Point cloud denoising method for aerospace part detection

Abstract

The application discloses a point cloud denoising method for aviation part detection, and belongs to the technical field of three-dimensional point cloud data processing, which comprises the following steps: obtaining original point cloud formed by three-dimensional optical scanning of an aviation part; preprocessing the original point cloud; extracting spatial coordinate features and local geometric features through a multi-scale feature encoder to generate point cloud feature weights; performing adaptive differentiable pooling sampling based on the point cloud feature weights to extract backbone sampling points; inferring a second-order continuous differentiable local manifold through a decoder and projecting reflected noise points back to the real physical surface of the aviation part; and performing consistent resampling on the point cloud after projection correction to obtain high-fidelity point cloud. The application has the effects of being capable of inhibiting secondary reflection noise in a chamfer area and background interference at a hole, maintaining R-angle and hole boundary geometric features, and improving the point cloud measurement precision of the aviation part.

Description

Technical Field

[0001] This application relates to the field of three-dimensional point cloud data processing technology, specifically a point cloud denoising method for the inspection of aerospace components. Background Technology

[0002] With the development of aerospace manufacturing technology, the requirements for machining and inspection accuracy of aerospace components such as aero-engine bladed disks, fuselage ring components, and large thin-walled bearings are constantly increasing. These components are typically characterized by large size ranges, complex structures, and numerous local details. Their surfaces often include chamfers, radius corners, precision grooves, weight-reduction holes, positioning holes, and complex irregular curved surfaces. To achieve efficient, non-contact inspection, 3D optical scanning equipment has been widely used to acquire point cloud data of aerospace component surfaces, and further used for the inspection of radius corners, hole positions, and geometric tolerances.

[0003] However, in actual scanning processes, complex noise can easily be mixed into the point cloud of aerospace components. On the one hand, aerospace components often use highly reflective metallic materials. When scanning chamfered areas, concave areas, or narrow grooves, the scanning beam is prone to multipath interference and secondary reflections, thus forming drift points or reflective noise clouds near the real surface, obscuring the real geometric boundaries. On the other hand, when the scanning beam passes through pore structures such as weight reduction holes and positioning holes, it may also collect the support plane or worktable surface below the component, forming background interference points that affect the identification of hole boundaries and hole positions.

[0004] Existing point cloud denoising methods, such as statistical filtering, radius filtering, and bilateral filtering, typically rely on fixed thresholds or local distance relationships to remove noise. While these methods can remove some outliers, they are prone to two types of problems when dealing with precision structures such as chamfers, radius corners, grooves, and hole boundaries of aerospace components: if the filtering strength is weak, secondary reflection noise and background interference are difficult to remove sufficiently; if the filtering strength is strong, the true geometric boundaries are easily over-smoothed, leading to blunting or distortion of measurement features such as radius corners and hole boundaries.

[0005] Therefore, it is necessary to provide a point cloud denoising method suitable for the inspection of aerospace components, which can adaptively distinguish between real surface points, reflected noise points and background interference points, and maintain the clarity of geometric features such as chamfers, R-angles, grooves and hole boundaries while removing complex noise, thereby improving the accuracy of subsequent dimensional inspection and geometric tolerance inspection. Summary of the Invention

[0006] This application provides a point cloud denoising method for the inspection of aerospace components, in order to solve the technical problem that existing methods are unable to simultaneously remove reflection noise and preserve precise geometric features in the point cloud of aerospace components, thereby affecting the accuracy of subsequent dimensional and geometric tolerance inspection.

[0007] According to one aspect of this application, a point cloud denoising method for the inspection of aerospace components is provided, comprising the following steps: S1, acquiring the original point cloud formed by three-dimensional optical scanning of the aerospace components, the original point cloud including points on the surface of the aerospace components, reflected noise points formed by secondary reflection from the chamfered area or concave area of ​​the aerospace components, and background interference points formed by the through-scan support plane at the pores of the aerospace components. S2, preprocess the original point cloud to obtain a preprocessed point cloud; S3. Input the preprocessed point cloud into the multi-scale feature encoder. Extract the spatial coordinate features and local geometric features of the preprocessed point cloud through the multi-scale feature encoder. Based on the spatial coordinate features and local geometric features, identify the difference between the real geometric manifold features and noise distribution features of the aerospace parts and generate point cloud feature weights. The point cloud feature weights are used to characterize the probability that the corresponding point belongs to the real physical surface of the aerospace parts. S4, based on the point cloud feature weights, performs adaptive differentiable pooling sampling, reduces the sampling weights of background interference points and reflection noise points, and extracts backbone sampling points from the preprocessed point cloud; S5, input the backbone sampling points into the decoder, and infer the second-order continuously differentiable local manifold in the neighborhood of the backbone sampling points through the decoder; S6. Determine the displacement vector of the reflected noise point relative to the real physical surface of the aerospace component based on the local manifold, and project the reflected noise point back to the real physical surface of the aerospace component based on the displacement vector; suppress background interference points based on the point cloud feature weights to restore the geometric boundaries of the chamfered area, concave area and / or hole boundary area of ​​the aerospace component, and obtain the projection correction point cloud; S7. Perform consistent resampling on the projection correction point cloud to obtain a high-fidelity point cloud for aerospace component inspection.

