Blade point cloud repairing method and system based on damage perception and multi-scale fractal reconstruction

Abstract

The application provides a blade point cloud repair method and system based on damage perception and multi-scale fractal reconstruction, and belongs to the field of point cloud data processing; solves the problems of difficult automatic positioning of defects and high-precision completion of damaged blades; the method comprises the following steps: collecting three-dimensional point cloud data of a damaged blade and preprocessing the three-dimensional point cloud data into original aligned blade point clouds; constructing a defect detection network and a multi-scale fractal reconstruction network and training; projecting the original aligned blade point clouds of the damaged blade into a two-dimensional gray feature map, processing the two-dimensional gray feature map by using the defect detection network to obtain a two-dimensional defect detection frame, and mapping the two-dimensional defect detection frame to a three-dimensional point cloud to obtain a three-dimensional defect surrounding area; cutting a defect point cloud segment according to the three-dimensional defect surrounding area, inputting the defect point cloud segment into the multi-scale fractal reconstruction network, generating a completed point cloud of the damaged blade, and fusing the completed point cloud with the original aligned blade point clouds to obtain a complete three-dimensional blade point cloud; and the application is applied to the repair of aero-engine blades.

Description

Technical Field

[0001] This application relates to the field of point cloud data processing technology, and in particular to a method and system for repairing leaf point clouds based on damage perception and multi-scale fractal reconstruction. Background Technology

[0002] Blades from deep-sea vehicles and aero engines operate under high pressure, high speed, and corrosive environments for extended periods. Subjected to fatigue, erosion, wear, and corrosion, they are prone to defects such as tip chipping, edge wear, and surface damage. Intelligent remanufacturing technology has become a key technology for blade repair, significantly improving blade lifespan and economic efficiency. Three-dimensional detection, damage assessment, and geometric reconstruction of damaged blades form the basis for subsequent remanufacturing repair and precision machining.

[0003] Current industrial repair processes use non-contact 3D scanning to obtain point clouds of the blade surface, then geometrically compare these point clouds with the nominal CAD model to determine the missing material areas. However, in-service blades undergo irreversible deformation, deviating from the original nominal model. Therefore, using the nominal CAD model as the repair benchmark leads to geometric mismatches and reduces repair accuracy.

[0004] Traditional methods such as geometric modeling, Boolean difference, surface fitting, and non-rigid registration heavily rely on geometric constraints or manual parameter adjustment. When blades have large-area defects, complex free-form surfaces, or abrupt changes in local curvature, these methods exhibit poor adaptability, slow convergence speed, high parameter sensitivity, and weak ability to preserve local features, failing to meet the requirements for automated repair of blades of different models, sizes, torsion angles, and thicknesses.

[0005] Deep learning-based point cloud inpainting methods can adaptively learn geometric distributions from massive datasets. Current neural network point cloud inpainting schemes fall into two categories. The first category is a global end-to-end single-stage network. This type of network takes the entire damaged model as input. For large-scale smooth surfaces like blades, using the entire damaged model as input makes it difficult for the network to focus on minute defects at the tip, leading edge, trailing edge, or surface, leading to distortion in the inpainted damaged area and unexpected geometric perturbations. The second category is a two-stage cascaded neural network, which first performs a global coarse reconstruction through a primary network, and then inputs the result into a secondary network for local refinement. This separate architecture requires two deep learning networks to be trained in series, resulting in high computational consumption, long training time, and slow processing speed for industrial deployment. More seriously, the secondary feature transfer across networks easily causes the loss of high-frequency geometric features, leading to large reconstruction errors. Therefore, a high-precision point cloud reconstruction technology based on a single-stage network is urgently needed.

[0006] A search of existing patent literature revealed the following: Patent CN106874628A discloses a method for repairing tip defects in aero-engine compressor blades, the core idea of ​​which is to perform three-dimensional reconstruction of the defective area at the blade tip. However, this solution is mainly applicable to specific tip defect scenarios and focuses on traditional geometric processing, lacking sufficient support for post-service overall creep, complex freeform surfaces, and deep learning-based completion. Patent CN110866969A discloses a three-dimensional reconstruction method for engine blades based on neural networks and point cloud registration, its core being global reconstruction or registration, without spatial perception and localization of small-scale defects. Patent CN115096891A discloses a blade defect identification method based on deep learning, but it mainly addresses two-dimensional defect identification or classification problems and cannot output a three-dimensional repair point cloud corresponding to the damaged area.

