A three-dimensional semantic feature extraction method and system for a machine case type part

Abstract

The application belongs to the field of high-end equipment manufacturing and intelligent manufacturing, and provides a three-dimensional semantic feature extraction method and system for machine case parts, wherein the method collects multi-view point clouds through a fixed line laser scanning head and a high-precision servo rotating platform, and arranges reference mark points; combined with the rotationally symmetrical axis extracted by principal component analysis and the mark points, coarse registration, fine registration and point cloud fusion are performed to generate a complete machine case surface three-dimensional point cloud model; based on the topological category identification of the machine case surface three-dimensional point cloud model, a geometric template is loaded, global parameters and local features with function semantics are extracted, and a structured three-dimensional semantic feature set is organized. The method can directly drive digital prototype updating, realize high-fidelity mapping from a physical entity to a digital model, and support virtual verification and closed-loop optimization in high-end equipment research and development.

Description

Technical Field

[0001] This invention belongs to the field of high-end equipment manufacturing and intelligent manufacturing, and involves industrial digital prototype generation and three-dimensional geometric modeling technology. Specifically, it relates to a method and system for extracting three-dimensional semantic features for aero-engine or gas turbine casing parts. Background Technology

[0002] Digital prototypes are the core carriers of high-end equipment R&D, and their accuracy depends on the high-fidelity digitization of the three-dimensional geometric features of parts. As a key load-bearing and connecting component of the power system, the complex geometric features of the aero-engine casing, such as its rotational profile, flanges, sealing grooves, and hole systems, directly determine the reliability of digital prototypes in virtual assembly, fluid dynamics simulation, and structural strength analysis. However, current technology still faces the following bottlenecks:

[0003] First, there is the traditional coordinate measuring machine (CMM) mapping technology, which uses contact-based point-by-point scanning. This technology is inefficient and cannot fully cover deep cavities, thin walls, or high curvature areas, making it impossible to efficiently acquire geometric data of the entire surface of the casing. Second, general-purpose 3D scanning systems, although they can quickly generate point clouds or triangular mesh models, lack engineering semantics in their output. They are difficult to parametrically extract typical features of the casing, such as mounting holes and sealing grooves, and cannot directly map them to the parametric model of the digital prototype to drive the digital prototype update. Third, there is a data gap between design and manufacturing. The actual manufacturing deviations of the casing are not effectively captured and fed back to the digital prototype, resulting in a mismatch between the geometric features of the digital prototype model and the physical entity, which weakens the authenticity of simulation and verification.

[0004] Therefore, there is an urgent need for a 3D feature recognition technology for the casing in digital prototype generation. Through high-fidelity, semantic 3D geometric feature extraction, the geometric features between the physical casing and the digital prototype can be accurately aligned, supporting virtual verification and closed-loop iterative optimization in the digital mainline. Summary of the Invention

[0005] To address the technical challenges of existing point cloud processing, such as the lack of semantic features, unavailable parameters, and the disconnect between design and manufacturing data, and to achieve fully automated, high-precision extraction from physical casings to structured semantic feature sets, directly serving the construction and updating of digital prototypes, this invention discloses a method and system for extracting three-dimensional semantic features for casing-type parts. This method includes multi-view high-precision point cloud acquisition, geometric feature recognition with functional semantics, and structured output, and is applicable to digital mainline application scenarios such as digital prototype construction, virtual verification, and design-manufacturing data closure.

[0006] Specifically, the technical solution for achieving the objective of this invention is as follows: On one hand, this invention discloses a method for extracting three-dimensional semantic features of casing-type parts, the method comprising the following steps: S1: Mount the casing parts on a high-precision servo rotary positioning platform, and use a fixed line laser scanning head to perform circumferential scanning on the casing parts according to a preset angle step size to collect multi-view three-dimensional point cloud data. S2: Arrange reference marks on the rigid protruding structure of the casing part for auxiliary registration; S3: Extract the rotational symmetry axis of the casing from the multi-view 3D point cloud data using principal component analysis. Based on the reference marker points and the rotational symmetry axis, perform registration and fusion processing on the multi-view 3D point cloud data to generate a 3D point cloud model of the casing surface. S4: Perform topological morphology category recognition on the three-dimensional point cloud model of the casing surface to obtain the topological morphology category. Based on the geometric template corresponding to the topological morphology category, extract global shape parameters and discrete local features with functional semantics from the three-dimensional point cloud model of the casing surface. S5: Organize the global shape parameters and discrete local features into a structured three-dimensional semantic feature set, which includes feature type, geometric parameters, spatial pose, functional semantics and quality indicators.

[0007] Furthermore, in step S1, the preset angle step size is 15°~30°, the overlap rate of the three-dimensional point cloud data of adjacent viewpoints is not less than 40%, and the position of the reference marker point avoids the thin-walled area, the curvature change area and the easily deformable area on the casing part.

[0008] Further, in step S2, the rigid protruding structure includes at least one of a process boss, a positioning boss, a flange end face, or a mounting edge, and the reference marking point is composed of a high reflectivity material or a high contrast color block.

[0009] Furthermore, the high-contrast color block is a circular target, comprising a black base and a white center.

[0010] Furthermore, in step S3, the registration and fusion process includes a coarse registration stage, a fine registration stage, and a point cloud fusion stage; In the coarse registration stage, an initial rigid body transformation matrix is ​​calculated based on the reference marker points and the rotation symmetry axis to perform preliminary alignment of the multi-view 3D point cloud data. In the fine registration stage, the point cloud aligned by the initial rigid body transformation matrix is ​​used as input. At least one of the following is selected from the key structural regions of the casing parts: flange connection surface, sealing groove periphery, hole distribution area and mounting edge. Points with significant geometric features are extracted as seed points and assigned weights positively correlated with curvature values. The weighted iterative nearest point algorithm is used to iteratively optimize and solve the fine transformation matrix based on the initial rigid body transformation matrix until the registration error is less than a preset threshold or the maximum number of iterations is reached. In the point cloud fusion stage, all viewpoint point clouds aligned by the fine transformation matrix are merged, deduplicated, and have holes compensated to output a complete 3D point cloud model of the casing surface.