[0008] Optionally, in step S2, preprocessing includes data format conversion, normal vector calculation, and initial downsampling.

[0009] Optionally, in step S2, the normal vector calculation includes: Determine the corresponding neighborhood point set for each point in the original point cloud; Construct a covariance matrix based on the three-dimensional coordinates of each neighboring point in the neighborhood point set; Eigenvalue decomposition is performed on the covariance matrix to obtain multiple eigenvalues ​​and eigenvectors corresponding to these eigenvalues. The eigenvector corresponding to the smallest eigenvalue is used as the normal vector of the corresponding point.

[0010] Optionally, in step S2, the initial downsampling includes at least one of voxel downsampling, farthest point sampling, and uniform sampling; Initial downsampling is used to reduce the amount of data in the original point cloud while preserving the geometric distribution of the large surface contours, chamfered areas, grooved areas, and hole boundary areas of aerospace components.

[0011] Optionally, in step S3, the multi-scale feature encoder includes a first-scale feature extraction branch, a second-scale feature extraction branch, and a third-scale feature extraction branch; The first scale feature extraction branch is used to perform neighborhood search on the preprocessed point cloud based on the number of first neighborhood points K1, and extract the first local point cloud features. The first local point cloud features are used to characterize the local geometric changes of the chamfer area, groove area and hole boundary area of ​​the aerospace parts. The second-scale feature extraction branch is used to perform neighborhood search on the preprocessed point cloud based on the number of second neighborhood points K2, and extract the second local point cloud features. The second local point cloud features are used to characterize the geometric continuity of the local curved surface region of the aerospace component. The third-scale feature extraction branch is used to perform neighborhood search on the preprocessed point cloud based on the number of third neighborhood points K3, and extract the third local point cloud features. The third local point cloud features are used to characterize the large-area outline and overall spatial structure of aerospace parts. K1 < K2 < K3; The multi-scale feature encoder fuses the first local point cloud features, the second local point cloud features, and the third local point cloud features to obtain fused point cloud features, and identifies the difference between the true geometric manifold features and noise distribution features of aerospace components based on the fused point cloud features.

[0012] Optionally, in step S3, the features extracted by the multi-scale feature encoder also include scan intensity features; Local geometric features include normal vector variation features and curvature features, and also include at least one of neighborhood distance variance features and point cloud density features; For points located in chamfered or concave regions that have abrupt changes in normal vectors, curvature, neighborhood distance variance, point cloud density, and / or scan intensity, the multi-scale feature encoder identifies them as candidate points for reflection noise. For points located in porous regions and with interlayer gaps in the depth direction between them and the point cloud of the main aerospace component, the multi-scale feature encoder identifies them as candidate points for background interference.

[0013] Optionally, in step S4, adaptive differentiable pooling sampling includes: An importance score for each point is generated based on the feature weights of the point cloud. Construct a differentiable soft assignment matrix based on importance scores; The preprocessed point cloud is sampled based on a differentiable soft allocation matrix to obtain backbone sampling points; The importance scores of background interference points and reflected noise points are lower than those of points on the surface of aerospace components.

[0014] Optionally, the differentiable soft assignment matrix is ​​generated based on the importance score of each point, and the matrix elements in the differentiable soft assignment matrix are used to characterize the weight of the corresponding point being sampled as a backbone sampling point, so that the adaptive differentiable pooling sampling process can be backpropagated through the loss function to adjust the sampling weight of each point in the preprocessed point cloud.

[0015] Optionally, in step S5, local neighborhood points are determined with the backbone sampling points as the center, and local embedding features are generated based on the spatial coordinate features, normal vector features, and curvature features of the backbone sampling points and local neighborhood points. The second-order continuously differentiable local manifold is then reconstructed based on the local embedding features by the decoder.

[0016] Optionally, in step S6, based on the positional deviation relationship between the reflected noise point and the local manifold, the displacement vector of the reflected noise point relative to the local manifold is predicted, and the reflected noise point is projected along the displacement vector to a position adjacent to the local manifold. In step S7, the projection correction point cloud is uniformly resampled so that the resampled high-fidelity point cloud maintains clear geometric boundaries and has uniform point cloud density in the chamfered area, groove area and hole boundary area of ​​aerospace parts.

