid coordinates corresponding to the original 3D point cloud data of the curved cylinder. The first formula is: , in, The original centroid coordinates, For the coordinates of the centroid to be reconstructed, , This represents the number of discrete points in the 3D point cloud of the curved cylinder after the first sampling or the number of discrete points in the 3D point cloud of the curved cylinder after the second sampling. The first point cloud data corresponding to the original curved cylinder Discretized points of the three-dimensional point cloud of the curved cylinder after the first sampling. The first point cloud data corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed The second sampled three-dimensional point cloud discrete points of the curved cylinder; The difference between each discrete point of the first sampled cylindrical three-dimensional point cloud and the original centroid coordinates is calculated to obtain multiple first translated cylindrical three-dimensional point cloud coordinates corresponding to the original cylindrical three-dimensional point cloud data. The difference between each discrete point of the second sampled cylindrical three-dimensional point cloud and the centroid coordinates to be reconstructed are calculated to obtain multiple second translated cylindrical three-dimensional point cloud coordinates corresponding to the cylindrical three-dimensional point cloud data to be reconstructed. The second formula is used to calculate the coordinates of the original three-dimensional point cloud data of the curved cylinder corresponding to multiple first-translated three-dimensional point cloud data and the coordinates of the curved cylinder corresponding to multiple second-translated three-dimensional point cloud data to be reconstructed, thereby obtaining the original scale parameters corresponding to the original three-dimensional point cloud data of the curved cylinder and the scale parameters to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed. The second formula is: , in, These are the original scale parameters. The scale parameters to be reconstructed are... , The first point cloud data corresponding to the original curved cylinder The coordinates of the first translated curved cylinder in three-dimensional point cloud. The first point cloud data corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed The coordinates of the second translated curved cylinder in three-dimensional point cloud. It is the 2-norm of a three-dimensional vector; The third equation is used to calculate the original scale parameters, the scale parameters to be reconstructed corresponding to the 3D point cloud data of the curved cylinder to be reconstructed, the multiple first translated 3D point cloud coordinates of the curved cylinder corresponding to the original 3D point cloud data, and the multiple second translated 3D point cloud coordinates of the curved cylinder corresponding to the 3D point cloud data of the curved cylinder to be reconstructed. Based on the calculation results, a first preprocessed 3D point cloud matrix of the curved cylinder corresponding to the original 3D point cloud data and a second preprocessed 3D point cloud matrix of the curved cylinder corresponding to the 3D point cloud data of the curved cylinder to be reconstructed are constructed. The third equation is: , in, The first point cloud data corresponding to the original curved cylinder The first preprocessed three-dimensional point cloud coordinates of the curved cylinder The first point cloud data corresponding to the original curved cylinder The second preprocessed three-dimensional point cloud coordinates of the curved cylinder The first point cloud data corresponding to the original curved cylinder The coordinates of the first translated curved cylinder in three-dimensional point cloud. The first point cloud data corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed The coordinates of the second translated curved cylinder in three-dimensional point cloud. These are the original scale parameters. These are the scale parameters to be reconstructed.

[0020] It should be understood that the original three-dimensional point cloud data of the curved cylinder (i.e., the original three-dimensional point cloud data of the curved cylinder and the three-dimensional point cloud data of the curved cylinder to be reconstructed) is subjected to adaptive preprocessing of pose and scale to obtain normalized point cloud data (i.e., the three-dimensional point cloud matrix of the curved cylinder after the first preprocessing and the three-dimensional point cloud matrix of the curved cylinder after the second preprocessing).

[0021] Specifically, the original 3D point cloud data (i.e., the original discrete points of the curved cylindrical 3D point cloud and the discrete points of the curved cylindrical 3D point cloud to be reconstructed) is sampled to a fixed number of points; the centroid coordinates of the sampled point cloud (i.e., the original centroid coordinates and the centroid coordinates to be reconstructed) are calculated; the centroid coordinates (i.e., the original centroid coordinates and the centroid coordinates to be reconstructed) are subtracted from all point cloud coordinates (i.e., the original centroid coordinates and the centroid coordinates to be reconstructed) to complete the decentering process, and the centroid coordinates (i.e., the original centroid coordinates and the centroid coordinates to be reconstructed) are recorded for subsequent reverse reconstruction; the maximum distance from the decentered point cloud to the origin is calculated, the point cloud coordinates are divided by the maximum distance to complete the scale normalization, and the maximum distance is recorded as the scale parameter (i.e., the original scale parameter and the scale parameter to be reconstructed).

[0022] It should be understood that the original 3D point cloud data of the curved cylinder (i.e., the original 3D point cloud discrete points of the curved cylinder and the 3D point cloud discrete points of the curved cylinder to be reconstructed) is downsampled with a fixed number of points, and centroid alignment decentering and scale normalization are performed in sequence to generate normalized point cloud data (i.e., the first preprocessed 3D point cloud matrix of the curved cylinder and the second preprocessed 3D point cloud matrix of the curved cylinder). At the same time, the centroid coordinates subtracted during decentering and the maximum distance parameter divided during scale normalization are recorded, i.e., the scale parameters (i.e. the original scale parameters and the scale parameters to be reconstructed).

[0023] Specifically, using the farthest point sampling algorithm or random sampling method, the original 3D point cloud data (i.e., the original discrete points of the curved cylindrical 3D point cloud and the discrete points of the curved cylindrical 3D point cloud to be reconstructed) of any initial point size are uniformly normalized to a preset number of key points; the centroid coordinates of the normalized point cloud (i.e., the original centroid coordinates and the centroid coordinates to be reconstructed) are calculated, and the centroid coordinates of all key points (i.e., the original discrete points of the curved cylindrical 3D point cloud and the discrete points of the curved cylindrical 3D point cloud to be reconstructed) are subtracted from the centroid coordinates (i.e., the original centroid coordinates and the centroid coordinates to be reconstructed) to complete the translation. Alignment is achieved by placing the centroid of the translated point cloud (i.e., the coordinates of the first and second translated cylindrical 3D point cloud) at the origin. The Euclidean distances from each key point to the origin are calculated, and the maximum distance value is selected as the scale parameter (i.e., the original scale parameter and the scale parameter to be reconstructed). The coordinates of the translated point cloud (i.e., the coordinates of the first and second translated cylindrical 3D point cloud) are simultaneously divided by this scale parameter (i.e., the original scale parameter and the scale parameter to be reconstructed) to complete the scale normalization.

[0024] It should be understood that, in order to ensure the uniformity of the tensor size of the network input and improve the efficiency of the baseline calculation, the original point cloud (i.e., the discrete points of the original curved cylindrical 3D point cloud and the discrete points of the curved cylindrical 3D point cloud to be reconstructed) is first resampled and normalized to the preset number of key points. Considering the need for representing local surface features, this invention sets the target sampling scale to be .

[0025] Specifically, the resampling process supports dynamic switching between two strategies. When the focus is on the uniformity of global point cloud coverage, the farthest point sampling algorithm is used. This algorithm selects the extreme value seed point by iteratively maintaining the minimum geometric distance from each point to the set of selected points. When the input scale... When there is a huge amount of data and an urgent need to accelerate data flow, a seamless switch to a random sampling strategy is made.

[0026] It should be understood that this invention incorporates an anomaly fallback protection mechanism. This mechanism will prevent issues arising when the number of input raw 3D point cloud data (i.e., the original discrete points of the curved cylindrical 3D point cloud and the discrete points of the curved cylindrical 3D point cloud to be reconstructed) exceeds a certain threshold. Less than the preset quantity In this case, the algorithm automatically employs a repeated random sampling strategy with replacement to replenish the number of points. This mechanism effectively avoids process crashes caused by memory addressing out-of-bounds errors during offline training or online inference.

