Defect point cloud segmentation method for turbine blade

Abstract

The invention relates to the technical field of defect identification, and discloses a defect point cloud segmentation method for a turbine blade, and the method comprises the steps: carrying out the feature self-adaptive bilateral filtering of point cloud data, and obtaining a source target point cloud; performing two-stage registration on the standard point cloud and the source target point cloud to obtain grid model data, and performing principal component analysis to obtain a feature vector; mapping the feature vectors projected to the three-dimensional space to a two-dimensional plane to obtain a two-dimensional point set, and fitting spline curves in the two-dimensional point set to obtain deformation average level and local fluctuation distribution features; the macroscopic angle inclination difference is quantified, and the overall inclination deformation amount is obtained; a multi-dimensional comparison frame is constructed based on the average deformation level, the local fluctuation distribution characteristics and the overall inclination deformation, and a deformation mode is identified based on the multi-dimensional comparison frame; performing cross validation on the deformation mode to obtain a dominant mechanism; according to the method, the turbine blade defect identification accuracy can be improved.

Description

Technical Field

[0001] This invention relates to the field of defect recognition technology, and in particular to a defect point cloud segmentation method for turbine blades. Background Technology

[0002] In the field of defect detection for complex curved surface workpieces such as turbine blades, existing technologies largely rely on manual operation by professionals. Accuracy is directly constrained by the experience level of personnel, complex background interference, and lighting conditions. The inspection process requires repeated adjustments to the observation angle to obtain clear feature information, leading to low efficiency, lengthy processing times, and a high risk of overlooking critical hidden dangers due to obstructed views or low recognition of minute defects. Furthermore, most technologies focus on post-defect detection after the defects have materialized, only conducting qualitative classification or quantitative assessment of specific defect types. This makes it difficult to capture subtle early warning signs of defect formation, failing to meet the needs of preventative maintenance for early risk warning.

[0003] In point cloud data processing and defect quantification, traditional bilateral filtering algorithms exhibit a cyclic dependency between normal vector estimation and point position updates, easily accumulating errors during iteration. This leads to the loss of geometric features or excessive smoothing. Furthermore, their fixed spatial and normal parameters lack adaptability, failing to respond to dynamic changes in the local geometric structure of the point cloud. This makes it difficult to effectively denoise while preserving edge sharpness and key geometric features. In addition, existing registration and defect analysis methods struggle to construct a multi-scale framework that collaboratively represents overall trends and local details. The segmentation accuracy of defect regions after point cloud registration is insufficient, and the inability to accurately identify deformation patterns through multi-dimensional angular differences results in low accuracy in the quantitative analysis of complex deformations such as blade deflection. This makes it difficult to reliably determine the defect-dominant mechanism and provides precise quantitative evidence for preventative maintenance. Summary of the Invention

[0004] This invention provides a method for defect point cloud segmentation for turbine blades to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for defect point cloud segmentation for turbine blades, comprising: S1. Perform feature-adaptive bilateral filtering on the point cloud data of the turbine blade to obtain the source and target point cloud of the turbine blade; S2. Perform two-level registration on the standard point cloud of the turbine blade and the source target point cloud to obtain the mesh model data of the turbine blade; S3. Perform principal component analysis on the source point cloud and target point cloud of the non-overlapping region in the grid model data to obtain the feature vector of the principal component in the turbine blade. S4. Map the feature vector projected to the three-dimensional space to the two-dimensional plane to obtain the two-dimensional point set of the turbine blade, and fit the spline curve in the two-dimensional point set to obtain the average deformation level and local fluctuation distribution characteristics of the turbine blade. S5. Quantify the macroscopic angle tilt difference in the mesh model data to obtain the overall tilt deformation of the defect area in the turbine blade; S6. Construct a multi-dimensional comparison framework for the turbine blade based on the average deformation level, the local fluctuation distribution characteristics, and the overall tilt deformation amount, and identify the deformation mode of the turbine blade based on the multi-dimensional comparison framework. S7. Cross-validate the deformation mode to obtain the dominant mechanism of the turbine blade.

[0006] In a preferred embodiment, the step of performing feature-adaptive bilateral filtering on the point cloud data of the turbine blade to obtain the source and target point cloud of the turbine blade includes: The points in the point cloud data are mapped to feature points, high curvature points, ordinary points and noise points to obtain the curvature value, noise fraction, curvature and noise threshold of the turbine blade. The adaptive weights of the turbine blades are obtained by adjusting the weight strategy of the bilateral filter based on the curvature value, the noise score, the curvature, and the noise threshold. The source and target point clouds of the turbine blades are obtained by bilateral filtering of the point cloud data based on the adaptive weights.

[0007] In a preferred embodiment, the step of performing bilateral filtering on the point cloud data based on the adaptive weights to obtain the source and target point cloud of the turbine blade includes: Based on the average displacement of the point cloud data, iterative parameter attenuation is performed on the point cloud data to obtain the source target point cloud of the turbine blade. The formula for calculating the average displacement is as follows:

[0008] In the formula, The average displacement is... The total number of points in the point cloud data. The point cloud ordinal number of the point cloud data. For point clouds In the The position of the next iteration. For point clouds In the The position of the next iteration.

[0009] In a preferred embodiment, the two-level registration of the standard point cloud of the turbine blade and the source / target point cloud to obtain the mesh model data of the turbine blade includes: Establish the corresponding constraint relationship between the standard point cloud of the turbine blade and the source target point cloud; The final value of the rigid body transformation of the defect point cloud in the corresponding constraint relationship is determined by random sampling; The final value of the rigid body transformation is projected onto Euclidean space to obtain the mesh model data of the turbine blade.

[0010] In a preferred embodiment, the step of performing principal component analysis on the source point cloud and target point cloud of the non-overlapping region in the mesh model data to obtain the feature vector of the principal components in the turbine blade includes: Extract the source point cloud and target point cloud of the non-overlapping region from the mesh model data; Based on the source point cloud and the target point cloud, a reference frame for the turbine blade is established. The covariance matrix of the point cloud in the grid model data is decomposed using the reference framework to obtain the eigenvectors of the turbine blades in the principal component directions.

[0011] In a preferred embodiment, the step of mapping the feature vector projected into three-dimensional space to a two-dimensional plane to obtain a two-dimensional point set of the turbine blade, and fitting the spline curve of the two-dimensional point set to obtain the average deformation level and local fluctuation distribution characteristics of the turbine blade, includes: Projecting the point cloud in the mesh model data onto the reference frame yields the three-dimensional coordinates of the point cloud in the mesh model data; Using the first principal component and the second principal component in the reference frame as projection planes, the three-dimensional coordinates are mapped to the projection planes to obtain the two-dimensional point set of the point cloud in the mesh model data; The spline curve of the point cloud in the grid model data is obtained by fitting the two-dimensional point set using cubic spline interpolation, and the local fluctuation distribution characteristics of the turbine blade are quantified by the tangent angle of the spline curve. The average deformation level of the turbine blade is evaluated based on the curvature changes and angular deviations of the local defect regions in the spline curve.

[0012] In a preferred embodiment, quantifying the macroscopic angular tilt differences in the mesh model data to obtain the overall tilt deformation of the defect region in the turbine blade includes: Robust plane fitting is performed on the separated defect point cloud and standard region point cloud in the mesh model data to obtain the defect normal vector of the defect point cloud and the standard normal vector of the standard region point cloud. The overall tilting deformation of the defect region in the turbine blade is calculated based on the defect normal vector and the standard normal vector, wherein the formula for calculating the overall tilting deformation is as follows:

[0013] In the formula, This refers to the overall tilt deformation. It is an inverse cosine function. Let be the defect normal vector. Let be the standard normal vector.

[0014] In a preferred embodiment, the step of constructing a multidimensional comparison framework for the turbine blade based on the average deformation level, the local fluctuation distribution characteristics, and the overall tilt deformation, and identifying the deformation mode of the turbine blade based on the multidimensional comparison framework, includes: A multidimensional comparison framework for the turbine blade is constructed based on the average deformation level, the local fluctuation distribution characteristics, and the overall tilting deformation. The deviation is constructed by combining the mean and standard deviation of the tangential angle difference at corresponding points in the spline curve with the overall tilt deformation and the reference angle of the standard normal vector. Calculate the local deformation average index and global deformation index of the corresponding point based on the deviation and the standard deviation; The deformation mode of the turbine blade is identified by using the local deformation average index and the global deformation index as reference quantitative indicators for the multidimensional comparison framework.