[0017] In summary, this application includes at least one of the following beneficial technical effects: This scheme distinguishes between the real geometric structure and noise distribution in the point cloud through multi-scale feature encoding. This allows the algorithm to move beyond relying solely on fixed distance thresholds or local density judgments, combining spatial relationships and local geometric changes at different scales to identify the differences between real precision structures such as chamfers, concave surfaces, grooves, and hole boundaries, and reflection noise and background interference. Adaptive differentiable pooling sampling reduces the impact of noise and background points on subsequent reconstruction processes, ensuring that the retained backbone sampling points better represent the real physical surface of aerospace components. Furthermore, local manifold reconstruction and projection correction guide reflection noise points that deviate from the real surface back to the vicinity of the reasonable geometric surface, rather than simply deleting or averaging them, thus preventing the blunting of key measurement features such as radius corners and hole boundaries. Finally, consistent resampling improves the uniformity of the point cloud density, enabling the output point cloud to remove secondary reflection noise from chamfered and concave areas, as well as background interference at holes, while maintaining the clarity and continuity of the real geometric boundaries of aerospace components. This provides a more reliable data foundation for subsequent dimensional inspection, radius corner measurement, hole position detection, and geometric tolerance evaluation, thereby improving the accuracy of dimensional and geometric tolerance inspection.

[0018] In addition to the purposes, features, and advantages described above, this application has other purposes, features, and advantages. A further detailed description of this application will be provided below with reference to the figures. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a point cloud denoising method used for the inspection of aerospace components. Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1 This application will be described in further detail.

[0021] This application discloses a point cloud denoising method for the inspection of aerospace components, comprising the following steps: S1, acquire the original point cloud formed by three-dimensional optical scanning of the aerospace parts. The original point cloud includes points on the surface of the aerospace parts, reflection noise points formed by secondary reflection from the chamfered or concave areas of the aerospace parts, and background interference points formed by the through-scan support plane at the pores of the aerospace parts. S2, preprocess the original point cloud to obtain a preprocessed point cloud; S3. Input the preprocessed point cloud into the multi-scale feature encoder. Extract the spatial coordinate features and local geometric features of the preprocessed point cloud through the multi-scale feature encoder. Based on the spatial coordinate features and local geometric features, identify the difference between the real geometric manifold features and noise distribution features of the aerospace parts and generate point cloud feature weights. The point cloud feature weights are used to characterize the probability that the corresponding point belongs to the real physical surface of the aerospace parts. S4, based on the point cloud feature weights, performs adaptive differentiable pooling sampling, reduces the sampling weights of background interference points and reflection noise points, and extracts backbone sampling points from the preprocessed point cloud; S5, input the backbone sampling points into the decoder, and infer the second-order continuously differentiable local manifold in the neighborhood of the backbone sampling points through the decoder; S6. Determine the displacement vector of the reflected noise point relative to the real physical surface of the aerospace component based on the local manifold, and project the reflected noise point back to the real physical surface of the aerospace component based on the displacement vector; suppress the background interference points based on the point cloud feature weights to restore the geometric boundaries of the chamfered area, concave area and / or hole boundary area of ​​the aerospace component, and obtain the projection correction point cloud. S7 performs consistent resampling on the projected correction point cloud to obtain a high-fidelity point cloud for the inspection of aerospace components.

[0022] In this embodiment, the aerospace components can be aero-engine bladed disks, fuselage ring parts, large thin-walled bearings, or other precision parts with chamfered, rounded, concave, grooved, or hole structures. The aerospace components are scanned by a three-dimensional optical scanning device to obtain the original point cloud. Because aerospace components are typically made of highly reflective metallic materials, the scanning beam is prone to secondary reflection in chamfered, concave, or narrow groove areas, causing drift points or reflection noise points near the actual surface. When the scanning beam passes through gaps such as weight reduction holes or positioning holes, it may also capture the underlying support plane or the surface of the worktable, thus forming background interference points.

[0023] After acquiring the raw point cloud, it is preprocessed to meet the requirements of subsequent feature extraction and network processing. Then, the preprocessed point cloud is input into a multi-scale feature encoder. By analyzing the spatial coordinate relationships and local geometric changes at different scales, the encoder distinguishes between the real geometric manifold and noise distribution of aerospace components. Real surface points typically exhibit good geometric continuity, while reflected noise points often deviate from the real surface, and background interference points are usually spatially separated from the main point cloud of the aerospace component.

[0024] After obtaining the point cloud feature weights, adaptive differentiable pooling sampling is performed based on these point cloud feature weights, so that the real surface points have higher sampling weights, while the sampling weights of reflection noise points and background interference points are reduced, thereby extracting backbone sampling points that can represent the real surface morphology of aerospace parts.

