A steam turbine blade point cloud model registration method, system, electronic device and storage medium based on coarse registration and fine registration progressive alignment

Abstract

The application discloses a turbine blade point cloud model registration method and system based on coarse registration and fine registration progressive alignment, an electronic device and a storage medium, and the method comprises the following steps: reading a theoretical model to construct an STL model; uniformly sampling the STL model to generate a theoretical point cloud; acquiring a measured scanning point cloud; simplifying the measured point cloud based on curvature similarity random sampling; establishing a spatial index of the theoretical point cloud based on a Kd tree to form a fast nearest point search mechanism; performing coarse registration by adopting normal distribution transformation; and performing fine registration by adopting an optimized nearest point iterative algorithm, wherein the fast nearest point search mechanism is utilized to search for corresponding point pairs in the fine registration. The application has the advantages of unified benchmark, high calculation efficiency and high registration accuracy.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional point cloud processing and industrial inspection technology, specifically a method, system, electronic device, and storage medium for registering turbine blade point cloud models based on progressive alignment of coarse and fine registration. Background Technology

[0002] Mechanical power units mostly adopt bladed structures. As the main working components of power machinery such as steam turbines, gas turbines, and aero engines, blades play a crucial role in the overall performance and safety stability of the machine. Blades have complex shapes, mostly twisted free-form surfaces, and are characterized by large size, large number, many types, complex structure, and high manufacturing difficulty.

[0003] The manufacturing precision of blades has a significant impact on rotor energy conversion efficiency and operational stability, and blade failures are frequent. Therefore, achieving high-precision inspection of blade profiles is crucial for ensuring the safety of power plants and improving manufacturing quality. Currently, blade profile quality evaluation faces challenges such as low measurement efficiency, inconsistent point cloud data quality, complex data processing, and insufficient accuracy and efficiency of registration algorithms. Meanwhile, quality control during blade manufacturing is hampered by complex correlations between indicators and unclear data distribution, making traditional control charts difficult to apply effectively. Standard template methods have limited accuracy, and coordinate measuring machines (CMMs) are inefficient. Non-contact measurement technologies, which balance accuracy and efficiency, are gradually becoming the mainstream method for blade inspection. After obtaining high-precision scanned point cloud data, how to effectively process the complex curved surface point cloud features of blades to achieve accurate detection of manufacturing errors has become a hot topic. However, there is currently no unified standard for existing blade profile quality evaluation technologies based on 3D point clouds, and point cloud data analysis methods are still immature.

[0004] Point cloud registration is a crucial step in detecting blade profile errors. Its goal is to spatially align the actual scanned blade point cloud data with the theoretical design model (such as a CAD model or STL model) for subsequent error analysis. Existing point cloud registration methods are mainly divided into coarse registration and fine registration. Coarse registration methods (such as feature-based registration and normal distribution transformation) can achieve global alignment even with large initial pose deviations, but their registration accuracy is limited. Fine registration methods (such as the nearest-point iteration algorithm) can achieve higher accuracy alignment, but they are sensitive to initial pose, prone to getting trapped in local optima, and have low computational efficiency when dealing with large amounts of point cloud data.

[0005] To address the aforementioned issues, existing research combines coarse and fine registration, employing a gradual "coarse-to-fine" strategy. For example, it first uses a normal distribution transformation for initial alignment, and then uses the nearest-point iteration algorithm for fine adjustment. However, existing methods still have the following shortcomings and limitations when practically applied to point cloud registration of steam turbine blades: First, the theoretical model and the reference point cloud are not consistent, and there is a lack of systematic processing procedures. Existing methods usually assume that the theoretical model has been discretized into a point cloud or directly use the original vertices of the CAD model as the reference point cloud. However, the theoretical model (such as the STL format) is a continuous surface composed of triangular patches. Simply extracting vertices or uniform sampling cannot fully express the geometric information of the surface. Furthermore, there is a lack of standardized processing procedures from the theoretical model to the point cloud data, resulting in inaccurate registration references.

[0006] Second, the large volume of blade point cloud data leads to low computational efficiency. High-precision 3D scanning of blade point cloud data typically contains millions or even tens of millions of points. Directly performing nearest-point iteration or normal distribution transformation calculations results in a massive computational burden, making it difficult to meet the real-time requirements of industrial online inspection. Existing point cloud simplification methods mostly rely on random sampling or voxel filtering, failing to fully consider the characteristics of drastic changes in blade curvature (such as blade edges and leading / leading edge regions). This easily leads to the retention of redundant points in flat areas and the loss of important geometric information in key high-curvature feature areas, affecting subsequent registration accuracy.

