Industrial structure surface optical data processing method, electronic device and storage medium

Abstract

This invention provides a method for processing optical data on the surface of industrial structural components, comprising: fusing detection data from at least one optical detection data source on the surface of the industrial structural component to obtain its point-level feature vector; performing denoising processing based on the abnormal risk assessment value of the point-level feature vector in a risk assessment model to obtain first denoised detection data; performing suppression processing on the spectral domain components of the risk frequency band data in the spectral domain of the first denoised detection data to obtain second detection data; generating the confidence level of the industrial structural component based on the abnormal risk assessment value of the detection points in the second and second detection data, and performing weighted filtering sampling based on the confidence level of the structural component to obtain third detection data; performing local response field estimation, defect region registration, and three-dimensional reconstruction based on the third detection data to obtain reconstructed data of the structural component, and extracting structural component feature data from the reconstructed data to generate structural component defect analysis results.

Description

Technical Field

[0001] This invention relates to the field of detection technology, and in particular to a method and apparatus for processing optical data on the surface of industrial structural components, an electronic device, and a storage medium. Background Technology

[0002] With the widespread application of laser scanning, structured light inspection, and multi-angle optical inspection in industrial quality inspection, acquiring spatial morphology and optical response information of structural component surfaces using non-contact methods has become an important means of analyzing surface defects such as cracks, holes, corrosion, scratches, and coating peeling. This type of inspection typically collects data such as the spatial coordinates of the inspection point, reflection intensity, polarization response, and transmission attenuation. Based on local response changes, it identifies abnormal areas on the structural component surface, thereby pinpointing the defect location and determining its extent.

[0003] Existing detection methods mostly employ fixed threshold filtering, outlier removal, point cloud smoothing, or 3D comparison based on a single reconstruction result to process the acquired data. The basic assumption of these methods is that local reflection changes, point offsets, and geometric discontinuities can directly correspond to the actual surface condition, thus enabling basic defect identification in regular surfaces or detection scenarios with weak optical interference.

[0004] However, in actual industrial field inspections, structural component surfaces often encounter multiple interfering factors, such as residual oil film, localized high-reflectivity coatings, scattering from semi-transparent materials, abrupt changes in edge curvature, and scanning posture disturbances. These factors can lead to optical interference phenomena in the inspection results, including reflected flypoints, scattering spurious points, localized defects, and boundary fractures. Simultaneously, the edges of genuine cracks, holes, or coating peeling can also exhibit abrupt changes in reflection intensity and localized geometric discontinuities. This creates the following core contradiction: if the inspection data is processed solely based on a fixed intensity threshold or geometric smoothing rules, genuine defect boundaries may be mistakenly deleted, or spurious anomalies caused by reflection and scattering may be retained.

[0005] Specifically, in the inspection of aluminum alloy coated structural components, thin oil films can create localized high-brightness reflection points, while scratch boundaries can also produce abrupt changes in reflection intensity—making them difficult to distinguish on a single intensity dimension. If a fixed reflection intensity threshold is used for filtering, oil film reflection points and actual scratch boundary points may be simultaneously identified as anomalies; further geometric smoothing may flatten scratch boundaries, causing defect boundaries to be truncated or contracted. In the inspection of composite material components, material scattering and delamination defects can cause local response instability. Without a mechanism to distinguish spatial continuity and disturbance sources, the system can easily mistakenly retain scattering pseudo-points or delete actual delamination boundaries. When such errors enter the subsequent area localization and severity determination stages, the defect range may be mistakenly expanded, reduced, or broken, leading to unstable results in the determination of defect type and severity.

[0006] Therefore, the problem with existing technologies is that the accuracy of surface defect analysis results needs to be improved in the context of structural component surface inspection. Summary of the Invention

[0007] This invention provides a method and apparatus for processing optical data on the surface of industrial structural components, which addresses the problem of improving the accuracy of surface defect analysis results in structural component surface inspection scenarios.

[0008] In a first aspect, embodiments of the present invention provide a method for processing optical data on the surface of an industrial structural component, comprising: fusing detection data from at least one optical detection data source on the surface of the industrial structural component to obtain its point-level feature vector; performing denoising processing based on the abnormal risk assessment value of the point-level feature vector in a risk assessment model to obtain first denoised detection data; performing suppression processing on the spectral domain components of the risk frequency band data in the spectral domain of the first denoised detection data to obtain second detection data; generating the confidence level of the industrial structural component based on the abnormal risk assessment value of the detection points in the second and second detection data, and performing weighted filtering sampling based on the confidence level of the structural component to obtain third detection data; performing local response field estimation, defect region registration, and three-dimensional reconstruction based on the third detection data to obtain reconstructed data of the structural component, and extracting structural component feature data from the reconstructed data to generate structural component defect analysis results.

[0009] In conjunction with the first aspect, the present invention provides a first possible implementation of the first aspect, wherein the process of fusing detection data from at least one optical detection data source on the surface of an industrial structural component to obtain its point-level feature vector includes: acquiring preprocessed detection data of detection points from at least one optical detection data source and their corresponding neighboring detection points, unifying them to the same coordinate system, and completing the registration and alignment of the detection data; this includes the point coordinates, reflection intensity, acquisition time, scanning pose, multi-angle light intensity response, polarization response, and transmission attenuation information of each detection point; based on the point coordinates, obtaining the average neighborhood distance characterizing the spatial dispersion of the detection point's neighborhood, the normal consistency characterizing the directional consistency of the detection point's neighborhood surface, the curvature change characterizing the curvature change of the detection point's neighborhood surface, and the feature vector characterizing the detection point relative to the surface of the detection point. The planar fitting residual represents the degree of deviation of the local plane in the neighborhood; the reflection intensity jump, representing the degree of abrupt change in the reflection intensity of the detection point relative to the neighborhood, is obtained based on the reflection intensity; the scattering dispersion, representing the degree of dispersion of the multi-angle light intensity response of the detection point, the multi-angle response consistency, representing the degree of matching between the multi-angle light intensity response of the detection point and the neighboring detection points, the polarization response difference, representing the degree of deviation of the polarization response of the detection point and the neighboring detection points, and the transmission attenuation difference, representing the degree of deviation of the transmission attenuation of the detection point and the neighboring detection points, are obtained based on the multi-angle light intensity response, polarization response difference, and transmission attenuation difference; the average neighborhood distance, normal consistency, curvature change, planar fitting residual, reflection intensity jump, scattering dispersion, multi-angle response consistency, polarization response difference, and transmission attenuation difference are combined to obtain a point-level feature vector.

[0010] In conjunction with the first aspect, this embodiment of the invention provides a second possible implementation of the first aspect, wherein the process of assessing the abnormal risk value in the risk assessment model based on the point-level feature vector includes: inputting the point-level feature vector into a pre-constructed risk assessment model to obtain its first risk output; generating point-level loss data of the detection points based on the first risk output, and determining a tail sample set based on the numerical distribution of the point-level loss data; calculating the conditional risk value based on the point-level loss data in the tail sample set, and performing gradient backpropagation on the risk assessment model based on the conditional risk value to obtain the risk gradient response of the detection points; Conditional Value at Risk (VaR) is satisfied: ; in, Indicates quantile parameter as Conditional Value at Risk Represents the tail sample set The number of samples in The point-level loss value corresponding to the i-th tail sample is represented by i, where i is a positive integer; the abnormal risk assessment value of the detection point is obtained by weighted summation of the first risk output and the risk gradient response based on preset weights.

[0011] In conjunction with the first aspect, this embodiment of the invention provides a third possible implementation of the first aspect, wherein the process of generating point-level loss data of the detection point based on the first risk output and determining the tail sample set based on the numerical distribution of the point-level loss data includes: determining whether the first risk output is located within a preset risk interval; if it is located, then performing loss mapping processing to obtain the point-level loss value of the detection point; if it is not located, then first performing boundary truncation processing, and then performing loss mapping processing based on the truncated first risk output to obtain the point-level loss value of the detection point; The point-level loss value satisfies: ; in, This represents the point-level loss value corresponding to the i-th detection point. This represents the first risk output value corresponding to the i-th detection point. The preset smoothing parameter is represented; the point-level loss values ​​are sorted in ascending order to determine the loss quantile threshold of the preset quantile parameter; the detection points with point-level loss values ​​not less than the loss quantile threshold are determined as tail samples and a tail sample set is formed.

[0012] In conjunction with the first aspect, this embodiment of the invention provides a fourth possible implementation of the first aspect, wherein the process of performing suppression processing on the spectral domain components of the risk band data of the first denoised detection data in the spectral domain to obtain the second detection data includes: performing nearest neighbor search on the detection points in the first denoised detection data to establish connection relationships between the detection points; determining the edge weights of the connection edges based on the spatial distance difference, reflection intensity difference, polarization response difference, and transmission attenuation difference between the detection points to obtain a nearest neighbor weighted graph; Edge weights satisfy: ; in, This represents the edge weight between the i-th detection point and the j-th detection point. Indicates the difference in spatial distance. This indicates poor reflection intensity. Indicates the difference in polarization response. Indicates the difference in transmission attenuation. , , and The scale parameters corresponding to the difference terms are respectively represented; the reflection intensity of the detection points in the first denoised detection data is arranged according to the order of the detection points to obtain the graph signal; a graph Laplacian matrix is ​​generated based on the nearest neighbor weighted graph, and eigenvalue decomposition is performed on the graph Laplacian matrix to obtain the eigenvector matrix; the graph signal is transformed from the spatial domain to the spectral domain based on the eigenvector matrix to obtain the spectral domain detection data. The transformation of the graph signal from the spatial domain to the spectral domain satisfies:

[0013] Where L represents the graph Laplacian matrix, D represents the degree matrix, W represents the edge weight matrix, and x represents the graph signal in the spatial domain. Let U represent the spectral domain detection data, T represent the eigenvector matrix, and T be the transpose operation. Detection points whose anomaly risk assessment value is not less than a preset risk threshold and are retained are obtained, resulting in retained risk detection points. Based on the node connectivity of these retained risk detection points in the nearest neighbor weighted graph, the adjacency distribution corresponding to each retained risk detection point is determined. The spectral domain response of the spectral domain detection data at each frequency component is statistically analyzed. Based on the adjacency distribution of these retained risk detection points in the nearest neighbor weighted graph, the risk projection weight corresponding to each frequency component is determined. The risk projection weight and spectral domain response are normalized, and the normalized risk projection is calculated. The product of the weights and the spectral response yields the frequency risk correlation value for each frequency component. Frequency components with a frequency risk correlation value not less than a preset frequency band determination threshold are identified as risk frequency components. Risk frequency components with consecutive frequency positions are combined to obtain risk frequency band data. Based on the risk frequency band data, the filter coefficients of the spectral filter at each frequency component are determined. The spectral detection data is multiplied by the corresponding filter coefficients at each frequency component to obtain suppressed spectral detection data. Based on the eigenvector matrix of the graph Laplacian matrix, the suppressed spectral detection data is inversely transformed to the spatial domain to obtain second detection data. The inverse transform satisfies: ; in, This represents the graph signal in the spatial domain of the enhanced detection data, where U represents the feature vector matrix. Represents an eigenvalue diagonal matrix. Let represent the spectral domain filter, x represent the spatial domain graph signal, and T represent the transpose operation.

[0014] In conjunction with the first aspect, this invention provides a fifth possible implementation of the first aspect, wherein the process of generating the confidence level of the industrial structural component based on the anomaly risk assessment values ​​of the detection points in the second detection data and performing weighted filtering sampling based on the structural component confidence level to obtain the third detection data includes: acquiring the local normal consistency, point density stability, curvature smoothness, reflection response consistency, and scattering stability of the detection points in the second detection data, and performing a weighted summation on them; correcting the weighted summation result based on the anomaly risk assessment values ​​to obtain the structural component confidence level; and using the detection points in the second detection data as a set of detection points to be sampled, and determining the sampling quantity based on a preset sampling quantity. The target sampling quantity is determined as follows: Detection points whose structural confidence levels meet preset initial selection conditions are selected from the set of detection points to be sampled, resulting in an initial set of selected detection points. Based on the detection points in the set of detection points to be sampled that are not included in the initial set, a remaining set of detection points to be sampled is obtained. The minimum distance between each detection point in the remaining set of detection points and the selected set is calculated, and a sampling priority value is calculated based on the corresponding structural confidence level. Detection points whose sampling priority values ​​meet preset sorting conditions are added to the selected set of detection points until the number of detection points in the selected set reaches the target sampling quantity. Third detection data is obtained based on the second detection data corresponding to each detection point in the selected set.

[0015] In conjunction with the first aspect, this invention provides a sixth possible implementation of the first aspect, wherein the process of obtaining the reconstructed data of the structural component by performing local response field estimation, defect region registration, and three-dimensional reconstruction based on the third detection data includes: acquiring the third detection data, which includes the point coordinates and reflection intensity of each detection point; establishing a neighborhood detection point set based on the point coordinates; acquiring the point coordinates and reflection intensity of the target detection point, as well as the point coordinates and reflection intensity of each neighboring detection point in the neighborhood detection point set corresponding to the target detection point; calculating distance weight data based on the point distance between the target detection point and each neighboring detection point; calculating a weighted sum of reflection intensity based on the distance weight data of the reflection intensity of each neighboring detection point, and calculating the weighted sum of reflection intensity based on the ratio of the weighted sum of reflection intensity to the sum of distance weight data. The system obtains the local response values ​​of the target detection points; combines the local response values ​​of each target detection point according to the detection point order to obtain the local response field; determines the response anomaly points based on the response value distribution of the detection points in the local response field, and determines the defect region based on the spatial connectivity between the response anomaly points; obtains the coordinates of the region boundary points of each defect region, and performs defect region registration processing based on the spatial correspondence between the region boundary point coordinates and the point coordinates of each detection point in the sampled detection data to obtain defect region registration data; performs surface reconstruction processing based on the point coordinates of each detection point in the sampled detection data to obtain the surface mesh data of the structural component; and maps the local response field of each defect region to the surface mesh data of the structural component based on the surface mesh data of the structural component and the defect region registration data to obtain the reconstructed data of the structural component.