[0025] The backbone sampling points are input into the decoder, which infers a second-order continuously differentiable local manifold within the neighborhood of the backbone sampling points. This local manifold is used to characterize the local geometry of the real physical surface of aerospace components. For reflection noise points that deviate from the real surface, the corresponding displacement vector is determined based on their positional relationship with the local manifold, and the reflection noise points are projected back to the vicinity of the real physical surface along this displacement vector. For background interference points formed by transmission scanning at apertures, they are suppressed according to the point cloud feature weights to reduce their impact on aperture boundary identification and subsequent measurements.

[0026] Finally, the projected point cloud is uniformly resampled to obtain a high-fidelity point cloud for aerospace component inspection. This high-fidelity point cloud can maintain the geometric boundary clarity of chamfered areas, concave areas, groove areas, and hole boundary areas while removing reflection noise and background interference, thus providing a reliable data foundation for subsequent inspection of radius corners, hole positions, groove contours, and geometric tolerances.

[0027] In one implementation, preprocessing includes data format conversion, normal vector calculation, and initial downsampling.

[0028] The original point cloud was acquired by a 3D optical scanning device. Different scanning devices may output different data formats, point cloud fields, and coordinate recording methods. Therefore, before entering the multi-scale feature encoder, the original point cloud is first converted into a unified point cloud data format that can be recognized and used by subsequent algorithms. After the data format conversion, normal vectors are calculated for the point cloud. Normal vectors reflect the orientation of local surfaces in the point cloud and provide geometric basis for subsequent identification of real surface points, reflected noise points, and background interference points. The data volume of scanned point clouds of aerospace components is usually large. Directly inputting all the original point clouds into the subsequent network would increase the computational burden and affect processing efficiency. Through initial downsampling, the number of point clouds can be reduced while preserving the overall contour and key geometric regions of the aerospace components, making the point cloud size more suitable for subsequent feature extraction, adaptive differentiable pooling sampling, and local manifold reconstruction.

[0029] In one implementation, step S2, the normal vector calculation includes: Determine the corresponding neighborhood point set for each point in the original point cloud; Construct a covariance matrix based on the three-dimensional coordinates of each neighboring point in the neighborhood point set; Eigenvalue decomposition is performed on the covariance matrix to obtain multiple eigenvalues ​​and eigenvectors corresponding to these eigenvalues. The eigenvector corresponding to the smallest eigenvalue is used as the normal vector of the corresponding point.

[0030] In this embodiment, normal vector calculation is used to obtain the orientation information of the local surface where each point in the point cloud is located. Specifically, for any point in the original point cloud, several neighboring points can be determined around that point using a neighborhood search method, thus forming a neighborhood point set. After determining the neighborhood point set, a covariance matrix is ​​constructed based on the three-dimensional coordinates of each neighboring point in the set. The covariance matrix is ​​used to characterize the distribution direction and dispersion of the neighboring points in three-dimensional space. For points located on the actual surface of the aerospace component, their neighboring points are usually distributed along the local surface, and the spatial distribution has a certain regularity; for reflection noise points formed by secondary reflection, their neighborhood distribution often deviates from the actual surface, and the dispersion is high; for background interference points, their neighboring points may form a planar distribution separated from the main point cloud of the aerospace component. Since the neighboring points are mainly distributed near the local surface, the direction corresponding to the minimum eigenvalue is usually the normal direction of the local surface, so the eigenvector corresponding to the minimum eigenvalue can be used as the normal vector of that point. In this way, the corresponding local surface orientation can be calculated for each point in the original point cloud. After obtaining the normal vector, the subsequent multi-scale feature encoder can use the normal vector information to determine the local geometric changes of the point cloud.

[0031] In step S2, the initial downsampling includes at least one of voxel downsampling, farthest point sampling, and uniform sampling; the initial downsampling is used to reduce the amount of data in the original point cloud and preserve the geometric distribution of the large surface contours, chamfered areas, groove areas, and hole boundary areas of the aerospace parts.

[0032] In one implementation, initial downsampling can employ voxel downsampling. Voxel downsampling divides the space containing the original point cloud into multiple voxel units and selects representative points within each voxel unit, thereby reducing the number of points in the cloud. Voxel downsampling allows the point cloud to maintain a relatively uniform spatial distribution and reduces the impact of excessively dense local point clouds on subsequent feature extraction.

[0033] In one implementation, initial downsampling can employ farthest-point sampling. Farthest-point sampling prioritizes spatially dispersed points, ensuring the sampled point cloud better covers the overall contour of the aerospace component. For aerospace components with large surface contours, concave structures, hole boundaries, and groove structures, farthest-point sampling helps maintain the spatial distribution relationships between different regions.

[0034] In one implementation, initial downsampling can employ uniform sampling. Uniform sampling selects points from the original point cloud at preset intervals or according to a preset sampling ratio to reduce the point cloud size and improve subsequent processing efficiency. In practical applications, voxel downsampling, farthest-point sampling, or uniform sampling can be selected based on the size of the aerospace component, the density of the scanned point cloud, and the accuracy requirements of subsequent detection; multiple methods can also be combined.