[0015] In a preferred embodiment, the formula for calculating the average local deformation index is as follows:

[0016] In the formula, The local deformation average index is... As a weighting factor, This refers to the overall tilt deformation. The maximum deviation angle in the overall tilt deformation; The formula for calculating the global deformation index is as follows:

[0017] In the formula, The global deformation index is... The average value of the tangential angle difference. This refers to the overall tilt deformation. The deviation amount, This represents the maximum deviation angle within the overall tilt deformation. It is the minimum deviation angle in the overall tilt deformation.

[0018] In a preferred embodiment, the cross-validation of the deformation modes to obtain the dominant mechanism of the turbine blade includes: The multi-dimensional quantitative indicators in the deformation mode are quantitatively analyzed to obtain the quantitative indicators of the turbine blade. Cross-validation of the quantitative indicators and the hierarchical consistency score yields the dominant mechanism of the turbine blade.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves accurate preprocessing of point cloud data through feature-adaptive bilateral filtering. This filtering method dynamically adjusts the weight strategy based on the curvature value, noise fraction, and geometric features of the point cloud, combined with an iterative parameter attenuation mechanism, effectively improving the geometric fidelity of the point cloud data. While filtering noise, it strictly preserves key geometric information such as high curvature features and edge features, providing high-quality source and target point cloud data for subsequent defect detection. Accurate mesh model data is constructed through two-level registration. Combined with principal component analysis and two-dimensional point set spline curve fitting, the average deformation level and local fluctuation distribution features can be accurately extracted. Then, by quantifying the macroscopic angle tilt difference, the overall tilt deformation is obtained. The constructed multi-dimensional comparison framework enables accurate identification of deformation patterns. Finally, cross-validation clarifies the defect-dominant mechanism, significantly improving the accuracy of turbine blade defect identification.

[0020] 2. This invention forms a complete solution from point cloud preprocessing to defect quantification and identification. Feature-adaptive bilateral filtering excels in preserving high curvature and edge features, effectively maintaining the smoothness and structural integrity of the point cloud surface and improving the reliability of data preprocessing. Based on a dual-angle representation analysis framework, the synergistic effect of projection fitting analysis and normal vector angle analysis achieves multi-dimensional and accurate quantification of blade deformation features, providing objective quantitative indicators for determining defect severity. Furthermore, the entire method can complete turbine blade defect detection without disassembly, meeting online inspection requirements, providing key technical support for preventive blade maintenance, ensuring the operational safety and reliability of the turbine system, and offering advantages in computational efficiency, improving the overall efficiency of the inspection process and providing feasibility for batch inspection applications in industrial scenarios. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a method for segmenting defect point clouds in turbine blades according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating defect identification in a defect point cloud segmentation method for turbine blades, provided as an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0023] This application provides a method for defect point cloud segmentation of turbine blades. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for defect point cloud segmentation of turbine blades can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0024] Reference Figure 1 The diagram shown is a flowchart illustrating a method for segmenting defect point clouds of turbine blades according to an embodiment of the present invention. In this embodiment, the method for segmenting defect point clouds of turbine blades includes: S1. Perform feature-adaptive bilateral filtering on the point cloud data of the turbine blade to obtain the source and target point cloud of the turbine blade; In this embodiment of the invention, the step of performing feature-adaptive bilateral filtering on the point cloud data of the turbine blade to obtain the source and target point cloud of the turbine blade includes: The points in the point cloud data are mapped to feature points, high curvature points, ordinary points and noise points to obtain the curvature value, noise fraction, curvature and noise threshold of the turbine blade. The adaptive weights of the turbine blades are obtained by adjusting the weight strategy of the bilateral filter based on the curvature value, the noise score, the curvature, and the noise threshold. The source and target point clouds of the turbine blades are obtained by bilateral filtering of the point cloud data based on the adaptive weights.

[0025] Specifically, point cloud noise is a key factor restricting the accuracy of fan blade detection. Industrial inspection's high sensitivity to geometric features (especially hard edges) requires point cloud processing to effectively denoise while strictly maintaining feature positions and sharpness. In 3D scenes without prior information, the core requirements include: neighbor noise removal, edge sharpness maintenance, and minimization of overall and internal point displacement. Traditional filtering methods easily lead to feature blurring or geometric distortion, while bilateral filtering, with its edge-aware characteristics, can suppress noise while maintaining edge sharpness and constraining point displacement.

[0026] Furthermore, voxel downsampling is used to balance point cloud density and preserve key geometric features. Combining the dimensions of normal variation and geometric morphology, a weighted combination method is employed to calculate point curvature, ensuring stable geometric feature recognition even with errors in the normal vector. Normal curvature is calculated through the angular variation of neighborhood normal vectors, while geometric curvature is obtained through principal component analysis of local point sets.

[0027] Furthermore, based on robust curvature information, we set the probability distribution characteristics of three key geometric descriptors—linearity, planarity, and sphericity—in the point cloud in space.

[0028] Furthermore, these geometric invariants are calculated directly from the spatial distribution of the points, without relying on pre-calculated normal vectors, thus fundamentally avoiding the cyclic dependency problem between normal vectors and point positions in traditional bilateral filtering. These feature points exhibit a clear multi-peak distribution characteristic, with each distribution peak corresponding to the statistical characteristics of a specific geometric structure type.

[0029] Furthermore, a classification boundary threshold set based on the feature distribution pattern discretizes the continuous 3D geometric descriptor into distinct feature categories. Each point represents the projection position and classification characteristics of the geometric feature vector of a single point in the point cloud onto the normalized feature space: red represents linear features, green corresponds to planar features, purple indicates corner features, and light blue represents volume features (remaining region). The dashed boundary lines L=0.65, P=0.35, P=0.6, and S=0.25 constitute the optimal classification hyperplane based on statistical learning. The core advantage of this geometric descriptor lies in its complete independence from the calculation accuracy of the normal vector, directly extracting essential geometric attributes from the spatial distribution of points through eigenvalue decomposition.

[0030] Furthermore, complementary feature information is constructed at different spatial resolutions (different neighborhood radii) through a multi-scale analysis strategy.

[0031] Furthermore, the multi-scale statistical graph, through stacked bar data, quantitatively presents the distribution of four types of point cloud features obtained based on geometric invariant classification at three feature analysis scales. The horizontal axis represents the feature scale parameter, and the vertical axis represents the absolute number of points for each feature type. The significant increase in the proportion of light-colored volumetric features with increasing scale reflects the prominent 3D filling characteristics of point clouds in large neighborhoods; the monotonically decreasing proportion of green planar features with increasing scale reveals the superiority of planar structure detection at fine scales; the steadily increasing proportion of orange linear features reflects the cumulative effect of edge features in coarse-scale analysis; and the scarcity of purple corner features across all scales, peaking at the medium scale, aligns with the essential attribute of corners as local geometric singularities. This statistical distribution verifies the cross-scale stability of the geometric invariant classification algorithm. The scale dependence of feature proportions provides crucial statistical priors for adaptive neighborhood selection and multi-scale fusion strategies, and provides a quantitative empirical basis for constructing a robust geometric prior knowledge graph and optimizing bilateral filtering parameters.

[0032] Furthermore, the cross-scale consistency heatmap quantifies the geometric feature classification across three spatial scales, with the values ​​in the matrix heatmap representing the correlation coefficients between feature classification results at two scales. Adjacent scales exhibit high consistency, while consistency decreases across the largest scale span. This heatmap reflects a reasonable transition from detailed features to global features, and the numerical level of the consistency coefficients indicates that the feature detection algorithm based on geometric invariants maintains good stability in multi-scale analysis. This progressive consistency pattern provides theoretical support for multi-scale parameter fusion and verifies the effectiveness of the algorithm in constructing stable geometric prior knowledge.

[0033] Furthermore, by integrating the analysis results at multiple scales, a stable and reliable feature recognition strategy is established: when a feature is consistently recognized at multiple scales, its geometric importance is reinforced; when a feature exhibits differences at different scales, independent feature analysis is performed at multiple scales to reduce local misjudgments that may be introduced by single-scale analysis, thereby providing more reliable prior knowledge for the subsequent adaptive filtering process.