[0016] Secondly, embodiments of the present invention also provide an electronic device, including: at least one processor; and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the processing method described above in the embodiments of the present invention by calling the program instructions.

[0017] Thirdly, embodiments of the present invention also provide a computer-readable storage medium storing computer instructions that cause the computer to perform the processing method described above in the embodiments of the present invention.

[0018] The method for processing optical data on the surface of industrial structural components provided by this invention systematically solves the technical problem of improving the accuracy of defect analysis results in the surface inspection scenario of industrial structural components through a staged progressive processing architecture: multi-source fusion risk assessment and denoising --> spectral domain risk-oriented suppression --> confidence-driven sampling --> local response field estimation and defect registration --> three-dimensional reconstruction and defect analysis.

[0019] Furthermore, this invention also has the following beneficial effects: multi-source optical data fusion to construct a comprehensive point-level feature vector. In constructing the point-level feature vector, this invention comprehensively integrates multiple optical detection data sources such as laser scanning, structured light scanning, and multi-angle optical detection, extracting multi-dimensional features including spatial geometric features (average neighborhood distance, normal consistency, curvature change, plane fitting residual), optical response features (reflection intensity jump, scattering dispersion, multi-angle response consistency), and material property response features (polarization response difference, transmission attenuation difference). This feature system can more completely characterize the physical properties and abnormal characteristics of the detection points, providing rich discriminative basis for subsequent risk assessment and data processing, enabling the system to maintain stable analytical performance even in complex industrial scenarios with multiple interferences such as oil films, high-reflectivity coatings, and semi-transparent scattering materials.

[0020] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

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

[0023] Figure 1 This is a flowchart illustrating the method for processing optical data on the surface of industrial structural components according to Embodiment 1 of the present invention. Figure 2 This is a flowchart illustrating the method for processing optical data on the surface of industrial structural components according to Embodiment 2 of the present invention. Figure 3 This is a schematic diagram of the structure of the electronic device according to Embodiment 3 of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] To improve the accuracy of defect analysis results in industrial structural component surface inspection scenarios, this invention employs a phased progressive processing architecture: multi-source fusion risk assessment and denoising → spectral domain risk-oriented suppression → confidence-driven sampling → local response field estimation and defect registration → 3D reconstruction and defect analysis. This approach systematically improves defect analysis results in industrial structural component surface inspection scenarios, such as increasing defect detection rate, reducing false alarm rate, reducing defect location error, improving defect type identification accuracy, and improving overall judgment accuracy, thereby systematically enhancing the accuracy of defect analysis results in industrial structural component surface inspection scenarios.

[0026] Example 1: This embodiment uses a U-shaped open-slot column / support component (virtual number: UC-SUP-2026-0042) manufactured through a combination of cold bending, welding, and machining as the inspection object. This structural component has a C / U-shaped cross-section, open at one end and closed at the other with a mounting flange. The areas to be inspected include the inner wall of the slot (inner surfaces of the two side panels and the bottom of the slot), the bend transition fillet, the outer edge, the cut edge at the open end, the inner wall of the bottom flange mounting hole, and the perimeter of the flange weld. Defect types to be inspected include surface pitting, micro-cracks at the bend fillet, weld porosity / slag inclusions, burrs at the open end, and cutting marks.

[0027] This embodiment collects detection data from three optical detection data sources: Data source A is a line laser scanner (Keyence LJ-X8000 series), acquiring approximately 2,134,580 points with a sampling interval of 0.05 mm. The coordinate system is the local coordinate system of the equipment (X along the scanning direction, Y for the feed direction, and Z for the column height direction). The main noise is flying spots at the edge of the slot opening and missing reflection blind spots in the deep cavity of the slot wall. Data source B is a structured light 3D camera (GOM ATOS Q), acquiring approximately 2,487,320 points with a sampling interval of 0.08 mm. The coordinate system is the camera coordinate system. The main noise is high reflectivity speckle on the outer edge and repeated points at the splicing seams of multiple views. Data source C is supplementary measurement data from a contact coordinate measuring machine (Hexagon Global S), acquiring approximately 15,680 points. The sampling interval is non-uniform (0.02 mm for bending fillets and mounting hole areas, skipping other areas). The coordinate system is the workpiece clamping coordinate system. There is almost no noise, but the coverage of the deep cavity inside the slot is limited.

[0028] Figure 1 This is a schematic flowchart of an industrial structural component surface optical data processing method provided in Embodiment 1 of the present invention. This embodiment includes the following steps: Step 100: Based on the detection data from at least one optical detection data source on the surface of the industrial structural component, fuse the detection data to obtain its point-level feature vector. In this embodiment, this step fuses the three heterogeneous, heterogeneous coordinate, and heterogeneous density optical detection data sources to construct a nine-dimensional point-level feature vector for each detection point, providing a multi-dimensional discrimination basis for subsequent risk assessment denoising and spectral domain risk suppression. The specific processing procedure is as follows.

[0029] (1) Multi-source detection data preprocessing and coordinate system First, preprocessed detection data from each detection point in three optical detection data sources were acquired, including point coordinates, reflection intensity, acquisition time, scanning pose, multi-angle light intensity response, polarization response, and transmission attenuation information for each detection point. This detection data was then unified into the same workpiece coordinate system W (Z-axis along the column height, X-axis along the groove width, Y-axis along the groove depth, and groove opening facing the +Y side). Coordinate system alignment employed a two-stage strategy: coarse registration using feature points combined with fine registration using Iterative Closest Point (ICP). Using the centers of the eight bottom flange mounting holes, the four corner points of the groove opening plane, and the centers of the two outer edge positioning surfaces provided by data source C as feature anchor points, the rigid transformation matrix from data sources A and B to the workpiece coordinate system W was solved using the quaternion method. After coarse registration, the maximum deviation for source A was 0.29 mm, and the maximum deviation for source B was 0.26 mm. Subsequently, using the merged point cloud as a reference, 50 iterations of ICP fine registration were performed, resulting in a final root mean square error (RMSE) of 0.016 mm.

[0030] After coordinate system alignment, a voxel mesh (voxel side length 0.10 mm) was created for the three-source fused point cloud. For multiple points within the same voxel, the one closest to the voxel center was retained, and the rest were deleted to remove duplicates in overlapping areas. After fusion, the total number of points was reduced from approximately 4,137,580 to approximately 1,124,860. Based on this, the point cloud bounding box was calculated: X∈[-62.0,62.0] mm, Y∈[0,48.5] mm, Z∈[0,485.0] mm (slot width 124 mm, slot depth 48.5 mm, column height 485 mm). The point cloud was translated to the centroid origin and normalized to a unit sphere according to the maximum diagonal length. Then, it was downsampled to 8192 detection points using Farthest Point Sampling (FPS) as the standard input point cloud P. (0) At the same time, the source label and the original normal information of each point are retained.

[0031] (2) Construction of point-level feature vectors Based on the detection data after unifying the coordinate system, a nine-dimensional point-level feature vector is constructed for each detection point. Specifically, based on the point coordinates of each detection point, the average neighborhood distance, representing the spatial dispersion of the detection point's neighborhood, is calculated. Specifically, for detection point i, its k=16 nearest neighbors are taken, and the average Euclidean distance between the detection point and each neighboring point is calculated. The normal consistency, representing the consistency of the orientation of the detection point's neighborhood surface, is calculated by taking the absolute value of the average of the dot products of the unit normal vectors of detection point i and each neighboring point. The curvature change, representing the degree of curvature change of the detection point's neighborhood surface, is calculated by the absolute value of the principal curvature difference of the locally fitted quadratic surface corresponding to detection point i. The plane fitting residual, representing the degree of deviation of the detection point from the local plane of its neighborhood, is calculated by the perpendicular distance from detection point i to the least squares fitted plane of its neighbors.

[0032] Taking detection point i=2048 (located in the rounded corner area of ​​the left vertical plate bend transition, from data source B) as an example, its average neighborhood distance is 0.0186 (normalized coordinate units, the same below), normal consistency is 0.337, curvature change is 0.129, and plane fitting residual is 0.0134. All indicators are significantly different from the corresponding values ​​in the plane area, reflecting that the area where this point is located has complex geometric discontinuities.

[0033] Based on the reflection intensity of each detection point, the reflection intensity jump, which characterizes the degree of abrupt change in the reflection intensity of the detection point relative to its neighborhood, is calculated. Specifically, it is the absolute value of the difference between the reflection intensity value of detection point i and the mean reflection intensity of its neighborhood points divided by the standard deviation of the neighborhood reflection intensity. The higher this ratio, the more discontinuous the optical response of the point relative to its neighborhood. Based on multi-angle light intensity response, polarization response, and transmission attenuation information, the following are calculated: scattering dispersion, which characterizes the dispersion of the multi-angle light intensity response at the detection point, i.e., the standard deviation of the light intensity response values ​​at the detection point under multiple illumination and reception angles; multi-angle response consistency, which characterizes the matching degree of the multi-angle light intensity response between the detection point and its neighboring detection points, defined as the mean of the Pearson correlation coefficients between the multi-angle light intensity response sequence of detection point i and the multi-angle light intensity response sequences of its neighboring points; polarization response difference, which characterizes the degree of deviation of the polarization response between the detection point and its neighboring detection points, defined as the Mahalanobis distance between the polarization state elliptic parameter of detection point i and the mean of the polarization state elliptic parameter of its neighboring points; and transmission attenuation difference, which characterizes the degree of deviation of the transmission attenuation between the detection point and its neighboring detection points, defined as the difference between the transmission attenuation coefficient of detection point i and the mean of the transmission attenuation coefficient of its neighboring points divided by the standard deviation of the transmission attenuation in the neighboring areas.

[0034] The nine feature components mentioned above—average neighborhood distance, normal consistency, curvature change, plane fitting residual, reflection intensity jump, scattering dispersion, multi-angle response consistency, polarization response difference, and transmission attenuation difference—are combined into a nine-dimensional vector, which serves as the point-level feature vector for the detection point. Taking a detection point i=5120 (located in the inner wall region of the bottom flange mounting hole, from data source A) as an example, its point-level feature vector is [0.0241, 0.121, 0.187, 0.0139, 0.734, 0.456, 0.512, 0.389, 0.602]. The values ​​of each component are relatively large and uniformly high, indicating that this point simultaneously exhibits multiple anomalous characteristics, including geometric discontinuities and abnormal optical responses.

[0035] Step 101: Based on the abnormal risk assessment value of the point-level feature vector in the risk assessment model, perform denoising processing to obtain the first denoised detection data. In this embodiment, this step utilizes a pre-built risk assessment model to perform anomaly risk assessment on the point-level feature vectors of each detection point. Based on the assessment results, selective denoising is performed to filter out high-confidence optical interference points (such as flying spots, speckles, edge oscillation pseudo-points, etc.) and retain high-risk detection points that may represent real defects. The specific processing procedure is as follows.

[0036] (1) Calculation of the first risk output The nine-dimensional point-level feature vectors of each detection point constructed in step 100 are input into the pre-constructed risk assessment model. In this embodiment, the risk assessment model adopts a point-level classification network with a Dynamic Graph CNN (DGCNN) as the backbone network. This network is pre-trained on labeled data of 4800 similar grooved column / support components, and the detection category set is 5 categories: {normal, pitting, microcracks, porosity / slag inclusions, and cuts / burrs}. The risk assessment model uses the output of forward propagation as the first risk output. (i is the index of the detection point, the same below), and its value range is in the interval [0,1]. The full point cloud risk statistics are: mean μ=0.128, standard deviation σ=0.205, 95th percentile=0.598.

[0037] (2) Calculation of point-level loss data and determination of tail samples Point-level loss data for each detection point is generated based on the first risk output. First, the first risk output of detection point i is determined. Whether it is located within the preset risk range [0,1]: If it is, directly perform loss mapping processing to obtain the point-level loss value of the detection point. If not located (i.e. If the value exceeds the interval boundary, boundary truncation is performed first, followed by loss mapping based on the truncated first risk output. The point-level loss value satisfies: in, This represents the point-level loss value corresponding to the i-th detection point. This indicates the first risk output value corresponding to this detection point. Preset smoothing parameters to prevent Numerical overflow occurs when the value approaches 1. Subsequently, the point-level loss values ​​of all detection points are sorted in ascending order, and the α quantile of the loss value is determined as the loss quantile threshold based on the preset quantile parameter α=0.90. Detection points with point-level loss values ​​not less than the loss quantile threshold are identified as tail samples, forming a tail sample set. .

[0038] (3) Calculation of Conditional Value at Risk and Risk Gradient Response Based on tail sample set Calculate the conditional value at risk (CVaR) using point-level loss data: ; in, The conditional value of risk is represented when the quantile parameter is α. | indicates the number of samples in the tail sample set. Let represent the point-level loss value corresponding to the i-th tail sample, where i is a positive integer. CVaR measures the conditional expectation of high-loss tail samples. Compared to the mean loss of all samples, CVaR shifts the focus of model optimization from a large number of low-information samples to the most difficult-to-determine boundary detection points. Gradient backpropagation is performed on the risk assessment model based on the CVaR value to obtain the risk gradient response for each detection point. .