[0035] In one implementation, in step S3, the multi-scale feature encoder includes a first-scale feature extraction branch, a second-scale feature extraction branch, and a third-scale feature extraction branch; The first scale feature extraction branch is used to perform neighborhood search on the preprocessed point cloud based on the number of first neighborhood points K1, and extract the first local point cloud features. The first local point cloud features are used to characterize the local geometric changes of the chamfer area, groove area and hole boundary area of ​​the aerospace parts. The second-scale feature extraction branch is used to perform neighborhood search on the preprocessed point cloud based on the number of second neighborhood points K2, and extract the second local point cloud features. The second local point cloud features are used to characterize the geometric continuity of the local curved surface region of the aerospace component. The third-scale feature extraction branch is used to perform neighborhood search on the preprocessed point cloud based on the number of third neighborhood points K3, and extract the third local point cloud features. The third local point cloud features are used to characterize the large-area outline and overall spatial structure of aerospace parts. K1 < K2 < K3; The multi-scale feature encoder fuses the first local point cloud features, the second local point cloud features, and the third local point cloud features to obtain fused point cloud features, and identifies the difference between the true geometric manifold features and noise distribution features of aerospace components based on the fused point cloud features.

[0036] In this embodiment, the first-scale feature extraction branch performs a neighborhood search based on the number of first neighboring points K1. Since K1 is relatively small, the first-scale feature extraction branch focuses more on local detail changes in the point cloud, making it suitable for characterizing geometric changes near chamfered regions, groove regions, and hole boundary regions. By using the first local point cloud features, boundary change information at the precision structure of aerospace components can be better preserved, reducing the risk of the true boundary being mistakenly weakened during subsequent denoising.

[0037] The second-scale feature extraction branch performs a neighborhood search based on the number of second-neighbor points, K2. Since K2 is greater than K1, the second-scale feature extraction branch can reflect the surface changes and geometric continuity of the point cloud over a larger local area. For concave surfaces, transition surfaces, and areas near radius-angles of aerospace components, the second local point cloud features are helpful in determining whether the region conforms to the continuous change trend of the actual component surface, thus providing a more stable local geometric basis for subsequent manifold reconstruction.

[0038] The third-scale feature extraction branch performs a neighborhood search based on the number of third-neighbor points, K3. Since K3 is greater than K2, the third-scale feature extraction branch can characterize the large-area outline and overall spatial structure of aerospace components from a wider range. Through the third local point cloud features, it is helpful to distinguish the point cloud of the main body of the aerospace component from the background plane, outliers far from the main body, or other spatially separated point cloud distributions.

[0039] After obtaining the first, second, and third local point cloud features, the multi-scale feature encoder fuses the features from different scales to obtain fused point cloud features. The fused point cloud features simultaneously contain local details, surface continuity, and overall spatial structure information, so that the generation of subsequent point cloud feature weights no longer depends on single-scale judgment.

[0040] In one specific implementation, the number of first neighboring points K1 can be 8, the number of second neighboring points K2 can be 16, and the number of third neighboring points K3 can be 24. In this case, the first-scale feature extraction branch selects the 8 nearest neighboring points centered on each point to be processed, used to extract subtle geometric changes in chamfered areas, grooved areas, and hole boundary areas; the second-scale feature extraction branch selects the 16 nearest neighboring points centered on each point to be processed, used to extract the continuous change trend of local curved surface areas; and the third-scale feature extraction branch selects the 24 nearest neighboring points centered on each point to be processed, used to extract the large-area contours and overall spatial structure of aerospace components.

[0041] In one implementation, in step S3, the features extracted by the multi-scale feature encoder also include scan intensity features; Local geometric features include normal vector variation features and curvature features, and also include at least one of neighborhood distance variance features and point cloud density features; For points located in chamfered or concave regions that have abrupt changes in normal vectors, curvature, neighborhood distance variance, point cloud density, and / or scan intensity, the multi-scale feature encoder identifies them as candidate points for reflection noise. For points located in porous regions and with interlayer gaps in the depth direction between them and the point cloud of the main aerospace component, the multi-scale feature encoder identifies them as candidate points for background interference.

[0042] In this embodiment, the multi-scale feature encoder can extract scanning intensity features in addition to spatial coordinate features and local geometric features. Scanning intensity features reflect the reflection state of the scanning beam during point cloud acquisition. When the surface of aerospace components is made of highly reflective metallic materials, secondary reflections are prone to occur in chamfered areas, concave areas, or narrow grooves, resulting in differences in the scanning intensity, spatial position, and local geometric distribution of some points compared to the actual surface points. Local geometric features can include normal vector variation features, curvature features, neighborhood distance variance features, and point cloud density features. Normal vector variation features reflect the changes in surface orientation between neighboring points; curvature features reflect the degree of curvature of the local surface; neighborhood distance variance features reflect the dispersion of the distribution of neighboring points; and point cloud density features reflect the density of points within a local area.