[0007] Third, the calculation accuracy of the nearest point from a point to the theoretical model is insufficient. During fine registration, it is necessary to repeatedly calculate the nearest point from each point in the scanned point cloud to the surface of the theoretical model. Existing methods mostly employ nearest neighbor search based on discrete point clouds (such as Kd-tree search for the nearest discrete point), rather than accurately calculating the nearest point from a point to the triangular facet surface. This results in errors in corresponding point pairs, limiting further improvements in registration accuracy.

[0008] Fourth, existing combination methods lack specific optimizations for the characteristics of turbine blades. Turbine blades have characteristics such as twisted free-form surfaces, uneven curvature distribution, and thin and sharp leading and trailing edges. General coarse and fine registration combination methods do not perform point cloud preprocessing and registration parameter optimization for these characteristics, making it difficult to balance registration accuracy and computational efficiency.

[0009] Therefore, there is an urgent need for a method and system for turbine blades that can systematically process theoretical models and scanned point clouds, and balance registration accuracy and computational efficiency, in order to solve the problems of inconsistent references, low computational efficiency and insufficient registration accuracy in existing technologies. Summary of the Invention

[0010] The purpose of this invention is to address the shortcomings of existing technologies by providing a method, system, electronic device, and storage medium for registering turbine blade point cloud models based on progressive alignment of coarse and fine registration, which features a unified reference, high computational efficiency, and high registration accuracy.

[0011] The technical objective of this invention is achieved through the following technical solution: A registration method for turbine blade point cloud models based on progressive alignment of coarse and fine registration includes the following steps: S1: Read the three-dimensional theoretical model information of the turbine blade and construct a high-precision blade STL model as a reference benchmark for point cloud registration; S2: Based on the STL model, a uniform sampling strategy is used to perform dense sampling on the model surface to generate a uniformly distributed surface point cloud with triangular patch indexes, which serves as the theoretical point cloud. S3: Obtain the measured scan point cloud of the steam turbine blade; S4: The point cloud simplification method based on curvature similarity random sampling simplifies the measured scan point cloud to obtain the simplified measured point cloud; S5: Based on the Kd-tree data structure, a spatial index is established for the theoretical point cloud to form a fast nearest point search mechanism; the fast nearest point search mechanism is used to quickly find the nearest point in the theoretical point cloud for each point in the simplified measured point cloud in the subsequent fine registration step. S6: A coarse registration method for point cloud data based on normal distribution transformation is adopted. Through grid division and probability density optimization, a preliminary global match is achieved between the theoretical point cloud and the simplified measured point cloud. S7: A point cloud data fine registration method based on an optimized nearest-point iterative algorithm is adopted. During the fine registration iteration process, the fast nearest-point search mechanism is used to find corresponding point pairs. The rigid body transformation between point pairs is calculated iteratively, and the fine registration of the point cloud is completed by combining covariance matrix decomposition and centroid translation estimation.

[0012] Further, in step S2, the method for uniformly acquiring surface point clouds based on the STL model includes: Based on the equidistant subdivision strategy of triangular facet side length information, the longest and second longest sides are segmented according to a preset distance D to generate auxiliary subdivision nodes; Within the patch, a parallel line grid generated by the segmentation points is constructed, and equidistant sampling is performed on each grid line to form a locally uniform point set; Summarize the uniformly sampled points of all triangular facets and establish the index relationship between points and facets.

[0013] Furthermore, in the fast nearest point search mechanism, the method for calculating the nearest point from a point to the STL model includes: A surface point set Q is constructed using the surface point cloud generated in step S2, wherein each sampling point in the surface point set Q has an index relationship with a corresponding triangular facet. Based on the given point P and surface point set Q in the simplified measured point cloud, the nearest neighbor point is obtained, and the candidate triangular facet is located. Calculate the perpendicular projection point P of point P onto the plane containing the candidate triangular facet. ty ; Calculation of projection point P based on vector cross product method tyGiven the spatial relationship between point P and the triangular facet, obtain the nearest point P to the triangular facet. min : If P ty If P is located inside the triangular facet, then P min For P ty ; If P ty If P is located outside any side of the triangular facet, then P min For P ty The projection point on that side; In other cases, calculate P. ty The distance to the three vertices of the triangular face is used, with the vertex having the smallest distance being P. min .