[0007] In summary, while existing technologies have proposed solutions and algorithms for blade tip local repair, global point cloud registration, 2D defect identification, and general point cloud completion, they all suffer from insufficient adaptability to industrial scenarios. There is an urgent need for an integrated and effective method that can achieve 3D scanning, point cloud preprocessing, 2D defect perception, 2D boundary-to-3D spatial inversion mapping, local defect truncation, multi-scale fractal completion, and high-fidelity solidified point cloud output. This method would transform the repair model from traditional "global blind completion" to "high-precision completion guided by defect perception," thus meeting the requirements of high-precision blade remanufacturing. Summary of the Invention

[0008] To address the challenge of automatically locating and accurately repairing defects in damaged blades under conditions of service deformation, size differences in defect areas, changes in freeform surface curvature, and abrupt changes in local edges, this application proposes a blade point cloud repair method and system based on damage perception and multi-scale fractal reconstruction.

[0009] The technical solution adopted in this application is as follows: a method for repairing leaf point clouds based on damage perception and multi-scale fractal reconstruction is proposed, including the following steps:

[0010] Step 1: Collect 3D point cloud data of the damaged blades of the aero-engine to be repaired and preprocess them to obtain the original aligned blade point cloud, and construct training samples containing various blade defects.

[0011] Step 2: Construct a defect detection network and a multi-scale fractal reconstruction network, and train the defect detection network and the multi-scale fractal reconstruction network based on training samples;

[0012] Step 3: Project the original aligned blade point cloud of the damaged blade into a two-dimensional grayscale feature map. Use the trained optimal defect detection network to process the two-dimensional grayscale feature map of the damaged blade to obtain a two-dimensional defect detection box. Map the two-dimensional defect detection box to a three-dimensional point cloud to obtain a three-dimensional defect-enclosed region. Trim the defect point cloud fragments according to the three-dimensional defect-enclosed region, and then input the defect point cloud fragments into the trained optimal multi-scale fractal reconstruction network to generate the complete point cloud of the damaged blade. Merge the complete point cloud of the damaged blade with the original aligned blade point cloud to obtain a complete three-dimensional blade point cloud.

[0013] Furthermore, in step one, a structured light 3D vision device is used to acquire the 3D point cloud of the damaged blade of the aero-engine to be repaired. The 3D point cloud is preprocessed, including statistical outlier removal, voxel mesh downsampling, and blade base removal. Then, solidified volume filling based on principal component analysis is performed to obtain a normalized original aligned blade point set.

[0014] Furthermore, statistical outlier removal involves calculating the average distance from each point in the point cloud to its neighboring points and then using an adaptive distance threshold to remove isolated noise points. Voxel grid downsampling utilizes a preset 3D grid size to spatially simplify and homogenize the density of the denoised point cloud, ensuring that scan data of different densities have a consistent spatial sampling scale before entering subsequent projection and reconstruction.

[0015] Furthermore, the steps for solidification volume filling based on principal component analysis include:

[0016] In the principal component analysis coordinate system, the point cloud after denoising and spatial sampling scale unification is aligned in attitude, and the aligned point cloud is projected onto the horizontal plane to construct a two-dimensional mesh index plane. The local extreme values ​​of the normal depth of the upper and lower surfaces in each mesh cell are extracted, and within the local extreme value interval, normal equidistant interpolation is performed at a preset step size to generate an internal filling lattice. Then, the internal filling lattice is mapped back to the original space using the inverse matrix of orthogonal transformation, and the empty shell point cloud of one side surface is converted into a three-dimensional lattice with internal solid thickness.

