Airplane skin damage identification method, device, equipment, storage medium and product

Abstract

Embodiments of the present application provide an aircraft skin damage identification method, device, equipment, storage medium and product. The method comprises: obtaining an initial point cloud of the surface of the aircraft skin through a structured light device, and performing data preprocessing on the initial point cloud to obtain a preliminary corrected point cloud; calculating the reflection intensity distribution of each region of the surface of the aircraft skin based on the preliminary corrected point cloud, and determining a compensated point cloud; extracting geometric features from the compensated point cloud, and performing denoising processing on the compensated point cloud to obtain a denoised point cloud; constructing a three-dimensional grid model representing the three-dimensional structure of the surface of the aircraft skin based on the denoised point cloud, and determining a damage candidate area of the surface of the aircraft skin; obtaining a point cloud subset associated with the damage candidate area and classifying the damage type thereof to obtain a classification result of the damage type; generating a damage annotation map according to the classification result; and calibrating the damage size parameter based on the damage annotation map to generate a final damage report.

Description

Technical Field

[0001] This application relates to the aerospace field, and in particular to a method, apparatus, device, storage medium, and product for identifying aircraft skin damage based on structured light. Background Technology

[0002] Aircraft skin is the core external structure of an aircraft. Minor damage such as cracks and dents on its surface can lead to structural failure due to the expansion of airflow and pressure during flight, directly affecting flight safety. Moreover, the efficiency of inspection determines the aircraft maintenance cycle. Inefficient inspection will increase downtime and increase operating costs. Therefore, high-precision and high-efficiency inspection is a core requirement for aviation maintenance.

[0003] Currently, the mainstream inspection technologies fall into two categories: traditional manual inspection and structured light inspection. Traditional manual inspection relies on operators to check each point with gauges, which is time-consuming and affected by the operator's experience and fatigue, making it difficult to guarantee the consistency and accuracy of the inspection. Structured light inspection, as a non-contact optical technology, obtains three-dimensional point cloud data of the skin surface by projecting specific light patterns and capturing deformations. Although it solves the problem of low manual efficiency, it suffers from technical problems such as susceptibility to interference from skin reflections and inefficient processing of distorted point clouds. On the one hand, skins are mostly made of highly reflective aluminum alloys, and ambient light or the reflected light from the structured light itself can mix into the acquired signal, causing the point cloud coordinates of the structured light inspection to deviate from the true position and fail to reflect the actual shape of the skin. Since the point cloud is the basis for damage identification, distortion directly undermines accuracy. On the other hand, existing technologies lack effective means to process distorted point clouds. They neither specifically remove invalid data points caused by reflections and equipment errors, nor do they rely on manual assistance for damage classification and size quantification, making it impossible to achieve micron-level high-precision measurement. The two issues mentioned above make it difficult for existing skin damage detection technologies to simultaneously address both "anti-reflective interference" and "high-precision automated identification," thus failing to meet the actual safety and efficiency requirements of aviation maintenance. There is an urgent need for a technical solution that can systematically solve these problems. Summary of the Invention

[0004] This application provides a method, apparatus, device, storage medium, and product for identifying aircraft skin damage, which can improve the accuracy of aircraft skin damage identification.

[0005] In a first aspect, this application provides a method for identifying damage to aircraft skin, the method comprising the following steps: The initial point cloud of the aircraft skin surface is acquired using a structured light device, and the initial point cloud is preprocessed to obtain a preliminary corrected point cloud. Based on the preliminary corrected point cloud, the reflection intensity distribution of each region on the aircraft skin surface is calculated, and the compensated point cloud is determined based on the calculated reflection intensity and the preset reflection intensity threshold. Geometric features are extracted from the compensated point cloud, and spatial domain filtering is used to denoise the compensated point cloud based on the geometric features to obtain a denoised point cloud. Based on the denoised point cloud, a three-dimensional mesh model representing the three-dimensional structure of the aircraft skin surface is constructed, and the damage candidate area of ​​the aircraft skin surface is determined from the three-dimensional mesh model by the point cloud local geometric feature analysis method. Obtain a subset of point clouds associated with the candidate damage region, and use a deep learning model to classify the damage type of the subset of point clouds to obtain the classification result of the damage type; Based on the classification results, the boundary information of the damage candidate regions is fused to generate a damage annotation map; and Based on the damage annotation map, an error compensation model constructed using a parameter fitting optimization algorithm is used to calibrate the damage size parameters in order to generate a final damage report.

[0006] Secondly, this application provides an aircraft skin damage identification device, the device comprising: The point cloud acquisition and preprocessing module is used to acquire the initial point cloud of the aircraft skin surface through a structured light device, and to perform data preprocessing on the initial point cloud to obtain a preliminary corrected point cloud. The reflection intensity analysis and compensation module is used to calculate the reflection intensity distribution of each region on the aircraft skin surface based on the preliminary correction point cloud, and to determine the compensated point cloud based on the calculated reflection intensity and a preset reflection intensity threshold. The point cloud denoising module is used to extract geometric features from the compensated point cloud and use a spatial domain filtering method to denoise the compensated point cloud based on the geometric features to obtain a denoised point cloud. The damage area localization module is used to construct a three-dimensional mesh model representing the three-dimensional structure of the aircraft skin surface based on the denoised point cloud, and to determine the damage candidate area of ​​the aircraft skin surface from the three-dimensional mesh model through the point cloud local geometric feature analysis method. The damage classification module is used to obtain a subset of point clouds associated with the damage candidate region, and to use a deep learning model to classify the damage type of the subset of point clouds to obtain the classification result of the damage type. The damage annotation generation module is used to generate a damage annotation map by fusing the boundary information of the damage candidate regions based on the classification results; and The damage report generation module is used to calibrate the damage size parameters based on the damage annotation map using an error compensation model constructed by an optimization algorithm based on parameter fitting, so as to generate a final damage report.

[0007] Thirdly, this application provides an aircraft skin damage identification device, the device comprising: a processor, and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the aircraft skin damage identification method as described above.

[0008] Fourthly, this application provides a computer storage medium storing computer program instructions, which, when executed by a processor, implement the aircraft skin damage identification method described above.

[0009] Fifthly, this application provides a computer program product, including a computer program that, when processed for execution, implements the method for identifying aircraft skin damage as described above.

[0010] The aircraft skin damage identification method, apparatus, equipment, storage medium, and product implemented in this application can address the high reflectivity and high precision requirements of aircraft skin by solving data quality issues through data preprocessing (e.g., polarization filtering) combined with dynamic acquisition parameter adjustment strategies (e.g., adaptive illumination compensation), improving damage identification accuracy through three-dimensional geometric feature extraction combined with deep learning, and achieving dimensional quantification calibration through an error compensation model constructed based on parameter fitting optimization algorithms (e.g., least squares iterative optimization). This forms a complete damage identification scheme adapted to aviation scenarios, effectively making up for the shortcomings of existing technologies in the field of aircraft skin damage detection and further improving the accuracy of aircraft skin damage identification. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating an aircraft skin damage identification method according to an embodiment of this application is shown. Figure 2 A flowchart illustrating a method for obtaining compensated point clouds according to an embodiment of this application is shown. Figure 3 A schematic diagram of the structure of an aircraft skin damage identification device according to an embodiment of this application is shown; Figure 4 A schematic diagram of the hardware structure of an aircraft skin damage identification device according to an embodiment of this application is shown. Detailed Implementation

[0013] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0014] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0015] To address the problems in the prior art, embodiments of this application provide a method, apparatus, device, storage medium, and product for identifying aircraft skin damage.