[0027] Specifically, after resampling, for a scale of The point cloud (i.e., the discrete points of the original curved cylindrical 3D point cloud and the discrete points of the curved cylindrical 3D point cloud to be reconstructed) undergoes centroid alignment and decentering. Calculation... Geometric centroid coordinates of key points in three-dimensional space (i.e., the original centroid coordinates and the centroid coordinates to be reconstructed), and uniformly subtract the centroid from the 3D coordinates of all key points (i.e., the original discrete points of the 3D point cloud of the curved cylinder and the discrete points of the 3D point cloud of the curved cylinder to be reconstructed). (i.e., the original centroid coordinates and the centroid coordinates to be reconstructed), so that the point cloud as a whole is translated to the origin. This step effectively eliminates the translation variance caused by the initial positions of different reconstructed targets in the real-world coordinate system. This invention requires the centroid vector to be... The original centroid coordinates and the centroid coordinates to be reconstructed are placed in a cache space for subsequent inverse spatial pose reconstruction calculations. Then, the decentralized data (i.e., the 3D point cloud coordinates of the first and second translated curved cylinders) are subjected to maximum distance scale normalization. The maximum Euclidean distance from each keypoint to the origin after decentralization is calculated and defined as the scale bounding radius. (i.e., scale parameters) (i.e., original scale parameters and scale parameters to be reconstructed). To avoid division-by-zero anomalies in extreme degraded data cases, if the calculated... (530), then force it to be set to a constant of 1.0. Divide the coordinates of each translated point cloud (i.e., the coordinates of the first translated curved cylinder 3D point cloud and the coordinates of the second translated curved cylinder 3D point cloud) by this constant. (i.e., the original scale parameters and the scale parameters to be reconstructed), so that the point cloud is strictly and proportionally mapped into the unit sphere space, resulting in normalized point cloud data (i.e., the first preprocessed curved cylindrical 3D point cloud matrix and the second preprocessed curved cylindrical 3D point cloud matrix). The final output of this preprocessing step is in the form of... The two-dimensional tensor, with the three-dimensional channel dimension first, is used to precisely match the one-dimensional convolution operation of the subsequent framework.

[0028] In the above embodiments, the original three-dimensional point cloud data of the curved cylinder and the three-dimensional point cloud data of the curved cylinder to be reconstructed are preprocessed to obtain the first preprocessed three-dimensional point cloud matrix of the curved cylinder, the second preprocessed three-dimensional point cloud matrix of the curved cylinder, the scale parameters to be reconstructed, and the centroid coordinates to be reconstructed. This effectively overcomes the problems of topological information loss and mean collapse caused by traditional pooling, and realizes efficient and high-precision direct regression from disordered point cloud to the geometric parameters of the curved cylinder, greatly improving the robustness of inverse reconstruction.

[0029] Optionally, as an embodiment of the present invention, such as Figure 2 As shown, the training model includes a dynamic graph convolutional neural network encoder and a dual-head multilayer perceptron network; The process of performing model analysis on the trained model using the first preprocessed three-dimensional point cloud matrix of the curved cylinder to obtain the cross-section prediction model includes: The dynamic graph convolutional neural network encoder performs feature analysis on the first preprocessed curved cylindrical 3D point cloud matrix to obtain a global feature descriptor. The global feature descriptor is decoupled and analyzed by the dual-head multilayer perceptron network to obtain the second predicted cross-sectional radius and the second predicted control point matrix. Import the actual control point matrix and the actual cross-section radius, and perform loss function analysis on the second predicted cross-section radius, the second predicted control point matrix, the actual control point matrix, and the actual cross-section radius to obtain the target loss function; The parameters of the trained model are updated according to the target loss function to obtain the cross-section prediction model.

[0030] Specifically, the normalized point cloud data (i.e., the first preprocessed curved cylindrical 3D point cloud matrix) is input into a pre-constructed dynamic graph convolutional neural network encoder (i.e., dynamic graph convolutional neural network encoder) to extract local and global geometric features and generate a global feature descriptor; the global feature descriptor is input into a dual-head multilayer perceptron network, and through the mutually decoupled centerline regression branch and radius regression branch, the joint regression outputs the centerline control points (i.e., the second predicted control point matrix) and cross-sectional radius (i.e., the second predicted cross-sectional radius) in the normalized coordinate system.

[0031] In the above embodiments, the cross-section prediction model is obtained by performing model analysis on the training model through the three-dimensional point cloud matrix of the curved cylinder after the first preprocessing. This effectively overcomes the problems of topological information loss and mean collapse caused by traditional pooling, and realizes efficient and high-precision direct regression of disordered point cloud to the geometric parameters of the curved cylinder. This greatly improves the robustness of inverse reconstruction, overcomes the information loss caused by single pooling, completely avoids the mean collapse problem of the predicted control points gathering towards the centroid, greatly reduces the training difficulty of the network, and improves the generalization ability to different scanning resolutions, noise and workpiece specifications in industrial sites.

[0032] Optionally, as an embodiment of the present invention, such as Figure 3 As shown, the dynamic graph convolutional neural network encoder includes a two-dimensional convolutional layer, a batch normalization layer, a LeakyReLU activation function layer, a first edge convolutional layer, a second edge convolutional layer, a third edge convolutional layer, a fourth edge convolutional layer, a one-dimensional convolutional layer, a max pooling layer, and an average pooling layer. The process of performing feature analysis on the first preprocessed curved cylindrical 3D point cloud matrix using the dynamic graph convolutional neural network encoder to obtain the global feature descriptor includes: The Euclidean distance algorithm is used to retrieve the coordinates of each of the three-dimensional point cloud of the first preprocessed curved cylinder in the first preprocessed curved cylinder three-dimensional point cloud matrix, and obtain the coordinates of multiple curved cylinder neighbor nodes corresponding to each of the three-dimensional point cloud coordinates of the first preprocessed curved cylinder. The fourth equation is used to calculate the 3D point cloud coordinates of each of the first preprocessed curved cylinders and the coordinates of multiple neighboring nodes of the curved cylinders corresponding to the 3D point cloud coordinates of each of the first preprocessed curved cylinders, thereby obtaining multiple original edge features corresponding to the 3D point cloud coordinates of each of the first preprocessed curved cylinders. The fourth equation is: , in, For the first The first preprocessed three-dimensional point cloud coordinates of the curved cylinder correspond to the first Original edge features, For vector concatenation, For the first The coordinates of the neighboring nodes of the curved cylinder. For the first The first preprocessed three-dimensional point cloud coordinates of the curved cylinder; The original edge features are convolved by the two-dimensional convolutional layer to obtain multiple convolved edge features corresponding to the three-dimensional point cloud coordinates of each of the first preprocessed curved cylinders. The fifth equation, the batch normalization layer, and the LeakyReLU activation function layer are used to calculate the multiple convolutional features corresponding to the 3D point cloud coordinates of each of the first preprocessed curved cylinders, respectively, to obtain the original local features corresponding to the 3D point cloud coordinates of each of the first preprocessed curved cylinders. The fifth equation is: , in, For the first The original local features corresponding to the 3D point cloud coordinates of the first preprocessed curved cylinder. For along the first A maximum pooling function for the coordinate dimensions of neighboring nodes of a curved cylinder. For the LeakyReLU activation function layer, For batch normalization layer, For the first The first preprocessed three-dimensional point cloud coordinates of the curved cylinder correspond to the first Features following each convolution; The first edge convolutional layer performs convolution processing on all the original local features, and constructs the first convolutional local feature matrix based on the convolution processing result. The second edge convolutional layer performs convolution processing on the local feature matrix after the first convolution to obtain the local feature matrix after the second convolution. The third edge convolutional layer performs convolution processing on the local feature matrix after the second convolution to obtain the local feature matrix after the third convolution. The fourth edge convolutional layer performs convolution processing on the local feature matrix after the third convolution to obtain the local feature matrix after the fourth convolution. The local feature matrices after the first convolution, the second convolution, the third convolution, and the fourth convolution are fused to obtain the original fused feature matrix. The original fused feature matrix is ​​mapped by the one-dimensional convolutional layer to obtain the mapped fused feature matrix. The mapped and fused feature matrix is ​​subjected to max pooling through the max pooling layer to obtain the first pooled and fused feature matrix. The average pooling layer is used to perform average pooling on the mapped fused feature matrix to obtain a second pooled fused feature matrix. The first pooled fused feature matrix and the second pooled fused feature matrix are concatenated to obtain the global feature descriptor.