[0014] Furthermore, in step S4, the point cloud simplification method based on curvature similarity random sampling includes: Based on the measured scanned point cloud, the neighborhood point set consisting of the k nearest neighbors of each point is obtained, and the covariance matrix is ​​constructed. ; The unit eigenvectors of the small eigenvalues ​​calculated based on the covariance matrix are used as the normal vectors of the corresponding points. The curvature similarity eigenvalues ​​s of each point are then obtained, as shown in the following formula: ; in, It is the absolute value of the dot product of the normal vector of this point and the normal vectors of other points in the neighborhood point set; Random sampling is performed in each category based on the calculated feature value s and the set simplification ratio to achieve point cloud simplification.

[0015] Furthermore, in step S5, establishing a fast nearest-point search mechanism based on the Kd-tree data structure includes: The theoretical point cloud is recursively divided into planes along each dimension to construct a three-dimensional Kd tree; Based on the relationship between the query point and the planes of each dimension, determine the subtree region to which it should fall, and calculate and obtain the candidate nearest point.

[0016] Furthermore, the coarse registration method for point cloud data based on normal distribution transformation in step S6 includes: The theoretical point cloud is divided into three-dimensional grid cells of fixed size, and a normal distribution model is constructed for the point set in each cell to describe the probability distribution characteristics of points in the local space. After performing an initial rigid body transformation on the simplified measured point cloud, it is mapped to the corresponding reference mesh cell, and the probability density of the transformed points is evaluated based on the statistical model within the cell. The objective optimization function is constructed by summing the probability densities of all transformed measured points in their corresponding grid cells: ; An optimization algorithm based on the Hessian matrix is ​​used to iteratively update the transformation parameters until the objective function converges, thus completing the coarse registration of the point cloud.

[0017] Furthermore, the point cloud data fine registration method based on the optimized nearest-point iterative algorithm in step S7 includes: Using the fast nearest point search mechanism, the nearest point in the theoretical point cloud is found for each point in the simplified measured point cloud, and a point-to-point correspondence is established. By constructing the covariance matrix based on all matching point pairs and solving for the optimal rotation matrix, rigid body rotation alignment is achieved. Calculate the centroid translation vector to complete the rigid body translation transformation; The average distance between corresponding point pairs is used as the error index for iterative optimization until the error converges.

[0018] A turbine blade point cloud model registration system based on progressive alignment of coarse and fine registration includes: The first construction unit is used to read the theoretical three-dimensional model information of the turbine blade and construct a high-precision theoretical model of the blade. The second construction unit is used to perform dense sampling on the model surface based on the STL model using a uniform sampling strategy to obtain a uniformly distributed surface point cloud with triangular facet indexes. The point cloud acquisition unit is used to acquire the measured scanned point cloud of the steam turbine blades; The computational optimization unit is used to simplify the measured scanned point cloud using a point cloud simplification method based on curvature similarity random sampling, and to construct a fast nearest point search mechanism based on the Kd-tree algorithm. Coarse registration unit is used to perform preliminary position alignment between theoretical point cloud and simplified measured point cloud based on normal distribution transformation method; The fine registration unit is used to find corresponding point pairs using the fast nearest point search mechanism during the iteration process based on the optimized nearest point iterative algorithm, and to perform fine alignment on the coarse registration results.

[0019] An electronic device includes 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 aforementioned method for registering turbine blade point cloud models based on progressive alignment of coarse and fine registration.

[0020] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for registering turbine blade point cloud models based on progressive alignment of coarse and fine registration.

[0021] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention achieves high-precision matching between turbine blade point clouds and theoretical models by integrating a progressive alignment strategy that combines coarse and fine registration. By combining curvature similarity simplification with Kd-tree accelerated search technology, this invention significantly improves the computational efficiency of point cloud registration, providing a reliable foundation for blade quality inspection and subsequent analysis, and helping enterprises efficiently complete blade quality control with limited resources.