[0043] For chamfered or concave regions, the real surface typically exhibits a continuously changing geometric structure; however, noise points formed by secondary reflections often deviate from the real surface and exhibit characteristics such as abrupt changes in normal vectors, curvature, abnormal neighborhood distance distribution, abnormal local density, or abnormal scan intensity. Multi-scale feature encoders can identify these points as candidate points for reflection noise based on these anomalous features. For aperture regions, after the scanning beam passes through weight reduction holes, positioning holes, or other structural gaps, it may acquire the support plane or worktable surface beneath the aerospace component. These points typically have interlayer gaps in the depth direction with the main point cloud of the aerospace component and may form a relatively continuous background plane distribution. Multi-scale feature encoders can identify these points as candidate points for background interference based on this spatial separation relationship. By combining local geometric features with scan intensity features, it is possible to more effectively distinguish between real surface points, candidate points for reflection noise, and candidate points for background interference of aerospace components, providing a reliable basis for subsequent generation of point cloud feature weights.

[0044] In one implementation, step S4, adaptive differentiable pooling sampling includes: An importance score for each point is generated based on the feature weights of the point cloud. Construct a differentiable soft assignment matrix based on importance scores; The preprocessed point cloud is sampled based on a differentiable soft allocation matrix to obtain backbone sampling points; The importance scores of background interference points and reflected noise points are lower than those of points on the surface of aerospace components.

[0045] In this embodiment, adaptive differentiable pooling sampling is used to extract backbone sampling points from the preprocessed point cloud that can represent the real physical surface of aerospace components. Since the preprocessed point cloud may still contain reflection noise points and background interference points, directly performing ordinary downsampling on the preprocessed point cloud can easily retain some noise points or background points, thus affecting the accuracy of subsequent local manifold reconstruction. An importance score for each point is generated based on the point cloud feature weights, enabling the sampling process to adaptively adjust according to the representational value of each point to the real physical surface.

[0046] For points located on the actual surface of aerospace components, their spatial coordinate features and local geometric features typically exhibit good continuity with the surrounding area, resulting in higher importance scores. For reflection noise points formed by secondary reflections in chamfered or concave areas, their spatial positions and local geometric changes usually deviate from the actual surface, leading to lower importance scores. Background interference points formed by scanning through apertures are spatially separated from the main point cloud of the aerospace component and are also assigned lower importance scores. After constructing a differentiable soft assignment matrix based on the importance scores of each point, the preprocessed point cloud can be sampled using this matrix. The soft assignment matrix reflects the weight relationship of different points selected as backbone sampling points, making it easier to retain actual surface points while reducing the sampling weights of reflection noise points and background interference points.

[0047] In one implementation, a differentiable soft assignment matrix is ​​generated based on the importance score of each point, and the matrix elements in the differentiable soft assignment matrix are used to characterize the weight of the corresponding point being sampled as a backbone sampling point, so that the adaptive differentiable pooling sampling process can be backpropagated through the loss function to adjust the sampling weight of each point in the preprocessed point cloud.

[0048] In this embodiment, after the multi-scale feature encoder generates point cloud feature weights based on the spatial coordinate features and local geometric features of each point in the preprocessed point cloud, it can further obtain the importance score of each point. The higher the importance score, the more likely the point belongs to the actual physical surface of the aerospace component; the lower the importance score, the more likely the point belongs to the reflection noise point or background interference point.

[0049] After generating a differentiable soft-assignment matrix based on the importance scores of each point, the elements in the matrix can represent the allocation relationship or sampling weight relationship between each point and the backbone sampling points. The sampling process is no longer a simple hard selection of points, but can express the contribution of each point to the backbone sampling points in the form of continuous weights. Real surface points, due to their higher importance scores, correspond to larger sampling weights in the soft-assignment matrix; reflection noise points and background interference points, due to their lower importance scores, correspond to smaller sampling weights in the soft-assignment matrix.

[0050] Since the sampling weights in the soft-assignment matrix are continuously generated based on importance scores, this sampling process can participate in the backpropagation of the loss function during model training. When subsequent local manifold reconstruction results, projection correction results, or resampling results deviate from the expected high-fidelity point cloud, the multi-scale feature encoder and related parameters in the sampling process can be adjusted inversely through the loss function to further increase the sampling weights of real surface points and further decrease the sampling weights of reflection noise points and background interference points. Thus, adaptive differentiable pooling sampling can not only extract backbone sampling points during forward processing but also automatically optimize sampling weights based on the denoising effect during training, improving the adaptability of the point cloud denoising model to different aerospace component structures and different noise distributions.