[0016] The method for identifying aircraft skin damage provided in the embodiments of this application will be introduced below.

[0017] Figure 1 A schematic flowchart of an aircraft skin damage identification method according to an embodiment of this application is shown. Figure 1 As shown, the method may include the following steps: Step S102: Acquire an initial point cloud of the aircraft skin surface using a structured light device, and preprocess the initial point cloud to obtain a preliminary corrected point cloud. In one embodiment, the initial point cloud data is acquired by scanning the skin surface using a structured light device, and reflected light is filtered through polarization filtering to initially reduce reflective interference, thereby obtaining a preliminary corrected point cloud.

[0018] A digital micromirror array of a structured light device projects a coded stripe pattern onto the skin surface. A binocular camera simultaneously acquires images of the deformed stripes. The 3D coordinates of each pixel are calculated based on the mapping relationship between the stripe phase offset and depth value, yielding raw point cloud data including reflective noise. For this raw point cloud data, orthogonal polarizers are installed in front of the binocular camera lens to acquire the light intensity values ​​of the horizontal and vertical polarization components. The polarization degree value of each sampling point is calculated. When the polarization degree value exceeds a preset polarization threshold, the point is identified as a specular reflection point and marked, resulting in a polarization-marked point cloud. Based on the spatial distribution of non-reflective points in the polarization-marked point cloud, the average Euclidean distance between each point and its k nearest neighbors is calculated. When the average distance is greater than a preset outlier threshold, the outlier point is removed. For the void regions formed after removing reflective and outlier points, interpolation is performed using a weighted average of the depth values ​​of neighboring non-reflective points to fill the voids, thus determining the preliminary corrected point cloud.

[0019] Specifically, in one embodiment, the structured light device uses a digital micromirror array as the projection unit. This array contains 1024×768 micromirrors, each with independently controllable deflection angles. A coded fringe pattern combining Gray code and phase shift is generated by controlling the micromirror deflection state. The Gray code sequence determines the number of fringe periods, and the phase shift sequence achieves sub-pixel precision positioning. The binocular camera uses an industrial-grade CMOS sensor with a resolution of 2048×1536 pixels and a frame rate of 60fps. The baseline distance between the two cameras is set between 200mm and 300mm. The fringe phase shift is calculated using a four-step phase shift method. Four fringe images with a phase difference of π / 2 are projected, and the gray values ​​of the four images are used to determine the fringe pattern. , , , Calculate the package phase φ = arctan[( - ) / ( - The absolute phase value is obtained by combining Gray code unwrapping, and then converted into actual depth coordinates through a pre-calibrated phase-depth mapping table. The binocular vision system further corrects the depth value through epipolar constraints to eliminate the systematic error of monocular measurement. A linear polarizer is used, with its polarization direction set to two orthogonal directions, horizontal and vertical. The specular reflection light on the skin surface has specific polarization characteristics, while the polarization degree of diffuse reflection light is low. The polarization degree is calculated by measuring the light intensity in the two polarization directions. When the skin is made of aluminum alloy, the polarization degree of the specular reflection area is usually greater than 0.6, while the polarization degree of the normal diffuse reflection area is less than 0.3. Therefore, the reflection point can be effectively distinguished according to the polarization degree threshold. In the k-nearest neighbor algorithm, the k value is adaptively adjusted according to the point cloud density, usually taking 8 to 12 neighboring points. Outlier point identification adopts a statistical method, calculating the distance from the centroid of the neighborhood. When the distance exceeds 2.5 times the standard deviation of the mean neighborhood distance, it is identified as an outlier point. Hole filling adopts the radial basis function interpolation method. Using the non-reflective points at the cavity boundary as control points, a radial basis function network is constructed. Interpolation coefficients are obtained by solving a system of linear equations, thereby reconstructing the depth of the cavity region. Curvature constraints are introduced during the interpolation process to ensure that the filled surface maintains geometric continuity with the original skin surface.

[0020] Through the above steps, this application eliminates interference from high reflectivity of the skin. This application designs a combination of polarization filtering and adaptive illumination compensation, which distinguishes specular reflection points by using orthogonal polarizers, and adjusts the exposure time and light source power according to the reflection intensity, thereby controlling the point cloud distortion rate within the threshold and adapting to the complex lighting environment of the hangar.

[0021] Step S104: Based on the preliminary corrected point cloud, calculate the reflection intensity distribution of each region on the aircraft skin surface, and determine the compensated point cloud based on the calculated reflection intensity and a preset reflection intensity threshold. In one embodiment, based on the reflection intensity distribution of the preliminary corrected point cloud, if the reflection intensity exceeds the preset reflection intensity threshold, an adaptive illumination compensation algorithm that adjusts the exposure time and light source power based on the difference in reflection intensity distribution is applied to adjust the acquisition parameters to obtain the compensated point cloud; if the reflection intensity does not exceed the preset reflection intensity threshold, the preliminary corrected point cloud is directly used as the compensated point cloud.

[0022] Reference Figure 2The document illustrates a flowchart of a method for obtaining a compensated point cloud according to an embodiment of this application, including: Step S202: Obtaining the optical properties of sampling points. Specifically, obtaining the original grayscale value corresponding to each sampling point in the preliminary corrected point cloud on the camera imaging plane. Step S204: Constructing a reflection intensity distribution map based on the optical properties of the sampling points. In one embodiment, based on a preset camera response function, the original grayscale value (e.g., an eight-bit grayscale value) is mapped to a linear light intensity value; subsequently, the surface reflection intensity coefficient is calculated by combining the angle between the surface normal vector at each sampling point and the incident direction of the structured light (e.g., the cosine of the angle); finally, a reflection intensity distribution map of the aircraft skin surface is constructed based on the reflection intensity coefficients of all sampling points. Step S206: Calculating the average reflection intensity of each local region on the aircraft skin surface based on the reflection intensity distribution map. Step S208: Dividing the reflection intensity regions based on the average reflection intensity. The average reflection intensity is compared with a preset reflection intensity threshold, and the aircraft skin surface is divided into different reflection intensity regions according to the comparison result. In one embodiment, when the average reflection intensity exceeds a preset reflection intensity threshold, the aircraft skin surface is divided into, for example, a strong reflection region, a medium reflection region, and a weak reflection region. Step S210: Determine the optimal set of acquisition parameters for different reflection intensity regions. For the divided different reflection intensity regions, a high dynamic range imaging strategy is adopted, setting multiple different combinations of exposure time and light source power for each. In one embodiment, the exposure time is distributed logarithmically from the preset shortest time to the longest time, and the light source power increases linearly from the lowest power to the rated power. For each set of parameters, corresponding image frames are acquired and weighted fusion analysis is performed to determine the optimal acquisition parameters corresponding to each reflection intensity region. Step S212: Use the determined optimal acquisition parameters to re-acquire and fuse data for different reflection intensity regions. Based on the optimal acquisition parameters determined for each region, the strong reflection region, weak reflection region, etc., are re-scanned and acquired. For the data at the boundary of the region caused by parameter differences, interpolation compensation processing is performed to eliminate the discontinuity of the stitching boundary, and finally the compensated point cloud is obtained. In one embodiment, a short exposure low power parameter set is used for the strong reflection area, and a long exposure high power parameter set is used for the weak reflection area. The transition zone at the boundary of the region is processed by bilinear interpolation to eliminate the gray-scale abrupt change at the splicing boundary and obtain the compensated point cloud.