[0022] 2. The turbine blade point cloud model registration method created in this invention is based on the STL model and provides a full-process registration system for point cloud acquisition, optimization and matching applicable to complex blade structures. That is, it realizes systematic modeling from multiple key links such as point cloud sampling, point cloud simplification, search acceleration and registration optimization, and provides a foundation for subsequent defect detection and quality assessment.

[0023] 3. This invention addresses the problem of inconsistency between scanned point clouds and theoretical model benchmarks. By integrating a progressive alignment strategy that combines coarse and fine registration, this invention helps achieve high-precision matching between point cloud data and theoretical models, improves the stability and accuracy of overall registration, and ensures the reliability of subsequent geometric analysis, defect detection, and quality assessment.

[0024] 4. This invention combines a curvature similarity simplification method with a Kd-tree accelerated search strategy to preprocess point cloud data before registration, effectively reducing the number of surface point clouds while retaining key information and significantly reducing the data size; at the same time, the Kd-tree structure improves the nearest point search efficiency, thereby accelerating the overall calculation speed in the registration process and solving the problem of low calculation efficiency in turbine blade point cloud registration. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of uniform sampling of triangular facets in this invention; Figure 3 This is a schematic diagram illustrating the Kd-tree partitioning principle in this invention; Figure 4 This is a schematic diagram of coarse registration using the normal distribution transformation algorithm in this invention; Figure 5 The flowchart of the optimized nearest point iterative algorithm in this invention is shown below. Figure 6 This is a detailed flowchart of the implementation of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0027] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0028] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0029] like Figure 1 — Figure 6 As shown, a method for registering a turbine blade point cloud model based on progressive alignment of coarse and fine registration includes the following steps: S1: Read the three-dimensional theoretical model information of the turbine blade and construct a high-precision blade STL model as a reference benchmark for point cloud registration; S2: Based on the STL model, a uniform sampling strategy is used to perform dense sampling on the model surface to generate a uniformly distributed surface point cloud with triangular patch indexes, which serves as the theoretical point cloud. S3: Obtain the measured scan point cloud of the steam turbine blade; S4: The point cloud simplification method based on curvature similarity random sampling simplifies the measured scan point cloud to obtain the simplified measured point cloud; S5: Based on the Kd-tree data structure, a spatial index is established for the theoretical point cloud to form a fast nearest point search mechanism; the fast nearest point search mechanism is used to quickly find the nearest point in the theoretical point cloud for each point in the simplified measured point cloud in the subsequent fine registration step. S6: A coarse registration method for point cloud data based on normal distribution transformation is adopted. Through grid division and probability density optimization, a preliminary global match is achieved between the theoretical point cloud and the simplified measured point cloud. S7: A point cloud data fine registration method based on an optimized nearest-point iterative algorithm is adopted. During the fine registration iteration process, the fast nearest-point search mechanism is used to find corresponding point pairs. The rigid body transformation between point pairs is calculated iteratively, and the fine registration of the point cloud is completed by combining covariance matrix decomposition and centroid translation estimation.

[0030] In this embodiment, S1: Read the three-dimensional theoretical model information of the turbine blade and construct a high-precision blade STL model as a reference benchmark for point cloud registration.

[0031] In practice, the creation of the STL model involves exporting a triangular mesh model in STL format from the CAD design files of the turbine blades (e.g., STEP, IGES formats). The STL model consists of a large number of triangular faces, each containing the spatial coordinates of three vertices and a face normal vector. In this embodiment, to ensure geometric accuracy while considering subsequent computational efficiency, the number of triangular faces in the STL model is controlled between 500,000 and 2 million, and a binary format is used during export to reduce file size. This STL model serves as the reference for subsequent registration.

[0032] In this embodiment, S2: Based on the STL model, a uniform sampling strategy is used to perform dense sampling on the model surface to generate a uniformly distributed surface point cloud with triangular patch indexes, which serves as the theoretical point cloud.

[0033] As a further improvement of the present invention, step S2, the method for uniformly acquiring surface point clouds based on an STL model, includes: For each triangular facet, obtain the encrypted point cloud, such as Figure 2 As shown, the method is as follows: Based on the equidistant subdivision strategy of triangular facet side length information, the length L of the longest side of the triangular facet is calculated, and the longest side and the second longest side are segmented according to the preset distance D, thereby generating high-density auxiliary subdivision nodes. Within the patch, a parallel line grid generated by the above-mentioned dividing points is constructed, and equidistant sampling at a preset distance D is performed on each grid line to form a local uniform point set; By aggregating the uniform sampling points of all triangular facets and establishing the index relationship between points and facets, a high-resolution, structured point cloud representation of the STL model surface is achieved.