[0039] (4) Abnormal risk assessment value and selective denoising The first risk is output based on preset weights. and risk gradient response We perform a weighted summation to obtain the anomaly risk assessment value for each detection point. In this embodiment, β=0.4 and γ= =0.6. This represents the anomaly risk assessment value of the i-th detection point. This represents the normalized first risk output value corresponding to the i-th detection point. This represents the normalized risk gradient response value corresponding to the i-th detection point. This indicates the preset fusion weight, which can be set to 0.5-0.8. For example, the R-value of detection point 2048 is 0.533, and the R-value of detection point 7312 is 0.807. Although their first risk output values ​​are close (0.863 vs 0.993), after fusing the risk gradient response, the R-value of point 7312 in the flying point region is significantly higher than that of point 2048 in the bend and rounded corner region, demonstrating the improved ability of CVaR-driven risk assessment to distinguish between interference and defect boundaries.

[0040] Denoising based on abnormal risk assessment values: (by...) Sort in descending order, and perform direct deletion operation with a preset deletion ratio of 5% (i.e., delete the top 409 detection points with the highest risk assessment value), starting from the standard input point cloud P. (0) Delete the above 409 points from (8192 points) to obtain the first denoised detection data P. ( ¹ ) (7783 points). The average risk of the remaining points after denoising is 0.084, which is 34.4% lower than the average of 0.128 before denoising.

[0041] This invention introduces a gradient backpropagation mechanism driven by Conditional Value at Risk (CVaR) into the risk assessment model, shifting the optimization objective of risk assessment from the mean loss of all samples to the tail samples with high loss. This allows the risk assessment model to prioritize the boundary detection points that are most difficult to determine during training or online inference. At the same time, by weighted fusion of the risk gradient response and the first risk output, the abnormal risk level of the detection point is comprehensively measured, resulting in a significant improvement in the distinguishability of the risk assessment value at the boundary between defects and interference compared to existing technologies.

[0042] Step 102: Perform suppression processing on the spectral domain components of the risk frequency band data in the first denoised detection data to obtain the second detection data. In this embodiment, since some interference signals that are difficult to distinguish from the actual defect response in the spatial domain still remain in the first denoised detection data, targeted suppression processing needs to be performed on the frequency components corresponding to the interference in the spectral domain. The specific process is as follows.

[0043] (1) Construction of weighted nearest neighbor graph For the first denoised detection data P ( ¹ ) The 7783 detection points in the sample were subjected to k-nearest neighbor search (taking... =16), establish the connection relationship between detection points. For each pair of detection points i and j with a connection relationship, based on their spatial distance difference... Poor reflection intensity Polarization response difference and transmission attenuation difference Determine the edge weight of the connecting edge The edge weights satisfy:

[0044] in, , , and These are the scale parameters for four different items: spatial distance, reflection intensity, polarization response, and transmission attenuation. In this embodiment... Take 1% of the maximum diagonal length of the bounding box, which is approximately 0.0173 (normalized coordinates). Take twice the standard deviation of the denoised reflection intensity; and Take twice the standard deviation of the corresponding data. For example, consider the edge connecting point i=5120 and its neighboring point j=5118: =0.011, =0.023, =0.015, =0.019, the edge weight is calculated. =0.872, indicating that the two points are highly similar in both spatial and optical response dimensions.

[0045] (2) Definition of Graph Signal and Spectral Domain Transformation The reflection intensity values ​​of each detection point in the first denoised detection data are arranged in order of detection points and defined as the graph signal x. The degree matrix D is calculated based on the edge weight matrix W. The graph Laplacian matrix L = DW is obtained. Eigenvalue decomposition is performed on L to obtain the eigenvalue diagonal matrix Λ = diag( , ..., And the eigenvector matrix U. Based on U, transform the spatial domain graph signal x to the spectral domain: in, For spectral domain detection data, This represents the transpose of the eigenvector matrix U. The frequency domain energy distribution is as follows: approximately 91.8% in the low-frequency band (k≤778), approximately 7.1% in the mid-frequency band (779≤k≤6223), and approximately 1.1% in the high-frequency band (k≥6224).

[0046] (3) Identification of risk frequency band data In step 101 ≥ The set H of retained risk detection points with a value of 0.5 and which have not been deleted. The adjacency distribution vector corresponding to each retained risk detection point is determined based on the node connectivity relationships of the retained risk detection points in the nearest neighbor weighted graph. Statistical spectral domain detection data. Spectral response at each frequency component Simultaneously, the adjacency distribution vector of each retained risk detection point is projected onto the spectral domain through U to obtain the risk projection weight. .right and After L2 normalization, the values ​​are multiplied to obtain the frequency risk correlation value. .Will Frequency components with a value ≥0.15 are identified as risk frequency components and combined into risk frequency band data. In this embodiment, high-frequency band A (k=6225~7156), high-frequency band B (k=5872~5989), and mid-high frequency band C (k=3341~3567) are identified. After merging, two effective risk frequency bands with a total of approximately 1276 frequency components are obtained.

[0047] (4) Spectral domain directional suppression and inverse transform For the frequency component k belonging to the risk band, set the filter coefficient. =0.05; For frequency components that do not belong to the risk frequency band, set... =1.0. (The rest of the text appears to be a mix of characters and symbols, possibly representing a corrupted or incomplete sentence.) Multiply The suppressed spectral domain detection data is obtained, and then the second detection data is obtained through inverse transformation:

[0048] in, Let x be the second detection data, Λ be the image signal of the first denoised detection data, h(Λ) be the eigenvalue diagonal matrix, and h(Λ) be the spectral domain filter. Taking k=7156 as an example: =0.821, which was determined to be a line laser scanning stripe disturbance. =0.05, the component is suppressed to 5% of its original value; taking k=142 as an example: =0.038, determined to be the overall cross-sectional shape information of the U-shaped channel --> =1.0, components are fully preserved. The Chamfer distance of the point cloud before and after suppression is 0.0031, and the maximum physical deviation is 0.043mm.

[0049] This invention constructs a nearest-neighbor weighted graph, incorporating spatial distance differences, reflection intensity differences, polarization response differences, and transmission attenuation differences among detection points into edge weight calculations to accurately characterize the multi-dimensional relationships between detection points. Reflection intensity is used as the graph signal, and the graph signal is transformed from the spatial domain to the spectral domain through graph Laplacian matrix eigenvalue decomposition. Furthermore, risk projection weights are determined based on the adjacency distribution of risk detection points retained from anomaly risk assessment values, and risk frequency bands are identified by combining the normalized product of spectral domain responses. Targeted suppression is applied only to the risk frequency band components, rather than uniform attenuation across the entire frequency band. Therefore, this invention can effectively reduce optical interference components such as specular reflection, scattering, and attitude disturbances while fully preserving the abrupt response of the true defect boundary. This solves the problem of blurred or lost defect boundaries caused by overall smoothing in existing technologies, ensuring that the boundary structure of the defect region remains clear after denoising.

[0050] Step 103: Generate the confidence level of the industrial structural component based on the anomaly risk assessment values ​​of the detection points in the second detection data and perform weighted filtering sampling based on the confidence level of the structural component to obtain the third detection data. In this embodiment, the second detection data has largely eliminated systematic optical interference components after spectral domain-oriented suppression processing, but a certain number of redundant detection points still exist in the data. This step uses weighted filtering sampling driven by structural confidence to prioritize and retain key detection points that contribute significantly to defect analysis.

[0051] (1) Generation of confidence scores for structural components Five quality indicators were obtained from the second detection data: local normal consistency, point density stability, curvature smoothness, reflection response consistency, and scattering stability. These indicators were then weighted and summed.

[0052] In this embodiment, the weights are: =0.20、 =0.15、 =0.20、 =0.25、 =0.20. Based on the anomaly risk assessment value in step 101. Corrected to obtain the confidence level of the structural component: , =2.0 Taking the detection point i'=3156 (the suspected pitting area on the inner wall of the bottom flange mounting hole) as an example: Q=0.592, R=0.931, C=1.697; taking the detection point i'=892 (the normal surface of the outer edge) as an example: Q=0.423, R=0.058, C=0.473.

[0053] (2) Confidence-driven weighted filtering sampling All 7783 detection points were used as the sampling set. ≥ Using the upper quartile (0.892) as the initial selection criterion, approximately 1946 initial detection points were selected. The minimum distance between the remaining detection points was then calculated. and combined Calculate the sampling priority value:

[0054] in =0.30, =0.70. Iterative selection The largest value is added to the selected set until the target number of 7783 is reached, resulting in the third detection data. After sampling, the density of key points in the bend transition fillet area increased by 2.4 times, the density of the inner wall of the bottom flange mounting hole increased by 1.9 times, and the retention rate of high-risk points reached 95.1%.

[0055] This invention generates a structural confidence score for each detection point by weighting and summing multiple quality indicators, including local normal consistency, point density stability, curvature smoothness, reflection response consistency, and scattering stability, in the enhanced detection data, and correcting them with anomaly risk assessment values. During downsampling, a sampling priority value is jointly constructed using the structural confidence score and the distance between points, prioritizing the selection of detection points with high confidence and uniform spatial distribution. This method effectively reduces the amount of data processed subsequently while prioritizing the retention of key detection points that reflect the true defect boundary structure, avoiding the risk of erroneously discarding key defect boundary points during uniform downsampling or random sampling, thus ensuring the morphological accuracy and analytical stability of the defect region.

[0056] Step 104: Based on the third detection data, perform local response field estimation, defect region registration, and three-dimensional reconstruction to obtain the reconstructed data of the structural component, and extract the structural component feature data from the reconstructed data to generate the structural component defect analysis results. In this embodiment, the first three steps have completed the signal cleaning and key point screening. This step moves into the stage of substantive defect identification, localization and three-dimensional characterization, and finally outputs complete defect analysis results.

[0057] (1) Calculation of local response field Obtain the coordinates and reflection intensity of each detection point in the third detection data. For each detection point i, select the k=16 nearest neighbors to form a neighborhood set. Distance weights are calculated using a Gaussian kernel based on the distance between points. (Corresponds to approximately 1.5mm of physical space) Calculate the local response value:

[0058] All By combining the detection points in sequence, the local response field L is obtained.

[0059] (2) Defect area determination and registration Local response value Exceed = +2 Detection points with a value of 0.684 are identified as response anomalies. For any two response anomalies, if the Euclidean distance is less than 0.684... =0.05 (approximately 3.5mm), determined to be a connected component of the same defect area. In this example, defect area #1, consisting of approximately 127 abnormal response points, was identified in the inner wall area of ​​the bottom flange mounting hole, and defect area #2, consisting of approximately 89 points, was identified in the left vertical plate bending transition fillet area. The coordinates of the α-shape boundary points of each defect area were extracted, and the nearest neighbor was found in the third detection data. For detection points (<0.005, approximately 0.35mm), the boundary coordinates are replaced with the corresponding original coordinates to obtain the registration data for the defect area.

[0060] (3) Three-dimensional reconstruction and defect analysis Based on the third detection data, a 4x upsampling surface reconstruction was performed using a Denoiser and Upsampler Network (DUP-Net), resulting in structural component surface mesh data with approximately 29,854 vertices. The local response fields of each defect region were mapped to the mesh vertices using k=4 distance-weighted interpolation, and the defect regions were labeled to obtain the reconstructed data. Feature data such as defect location coordinates, boundary contours, response field intensity distribution, geometric depth / height deviations, and spatial dimensions were extracted to generate a complete defect analysis result covering defect location, extent, type, grade, and optical property deviations.

[0061] This invention performs local response field estimation on sampled detection data, fusing reflection intensity information from each detection point and its neighboring points using a distance-weighted method to obtain a spatially continuous local response field. Based on the spatial connectivity of anomalies in the response field, defect regions are determined, and defect region registration is performed using the spatial correspondence between the region boundary point coordinates and the original detection point coordinates. Simultaneously, surface reconstruction is performed based on the sampled detection point coordinates, mapping the registered defect region's local response field to the reconstructed structural component surface mesh data. This process organically combines optical response analysis, defect spatial localization, and geometric topography reconstruction, overcoming the problem of unstable defect type and grade determination caused by relying solely on topography or reflection intensity information. It can output complete analysis results of defect location, range, type, grade, and optical characteristic deviations within a unified framework, adapting to the diverse inspection needs of industrial structural components such as oil films, coatings, and scattering materials.

[0062] In this embodiment, the tested object UC-SUP-2026-0042 was ultimately determined to be unqualified, with two defects detected: Defect #1 is located on the inner wall of the bottom flange mounting hole, the type is pitting corrosion, the pit depth is about 0.18mm, the diameter is about 1.4mm, and the area is about The first defect is classified as medium and it is recommended to repair it by welding and then re-inspect it. The second defect is located at the bend transition fillet of the left vertical plate. It is a microcrack with a length of about 2.8 mm, a maximum width of about 0.06 mm, and a direction along the tangent of the bend fillet. It is classified as severe and it is recommended to confirm it by magnetic particle testing.

[0063] The above-described industrial structural component surface optical data processing method, combined with this embodiment, employs a phased progressive processing architecture: multi-source fusion risk assessment and denoising → spectral domain risk orientation suppression → confidence-driven sampling → local response field estimation and defect registration → 3D reconstruction and defect analysis. This approach improves defect detection rate, reduces false alarm rate, decreases defect location error, enhances defect type identification accuracy, and improves overall judgment accuracy, thereby systematically improving the accuracy of defect analysis results in industrial structural component surface inspection scenarios. Furthermore, this embodiment also offers the following advantages: multi-source optical data fusion to construct comprehensive point-level feature vectors. In constructing point-level feature vectors, this invention comprehensively integrates multiple optical detection data sources, including laser scanning, structured light scanning, and multi-angle optical detection, extracting multi-dimensional features including spatial geometric features (average neighborhood distance, normal consistency, curvature change, plane fitting residual), optical response features (reflection intensity jump, scattering dispersion, multi-angle response consistency), and material property response features (polarization response difference, transmission attenuation difference). This feature system can more completely characterize the physical properties and abnormal characteristics of the detection points, providing rich criteria for subsequent risk assessment and data processing, and enabling the system to maintain stable analytical performance in complex industrial scenarios with multiple interferences such as oil film, high-reflection coating, and semi-transparent scattering materials.