[0033] Understandably, the local neighbor nodes (i.e., the coordinates of the curved cylindrical neighbor nodes) of each point are dynamically calculated in the feature space to construct a local adjacency graph; using multi-layer edge convolution operators, the relative difference and absolute features of the center point features (i.e., the first preprocessed curved cylindrical 3D point cloud matrix) and the neighbor features (i.e., the coordinates of the curved cylindrical neighbor nodes) are concatenated to extract local geometric shape information; the edge convolution output features (i.e., the original fused feature matrix) of different depths are concatenated at multiple scales and dimensionality reduced to a unified dimension; max pooling and average pooling are performed simultaneously along the point dimension, and the two pooling results (i.e., the fused feature matrix after the first pooling and the fused feature matrix after the second pooling) are concatenated laterally along the channel dimension to generate the global feature descriptor.

[0034] Understandably, the normalized point cloud data (i.e., the first preprocessed curved cylindrical 3D point cloud matrix) is input into the pre-constructed dynamic graph convolutional neural network encoder (i.e., the dynamic graph convolutional neural network encoder). A K-nearest neighbor local adjacency graph is dynamically constructed in the feature space. The local geometric features of the point cloud are extracted layer by layer through multi-layer edge convolution. The edge convolution output features at different depths (i.e., the local feature matrix after the first convolution, the local feature matrix after the second convolution, the local feature matrix after the third convolution, and the local feature matrix after the fourth convolution) are multi-scale spliced ​​and one-dimensional convolution dimensionality reduction mapping is performed. Finally, dual pooling aggregation of max pooling and average pooling is performed simultaneously along the point dimension to generate a global feature descriptor.

[0035] Specifically, at each layer of the feature space, Euclidean distance is used to retrieve several nearest neighbor nodes (i.e., coordinates of curved cylindrical neighbor nodes) for each point, constructing a dynamic local adjacency graph in the current feature space. Using edge convolution operators, the relative difference vector between the center point feature (i.e., the 3D point cloud matrix of the curved cylinder after the first preprocessing) and its neighbor node features (i.e., coordinates of curved cylindrical neighbor nodes), along with the center point's own features, is concatenated to construct a 3D edge feature tensor. Then, through 2D convolution (i.e., 2D convolutional layer), batch normalization (i.e., batch normalization layer), and a nonlinear activation function (i.e., LeakyReLU activation function layer), the maximum value is taken along the neighbor dimension to complete the aggregation of local neighborhood morphological features. A cascaded structure containing multiple layers of edge convolution operators is constructed. The feature maps output by multi-layer edge convolution (i.e., the local feature matrices after the first, second, third, and fourth convolutions) are fully concatenated along the channel dimension to preserve multi-scale receptive field information. The concatenated high-dimensional features (i.e., the original fused feature matrix) are then projected onto the preset embedding feature dimension through a one-dimensional convolutional layer. Along the number dimension of the point cloud, max pooling and average pooling operations are performed on the embedding feature matrix (i.e., the fused feature matrix after mapping). The max pooling feature vector (i.e., the fused feature matrix after the first pooling) and the average pooling feature vector (i.e., the fused feature matrix after the second pooling) are then horizontally concatenated to generate a global feature descriptor.

[0036] Specifically, the normalized point cloud data output (i.e., the three-dimensional point cloud matrix of the curved cylinder after the first preprocessing) is input into a pre-constructed dynamic graph convolutional neural network encoder (i.e., dynamic graph convolutional neural network encoder) for feature extraction. This invention abandons the global single pooling defect of traditional feature extraction networks and instead extracts spatial geometric shape information by dynamically constructing local connection graphs.

[0037] First, this invention performs block division in the feature space. To avoid GPU memory overflow caused by calculating massive distance matrices all at once when dealing with large batches of point cloud inputs, this invention employs a feature block mechanism. In each batch of calculations, the feature block with the smallest absolute Euclidean distance is selected. With a number of neighboring nodes, the present invention preferably sets the number of nearest neighbors. The nearest neighbor search computation ultimately outputs a high-dimensional neighbor index tensor. Next, the network uses an edge convolution operator to extract local high-frequency features based on this index tensor, for any given center point feature vector. (i.e., the 3D point cloud matrix of the curved cylinder after the first preprocessing), extract based on its corresponding neighbor index. Feature vectors of neighboring nodes (i.e., the coordinates of neighboring nodes of the curved cylinder). Subtract the center point feature (i.e., the 3D point cloud matrix of the curved cylinder after the first preprocessing) from the features of neighboring nodes (i.e., the coordinates of neighboring nodes of the curved cylinder) to obtain the relative geometric difference vector. The relative difference vector is compared with the absolute eigenvector of the center point itself. The edges are concatenated along the channel dimension. The concatenated structure forms a 3D edge feature tensor containing both local relative positional variations and absolute spatial positional information. This edge feature tensor is then sequentially fed into a kernel with a kernel size of [missing information]. The network employs two-dimensional convolutional layers and channel batch normalization layers. The feature flow is then non-linearly activated by a leaky linear rectified function with a negative slope of 0.2, and max pooling is performed along the neighbor number dimension to generate local topological convergence features. A deep feature cascade structure is constructed, consisting of four stacked edge convolution operators with output local channel numbers set to 64, 64, 128, and 256 respectively. The network concatenates the feature maps output by these four layers (i.e., the local feature matrices after the first, second, third, and fourth convolutions) along the channel dimension to obtain a 512-dimensional multi-scale local dense feature matrix (i.e., the original fused feature matrix) containing different receptive field scales. Subsequently, by using one-dimensional convolution combined with batch normalization and a linear rectified function with leakage, the spliced ​​feature projection is reduced to a unified embedded feature space of 1024 dimensions. After the dimensionality reduction mapping is completed, dual-pooling global feature aggregation is performed along the dimension of the number of key points in the point cloud. The network simultaneously calculates the global maximum value of the 1024-dimensional embedded feature to extract local geometric features, and calculates the global average value. The algorithm concatenates the feature vector of the maximum value and the feature vector of the average value laterally, and finally outputs a global feature descriptor with a fixed length of 2048. This dual-pooling mechanism ensures to the greatest extent that the high-frequency topological information representing the high curvature region and boundary endpoints of the curved cylinder is not flattened by mean normalization.

[0038] In the above embodiments, the three-dimensional point cloud matrix of the first preprocessed curved cylinder is analyzed by a dynamic graph convolutional neural network encoder to obtain a global feature descriptor. This eliminates the global single pooling defect of traditional feature extraction networks, extracts information about spatial geometric shape, and ensures to the greatest extent that the high-frequency topological information representing the high curvature region and boundary endpoints of the curved cylinder is not flattened by mean-averaging.

[0039] Optionally, as an embodiment of the present invention, the dual-head multilayer perceptron network includes a first MLP multilayer perceptron subnetwork and a second MLP multilayer perceptron subnetwork. The process of decoupling and analyzing the global feature descriptor through the dual-head multilayer perceptron network to obtain the second predicted cross-sectional radius and the second predicted control point matrix includes: The central axis features are obtained by performing central axis regression on the global feature descriptor through the first MLP multilayer perceptron sub-network. The central axis features are reshaped to obtain the second predictive control point matrix; The second predicted cross-sectional radius is obtained by performing radius regression processing on the global feature descriptor through the second MLP multilayer perceptron subnetwork.

[0040] It should be understood that the central axis of the curved cylinder is parametrically represented by a cubic B-spline curve; the central axis regression branch (i.e., the first MLP multilayer perceptron sub-network) outputs the coordinates of four ordered control points (i.e., the second predictive control point matrix) used to determine the topological trend of the cubic B-spline curve; the radius regression branch (i.e., the second MLP multilayer perceptron sub-network) independently outputs a single value as the cross-sectional radius of the curved cylinder (i.e., the second predictive cross-sectional radius).