[0034] In this embodiment, S3: Obtain the measured scan point cloud of the turbine blade.

[0035] In practical applications, the data acquisition methods for the measured scanning point cloud include: acquiring leaf scanning three-dimensional point cloud data based on high-resolution digital imaging technology and blue LED light emission technology.

[0036] Specifically, a blue light 3D scanner is used to clean the turbine blades to be inspected and fix them on a rotating worktable. The blue light 3D scanner projects a coded structured grating onto the blade surface, and two high-resolution industrial cameras simultaneously acquire grating images of the deformed blade surface from different angles. The 3D coordinates corresponding to each pixel are calculated using the phase-shifting method and triangulation principle to generate a single-view point cloud. A single scan covers approximately 1 / 3 of the blade area, acquiring approximately 3 to 5 million points, with a scanning accuracy better than 0.02 mm. To obtain the complete blade profile (including the blade base, blade back, leading and trailing edges, blade root, and blade tip), the rotating worktable performs 5 to 8 scans. The multi-view point clouds are fused using the scanner's built-in marker point stitching algorithm (or automatic feature stitching) to obtain a complete measured point cloud of the blade, which is output in PLY or XYZ format. In this embodiment, the final measured point cloud contains approximately 2 to 8 million points.

[0037] In this embodiment, S4: The point cloud simplification method based on curvature similarity random sampling simplifies the measured scan point cloud to obtain the simplified measured point cloud.

[0038] In practical implementation, step S4, the point cloud simplification method based on curvature similarity random sampling, includes: Based on the measured scan point cloud, the neighborhood point set consisting of the k nearest neighbors of each point is obtained (k=30 in this embodiment), and the covariance matrix C3×3 is constructed. The covariance matrix is ​​decomposed into eigenvalues. The unit eigenvector corresponding to the smallest eigenvalue is taken as the normal vector n of that point. The curvature similarity eigenvalue s of each point is obtained, as shown in the following formula: ; in, The value of s is the absolute value of the dot product between the normal vector of this point and the normal vectors of other points in the neighborhood point set. Feature s reflects the similarity of the curvature of the point cloud, and its value is between 0 and 1. The closer s is to 1, the flatter the local region is; the closer s is to 0, the more drastic the curvature changes.

[0039] Random sampling is performed within each category based on the calculated feature value 's' and the set simplification ratio to achieve point cloud simplification. In this embodiment, the point cloud is divided into three categories according to the 's' value: high curvature (s < 0.3), medium curvature (0.3 ≤ s < 0.7), and low curvature (s ≥ 0.7). In this embodiment, the overall simplification ratio is set to 60%, and the retention ratios for each category are as follows: 80% for high curvature, 50% for medium curvature, and 20% for low curvature. Random sampling is performed within each category, and points are selected according to their respective retention ratios to obtain the final simplified measured point cloud. It should be noted that the above values ​​can be adjusted according to the actual blade curvature distribution and accuracy requirements, and those skilled in the art can understand and reasonably modify them.

[0040] In this embodiment, S5: A spatial index is established on the theoretical point cloud based on the Kd-tree data structure to form a fast nearest point search mechanism; the fast nearest point search mechanism is used to quickly find the nearest point in the theoretical point cloud for each point in the simplified measured point cloud in the subsequent fine registration step.

[0041] In step S5, establishing a fast nearest-point search mechanism based on the Kd-tree data structure includes: The theoretical point cloud is recursively divided into planes along each dimension to construct a three-dimensional Kd tree; Based on the relationship between the query point and the planes of each dimension, determine the subtree region to which it should fall, and calculate and obtain the candidate nearest point.

[0042] like Figure 3 As shown, the specific method for constructing a three-dimensional Kd-tree is as follows: The theoretical point cloud is recursively divided along the x, y, and z directions. At each division, the median of the coordinates in the current dimension is selected as the dividing point, splitting the point cloud into left and right subtrees with approximately equal numbers of points in each subtree. This process is repeated until each leaf node contains fewer points than a preset granularity threshold (e.g., 20 points), completing the construction of the 3D Kd-tree.