[0064] The specific implementation methods and cooperative working processes of various technical features in the industrial structural component surface optical data processing method disclosed in the embodiments of the present invention have been fully disclosed and described. Those skilled in the art should understand that the parameter values ​​involved in the above embodiments are merely illustrative examples and can be adaptively adjusted according to the structural type of the object being detected, the required detection accuracy, and the type and quantity of optical detection data sources, without departing from the protection scope defined by the claims of the present invention.

[0065] Example 2 like Figure 2 As shown, Figure 2 This is a schematic flowchart of the surface optical data processing method for industrial structural components provided in Embodiment 2 of the present invention. Embodiment 2 provides another specific example of a surface optical data processing method for industrial structural components. The method includes the following steps.

[0066] Step S201: Obtain preprocessed detection data of multiple detection points on the surface of industrial structural parts, and extract point-level feature vectors based on the preprocessed detection data of the detection points; The process of extracting point-level feature vectors from preprocessed detection data based on detection points includes: Acquire preprocessed detection data of the detection point and its corresponding neighboring detection points, including the point coordinates, reflection intensity, acquisition time and scanning pose of each detection point, and including at least two of the following: multi-angle light intensity response, polarization response and transmission attenuation information. Based on the point coordinates, we obtain the average neighborhood distance, which characterizes the spatial dispersion of the detection point's neighborhood; the normal consistency, which characterizes the orientation consistency of the detection point's neighborhood surface; the curvature change, which characterizes the curvature change of the detection point's neighborhood surface; and the plane fitting residual, which characterizes the deviation of the detection point from the local plane of its neighborhood. The reflection intensity jump is obtained based on the reflection intensity, which characterizes the degree of abrupt change in the reflection intensity of the detection point relative to the neighborhood reflection intensity. Based on multi-angle light intensity response, polarization response and transmission attenuation information, the following are obtained: scattering dispersion, multi-angle response consistency, polarization response difference, and transmission attenuation difference, which characterize the dispersion of multi-angle light intensity response of the detection point, multi-angle response consistency, polarization response difference, and transmission attenuation difference. The average neighborhood distance, normal consistency, curvature change, plane fitting residual, reflection intensity jump, scattering dispersion, multi-angle response consistency, polarization response difference, and transmission attenuation difference are combined in a preset order to obtain a point-level feature vector.

[0067] Specifically, preprocessed detection data from multiple detection points on the surface of industrial structural components is acquired. This preprocessed detection data is the data after coordinate unification, multi-view registration, range constraint, point density constraint, and voxelization. It includes the point coordinates, reflection intensity, acquisition time, and scanning pose of each detection point, and at least two of the following: multi-angle light intensity response, polarization response, and transmission attenuation information.

[0068] For any given detection point, this detection point is designated as the target detection point. Based on the coordinates of the target detection point, detection points in the preprocessed detection data that satisfy preset neighborhood conditions are selected as the neighborhood detection point set. The preset neighborhood conditions can be that the distance between points is no greater than a preset neighborhood radius, or that there is a preset number of detection points closest to the target detection point. Specifically, this can be set as follows: selecting detection points with a distance of no more than 0.5mm-1.5mm between them as neighborhood detection points, centered on the target detection point; or selecting 12-32 detection points closest to the target detection point as neighborhood detection points. For laser scanning data with high detection point density, it is preferable to select the 24 closest detection points as neighborhood detection points.

[0069] The distance data between points is calculated based on the point coordinates of the target detection point and the point coordinates of the neighboring detection points. The mean of the distance data between points is used as the average neighborhood distance to characterize the spatial dispersion of the detection points in the neighborhood of the target detection point.

[0070] A local covariance matrix is ​​constructed based on the coordinates of neighboring detection points. The local normal data of the target detection point is determined based on the eigenvectors of the local covariance matrix. Based on the statistical values ​​of the angle between the local normal data of the target detection point and the local normal data of the neighboring detection points, the normal consistency of the target detection point, which represents the degree of consistency of the orientation of the neighboring surface of the target detection point, is obtained.

[0071] The local curvature data of the target detection point is determined based on the eigenvalues ​​of the local covariance matrix. Based on the statistical difference between the local curvature data of the target detection point and the local curvature data of the neighboring detection points, the curvature change of the target detection point, which represents the degree of bending change of the neighboring surface of the target detection point, is obtained.

[0072] Local plane fitting is performed based on the coordinates of neighboring detection points to obtain local plane parameters. Based on the distance from the coordinates of the target detection point to the local plane, the plane fitting residual, which characterizes the degree of deviation of the target detection point from the neighboring local plane, is obtained.

[0073] The reflection intensity difference data is calculated based on the reflection intensity of the target detection point and the reflection intensity of the neighboring detection points. The statistical results of the reflection intensity difference data are used as the reflection intensity jump to characterize the degree of abrupt change in the reflection intensity of the target detection point relative to the neighboring detection points.

[0074] When the preprocessed detection data includes multi-angle light intensity response, the light intensity discrete statistical results are calculated based on the light intensity values ​​of the target detection point at multiple acquisition angles to obtain the scattering discreteness characterizing the degree of dispersion of the multi-angle light intensity response of the target detection point; and based on the light intensity difference data between the target detection point and the neighboring detection points at the same acquisition angle, the multi-angle response consistency characterizing the degree of matching between the target detection point and the neighboring detection points in the multi-angle light intensity response is obtained.

[0075] When the preprocessed detection data includes polarization response, the polarization response difference data is calculated based on the polarization response of the target detection point and the polarization response of the neighboring detection points. The statistical results of the polarization response difference data are used as the polarization response difference to characterize the degree of deviation between the target detection point and the neighboring detection points in polarization response.

[0076] When the preprocessed detection data includes transmission attenuation information, the transmission attenuation difference data is calculated based on the transmission attenuation information of the target detection point and the transmission attenuation information of the neighboring detection points. The statistical result of the transmission attenuation difference data is used as the transmission attenuation difference characterizing the degree of deviation between the target detection point and the neighboring detection points in terms of transmission attenuation information.

[0077] The average neighborhood distance, normal consistency, curvature change, plane fitting residual, reflection intensity jump, scattering dispersion, multi-angle response consistency, polarization response difference, and transmission attenuation difference are combined in sequence to obtain the point-level feature vector of the target detection point. This process is repeated for multiple detection points in the preprocessed detection data to obtain point-level feature vectors corresponding to each detection point. In one embodiment, to avoid the influence of different dimensions on the risk assessment model input, each of the above features can be normalized before combination, and the normalized features are combined in sequence to obtain the point-level feature vector. Max-min normalization can be used for normalization.

[0078] Example: Taking an aluminum alloy coated structural component as an example, pre-processing inspection data of the component surface is obtained. This pre-processing inspection data includes 52,000 inspection points with an average spacing of 0.20 mm between them. For the target inspection point numbered P12864, its coordinates are (126.40 mm, 74.80 mm, 2.16 mm), the reflection intensity is 0.72, the acquisition time is 10.245 s, and the scanning pose is the pose data of the corresponding scanning head at that moment. This inspection point also has multi-angle light intensity responses of 0.68, 0.74, and 0.71 at three acquisition angles, a polarization response of 0.33, and a transmission attenuation information of 0.18.

[0079] Using P12864 as the target detection point, the 24 nearest detection points were selected as the neighborhood detection point set. Based on the point distance between P12864 and the 24 neighborhood detection points, the average neighborhood distance was found to be 0.46 mm. Local normal and curvature data were obtained from the coordinates of the neighborhood detection points, resulting in a normal consistency of 0.91 and a curvature variation of 0.037. The plane fitting residual was found to be 0.018 mm based on the local plane fitting results. The reflection intensity difference between P12864 and the neighborhood detection points was found to be 0.21. The scattering dispersion was found to be 0.026 and the multi-angle response consistency was found to be 0.88 based on the multi-angle light intensity response. The polarization response difference was found to be 0.045 and the transmission attenuation difference was found to be 0.032 based on the polarization response and transmission attenuation information. According to a preset order, the above data is combined into the point-level feature vector of P12864: [0.46, 0.91, 0.037, 0.018, 0.21, 0.026, 0.88, 0.045, 0.032].

[0080] Through the above technical solution, this application solves the problem that single geometric information or single reflection information in the surface inspection data of industrial structural parts is difficult to characterize the real defect state. Based on the point coordinates, reflection intensity and related optical response of the detection point, point-level feature vectors are extracted, which can characterize the spatial discreteness, surface orientation change and local plane deviation of the detection point neighborhood, as well as the abrupt change of reflection intensity, scattering change, polarization change and transmission attenuation change, providing a stable data foundation for subsequent anomaly risk assessment and defect boundary identification.

[0081] Step S202: Input the point-level feature vector into the pre-constructed risk assessment model to obtain the abnormal risk score of the detection point, remove the detection points whose abnormal risk score is not less than the preset risk threshold and do not meet the preset defect boundary continuity condition, and retain the pre-processed detection data of the remaining detection points as denoised detection data. The process of inputting point-level feature vectors into a pre-built risk assessment model to obtain anomaly risk scores for detection points includes: The point-level feature vectors of the detection points are input into the pre-constructed risk assessment model to obtain its initial risk output; Point-level loss data for detection points is generated based on the initial risk output, and the tail sample set is determined based on the numerical distribution of the point-level loss data. The conditional risk value is calculated based on the point-level loss data in the tail sample set, and gradient backpropagation is performed on the risk assessment model based on the conditional risk value to obtain the risk gradient response of the detection point. Conditional Value at Risk (VaR) is satisfied: ; in, Indicates quantile parameter as Conditional Value at Risk Represents the tail sample set The number of samples in This represents the point-level loss value corresponding to the i-th tail sample, where i is a positive integer; The abnormal risk score of the detection point is obtained by weighted summation of the initial risk output and the risk gradient response based on preset weights.

[0082] The abnormal risk score meets the following requirements: ; in, This represents the anomaly risk score for the i-th detection point. This represents the normalized initial risk output value corresponding to the i-th detection point. This represents the normalized risk gradient response value corresponding to the i-th detection point. This indicates the preset fusion weight. The preset fusion weight can be set to 0.5-0.8.

[0083] The process of generating point-level loss data for detection points based on the initial risk output, and determining the tail sample set based on the numerical distribution of the point-level loss data, includes: Determine whether the initial risk output is within the preset effective risk range; The preset effective risk interval is used to limit the effective range of the risk assessment model output value. If the risk assessment model output value has been normalized, the preset effective risk interval is set to [0,1]; if the numerical stability of the log loss calculation is considered, the initial risk output after boundary truncation can be limited to the interval [0.001,0.999].

[0084] If it is located, then perform loss mapping processing to obtain the point-level loss value of that detection point; If it is not located, first perform boundary truncation processing, and then perform loss mapping processing based on the initial risk output after truncation to obtain the point-level loss value of the detection point; The point-level loss value satisfies: ; in, This represents the point-level loss value corresponding to the i-th detection point. This represents the initial risk output value corresponding to the i-th detection point. This represents the preset smoothing parameter; the preset smoothing parameter is used to avoid invalid values ​​in logarithmic operations during point-level loss value calculation, and can be set to... - The preferred setting is .

[0085] The point-level loss values ​​are sorted in ascending order to determine the loss quantile threshold of the preset quantile parameter; the preset quantile parameter can be set to 0.85-0.95.

[0086] Detection points with point-level loss values ​​not less than the loss quantile threshold are identified as tail samples, and a tail sample set is formed.

[0087] Specifically, the point-level feature vectors of multiple detection points obtained in step S201, as well as the preprocessed detection data corresponding to each detection point, are obtained.

[0088] By inputting the point-level feature vectors of each detection point into the pre-constructed risk assessment model, the initial risk output corresponding to each detection point, representing the degree of interference caused by the detection point to subsequent defect identification, is obtained.

[0089] Point-level loss data for each detection point is generated based on the initial risk output. Specifically, when the initial risk output is within a preset effective risk range, loss mapping processing is directly performed on the initial risk output to obtain the point-level loss value; when the initial risk output is not within the preset effective risk range, boundary truncation processing is first performed on the initial risk output, and then loss mapping processing is performed to obtain the point-level loss value.

[0090] The point-level loss values ​​of each detection point are sorted in ascending order to determine the loss quantile threshold corresponding to the preset quantile parameter. Detection points with point-level loss values ​​not less than the loss quantile threshold are identified as tail samples, forming a tail sample set.

[0091] The conditional risk value is calculated based on the point-level loss values ​​in the tail sample set, and gradient backpropagation is performed on the risk assessment model based on the conditional risk value to obtain the risk gradient response of each detection point, which represents the degree of influence of the point-level feature vector of the detection point on the conditional risk value.

[0092] The initial risk output and the risk gradient response are normalized respectively, and the normalized initial risk output and the normalized risk gradient response are weighted and summed based on preset weights to obtain the abnormal risk score of each detection point.

[0093] Detection points with abnormal risk scores not less than a preset risk threshold are identified as candidate rejection detection points. Based on the spatial connectivity, curvature change, and reflection intensity change relationships between the candidate rejection detection points and their neighboring detection points, it is determined whether the candidate rejection detection points meet the preset defect boundary continuity conditions.

[0094] The preset risk threshold can be set to 0.65-0.80. In one implementation, if the abnormal risk score output by the risk assessment model is normalized to the 0-1 range, the preset risk threshold is set to 0.70; when there are many highly reflective areas on the surface of the structural component, the preset risk threshold can be increased to 0.75 to reduce the false rejection of true boundary points.