[0023] Specifically, in one implementation, the camera response function is obtained through pre-calibration. Multiple images are acquired under different lighting conditions using a standard grayscale card, establishing a mapping relationship between grayscale values ​​and actual light intensity. This function typically exhibits non-linear characteristics, with a steeper slope in the low grayscale range and tending towards saturation in the high grayscale range. For the 0-255 grayscale range of an eight-bit image, the response function maps it to a normalized 0-1 light intensity value range, achieving accurate conversion from image space to photometric space. The calculation of the surface reflection intensity coefficient requires comprehensive consideration of the light source geometry and material reflection characteristics. According to Lambert's law of reflection, reflection intensity is proportional to the cosine of the incident angle, but the aluminum alloy skin surface exhibits both diffuse and specular reflection characteristics. By calculating the dot product of each point cloud normal vector and the incident direction of the light source, the cosine value is obtained, and then multiplied by the normalized light intensity value corresponding to that point to obtain the reflection intensity coefficient. Combining the spatial coordinates with the reflection intensity coefficient forms a reflection intensity distribution map containing positional and optical information. The reflection region division employs an adaptive thresholding method based on statistical distribution. First, the histogram distribution of the reflectance intensity across the entire skin surface is calculated. A Gaussian mixture model is used to identify three distribution peaks, corresponding to weak, moderate, and strong reflectance areas, respectively. When the mean reflectance intensity exceeds a preset threshold, it indicates a large area of ​​strong reflection requiring compensation. The threshold is typically set to 0.7 of the normalized intensity; this value was obtained through extensive experimental statistics and effectively distinguishes between cases requiring compensation and those not. High dynamic range imaging parameter settings follow a logarithmic interval principle to ensure coverage of the entire dynamic range. The exposure time series increases in powers of 2, starting with the shortest exposure time and doubling each time until the longest exposure time is reached. The light source power increases linearly, starting with the lowest power and increasing at fixed steps to the rated power. This combination method obtains image sequences with different exposure levels, providing rich brightness level information for subsequent fusion. Each set of parameters corresponds to one image frame, forming an exposure sequence image set. The weighted fusion process employs a pixel reliability-based weight allocation strategy. For each pixel location, its signal-to-noise ratio (SNR) in different exposure images is calculated; pixels with high SNR are given greater weight, while pixels near saturation or underexposed have reduced weight. The weighting function employs a Gaussian distribution, with the highest weight at the middle gray levels and gradually decreasing weight at the extremes. A weighted average is used to obtain a fused high dynamic range image, containing complete detail information from dark to bright areas. Based on the brightness characteristics of different reflective regions, the optimal exposure parameter combination for each region is extracted from the fused image. Spatial continuity of parameters needs to be considered when re-acquiring data from different regions. Short exposure and low power parameter sets are used in strong reflective areas to reduce the number of saturated pixels and preserve highlight details; long exposure and high power parameter sets are used in weak reflective areas to increase the signal intensity in dark areas and enhance detail visibility; and compromise parameters are used in medium reflective areas to balance the performance of bright and dark areas. An overlapping transition zone is set between adjacent regions, with a width of 10% to 20% of the single acquisition field of view. Acquisition parameters are gradually adjusted within the transition zone to achieve a smooth transition.Bilinear interpolation is a crucial step in eliminating stitching artifacts. Within the transition zone at the boundary of regions, the grayscale value of each pixel is jointly determined by the acquisition results of the two adjacent regions. Interpolation weights are calculated based on the distance from the pixel to the region boundary, with closer pixels receiving larger weights. This gradual transition eliminates abrupt grayscale changes caused by different exposure parameters, resulting in visually continuous point cloud data. When inspecting the wing skin of a certain aircraft model, the large curvature of the wing's leading edge leads to significant differences in reflectivity at different locations. The top of the leading edge, near-vertical incidence, has the highest reflectivity and is prone to overexposure; while the side regions have larger incidence angles and lower reflectivity. The aforementioned adaptive compensation method automatically selects appropriate acquisition parameters for different regions, ensuring uniform and consistent data quality across the entire wing surface. Furthermore, the quality assessment of the compensated point cloud is achieved by calculating two indicators: point cloud density uniformity and grayscale distribution range. Density uniformity reflects the sufficiency of sampling in different regions, while grayscale distribution range reflects the utilization rate of the dynamic range. After adaptive illumination compensation, the standard deviation of the point cloud density decreases, and the effective grayscale range increases, significantly improving the data foundation for subsequent damage identification.

[0024] Step S106: Extract geometric features from the compensated point cloud, and use a spatial domain filtering method to denoise the compensated point cloud based on the geometric features to obtain a denoised point cloud. In one embodiment, geometric feature vectors are extracted from the compensated point cloud, and a two-dimensional Gaussian filtering algorithm with a preset standard deviation range is used to remove noise interference to determine the denoised point cloud.

[0025] For each sampling point in the compensated point cloud, its k nearest neighbors within a preset radius are searched to form a neighborhood point set. A covariance matrix is ​​constructed for the neighborhood point set and eigenvalue decomposition is performed. The eigenvector corresponding to the smallest eigenvalue is used as the normal vector of the sampling point. The orientation of the normal vector is propagated to ensure that the angle between the normal vectors of adjacent points is less than 90 degrees, resulting in a normal vector field with consistent orientation. Based on the normal vector field, a quadratic polynomial surface z = ax² + bxy + cy² + dx + ey + f is fitted in the neighborhood of each sampling point. The coefficients a to f are solved using the least squares method. The second-order partial derivatives are calculated based on the surface equation to form the Hessian matrix, whose eigenvalues ​​are the principal curvature and secondary curvature. At the same time, the standard deviation of the distance from the point to the centroid of the neighborhood point set is calculated as the neighborhood distance deviation, resulting in a geometric eigenvector containing the normal vector, principal curvature value, and distance deviation. For each component of the geometric feature vector, a corresponding feature map is constructed. A two-dimensional Gaussian filter kernel is used to perform convolution operation on the feature map. The standard deviation of the filter kernel is adaptively adjusted within a preset range according to the local density of the point cloud. If the change in the feature value of a point before and after filtering exceeds a preset threshold, it is marked as a noise point and removed from the point cloud. The remaining points constitute a denoised point cloud.