[0041] It should be understood that the global feature descriptor is fed into two parallel fully connected regression branches. The central axis regression branch (i.e., the first MLP multilayer perceptron subnetwork) outputs the coordinates of four ordered control points (i.e., the second predicted control point matrix) that determine the parameterization direction of the cubic B-spline curve of the central axis. The radius regression branch (i.e., the second MLP multilayer perceptron subnetwork) outputs the value as the normalized cross-sectional radius (i.e., the second predicted cross-sectional radius).

[0042] Specifically, the central axis of the curved cylinder is parametrically represented by a cubic B-spline curve. Since determining a cubic B-spline curve with clear tangential and curvature continuity requires at least four ordered control points, the final linear output of the central axis regression branch (i.e., the first MLP multilayer perceptron sub-network) is reshaped to obtain a coordinate matrix of four curve control points with a strict spatial topological order (i.e., the second predicted control point matrix). The final linear output of the radius regression branch (i.e., the second MLP multilayer perceptron sub-network) is a one-dimensional parameter representing the scalar value of the cross-sectional radius (i.e., the second predicted cross-sectional radius).

[0043] Specifically, the global feature descriptor with a length of 2048 after dimensionality reduction is fed in parallel into the decoupled central axis regression branch (i.e., the first MLP multilayer perceptron subnetwork) and radius regression branch (i.e., the second MLP multilayer perceptron subnetwork). Both fully connected regression branches (i.e., the first MLP multilayer perceptron subnetwork and the second MLP multilayer perceptron subnetwork) adopt a multilayer perceptron structure and contain several hidden layers connected in sequence. The node dimensions are preferably 512 and 256. In order to prevent overfitting, a one-dimensional batch normalization operation, a non-linear activation function, and a Dropout random discarding mechanism are added between each hidden layer.

[0044] The final fully connected layer of the first central axis regression branch (i.e., the first MLP multilayer perceptron subnetwork) precisely maps high-dimensional features (i.e., global feature descriptors) into a dimensionless 12-dimensional continuous vector (i.e., central axis features). This 12-dimensional vector outputs a normalized coordinate system curve control point matrix through a data reshaping operation. (i.e., the second predictive control point matrix), this control point matrix The four coordinates with a strict spatial sequence are uniquely represented by the four cubic B-spline fitting models required to determine the central axis. Since, mathematically, determining a cubic B-spline curve with clear tangential and curvature continuity requires at least four ordered control points, this invention preferably sets the network joint solution output to four control points to minimize the number of network parameters while ensuring curve fitting accuracy.

[0045] The final fully connected layer of the second radius regression branch (i.e., the second MLP multilayer perceptron sub-network) independently maps homogeneous features (i.e., global feature descriptors) to a single one-dimensional scalar value, which is the fitting cross-sectional radius in the normalized feature space. (i.e., the radius of the second predicted section).

[0046] In the above embodiments, a dual-head multilayer perceptron network is used to decouple and analyze the global feature descriptors to obtain the second predicted cross-sectional radius and the second predicted control point matrix, which prevents overfitting and ensures the accuracy of curve fitting while minimizing the number of network parameters.

[0047] Optionally, as an embodiment of the present invention, the process of performing loss function analysis on the second predicted cross-section radius, the second predicted control point matrix, the true control point matrix, and the true cross-section radius to obtain the target loss function includes: The target loss function is obtained by calculating the second predicted cross-sectional radius, the second predicted control point matrix, the true control point matrix, and the true cross-sectional radius using the sixth equation. The sixth equation is: , in, , , in, Let be the target loss function. and All are loss weighting coefficients. The order-preserving MSE loss function for control points. The radius-smoothing L1 loss function is used. The second predicted control point matrix is ​​the first... The second predictive control point Each dimension coordinate, The first in the true control point matrix The first real control point Each dimension coordinate, The radius of the second predicted section. This is the actual cross-sectional radius.

[0048] It should be understood that during the model training phase, the dynamic graph convolutional neural network encoder and the dual-head multilayer perceptron network are jointly optimized end-to-end through a set composite loss function. The composite loss function (i.e., the target loss function) is composed of a weighted sum of control point loss and radius loss. For the centerline control points with strict topological order, the control point loss is calculated using the mean square error strictly aligned to the index (i.e., the control point order-preserving MSE loss function). For the cross-sectional radius, the radius loss is calculated using smoothed L1 loss or L1 distance (i.e., the radius smoothed L1 loss function).

[0049] It should be understood that during the model training phase, the network is jointly trained end-to-end using a composite loss function consisting of a mean squared error control point loss strictly aligned to index and a smoothed L1 radius loss.

[0050] Specifically, the control point loss (i.e., the control point order-preserving MSE loss function) is calculated using the mean squared error strictly aligned to the index to avoid the disruption of the curve topology caused by disordered distance metrics; the radius loss (i.e., the radius smoothing L1 loss function) is calculated using either smoothing L1 loss or L1 loss; the control point loss (i.e., the control point order-preserving MSE loss function) and the radius loss (i.e., the radius smoothing L1 loss function) are weighted and summed to obtain the total composite loss (i.e., the target loss function), and the network is trained by backpropagation using the total composite loss (i.e., the target loss function) to update the network parameters.

[0051] Specifically, during the end-to-end forward inference and backward training phases of the model, to ensure the fidelity of the topological orientation of the parameterized generated curves, the use of unordered distance metrics that would disrupt the Euclidean order is strictly prohibited, as is the loss of the central axis control points. (i.e., the control point order-preserving MSE loss function) strictly uses a one-to-one mean square error constraint calculated with absolute alignment of matrix indices, thereby strongly penalizing disordered network output trends and radius loss. The radius-smoothing L1 loss function is calculated using the smoothed L1 loss algorithm to effectively enhance the network's robustness against outliers and Gaussian noise points on the scanned surface. The training process employs an adaptive moment estimation optimizer, with an initial learning rate set at... The momentum weight decay coefficient is It is equipped with a cosine annealing periodic learning rate scheduler, and the composite objective function (i.e., the objective loss function) for joint training is defined as: In this invention, the control point weight coefficient With radius weighting coefficient (i.e., loss weight coefficients) are all equivalently set to 1.0 by default. The neural network engine uses this composite objective function. The differential gradient (i.e., the target loss function) is automatically calculated and backpropagated to update the weight parameters of all graph convolutional and densely connected layers.

[0052] In the above embodiments, loss function analysis is performed on the second predicted cross-section radius, the second predicted control point matrix, the real control point matrix, and the real cross-section radius to obtain the target loss function. This greatly improves the robustness of inverse reconstruction, overcomes the information loss caused by single pooling, completely avoids the mean collapse problem of predicted control points clustering towards the centroid, greatly reduces the training difficulty of the network, and improves the generalization ability to different scanning resolutions, noise levels, and workpiece specifications in industrial settings.

[0053] Optionally, as an embodiment of the present invention, such as Figure 4 As shown, the process of reconstructing and analyzing the first predicted cross-sectional radius, the scale parameters to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, the centroid coordinates to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and obtaining the point cloud reconstruction result of the curved cylinder structure based on the reconstruction analysis results includes: The central axis control point matrix is ​​obtained by calculating the scale parameters to be reconstructed, the centroid coordinates to be reconstructed corresponding to the 3D point cloud data of the cylindrical body to be reconstructed, and the first predicted control point matrix corresponding to the 3D point cloud data of the cylindrical body to be reconstructed using the seventh equation. The seventh equation is as follows: , in, This is the matrix of control points for the central axis. This is the matrix of the first predicted control points. The scale parameters to be reconstructed are... The coordinates of the centroid to be reconstructed; The target cross-sectional radius is obtained by calculating the scale parameters to be reconstructed and the first predicted cross-sectional radius corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed using the eighth equation. The eighth equation is: , in, The radius of the target cross section, The radius of the first predicted section. The scale parameters to be reconstructed; The geometric parameters of the curved cylinder are constructed using the central axis control point matrix and the target cross-sectional radius, and the geometric parameters of the curved cylinder are used as the point cloud reconstruction result of the curved cylinder structure.