[0051] In one implementation, in step S5, local neighborhood points are determined with the backbone sampling points as the center, and local embedding features are generated based on the spatial coordinate features, normal vector features, and curvature features of the backbone sampling points and local neighborhood points. The second-order continuously differentiable local manifold is reconstructed based on the local embedding features by the decoder.

[0052] In this embodiment, the backbone sampling points are points obtained after adaptive differentiable pooling sampling, which can well characterize the real physical surface of aerospace components. To further recover the local geometry of aerospace components, corresponding local neighborhood points can be determined around each backbone sampling point. After determining the local neighborhood points, local embedding features are generated by combining the spatial coordinate features, normal vector features, and curvature features of the backbone sampling points and local neighborhood points. The spatial coordinate features are used to characterize the positional relationship of each point within the local neighborhood, the normal vector features are used to characterize the directional changes of the local surface, and the curvature features are used to characterize the bending changes of areas such as chamfers, concave surfaces, grooves, or hole boundaries.

[0053] The decoder reconstructs a second-order continuously differentiable local manifold based on local embedding features. The local manifold represents the continuous geometry of the real physical surface of aerospace components within a local region. For large surface areas, the local manifold can characterize relatively smooth surface structures; for chamfered regions, concave regions, groove regions, and hole boundary regions, the local manifold can characterize continuous surface curvature and boundary changes. By reconstructing the second-order continuously differentiable local manifold, the orientation and curvature changes of the local surface can be preserved during denoising, avoiding the simple smoothing away of real radius corners, groove boundaries, or hole boundaries.

[0054] In one implementation, in step S6, based on the positional deviation relationship between the reflected noise point and the local manifold, the displacement vector of the reflected noise point relative to the local manifold is predicted, and the reflected noise point is projected along the displacement vector to a position adjacent to the local manifold; in step S7, the projected correction point cloud is uniformly resampled so that the resampled high-fidelity point cloud maintains clear geometric boundaries and has uniform point cloud density in the chamfered area, groove area and hole boundary area of ​​aerospace parts.

[0055] In this embodiment, after the local manifold reconstruction is completed, it can be used as a local geometric reference for the actual physical surface of the aerospace component. For reflection noise points located in chamfered or concave regions, since they are usually formed by secondary reflection of the scanning beam, these points will be spatially offset relative to the actual surface.

[0056] The displacement vector represents the direction and amount of correction for reflected noise points to their proximity to the local manifold. Based on this displacement vector, reflected noise points deviating from the actual physical surface can be projected onto the vicinity of the local manifold, causing them to converge towards the actual surface of the aerospace component. Consistent resampling is used to adjust the point spacing and density distribution of the point cloud after projection correction, resulting in a more uniform point cloud density in different regions of the high-fidelity point cloud. For example, redundant points can be appropriately reduced in areas with excessively dense point clouds; in areas where the point cloud is sparse due to noise suppression or projection correction, compensatory sampling can be performed based on the local manifold, thereby improving the overall uniformity of the point cloud. After consistent resampling, the high-fidelity point cloud maintains clearer geometric boundaries in the chamfered, grooved, and hole boundary regions of aerospace components, and reduces the impact of background interference points and reflected noise points on subsequent measurements.

[0057] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A point cloud denoising method for the inspection of aerospace components, characterized in that, Includes the following steps: S1, acquire the original point cloud formed by three-dimensional optical scanning of the aerospace parts. The original point cloud includes points on the surface of the aerospace parts, reflection noise points formed by secondary reflection from the chamfered or concave areas of the aerospace parts, and background interference points formed by the through-scan support plane at the pores of the aerospace parts. S2, preprocess the original point cloud to obtain a preprocessed point cloud; S3. Input the preprocessed point cloud into the multi-scale feature encoder. Extract the spatial coordinate features and local geometric features of the preprocessed point cloud through the multi-scale feature encoder. Based on the spatial coordinate features and local geometric features, identify the difference between the real geometric manifold features and noise distribution features of the aerospace parts and generate point cloud feature weights. The point cloud feature weights are used to characterize the probability that the corresponding point belongs to the real physical surface of the aerospace parts. S4, based on the point cloud feature weights, performs adaptive differentiable pooling sampling, reduces the sampling weights of background interference points and reflection noise points, and extracts backbone sampling points from the preprocessed point cloud; S5, input the backbone sampling points into the decoder, and infer the second-order continuously differentiable local manifold in the neighborhood of the backbone sampling points through the decoder; S6. Determine the displacement vector of the reflected noise point relative to the real physical surface of the aerospace component based on the local manifold, and project the reflected noise point back to the real physical surface of the aerospace component based on the displacement vector; suppress the background interference points based on the point cloud feature weights to restore the geometric boundaries of the chamfered area, concave area and / or hole boundary area of ​​the aerospace component, and obtain the projection correction point cloud. S7 performs consistent resampling on the projected correction point cloud to obtain a high-fidelity point cloud for the inspection of aerospace components.