[0095] The preset risk threshold can be constructed based on the mean and standard deviation of the abnormal risk scores of the detection points; The preset risk threshold is satisfied as follows: ; in, This indicates a preset risk threshold. This represents the mean of the abnormal risk scores. The standard deviation of the abnormal risk score is represented by the following: This represents the threshold adjustment coefficient, and an example value for the threshold adjustment coefficient can be 3.

[0096] The preset defect boundary continuity condition can be set as follows: spatial connectivity satisfies that the candidate detection point has at least 3 spatially connected detection points within its neighborhood; curvature variation satisfies that the curvature variation direction between the candidate detection point and the connected detection points is consistent; and reflection intensity variation satisfies that the reflection intensity difference is continuously distributed. When the above conditions are met, the candidate detection point is determined to satisfy the preset defect boundary continuity condition. For slender defects such as scratches and cracks, the number of connected detection points can be set to at least 3-5; for the boundary of holes or corrosion pits, the number of connected detection points can be set to at least 6-10.

[0097] If a candidate detection point does not meet the preset defect boundary continuity condition, the candidate detection point is removed; if the candidate detection point meets the preset defect boundary continuity condition, the candidate detection point is retained. For detection points with anomaly risk scores less than the preset risk threshold, their preprocessed detection data is retained.

[0098] The preprocessed detection data corresponding to all retained detection points are combined to obtain the denoised detection data.

[0099] Example: Taking aluminum alloy coated structural components as an example, step S201 obtains point-level feature vectors corresponding to 52,000 detection points, and inputs these point-level feature vectors into the pre-built risk assessment model.

[0100] The risk assessment model outputs the initial risk output for each detection point. This initial risk output is mapped to point-level loss values ​​and sorted in ascending order. Assuming a preset quantile parameter of 0.90, the loss values ​​in the top 90% after sorting are used as the loss quantile threshold. The 5200 detection points with point-level loss values ​​not less than this loss quantile threshold are identified as tail samples.

[0101] Conditional Value at Risk (VaR) is calculated based on the point-level loss values ​​of the 5200 tail samples, and gradient backpropagation is performed on the risk assessment model to obtain the risk gradient response for each detection point. Subsequently, the initial risk output and the risk gradient response are normalized and weighted, and then summed to obtain the anomaly risk score for each detection point.

[0102] In this example, the preset risk threshold is 0.70. A total of 3180 detection points with an abnormal risk score of not less than 0.70 were identified. For these detection points, it is further determined whether they meet the preset defect boundary continuity condition.

[0103] For example, the abnormal risk score of detection point P12864 is 0.74. There are 9 detection points connected to it in its neighborhood. These detection points are distributed along the same scratch boundary, and the curvature change and reflection intensity change are continuous. Therefore, it is determined that P12864 meets the preset defect boundary continuity condition, and its preprocessed detection data is retained.

[0104] For example, the abnormal risk score of detection point P24517 is 0.81, but there is only one connected detection point in its neighborhood with an abnormal risk score not less than the preset risk threshold, and the reflection intensity of this point shows an isolated abrupt change relative to the neighboring detection points. Therefore, it is determined that P24517 does not meet the preset defect boundary continuity condition, and its preprocessed detection data is removed.

[0105] Ultimately, among the 3180 candidate rejection detection points, 820 detection points were retained because they met the preset defect boundary continuity condition, and 2360 detection points were rejected; the remaining 48820 detection points with anomaly risk scores less than the preset risk threshold were directly retained. The preprocessed detection data corresponding to the 49640 retained detection points were combined to obtain the denoised detection data.

[0106] Through the above technical solution, Embodiment 2 of the present invention solves the problem that existing fixed threshold denoising methods easily confuse optical pseudopoints with real defect boundary points. It obtains the abnormal risk score of the detection point based on the point-level feature vector, and selectively removes the detection point by combining the preset risk threshold and the defect boundary continuity condition. This can remove high-risk pseudo-anomalies formed by reflection, scattering or attitude disturbance, and retain continuous abnormal detection points on the boundaries of cracks, holes, corrosion or coating peeling, thereby obtaining denoised detection data that is more suitable for subsequent image processing and defect analysis.

[0107] The process of pre-building the risk assessment model includes: Before performing optical defect analysis on the surface of industrial structural components, a risk assessment model is pre-built. The risk assessment model is used to output the initial risk output of the corresponding detection point based on the point-level feature vector of the detection point, and to provide basic data for subsequent calculation of abnormal risk scores.

[0108] Specifically, historical surface inspection data of structural components is acquired. This data includes the coordinates, reflection intensity, acquisition time, and scanning pose of multiple historical inspection points, and includes at least two of the following: multi-angle light intensity response, polarization response, and transmission attenuation information. This historical surface inspection data can originate from inspected metal structural components, coated structural components, composite material structural components, or semi-transparent engineering material components.

[0109] Historical structural component surface inspection data undergoes preprocessing consistent with the online inspection process to obtain historical preprocessed inspection data. Preprocessing includes coordinate unification, multi-view registration, range constraints, point density constraints, and voxelization, ensuring that the historical preprocessed inspection data has the same data structure as the preprocessed inspection data used in online inspection.

[0110] Based on the point coordinates, reflection intensity, multi-angle light intensity response, polarization response, and transmission attenuation information of each historical detection point in the historical preprocessed detection data, historical point-level feature vectors are extracted for each historical detection point. These historical point-level feature vectors include average neighborhood distance, normal consistency, curvature change, plane fitting residual, reflection intensity jump, scattering dispersion, multi-angle response consistency, polarization response difference, and transmission attenuation difference.

[0111] Obtain the verification results corresponding to historical inspection points, and generate risk labeling data for these historical inspection points based on the verification results. Verification results can be obtained through manual verification, comparison with standard samples, or consistency judgment based on repeated scans. If a historical inspection point is a spatially isolated reflection flypoint, scattering spurious point, or attitude disturbance point, or causes a false defect area in subsequent defect boundary identification, then the historical inspection point is marked as a high-risk inspection point; if a historical inspection point is located on the surface of a real structural component or on a real defect boundary, and shows consistency in position and response during repeated scans, then the historical inspection point is marked as a low-risk inspection point.

[0112] A risk assessment model is trained using historical point-level feature vectors as input and risk-labeled data as supervision information. The risk assessment model can be implemented using a multilayer perceptron, point feature network, or graph feature network. Its input is the point-level feature vector of the detection point, and its output is the initial risk output corresponding to that detection point.

[0113] During training, point-level loss data is generated based on the initial risk output and risk labeling data of the risk assessment model. These point-level loss values ​​are then sorted in ascending order to determine the loss quantile threshold corresponding to the preset quantile parameters. Historical detection points with point-level loss values ​​not less than the loss quantile threshold are identified as tail samples, and the conditional value of risk (VoV) is calculated based on the point-level loss data of these tail samples. The VoV is used to enhance the risk assessment model's ability to identify high-loss tail samples.

[0114] The risk assessment model's parameters are updated based on point-level loss data and conditional value of risk until the model training meets a preset stopping condition. The preset stopping condition can be that the training epochs have been reached, the validation set loss no longer decreases, or the high-risk detection point identification results in the validation set meet preset requirements.

[0115] The trained risk assessment model is stored as a pre-built risk assessment model. During online detection, the pre-built risk assessment model is invoked, and the point-level feature vector obtained in step S201 is input into the model to obtain the initial risk output corresponding to the detection point, and further anomaly risk scores are generated.

[0116] In a preferred embodiment of the present invention, the preset stopping conditions of the risk assessment model include 50-300 training epochs and the validation set loss changing by less than a certain amount within 5-20 consecutive epochs. - The risk assessment model training must meet at least one of the following criteria: a recall rate of at least 0.85 for high-risk detection points, a precision rate of at least 0.80, an F1 score of at least 0.82, and a validation set loss that does not decrease for 5-15 consecutive rounds, or a training time of 2-24 hours. Preferably, when the training rounds reach 120, the validation set loss does not decrease for 10 consecutive rounds, or the recall rate of high-risk detection points in the validation set reaches 0.90 and the F1 score reaches 0.85, the training of the risk assessment model is stopped, and the current model parameters are used as the pre-built risk assessment model parameters.

[0117] Step S203: Construct a nearest neighbor weighted graph based on the denoised detection data, and use the reflection intensity in the denoised detection data of the detection points as the graph signal. Based on the nearest neighbor weighted graph, transform the graph signal from the spatial domain to the spectral domain to obtain the spectral domain detection data. Step S203 includes: performing a nearest neighbor search on the detection points in the denoised detection data to establish the connection relationship between the detection points; The edge weights of the connecting edges are determined based on the spatial distance difference, reflection intensity difference, polarization response difference, and transmission attenuation difference between the detection points, resulting in a nearest neighbor weighted graph. Edge weights satisfy: ; in, This represents the edge weight between the i-th detection point and the j-th detection point. Indicates the difference in spatial distance. This indicates poor reflection intensity. Indicates the difference in polarization response. Indicates the difference in transmission attenuation. , , and These represent the scale parameters for the corresponding difference terms; the edge weight scale parameter is used to construct the nearest neighbor weighted graph. The spatial distance scale parameter can be set according to the average spacing between detection points. For example, when the average spacing between detection points is 0.2 mm, the spatial distance scale parameter can be set to 0.5 mm-1.0 mm; the reflection intensity difference scale parameter can be set to 0.05-0.20; the polarization response difference scale parameter can be set to 0.02-0.10; and the transmission attenuation difference scale parameter can be set to 0.02-0.10. When the detection data does not contain polarization response or transmission attenuation information, the edge weight of the connecting edge is determined only based on the spatial distance difference and the reflection intensity difference.

[0118] Arrange the reflection intensity of the detection points in the denoised detection data according to the order of the detection points to obtain the image signal; A graph Laplacian matrix is ​​generated based on a nearest-neighbor weighted graph, and eigenvalue decomposition is performed on the graph Laplacian matrix to obtain an eigenvector matrix. Based on the eigenvector matrix, the graph signal is transformed from the spatial domain to the spectral domain to obtain spectral domain detection data. The transformation of the graph signal from the spatial domain to the spectral domain satisfies:

[0119] Where L represents the graph Laplacian matrix, D represents the degree matrix, W represents the edge weight matrix, and x represents the graph signal in the spatial domain. Let U represent the spectral domain detection data, and U represent the eigenvector matrix.

[0120] Specifically, the denoising detection data obtained in step S202 is acquired. The denoising detection data includes the point coordinates, reflection intensity, polarization response, and transmission attenuation information of the retained detection points.

[0121] Based on the coordinates of each detection point in the denoised detection data, a nearest neighbor search is performed on each detection point to obtain the set of neighboring detection points corresponding to each detection point, and a nearest neighbor connection edge is established according to the connection relationship between the detection point and the neighboring detection points.

[0122] For each nearest neighbor connection edge, the edge weight, representing the degree of correlation between the two detection points in terms of spatial location and optical response, is calculated based on the spatial distance difference, reflection intensity difference, polarization response difference, and transmission attenuation difference between the two detection points at the two ends of the connection edge. The smaller the spatial distance difference, the smaller the reflection intensity difference, the smaller the polarization response difference, and the smaller the transmission attenuation difference, the larger the edge weight of the connection edge.

[0123] Based on all detection points, nearest neighbor edges, and the edge weights of each nearest neighbor edge, a nearest neighbor weighted graph is obtained. In the nearest neighbor weighted graph, the nodes correspond to the detection points in the denoised detection data, the connecting edges correspond to the nearest neighbor relationships between detection points, and the edge weights correspond to the degree of local correlation between detection points.

[0124] Arrange the reflection intensity of each detection point in the denoised detection data according to the order of the detection points to obtain a graphical signal that represents the distribution of reflection intensity of each detection point in the nearest neighbor weighted graph in the denoised detection data.

[0125] The edge weight matrix and degree matrix are generated based on the nearest neighbor weighted graph, and the graph Laplacian matrix is ​​obtained from the edge weight matrix and degree matrix; then, eigenvalue decomposition is performed on the graph Laplacian matrix to obtain the eigenvector matrix.

[0126] Based on the feature vector matrix, the graph signal is transformed from the spatial domain to the spectral domain to obtain spectral domain detection data that characterizes the distribution of reflection intensity on different frequency components of the nearest neighbor weighted graph.

[0127] Example: Taking the denoising test data of the aluminum alloy coated structural component obtained in step S202 as an example, denoising test data of 49,640 retained test points are obtained. Each test point includes point coordinates and reflection intensity, and some test points also include polarization response and transmission attenuation information.

[0128] For each detection point, the 12 nearest detection points are selected as the neighborhood detection point set, and nearest neighbor connections are established between the target detection point and its 12 neighborhood detection points. For the detection point numbered P12864, its neighborhood detection points include 12 detection points such as P12831, P12844, and P12859.

[0129] Taking the connecting edge between P12864 and P12859 as an example, their spatial distance is 0.31 mm, the difference in reflection intensity is 0.04, the difference in polarization response is 0.012, and the difference in transmission attenuation is 0.008. Based on these differences, the edge weight of this connecting edge is calculated to be 0.86. For detection points whose spatial distance exceeds the neighborhood range or whose optical response difference exceeds a preset scale, no nearest neighbor connecting edge is established.

[0130] Based on 49,640 detection points, the nearest neighbor edges corresponding to each detection point, and the edge weights of each edge, a nearest neighbor weighted graph is obtained. Then, the reflection intensities of the 49,640 detection points are arranged in order of their detection point numbers, resulting in a graph signal of length 49,640. For example, the reflection intensity corresponding to P12864 is 0.72, so its corresponding value in the graph signal is 0.72.