[0034] In this embodiment of the invention, the step of performing bilateral filtering on the point cloud data based on the adaptive weights to obtain the source and target point cloud of the turbine blade includes: Based on the average displacement of the point cloud data, iterative parameter attenuation is performed on the point cloud data to obtain the source target point cloud of the turbine blade. The formula for calculating the average displacement is as follows:

[0035] In the formula, The average displacement is... The total number of points in the point cloud data. The point cloud ordinal number of the point cloud data. For point clouds In the The position of the next iteration. For point clouds In the The position of the next iteration.

[0036] Furthermore, the iterative optimization of adaptive weight calculation is a core technical step in solving the traditional bilateral filtering normal vector-position cyclic dependency problem. Based on previous feature detection results, each point in the point cloud is mapped to four geometric categories using pre-calculated curvature information and a multi-scale feature detection function: feature point, high curvature point, ordinary point, and noise point.

[0037] Furthermore, feature-adaptive bilateral filtering constructs a multi-dimensional enhanced weight calculation framework. Through a hierarchical parameter adjustment strategy, it optimizes the spatial and normal weight parameters in traditional filtering, while introducing feature-aware weights: geometric feature weights and curvature-guided weights. An adaptive weight adjustment strategy is employed for different types of points.

[0038] Furthermore, the parameter pairs in the parameter adjustment strategy are adaptively selected based on the local density features and geometric classification of the points. This avoids the over-smoothing or under-smoothing problems that can occur when optimizing point clouds with different geometric characteristics using fixed parameter settings.

[0039] Furthermore, as the most important component of the geometric structure, the weight adjustment of feature points follows an extreme protection principle. A feature enhancement factor greatly strengthens the weight of the feature center point; a similarity threshold is introduced, calculating the angular difference distribution between the k-nearest neighbor normal vectors of each point and the center point's normal vector, and using its 25th percentile as the similarity judgment threshold. This ensures that only the quarter-nearest neighbor points closest to the center point's normal vector are considered geometrically similar and maintain a high similarity neighbor weight; for other cases, a neighbor weight decay factor suppresses the influence of dissimilar neighbor points. This asymmetric weight allocation mechanism ensures that feature points retain their original positions and geometric features more during the filtering process, rather than being affected by the averaging effect of neighboring points.

[0040] Furthermore, the curvature-guided weights dynamically adjust the filtering intensity based on the complexity of the local geometry, reducing the smoothing effect in high curvature regions and enhancing the smoothing effect in low curvature regions, thereby achieving adaptive protection of geometric features.

[0041] Furthermore, an iterative parameter decay strategy based on actual displacement changes is introduced, and the convergence state is determined through dynamic parameter adjustment. Specifically, as the number of iterations increases, the filter intensity decays exponentially: the algorithm determines convergence and stops iterating.

[0042] Furthermore, local geometric distortions are prevented by limiting the maximum displacement amplitude of individual points. When the calculated displacement exceeds a set maximum allowable displacement threshold, the algorithm normalizes the displacement vector and then multiplies it by the maximum allowable displacement value. This ensures that even if cumulative errors occur during local filtering, the algorithm can still maintain the overall geometric integrity of the point cloud.

[0043] Furthermore, through this series of weight designs and improved calculation formulas, accurate identification and differentiated processing of different geometric features were successfully achieved. The processing results of the original point cloud, the traditional bilateral filtering method, and the improved algorithm are presented through a comparison of front, side, and top views. A unified color mapping based on curvature values ​​is used in the figures, with warm colors representing high-curvature feature regions and cool colors representing low-curvature smooth surfaces.

[0044] In summary, feature-adaptive bilateral filtering of turbine blade point cloud data to obtain source and target point clouds can accurately handle point cloud noise while preserving key geometric features. This filtering method first maps the point cloud into feature points, high-curvature points, ordinary points, and noise points. It then dynamically adjusts the bilateral filtering weight strategy based on parameters such as curvature value and noise fraction to avoid over-smoothing or under-smoothing issues caused by fixed parameters. Simultaneously, an iterative parameter attenuation mechanism based on average displacement ensures stable convergence of the filtering process and reduces point cloud geometric distortion.

[0045] In summary, this process significantly improves the geometric fidelity of point cloud data. While effectively removing noise, it efficiently preserves key structural features such as high curvature regions and edges of the blade, avoiding the feature blurring or curvature information loss problems that are prone to occur in traditional filtering. This lays a high-quality data foundation for subsequent two-level registration between standard point clouds and source target point clouds, grid model data construction, and defect feature extraction, ensuring the accuracy of subsequent defect detection and helping to improve the overall accuracy of turbine blade defect identification. At the same time, it maintains the continuity of the point cloud surface, providing reliable data support for the accurate analysis of blade defects.

[0046] S2. Perform two-level registration on the standard point cloud of the turbine blade and the source target point cloud to obtain the mesh model data of the turbine blade; In this embodiment of the invention, the step of performing two-level registration of the standard point cloud of the turbine blade and the source target point cloud to obtain the mesh model data of the turbine blade includes: Establish the corresponding constraint relationship between the standard point cloud of the turbine blade and the source target point cloud; The final value of the rigid body transformation of the defect point cloud in the corresponding constraint relationship is determined by random sampling; The final value of the rigid body transformation is projected onto Euclidean space to obtain the mesh model data of the turbine blade.

[0047] Specifically, after completing the feature-adaptive bilateral filtering process, the optimized target defect point cloud and source standard point cloud are obtained. To achieve accurate defect detection, the two need to be registered and aligned. This paper adopts a two-level registration strategy: for the pre-processed and filtered standard point cloud and target point cloud, a correspondence constraint is established using feature space; initial values ​​of rigid body transformation in space are solved by random sampling; and the final transformation is obtained in the world coordinate system (in Euclidean space) using iterative nearest points and applied to the defect point cloud to align the mesh model data of defective and non-defective parts.

[0048] Furthermore, in the point cloud registration quality assessment, the accurate division of overlapping and non-overlapping regions is based on the spatial proximity measurement theory between registered point clouds. The accurate segmentation of regions is achieved by constructing a bidirectional nearest neighbor mapping relationship between two registered point clouds.

[0049] Furthermore, within the mathematical framework of bidirectional nearest neighbor mapping, the distance field operator constructs a complete spatial mapping relationship between the target point cloud and the source point cloud.

[0050] Furthermore, the core function of the segmentation operator is to transform continuous geometric distance information into discrete topological region identifiers in order to achieve accurate analysis of complex geometric structures.

[0051] Furthermore, the introduction of this bidirectional verification mechanism eliminates the geometric inconsistencies that may arise from unidirectional mapping, ensuring the topological integrity of overlapping region identification. The setting of the spatial tolerance empirical parameter comprehensively considers point cloud sampling density, registration accuracy, and the scale of geometric deformation features, providing the algorithm with flexibility to adapt to different application scenarios. Through spatial connectivity analysis of the non-overlapping point set, the algorithm achieves geometric clustering of discrete outliers, providing a precise spatial segmentation basis for subsequent quantitative analysis of point cloud angle deviations.

[0052] In summary, two-level registration of the standard point cloud and the source / target point cloud of the turbine blade to obtain mesh model data enables high-precision alignment of the two types of point clouds, providing a precise geometric benchmark for subsequent defect analysis. This registration process first establishes the corresponding constraint relationship between the standard point cloud and the source / target point cloud, determines the final value of the rigid body transformation of the defect point cloud through random sampling, and then projects it onto Euclidean space to generate mesh model data. This effectively avoids the error accumulation problem that easily occurs with single registration, improving the stability and accuracy of point cloud alignment.

[0053] In summary, this process accurately integrates the baseline geometric information of the standard point cloud with the actual morphological data of the source target point cloud. The generated mesh model data clearly presents the overall and local geometry of the blade, especially ensuring the spatial position accuracy of the point cloud in non-overlapping areas. This provides a reliable data carrier for subsequent principal component analysis to extract feature vectors and quantify macroscopic angle tilt differences. Simultaneously, high-quality mesh model data reduces interference factors in subsequent defect feature extraction, ensuring the accuracy of parameter calculations such as average deformation level and local fluctuation distribution characteristics. This lays a solid foundation for constructing a multi-dimensional comparison framework and identifying deformation patterns, ultimately contributing to improved overall accuracy of turbine blade defect detection.