[0011] Further, in step S4, global shape parameters with functional semantics and discrete local features are extracted from the three-dimensional point cloud model of the casing surface, including: S41: Denoise, normal vector estimation and adaptive downsampling are performed on the three-dimensional point cloud model of the casing surface to generate a preprocessed point cloud; S42: Sample the preprocessed point cloud to generate multiple point cloud local blocks, and input the point cloud local blocks into a trained deep learning classifier to identify the overall topological morphology category of the casing parts, and load the corresponding geometric template and feature extraction rules from the preset category template library according to the identification results. S43: Based on the geometric template and the feature extraction rules, perform model fitting on the key geometric elements in the preprocessed point cloud. The model fitting includes: fitting a cylindrical model to the rotary contour, fitting a planar model to the flange end face, and fitting a circular trajectory model to the sealing groove. S44: Extract global shape parameters based on the model fitting results. The global shape parameters include the gyration radius sequence segmented along the axial direction, the outer diameter and thickness of each flange, the number and height of the process bosses, and the overall axial length. S45: Identify discrete local features under the constraints of the global shape parameters. The discrete local features include the mounting holes on the flange end face, the centerline trajectory and cross-sectional dimension parameters of the sealing groove, the center coordinates and diameter of the weight reduction hole, and the inlet and outlet spatial coordinates of the internal flow channel. S46: Based on the predefined functional semantic mapping relationship in the feature extraction rules, bind a corresponding functional semantic label to each identified feature.

[0012] Furthermore, the deep learning classifier is a lightweight model optimized with the PointNet++ architecture and deployed on edge computing devices. Its input is a local patch of point cloud, and its output is a probability distribution of topological morphology categories.

[0013] In an improved embodiment of the above-described method for extracting three-dimensional semantic features of casing-type parts, the method further includes: S6: Map the structured 3D semantic feature set to the parametric feature tree of the initial digital prototype model, update the model parameters in a feature-driven manner, and generate an editable digital prototype that retains the original modeling constraints.

[0014] On the other hand, the present invention provides a three-dimensional semantic feature extraction system for casing-type parts to implement the above method. The system includes a multi-view three-dimensional data acquisition module, an intelligent recognition module, and a data interaction and processing layer.

[0015] Specifically, the multi-view 3D data acquisition module executes the above steps S1 and S2 to acquire multi-view 3D point cloud data of the casing parts, including a fixed line laser scanning head, a high-precision servo rotary positioning platform, an auxiliary positioning tool for setting up reference markers, a multi-angle lighting system, and a synchronous controller. The intelligent recognition module executes the above steps S3 and S4 to receive the multi-view 3D point cloud data, perform point cloud registration and fusion, topological morphology recognition, global and discrete feature extraction with functional semantics, and generate a structured 3D semantic feature set. The data interaction and processing layer executes the above step S5, connecting the multi-view 3D data acquisition module and the intelligent recognition module to realize real-time transmission of point cloud data, command downlink control, and status synchronization.

[0016] Furthermore, the multi-view 3D data acquisition module is externally equipped with a light-shielding structure to suppress ambient light interference, and the intelligent recognition module is deployed on an edge computing device to support unattended automatic triggering.

[0017] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least: 1. By combining multi-view line laser scanning with three-stage point cloud registration and fusion, the system can completely and accurately capture the complex geometric features of the casing surface (including rotational profiles, flanges, sealing grooves and hole systems, etc.), overcoming the limitations of traditional contact measurement in covering deep cavity and thin-walled areas, and providing a high-precision, full-coverage geometric data foundation for the construction of digital prototypes.

[0018] 2. The extracted features are output in a structured form with functional semantics, which can be directly mapped to the parametric feature tree of the digital prototype, driving automatic model updates or virtual assembly verification. This eliminates the need for manual reverse modeling, significantly shortening the conversion cycle from physical prototype to digital prototype and improving the realism and engineering usability of the digital prototype.

[0019] 3. The initial digital prototype is corrected by reverse engineering based on the measured geometric features of the real casing, and the manufacturing deviation is explicitly fed back to the design model to make up for the geometric mismatch between the theoretical design and the physical entity, so that the digital prototype is closer to the actual state, thereby supporting higher-fidelity virtual simulation analysis of fluid, structure and other aspects.

[0020] 4. The system architecture integrates scanning, registration, recognition and output functions, and supports end-to-end automated processing from raw point cloud to structured semantic feature set. It can complete the full feature digitization of the casing in a single unattended process, meeting the needs of rapid iteration and efficient verification in the research and development of high-end equipment.

[0021] 5. Unlike general 3D reconstruction systems, this invention focuses on the core requirements of casing-type parts in the digital prototype. Through the collaborative optimization of dedicated hardware configuration (such as high-contrast benchmarks and servo rotation platforms) and software algorithms (such as template-guided geometric fitting and semantic annotation), it effectively solves the key bottleneck of consistency between the geometric features of the physical entity and the digital prototype, and has clear engineering application value. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the three-dimensional semantic feature extraction method for casing-type parts according to the present invention; Figure 2 This is an architecture diagram of the three-dimensional semantic feature extraction system for casing-type parts according to the present invention; Figure 3 This is a schematic diagram of the multi-view three-dimensional data acquisition module of the present invention; Among them, 201 is a multi-view 3D data acquisition module; 202 is an intelligent recognition module; 203 is a data interaction and processing layer; 301 is a 3D laser scanner; and 302 is a high-precision servo rotation mechanism. Detailed Implementation

[0024] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0025] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features of the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] This invention discloses a method for extracting three-dimensional semantic features of casing-type parts. This method can achieve fully automatic and high-precision extraction from physical casing to structured semantic feature set by multi-view high-precision point cloud acquisition, geometric feature recognition with functional semantics and structured output. It can directly serve the construction and updating of digital prototypes and is suitable for digital mainline application scenarios such as digital prototype construction, virtual verification and design-manufacturing data closed loop.

[0027] Specifically, such as Figure 1 As shown, the method includes the following steps: S1: The casing parts are mounted on a high-precision servo rotary positioning platform, and a fixed line laser scanning head is used to perform circumferential scanning on the casing parts according to a preset angle step size to collect multi-view three-dimensional point cloud data.