[0043] Specifically, the nearest point search method is as follows: Starting from the root node, based on the comparison between the coordinates of the query point in the current partition dimension and the node coordinates, determine whether to enter the left or right subtree. Recursively proceed level by level until a leaf node is reached. Within the local region of that leaf node, calculate the Euclidean distances between all points and the query point, and record the current nearest distance and the nearest point. Then, backtrack: check if the distance from the hyperplane of another subtree to the query point is less than the current nearest distance. If it is less, continue searching in that subtree; otherwise, prune the branch. Finally, output the global nearest point.

[0044] As a further improvement to this invention, the method for calculating the nearest point from a point to the STL model in the fast nearest point search mechanism includes: A surface point set Q is constructed using the theoretical point cloud generated in step S2, wherein each sampling point in the surface point set Q has an index relationship with a corresponding triangular facet. Based on a given point P (from a simplified measured point cloud) and a point set Q, the nearest neighbor point is obtained, and candidate triangular faces are located. Calculate the perpendicular projection point P of point P onto the plane containing the candidate triangular facet. ty ; Calculation of projection point P based on vector cross product method ty Given the spatial relationship between point P and triangle ABC, find the nearest point P to triangle ABC. min The specific rules are as follows: If point P ty If P is located inside triangle ABC, then P min For P ty ; If point P ty If P is located outside side AB, then min For P ty The projection point on side AB; If point P ty If P is located outside side BC, then min For P ty The projection point on side BC; If point P ty If P is located outside side AC, then P min For P ty The projection point on edge AC; Except for the cases mentioned above, calculate P. ty The distances to the three vertices A, B, and C of the triangular face are calculated, with the vertex having the smallest distance being designated as P. min .

[0045] In this embodiment, S6: A coarse registration method for point cloud data based on normal distribution transformation is adopted to achieve preliminary global matching between the theoretical point cloud and the simplified measured point cloud through grid division and probability density optimization.

[0046] As a further improvement to the present invention, the overall process is as follows: Figure 4 As shown; in step S6, the coarse registration method for point cloud data based on normal distribution transformation includes: Construct a normal distribution grid model of the reference point cloud: The theoretical point cloud is divided into fixed-size 3D mesh cells (in this embodiment, the mesh side length is 5mm). The mean vector q and covariance matrix C are calculated for the point set within each cell, and a normal distribution model is constructed. The probability distribution of each point within a cell is as follows: ; In the formula: For each 3D point within the mesh The probability represented by the normal distribution; is the coordinate vector of points within the grid; The mean vector of coordinates within the grid; Let be the covariance matrix of all points within the grid.

[0047] Initial transformation and probability evaluation: After performing an initial rigid body transformation on the measured point cloud, it is mapped to the corresponding reference mesh cell, and the probability density of the transformed points is evaluated based on the statistical model within the cell. In this embodiment, the initial rigid body transformation is set as an identity matrix.

[0048] Construct the objective optimization function: The objective optimization function is constructed by summing the probability densities of all transformed measured points in their corresponding grid cells: ; Iterative optimization: The transformation parameters are iteratively updated using an optimization algorithm based on the Hessian matrix. The coarse registration of the point cloud is completed until the objective function converges (in this embodiment, the change between two adjacent iterations is less than 1×10⁻⁶ or the number of iterations reaches 30).

[0049] In this embodiment, S7: A point cloud data fine registration method based on an optimized nearest-point iterative algorithm is adopted. During the fine registration iteration process, the fast nearest-point search mechanism is used to find corresponding point pairs. The rigid body transformation between point pairs is calculated iteratively, and the fine registration of the point cloud is completed by combining covariance matrix decomposition and centroid translation estimation.

[0050] As a further improvement to the present invention, the overall process is as follows: Figure 5 As shown; in step S7, the point cloud data fine registration method based on the optimized nearest-point iterative algorithm includes: Using the fast nearest point search mechanism established in step S5, the nearest point in the theoretical point cloud is found for each point in the simplified measured point cloud, and a point pair correspondence is established. By constructing the covariance matrix based on all matching point pairs and solving for the optimal rotation matrix, rigid body rotation alignment is achieved. Calculate the centroid translation vector to complete the rigid body translation transformation; The average distance between corresponding point pairs is used as the error index for iterative optimization until the error converges.