[0131] An edge weight matrix and a degree matrix are generated based on a nearest-neighbor weighted graph. Then, a graph Laplacian matrix is ​​obtained, and eigenvalue decomposition is performed on the graph Laplacian matrix to obtain an eigenvector matrix. Finally, the eigenvector matrix is ​​used to perform a spectral domain transformation on the graph signal to obtain spectral domain detection data. This spectral domain detection data serves as input for subsequent determination of risk frequency band data.

[0132] Through the above technical solution, Embodiment 2 of the present invention solves the problem that the detection points after denoising are still difficult to directly represent the local correlation and global disturbance distribution. Based on the denoised detection data, a nearest neighbor weighted graph is constructed, and the reflection intensity is mapped onto the nearest neighbor weighted graph as a graph signal. This can not only preserve the adjacency relationship between detection points formed by spatial distance and optical response, but also convert the dispersed reflection disturbances in the spatial domain into frequency components in the spectral domain, thereby providing a data foundation for subsequent identification of risk frequency bands and execution of spectral domain suppression.

[0133] Step S204: Based on the spectral domain detection data and the adjacency distribution of retained detection points with anomaly risk scores not less than a preset risk threshold in the nearest neighbor weighted graph, determine the risk frequency band data, perform suppression processing on the spectral domain components of the risk frequency band data, and inversely transform the suppressed spectral domain detection data to the spatial domain to obtain enhanced detection data. The process of determining risk frequency band data includes: obtaining detection points whose abnormal risk scores are not less than a preset risk threshold and are retained, thus obtaining retained risk detection points; Based on the node connectivity of the retained risk detection points in the nearest neighbor weighted graph, the adjacency distribution corresponding to the retained risk detection points is determined. Spectral domain response of statistical spectral domain detection data at each frequency component; Based on the adjacency distribution of retained risk detection points in the nearest neighbor weighted graph, the risk projection weights corresponding to each frequency component are determined. The risk projection weights and spectral response are normalized, and the product of the normalized risk projection weights and spectral response is calculated to obtain the frequency risk correlation value of each frequency component. Frequency components with a frequency risk correlation value not less than a preset frequency band determination threshold are identified as risk frequency components, and risk frequency components with continuous frequency positions are combined to obtain risk frequency band data.

[0134] In this second embodiment, the spectral domain components of the risk band data are further suppressed, and the suppressed spectral domain detection data is inversely transformed to the spatial domain to obtain enhanced detection data. This process includes: The filter coefficients of the spectral domain filter at each frequency component are determined based on the risk band data. Multiply the spectral domain detection data by the corresponding filter coefficients at each frequency component to obtain the suppressed spectral domain detection data; The eigenvector matrix based on the graph Laplacian matrix is ​​used to inversely transform the suppressed spectral domain detection data to the spatial domain to obtain enhanced detection data. The inverse transform satisfies: ; in, This represents the graph signal in the spatial domain of the enhanced detection data, where U represents the feature vector matrix. Represents an eigenvalue diagonal matrix. This represents a spectral domain filter.

[0135] Specifically, the spectral domain detection data and nearest neighbor weighted graph obtained in step S203 are acquired, as well as the abnormal risk scores and preset risk thresholds for each detection point obtained in step S202.

[0136] From the denoised detection data, detection points with anomaly risk scores not less than a preset risk threshold are selected and retained as retained risk detection points. These retained risk detection points are those that were not eliminated in the previous processing because they met the defect boundary continuity condition.

[0137] Based on the node connectivity of the retained risk detection points in the nearest neighbor weighted graph, the adjacency distribution of the retained risk detection points is determined. This adjacency distribution includes the number of neighboring detection points connected to each retained risk detection point and the corresponding edge weights, which characterize the degree of clustering of the retained risk detection points in the nearest neighbor weighted graph.

[0138] The spectral response of the statistical spectral domain detection data at each frequency component is analyzed, and the risk projection weight corresponding to each frequency component is determined based on the adjacency distribution of the retained risk detection points. Frequency components with a higher correlation to the adjacency distribution of the retained risk detection points have higher risk projection weights.

[0139] The risk projection weights and spectral responses corresponding to each frequency component are normalized, and the product of the normalized risk projection weights and the normalized spectral responses is calculated to obtain the frequency risk correlation value for each frequency component. The frequency risk correlation value characterizes the degree to which the corresponding frequency component is simultaneously affected by both the risk detection point distribution and the spectral response.

[0140] The frequency risk correlation value of each frequency component is determined to be no less than a preset frequency band determination threshold. If it is not less, the corresponding frequency component is identified as a risk frequency component. Risk frequency components with consecutive frequency positions are combined to obtain risk frequency band data. The preset frequency band determination threshold is used to determine the risk frequency components from each frequency component. After normalizing the risk projection weight and spectral response, their product is used as the frequency risk correlation value. The preset frequency band determination threshold can be set to 0.60-0.75, preferably 0.65. When the frequency risk correlation values ​​of at least three consecutive frequency components are not less than the preset frequency band determination threshold, these frequency components are combined to form risk frequency band data.

[0141] The filter coefficients of the spectral domain filter are determined based on the risk frequency band data, and suppression processing is performed on the spectral domain components corresponding to the risk frequency band data to obtain suppressed spectral domain detection data. For spectral domain components that do not belong to the risk frequency band data, their spectral domain components are kept unchanged or only a preset weak adjustment is performed. For spectral domain components that belong to the risk frequency band data, the filter coefficient can be set to 0.3-0.7; for spectral domain components that do not belong to the risk frequency band data, the filter coefficient is set to 1.

[0142] Based on the eigenvector matrix corresponding to the graph Laplacian matrix in step S203, the suppressed spectral domain detection data is inversely transformed to the spatial domain to obtain enhanced detection data. The enhanced detection data retains the correspondence between detection points in the denoised detection data and updates the reflection intensity response of the detection points, which is used for subsequent structural component confidence calculation and sampling processing.

[0143] Example: Taking the aforementioned aluminum alloy coated structural component as an example, step S203 obtains spectral domain detection data and nearest neighbor weighted graphs corresponding to 49,640 detection points. In step S202, the preset risk threshold is 0.70. Among them, 820 detection points have an abnormal risk score of not less than 0.70 and are retained because they meet the defect boundary continuity condition. Therefore, these 820 detection points are regarded as retained risk detection points.

[0144] The node connectivity relationships of the 820 retained risk detection points in the nearest neighbor weighted graph were read. Statistical analysis showed that one group of retained risk detection points was concentrated near the scratch boundary on the structural component surface, with an average of 10 adjacent detection points and an average connection edge weight of 0.78, indicating that the risk detection points in this area have a continuous adjacent distribution.

[0145] The spectral response of the statistical spectral domain detection data at each frequency component is calculated, and the frequency risk correlation value is calculated in combination with the adjacency distribution of the retained risk detection points. Assuming a preset frequency band determination threshold of 0.65, the frequency risk correlation values ​​of the 760th to 815th frequency components are all determined to be no less than 0.65, and these frequency components are continuous in frequency position; therefore, they are combined into a risk frequency band data segment.

[0146] Subsequently, suppression processing is performed on the spectral components corresponding to the risk frequency band data. For example, the corresponding spectral components are reduced to 40% of the original response by a preset filtering coefficient, while other spectral components remain unchanged. The suppressed spectral detection data is then inversely transformed to the spatial domain to obtain enhanced detection data.

[0147] In the enhanced detection data, the detection point number and coordinates remain consistent with the denoised detection data, and the reflection intensity response of the detection points is updated after spectral domain suppression. For example, the reflection intensity of detection point P12864 in the denoised detection data is 0.72, and after risk band suppression and inverse transform, its reflection intensity in the enhanced detection data is 0.68. This enhanced detection data serves as the input for subsequently generating the structural component confidence score.

[0148] Through the above technical solution, Embodiment 2 of the present invention solves the problem that existing smoothing processes easily weaken both optical interference and real defect boundaries. Based on spectral domain detection data and the adjacency distribution of retained risk detection points in the nearest neighbor weighted graph, risk frequency band data is determined, and the corresponding spectral domain components are suppressed. This can both directionally weaken the risk frequency components caused by reflection, scattering, and attitude disturbances, and avoid overall smoothing of real boundary information in non-risk frequency bands, thereby obtaining enhanced detection data that preserves the defect structure and reduces the influence of optical disturbances.

[0149] Step S205: Generate structural component confidence based on enhanced detection data and the abnormal risk scores of detection points in enhanced detection data, and perform weighted filtering sampling based on structural component confidence to obtain sampled detection data; The process of generating structural component confidence scores based on enhanced detection data and the anomaly risk scores of each detection point within the enhanced detection data, and then performing weighted filtering sampling based on these structural component confidence scores to obtain sampled detection data, includes: The local normal consistency, point density stability, curvature smoothness, reflection response consistency, and scattering stability of the detection points in the enhanced detection data are obtained and then summed using weighted averages. The weighted summation result is corrected based on the anomaly risk score to obtain the confidence level of the structural component; The detection points in the enhanced detection data are used as the set of detection points to be sampled, and the target sampling number is determined based on the preset sampling number. Select the structural component confidence level that meets the preset initial selection conditions from the set of sampling detection points as the initial selected detection points, and obtain the set of selected detection points; Based on the detection points in the set of detection points to be sampled that have not been included in the set of selected detection points, the set of remaining detection points to be sampled is obtained; Calculate the minimum distance between each detection point in the remaining set of unsampled detection points and the selected set of detection points, and calculate the sampling priority value in combination with the corresponding structural confidence level; Add the detection points whose sampling priority values ​​meet the preset sorting conditions to the selected detection point set until the number of detection points in the selected detection point set reaches the target sampling number; Based on the enhanced detection data corresponding to each detection point in the selected detection point set, the sampled detection data is obtained.

[0150] Specifically, the enhanced detection data obtained in step S204, and the anomaly risk score corresponding to each detection point in the enhanced detection data, are acquired. The enhanced detection data includes the point coordinates, reflection intensity, and optical response data after spectral suppression for each detection point.

[0151] Based on the neighboring detection points of each detection point in the enhanced detection data, the local normal consistency, point density stability, curvature smoothness, reflection response consistency, and scattering stability of each detection point are calculated, and combined with the anomaly risk score corresponding to the detection point, the confidence level of the structural component is generated.

[0152] Among them, the more the local normal consistency, point density stability, curvature smoothness, reflection response consistency, and scattering stability meet the preset stability conditions, the higher the confidence level of the structural component; the higher the anomaly risk score, the lower the confidence level of the structural component.

[0153] Local normal consistency is used to characterize the degree of consistency between the surface normal directions of the target detection point and its neighboring detection points. Local normal data is determined based on the point coordinates of the target detection point and its neighboring detection points, and the angle variation between the local normal data of the target detection point and the local normal data of its neighboring detection points is statistically analyzed; the smaller the angle variation, the higher the local normal consistency.

[0154] Point density stability is used to characterize whether the distribution of detection points within the neighborhood of a target detection point is uniform. The number of detection points within the neighborhood of the target detection point is counted and compared with a preset number of neighborhood points or the average number of neighborhood points; the smaller the difference, the higher the point density stability.

[0155] Curvature smoothness is used to characterize whether the surface curvature change within the neighborhood of a target detection point is continuous. Local curvature data is calculated based on the coordinates of the target detection point and neighboring detection points, and the difference between the local curvature data of the target detection point and the local curvature data of the neighboring detection points is statistically analyzed; the smaller the curvature difference, the higher the curvature smoothness.

[0156] Reflection response consistency is used to characterize whether the change in reflection intensity between the target detection point and its neighboring detection points is continuous. The reflection intensity difference is calculated based on the reflection intensity of the target detection point and the reflection intensity of its neighboring detection points; the smaller the reflection intensity difference, the higher the reflection response consistency.

[0157] Scattering stability is used to characterize whether the fluctuations in the multi-angle light intensity response of a target detection point are stable. Multi-angle light intensity difference data are calculated based on the multi-angle light intensity responses of the target detection point and neighboring detection points, and their dispersion is statistically analyzed; the smaller the dispersion, the higher the scattering stability.

[0158] Normalization was performed on local normal consistency, point density stability, curvature smoothness, reflection response consistency, and scattering stability to ensure that each index falls within the same numerical range. Then, the normalized indices were weighted and summed according to preset weights to obtain the initial confidence level of the target detection point.

[0159] The weighted summation result is corrected based on the anomaly risk score to obtain the confidence level of the structural component: ; in, This represents the confidence level of the structural component at the i-th detection point. This represents the weighted summation result of the i-th detection point. Let represent the abnormal risk score of the i-th detection point, and b represent the risk correction coefficient, which takes a value between 0 and 1.

[0160] The detection points in the enhanced detection data are used as the set of detection points to be sampled, and the target sampling number is determined based on a preset sampling quantity or a preset sampling ratio. The preset sampling ratio can be set to 20%-40% of the total number of detection points in the enhanced detection data; when the number of detection points is large, the target sampling number can also be directly set to 8,000-20,000 detection points. For example, when the enhanced detection data includes 49,640 detection points, the target sampling number can be set to 12,000 detection points.

[0161] From the set of test points to be sampled, test points whose structural component confidence levels meet the preset initial selection criteria are selected as the initial selected test points, thus obtaining the selected test point set; simultaneously, test points not included in the selected test point set are the remaining test points to be sampled. The preset initial selection criteria can be set as follows: test points whose structural component confidence levels rank in the top 1%-5% are selected as the initial selected test points; or test points whose structural component confidence levels are not less than 0.85 are selected as the initial selected test points.

[0162] Calculate the minimum distance between each detection point in the remaining set of unsampled detection points and the selected set of detection points to obtain the spatial coverage distance, and calculate the sampling priority value based on the spatial coverage distance and the confidence level of the structural components.