[0054] It should be understood that the coordinates of the central axis control points (i.e., the first predicted control point matrix) in the normalized coordinate system output by the network are multiplied by the scale parameter (i.e. the scale parameter to be reconstructed), and the centroid coordinates (i.e. the centroid coordinates to be reconstructed) are added to restore the coordinates of the central axis control points (i.e., the central axis control point matrix) in the real-world coordinate system; the normalized cross-sectional radius (i.e., the first predicted cross-sectional radius) output by the network is multiplied by the scale parameter (i.e. the scale parameter to be reconstructed) to restore the actual cross-sectional radius size (i.e., the target cross-sectional radius) in the real-world coordinate system.

[0055] Specifically, using the recorded centroid coordinates (i.e., the centroid coordinates to be reconstructed) and scale parameters (i.e. the scale parameters to be reconstructed), the ordered control point coordinates (i.e., the first predicted control point matrix) and normalized cross-sectional radius (i.e., the first predicted cross-sectional radius) output by the network are inversely normalized to obtain the high-precision geometric parameters of the curved cylinder (i.e., the central axis control point matrix and the target cross-sectional radius) in the real-world coordinate system.

[0056] It should be understood that the normalized control point coordinates output by the network (i.e., the first predicted control point matrix) are multiplied by the scale parameter (i.e. the scale parameter to be reconstructed), and the centroid coordinates (i.e. the centroid coordinates to be reconstructed) are added to restore the absolute three-dimensional coordinates of the central axis control points in the world coordinate system (i.e., the central axis control point matrix). The normalized cross-sectional radius output by the network (i.e., the first predicted cross-sectional radius) is multiplied by the scale parameter (i.e. the scale parameter to be reconstructed) to restore the actual cross-sectional radius size in the world coordinate system (i.e., the target cross-sectional radius).

[0057] Specifically, let the matrix of control points of the central axis in the restored absolute world coordinate system be... (i.e., the matrix of control points along the central axis), its inverse affine transformation calculation formula is: Let the radius of the solid cross section in the restored absolute world coordinate system be... (i.e., the radius of the target cross section) is Its algebraic calculation formula is: Through this mathematical mapping loop, the automatic direct conversion from disordered discrete scan point clouds to high-precision real-world geometric parameters is achieved.

[0058] In the above embodiments, the first predicted cross-sectional radius, the scale parameters to be reconstructed, the centroid coordinates to be reconstructed, and the first predicted control point matrix are reconstructed and analyzed. Based on the reconstruction analysis results, the point cloud reconstruction results of the curved cylindrical structure are obtained. This completely avoids the mean collapse problem of the predicted control points gathering towards the centroid, greatly reduces the training difficulty of the network, and improves the generalization ability to different scanning resolutions, noise levels, and workpiece specifications in industrial sites.

[0059] Optionally, as another embodiment of the present invention, this invention belongs to the field of reverse engineering and 3D point cloud processing technology, specifically relating to a high-precision reconstruction method for curved cylindrical point clouds based on a dynamic graph convolutional network. This method first acquires the original point cloud of the curved cylinder, performs fixed-point sampling, decentralization, and scale normalization preprocessing based on maximum distance; then, it constructs a dynamic graph convolutional encoder containing multi-layer edge convolution (EdgeConv) operators to dynamically extract local geometric features of the point cloud, and performs multi-scale feature stitching and global double pooling aggregation; next, it inputs the aggregated features into a dual-head multilayer perceptron network, jointly solves and outputs the central axis control points and radius parameters of the curved cylinder; finally, it uses a strictly order-preserving mean square error and smoothed L1 combined loss function to perform end-to-end optimization training on the network, and performs reverse reconstruction in the world coordinate system. This invention effectively overcomes the problems of topological information loss and mean collapse caused by traditional pooling, achieving efficient and high-precision direct regression of disordered point clouds to the geometric parameters of curved cylinders, greatly improving the robustness of reverse reconstruction.

[0060] Optionally, as another embodiment of the present invention, the present invention aims to solve the technical problems of high sensitivity of staged geometric fitting, spatial topological information annihilation and "mean collapse" caused by traditional single pooling mechanism, and the destruction of the order of control points by existing distance metric algorithms, and to achieve efficient, robust and order-preserving end-to-end direct regression of disordered 3D point cloud to control points and radius parameters of the central axis of curved cylinder.

[0061] Alternatively, as another embodiment of the present invention, compared with the prior art, the present invention has the following significant advantages: This invention completely overcomes the pain points of "mean collapse" and topology loss: It abandons the global max-pooling architecture of traditional PointNet and adopts the edge convolution operator of a dynamic graph convolutional network. By dynamically constructing a local adjacency graph in the feature space, it explicitly models the relative spatial topological relationships between points. Especially in the specific scenario of curved cylinder reconstruction, max pooling can extract high-frequency topological features with drastic curvature changes (such as pipe endpoints and curved boundaries), while average pooling can preserve the low-frequency spatial distribution trend of the pipe as a whole. The combination of the two produces unexpected technical effects, overcoming the information loss caused by single pooling. Under the optimization pressure of mean squared error loss, it completely avoids the "mean collapse" problem of predicted control points clustering towards the centroid.

[0062] Strict topological orientation-preserving regression is achieved: This invention addresses the strict sequential requirement of parameterizing the cubic B-spline curve of the central axis, determines the optimal architecture of the network regression output with 4 control points, and innovatively introduces the mean square error of strictly one-to-one alignment of the index as the objective function of the control points, without using unordered distance metrics, thereby strongly constraining the network's ability to identify the tangent direction and extension direction of the curve axis, and ensuring the topological fidelity of the end-to-end parameter regression.

[0063] Exhibiting excellent size and density adaptive robustness: This invention incorporates an adaptive data preprocessing module in the initial stage. This module standardizes scanned point clouds of arbitrary density through downsampling and eliminates the influence of initial pose and absolute size using dynamically captured centroid and maximum distance parameters. This allows the network to focus solely on learning morphological features within a unified dimensionless normalized space, significantly reducing the training difficulty and improving the algorithm's generalization ability to different scanning resolutions, noise levels, and workpiece specifications in industrial environments.

[0064] High-efficiency direct output, facilitating industrial closed-loop integration: The end-to-end framework of this invention is completely different from the traditional step-by-step approach of "component detection + model fitting". It adopts a parallel dual-head regression architecture, which simplifies the calculation process. The inference process only needs to run a single forward propagation and a simple algebraic backpropagation to directly output the centerline control points and radii readable by standard CAD systems, which greatly facilitates the integration of fully automatic closed-loop measurement, grasping or pipeline processing processes.

[0065] Alternatively, as another embodiment of the present invention, such as Figure 5 As shown, the present invention includes the following steps: Step 1 (S110, S120): Adaptive data preprocessing: Obtain the original 3D point cloud data of the curved cylinder to be reconstructed, perform fixed-point downsampling on it, and sequentially perform centroid alignment decentralization and scale normalization to generate normalized point cloud data. At the same time, record the centroid coordinates subtracted during decentralization and the maximum distance parameter divided during scale normalization, i.e., the scale parameter. Step 2 (S130): Dynamic graph convolution local and global multi-scale feature extraction: The normalized point cloud data is input into the pre-constructed dynamic graph convolutional neural network encoder. A K-nearest neighbor local adjacency graph is dynamically constructed in the feature space. The local geometric features of the point cloud are extracted layer by layer through multi-layer edge convolution. The edge convolution output features of different depths are multi-scale spliced ​​and one-dimensional convolution dimensionality reduction mapping is performed. Finally, dual pooling aggregation of max pooling and average pooling is performed simultaneously along the point dimension to generate a global feature descriptor. Step 3 (S141, S142), Parallel regression with dual-headed parameters: The global feature descriptor is fed into two parallel fully connected regression branches respectively. The midline regression branch outputs the coordinates of four ordered control points that determine the parameterization direction of the cubic B-spline curve of the midline, and the radius regression branch outputs the value as the normalized cross-sectional radius. Step 4 (S150, S160), Composite Loss Optimization and Spatial Pose Inverse Reconstruction: In the model training phase (S150), the network is jointly trained end-to-end using a composite loss function consisting of a mean square error control point loss with strict index alignment and a smoothed L1 radius loss; in the model inference phase (S160), the centroid coordinates and scale parameters recorded in Step 1 are used to perform inverse normalization calculations on the ordered control point coordinates and normalized cross-sectional radii output by the network, thereby reconstructing the high-precision geometric parameters of the curved cylinder in the real-world coordinate system.