[0017] Furthermore, in step one, a training sample containing surface defects, edge defects, leaf tip defects, and composite defects is constructed using template elimination. The specific steps include:

[0018] Read the complete undamaged blade point cloud and the defect template point cloud; select the defect template rotation angle from 0°, 90°, 180° and 270°, scale the defect template according to the scaling factor, and select the translation position of the defect template in the principal component analysis coordinate system of the complete undamaged blade point cloud; search for the defect template coverage points based on KD-Tree and subtract them from the complete undamaged blade point cloud to obtain the damaged blade point cloud sample; the defect template includes surface wedge template, edge wedge template and blade tip cutting template.

[0019] Furthermore, the defect detection network in step two adopts a single-class defect detection network, and the multi-scale fractal reconstruction network includes a multi-resolution encoder and a point pyramid decoder.

[0020] In step three, the cropped defect point cloud fragments are processed by point number alignment, principal component alignment and normalization, and defect point cloud fragments at different resolution scales are obtained by iterative sampling of the farthest point.

[0021] The multi-resolution encoder extracts local geometric features from defect point cloud fragments at different resolution scales from the input and splices them together to form a global feature vector. The point pyramid decoder receives the global feature vector and performs a progressive decoding operation from coarse to fine, sequentially generating primary center points, secondary center points, and final detail points as the output of the completed point cloud.

[0022] Furthermore, step three, which projects the original aligned point cloud of the damaged blade into a two-dimensional grayscale feature map, includes:

[0023] The horizontal projection coordinates of the original aligned leaf point cloud are retained. The minimum points of the corresponding dimensions are subtracted from the horizontal projection coordinates to translate them to the non-negative interval. The coordinates are then multiplied by a preset scaling factor to convert them into discrete two-dimensional pixel coordinates. The frequency of spatial points that coincide with each pixel coordinate is counted. The frequency of each pixel is linearly stretched to the grayscale interval, and the final two-dimensional grayscale feature map is output.

[0024] Furthermore, the specific steps in step three of mapping the two-dimensional defect detection box to the three-dimensional point cloud to obtain the three-dimensional defect bounding region include: firstly, extracting the diagonal limit pixel boundary of the two-dimensional defect detection box, and combining the center position and width and height dimensions of the two-dimensional defect detection box to expand the pixel boundary of the detection region outward; then, dividing the extreme value of the expanded two-dimensional discrete pixel boundary by a preset scale factor, and compensating for the corresponding three-dimensional physical space translation, thereby inversely solving and restoring it to the absolute physical boundary on the three-dimensional spatial principal plane; finally, using the inversely calculated physical boundary of the three-dimensional spatial principal plane, combined with the global minimum depth boundary and maximum depth boundary of the three-dimensional point cloud of the damaged blade in the vertical normal direction, jointly constructing a three-dimensional spatial logical bounding box that penetrates the thickness of the blade, and performing spatial truncation in the original aligned blade point cloud to output the local three-dimensional defect bounding box.

[0025] Furthermore, the specific steps for fusing the completed point cloud of the damaged blade with the original aligned blade point cloud in step three to obtain a complete 3D blade point cloud include: performing an inverse normalization operation on the predicted completed point cloud; using the orthogonal inverse transformation matrix of principal component analysis to restore the inverse normalized completed point cloud to its original spatial pose; and using spatial nearest neighbor search to find and remove overlapping redundant points between the completed point cloud and the undamaged point cloud sectional area, outputting a complete 3D blade point cloud.

[0026] This application also proposes a blade point cloud repair system based on damage perception and multi-scale fractal reconstruction, which includes the following modules:

[0027] Point cloud acquisition module: used to acquire three-dimensional point cloud data of the damaged blades of the aero-engine to be repaired;

[0028] Preprocessing module: Used to process the 3D point cloud data of the damaged blade into the original aligned blade point cloud;

[0029] 2D projection module: used to project the original aligned leaf point cloud into a 2D grayscale feature map;

[0030] Defect perception module: Used to process the two-dimensional grayscale feature map of the damaged blade through the trained optimal defect detection network to obtain a two-dimensional defect detection box;

[0031] 3D positioning module: used to map the 2D defect detection box to the 3D point cloud to obtain the 3D defect-enclosed region;

[0032] Defect clipping module: It is used to clip defect point cloud fragments based on the 3D defect-enclosed region, and perform point number alignment, principal component alignment and normalization on the clipped defect point cloud fragments. It obtains defect point cloud fragments at different resolution scales by iterative sampling of the farthest point.