[0163] Detection points whose sampling priority values ​​meet the preset sorting criteria are added to the selected detection point set, and the remaining unsampled detection point set is updated. The spatial coverage distance calculation, sampling priority value calculation, and detection point addition process are repeated until the number of detection points in the selected detection point set reaches the target sampling quantity. The preset sorting criteria can be set as follows: the detection point with the highest sampling priority value is determined as the target detection point; or the top-ranked detection point with the highest sampling priority value is added to the selected detection point set sequentially. The sampling priority value is jointly determined by the spatial coverage distance and the structural component confidence level, where the weight of the spatial coverage distance can be set to 0.4-0.7, and the weight of the structural component confidence level can be set to 0.3-0.6.

[0164] The enhanced detection data corresponding to each detection point in the selected detection point set are combined to obtain sampled detection data. The sampled detection data retains the point coordinates, reflection intensity, and corresponding structural confidence of the detection points, which are used for subsequent local response field estimation, defect region registration, and auxiliary 3D reconstruction.

[0165] Example: Taking the aforementioned aluminum alloy coated structural component as an example, step S204 yields enhanced detection data corresponding to 49,640 detection points. The coordinates of each detection point, the enhanced reflection intensity, and the corresponding anomaly risk score are read.

[0166] For detection point P12864, its enhanced reflection intensity is 0.68, and its anomaly risk score is 0.74. Based on its neighboring detection points, the following are calculated: local normal consistency is 0.92, point density stability is 0.87, curvature smoothness is 0.84, reflection response consistency is 0.82, and scattering stability is 0.80. Combining this with its anomaly risk score, the structural confidence level for P12864 is 0.76.

[0167] For detection point P31042, its enhanced reflection intensity is 0.51, its anomaly risk score is 0.22, its local normal consistency is 0.95, its point density stability is 0.91, its curvature smoothness is 0.93, its reflection response consistency is 0.90, its scattering stability is 0.88, and its structural confidence level is 0.89.

[0168] In this example, the target sampling number is set to 12,000 detection points. First, detection points with high confidence in structural components are selected from 49,640 detection points as the initial selected detection points. Then, the minimum distance between the remaining detection points and the set of selected detection points is calculated, and the sampling priority value is calculated in combination with the confidence in structural components.

[0169] For example, although detection point P12864 has a high anomaly risk score, its structural component confidence score is 0.76, and its distance from the nearest detection point in the selected detection point set is 0.62 mm. Therefore, its sampling priority value meets the preset sorting conditions and it is added to the selected detection point set. Detection point P41720 has a structural component confidence score of 0.41, and its distance from the selected detection point set is 0.18 mm. It does not meet the preset sorting conditions and is not selected for this round of sampling.

[0170] Repeat the above process until the selected detection point set reaches 12,000 detection points. Finally, combine the enhanced detection data corresponding to these 12,000 detection points to obtain sampled detection data, which is used as input for the subsequent step S206.

[0171] Through the above technical solution, Embodiment 2 of the present invention solves the problem that downsampling processing is prone to losing key defect boundary points or retaining redundant low-value points. Based on enhanced detection data and abnormal risk scores, structural component confidence is generated, and weighted filtering sampling is performed according to the structural component confidence. This can both prioritize the retention of detection points with stable surface structure, reliable optical response and representativeness of defect boundaries, and reduce redundant detection points in flat areas or low-confidence areas, thereby obtaining sampling detection data that balances data compression and defect structure preservation.

[0172] Step S206: Based on the sampled detection data, perform local response field estimation, defect region registration and assisted three-dimensional reconstruction to obtain structural component reconstruction data. Extract structural component feature data based on the structural component reconstruction data, and output structural component defect analysis results based on the correspondence between structural component feature data and preset judgment conditions.

[0173] The process of obtaining structural component reconstruction data by performing local response field estimation, defect region registration, and assisted 3D reconstruction based on sampled detection data includes: Acquire sampling and detection data, including the point coordinates and reflection intensity of each detection point; A neighborhood detection point set is established based on the point coordinates, and a local weighted interpolation process is performed based on the reflection intensity and the neighborhood detection point set to obtain the local response field; Anomalies are identified based on the distribution of response values ​​at detection points in the local response field, and defect areas are determined based on the spatial connectivity between these anomalies. Obtain the coordinates of the boundary points of each defect region, and perform defect region registration processing based on the spatial correspondence between the boundary point coordinates and the point coordinates of each detection point in the sampled detection data to obtain defect region registration data; Surface reconstruction processing is performed based on the point coordinates of each detection point in the sampled detection data to obtain the surface mesh data of the structural component. Based on the surface mesh data of the structural component and the registration data of the defect region, the local response field of each defect region is mapped to the surface mesh data of the structural component to obtain the structural component reconstruction data.

[0174] The process of obtaining the local response field by performing local weighted interpolation based on reflection intensity and the set of neighboring detection points includes: Obtain the point coordinates and reflection intensity of the target detection point, as well as the point coordinates and reflection intensity of each neighboring detection point in the set of neighboring detection points corresponding to the target detection point; Distance weight data is calculated based on the point distance between the target detection point and each neighboring detection point. Distance weight data is calculated based on the point distance between the target detection point and each neighboring detection point. The weighted sum of reflection intensity is calculated based on the distance weight data of the reflection intensity of each neighboring detection point, and the local response value of the target detection point is obtained based on the ratio of the weighted sum of reflection intensity to the sum of distance weight data. The local response values ​​of each target detection point are combined in the order of the detection points to obtain the local response field.

[0175] Specifically, the sampling and detection data obtained in step S205 is acquired. This data includes the point coordinates, reflection intensity, and structural confidence level of each sampling and detection point.

[0176] Based on the coordinates of each sampling detection point, a neighborhood detection point set is established. Then, based on the reflection intensity of each detection point in the neighborhood detection point set, local weighted interpolation processing is performed to obtain local response field data that characterizes the continuous variation of reflection intensity in the local spatial range within the sampling detection data.

[0177] Based on the distribution of response values ​​at each sampling detection point in the local response field data, sampling detection points whose response values ​​meet preset abnormal response conditions are selected to obtain response anomaly detection points. The preset abnormal response conditions include that the deviation of the response value from the mean of the neighborhood response is not less than a preset response deviation threshold. The preset abnormal response conditions can be set as follows: the deviation of the local response value of the sampling detection point from the mean of the neighborhood response is not less than 0.15-0.25, or this deviation is not less than twice the standard deviation of the neighborhood response.

[0178] Based on the spatial connectivity between response anomaly detection points, interconnected response anomaly detection points that meet a preset threshold for the number of regional points are grouped into defect region data. Isolated response anomaly detection points that do not meet the preset threshold are not used to generate defect region data. The preset threshold for the number of regional points can be set to 20-50 response anomaly detection points. For high-resolution scan data, it can be set to 30 response anomaly detection points.

[0179] Obtain the coordinates of the boundary points corresponding to each defect area data, and perform defect area registration processing based on the spatial correspondence between the boundary point coordinates and the coordinates of each sampling detection point in the sampling detection data to obtain defect area registration data that represents the correspondence between the defect area data and the spatial position on the surface of the structural component.

[0180] Surface reconstruction processing is performed based on the point coordinates of each sampling and detection point in the sampling and detection data to obtain the surface mesh data of the structural component; then the defect area registration data and local response field data are mapped to the surface mesh data of the structural component to obtain the structural component reconstruction data.

[0181] Structural feature data is extracted based on the reconstructed structural component data. This structural feature data includes at least one of the following: location, area, boundary length, reflectivity deviation, scattering variation, defect boundary continuity, and surface roughness auxiliary parameters of the defect region.

[0182] The structural component feature data is matched with preset judgment conditions to determine the defect type and defect level corresponding to the defect area, and the structural component defect analysis results are output. The structural component defect analysis results include defect location, defect type, defect range, defect level, and corresponding optical property deviation. Preset judgment conditions may include: when the defect area is elongated and continuous, with a boundary continuity of not less than 0.75 and a reflectivity deviation of not less than 0.15, it is judged as a scratch or crack defect; when the defect area is closed or nearly closed and the local depth change is not less than 0.05 mm, it is judged as a hole or pit defect; when the surface roughness auxiliary parameter of the defect area is not less than 0.03 mm and the scattering change is not less than 0.20, it is judged as a corrosion or roughness damage defect; when the reflectivity deviation and polarization response difference both exceed the corresponding thresholds, it is judged as a coating peeling or contamination adhesion defect.

[0183] A preferred example of the present invention is the aforementioned aluminum alloy coated structural component, wherein step S205 yields sampling and detection data corresponding to 12,000 sampling and detection points. Each sampling and detection point includes point coordinates, reflection intensity, and structural component confidence level.

[0184] Centered on each sampling detection point, a set of detection points within a radius of 0.8 mm is selected as the neighborhood detection point set, and local weighted interpolation is performed based on the reflection intensity of the neighborhood detection points to obtain the local response field data.

[0185] In the local response field data, a set of response anomaly detection points were detected, whose reflection intensity response deviated from the mean of the neighborhood response by more than 0.18, and these response anomaly detection points were spatially continuous. Further analysis determined that this continuous region contained 146 response anomaly detection points, exceeding the preset region point count threshold of 30, and therefore it was identified as a defect region.

[0186] The coordinates of the boundary points of the defect area data are obtained, and spatial correspondence is established between these boundary point coordinates and the coordinates of the sampling detection points in the sampling detection data to obtain the defect area registration data. Subsequently, surface reconstruction is performed based on the point coordinates of 12,000 sampling detection points to obtain the surface mesh data of the structural component. The defect area registration data and local response field data are then mapped onto the surface mesh data of the structural component to obtain the reconstructed structural component data.

[0187] Based on the reconstructed data of this structural component, the structural feature data of the defect area were extracted: the defect center position is (126.7mm, 75.1mm, 2.14mm), the defect area is 5.8mm², the boundary length is 12.6mm, the reflectivity deviation is 0.24, the defect boundary continuity is 0.86, and the surface roughness auxiliary parameter is 0.031mm.

[0188] The structural component feature data was matched with preset judgment conditions. Since the defect area has a continuous, slender boundary, a reflectivity deviation higher than the preset reflectivity deviation threshold, and a boundary continuity higher than the preset boundary continuity threshold, the defect area was classified as a scratch-type defect, with a medium defect level. The structural component defect analysis results were output: defect location (126.7mm, 75.1mm, 2.14mm), defect type: scratch, defect range: 5.8mm², defect level: medium, corresponding optical characteristic deviation: reflectivity deviation 0.24.

[0189] Through the above technical solution, Embodiment 2 of the present invention solves the problem that the determination of defect location, range and level is unstable when relying solely on discrete detection points or single reconstruction results in defect analysis. Based on the sampled detection data, it performs local response field estimation, defect region registration and assisted three-dimensional reconstruction, and extracts structural feature data from the structural component reconstruction data. It can not only correspond discrete optical response with the spatial position of the structural component surface, but also make a comprehensive judgment by combining the geometric shape and optical characteristics of the defect region, thereby outputting structural component defect analysis results that include defect location, type, range and level.

[0190] Example 3 Figure 3 This is a schematic diagram of the structure of the electronic device according to Embodiment 3 of the present invention. Figure 3 The preferred embodiment of the electronic device of the present invention shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this application.

[0191] like Figure 3 As shown, the electronic device is presented in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: one or more processors 310, memory 330, and communication bus 340 connecting different system components (including memory 330 and processing unit 310).

[0192] Communication bus 340 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, Industry Standard Architecture (ISA) buses, Micro Channel Architecture (MAC) buses, Enhanced ISA buses, Video Electronics Standards Association (VESA) local buses, and Peripheral Component Interconnect (PCI) buses.

[0193] Electronic devices typically include a variety of computer-readable media. These media can be any available media that can be accessed by the electronic device, including volatile and non-volatile media, and removable and non-removable media.

[0194] Memory 330 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The electronic device may further include other removable / non-removable, volatile / non-volatile computer system storage media. Memory 330 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions and steps of the embodiments of this application.

[0195] A program / utility having a set (at least one) of program modules can be stored in memory 930. Such program modules include—but are not limited to—an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. The program modules typically perform the functions and / or methods described in the embodiments of this application.

[0196] Processor 310 executes various functional applications and data processing by running programs stored in memory 330, such as implementing embodiments of this application. Figure 1 The processing method provided in the illustrated embodiment.

[0197] This application provides a non-transitory computer-readable storage medium that stores computer instructions, which cause the computer to execute embodiments of this application. Figure 1 The processing method provided in the illustrated embodiment.

[0198] The aforementioned computer-readable storage medium may be any combination of one or more computer-readable media. A computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium may be any tangible medium that contains or stores a program that may be used by or in connection with an instruction execution system, apparatus, or device.

[0199] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit programs for use by or in connection with an instruction execution system, apparatus, or device.

[0200] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0201] Computer program code for performing the operations of the embodiments of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0202] In this embodiment of the invention, to further reveal the beneficial effects of the invention, the following comparison of beneficial effects is provided: Comparative test setup: Scheme M0 (baseline): After the original multi-source point clouds are directly stitched together, a conventional DGCNN network is used for defect classification and region determination, without executing the processing flow of this invention.

[0203] Solution M1 (traditional denoising): Multi-source ICP registration + statistical outlier removal (SOR, k=16, α=1.5) + DGCNN judgment.

[0204] Option M2: Frequency domain directional interference suppression is not performed in the processing steps of this invention.

[0205] Option M3: Weighted filtering sampling is not performed in the processing steps of this invention.

[0206] Method M4: No three-dimensional reconstruction is performed in the processing steps of this invention.

[0207] Method M5 (Complete solution of the present invention).