[0051] Example 2 A turbine blade point cloud model registration system based on progressive alignment of coarse and fine registration includes: The first construction unit is used to read the theoretical three-dimensional model information of the turbine blade and construct a high-precision theoretical model of the blade. The second construction unit is used to perform dense sampling on the model surface based on the STL model using a uniform sampling strategy to obtain a uniformly distributed surface point cloud with triangular facet indexes. The point cloud acquisition unit is used to acquire the measured scanned point cloud of the steam turbine blades; The computational optimization unit is used to simplify the measured scanned point cloud using a point cloud simplification method based on curvature similarity random sampling, and to construct a fast nearest point search mechanism based on the Kd-tree algorithm. Coarse registration unit is used to perform preliminary position alignment between theoretical point cloud and simplified measured point cloud based on normal distribution transformation method; The fine registration unit is used to find corresponding point pairs using the fast nearest point search mechanism during the iteration process based on the optimized nearest point iterative algorithm, and to perform fine alignment on the coarse registration results.

[0052] The implementation of each unit and its function in the above system is completely consistent with steps S1 to S7 in the aforementioned method embodiment 1. For details, please refer to the description of steps S1 to S7, which will not be repeated here.

[0053] Example 3 An electronic device includes 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 turbine blade point cloud model registration method based on progressive alignment of coarse and fine registration as described in Embodiment 1.

[0054] Example 4 A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the turbine blade point cloud model registration method based on progressive alignment of coarse and fine registration as described in Embodiment 1.

[0055] The technical solutions provided by the embodiments of the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the embodiments of the present invention. The descriptions of the embodiments above are only for helping to understand the principles of the embodiments of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for registering point cloud models of steam turbine blades based on progressive alignment of coarse and fine registration, characterized in that, Includes the following steps: S1: Read the three-dimensional theoretical model information of the turbine blade and construct a high-precision blade STL model as a reference benchmark for point cloud registration; S2: Based on the STL model, a uniform sampling strategy is used to perform dense sampling on the model surface to generate a uniformly distributed surface point cloud with triangular patch indexes, which serves as the theoretical point cloud. S3: Obtain the measured scan point cloud of the steam turbine blade; S4: The point cloud simplification method based on curvature similarity random sampling simplifies the measured scan point cloud to obtain the simplified measured point cloud; S5: Based on the Kd-tree data structure, a spatial index is established for the theoretical point cloud to form a fast nearest point search mechanism; the fast nearest point search mechanism is used to quickly find the nearest point in the theoretical point cloud for each point in the simplified measured point cloud in the subsequent fine registration step. S6: A coarse registration method for point cloud data based on normal distribution transformation is adopted. Through grid division and probability density optimization, a preliminary global match is achieved between the theoretical point cloud and the simplified measured point cloud. S7: A point cloud data fine registration method based on an optimized nearest-point iterative algorithm is adopted. During the fine registration iteration process, the fast nearest-point search mechanism is used to find corresponding point pairs. The rigid body transformation between point pairs is calculated iteratively, and the fine registration of the point cloud is completed by combining covariance matrix decomposition and centroid translation estimation.

2. The method for registering turbine blade point cloud models based on progressive alignment of coarse and fine registration as described in claim 1, characterized in that, In step S2, the method for uniformly acquiring surface point clouds based on the STL model includes: Based on the equidistant subdivision strategy of triangular facet side length information, the longest and second longest sides are segmented according to a preset distance D to generate auxiliary subdivision nodes; Within the patch, a parallel line grid generated by the segmentation points is constructed, and equidistant sampling is performed on each grid line to form a locally uniform point set; Summarize the uniformly sampled points of all triangular facets and establish the index relationship between points and facets.

3. The method for registering turbine blade point cloud models based on progressive alignment of coarse and fine registration as described in claim 1, characterized in that, In the fast nearest point search mechanism, the method for calculating the nearest point from a point to the STL model includes: A surface point set Q is constructed using the surface point cloud generated in step S2, wherein each sampling point in the surface point set Q has an index relationship with a corresponding triangular facet. Based on the given point P and surface point set Q in the simplified measured point cloud, the nearest neighbor point is obtained, and the candidate triangular facet is located. Calculate the perpendicular projection point P of point P onto the plane containing the candidate triangular facet. ty ; Calculation of projection point P based on vector cross product method ty Given the spatial relationship between point P and the triangular facet, obtain the nearest point P to the triangular facet. min : If P ty If P is located inside the triangular facet, then P min For P ty ; If P ty If P is located outside any side of the triangular facet, then P min For P ty The projection point on that side; In other cases, calculate P. ty The distance to the three vertices of the triangular face is used, with the vertex having the smallest distance being P. min .