[0026] Specifically, in one implementation, k-nearest neighbor search achieves fast querying by constructing a KD-tree data structure for the point cloud. For typical million-level point cloud data on skin surfaces, the KD-tree can reduce the complexity of neighborhood search from O(n) to O(logn). The choice of k value is related to the point cloud density; k is 8 to 12 in dense regions and 15 to 20 in sparse regions, ensuring that enough neighboring points participate in feature calculation. The construction of the covariance matrix is ​​based on the spatial distribution characteristics of the neighboring points. After centering the set of neighboring points, a 3×3 covariance matrix is ​​calculated, whose elements are the covariances of each coordinate component. Three eigenvalues ​​are obtained through the Jacobi iteration method or QR decomposition. ≥ ≥ and the corresponding eigenvectors. Minimum eigenvalue. The corresponding eigenvector is the normal vector of that point, reflecting the direction of minimum change in the neighborhood point set. The orientation of the normal vector is unified through a propagation algorithm, starting from the seed point and gradually adjusting the orientation of the normal vectors of adjacent points to make the included angle less than 90 degrees. The quadratic polynomial surface fitting uses the least squares method to solve the overdetermined system of equations. For n points in the neighborhood (… , , Construct an n×6 coefficient matrix A, where each row is [ ², , ², , [1], by solving the normal equation Ax= b yields the polynomial coefficients. The Hessian matrix is ​​obtained from the second-order partial derivatives. ²z x²、 ²z y、 ²z y² is composed of eigenvalues and Principal curvature reflects the degree of curvature of the skin surface. Damaged areas typically exhibit abnormal curvature, such as negative principal curvature at pits and drastic curvature changes at crack edges. The construction of the two-dimensional Gaussian filter kernel considers the anisotropic distribution of the point cloud. The filter kernel size is adaptively adjusted according to the local point density, using a 3×3 kernel in dense areas and expanding to 5×5 or 7×7 in sparse areas. The standard deviation σ is dynamically adjusted within a preset range, typically taking 1.5 to 2.5 times the local average point spacing. Feature maps are constructed for the three feature components: normal vector, principal curvature, and distance deviation, and independent filtering is performed. Noise points are identified using a multi-feature comprehensive evaluation. When the angle between the normal vector and the average normal vector of a point exceeds 30 degrees, or the principal curvature value exceeds three times the standard deviation of the neighborhood mean, or the distance deviation exceeds a threshold, it is marked as a potential noise point. If two or more of the three features are abnormal, it is confirmed as a noise point and removed, achieving a robust denoising effect.

[0027] Step S108: Construct a three-dimensional mesh model representing the three-dimensional structure of the aircraft skin surface based on the denoised point cloud, and determine the damage candidate areas of the aircraft skin surface from the three-dimensional mesh model using a point cloud local geometric feature analysis method. In one embodiment, a three-dimensional mesh model is constructed for the denoised point cloud, and a curvature analysis algorithm based on the principal curvature calculated from the point cloud neighborhood is used to detect potential damage areas and determine damage candidate areas.

[0028] For denoising point clouds, a 3D mesh is constructed using the Delaunay triangulation algorithm. Each point in the point cloud is treated as a vertex using a point-by-point insertion method. Adjacent vertices are connected to form triangular patches based on the criterion that no other vertex is contained within the circumcircle of any triangle, resulting in a 3D mesh model containing the topological relationships of vertices, edges, and patches. Based on this 3D mesh model, for each vertex, all directly connected adjacent vertices are searched. A quadratic surface is fitted to these adjacent vertices using the least squares method, and the principal curvature of the surface at that vertex is calculated. and second curvature A curvature feature map is constructed based on the absolute values ​​of Gaussian curvature and mean curvature. For this curvature feature map, when the absolute value of the Gaussian curvature or the absolute value of the mean curvature of a vertex exceeds a preset threshold, that vertex is marked as a damage seed point. Starting from the seed point, the map expands towards adjacent vertices. If the difference between the curvature value of an adjacent vertex and the curvature value of the seed point is less than a preset tolerance, they are merged into the same region. Multiple connected regions are formed through iterative expansion, thus determining the damage candidate region.

[0029] Specifically, in one implementation, Delaunay triangulation is achieved using an incremental insertion algorithm. Initially, a convex hull containing all point clouds is constructed. Then, internal points are inserted one by one. For each inserted point, a triangle containing that point is found, deleted, and connected to the new point to form a new triangular facet. The Delaunay property is preserved through edge flipping operations; that is, the circumcircle of any triangle does not contain other vertices. This property ensures the quality of the triangular mesh and avoids excessively long triangles. The curvature feature is calculated based on local surface fitting. For each vertex in the mesh, a ring of directly connected adjacent vertices is collected, typically containing 6 to 8 vertices. The coordinates of these vertices are transformed to a local coordinate system with the target vertex as the origin and the normal vector as the z-axis. Then, a quadratic polynomial surface z = ax² + 2bxy + cy² is fitted. The coefficients a, b, and c are solved using the least squares method to construct a 2×2 shape operator matrix, whose eigenvalues ​​are the principal curvatures. and Gaussian curvature K = × Reflecting the inherent curvature of the surface, the average curvature H = ( + The value of 1 / 2 represents the external curvature of the surface. It should be noted that the selection of damage seed points uses a dual threshold determination. When the absolute value of the Gaussian curvature exceeds three times the standard deviation of the normal curvature range of the skin, or the absolute value of the average curvature exceeds a preset threshold, the vertex is marked as a potential seed point. For pit-type damage, the Gaussian curvature is negative and has a large absolute value; for protrusion-type damage, the Gaussian curvature is positive; crack edges exhibit a sharp change in average curvature. Exemplarily, the region growth process uses a priority queue to manage the vertices to be processed. Starting from the seed point, all its adjacent vertices are added to the queue, sorted according to curvature similarity. The first vertex in the queue is taken out, and the difference between its curvature value and the curvature of the seed point is calculated. If the difference is less than 20% of the preset tolerance range, the vertex is merged into the current region, and its unvisited adjacent vertices are added to the queue. This process is repeated until the queue is empty, forming a connected damage candidate region. Preferably, the formed candidate region undergoes post-processing optimization. The geometric features of each candidate region, such as area, perimeter, and compactness, are calculated. Regions with excessively small areas may be spurious damage caused by noise and are therefore eliminated. For elongated candidate regions, their main direction is identified through skeleton extraction to determine whether they are crack-type damage. This multi-level screening improves the accuracy of damage candidate region identification.

[0030] The above steps achieve automated and high-precision point cloud processing. Using two-dimensional Gaussian filtering and a damage detection algorithm based on neighborhood principal curvature, a three-dimensional mesh is constructed through Delaunay triangulation. Combined with curvature thresholding, candidate damage regions are located, achieving automated noise removal and damage localization, and reducing localization errors.

[0031] Step S110: Obtain a subset of point clouds associated with the damage candidate region, and use a deep learning model to classify the damage type of the point cloud subset to obtain a classification result of the damage type. In one embodiment, a subset of point clouds within the damage candidate region is obtained, and a deep learning model is used to classify the damage type to obtain a classification result.