[0066] Alternatively, as another embodiment of the present invention, such as Figure 6 As shown, this invention includes an adaptive data preprocessing module, which internally comprises a sampling calculation unit, a translation and alignment unit, and a geometric scaling unit. The sampling calculation unit is responsible for reading data acquired by the underlying 3D scanning device. The unit processes raw point cloud data to a standard scale and performs a task to intercept point sparsity anomalies. When an anomaly is triggered, the unit resamples the raw data to a standard scale based on an internal algorithm. The translation alignment unit and the geometry scaling unit are sequentially applied to units with a scale of... The point cloud is subjected to centroid decentering and maximum distance division to output a shape that conforms to the standard-scale distribution law. The normalized tensor. The adaptive data preprocessing module simultaneously injects and stores the geometric centroid vector in the global memory pool. and the radius of the scale Once the parameters are restored to the parameter cache, the system's dynamic graph convolutional feature extraction module deeply embeds a spatial nearest neighbor dynamic graph construction operator and a multi-layer cascaded feature edge convolution operator. This feature extraction module utilizes the network convolution kernel weight sharing mechanism to process massive point cloud data and dynamically search for each node. This module combines the relative geometric difference vectors to identify the nearest neighbors in the feature space. with absolute eigenvectors This module achieves sensitive capture of local high-frequency boundary information of curved cylinders. Simultaneously, it closely integrates multi-scale feature channel stitching with a combined max and average double pooling algorithm. Through the aforementioned double pooling data flow design, the module ultimately generates and outputs a global feature descriptor of length 2048, achieving lossless encoding of the target's global low-frequency distribution pattern. The system's dual-headed parallel parameter regression module contains two perceptron regression sub-branches with absolutely decoupled parameters at the backpropagation stage. The first sub-branch, the central axis regression branch, is dedicated to receiving the upstream global feature descriptor to reconstruct the spatial backbone topology of the curve. This branch ultimately outputs a normalized control point matrix with a strict sequential geometric sequence through a fully connected mapping. The second sub-branch, the radius regression branch, is dedicated to calculating the radial dimension properties of the corresponding curved cylinder cross-section based on the aforementioned common global feature descriptors. This branch independently regresses and outputs a one-dimensional normalized radius scalar. The system's model optimization and spatial reconstruction module has dual functions: automatic monitoring of the underlying network training and algebraic calculation of the absolute pose of the upper-layer model coordinates. During the offline training phase of the neural network, this module calls a custom-written composite loss function to accurately calculate the order-preserving mean square error value. With robust smoothing of L1 loss The module then applies the composite objective function. The module calculates gradients and drives the deep learning framework to perform backward parameter updates. During online deployment and forward inference on the edge, this module calls the matrix tensor operation unit to perform inverse restoration of spatial affine transformations. The operation unit uses the previously cached centroid coordinates. With scale parameters It directly outputs a matrix of control points for the central axis of the entity with absolute physical dimensions to the business system. With the radius of the solid cross section .

[0067] Alternatively, as another embodiment of the present invention, such as Figure 7 and 8 As shown, to rigorously verify the physical effectiveness of this invention in completely solving the persistent problem of "mean collapse" in parametric point cloud regression tasks in a real industrial environment, this invention conducted a detailed algorithm performance ablation experiment. The experimental platform was loaded with a large-scale digital twin dataset containing 10,000 complex curved cylindrical point cloud samples, which was constructed independently. The specific construction method of this dataset is as follows: First, a cubic B-spline curved cylindrical model with principal curvature radii distributed in a specific range and cross-sectional radii conforming to reasonable manufacturing tolerance ranges is randomly generated in 3D simulation software; then, point clouds are generated by uniformly discrete sampling on the model surface, and a standard deviation of is manually added. Gaussian noise was used to simulate the measurement physical errors of a real laser scanner; finally, the 10,000 samples were rigorously divided into training, validation, and test sets in an 8:1:1 ratio, fully ensuring the diversity and coverage of the model training sample library in three-dimensional space. All pipe point cloud samples selected for testing were processed by the rigorous adaptive resampling engine proposed in this invention and uniformly standardized. With the input size, the network underwent a total of 100 epochs of end-to-end joint training using the innovative dual-branch feature network architecture described above. Detailed experimental record forms and corresponding documentation are available. Figure 7 The convergence trend curve clearly demonstrates the excellent optimization properties of this model. Under the smooth scheduling of the cosine annealing learning rate update decay strategy, the training set loss curve during network training and the validation set loss curve during independent validation exhibit highly synchronized and stable decreasing curves. Furthermore, throughout the entire deep iteration cycle, there was no oscillation or divergence in the loss value caused by sample overfitting or gradient explosion. When the network optimization cycle enters the final stage, the oscillation standard deviation of the feature loss drastically decreases to less than [value missing]. The objective data, demonstrating rigorous mathematical logic, proves that the core regression model has entered a state of extremely stable parameter convergence and minimum value, with the optimal validation loss reaching a minimum of 0.001399. Regarding the specific evaluation index of geometric reconstruction accuracy, the root mean square prediction error (RMSE) of the pose of the control points on the central axis curve mapped to the normalized feature space dimension is as low as 0.0372. This accuracy objectively proves the spatial geometric positioning capability of this invention. Compared to conventional baseline reference networks, which are prone to mean collapse, this invention, relying on an innovative dynamic graph network and order-preserving error combination architecture, achieves a significant qualitative improvement in fitting accuracy at spatial endpoints and in complex curved regions. Furthermore, the cross-sectional radius index output by the decoupled branch shows a monotonically decreasing and convergent smooth L1 distance loss to an extremely low order of magnitude of 0.000017. This result strongly verifies that the parallel dual-branch scheme effectively avoids the extremely high parameter fitting accuracy caused by the mutual interference between spatial topological features and dimensional scalar features. Figure 8 The qualitative analysis of the final multi-dimensional three-dimensional visualization results of the reverse surface reconstruction demonstrates the generalization robustness of this invention. Figure 8 (a) shows the actual point cloud and serves as a reference for the ground truth value; when faced with the same data... Figure 8 (b) The traditional network prediction results suffer from severe mean collapse, with control points incorrectly squeezed into the central region; while for any unconventional pose or extreme curved cylinder with extreme irregularity and multiple continuous bending within the sample library, Figure 8(c) The four-point array generated by the network prediction of this invention perfectly fits the physical target. The network reconstructs the smooth trend of the central axis curvature, representing the real physical form, with high fidelity. In particular, the endpoints of the two poles of the central axis curve in the generated surface are positioned with extreme precision in the three-dimensional Euclidean coordinate system without any visible offset. This visual evidence conclusively proves that the dynamic graph network architecture constructed by this invention successfully overcomes the theoretical dead end of traditional single global max pooling structures, which are prone to feature dimensionality reduction and mean flattening. In summary, the data and image comparison confirm that this invention, while dealing with point clouds containing outliers and noise, exhibits unparalleled high fidelity reproduction and strong noise resistance in maintaining the topological manifold direction of the target spatial structure. This invention and system have significant and intuitive value for industrial automation production line-level engineering applications and replacement of existing algorithms.

[0068] Figure 9 This is a block diagram of a point cloud reconstruction device for a curved cylindrical structure provided in an embodiment of the present invention.