[0054] S3. Perform principal component analysis on the source point cloud and target point cloud of the non-overlapping region in the grid model data to obtain the feature vector of the principal component in the turbine blade. In this embodiment of the invention, the step of performing principal component analysis on the source point cloud and target point cloud of the non-overlapping region in the mesh model data to obtain the feature vector of the principal component in the turbine blade includes: Extract the source point cloud and target point cloud of the non-overlapping region from the mesh model data; Based on the source point cloud and the target point cloud, a reference frame for the turbine blade is established. The covariance matrix of the point cloud in the grid model data is decomposed using the reference framework to obtain the eigenvectors of the turbine blades in the principal component directions.

[0055] Specifically, angular deviations in complex curved workpieces typically exhibit multi-scale, multi-mechanism composite deformation modes, making it difficult for a single geometric description method to fully capture the complete information of their deformation characteristics. After achieving high-precision point cloud registration and separating defect regions, refined identification and quantitative analysis of deformation characterization become crucial technical aspects. Based on this, this section proposes a parallel analysis framework based on dual-angle characterization. This framework utilizes two independent and geometrically complementary computational modules—projection fitting analysis and normal vector angle analysis—to achieve multi-dimensional collaborative characterization of defect deformation.

[0056] Furthermore, the core of point cloud angle deviation calculation lies in establishing a suitable geometric reference frame. For the source and target point clouds from which non-overlapping regions have been extracted, a unified reference coordinate system is established through principal component analysis to eliminate attitude differences between different workpieces (this coordinate system will be used as the benchmark frame for all subsequent geometric analyses). Eigenvalue decomposition is performed on the covariance matrix of the point cloud to obtain eigenvectors for the three principal component directions. These principal component vectors constitute the local coordinate system of the point cloud. Here, represents the main extension direction of the point cloud, and forms the projection plane perpendicular to the main extension direction.

[0057] Furthermore, in order to better characterize the cross-sectional features and curve fitting analysis of point clouds and to compare the angular deviations between different point clouds, this method uses the plane formed by the second principal component and the third principal component of the target point cloud as the projection plane to map the projected three-dimensional points onto the two-dimensional plane.

[0058] Furthermore, for the two-dimensional point set projected onto the reference coordinate system, a cubic B-spline curve is used for fitting. After obtaining the spline curve, the tangent angle at each point on the curve is calculated to quantify the local geometric changes.

[0059] Furthermore, projection fitting analysis is used to calculate the tangent angle deviation between the two spline curves at the same parameter positions, and then normalization is performed. Based on data such as the curvature change and mean value of the angle deviation in local defect deformation, the average level of deformation and the characteristics of local fluctuation distribution are analyzed and evaluated.

[0060] In summary, principal component analysis (PCA) of the source and target point clouds in non-overlapping regions of the mesh model data to obtain principal component eigenvectors can accurately capture the core dimensional features of the blade geometry, establishing a standardized reference framework for subsequent deformation analysis. This process first extracts the point clouds of non-overlapping regions—these regions are key carriers of defect features—and then establishes a blade baseline framework based on them. By decomposing the eigenvalues ​​of the covariance matrix, the eigenvectors of the principal component directions are clarified, effectively removing redundant information from the point cloud data and focusing on the core dimensions reflecting the blade's geometric morphology.

[0061] In summary, the generated principal component eigenvectors clearly characterize the blade's main extension direction and the projection plane perpendicular to that direction. This provides a precise spatial coordinate reference for subsequently mapping the 3D eigenvectors to a 2D plane and constructing a 2D point set, ensuring the fidelity of geometric information during the 2D transformation process. Simultaneously, these eigenvectors provide a unified geometric benchmark for quantifying local fluctuations and overall deformation of the blade, avoiding analytical errors caused by inconsistent reference frames. This ensures the accuracy of subsequent steps such as spline curve fitting and deformation level assessment, providing reliable feature basis for ultimately constructing a multi-dimensional comparison framework and identifying deformation patterns, thus helping to improve the accuracy of defect detection.

[0062] S4. Map the feature vector projected to the three-dimensional space to the two-dimensional plane to obtain the two-dimensional point set of the turbine blade, and fit the spline curve in the two-dimensional point set to obtain the average deformation level and local fluctuation distribution characteristics of the turbine blade. In this embodiment of the invention, the step of mapping the feature vector projected onto a three-dimensional space to a two-dimensional plane to obtain a two-dimensional point set of the turbine blade, and fitting the spline curve of the two-dimensional point set to obtain the average deformation level and local fluctuation distribution characteristics of the turbine blade, includes: Projecting the point cloud in the mesh model data onto the reference frame yields the three-dimensional coordinates of the point cloud in the mesh model data; Using the first principal component and the second principal component in the reference frame as projection planes, the three-dimensional coordinates are mapped to the projection planes to obtain the two-dimensional point set of the point cloud in the mesh model data; The spline curve of the point cloud in the grid model data is obtained by fitting the two-dimensional point set using cubic spline interpolation, and the local fluctuation distribution characteristics of the turbine blade are quantified by the tangent angle of the spline curve. The average deformation level of the turbine blade is evaluated based on the curvature changes and angular deviations of the local defect regions in the spline curve.

[0063] Specifically, the first step is to define the specific structure of the reference frame. This frame is established with the turbine blade's mounting reference plane as a reference and includes three mutually perpendicular coordinate axes (X-axis, Y-axis, and Z-axis) and a fixed origin. The X-axis is parallel to the central axis from the blade root to the blade tip, the Y-axis is perpendicular to the blade's pressure surface, and the Z-axis is perpendicular to the plane formed by the X and Y axes. The position and orientation of the entire frame are fixed, used to unify the coordinate positioning of the point cloud data in the mesh model. All point cloud points are completely extracted from the mesh model data, ensuring coverage of the turbine blade's root, blade body, blade tip, and edge regions, without omitting point cloud information from any critical parts. For each extracted point cloud point, a high-precision distance measuring tool is used to measure the vertical distance from that point to the reference frame's X-axis, Y-axis, and Z-axis. During the measurement process, it is ensured that the measurement direction of the distance measuring tool is strictly parallel to the corresponding coordinate axis to avoid data errors caused by directional deviations. The measured vertical distance to the X-axis is used as the coordinate value of the point cloud point in the X-axis direction of the reference frame, the vertical distance to the Y-axis is used as the coordinate value in the Y-axis direction, and the vertical distance to the Z-axis is used as the coordinate value in the Z-axis direction. Each point cloud point obtains its corresponding three coordinate values ​​in this way. These coordinate values ​​of all point cloud points are arranged in the order of the original point cloud number and together constitute the three-dimensional coordinates of the point cloud in the mesh model data.

[0064] Furthermore, a comprehensive analysis of the 3D coordinates of the point cloud in the mesh model data is performed. The dispersion of all 3D coordinates along the X, Y, and Z axes of the baseline frame is statistically analyzed. The dispersion is determined by calculating the distribution range of coordinate values ​​in each direction (maximum value minus minimum value). The direction with the largest distribution range is the first principal component in the baseline frame, reflecting the most significant distribution trend of the point cloud data. After determining the first principal component, the direction containing the first principal component is excluded, and the dispersion of the remaining two directions (e.g., the directions in the X and Y axes excluding the first principal component and the Z axis) is analyzed. The direction with the second largest distribution range is selected as the second principal component, and the direction of the second principal component must be strictly perpendicular to the direction of the first principal component. The perpendicularity is verified by ensuring that the vector dot product of the two directions is zero. The first and second principal components together constitute the two-dimensional plane used for projection. Each 3D coordinate in the point cloud of the mesh model data is processed individually. First, the direction perpendicular to the projection plane in the reference frame is determined (i.e., the direction of the third principal component, which is also perpendicular to the first and second principal components). Then, the coordinate components of the 3D coordinates in the direction of the third principal component are removed. Specifically, the coordinate values ​​corresponding to this direction are set to zero, retaining only the coordinate components of the 3D coordinates in the directions of the first and second principal components. The retained coordinate components in the first principal component direction are used as the x-coordinates, and the coordinate components in the second principal component direction are used as the y-coordinates, combining them to form a 2D data set. All 3D coordinates are converted into corresponding 2D data in the same way. These 2D data sets are arranged in the order of the original point cloud, and the resulting set is the 2D point set of the point cloud in the mesh model data.