[0032] The bounding box coordinates of the damage candidate region are obtained. Based on the bounding box, a subset of the point cloud containing the damage region and its surrounding neighborhood is cropped from the original point cloud. This subset of the point cloud is voxelized and converted into a 3D voxel mesh at a preset resolution. Each voxel records the density and normal vector distribution information of its internal points, resulting in a voxelized feature tensor. Data augmentation is performed on the voxelized feature tensor through rotation, mirroring, and scaling transformations. A 3D convolutional neural network is used to extract multi-scale spatial features. The convolutional layers extract edge contour features from the low layer, local shape patterns from the middle layer, and global semantic features from the high layer. Feature maps from different levels are fused through a feature pyramid to obtain a multi-scale feature representation. Based on the multi-scale feature representation, it is mapped to the damage category space through a fully connected layer. The softmax function is used to calculate the probability distribution of each damage category, including cracks, pits, corrosion, scratches, and bulges. The category with the highest probability is selected as the predicted category, and its confidence score is recorded to obtain a preliminary classification result. For the preliminary classification results, if the confidence score is lower than a preset threshold, the geometric feature descriptors of the candidate region are extracted, including surface roughness, depth distribution histogram and shape compactness, and input into a pre-trained support vector machine for secondary discrimination. The discrimination results of deep learning and geometric features are then fused to determine the final classification result.

[0033] Specifically, in one implementation, point cloud clipping employs an axis-aligned bounding box method to achieve accurate region extraction. Based on the boundary vertex coordinates of the damage candidate region, minimum and maximum coordinate values ​​are calculated to form a 3D bounding box. To preserve the contextual information of the damage edges, each dimension of the bounding box is expanded by 10% to 15%, ensuring that the clipped point cloud subset contains complete damage features and its transition regions. For typical skin damage, the clipped point cloud subset typically contains 5000 to 20000 points, ensuring both the integrity of local details and controlling the computational complexity of subsequent processing. Specifically, voxelization is the key process for converting irregular point clouds into regular 3D meshes. A voxel resolution of 2mm × 2mm × 2mm is set, a scale capable of capturing millimeter-level damage details. For each voxel, the number of points it contains is counted as a density value, and the average normal vector of these points is calculated as the directional feature. Empty voxels are assigned a value of zero, and occupied voxels are normalized to the 0-1 range based on the point density. In this way, sparse point cloud data is transformed into a dense 32×32×32 or 64×64×64 three-dimensional tensor representation, suitable for convolutional neural network processing. Voxelization preserves the spatial topology and local geometric features of the point cloud, providing a standardized input format for deep learning. It is important to note that data augmentation strategies are crucial for improving the model's generalization ability. Rotation augmentation is performed separately along three axes, with rotation angles randomly sampled within the range of -30 degrees to 30 degrees; mirror augmentation flips the data along the principal plane; and scaling is adjusted within the range of 0.8 to 1.2 times to simulate scanning effects at different distances. Furthermore, random point dropping and Gaussian noise injection are introduced to simulate data loss and measurement errors in actual detection. Each training sample generates 8 to 10 variants by combining different augmentation operations, significantly expanding the diversity of the training dataset. Exemplarily, the 3D convolutional neural network architecture employs a progressive feature extraction design. The network contains five convolutional blocks, each consisting of two 3×3×3 convolutional layers, a batch normalization layer, and a ReLU activation function. The first convolutional block extracts 32-dimensional features, primarily capturing edge and corner information. The second and third blocks extract 64-dimensional and 128-dimensional features respectively, identifying local shape patterns such as depressions and bulges. The fourth and fifth blocks extract 256-dimensional high-level semantic features to understand the overall morphology of the damage. Spatial resolution is reduced and receptive field increased through 2×2×2 max pooling between convolutional layers. The feature pyramid network fuses feature maps from different levels through lateral connections and upsampling operations, forming a 256-dimensional feature vector containing multi-scale information. Preferably, the damage classification system establishes five main categories based on aviation maintenance standards. Crack-type damage is characterized by a linearly extending, elongated shape with drastic depth variations; pit-type damage presents as circular or elliptical depressions with uniform depth; corrosion-type damage exhibits increased surface roughness and irregular boundaries; scratch-type damage consists of shallow linear marks with relatively shallow depth; and bulge-type damage manifests as localized convex deformation.The fully connected layer maps 256-dimensional feature vectors to a 5-dimensional class space and calculates the posterior probability of each class using the softmax function. The training process employs the cross-entropy loss function and a stochastic gradient descent optimizer with momentum. The initial learning rate is set to 0.01, decaying to 0.1 times the original rate every 30 epochs. In one possible implementation, the confidence threshold is determined using statistical methods on the validation set. Confidence scores for all predictions on the validation set are collected, and a curve showing the relationship between confidence and accuracy is plotted. When the confidence score is below 0.7, the classification accuracy drops significantly; therefore, 0.7 is set as the threshold for triggering secondary discrimination. For low-confidence samples, supplementary geometric feature descriptors are extracted for auxiliary discrimination. It is understood that the calculation of geometric feature descriptors is based on the statistical characteristics of the damaged area. Surface roughness is obtained by calculating the standard deviation of the distance from a point to the fitted plane; the depth distribution histogram quantifies the depth values ​​into 20 intervals, and the distribution of points in each interval is statistically analyzed; shape compactness is defined as the cube root of the ratio of volume to surface area, reflecting the regularity of the damage. These features form a 15-dimensional vector, which is input into a pre-trained support vector machine (SVM) for classification. The SVM uses a radial basis function kernel, and the kernel parameters and penalty coefficients are optimized through grid search. Furthermore, a fusion discrimination strategy comprehensively considers the classification results of deep learning and geometric features. When the two methods predict the same category, the category is directly output; when the predictions are inconsistent, a weighted vote based on their respective confidence levels determines the final category. The weight of the deep learning result is set to 0.6, and the weight of the geometric feature result is set to 0.4; this ratio was obtained through extensive experimental verification. This multi-model fusion mechanism improves classification accuracy, especially for corrosion-related damage with blurred boundaries. For example, when a suspected damage to the fuselage skin is detected, the system first identifies it as a pit using a deep learning model with a confidence level of 0.65. Since the confidence level is below the threshold, geometric feature analysis is triggered, revealing an abnormally high surface roughness and a multi-peak depth distribution, which is more consistent with corrosion damage characteristics. Finally, the two discrimination results are fused, and the system outputs a classification of corrosion-related damage, marking the areas requiring focused inspection.

[0034] Step S112: Based on the classification results, fuse the boundary information of the damage candidate regions to generate a damage annotation map. In one embodiment, based on the classification results, fuse the boundary information of the damage candidate regions. If the rate of change of boundary curvature is higher than the change threshold, it is marked as crack damage; if the rate of change of boundary curvature is not higher than the change threshold, it is marked as other damage types according to the classification results of the deep learning model, thus obtaining a damage annotation map.