2. The point cloud denoising method for aircraft component inspection according to claim 1, characterized in that: In step S2, preprocessing includes data format conversion, normal vector calculation, and initial downsampling.

3. The point cloud denoising method for aircraft component inspection according to claim 2, characterized in that: In step S2, the normal vector calculation includes: Determine the corresponding neighborhood point set for each point in the original point cloud; Construct a covariance matrix based on the three-dimensional coordinates of each neighboring point in the neighborhood point set; Eigenvalue decomposition is performed on the covariance matrix to obtain multiple eigenvalues ​​and eigenvectors corresponding to these eigenvalues. The eigenvector corresponding to the smallest eigenvalue is used as the normal vector of the corresponding point.

4. The point cloud denoising method for the inspection of aerospace components according to claim 2, characterized in that: In step S2, the initial downsampling includes at least one of voxel downsampling, farthest point sampling, and uniform sampling; Initial downsampling is used to reduce the amount of data in the original point cloud while preserving the geometric distribution of the large surface contours, chamfered areas, grooved areas, and hole boundary areas of aerospace components.

5. The point cloud denoising method for aircraft component inspection according to claim 1, characterized in that: In step S3, the multi-scale feature encoder includes a first-scale feature extraction branch, a second-scale feature extraction branch, and a third-scale feature extraction branch; The first scale feature extraction branch is used to perform neighborhood search on the preprocessed point cloud based on the number of first neighborhood points K1, and extract the first local point cloud features. The first local point cloud features are used to characterize the local geometric changes of the chamfer area, groove area and hole boundary area of ​​the aerospace parts. The second-scale feature extraction branch is used to perform neighborhood search on the preprocessed point cloud based on the number of second neighborhood points K2, and extract the second local point cloud features. The second local point cloud features are used to characterize the geometric continuity of the local curved surface region of the aerospace component. The third-scale feature extraction branch is used to perform neighborhood search on the preprocessed point cloud based on the number of third neighborhood points K3, and extract the third local point cloud features. The third local point cloud features are used to characterize the large-area outline and overall spatial structure of aerospace parts. K1 < K2 < K3; The multi-scale feature encoder fuses the first local point cloud features, the second local point cloud features, and the third local point cloud features to obtain fused point cloud features, and identifies the difference between the true geometric manifold features and noise distribution features of aerospace components based on the fused point cloud features.

6. The point cloud denoising method for aircraft component inspection according to claim 1, characterized in that: In step S3, the features extracted by the multi-scale feature encoder also include scan intensity features; Local geometric features include normal vector variation features and curvature features, and also include at least one of neighborhood distance variance features and point cloud density features; For points located in chamfered or concave regions that have abrupt changes in normal vectors, curvature, neighborhood distance variance, point cloud density, and / or scan intensity, the multi-scale feature encoder identifies them as candidate points for reflection noise. For points located in porous regions and with interlayer gaps in the depth direction between them and the point cloud of the main aerospace component, the multi-scale feature encoder identifies them as candidate points for background interference.

7. The point cloud denoising method for aircraft component inspection according to claim 1, characterized in that: In step S4, adaptive differentiable pooling sampling includes: An importance score for each point is generated based on the feature weights of the point cloud. Construct a differentiable soft assignment matrix based on importance scores; The preprocessed point cloud is sampled based on a differentiable soft allocation matrix to obtain backbone sampling points; The importance scores of background interference points and reflected noise points are lower than those of points on the surface of aerospace components.

8. The point cloud denoising method for aircraft component inspection according to claim 7, characterized in that: The differentiable soft assignment matrix is ​​generated based on the importance score of each point, and the matrix elements in the differentiable soft assignment matrix are used to characterize the weight of the corresponding point being sampled as a backbone sampling point, so that the adaptive differentiable pooling sampling process can be backpropagated through the loss function to adjust the sampling weight of each point in the preprocessed point cloud.

9. The point cloud denoising method for aircraft component inspection according to claim 1, characterized in that: In step S5, local neighborhood points are determined with the backbone sampling points as the center. Local embedding features are generated based on the spatial coordinate features, normal vector features, and curvature features of the backbone sampling points and local neighborhood points. The second-order continuously differentiable local manifold is reconstructed by the decoder based on the local embedding features.

10. The point cloud denoising method for aircraft component inspection according to claim 9, characterized in that: In step S6, based on the positional deviation relationship between the reflected noise point and the local manifold, the displacement vector of the reflected noise point relative to the local manifold is predicted, and the reflected noise point is projected along the displacement vector to a position adjacent to the local manifold. In step S7, the projection correction point cloud is uniformly resampled so that the resampled high-fidelity point cloud maintains clear geometric boundaries and has uniform point cloud density in the chamfered area, groove area and hole boundary area of ​​aerospace parts.

Images