[0065] Furthermore, the two-dimensional point set is first observed as a whole to determine its distribution range and trend. From the beginning of the point set (corresponding to the point cloud at the root of the turbine blade) to the end (corresponding to the point cloud at the tip of the turbine blade), points that completely cover the beginning, middle, and end of the point set are selected as the base points for fitting, at a frequency of one point every five original points. Simultaneously, it is ensured that the selected base points include all obvious inflection points in the point set (i.e., locations where the distribution direction of the points changes). The number of base points must meet the requirements of subsequent cubic spline interpolation fitting; typically, 20-30 base points are selected. The order of the cubic spline interpolation is determined based on the number of selected base points, following the principle that the order should be 7 less than the number of base points (e.g., 20 base points correspond to a 3rd-order curve, 30 base points correspond to a 5th-order curve). The determination of the order must ensure that the curve can smoothly transition while accurately matching the distribution pattern of the base points. In a two-dimensional plane, control vertices for cubic spline interpolation are set, with the number of control vertices matching the curve order (e.g., 4 control vertices for a 3rd-order curve, 6 control vertices for a 5th-order curve). The control vertices are initially distributed along the trend line formed by the base points. Then, the positions of the control vertices are adjusted one by one, observing the fit between the curve and the base points after each adjustment. If the curve deviates from a base point by more than 0.1 mm, the control vertices near that base point are fine-tuned until the maximum deviation of the curve from all base points does not exceed 0.1 mm. The resulting curve is the spline curve of the point cloud in the mesh model data. On the spline curve, sampling points are evenly selected from the start to the end of the curve, with one sampling point per millimeter. For each sampling point, a two-dimensional coordinate system protractor is used to determine the direction of the tangent to the curve at that point. The center of the protractor coincides with the sampling point, and the 0-degree mark of the protractor is aligned with the positive X-axis of the two-dimensional plane. The angle between the tangent direction and the 0-degree mark is read; this angle is the tangent angle at that sampling point. Arrange the tangent angles of all sampling points in the sampling order and calculate the difference between the tangent angles of adjacent sampling points. If the difference is within 5°, it indicates that the corresponding turbine blade surface is relatively smooth; if the difference exceeds 10°, it indicates that there are obvious undulations at the corresponding position. By statistically analyzing the number and distribution of sampling points with differences in different ranges, the amplitude and frequency of angle changes are analyzed, thereby realizing the quantification of the local fluctuation distribution characteristics of turbine blades using the tangent angle of spline curves.

[0066] Furthermore, the spline curve is first compared with a pre-established standard spline curve for turbine blades. Segments on the curve that deviate from the standard curve by more than 0.3 mm are identified as local defect areas. Simultaneously, the starting point (where the defect begins to deviate from the standard curve) and ending point (where the defect returns to the standard curve) of each local defect area are marked on the spline curve, ensuring that the marked starting and ending points accurately correspond to the defect boundaries. Within each local defect area, measurement points are selected at intervals of 0.2 mm, sequentially from the starting point to the ending point. For each measurement point, a ruler is placed against the curve near that point, and the degree of deviation between the curve and the ruler's edge is observed. A large deviation indicates a large curvature at that point, while a small deviation indicates a small curvature. The curvature of each measurement point is determined by comparing it with a standard curvature comparison card (marking the distance of the curve from the ruler corresponding to different curvatures). Record the curvature values ​​of all measurement points within the same local defect area, identify the maximum and minimum curvature values, and subtract the minimum curvature value from the maximum curvature value to obtain the range of curvature variation within the area. By comparing the increase and decrease of curvature values ​​at different measurement points, determine whether the curvature gradually increases, gradually decreases, or fluctuates from the starting point to the ending point, thus gaining a complete understanding of the curvature changes within the area. Use the average tangent angle of the first 10 sampling points before the starting point of the local defect area as the starting reference angle, and the average tangent angle of the last 10 sampling points as the ending reference angle. For each sampling point within the area, subtract the corresponding reference angle from its tangent angle (starting reference angle for the initial segment, ending reference angle for the ending segment, and the average of the two reference angles for the middle segment as the comparison reference). The calculated difference is the angle deviation of that sampling point. Record the angle deviation value for each sampling point, and count the number of sampling points with positive and negative deviations, as well as the number of sampling points with an absolute deviation exceeding 5°, to obtain the overall distribution of angle deviations within the area. The curvature variation range values ​​of all local defect areas are collected. These values ​​are summed and divided by the total number of local defect areas to obtain the average curvature variation range across all local defect areas. Simultaneously, the angular deviation values ​​of all sampling points within each local defect area are collected. The absolute values ​​of these values ​​are summed and divided by the total number of sampling points to obtain the average angular deviation across all local defect areas. If the average curvature variation range is less than 2 and the average angular deviation is less than 3°, the average deformation level of the turbine blade is considered low. If the average curvature variation range is between 2 and 5 and the average angular deviation is between 3° and 8°, the average deformation level is considered medium. If the average curvature variation range is greater than 5 and the average angular deviation is greater than 8°, the average deformation level is considered high. Through this comprehensive assessment, the final average deformation level of the turbine blade is determined.

[0067] In summary, mapping feature vectors projected into three-dimensional space onto a two-dimensional plane to obtain a two-dimensional point set for turbine blades, and then fitting spline curves to this point set, can efficiently transform complex three-dimensional geometric information and accurately extract blade deformation features. The mapping process relies on the baseline framework obtained from principal component analysis, selecting the first and second principal components to construct the projection plane. This ensures that the two-dimensional point set accurately retains key geometric information from the three-dimensional feature vectors, avoiding feature loss due to dimensional transformation and laying a reliable data foundation for subsequent analysis.

[0068] In summary, spline curve fitting can accurately characterize the geometry of a two-dimensional point set. The tangent angle of the curve can quantify the local fluctuation distribution characteristics of the blade. Combined with the curvature changes and angular deviations in local defect areas, the average deformation level can be objectively assessed. This process achieves the quantification and visualization of blade deformation characteristics, capturing both subtle local fluctuations and clarifying the overall deformation trend. It provides precise deformation feature data support for subsequent quantification of macroscopic tilt differences and the construction of a multi-dimensional comparison framework, effectively improving the accuracy of deformation pattern recognition, assisting in the subsequent determination of the defect-dominant mechanism, and ensuring the accuracy of turbine blade defect detection.

[0069] S5. Quantify the macroscopic angle tilt difference in the mesh model data to obtain the overall tilt deformation of the defect area in the turbine blade; In this embodiment of the invention, quantifying the macroscopic angular tilt differences in the mesh model data to obtain the overall tilt deformation of the defect region in the turbine blade includes: Robust plane fitting is performed on the separated defect point cloud and standard region point cloud in the mesh model data to obtain the defect normal vector of the defect point cloud and the standard normal vector of the standard region point cloud. The overall tilting deformation of the defect region in the turbine blade is calculated based on the defect normal vector and the standard normal vector, wherein the formula for calculating the overall tilting deformation is as follows:

[0070] In the formula, This refers to the overall tilt deformation. It is an inverse cosine function. Let be the defect normal vector. Let be the standard normal vector.

[0071] Specifically, the core task of the normal vector angle characterization module is to extract the overall geometric tendency of the defect region and the standard region, and quantify the macroscopic angular tilt differences between the regions. For the separated defect point cloud and standard region point cloud, robust plane fitting is performed to obtain their respective principal plane unit normal vectors.

[0072] Furthermore, absolute value operations are used to eliminate the uncertainty of the normal vector direction, ensuring the uniqueness and comparability of angle measurements. The rotation angle of the principal plane of the defect area relative to the principal plane of the standard area is calculated using the vector angle formula. This rotation angle corresponds to the overall tilt deformation of the workpiece in that area.