[0035] Based on the classification results, a sequence of boundary contour points of the damage candidate region is extracted. A contour simplification algorithm is used to retain key feature points. For each point in the feature point sequence, the angle change between it and its adjacent points is calculated. The local curvature is obtained by dividing the angle change value by the arc length between points. The curvature change rate is obtained by dividing the difference in curvature between adjacent points by the distance between points, thus constructing a boundary curvature change rate sequence. For this boundary curvature change rate sequence, if the change rate values ​​of multiple consecutive points exceed a preset change threshold, the damage region is re-marked as crack damage. If the change rate values ​​do not continuously exceed the threshold, the original classification result output by the deep learning model is retained, resulting in a corrected damage category. Based on the corrected damage category, a preset color coding value is assigned to each damage type. The point cloud within the damage region is marked according to the corresponding color value. The marking results are superimposed onto the three-dimensional coordinate space of the original skin point cloud through color rendering, forming a damage annotation map that can distinguish different damage types.

[0036] Specifically, in one implementation, boundary contour extraction is achieved through an eight-neighborhood tracking algorithm. Starting from any boundary point of the damaged area, adjacent boundary points are searched in a clockwise or counterclockwise direction to form an ordered sequence of contour points. Contour simplification employs a recursive segmentation method, setting a distance threshold of 0.5 mm, retaining feature points with significant curvature changes, and eliminating redundant intermediate points. The simplified contour retains the original shape features while reducing subsequent computational load. Specifically, the calculation of local curvature is based on the geometric relationship of discrete points. For a point Pi on the contour, it forms two vectors with its preceding and following points Pi-1 and Pi+1. The curvature κ = θ / s is calculated using the relationship between the vector angle θ and the arc length s. The rate of change of curvature is obtained by dividing the curvature difference Δκ between two adjacent points by the distance Δs between the points, i.e., dκ / ds = Δκ / Δs. This discretized calculation method is suitable for processing point cloud data and can effectively capture the geometric change features of the contour. Crack-type damage typically exhibits a sharp change in curvature at the boundary, while the boundaries of other types of damage are relatively smooth. It should be noted that the determination of continuous exceeding the threshold adopts a sliding window detection method. The window size is set to 5 to 7 consecutive points. When more than 80% of the points within the window have a curvature change rate exceeding a preset threshold, it is considered that a continuous exceeding of the threshold has occurred. The preset threshold is obtained through statistical analysis of a large number of samples, typically set to 3 times the standard deviation of the mean of the normal curvature change rate. This determination method can identify the linear extension characteristics of cracks while avoiding misjudgments caused by individual noise points. For example, color encoding uses the HSV color space to achieve differentiated display. Crack damage is assigned a hue value of 0 degrees corresponding to the red family, pit damage to 240 degrees corresponding to the blue family, corrosion damage to 60 degrees corresponding to the yellow family, scratch damage to 120 degrees corresponding to the green family, and bulge damage to 300 degrees corresponding to the purple family. Saturation is set to the maximum value to ensure vivid colors, and brightness is adjusted according to the damage depth; the greater the depth, the lower the brightness. Preferably, the rendering of the damage annotation map uses point cloud coloring technology. For each point within a damaged area, a corresponding RGB color value is assigned according to its damage type. Undamaged areas retain their original grayscale display, forming a clear visual contrast. 3D rendering is achieved through OpenGL or DirectX graphics interfaces, supporting interactive operations such as rotation, scaling, and translation, facilitating observation of damage distribution from different angles. Rendering results can be exported as standard PLY or OBJ format files, containing vertex coordinates and color information.

[0037] Step S114: Based on the damage annotation map, the damage size parameters are calibrated using an error compensation model constructed by a parameter fitting optimization algorithm to generate a final damage report. In one embodiment, damage size parameters including damage length, width, and depth are quantified from the damage annotation map, and an iterative optimization algorithm based on the least squares method is used to calibrate the measurement error to determine the final damage report.

[0038] The 3D point cloud coordinates of each damaged area are extracted from the damage annotation map. Principal component analysis is used to determine the principal axis direction of the damage. The maximum span of the point cloud along the principal axis is calculated as the damage length, the maximum span perpendicular to the principal axis is calculated as the width, and the maximum vertical distance from the point cloud to the fitted plane of the surrounding undamaged area is calculated as the depth, thus obtaining the initial size parameters. A measurement error model is established based on the initial size parameters. The linear relationship between multiple measurements and standard reference values ​​is fitted using the least squares method. The residual value of each measurement point is calculated, and the regression coefficient is adjusted according to the residual magnitude. Iterative calculation is performed until the sum of squared residuals converges to below a preset threshold, obtaining the corrected size values. A damage detection report is compiled based on the corrected size values, recording the location coordinates, damage type, length, width, and depth values ​​of each damage. The damage level is determined by comparing the damage size with a preset grading standard, and a final damage report containing all damage information is generated.

[0039] Specifically, in practical applications, principal component analysis (PCA) extracts the principal axes by constructing the covariance matrix of the point cloud of the damaged region. The 3D coordinates of all points within the damaged region are normalized to the centroid coordinate system, and a 3×3 covariance matrix is ​​calculated. Three orthogonal principal directions are obtained through eigenvalue decomposition. The eigenvector corresponding to the largest eigenvalue is the principal axis direction of the damage, which is usually consistent with the direction of damage extension. All points are projected along the principal axis; the difference between the maximum and minimum projection values ​​is the damage length. Specifically, a robust fitting method is used to determine the reference plane. Undamaged point clouds within a radius of 2 to 3 times the damage size around the damaged region are selected. Outliers are removed using a random sampling consensus algorithm, and the remaining points are fitted to a plane. The fitted plane equation is ax + by + cz + d = 0, where the coefficients are solved using the least squares method. The damage depth is defined as the maximum vertical distance from a point within the damaged region to this reference plane; a positive value indicates a convexity, and a negative value indicates a depression. This local reference plane method can adapt to changes in skin curvature, improving the accuracy of depth measurement. It should be noted that the measurement error model considers both systematic and random errors. Systematic errors mainly originate from the calibration deviation of the scanning equipment and environmental factors, manifesting as a linear deviation between the measured value and the reference value. Random errors are caused by measurement noise and follow a normal distribution. By collecting multiple measurement data of the same damage, a linear regression model y=ax+b+ε is established, where y is the measured value, x is the reference value, a and b are regression coefficients, and ε is the random error term. For example, the iterative optimization process uses the weighted least squares method. In the initial iteration, all measurement points have equal weights, and the residual ri=yi-axi-b is calculated. The weights wi=1 / (1+|ri| / σ) are updated according to the residual magnitude, where σ is the standard deviation of the residuals. The regression coefficients are recalculated using the new weights, and this process is repeated until the coefficient change between two adjacent iterations is less than a preset threshold, typically set to 0.001. After 3 to 5 iterations, the sum of squared residuals usually converges to a stable value. Preferably, a four-level classification standard is used for damage level determination. Level 1 damage is minor surface damage with a depth of less than 0.5 mm; Level 2 damage has a depth between 0.5 mm and 2 mm; Level 3 damage has a depth between 2 mm and 5 mm; and Level 4 damage has a depth exceeding 5 mm or penetrates the skin. The report also records the area of ​​the damage, obtained through convex hull calculation of the point cloud within the damaged area, providing comprehensive quantitative basis for maintenance decisions.