[0028] When performing circumferential scanning on the casing parts, the preset angle step can be set to 15°~30°, the overlap rate of the three-dimensional point cloud data of adjacent viewpoints is not less than 40%, and the position of the reference marker point avoids the thin-walled area, curvature change area and easily deformable area on the casing parts.

[0029] S2: Reference marking points for auxiliary registration are arranged on the rigid protruding structure of the casing part. The rigid protruding structure includes at least one of a process boss, a positioning boss, a flange end face, or a mounting edge. The reference marking points are made of a high-reflectivity material or a high-contrast color block, so that they form a distinguishable grayscale or intensity difference with the casing surface in the 3D scan image to ensure reliable identification. Preferably, the high-contrast color block is a circular target, including a black base and a white center, with a diameter of 3mm to 8mm and a surface flatness better than ±0.02mm.

[0030] The rigid structure possesses high rigidity and outstanding spatial stability, which can effectively avoid deformation problems caused by scanning vibration or tooling clamping external forces. Furthermore, relying on the dual reference of the reference marker point and the geometric symmetry of the casing, the initial relative position of the point cloud from each viewpoint can be accurately determined.

[0031] S3: Extract the rotational symmetry axis of the casing from the multi-view 3D point cloud data using principal component analysis. Based on the reference marker points and the rotational symmetry axis, perform registration and fusion processing on the multi-view 3D point cloud data to generate a 3D point cloud model of the casing surface.

[0032] Specifically, the registration and fusion process employs a high-precision registration algorithm to iteratively optimize the casing point cloud from all perspectives, eliminating spatial positional differences caused by multi-view scanning and ultimately generating a complete and continuous 3D point cloud of the casing surface. This process includes three stages: coarse registration, fine registration, and point cloud fusion.

[0033] The first stage (coarse registration stage): Based on the reference markers and the axis of rotational symmetry, an initial rigid body transformation matrix is ​​calculated to perform preliminary alignment of the multi-view 3D point cloud data. Specifically, this stage is used for point cloud pair matching and corresponding point search, and includes the following processes: First, feature point sets are extracted from multi-view 3D point cloud data (source point cloud). Feature point selection should focus on reference marker points with significant geometric recognizability on the surface of the casing, including but not limited to edge points of casing bosses, corner points of flange surfaces, and curvature change points of arc surfaces. At the same time, a reference point cloud is determined—the point cloud under the scanning view of the casing body is selected as the initial reference point cloud.

[0034] Secondly, in the overlapping area of ​​the source point cloud and the reference point cloud, that is, the casing parts that can be acquired from adjacent scanning perspectives, such as a certain process boss or annular flange surface, the Euclidean distance and normal vector similarity are used as dual indicators for judgment: the Euclidean distance between each feature point in the source point cloud and the point in the reference point cloud is calculated, and the candidate point with the smallest distance is selected; then, by comparing the consistency of the normal vector direction between the candidate point and the feature point of the source point, the final source point-corresponding point matching pair is determined, forming a point pair set.

[0035] Furthermore, considering the rotational symmetry of the casing, additional geometric feature constraints are introduced: using the center coordinates of the casing's arc surface and the normal vector direction of the flange surface as constraints, the search range of corresponding points is limited, avoiding incorrect matching caused by the single point cloud feature of the smooth area on the casing surface, and ensuring the matching accuracy of the point pair set.

[0036] The second stage (the fine registration stage): Using the point cloud aligned by the initial rigid body transformation matrix as input, at least one of the following is selected from the key structural regions of the casing parts: flange connection surface, sealing groove periphery, hole distribution area, and mounting edge. Points with significant geometric features are extracted as seed points and assigned weights positively correlated with curvature values. The weighted iterative nearest point (W-ICP) algorithm is used to iteratively optimize and solve the fine transformation matrix based on the initial rigid body transformation matrix until the registration error is less than a preset threshold or the maximum number of iterations is reached.

[0037] The third stage (point cloud fusion stage): merge, deduplicate and compensate for holes in all viewpoint point clouds aligned by the fine transformation matrix, and output a complete 3D point cloud model of the casing surface.

[0038] The specific implementation processes for the second and third phases are as follows: First, solve for the transformation matrix: Based on the set of point pairs obtained in the first stage, the objective function is constructed using the least squares method. The objective is to minimize the spatial positional deviation between the source point cloud and the corresponding points of the reference point cloud after transformation. The rigid body transformation matrix that satisfies this objective is solved. This matrix contains 3 translational degrees of freedom and 3 rotational degrees of freedom, corresponding to the translation amount and rotation angle in the X, Y, and Z directions of space, respectively. It can completely describe the spatial attitude adjustment actions that the source point cloud needs to perform in order to move closer to the position of the reference point cloud.

[0039] Considering that the casing is a large and precision component, and its bosses, flanges and other structures are core functional areas with higher requirements for registration accuracy, higher weights are given to the "source point-corresponding point" matching pairs of key parts such as the edge of the boss and the sealing surface of the flange when solving the transformation matrix. This ensures that the solution of the transformation matrix prioritizes meeting the registration accuracy requirements of the core structure and avoids the core area from being affected by overall registration deviation in subsequent applications.

[0040] Secondly, point cloud transformation and error calculation are performed. Using the rigid body transformation matrix obtained by solving the transformation matrix, a spatial transformation operation is performed on the source point cloud. That is, according to the translation and rotation angle in the matrix, all points of the source point cloud are mapped to the coordinate system of the reference point cloud, thus completing the spatial attitude adjustment of the source point cloud.

[0041] After the transformation is completed, the average distance error between corresponding point pairs in the adjusted source point cloud and the reference point cloud is calculated. The root mean square error (RMSE) is used as the quantification index, and the calculation formula is as follows: ,in Let i be the i-th point of the transformed source point cloud. Let be the i-th corresponding point in the reference point cloud, and n be the total number of matching pairs.

[0042] The calculated average distance error is compared with the preset accuracy threshold. If the error is greater than the accuracy threshold, it means that there is still a significant positional deviation between the source point cloud and the reference point cloud, and the next iteration needs to be performed. If the error is less than or equal to the accuracy threshold, the current iteration is terminated, and the source point cloud and the reference point cloud are accurately registered.