[0073] Furthermore, from the perspective of deformation mechanism, the overall tilt deformation mainly captures the average effect of rigid body rotation and large-scale bending deformation, providing a macroscopic characteristic description of the geometric deviation of the defect area, and forming cross-validation with the local bending analysis results of projection fitting from a global geometric perspective.

[0074] In summary, quantifying the macroscopic tilt differences in the mesh model data to obtain the overall tilt deformation of the turbine blade defect region can accurately capture the macroscopic geometric deviation of the defect region, providing key quantitative indicators for defect assessment. This process first performs robust plane fitting on the separated defect point cloud and standard region point cloud to obtain the defect normal vector and standard normal vector respectively. Then, the angle between the two is calculated using the vector angle formula to eliminate uncertainty in the normal vector direction, ensuring that the measurement results are unique and comparable.

[0075] In summary, the obtained overall tilt deformation objectively reflects the macroscopic tilt degree of the defect area relative to the standard area, accurately characterizes overall deformation effects such as rigid body rotation and large-scale bending, and supplements macroscopic deformation information not covered by local fluctuation features. This indicator complements the average deformation level and local fluctuation distribution characteristics, providing core macroscopic parameter support for constructing a multidimensional comparison framework. It avoids misjudgment of overall deformation caused by analysis of single local features, improves the comprehensiveness and accuracy of deformation pattern recognition, and provides reliable macroscopic quantitative basis for subsequent cross-validation to clarify the defect-dominant mechanism and determine the severity of defects, thus helping to improve the accuracy of turbine blade defect detection.

[0076] S6. Construct a multi-dimensional comparison framework for the turbine blade based on the average deformation level, the local fluctuation distribution characteristics, and the overall tilt deformation amount, and identify the deformation mode of the turbine blade based on the multi-dimensional comparison framework. In this embodiment of the invention, the step of constructing a multidimensional comparison framework for the turbine blade based on the average deformation level, the local fluctuation distribution characteristics, and the overall tilt deformation, and identifying the deformation mode of the turbine blade based on the multidimensional comparison framework, includes: A multidimensional comparison framework for the turbine blade is constructed based on the average deformation level, the local fluctuation distribution characteristics, and the overall tilting deformation. The deviation is constructed by combining the mean and standard deviation of the tangential angle difference at corresponding points in the spline curve with the overall tilt deformation and the reference angle of the standard normal vector. Calculate the local deformation average index and global deformation index of the corresponding point based on the deviation and the standard deviation; The deformation mode of the turbine blade is identified by using the local deformation average index and the global deformation index as reference quantitative indicators for the multidimensional comparison framework.

[0077] In this embodiment of the invention, the formula for calculating the average local deformation index is as follows:

[0078] In the formula, The local deformation average index is... As a weighting factor, This refers to the overall tilt deformation. The maximum deviation angle in the overall tilt deformation; The formula for calculating the global deformation index is as follows:

[0079] In the formula, The global deformation index is... The average value of the tangential angle difference. This refers to the overall tilt deformation. The deviation amount, This represents the maximum deviation angle within the overall tilt deformation. It is the minimum deviation angle in the overall tilt deformation.

[0080] Specifically, we first clarify the average deformation level of the turbine blade (i.e., the average deformation value of each monitoring point of the blade), the local fluctuation distribution characteristics (i.e., the discrete distribution of the deformation value of each monitoring point relative to the average deformation level), and the overall tilt deformation (i.e., the quantitative result of the tilt direction and tilt degree of the blade in space). We take these three elements as the core dimensions, determine the specific description content under each dimension, and ensure that each dimension accurately reflects the corresponding deformation information, thereby constructing a multi-dimensional comparison framework for turbine blades.

[0081] Furthermore, the spline curve of the turbine blade after deformation is obtained, and the corresponding points on the curve are found. The difference in tangential angle between each corresponding point and the same position point of the undeformed standard spline curve is calculated. The average value of the tangential angle difference is obtained by summing all the tangential angle differences and dividing by the total number of corresponding points. Then, the standard deviation of the tangential angle difference is obtained by taking the square root of the sum of the squares of the differences between each tangential angle difference and the average value, dividing by the total number of corresponding points. At the same time, the specific value of the overall tilt deformation and the reference angle of the standard normal vector are determined (the standard normal vector is the normal vector at the corresponding position on the surface of the undeformed blade, and its reference angle is the angle between the normal vector and the preset reference direction). These four elements are integrated according to the set logic and included in the calculation to construct the deviation.

[0082] Furthermore, the constructed deviation and standard deviation of tangential angle difference are obtained. For a single corresponding point, the deviation of the point is correlated with the standard deviation of tangential angle difference according to the set numerical combination method to obtain the local deformation average index of the corresponding point. For all corresponding points, the deviations of all points are added together and then divided by the total number of corresponding points to obtain the mean deviation. Then, the mean deviation is correlated with the standard deviation of tangential angle difference according to the set numerical combination method to obtain the global deformation index.

[0083] Furthermore, the local deformation average index and the global deformation index are used as reference quantitative indicators for the multidimensional comparison framework. The correspondence between these two indicators and each dimension of the framework is clarified. By observing the distribution of the local deformation average index at corresponding points (e.g., the positions of high and low values) and the specific values ​​of the global deformation index, combined with the characteristics of each dimension of the multidimensional comparison framework, it is determined whether the blade exhibits local bulges, local depressions, or overall bending, thereby identifying the deformation mode of the turbine blade. In summary, constructing a multidimensional comparison framework for turbine blades based on the average deformation level, local fluctuation distribution characteristics, and overall tilt deformation, and identifying deformation modes accordingly, enables multidimensional collaborative analysis of blade deformation characteristics, improving the accuracy and comprehensiveness of deformation mode identification. This framework integrates key information from subtle local deformations (local fluctuations, average deformation level) and macroscopic overall deformation (overall tilt deformation), forming a complementary and systematic feature analysis system, avoiding the loss of deformation information or misjudgment caused by single-dimensional feature analysis.

[0084] Specifically, the overall tilt deformation is derived from the quantitative result obtained by measuring the tilt direction and degree of the turbine blade as a whole in space. This result is the formula for calculating the average local deformation index. .

[0085] Furthermore, the maximum deviation angle is derived by first calculating the difference in tangential angles between corresponding points on the spline curve after turbine blade deformation and the same points on the undeformed standard spline curve, and then selecting the largest value from all tangential angle differences. This value is the value in the formula for calculating the local deformation average index. And in the global deformation index calculation formula .

[0086] Furthermore, the weighting factor is a fixed value pre-set based on the turbine blade's application scenario, material properties, or specific deformation monitoring requirements. This value is the value in the formula for calculating the local deformation average index. .

[0087] Furthermore, the source of the deviation is the result obtained by integrating and calculating the mean difference of the tangential angle and the standard deviation of the tangential angle difference between corresponding points of the spline curve, together with the overall tilt deformation and the reference angle of the standard normal vector according to the set logic. This result is σ in the global deformation index calculation formula.

[0088] Furthermore, the mean tangential angle difference is derived by first calculating the tangential angle difference between corresponding points on the spline curve after turbine blade deformation and the same points on the undeformed standard spline curve, then summing all tangential angle differences, and finally dividing the sum by the total number of corresponding points. The final value obtained is the value in the global deformation index calculation formula. .

[0089] Furthermore, the minimum deviation angle is derived by first calculating the difference in tangential angles between corresponding points on the spline curve after turbine blade deformation and the same points on the undeformed standard spline curve, and then selecting the smallest value from all tangential angle differences. This value is the value in the global deformation index calculation formula. .

[0090] Furthermore, the formula for calculating the local deformation average index is obtained by multiplying the weighting factor by the overall tilting deformation, then multiplying (1 minus the weighting factor) by the maximum deviation angle, and finally adding the two products together. This comprehensively considers the overall tilting deformation of the turbine blade and the local maximum deviation angle, and adjusts the degree of influence of the two on the result through the weighting factor. Finally, the local deformation average index that reflects the average level of local deformation at a single corresponding point of the turbine blade is obtained.

[0091] Furthermore, the formula for calculating the global deformation index combines 1.41 with the deviation amount and then divides the resulting product by the average tangential angle difference. This combines the degree of turbine blade deformation deviation reflected by the deviation amount with the average level of tangential angle differences at corresponding points of the turbine blade reflected by the average tangential angle difference, ultimately yielding a global deformation index that reflects the overall deformation degree of the turbine blade.