[0040] The above steps establish a digital collaborative link for detection and maintenance. The output includes a standardized report containing "damage location + type + size + level," supporting direct integration with China Eastern Airlines' hangar MRO system, SAP system, and digital twin module, enabling real-time synchronization of detection data and reducing data entry time.

[0041] Therefore, the solution proposed in this application is suitable for the collaborative needs of hangars in multiple scenarios. It ensures technical compatibility with hangar intelligent equipment, such as supporting on-site viewing of inspection results by handheld mobile devices, and damage data can trigger AGV scheduling, while also meeting the hangar's explosion-proof and electromagnetic interference resistance requirements.

[0042] The proposed solution addresses the high reflectivity and high precision requirements of aircraft skin by using polarization filtering and adaptive illumination compensation to solve data quality issues, improving damage identification accuracy through three-dimensional geometric feature extraction and deep learning, and achieving dimensional quantization calibration through least squares iterative optimization. This forms a complete damage identification solution adapted to aviation scenarios, effectively compensating for the shortcomings of existing technologies in aircraft skin damage detection and further improving the accuracy of aircraft skin damage identification.

[0043] Specifically, the above-described method of this application has the following technical advantages: Anti-reflective capabilities are better suited for aviation scenarios: Compared with detection technologies without anti-reflective design, this application adds polarization filtering and adaptive illumination compensation, which can improve the recognition rate of skin reflection points and the efficiency of point clouds; compared with image ultrasonic fusion technology, it does not rely on ultrasonic equipment, eliminates reflection through a pure optical solution, and has better equipment portability, making it suitable for the small maintenance space of hangars.

[0044] Higher damage localization accuracy: Compared with detection technology that relies on two-dimensional feature binarization, this application can locate microcracks at the 0.1mm level through three-dimensional point cloud curvature analysis. The candidate area recognition accuracy is high, far exceeding the macroscopic damage recognition rate of the former. Compared with image ultrasonic fusion technology, three-dimensional mesh modeling combined with principal curvature analysis can directly reflect the surface morphology of the skin, reduce damage localization error, and meet the micron-level requirements of aviation maintenance.

[0045] More intelligent and accurate classification and quantification: This application uses a three-dimensional convolutional neural network (CNN) with secondary discrimination to improve the accuracy of damage classification, especially for corrosion damage with blurred boundaries; at the same time, the addition of least squares iterative optimization reduces the error in size measurement and can be directly used for aircraft material selection.

[0046] Enhanced digital collaboration capabilities: This application report can directly connect to the hangar MRO system, automatically associate maintenance needs with work cards, and synchronize to the digital twin module to achieve virtual damage mapping; it supports real-time data retrieval by handheld mobile devices, eliminating the need for maintenance personnel to travel to and from the control room, and reducing the inspection time for a single aircraft from 4 hours to 1.5 hours, aligning with the hangar's goals of efficient and paperless operation and maintenance.

[0047] Figure 3 A schematic diagram of an aircraft skin damage identification device 300 according to an embodiment of this application is shown. The aircraft skin damage identification device 300 includes the following modules: The point cloud acquisition and preprocessing module 302 is used to acquire the initial point cloud of the aircraft skin surface through a structured light device, and to perform data preprocessing on the initial point cloud to obtain a preliminary corrected point cloud.

[0048] The reflection intensity analysis and compensation module 304 is used to calculate the reflection intensity distribution of each region on the aircraft skin surface based on the preliminary correction point cloud, and to determine the compensated point cloud based on the calculated reflection intensity and a preset reflection intensity threshold.

[0049] The point cloud denoising module 306 is used to extract geometric features from the compensated point cloud and perform denoising processing on the compensated point cloud based on the geometric features using a spatial domain filtering method to obtain a denoised point cloud.

[0050] The damage area localization module 308 is used to construct a three-dimensional mesh model representing the three-dimensional structure of the aircraft skin surface based on the denoised point cloud, and to determine the damage candidate area of ​​the aircraft skin surface from the three-dimensional mesh model through the point cloud local geometric feature analysis method.

[0051] The damage classification module 310 is used to obtain a subset of point clouds associated with the damage candidate region, and to use a deep learning model to classify the damage type of the subset of point clouds to obtain the classification result of the damage type.

[0052] The damage annotation generation module 312 is used to generate a damage annotation map by fusing the boundary information of the damage candidate region based on the classification result.

[0053] The damage report generation module 314 is used to calibrate the damage size parameters based on the damage annotation map using an error compensation model constructed by an optimization algorithm based on parameter fitting, so as to generate a final damage report.

[0054] Figure 4 A schematic diagram of the hardware structure of an aircraft skin damage identification device 400 according to an embodiment of this application is shown.

[0055] The aircraft skin damage identification device 400 may include a processor 402 and a memory 404 storing computer program instructions.

[0056] Specifically, the processor 402 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0057] Memory 404 may include mass storage for data or instructions. For example, and not limitingly, memory 404 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. In one instance, memory 404 may include removable or non-removable (or fixed) media, or memory 404 may be non-volatile solid-state storage. Memory 404 may be internal or external to the integrated gateway disaster recovery device.

[0058] Memory 404 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Thus, generally, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to one aspect of this disclosure.

[0059] The processor 402 reads and executes the computer program instructions stored in the memory 404 to achieve... Figure 1 The method in the illustrated embodiment...

[0060] In one example, the device may also include a communication interface 406 and a bus 410. For example, Figure 4 As shown, the processor 402, memory 404, and communication interface 406 are connected through bus 410 and complete communication with each other.

[0061] The communication interface 406 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0062] Bus 410 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 410 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0063] Furthermore, in conjunction with the aircraft skin damage identification method in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the aircraft skin damage identification methods in the above embodiments.

[0064] This application also provides a computer program product, including a computer program that, when executed, implements any of the aircraft skin damage identification methods described in the above embodiments.

[0065] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0066] The functional blocks shown in the above-described block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, read-only memory (ROM), flash memory, erasable read-only memory (EROM), floppy disks, compact disc read-only memory (CD-ROM), optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0067] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0068] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0069] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for identifying damage to aircraft skin, characterized in that, The method includes the following steps: The initial point cloud of the aircraft skin surface is obtained using structured light equipment; Calculate the polarization characteristics of each sampling point in the initial point cloud; and Based on the comparison results between the polarization features and the preset polarization feature threshold, the initial point cloud is corrected to obtain a preliminary corrected point cloud; Based on the preliminary corrected point cloud, the reflection intensity distribution of each region on the aircraft skin surface is calculated. In response to the calculated reflection intensity exceeding a preset reflection intensity threshold, a dynamic acquisition parameter adjustment strategy is applied to obtain a compensated point cloud; and In response to the calculated reflection intensity not exceeding the preset reflection intensity threshold, the preliminary corrected point cloud is used as the compensated point cloud; Geometric features are extracted from the compensated point cloud, and spatial domain filtering is used to denoise the compensated point cloud based on the geometric features to obtain a denoised point cloud. Based on the denoised point cloud, a three-dimensional mesh model representing the three-dimensional structure of the aircraft skin surface is constructed, and the damage candidate area of ​​the aircraft skin surface is determined from the three-dimensional mesh model by the point cloud local geometric feature analysis method. Obtain a subset of point clouds associated with the candidate damage region, and use a deep learning model to classify the damage type of the subset of point clouds to obtain the classification result of the damage type; Based on the classification results, the boundary information of the damage candidate regions is fused to generate a damage annotation map; and Based on the damage annotation map, an error compensation model constructed using a parameter fitting-based optimization algorithm is used to calibrate the damage size parameters in order to generate the final damage report.