[0043] Optimal, perform iterative loops and global optimization. Repeat the above cycle of "corresponding point search - transformation matrix solution - point cloud transformation and error calculation". In each iteration, based on the source point cloud after the previous transformation, re-search for corresponding points, solve the new transformation matrix and adjust the point cloud position so that the average distance error between the source point cloud and the reference point cloud gradually decreases until the error meets the accuracy threshold, and the registration of a single set of source point cloud and reference point cloud is completed.

[0044] To address the registration requirements of multi-view casing point clouds, this invention employs a global iterative optimization strategy: First, iterative registration of all pairwise point clouds is completed sequentially according to the order of adjacent scanning views; then, with the goal of minimizing the global overall error of all view point clouds, i.e., the sum of errors of all adjacent registered point clouds, the rigid body transformation matrix of each view point cloud is jointly adjusted to eliminate the "cumulative error" caused by the accumulation of pairwise registration, ultimately ensuring that the casing point clouds of all views are seamlessly stitched together in the same global coordinate system, thus forming a continuous and complete three-dimensional point cloud model of the casing surface.

[0045] S4: Perform topological morphology category recognition on the three-dimensional point cloud model of the casing surface to obtain the topological morphology category. Based on the geometric template corresponding to the topological morphology category, extract global shape parameters and discrete local features with functional semantics from the three-dimensional point cloud model of the casing surface.

[0046] In some embodiments, extracting global shape parameters and discrete local features with functional semantics from the three-dimensional point cloud model of the casing surface includes the following steps: S41: Denoise, normal vector estimation and adaptive downsampling are performed on the three-dimensional point cloud model of the casing surface to generate a preprocessed point cloud.

[0047] In practice, an improved statistical filtering algorithm can be used to remove outliers from the cassette point cloud, which includes the following steps: First, based on the density of the cassette point cloud, determine the number of neighboring points k, and calculate the average distance from each point to its k neighboring points. Secondly, based on the overall noise distribution of the cassette point cloud, a standard deviation threshold is dynamically set, and points whose distance exceeds the threshold are identified as noise and deleted. Finally, an adaptive voxel downsampling method is used to simplify the denoised point cloud: for fine structures such as thin walls of the casing and edges of bosses, or smooth curved surfaces of the main body, the voxel size is set separately. This reduces the amount of data while ensuring that fine structural features are not lost, thus improving the efficiency of subsequent calculations.

[0048] S42: Local region sampling is performed on the preprocessed point cloud to generate multiple point cloud local blocks, and the point cloud local blocks are input into a trained deep learning classifier to identify the overall topological morphology category of the casing parts, and the corresponding geometric templates and feature extraction rules are loaded from a preset category template library according to the identification results; wherein, the deep learning classifier is a lightweight model optimized by the PointNet++ architecture and deployed on an edge computing device, its input is the point cloud local blocks, and its output is the probability distribution of topological morphology categories.

[0049] In practice, the preprocessing process of the input point cloud using a trained deep learning classifier is as follows: First, local geometric attribute extraction is performed. That is, for the preprocessed point cloud, multi-dimensional local geometric attributes can be calculated point by point to provide data support for subsequent shape recognition. Specifically, this includes: 1) Normal vector calculation: The tangent plane of the k-neighborhood point set of the point is fitted by the least squares method, and the normal vector of the tangent plane is solved by singular value decomposition (SVD); combined with the rotational symmetry characteristics of the casing, the orientation of the normal vector is unified with the direction of the casing axis as the reference to avoid directional deviation in subsequent region segmentation. 2) Curvature calculation: Based on the spatial rate of change of the normal vector, the discrete curvature estimation method is used to solve the Gaussian curvature and average curvature of each point: the region of Gaussian curvature is determined as a flat region, a circular arc surface region, and a region with sharp edges and corners; 3) Convexity / Concaveness Determination: Calculate the average distance deviation between a point and its k-neighboring point set, and determine whether it is a convex or concave region, thus achieving a preliminary distinction between the concave and convex structures of the casing.

[0050] Secondly, hierarchical identification of the overall geometry of the casing is performed. Based on the distribution patterns of local geometric attributes, a hierarchical clustering method combined with geometric rule matching is adopted to achieve accurate identification of the overall geometry of the casing in three levels, specifically including: 1) Basic Topological Morphology Recognition: By analyzing the distribution range, connectivity, and spatial symmetry of different geometric attribute regions in the cassette point cloud, the basic topological morphology is identified. If there are continuous annular high curvature edges in the point cloud, and the region enclosed by the edges is a constant curvature surface of revolution, then the casing is determined to be a rotationally symmetric topological shape. If, based on a rotationally symmetric topological form, there are three or more independent convex regions, and each convex region is evenly distributed circumferentially along the rotation axis and connected to the main rotation surface through a flat transition surface, then it is determined to be a rotationally symmetric form with bosses / flanges. If a point cloud contains multiple topological branches and the geometric properties of each branch are significantly different, the position coordinates and connection angles of the branch nodes can be extracted to determine that it is a composite topological form, which is suitable for irregularly shaped casings.

[0051] 2) Core structure region segmentation: Based on the abrupt boundary changes in local geometric properties, an improved region growing algorithm is used to segment the cassette point cloud and define the core structure region. Using the edge points of Gaussian curvature as seed points, and setting the growth condition to "Gaussian curvature difference between adjacent points", the flange surface area, boss area, and mounting edge area of ​​the casing are segmented. Using points with Gaussian curvature fluctuations as seed points, and setting the growth condition to "angle between the normal vectors of adjacent points", the main rotating surface region and the arc transition surface region of the casing are segmented. Points that are concave or convex are used as seed points. The growth condition is set to "difference in distance between adjacent points". The weight reduction hole / groove area and the sealing groove area are segmented to achieve precise division of the core structure.

[0052] After identifying the overall topological morphology category of the casing components, geometric templates and feature extraction rules can be matched and loaded using a constructed casing morphology library. This morphology library contains preset typical casing morphology categories, each corresponding to a specific topological morphology and core structure distribution rules. The number, size, and distribution relationship of the core structure regions defined above are matched with the category rules in the morphology library to determine the overall morphology category of the casing, providing clear morphological constraints for subsequent key geometric model fitting.