[0069] Alternatively, as another embodiment of the present invention, such as Figure 9 As shown, a point cloud reconstruction device with a curved cylindrical structure includes: The point cloud data acquisition module is used to obtain the original 3D point cloud data of the curved cylinder and the 3D point cloud data of the curved cylinder to be reconstructed from the 3D scanning device. The preprocessing module is used to preprocess the original curved cylindrical 3D point cloud data and the curved cylindrical 3D point cloud data to be reconstructed to obtain a first preprocessed curved cylindrical 3D point cloud matrix corresponding to the original curved cylindrical 3D point cloud data, a second preprocessed curved cylindrical 3D point cloud matrix corresponding to the curved cylindrical 3D point cloud data to be reconstructed, a scale parameter to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed, and centroid coordinates to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed. The model analysis module is used to construct a training model and perform model analysis on the training model using the first preprocessed curved cylindrical 3D point cloud matrix to obtain a cross-section prediction model. The prediction module is used to predict the second preprocessed curved cylinder three-dimensional point cloud matrix through the cross-section prediction model to obtain the first predicted cross-section radius corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed. The reconstruction result module is used to perform reconstruction analysis on the first predicted cross-sectional radius, the scale parameters to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, the centroid coordinates to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and to obtain the point cloud reconstruction result of the curved cylinder structure based on the reconstruction analysis result.

[0070] Optionally, another embodiment of the present invention provides a point cloud reconstruction system for a curved cylindrical structure, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the point cloud reconstruction method for the curved cylindrical structure as described above. This system can be a computer or similar system.

[0071] Optionally, another embodiment of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the point cloud reconstruction method for curved cylindrical structures as described above.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0073] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described apparatus and unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0074] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0075] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.

[0076] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0077] If the integrated unit is implemented as 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, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for reconstructing point clouds of curved cylindrical structures, characterized in that, Includes the following steps: Obtain the original 3D point cloud data of the curved cylinder and the 3D point cloud data of the curved cylinder to be reconstructed from the 3D scanning equipment; The original three-dimensional point cloud data of the curved cylinder and the three-dimensional point cloud data of the curved cylinder to be reconstructed are preprocessed to obtain a first preprocessed three-dimensional point cloud matrix of the curved cylinder corresponding to the original three-dimensional point cloud data, a second preprocessed three-dimensional point cloud matrix of the curved cylinder corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, a scale parameter to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and centroid coordinates to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed. A training model is constructed, and the training model is analyzed using the first preprocessed three-dimensional point cloud matrix of the curved cylinder to obtain a cross-section prediction model. The cross-section prediction model is used to predict the second preprocessed three-dimensional point cloud matrix of the curved cylinder to obtain the first predicted cross-section radius corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed. The first predicted cross-sectional radius, the scale parameters to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, the centroid coordinates to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed are reconstructed and analyzed. The reconstruction result of the curved cylinder structure point cloud is obtained based on the reconstruction analysis result.

2. The method for reconstructing point clouds of curved cylindrical structures according to claim 1, characterized in that, The original three-dimensional point cloud data of the curved cylinder includes multiple discrete points of the original three-dimensional point cloud of the curved cylinder, and the three-dimensional point cloud data of the curved cylinder to be reconstructed includes multiple discrete points of the three-dimensional point cloud of the curved cylinder to be reconstructed. The process of preprocessing the original curved cylindrical 3D point cloud data and the curved cylindrical 3D point cloud data to be reconstructed to obtain a first preprocessed curved cylindrical 3D point cloud matrix corresponding to the original curved cylindrical 3D point cloud data, a second preprocessed curved cylindrical 3D point cloud matrix corresponding to the curved cylindrical 3D point cloud data to be reconstructed, the scale parameters to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed, and the centroid coordinates to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed includes: According to the preset sampling scale, all the original three-dimensional point cloud discrete points of the curved cylinder are sampled and processed to obtain multiple first-sampled three-dimensional point cloud discrete points of the curved cylinder corresponding to the original three-dimensional point cloud data of the curved cylinder. According to the preset sampling scale, all discrete points of the three-dimensional point cloud of the cylindrical body to be reconstructed are sampled to obtain multiple discrete points of the three-dimensional point cloud of the cylindrical body after second sampling, which correspond to the three-dimensional point cloud data of the cylindrical body to be reconstructed. The first formula is used to calculate the original centroid coordinates corresponding to the original 3D point cloud data of the curved cylinder and the original centroid coordinates corresponding to the original 3D point cloud data of the curved cylinder, as well as the original centroid coordinates corresponding to the original 3D point cloud data of the curved cylinder and the original centroid coordinates corresponding to the original 3D point cloud data of the curved cylinder. The first formula is: ， in, The original centroid coordinates, For the coordinates of the centroid to be reconstructed, , This represents the number of discrete points in the 3D point cloud of the curved cylinder after the first sampling or the number of discrete points in the 3D point cloud of the curved cylinder after the second sampling. The first point cloud data corresponding to the original curved cylinder Discretized points of the three-dimensional point cloud of the curved cylinder after the first sampling. The first point cloud data corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed The second sampled three-dimensional point cloud discrete points of the curved cylinder; The difference between each discrete point of the first sampled cylindrical three-dimensional point cloud and the original centroid coordinates is calculated to obtain multiple first translated cylindrical three-dimensional point cloud coordinates corresponding to the original cylindrical three-dimensional point cloud data. The difference between each discrete point of the second sampled cylindrical three-dimensional point cloud and the centroid coordinates to be reconstructed are calculated to obtain multiple second translated cylindrical three-dimensional point cloud coordinates corresponding to the cylindrical three-dimensional point cloud data to be reconstructed. The second formula is used to calculate the coordinates of the original three-dimensional point cloud data of the curved cylinder corresponding to multiple first-translated three-dimensional point cloud data and the coordinates of the curved cylinder corresponding to multiple second-translated three-dimensional point cloud data to be reconstructed, thereby obtaining the original scale parameters corresponding to the original three-dimensional point cloud data of the curved cylinder and the scale parameters to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed. The second formula is: ， in, These are the original scale parameters. The scale parameters to be reconstructed are... , The first point cloud data corresponding to the original curved cylinder The coordinates of the first translated curved cylinder in three-dimensional point cloud. The first point cloud data corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed The coordinates of the second translated curved cylinder in three-dimensional point cloud. It is the 2-norm of a three-dimensional vector; The third equation is used to calculate the original scale parameters, the scale parameters to be reconstructed corresponding to the 3D point cloud data of the curved cylinder to be reconstructed, the multiple first translated 3D point cloud coordinates of the curved cylinder corresponding to the original 3D point cloud data, and the multiple second translated 3D point cloud coordinates of the curved cylinder corresponding to the 3D point cloud data of the curved cylinder to be reconstructed. Based on the calculation results, a first preprocessed 3D point cloud matrix of the curved cylinder corresponding to the original 3D point cloud data and a second preprocessed 3D point cloud matrix of the curved cylinder corresponding to the 3D point cloud data of the curved cylinder to be reconstructed are constructed. The third equation is: ， in, The first point cloud data corresponding to the original curved cylinder The first preprocessed three-dimensional point cloud coordinates of the curved cylinder The first point cloud data corresponding to the original curved cylinder The second preprocessed three-dimensional point cloud coordinates of the curved cylinder The first point cloud data corresponding to the original curved cylinder The coordinates of the first translated curved cylinder in three-dimensional point cloud. The first point cloud data corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed The coordinates of the second translated curved cylinder in three-dimensional point cloud. These are the original scale parameters. These are the scale parameters to be reconstructed.

3. The method for reconstructing point clouds of curved cylindrical structures according to claim 1, characterized in that, The training model includes a dynamic graph convolutional neural network encoder and a dual-head multilayer perceptron network. The process of performing model analysis on the trained model using the first preprocessed three-dimensional point cloud matrix of the curved cylinder to obtain the cross-section prediction model includes: The dynamic graph convolutional neural network encoder performs feature analysis on the first preprocessed curved cylindrical 3D point cloud matrix to obtain a global feature descriptor. The global feature descriptor is decoupled and analyzed by the dual-head multilayer perceptron network to obtain the second predicted cross-sectional radius and the second predicted control point matrix. Import the actual control point matrix and the actual cross-section radius, and perform loss function analysis on the second predicted cross-section radius, the second predicted control point matrix, the actual control point matrix, and the actual cross-section radius to obtain the target loss function; The parameters of the trained model are updated according to the target loss function to obtain the cross-section prediction model.