2. The aircraft skin damage identification method according to claim 1, characterized in that, Calculating the polarization features of each sampling point in the initial point cloud includes: Obtain the vertical polarization component intensity and the horizontal polarization component intensity at each sampling point; and The degree of polarization of each sampling point is calculated based on the intensity of the vertical polarization component and the intensity of the horizontal polarization component.

3. The aircraft skin damage identification method according to claim 2, characterized in that, The vertical polarization component intensity and horizontal polarization component intensity of each sampling point are acquired by a binocular camera equipped with orthogonal polarizers in the structured light device.

4. The aircraft skin damage identification method according to claim 1, characterized in that, The calculation of the reflection intensity distribution in each region of the aircraft skin surface includes: Obtain the optical attribute information of each sampling point in the preliminary correction point cloud; and The optical property information is converted into reflection intensity values, and a reflection intensity distribution map of the aircraft skin surface is constructed based on the reflection intensity values ​​of all sampling points.

5. The aircraft skin damage identification method according to claim 4, characterized in that, The optical property information includes at least grayscale values.

6. The aircraft skin damage identification method according to claim 5, characterized in that, The dynamic acquisition parameter adjustment strategy includes: The aircraft skin surface is divided into multiple reflection intensity regions based on the aforementioned reflection intensity distribution; Multiple sets of acquisition parameters are configured for regions with different reflection intensities, and the optimal acquisition parameters for each region are determined through a weighted fusion strategy; and Based on the optimal acquisition parameters, data is reacquired and fused in the corresponding reflection intensity region to generate the compensated point cloud.

7. The aircraft skin damage identification method according to any one of claims 1 to 3, characterized in that, The spatial domain filtering method includes: Based on the spatial distribution characteristics of the compensated point cloud, the parameters of the filter kernel are adaptively configured.

8. The aircraft skin damage identification method according to any one of claims 1 to 3, characterized in that, The spatial domain filtering includes Gaussian filtering.

9. The aircraft skin damage identification method according to any one of claims 1 to 3, characterized in that, The candidate damage regions for the aircraft skin surface were determined from the three-dimensional mesh model using point cloud local geometric feature analysis, including: For each vertex of the three-dimensional mesh model, fit a quadratic surface to its adjacent vertices, and calculate the Gaussian curvature and mean curvature at that vertex; Vertices whose absolute Gaussian curvature or average absolute curvature exceeds the corresponding preset threshold are marked as damage seed points. Starting from each damage seed point, iteratively expand to its adjacent vertices, merging vertices whose curvature values ​​differ from the seed points by less than a preset tolerance into the same region, ultimately generating a connected damage candidate region.

10. The aircraft skin damage identification method according to any one of claims 1 to 3, characterized in that, A deep learning model is used to classify the damage types of the point cloud subset, and the classification results of the damage types include: Obtain the probability distribution of each damage category output by the deep learning model; Based on the probability distribution, a preliminary classification result and its confidence level are determined; and In response to the confidence level being lower than a preset confidence threshold, auxiliary geometric features of the point cloud subset are extracted, and the preliminary classification result is verified using a discriminant model to determine the classification result.

11. The aircraft skin damage identification method according to claim 10, characterized in that, The discriminant model is a pre-trained support vector machine.

12. The aircraft skin damage identification method according to any one of claims 1 to 3, characterized in that, Based on the classification results, the boundary information of the damage candidate regions is fused to generate a damage annotation map, including: Analyze the rate of change of geometric features of the boundary of the damage candidate region; Compare the rate of change of the geometric feature with a preset rate of change threshold; and Based on the comparison results and the classification results, the candidate damage regions are marked to generate the damage annotation map.

13. The aircraft skin damage identification method according to any one of claims 1 to 3, characterized in that, The damage size parameters include the length, width, and depth of the damage.

14. The aircraft skin damage identification method according to any one of claims 1 to 3, characterized in that, Based on the damage annotation map, an error compensation model constructed using a parameter fitting-based optimization algorithm is used to calibrate the damage size parameters to generate a final damage report, including: The three-dimensional point cloud coordinates of each of the damage candidate regions are extracted from the damage annotation map to obtain the initial damage size parameters; An error compensation model based on the least squares method is established to correct the initial damage size parameters, resulting in corrected damage size parameters; and Based on the corrected damage size parameters, the damage level is determined and the final damage report is generated.

15. An aircraft skin damage identification device, performing the aircraft skin damage identification method as described in claim 1, characterized in that, The device includes: The point cloud acquisition and preprocessing module is used to acquire the initial point cloud of the aircraft skin surface through a structured light device, and to perform data preprocessing on the initial point cloud to obtain a preliminary corrected point cloud. The reflection intensity analysis and compensation module is used to calculate the reflection intensity distribution of each region on the aircraft skin surface based on the preliminary correction point cloud, and to determine the compensated point cloud based on the calculated reflection intensity and a preset reflection intensity threshold. The point cloud denoising module is used to extract geometric features from the compensated point cloud and use a spatial domain filtering method to denoise the compensated point cloud based on the geometric features to obtain a denoised point cloud. The damage area localization module is used to construct a three-dimensional mesh model representing the three-dimensional structure of the aircraft skin surface based on the denoised point cloud, and to determine the damage candidate area of ​​the aircraft skin surface from the three-dimensional mesh model through the point cloud local geometric feature analysis method. The damage classification module is used to obtain a subset of point clouds associated with the damage candidate region, and to use a deep learning model to classify the damage type of the subset of point clouds to obtain the classification result of the damage type. The damage annotation generation module is used to generate a damage annotation map by fusing the boundary information of the damage candidate regions based on the classification results; and The damage report generation module is used to calibrate the damage size parameters based on the damage annotation map using an error compensation model constructed by an optimization algorithm based on parameter fitting, so as to generate a final damage report.

16. An aircraft skin damage identification device, characterized in that, The device includes: a processor and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the aircraft skin damage identification method as described in any one of claims 1-14.

17. A computer-readable storage medium, characterized in that, The computer storage medium stores computer program instructions, which, when executed by a processor, implement the aircraft skin damage identification method as described in any one of claims 1-14.

18. A computer program product, characterized in that, Includes a computer program, which, when executed, implements the aircraft skin damage identification method as described in any one of claims 1-14.

Images