[0053] S43: Based on the geometric template and the feature extraction rules, perform model fitting on the key geometric elements in the preprocessed point cloud. The model fitting includes fitting a cylindrical model to the rotary profile, fitting a planar model to the flange end face, and fitting a circular trajectory model to the sealing groove.

[0054] In practice, for the determined overall shape category of the casing, a suitable model (containing geometric templates and feature extraction rules) is selected from the basic geometric model library. Through the process of "initial value determination - accurate fitting - accuracy verification", the model representation of the core structure is realized. First, initial values ​​for the fitting are determined: based on the point cloud coordinates of the core structural region and combined with the design principles of the casing structure, the initial parameters of the fitting model are calculated to ensure the subsequent fitting convergence speed. For the main body of the surface of revolution: Principal component analysis (PCA) is used to solve for the principal direction of the point cloud in the core structural region, and the principal direction is used as the initial direction of the axis; by calculating the projection of the point cloud onto the plane perpendicular to the principal direction, the center of the projection circle is fitted to obtain the initial position of the axis; the radius distribution of the projection circle is statistically analyzed, and the average value is taken as the initial value of the radius or the initial value of the large / small end radius; For the flange / boss plane: the least squares method is used to fit the tangent plane to the point cloud in the region, the normal vector of the fitted plane is used as the initial value of the plane normal vector, and the distance from the plane to the origin of the coordinate system is used as the initial value of the intercept. For a circular arc transition surface: based on the curvature distribution of points within the region, the reciprocal of the average curvature is taken as the initial value of the arc radius; the initial value of the arc center is determined by fitting three feature points on the arc.

[0055] Secondly, high-precision model fitting is performed: the least squares method combined with robust M-estimation is used to reduce the impact of point cloud noise and local distortion on fitting accuracy, thus achieving high-precision fitting. Cylindrical / conical surface fitting: Construct an objective function that minimizes the distance error from a point to the surface, introduce robust weights, and solve for the optimal parameters using a nonlinear least squares optimization algorithm (Levenberg-Marquardt, LM) iterative algorithm. Set the number of iterations to ensure convergence. Plane / circular surface fitting: The weighted least squares method is used to assign low weights to edge distortion points and high weights to points in the central region. The fitting results are the plane equation Ax+By+Cz+D=0, the center coordinates (x0,y0,z0) and radius R of the circular arc. Combinatorial structural fitting: For a casing containing multiple sub-surfaces, each sub-surface is fitted separately first, then geometric constraints are introduced to construct a constraint optimization objective function, and the parameters of each sub-surface are adjusted to ensure the continuity between sub-surfaces.

[0056] Finally, the fitting accuracy is verified: the mean distance error (MAE) and maximum distance error (MaxE) between the point cloud of the core structure region and the fitted model are calculated to verify the fitting accuracy. Set accuracy thresholds: Set accuracy thresholds for critical and non-critical structures of the casing respectively; if the error exceeds the threshold, readjust the initial fitting value and repeat the fitting process in step 2; if the error meets the threshold requirements, determine the final critical geometric model as the basis for parameter extraction.

[0057] S44: Extract global shape parameters based on the model fitting results. The global shape parameters include the gyration radius sequence segmented along the axial direction, the outer diameter and thickness of each flange, the number and height of the process bosses, and the overall axial length.

[0058] In practical implementation, based on the fitted key geometric model, quantitative parameters and qualitative features that can uniquely characterize the global shape of the casing are extracted, forming a complete global shape feature parameter system, specifically including four types of parameters: 1) Core parameters for rotation, including Parameters of the main rotating surface: inner diameter d1, outer diameter d, and height H of the cylindrical surface; large end diameter D1, small end diameter D2, cone length L, and cone angle α of the conical surface; spatial coordinates (x0, y0, z0) and direction vector (a, b, c) of the rotation axis. Parameters of the ring structure: inner diameter d_f1, outer diameter d_f2, and width W_f of the flange face; groove width W_g, groove depth H_g, and distance from the center of the circle to the axis of rotation R_g.

[0059] 2) Distributed parameters for bosses / flanges, including: Boss parameters: number of bosses N, circumferential distribution angle θ along the axis of rotation, spatial coordinates (x_t, y_t, z_t) of the boss center, height H_t of the boss, and diameter d_t. Flange parameters: number of flanges M, axial distance between adjacent flange faces L_f, perpendicularity deviation of flange face to axis of rotation f, flatness deviation of flange face f_p.

[0060] 3) Overall dimensions and topology parameters, including: Overall dimensional parameters: maximum outer diameter D_max, minimum inner diameter d_min, total length L_total, and maximum axial height H_max of the casing; Topological characteristic parameters: number of core structural regions K, surface area ratio of each region η, number of concave structures P, and distribution density of concave structures ρ.

[0061] 4) Morphological correlation parameters, including: The correlation features between key geometric models are extracted, including: the transition radius R_trans between the cylindrical and conical surfaces, the connection method between the boss and the main rotating surface, the axial distance L_tf between the boss and the flange surface, the parallelism deviation f_par between the flange surface and the end face of the casing, and the coaxiality deviation f_co of each rotating surface. This forms a complete set of global shape feature parameters for the casing, providing core data support for the subsequent intelligent identification of casing parts (such as matching with a standard model library).

[0062] S45: Identify discrete local features under the constraints of the global shape parameters. The discrete local features include the mounting holes on the flange end face, the centerline trajectory and cross-sectional dimension parameters of the sealing groove, the center coordinates and diameter of the weight reduction hole, and the inlet and outlet spatial coordinates of the internal flow channel.

[0063] In practical implementation, accurate identification of discrete features requires the three-dimensional point cloud of the casing surface as the basic data source, combined with the constraint information provided by the global shape parameters. The two have a collaborative support relationship of "data source-constraint condition".