4. The method for registering turbine blade point cloud models based on progressive alignment of coarse and fine registration as described in claim 1, characterized in that, In step S4, the point cloud simplification method based on curvature similarity random sampling includes: Based on the measured scanned point cloud, the neighborhood point set consisting of the k nearest neighbors of each point is obtained, and the covariance matrix is ​​constructed. ; The unit eigenvectors of the small eigenvalues ​​calculated based on the covariance matrix are used as the normal vectors of the corresponding points. The curvature similarity eigenvalues ​​s of each point are then obtained, as shown in the following formula: ； in, It is the absolute value of the dot product of the normal vector of this point and the normal vectors of other points in the neighborhood point set; Random sampling is performed in each category based on the calculated feature value s and the set simplification ratio to achieve point cloud simplification.

5. The method for registering turbine blade point cloud models based on progressive alignment of coarse and fine registration as described in claim 1, characterized in that, In step S5, establishing a fast nearest-point search mechanism based on the Kd-tree data structure includes: The theoretical point cloud is recursively divided into planes along each dimension to construct a three-dimensional Kd tree; Based on the relationship between the query point and the planes of each dimension, determine the subtree region to which it should fall, and calculate and obtain the candidate nearest point.

6. The method for registering turbine blade point cloud models based on progressive alignment of coarse and fine registration according to claim 1, characterized in that, The coarse registration method for point cloud data based on normal distribution transformation in step S6 includes: The theoretical point cloud is divided into three-dimensional grid cells of fixed size, and a normal distribution model is constructed for the point set in each cell to describe the probability distribution characteristics of points in the local space. After performing an initial rigid body transformation on the simplified measured point cloud, it is mapped to the corresponding reference mesh cell, and the probability density of the transformed points is evaluated based on the statistical model within the cell. The objective optimization function is constructed by summing the probability densities of all transformed measured points in their corresponding grid cells: ； An optimization algorithm based on the Hessian matrix is ​​used to iteratively update the transformation parameters until the objective function converges, thus completing the coarse registration of the point cloud.

7. The method for registering turbine blade point cloud models based on progressive alignment of coarse and fine registration according to claim 1, characterized in that, The point cloud data fine registration method based on the optimized nearest-point iterative algorithm in step S7 includes: Using the fast nearest point search mechanism, the nearest point in the theoretical point cloud is found for each point in the simplified measured point cloud, and a point-to-point correspondence is established. By constructing the covariance matrix based on all matching point pairs and solving for the optimal rotation matrix, rigid body rotation alignment is achieved. Calculate the centroid translation vector to complete the rigid body translation transformation; The average distance between corresponding point pairs is used as the error index for iterative optimization until the error converges.

8. A registration system for a turbine blade point cloud model based on progressive alignment of coarse and fine registration, characterized in that, include: The first construction unit is used to read the theoretical three-dimensional model information of the turbine blade and construct a high-precision theoretical model of the blade. The second construction unit is used to perform dense sampling on the model surface based on the STL model using a uniform sampling strategy to obtain a uniformly distributed surface point cloud with triangular facet indexes. The point cloud acquisition unit is used to acquire the measured scanned point cloud of the steam turbine blades; The computational optimization unit is used to simplify the measured scanned point cloud using a point cloud simplification method based on curvature similarity random sampling, and to construct a fast nearest point search mechanism based on the Kd-tree algorithm. Coarse registration unit is used to perform preliminary position alignment between theoretical point cloud and simplified measured point cloud based on normal distribution transformation method; The fine registration unit is used to find corresponding point pairs using the fast nearest point search mechanism during the iteration process based on the optimized nearest point iterative algorithm, and to perform fine alignment on the coarse registration results.

9. An electronic device 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 turbine blade point cloud model registration method based on progressive alignment of coarse and fine registration as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the turbine blade point cloud model registration method based on progressive alignment of coarse and fine registration as described in any one of claims 1 to 7.

Images