4. The method for reconstructing point clouds of curved cylindrical structures according to claim 3, characterized in that, The dynamic graph convolutional neural network encoder includes a two-dimensional convolutional layer, a batch normalization layer, a LeakyReLU activation function layer, a first edge convolutional layer, a second edge convolutional layer, a third edge convolutional layer, a fourth edge convolutional layer, a one-dimensional convolutional layer, a max pooling layer, and an average pooling layer. The process of performing feature analysis on the first preprocessed curved cylindrical 3D point cloud matrix using the dynamic graph convolutional neural network encoder to obtain the global feature descriptor includes: The Euclidean distance algorithm is used to retrieve the coordinates of each of the three-dimensional point cloud of the first preprocessed curved cylinder in the first preprocessed curved cylinder three-dimensional point cloud matrix, and obtain the coordinates of multiple curved cylinder neighbor nodes corresponding to each of the three-dimensional point cloud coordinates of the first preprocessed curved cylinder. The fourth equation is used to calculate the 3D point cloud coordinates of each of the first preprocessed curved cylinders and the coordinates of multiple neighboring nodes of the curved cylinders corresponding to the 3D point cloud coordinates of each of the first preprocessed curved cylinders, thereby obtaining multiple original edge features corresponding to the 3D point cloud coordinates of each of the first preprocessed curved cylinders. The fourth equation is: ， in, For the first The first preprocessed three-dimensional point cloud coordinates of the curved cylinder correspond to the first Original edge features, For vector concatenation, For the first The coordinates of the neighboring nodes of the curved cylinder. For the first The first preprocessed three-dimensional point cloud coordinates of the curved cylinder; The original edge features are convolved by the two-dimensional convolutional layer to obtain multiple convolved edge features corresponding to the three-dimensional point cloud coordinates of each of the first preprocessed curved cylinders. The fifth equation, the batch normalization layer, and the LeakyReLU activation function layer are used to calculate the multiple convolutional features corresponding to the 3D point cloud coordinates of each of the first preprocessed curved cylinders, respectively, to obtain the original local features corresponding to the 3D point cloud coordinates of each of the first preprocessed curved cylinders. The fifth equation is: ， in, For the first The original local features corresponding to the 3D point cloud coordinates of the first preprocessed curved cylinder. For along the first A maximum pooling function for the coordinate dimensions of neighboring nodes of a curved cylinder. For the LeakyReLU activation function layer, For batch normalization layer, For the first The first preprocessed three-dimensional point cloud coordinates of the curved cylinder correspond to the first Features following each convolution; The first edge convolutional layer performs convolution processing on all the original local features, and constructs the first convolutional local feature matrix based on the convolution processing result. The second edge convolutional layer performs convolution processing on the local feature matrix after the first convolution to obtain the local feature matrix after the second convolution. The third edge convolutional layer performs convolution processing on the local feature matrix after the second convolution to obtain the local feature matrix after the third convolution. The fourth edge convolutional layer performs convolution processing on the local feature matrix after the third convolution to obtain the local feature matrix after the fourth convolution. The local feature matrices after the first convolution, the second convolution, the third convolution, and the fourth convolution are fused to obtain the original fused feature matrix. The original fused feature matrix is ​​mapped by the one-dimensional convolutional layer to obtain the mapped fused feature matrix. The mapped fused feature matrix is ​​subjected to max pooling through the max pooling layer to obtain the first pooled fused feature matrix. The average pooling layer is used to perform average pooling on the mapped fused feature matrix to obtain a second pooled fused feature matrix. The first pooled fused feature matrix and the second pooled fused feature matrix are concatenated to obtain the global feature descriptor.

5. The method for reconstructing point clouds of curved cylindrical structures according to claim 3, characterized in that, The dual-head multilayer perceptron network includes a first MLP multilayer perceptron subnetwork and a second MLP multilayer perceptron network. The process of decoupling and analyzing the global feature descriptor through the dual-head multilayer perceptron network to obtain the second predicted cross-sectional radius and the second predicted control point matrix includes: The central axis features are obtained by performing central axis regression on the global feature descriptor through the first MLP multilayer perceptron sub-network. The central axis features are reshaped to obtain the second predictive control point matrix; The second predicted cross-sectional radius is obtained by performing radius regression processing on the global feature descriptor through the second MLP multilayer perceptron subnetwork.

6. The method for reconstructing point clouds of curved cylindrical structures according to claim 3, characterized in that, The process of performing loss function analysis on the second predicted cross-sectional radius, the second predicted control point matrix, the actual control point matrix, and the actual cross-sectional radius to obtain the target loss function includes: The target loss function is obtained by calculating the second predicted cross-sectional radius, the second predicted control point matrix, the true control point matrix, and the true cross-sectional radius using the sixth equation. The sixth equation is: ， in, , , in, Let be the target loss function. and All are loss weighting coefficients. The order-preserving MSE loss function for control points. The radius-smoothing L1 loss function is used. The second predicted control point matrix is ​​the first... The second predictive control point Each dimension coordinate, The first in the true control point matrix The first real control point Each dimension coordinate, The radius of the second predicted section. This is the actual cross-sectional radius.

7. The method for reconstructing point clouds of curved cylindrical structures according to claim 1, characterized in that, The process of reconstructing and analyzing the first predicted cross-sectional radius, the scale parameters to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, the centroid coordinates to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and obtaining the point cloud reconstruction result of the curved cylinder structure based on the reconstruction analysis results includes: The central axis control point matrix is ​​obtained by calculating the scale parameters to be reconstructed, the centroid coordinates to be reconstructed corresponding to the 3D point cloud data of the cylindrical body to be reconstructed, and the first predicted control point matrix corresponding to the 3D point cloud data of the cylindrical body to be reconstructed using the seventh equation. The seventh equation is as follows: ， in, This is the matrix of control points for the central axis. This is the matrix of the first predicted control points. The scale parameters to be reconstructed are... The coordinates of the centroid to be reconstructed; The target cross-sectional radius is obtained by calculating the scale parameters to be reconstructed and the first predicted cross-sectional radius corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed using the eighth equation. The eighth equation is: ， in, The target cross-sectional radius, The radius of the first predicted section. The scale parameters to be reconstructed; The geometric parameters of the curved cylinder are constructed using the central axis control point matrix and the target cross-sectional radius, and the geometric parameters of the curved cylinder are used as the point cloud reconstruction result of the curved cylinder structure.

8. A point cloud reconstruction device with a curved cylindrical structure, characterized in that, include: The point cloud data acquisition module is used to obtain the original 3D point cloud data of the curved cylinder and the 3D point cloud data of the curved cylinder to be reconstructed from the 3D scanning device. The preprocessing module is used to preprocess the original curved cylindrical 3D point cloud data and the curved cylindrical 3D point cloud data to be reconstructed to obtain a first preprocessed curved cylindrical 3D point cloud matrix corresponding to the original curved cylindrical 3D point cloud data, a second preprocessed curved cylindrical 3D point cloud matrix corresponding to the curved cylindrical 3D point cloud data to be reconstructed, a scale parameter to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed, and centroid coordinates to be reconstructed corresponding to the curved cylindrical 3D point cloud data to be reconstructed. The model analysis module is used to construct a training model and perform model analysis on the training model using the first preprocessed curved cylindrical 3D point cloud matrix to obtain a cross-section prediction model. The prediction module is used to predict the second preprocessed curved cylinder three-dimensional point cloud matrix through the cross-section prediction model to obtain the first predicted cross-section radius corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed. The reconstruction result module is used to perform reconstruction analysis on the first predicted cross-sectional radius, the scale parameters to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, the centroid coordinates to be reconstructed corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and the first predicted control point matrix corresponding to the three-dimensional point cloud data of the curved cylinder to be reconstructed, and to obtain the point cloud reconstruction result of the curved cylinder structure based on the reconstruction analysis result.

9. A point cloud reconstruction device for a curved cylindrical structure, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the point cloud reconstruction method for curved cylindrical structures as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the point cloud reconstruction method for curved cylindrical structures as described in any one of claims 1 to 7.

Images