[0064] Among them, hole position features: extract the regular hole system on the surface of the casing, identify its spatial position, distribution rules and geometric parameters, and verify the regularity of the hole position through geometric constraints; Rib / Boss Features: Locate the rib or boss structure on the surface of the casing, extract its geometric parameters and connection relationship with the surrounding area; Mounting edge features: Identify the planar area of ​​the mounting edge and extract its flatness, edge straightness, and mating surface features with other components; Internal features: If the scanning range includes the inner cavity of the casing, extract the geometry and key structures of the flow channels within the cavity.

[0065] S46: Based on the predefined functional semantic mapping relationship in the feature extraction rules, bind a corresponding functional semantic label to each identified feature.

[0066] In practice, after completing multi-view point cloud registration and fusion of the chassis, global shape feature analysis, and accurate identification of discrete features, the extracted 3D geometric features need to be converted into parametric descriptions and associated with the chassis's design logic to form a 3D feature library with semantic information. This includes the following processes: ① Overall geometric morphological features, i.e., the parametric expression of the macroscopic structure of the casing: The overall geometric morphological features are the core macroscopic features of the casing for "global load-bearing capacity, assembly adaptation, and model differentiation". Based on the complete 3D point cloud of the casing after registration and fusion, they are extracted through the "local geometric attribute calculation - hierarchical morphological recognition - key model fitting" process in the global shape feature analysis stage. Specifically, they include 6 types of features: The first category is the basic characteristic of the main body's rotation: Feature function positioning: It constitutes the main frame of the casing and undertakes the core functions of "enclosing the internal rotor assembly, transmitting axial / radial loads, and forming airflow / fluid channels". It is the most basic macroscopic structure of the casing. Parametric parameters: Cylindrical surface: inner diameter d1, outer diameter d2, axial length L, axial spatial coordinates (x0, y0, z0), axial direction vector (a, b, c); Conical surface: large end diameter D1, small end diameter D2, cone length L_ cone Cone angle α, coaxiality Φ with the main cylindrical surface f Composite structure: Axial connection position z of sub-surfaces joint Transition fillet radius R trans ; The second category is characterized by flange connections: Feature function positioning: The assembly interface between the casing and adjacent components, undertaking the functions of "detachable connection and leak-proof sealing"; Parametric parameters: Basic dimensions: Inner diameter d f1 outer diameter d f2 axial width W f Axial installation position z f Accuracy parameter: end face flatness b f Perpendicularity to the main body's rotation axis ⊥ f Axial distance L between adjacent flanges f-f ; The third category is process / positioning boss features: Feature function positioning: Divided into process bosses and positioning bosses, evenly distributed along the circumference of the casing; Parameterized parameters: Basic dimensions: Quantity N t Top cross-sectional shape, circular diameter d t / Square side length S t Height H t Distribution and accuracy: circumferential distribution angle θ t , center coordinates (x t ,y t ,z t ), Parallelism between the top surface and the main body axis is parallel to t; The fourth type is axial transition characteristics: Feature function positioning: connecting different main structures of the casing, undertaking the functions of reducing stress concentration, optimizing fluid channels, and improving structural strength, and divided into rounded transition and right angle transition; Parametric parameters: Circular transition: Transition radius R trans axial length L trans Tangency with adjacent structures; Right angle transition: Chamfer dimension C / rounding radius R, direction of normal vector of transition surface; The fifth category is overall size and topological features: Feature function positioning: Characterizes the relationship between the overall size and macroscopic structure of the casing, and is the core basis for distinguishing models, assembly positioning, and selecting processing equipment; Parametric parameters: Overall dimensions: Maximum outer diameter D max , minimum inner diameter d min Total axial length L total Maximum radial height H max Topological relationships: number of core structures K, area ratio of each structure, and connection method; The sixth category is overall precision features: Key performance indicators for ensuring the "high-precision assembly, stable operation, and low loss" of the casing directly affect its performance and lifespan. Parametric parameters: Coaxiality: Coaxiality Φco between the main body's rotating surface and the flange positioning hole; Parallelism: Parallelism ∥ff between adjacent flange end faces; ② Discrete Features: Discrete features are the core local features for the realization of local functions, process adaptation, and lightweighting of the chassis. Based on the complete 3D point cloud of the chassis after registration and fusion, and combined with the constraints of global shape parameters, they are extracted through the "local point cloud extraction - clustering recognition - constraint verification" process of the discrete feature accurate identification stage. Specifically, they include 5 types of features: The first type is characterized by weight-reducing holes: Feature function positioning: While ensuring structural strength, reduce the weight of the casing and reduce operating inertia, and mostly distribute them on the flange surface and non-load-bearing areas of the main body; Parameterized parameters: Basic dimensions: Quantity N h aperture d h Hole depth H h Cross-sectional shape; distribution and accuracy: center coordinates (x h ,y h ,z h ), circumferential distribution angle θ h Parallelism between axis and main axis h Hole wall roughness Ra; The second category is the feature of the sealing groove: Feature function positioning: Install seals to achieve the "leakage prevention" function of the casing assembly interface; Parameterized parameters: Basic dimensions: Quantity N g , slot width W g , groove depth H g , groove bottom radius R g Distribution and precision: Distance L from the center of the groove to the flange end face g-f Coaxiality Φ g-co Perpendicularity between side and bottom surfaces ⊥ g ; The third type is the feature of the locating pin hole: Feature function positioning: It works with positioning pins to achieve "high-precision assembly positioning" of the casing and other components, and is mostly distributed on the top of the boss and the flange surface; Parameterized parameters: Basic dimensions: Quantity N p aperture d p Hole depth H p Accuracy parameters: center distance from the reference boss Lp-t, perpendicularity of the axis to the top surface of the boss ⊥p, center distance between adjacent pin holes L p-p ; The fourth category is threaded hole features: Feature function positioning: The casing is detachably connected to other components by bolts, which are responsible for the functions of "connection strength and repeated disassembly and assembly". They are mostly distributed on the flange face and boss. Parameterized parameters: Basic dimensions: Quantity N th Thread specifications, thread hole depth H th Including effective thread depth H th-eff Bottom hole diameter d th-dr Distribution and accuracy: Diameter D of the central distribution circle th-cir Circumferential distribution angle θ th perpendicularity of axis to flange face ⊥ th ; The fifth category is the characteristic of process gaps: Feature function positioning: to meet the process requirements of casing machining or assembly, and avoid process interference; Parameterized parameters: Basic dimensions: Quantity N n Shape, square, side length S n / Arc radius R n Depth H n Position and precision: center coordinates (x n ,y n ,z n ), distance L between the edge and adjacent structures n-adj Surface roughness Ra; S5: Organize the global shape parameters and discrete local features into a structured three-dimensional semantic feature set, which includes feature type, geometric parameters, spatial pose, functional semantics and quality indicators, to support at least one downstream task in quality assessment, reverse modeling or digital prototype updating.