[0208] Evaluation indicators: Defect detection rate R_d (Recall): Number of correctly detected labeled defects / Total number of labeled defects False Alarm Rate R_f: Number of falsely reported defective areas / Total number of areas identified as defective Defect location error E_loc: The average three-dimensional distance (mm) between the detected defect center and the marked center. Defect type identification accuracy (Acc_cls): The percentage of samples that correctly identify the defect type. Overall accuracy (Acc_all): The percentage of samples that are both correctly classified and correctly located. Repeatability C_r: The consistency rate of judgment among three independent scans of the same workpiece. Multi-condition stability S_env: Normalized inverse of the standard deviation of Acc_all under five noise / disturbance conditions. The improvement in defect analysis accuracy achieved by this invention is described below: Overall accuracy comparison (480 actual workpieces, five types of defects), see Table 1 below.

[0209] Table 1 Comparison of Core Accuracy Indicators for Each Test Scheme

[0210] Improvement ratio relative to the baseline scheme (M5 Complete scheme of this invention) Improvement in defect detection rate R_d: (94.7 71.3) / 71.3 = 32.8% Conclusion: The defect detection rate of the present invention is improved by 32.8% compared with the baseline test method.

[0211] The false alarm rate R_f decreases: (18.6 4.3) / 18.6 = 76.9% Conclusion: The false alarm rate was reduced by 76.9%, significantly reducing invalid manual verification.

[0212] Defect location error E_loc decreased: (1.42 0.31) / 1.42 = 78.2% Conclusion: Defect location accuracy improved (error reduced) by 78.2%.

[0213] Improved Acc_cls accuracy for defect type identification: (91.6 68.5) / 68.5 = 33.7% Conclusion: The accuracy of defect type identification was improved by 33.7%.

[0214] Overall accuracy (Accc_all) improved: (89.3 62.8) / 62.8 = 42.2% Conclusion: The overall accuracy of structural component defect analysis was improved by 42.2%.

[0215] Compared to the traditional noise reduction testing solution M1, the overall accuracy is improved: (89.3 69.6) / 69.6 = 28.3% Conclusion: Compared with the "registration + SOR + DGCNN" process commonly used in industrial settings, the overall accuracy is improved by 28.3%.

[0216] The comparison of accuracy by defect type (M5 complete solution vs. M0 baseline) is shown in Table 2 below.

[0217]

[0218] Table 2 Comparison of Classification Detection Rate and Identification Accuracy Note: The improvement in microcracks was the largest (+44.8%), mainly due to the suppression of measurement fringes in the bend fillet region by frequency domain directional perturbation, and the ability of 3D reconstruction to recover subtle geometric anomalies.

[0219] A comparison of the accuracy of different detection areas (M5 complete solution) is shown in Table 3 below.

[0220] Table 3. Accuracy of Comprehensive Judgment of Key Areas for Channel-Shaped Structural Components (Acc_all) Conclusion: The present invention significantly improves the accuracy of traditionally difficult areas for inspection, such as deep cavities, rounded corners, and inner walls of holes in slotted structural parts, with an improvement of up to 64.5% in the area of ​​rounded corners at bends.

[0221] The foregoing has described specific embodiments of the present invention. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0222] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of the different embodiments or examples.

[0223] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of embodiments of the present invention, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0224] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of the invention pertain.

[0225] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0226] In the embodiments provided in this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, modules, or units, and may be electrical, mechanical, or other forms.

[0227] Furthermore, in the various embodiments of the present invention, the functional units can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0228] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the implementation methods of the present invention, and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the present invention.

Claims

1. A method for processing optical data on the surface of industrial structural components, characterized in that, include: The detection data from at least one optical detection data source on the surface of industrial structural parts are fused to obtain their point-level feature vectors. Based on the abnormal risk assessment value of the point-level feature vector in the risk assessment model, denoising processing is performed to obtain the first denoised detection data; Suppression processing is performed on the spectral domain components of the risk frequency band data in the spectral domain of the first denoised detection data to obtain the second detection data; The confidence level of the industrial structural component is generated based on the abnormal risk assessment value of the detection points in the second detection data and weighted filtering sampling is performed based on the confidence level of the structural component to obtain the third detection data. Based on the third detection data, local response field estimation, defect region registration, and three-dimensional reconstruction are performed to obtain the reconstructed data of the structural component. The structural component feature data in the reconstructed data are then extracted to generate the structural component defect analysis results.

2. The method for processing optical data on the surface of industrial structural components according to claim 1, characterized in that, The process of fusing detection data from at least one optical detection data source on the surface of industrial structural components to obtain their point-level feature vectors includes: Preprocessed detection data of at least one detection point from an optical detection data source and its corresponding neighboring detection points are acquired, unified to the same coordinate system, and the detection data is registered and aligned. This includes the point coordinates, reflection intensity, acquisition time, scanning pose, multi-angle light intensity response, polarization response, and transmission attenuation information of each detection point. Based on the point coordinates, we obtain the average neighborhood distance, which characterizes the spatial dispersion of the detection point's neighborhood; the normal consistency, which characterizes the orientation consistency of the detection point's neighborhood surface; the curvature change, which characterizes the curvature change of the detection point's neighborhood surface; and the plane fitting residual, which characterizes the deviation of the detection point from the local plane of its neighborhood. Based on the reflection intensity, a reflection intensity jump is obtained that characterizes the degree of abrupt change in the reflection intensity of the detection point relative to the neighborhood reflection intensity. Based on the multi-angle light intensity response, polarization response, and transmission attenuation information, the following are obtained: scattering dispersion, which characterizes the degree of dispersion of the multi-angle light intensity response of the detection point; multi-angle response consistency, which characterizes the degree of matching between the multi-angle light intensity response of the detection point and that of neighboring detection points; polarization response difference, which characterizes the degree of deviation between the polarization response of the detection point and that of neighboring detection points; and transmission attenuation difference, which characterizes the degree of deviation between the transmission attenuation of the detection point and that of neighboring detection points. The average neighborhood distance, normal consistency, curvature change, plane fitting residual, reflection intensity jump, scattering dispersion, multi-angle response consistency, polarization response difference, and transmission attenuation difference are combined to obtain a point-level feature vector.

3. The method for processing optical data on the surface of industrial structural components according to claim 1, characterized in that, The process of assessing abnormal risk values ​​in the risk assessment model based on the point-level feature vectors includes: The point-level feature vector is input into a pre-constructed risk assessment model to obtain its first risk output. Based on the first risk output, point-level loss data for the detection points is generated, and the tail sample set is determined based on the numerical distribution of the point-level loss data. The conditional risk value is calculated based on the point-level loss data in the tail sample set, and gradient backpropagation is performed on the risk assessment model based on the conditional risk value to obtain the risk gradient response of the detection point. Conditional Value at Risk (VaR) is satisfied: ; in, Indicates the quantile parameter is Conditional Value at Risk Represents the tail sample set The number of samples in This represents the point-level loss value corresponding to the i-th tail sample, where i is a positive integer; The abnormal risk assessment value of the detection point is obtained by weighted summation of the first risk output and the risk gradient response based on preset weights.

4. The method for processing optical data on the surface of industrial structural components according to claim 3, characterized in that, The process of generating point-level loss data for detection points based on the first risk output, and determining the tail sample set based on the numerical distribution of the point-level loss data, includes: Determine whether the first risk output is within the preset risk range; If it is located, then perform loss mapping processing to obtain the point-level loss value of that detection point; If it is not located, first perform boundary truncation processing, and then perform loss mapping processing based on the first risk output after truncation to obtain the point-level loss value of the detection point; The point-level loss value satisfies: ; in, This represents the point-level loss value corresponding to the i-th detection point. This represents the first risk output value corresponding to the i-th detection point. Indicates the preset smoothing parameters; The point-level loss values ​​are sorted in ascending order to determine the loss quantile threshold of the preset quantile parameter; the detection points with point-level loss values ​​not less than the loss quantile threshold are determined as tail samples and a tail sample set is formed.

5. The method for processing optical data on the surface of industrial structural components according to claim 1, characterized in that, The process of performing suppression processing on the spectral domain components of the risk frequency band data of the first denoised detection data in the spectral domain to obtain the second detection data includes: Perform a nearest neighbor search on the detection points in the first denoised detection data to establish the connection relationship between the detection points; The edge weights of the connecting edges are determined based on the spatial distance difference, reflection intensity difference, polarization response difference, and transmission attenuation difference between the detection points, resulting in a nearest neighbor weighted graph. Edge weights satisfy: ; in, This represents the edge weight between the i-th detection point and the j-th detection point. Indicates the difference in spatial distance. This indicates poor reflection intensity. Indicates the difference in polarization response. Indicates the difference in transmission attenuation. , , and These represent the scale parameters for the corresponding difference terms; Arrange the reflection intensity of the detection points in the first denoised detection data according to the order of the detection points to obtain the image signal; A graph Laplacian matrix is ​​generated based on a nearest-neighbor weighted graph, and eigenvalue decomposition is performed on the graph Laplacian matrix to obtain an eigenvector matrix. Based on the eigenvector matrix, the graph signal is transformed from the spatial domain to the spectral domain to obtain spectral domain detection data. The transformation of the graph signal from the spatial domain to the spectral domain satisfies: Where L represents the graph Laplacian matrix, D represents the degree matrix, W represents the edge weight matrix, and x represents the graph signal in the spatial domain. Let U represent the spectral domain detection data, T represent the eigenvector matrix, and T be the transpose operation. Obtain the detection points whose abnormal risk assessment value is not less than the preset risk threshold and are retained, and obtain the retained risk detection points; Based on the node connection relationship of the retained risk detection points in the nearest neighbor weighted graph, the adjacency distribution corresponding to the retained risk detection points is determined; Spectral domain response of statistical spectral domain detection data at each frequency component; Based on the adjacency distribution of the retained risk detection points in the nearest neighbor weighted graph, the risk projection weight corresponding to each frequency component is determined. The risk projection weights and spectral response are normalized, and the product of the normalized risk projection weights and spectral response is calculated to obtain the frequency risk correlation value of each frequency component. Frequency components with a frequency risk correlation value not less than a preset frequency band determination threshold are identified as risk frequency components, and risk frequency components with continuous frequency positions are combined to obtain risk frequency band data. The filtering coefficients of the spectral domain filter at each frequency component are determined based on the risk frequency band data. The spectral domain detection data is multiplied by the corresponding filter coefficients at each frequency component to obtain the suppressed spectral domain detection data. The eigenvector matrix based on the graph Laplacian matrix is ​​used to inversely transform the suppressed spectral domain detection data to the spatial domain to obtain the second detection data. The inverse transform satisfies: ; in, This represents the graph signal in the spatial domain of the enhanced detection data, where U represents the feature vector matrix. Represents an eigenvalue diagonal matrix. Let represent the spectral domain filter, x represent the spatial domain graph signal, and T represent the transpose operation.

6. The method for processing optical data on the surface of industrial structural components according to claim 1, characterized in that, The process of generating the confidence level of the industrial structural component based on the anomaly risk assessment value of the detection points in the second detection data and performing weighted filtering sampling based on the confidence level of the structural component to obtain the third detection data includes: The local normal consistency, point density stability, curvature smoothness, reflection response consistency, and scattering stability of the detection points in the second detection data are obtained and then summed using weighted averages. The weighted summation result is corrected based on the aforementioned abnormal risk assessment value to obtain the confidence level of the structural component; The detection points in the second detection data are used as the set of detection points to be sampled, and the target sampling number is determined based on the preset sampling number. Select the structural component confidence level that meets the preset initial selection conditions from the set of sampling detection points as the initial selected detection points, and obtain the set of selected detection points; Based on the detection points in the set of detection points to be sampled that have not been included in the set of selected detection points, the set of remaining detection points to be sampled is obtained; Calculate the minimum distance between each detection point in the remaining set of unsampled detection points and the selected set of detection points, and calculate the sampling priority value in combination with the corresponding structural confidence level; Add the detection points whose sampling priority values ​​meet the preset sorting conditions to the selected detection point set until the number of detection points in the selected detection point set reaches the target sampling number; The third detection data is obtained based on the second detection data corresponding to each detection point in the selected detection point set.

7. The method for processing optical data on the surface of industrial structural components according to claim 1, characterized in that, The process of estimating the local response field, registering the defect region, and reconstructing three dimensions based on the third detection data to obtain the reconstructed data of the structural component includes: Acquire third detection data, which includes the point coordinates and reflection intensity of each detection point; Establish a neighborhood detection point set based on point coordinates; Obtain the point coordinates and reflection intensity of the target detection point, as well as the point coordinates and reflection intensity of each neighboring detection point in the set of neighboring detection points corresponding to the target detection point; Distance weight data is calculated based on the point distance between the target detection point and each neighboring detection point. The weighted sum of reflection intensity is calculated based on the distance weight data of the reflection intensity of each neighboring detection point, and the local response value of the target detection point is obtained based on the ratio of the weighted sum of reflection intensity to the sum of distance weight data. The local response values ​​of each target detection point are combined in the order of the detection points to obtain the local response field; Anomalies are identified based on the distribution of response values ​​at detection points in the local response field, and defect areas are determined based on the spatial connectivity between these anomalies. Obtain the coordinates of the boundary points of each defect region, and perform defect region registration processing based on the spatial correspondence between the boundary point coordinates and the point coordinates of each detection point in the sampled detection data to obtain defect region registration data; Surface reconstruction processing is performed based on the point coordinates of each detection point in the sampled detection data to obtain the surface mesh data of the structural component. Based on the surface mesh data of the structural component and the registration data of the defect region, the local response field of each defect region is mapped to the surface mesh data of the structural component to obtain the reconstructed data of the structural component.

8. An electronic device, characterized in that, include: At least one processor; And at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor invokes the program instructions to perform the processing method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause the computer to perform the processing method as described in any one of claims 1 to 7.

Images