[0092] Furthermore, in the formula for calculating the average local deformation index, when the weighting factor remains constant, if the overall tilt deformation increases, the average local deformation index will increase accordingly; if the overall tilt deformation decreases, the average local deformation index will decrease accordingly.

[0093] Furthermore, in the formula for calculating the average local deformation index, when the weighting factor remains constant, if the maximum deviation angle increases, the average local deformation index will increase accordingly; if the maximum deviation angle decreases, the average local deformation index will decrease accordingly.

[0094] Furthermore, in the formula for calculating the local deformation average index, when the overall tilt deformation and the maximum deviation angle remain unchanged, if the weighting factor increases, the influence of the overall tilt deformation on the local deformation average index will be enhanced, while the influence of the maximum deviation angle on the local deformation average index will be weakened; if the weighting factor decreases, the influence of the overall tilt deformation on the local deformation average index will be weakened, while the influence of the maximum deviation angle on the local deformation average index will be enhanced.

[0095] Furthermore, in the formula for calculating the global deformation index, when the average difference in tangential angles remains constant, if the deviation increases, the global deformation index will increase accordingly; if the deviation decreases, the global deformation index will decrease accordingly.

[0096] Furthermore, in the formula for calculating the global deformation index, when the deviation remains constant, if the average difference in tangential angles increases, the global deformation index will decrease accordingly; if the average difference in tangential angles decreases, the global deformation index will increase accordingly.

[0097] Furthermore, in the formula for calculating the global deformation index, when the deviation remains constant, if the increase in the maximum deviation angle leads to an increase in the average tangential angle difference, the global deformation index will decrease accordingly; if the decrease in the minimum deviation angle leads to an increase in the average tangential angle difference, the global deformation index will also decrease accordingly; if the decrease in the maximum deviation angle leads to a decrease in the average tangential angle difference, the global deformation index will increase accordingly; if the increase in the minimum deviation angle leads to a decrease in the average tangential angle difference, the global deformation index will also increase accordingly.

[0098] In summary, during the identification process, a deviation is constructed by using the mean and standard deviation of the tangential angle difference and the overall tilt deformation. Combined with a hierarchical consistency score as a quantitative reference indicator, this method can accurately distinguish the geometric differences of different deformation modes and effectively identify various deformation types such as rigid body rotation and local non-uniform deformation. This process provides an objective and quantitative basis for determining deformation modes, replacing traditional subjective judgments based on experience. It ensures the reliability and consistency of the identification results, laying a solid foundation for subsequent cross-validation of deformation modes to clarify the defect-dominant mechanism. This directly contributes to improving the overall accuracy of turbine blade defect detection and analysis, providing scientific support for defect assessment and maintenance decisions.

[0099] S7. Cross-validate the deformation mode to obtain the dominant mechanism of the turbine blade.

[0100] In this embodiment of the invention, the cross-validation of the deformation mode to obtain the dominant mechanism of the turbine blade includes: The multi-dimensional quantitative indicators in the deformation mode are quantitatively analyzed to obtain the quantitative indicators of the turbine blade. Cross-validation of the quantitative indicators and the hierarchical consistency score yields the dominant mechanism of the turbine blade.

[0101] Specifically, based on the aforementioned dual-angle characterization analysis framework, a systematic statistical analysis and quantitative evaluation of the separated defect regions are conducted. For the statistical analysis of various deflections in the self-collected dataset, this process obtains standardized geometric representations of the defect and standard regions through spline curve fitting. Then, key statistical parameters are calculated based on the normalized angle sequence. Finally, objective classification of defect deformation patterns is achieved by combining normal vector angle analysis. This includes extracting fitted lines for different defects, statistically analyzing angle differences, normalizing curvature distribution, and assessing differences in normal vector angle changes.

[0102] Specifically, this study first addresses the cyclic dependency between normal vectors and point positions encountered by traditional bilateral filtering in point cloud processing, proposing a feature-adaptive bilateral filtering algorithm. This algorithm constructs a multi-level technical architecture and employs a robust curvature estimation method to weightedly combine normal curvature with geometric curvature based on principal component analysis, forming a geometric descriptor system independent of the filtering process, fundamentally breaking the dependence on the accuracy of the initial normal vector. The algorithm further introduces a multi-scale feature detection mechanism, identifying key geometric features at different spatial scales through geometric invariants such as linearity, planarity, and sphericity, providing reliable prior knowledge for subsequent adaptive processing.

[0103] Furthermore, the algorithm extends the traditional dual-weight model into an adaptive weight system encompassing spatial and normal weights. Combined with a feature-aware weight adjustment strategy and a dynamic parameter selection mechanism, it achieves differentiated processing for regions with different geometric characteristics. By employing a hierarchical processing strategy, the point cloud is divided into four layers: feature points, high-curvature points, ordinary points, and noise points. Each layer uses specially optimized filtering strength and parameter configurations, and stable convergence control is achieved through iterative parameter decay and average displacement monitoring. Quantitative evaluation results fully validate the algorithm's effectiveness. Compared with traditional methods, it achieves a significant improvement in geometric fidelity, with improvements exceeding 13% in both root mean square error and mean absolute error, and further optimization of the signal-to-noise ratio. It exhibits a particularly significant advantage in feature preservation, achieving performance improvements of nearly 90% and 21.00% in high-curvature and edge feature preservation, respectively, directly verifying the algorithm's breakthrough progress in solving the problem of excessive smoothing of geometric features in traditional methods. Simultaneously, while maintaining a steady improvement in surface smoothness and edge preservation capabilities, computational efficiency is also significantly improved.

[0104] Furthermore, supported by high-quality point cloud processing technology, this study established a method for analyzing defects in complex curved surface workpieces with dual-angle representation, and constructed a preventive detection method system based on feature extraction. Through high-precision point cloud registration and defect region separation, the three-dimensional geometric deformation problem is transformed into a multi-scale angular difference representation and pattern recognition problem under a unified reference coordinate system. This method establishes a standardized geometric reference frame through principal component analysis, combines cubic spline curve fitting to achieve precise quantification of local tangent angles, and employs plane fitting to extract the normal vector angle representation of the overall geometric tendency, forming a parallel and complementary architecture of projection fitting analysis and normal vector angle analysis. Based on the cross-analysis mechanism of absolute difference index and relative fluctuation index, quantitative analysis of blade defect characteristics is achieved, promoting the transformation from empirical judgment to objective quantitative analysis.

[0105] Furthermore, statistical analysis and result evaluation verified the effectiveness and reliability of this technical system in identifying defect deformation mechanisms, providing a systematic technical solution for quality control and process optimization in complex manufacturing environments. This technical system can achieve early warning of potential bending risks during routine preventative maintenance without disassembly, providing key quantitative indicators and early warning capabilities for preventative maintenance of blade damage.

[0106] In summary, cross-validating turbine blade deformation modes to determine the dominant defect mechanism enables precise tracing of the deformation's essence, providing a core basis for defect analysis and maintenance decisions. This process involves quantitative analysis of multi-dimensional quantitative indicators in the deformation modes (such as average deformation level, local fluctuation distribution characteristics, overall tilt deformation, and hierarchical consistency score), combining the consistency of dual-perspective characterization with the reliability of each analysis result to form a multi-dimensional cross-validation system. This effectively avoids potential biases or misjudgments that may arise from single-indicator analysis.

[0107] In summary, cross-validation can accurately identify the dominant mechanism of defect deformation, clarifying whether the deformation is caused by factors such as overall rigid rotation, local non-uniform distortion, or multi-physics coupling, providing quantitative support for objectively determining the severity of defects. This result not only verifies the accuracy of deformation pattern recognition but also provides targeted directions for subsequent quality control, process optimization, and preventative maintenance—for example, developing differentiated maintenance strategies for different dominant mechanisms to avoid blind maintenance. Simultaneously, this process further solidifies the reliability of the entire defect detection workflow, ensuring consistent results across all stages from point cloud processing to deformation analysis, directly improving the scientific rigor and practicality of turbine blade defect detection, and guaranteeing the operational safety and reliability of the turbine system.