[0067] In an improved embodiment of the above-described method for extracting three-dimensional semantic features of casing-type parts, the method further includes: S6: Map the structured 3D semantic feature set to the parametric feature tree of the initial digital prototype model, update the model parameters in a feature-driven manner, and generate an editable digital prototype that retains the original modeling constraints.

[0068] When updating the model, the three-dimensional feature data with semantic information can be geometrically aligned with the initial model of the digital prototype. Through feature-level fusion or deviation compensation, an updated version of the digital prototype that is consistent with the geometric features of the physical casing can be generated.

[0069] In addition, digital prototype compatibility verification was conducted: based on the extracted 3D features, the geometric compatibility between the digital prototype and the physical casing was verified, which helped R&D personnel quickly identify design-manufacturing differences and optimize the virtual assembly and simulation accuracy of the digital prototype.

[0070] Based on the same inventive concept, this invention also provides a three-dimensional semantic feature extraction system for casing-type parts, as described in the following embodiments. Since the principle of the three-dimensional semantic feature extraction system for casing-type parts is similar to the three-dimensional semantic feature extraction method for casing-type parts disclosed in the above embodiments, the implementation of the three-dimensional semantic feature extraction system for casing-type parts can refer to the implementation of the three-dimensional semantic feature extraction method for casing-type parts, and repeated details will not be elaborated further. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0071] Figure 2 This is a structural block diagram of a three-dimensional semantic feature extraction system for casing-type parts disclosed in an embodiment of the present invention, as shown below. Figure 2 As shown, the system includes a multi-view 3D data acquisition module 201, an intelligent recognition module 202, and a data interaction and processing layer 203. The structure is described below.

[0072] Specifically, the multi-view 3D data acquisition module 201 performs the above steps S1 and S2 to acquire multi-view 3D point cloud data of the casing parts, including a fixed line laser scanning head, a high-precision servo rotary positioning platform, an auxiliary positioning tool for setting reference markers, a multi-angle lighting system, and a synchronization controller.

[0073] like Figure 3 As shown, the fixed line laser scanning head can be a 3D laser scanner 301, which has high-density point cloud acquisition capabilities and a scanning range covering the entire surface of the casing parts. The high-precision servo rotation mechanism 302 is used to drive or rotate the casing parts mounted on it according to set rules, ensuring that the overlapping area of ​​adjacent data (not less than 40%) meets the complete stitching requirements during multi-view scanning, covering the entire circumference and feature area of ​​the casing to obtain complete 3D point cloud data of the casing surface. The multi-angle illumination system is used to provide illumination for the casing parts to eliminate surface reflection interference and adapt to different material properties. The auxiliary positioning tool is used to set reference marker points on the casing parts for initial coarse registration of the multi-view point clouds, ensuring data stitching consistency. The synchronization controller simultaneously acquires and stores data from the laser scanner and the rotating platform, coordinating the synchronous movements of the scanner and the rotating platform.

[0074] The intelligent recognition module 202 executes the above steps S3 and S4, which are used to receive the multi-view 3D point cloud data, perform point cloud registration and fusion, topological morphology recognition, global and discrete feature extraction with functional semantics, and generate a structured 3D semantic feature set. The intelligent recognition module 202 includes a point cloud registration and fusion module (implementing step S3), a 3D geometric feature intelligent extraction module (implementing step S4), a feature semanticization module (implementing step S5), and a digital prototype mapping module (implementing step S6).

[0075] The data interaction and processing layer 203 connects the multi-view 3D data acquisition module and the intelligent recognition module, and is used to realize the real-time transmission of point cloud data, command downlink control and status synchronization.

[0076] Furthermore, the multi-view 3D data acquisition module is externally equipped with a light-shielding structure to suppress ambient light interference, and the intelligent recognition module is deployed on an edge computing device to support unattended automatic triggering.

[0077] The embodiments of the present invention achieve the following technical effects: 1. By combining multi-view line laser scanning with three-stage point cloud registration and fusion, the system can completely and accurately capture the complex geometric features of the casing surface (including rotational profiles, flanges, sealing grooves and hole systems, etc.), overcoming the limitations of traditional contact measurement in covering deep cavity and thin-walled areas, and providing a high-precision, full-coverage geometric data foundation for the construction of digital prototypes.

[0078] 2. The extracted features are output in a structured form with functional semantics, which can be directly mapped to the parametric feature tree of the digital prototype, driving automatic model updates or virtual assembly verification. This eliminates the need for manual reverse modeling, significantly shortening the conversion cycle from physical prototype to digital prototype and improving the realism and engineering usability of the digital prototype.

[0079] 3. The initial digital prototype is corrected by reverse engineering based on the measured geometric features of the real casing, and the manufacturing deviation is explicitly fed back to the design model to make up for the geometric mismatch between the theoretical design and the physical entity, so that the digital prototype is closer to the actual state, thereby supporting higher-fidelity virtual simulation analysis of fluid, structure and other aspects.

[0080] 4. The system architecture integrates scanning, registration, recognition and output functions, and supports end-to-end automated processing from raw point cloud to structured semantic feature set. It can complete the full feature digitization of the casing in a single unattended process, meeting the needs of rapid iteration and efficient verification in the research and development of high-end equipment.

[0081] 5. Unlike general 3D reconstruction systems, this invention focuses on the core requirements of casing-type parts in the digital prototype. Through the collaborative optimization of dedicated hardware configuration (such as high-contrast benchmarks and servo rotation platforms) and software algorithms (such as template-guided geometric fitting and semantic annotation), it effectively solves the key bottleneck of consistency between the geometric features of the physical entity and the digital prototype, and has clear engineering application value.

[0082] Obviously, those skilled in the art should understand that the modules or steps of the above-described embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of 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 extracting three-dimensional semantic features of casing-type parts, characterized in that, The method includes: The casing parts are mounted on a high-precision servo rotary positioning platform, and a fixed line laser scanning head is used to perform circumferential scanning on the casing parts according to a preset angle step size to collect multi-view three-dimensional point cloud data. Reference marking points for auxiliary registration are arranged on the rigid protruding structure of the casing parts; The rotational symmetry axis of the casing is extracted from the multi-view 3D point cloud data using principal component analysis. Based on the reference marker points and the rotational symmetry axis, the multi-view 3D point cloud data is registered and fused to generate a 3D point cloud model of the casing surface. The topological morphology category of the three-dimensional point cloud model of the casing surface is identified to obtain the topological morphology category. Based on the geometric template corresponding to the topological morphology category, global shape parameters and discrete local features with functional semantics are extracted from the three-dimensional point cloud model of the casing surface. The global shape parameters and discrete local features are organized into a structured three-dimensional semantic feature set, which includes feature type, geometric parameters, spatial pose, functional semantics, and quality indicators.

2. The method for extracting three-dimensional semantic features of casing-type parts according to claim 1, characterized in that, The preset angle step size is 15°~30°, the overlap rate of the three-dimensional point cloud data of adjacent viewpoints is not less than 40%, and the position of the reference marker point avoids the thin-walled area, curvature change area and easily deformable area on the casing part.

3. The method for extracting three-dimensional semantic features of casing-type parts according to claim 1, characterized in that, The rigid protruding structure includes at least one of a process boss, a positioning boss, a flange end face, or a mounting edge, and the reference marking point is composed of a high reflectivity material or a high contrast color block.

4. The method for extracting three-dimensional semantic features of casing-type parts according to claim 3, characterized in that, The high-contrast color block is a circular target, consisting of a black base and a white center.

5. The method for extracting three-dimensional semantic features of casing-type parts according to claim 1, characterized in that, The registration and fusion process includes a coarse registration stage, a fine registration stage, and a point cloud fusion stage. In the coarse registration stage, an initial rigid body transformation matrix is ​​calculated based on the reference marker points and the rotation symmetry axis to perform preliminary alignment of the multi-view 3D point cloud data. In the fine registration stage, the point cloud aligned by the initial rigid body transformation matrix is ​​used as input. At least one of the following is selected from the key structural regions of the casing parts: flange connection surface, sealing groove periphery, hole distribution area and mounting edge. Points with significant geometric features are extracted as seed points and assigned weights positively correlated with curvature values. The weighted iterative nearest point algorithm is used to iteratively optimize and solve the fine transformation matrix based on the initial rigid body transformation matrix until the registration error is less than a preset threshold or the maximum number of iterations is reached. In the point cloud fusion stage, all viewpoint point clouds aligned by the fine transformation matrix are merged, deduplicated, and have holes compensated to output a complete 3D point cloud model of the casing surface.

6. The method for extracting three-dimensional semantic features of casing-type parts according to claim 1, characterized in that, Extracting global shape parameters and discrete local features with functional semantics from the three-dimensional point cloud model of the casing surface includes: The three-dimensional point cloud model of the casing surface is subjected to denoising, normal vector estimation and adaptive downsampling to generate a preprocessed point cloud. The preprocessed point cloud is sampled locally to generate multiple point cloud local blocks, and the point cloud local blocks are input into a trained deep learning classifier to identify the overall topological morphology category of the casing parts. Based on the identification results, the corresponding geometric templates and feature extraction rules are loaded from a preset category template library. Based on the geometric template and the feature extraction rules, the key geometric elements in the preprocessed point cloud are fitted with models. The model fitting includes: fitting a cylindrical model to the rotary contour, fitting a planar model to the flange end face, and fitting a circular trajectory model to the sealing groove. Global shape parameters are extracted based on the model fitting results. These global shape parameters include the gyration radius sequence segmented along the axial direction, the outer diameter and thickness of each flange, the number and height of the process bosses, and the overall axial length. Discrete local features are identified under the constraints of the global shape parameters. These discrete local features include the mounting holes on the flange end face, the centerline trajectory and cross-sectional dimension parameters of the sealing groove, the center coordinates and diameter of the weight reduction hole, and the inlet and outlet spatial coordinates of the internal flow channel. Based on the predefined functional semantic mapping relationship in the feature extraction rules, a corresponding functional semantic label is bound to each identified feature.

7. The method for extracting three-dimensional semantic features of casing-type parts according to claim 6, characterized in that, The deep learning classifier is a lightweight model optimized with the PointNet++ architecture and deployed on edge computing devices. Its input is a local patch of point cloud, and its output is a probability distribution of topological morphology categories.

8. The method for extracting three-dimensional semantic features of casing-type parts according to claim 1, characterized in that, The method further includes: The structured 3D semantic feature set is mapped to the parametric feature tree of the initial digital prototype model, and the model parameters are updated in a feature-driven manner to generate an editable digital prototype that retains the original modeling constraints.

9. A three-dimensional semantic feature extraction system for casing-type parts, characterized in that, The system for implementing the method as described in any one of claims 1 to 8 comprises: The multi-view 3D data acquisition module is used to acquire multi-view 3D point cloud data of the casing parts, including a fixed line laser scanning head, a high-precision servo rotary positioning platform, an auxiliary positioning tool for setting up reference markers, a multi-angle lighting system, and a synchronization controller. The intelligent recognition module is used to receive the multi-view 3D point cloud data, perform point cloud registration and fusion, topological morphology recognition, global and discrete feature extraction with functional semantics, and generate a structured 3D semantic feature set. The data interaction and processing layer connects the multi-view 3D data acquisition module and the intelligent recognition module, and is used to realize the real-time transmission of point cloud data, command downlink control and status synchronization.

10. The three-dimensional semantic feature extraction system for casing-type parts according to claim 9, characterized in that, The multi-view 3D data acquisition module is externally equipped with a light-shielding structure to suppress ambient light interference. The intelligent recognition module is deployed on an edge computing device and supports unattended automatic triggering.

Images