[0108] In summary, this invention constructs a complete process for point cloud bending defect analysis, covering five core aspects such as... Figure 2 A schematic diagram of a defect identification method for turbine blade defect point cloud segmentation is shown: First, raw point cloud data is acquired through data collection to lay the foundation for subsequent analysis; then, the point cloud processing stage is entered, where multi-scale feature detection and point cloud segmentation systems are used to enhance edge preservation, and feature-adaptive bilateral filters are combined to optimize data quality; subsequently, two-level registration is adopted, first coarse registration and then fine registration, to establish a spatially aligned registered point cloud; the defect detection stage is based on point-pair spatial characteristic analysis, defining classification criteria to separate overlapping points, and using the geometric anomalies of bending defects to identify and locate defects; finally, the analytical analysis and calculation stage is entered, where the projection fitting module constructs a reference frame, fits the geometric contour and calculates the tangent angle, the bending region extraction module processes surface defects, and the normal vector angle module extracts normalized normal vectors and calculates flow direction angles through plane segmentation. By statistically analyzing the average rotation angle, angle deviation and other results, multi-scale bending characteristic evaluation and analysis are achieved.

[0109] In the several embodiments provided by this invention, it should be understood that the disclosed method can be implemented in other ways.

[0110] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0111] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, and technology that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for defect point cloud segmentation for turbine blades, characterized in that, The method includes: S1. Perform feature-adaptive bilateral filtering on the point cloud data of the turbine blade to obtain the source and target point cloud of the turbine blade; S2. Perform two-level registration on the standard point cloud of the turbine blade and the source target point cloud to obtain the mesh model data of the turbine blade; S3. Perform principal component analysis on the source point cloud and target point cloud of the non-overlapping region in the grid model data to obtain the feature vector of the principal component in the turbine blade. S4. Map the feature vector projected to the three-dimensional space to the two-dimensional plane to obtain the two-dimensional point set of the turbine blade, and fit the spline curve in the two-dimensional point set to obtain the average deformation level and local fluctuation distribution characteristics of the turbine blade. S5. Quantify the macroscopic angle tilt difference in the mesh model data to obtain the overall tilt deformation of the defect area in the turbine blade; S6. Construct a multi-dimensional comparison framework for the turbine blade based on the average deformation level, the local fluctuation distribution characteristics, and the overall tilt deformation amount, and identify the deformation mode of the turbine blade based on the multi-dimensional comparison framework. S7. Cross-validate the deformation mode to obtain the dominant mechanism of the turbine blade.

2. The defect point cloud segmentation method for turbine blades as described in claim 1, characterized in that, The step of performing feature-adaptive bilateral filtering on the point cloud data of the turbine blades to obtain the source and target point clouds of the turbine blades includes: The points in the point cloud data are mapped to feature points, high curvature points, ordinary points and noise points to obtain the curvature value, noise fraction, curvature and noise threshold of the turbine blade. The adaptive weights of the turbine blades are obtained by adjusting the weight strategy of the bilateral filter based on the curvature value, the noise score, the curvature, and the noise threshold. The source and target point clouds of the turbine blades are obtained by bilateral filtering of the point cloud data based on the adaptive weights.

3. The defect point cloud segmentation method for turbine blades as described in claim 2, characterized in that, The step of performing bilateral filtering on the point cloud data based on the adaptive weights to obtain the source and target point cloud of the turbine blade includes: Based on the average displacement of the point cloud data, iterative parameter attenuation is performed on the point cloud data to obtain the source target point cloud of the turbine blade. The formula for calculating the average displacement is as follows: In the formula, The average displacement is... The total number of points in the point cloud data. The point cloud ordinal number of the point cloud data. For point clouds In the The position of the next iteration. For point clouds In the The position of the next iteration.

4. The defect point cloud segmentation method for turbine blades as described in claim 1, characterized in that, The two-level registration of the standard point cloud of the turbine blade and the source / target point cloud to obtain the mesh model data of the turbine blade includes: Establish the corresponding constraint relationship between the standard point cloud of the turbine blade and the source target point cloud; The final value of the rigid body transformation of the defect point cloud in the corresponding constraint relationship is determined by random sampling; The final value of the rigid body transformation is projected onto Euclidean space to obtain the mesh model data of the turbine blade.

5. The defect point cloud segmentation method for turbine blades as described in claim 1, characterized in that, The process of performing principal component analysis on the source point cloud and target point cloud of the non-overlapping region in the mesh model data to obtain the feature vectors of the principal components in the turbine blade includes: Extract the source point cloud and target point cloud of the non-overlapping region from the mesh model data; Based on the source point cloud and the target point cloud, a reference frame for the turbine blade is established. The covariance matrix of the point cloud in the grid model data is decomposed using the reference framework to obtain the eigenvectors of the turbine blades in the principal component directions.

6. The defect point cloud segmentation method for turbine blades as described in claim 5, characterized in that, The process of mapping the feature vector projected into three-dimensional space to a two-dimensional plane to obtain a two-dimensional point set of the turbine blade, and fitting the spline curve of the two-dimensional point set to obtain the average deformation level and local fluctuation distribution characteristics of the turbine blade, includes: Projecting the point cloud in the mesh model data onto the reference frame yields the three-dimensional coordinates of the point cloud in the mesh model data; Using the first principal component and the second principal component in the reference frame as projection planes, the three-dimensional coordinates are mapped to the projection planes to obtain the two-dimensional point set of the point cloud in the mesh model data; The spline curve of the point cloud in the grid model data is obtained by fitting the two-dimensional point set using cubic spline interpolation, and the local fluctuation distribution characteristics of the turbine blade are quantified by the tangent angle of the spline curve. The average deformation level of the turbine blade is evaluated based on the curvature changes and angular deviations of the local defect regions in the spline curve.

7. The defect point cloud segmentation method for turbine blades as described in claim 1, characterized in that, The quantification of macroscopic angular tilt differences in the mesh model data yields the overall tilt deformation of the defect region in the turbine blade, including: Robust plane fitting is performed on the separated defect point cloud and standard region point cloud in the mesh model data to obtain the defect normal vector of the defect point cloud and the standard normal vector of the standard region point cloud. The overall tilting deformation of the defect region in the turbine blade is calculated based on the defect normal vector and the standard normal vector, wherein the formula for calculating the overall tilting deformation is as follows: In the formula, This refers to the overall tilt deformation. It is an inverse cosine function. Let be the defect normal vector. Let be the standard normal vector.

8. The defect point cloud segmentation method for turbine blades as described in claim 7, characterized in that, The process of constructing a multidimensional comparison framework for the turbine blade based on the average deformation level, the local fluctuation distribution characteristics, and the overall tilt deformation, and identifying the deformation mode of the turbine blade based on the multidimensional comparison framework, includes: A multidimensional comparison framework for the turbine blade is constructed based on the average deformation level, the local fluctuation distribution characteristics, and the overall tilting deformation. The deviation is constructed by combining the mean and standard deviation of the tangential angle difference at corresponding points in the spline curve with the overall tilt deformation and the reference angle of the standard normal vector. Calculate the local deformation average index and global deformation index of the corresponding point based on the deviation and the standard deviation; The deformation mode of the turbine blade is identified by using the local deformation average index and the global deformation index as reference quantitative indicators for the multidimensional comparison framework.

9. The defect point cloud segmentation method for turbine blades as described in claim 8, characterized in that, The formula for calculating the average local deformation index is as follows: In the formula, The local deformation average index is... As a weighting factor, This refers to the overall tilt deformation. The maximum deviation angle in the overall tilt deformation; The formula for calculating the global deformation index is as follows: In the formula, The global deformation index is... The average value of the tangential angle difference. This refers to the overall tilt deformation. The deviation amount, This represents the maximum deviation angle within the overall tilt deformation. It is the minimum deviation angle in the overall tilt deformation.

10. The defect point cloud segmentation method for turbine blades as described in claim 9, characterized in that, The cross-validation of the deformation modes to obtain the dominant mechanism of the turbine blade includes: The multi-dimensional quantitative indicators in the deformation mode are quantitatively analyzed to obtain the quantitative indicators of the turbine blade. Cross-validation of the quantitative indicators and the hierarchical consistency score yields the dominant mechanism of the turbine blade.

Images