[0033] Fractal Reconstruction Module: Used to process defect point cloud segments at different resolution scales through the trained optimal multi-scale fractal reconstruction network to obtain the complete point cloud of the damaged leaf.

[0034] The fusion output module is used to fuse the completed point cloud of the damaged blade with the original aligned point cloud to obtain a complete 3D blade point cloud. It is then exported in a standard general point cloud format, directly packaged, and uploaded and pushed to the CNC control terminal of the laser cladding intelligent remanufacturing machine tool.

[0035] The advantages of this application over the prior art are as follows:

[0036] 1. A point cloud feature mapping and solidification mechanism that highly fits the physical truth is constructed. In response to the single-sided scanning defect caused by structured light scanning, this application proposes a one-dimensional normal adaptive solidification interpolation algorithm based on depth extremum constraints. Combined with dimensionality reduction projection and density clustering, it realizes self-supervised extraction with zero manual annotation cost.

[0037] 2. It achieves deep decoupling between spatial perception and localization completion, reducing the massive end-to-end 3D full-scene reconstruction problem to a smaller-scale problem of target bounding box detection and local reconstruction of independent defect regions. While significantly improving the dynamic allocation of computing resources, it avoids the unexpected geometric perturbations caused by neural networks to lossless regions.

[0038] 3. A closed-loop reconstruction of pyramid cascade fractal constraints from coarse to fine is achieved. The model follows the natural growth law of "topological skeleton first, then detailed curvature", which restores the original morphological features of complex freeform surfaces to a very high degree. It can be directly connected to industrial-grade additive manufacturing equipment to perform adaptive repair. Attached Figure Description

[0039] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0040] Figure 1 This is the overall flowchart of the point cloud repair method for aero-engine blades in this application.

[0041] Figure 2 This is a schematic diagram of the point cloud repair system for aero-engine blades in this application.

[0042] Figure 3 This is a schematic diagram of the original 3D scanned blade point cloud model. In the figure, B1~B5 represent the 3D scanned point cloud models of five different blades.

[0043] Figure 4 This is a flowchart of the volume filling process proposed in this application.

[0044] Figure 5 Flowchart for constructing training samples for this application.

[0045] Figure 6 This is a schematic diagram of a defect template structure. The diagram shows two types of defect mesh templates, filled with a 3D point cloud, and used in blade repair.

[0046] Figure 7 To generate a schematic diagram for the training dataset, (a) shows a schematic diagram of surface defects at different locations on different blades, (b) shows a schematic diagram of edge defects at different locations on different blades, (c) shows a schematic diagram of tip defects at different locations on different blades, and (d) shows a schematic diagram of compound defects at different locations on different blades.

[0047] Figure 8This is a flowchart of the operation of the defect perception module in this application.

[0048] Figure 9 for Figure 2 Two-dimensional projection feature maps and defect label diagrams of five different blades.

[0049] Figure 10 This is a flowchart illustrating the operation of the multi-scale fractal reconstruction module in this application.

[0050] Figure 11 for Figure 2 A schematic diagram comparing the defect detection and point cloud completion results of five different types of blades. Detailed Implementation

[0051] like Figures 1 to 11 As shown, this application provides a method for repairing leaf point clouds based on damage perception and multi-scale fractal reconstruction, including the following steps:

[0052] Step 1: Collect 3D point cloud data of damaged blades and preprocess it to obtain the original aligned blade point cloud. Construct training samples containing various blade defects, specifically including:

[0053] Step S101: Use a structured light 3D vision device to acquire the 3D point cloud of the damaged blade of the aero-engine to be repaired;

[0054] Step S102: Preprocess the three-dimensional point cloud to obtain the aligned blade point cloud, hereinafter referred to as the original aligned blade point cloud; and generate blade point cloud training samples from the complete and undamaged blade point cloud using the template elimination method.

[0055] Step 2: Construct a defect detection network and a multi-scale fractal reconstruction network, and train these networks using training samples. Specifically, this includes:

[0056] Step S103: Project the leaf point cloud training samples into a two-dimensional grayscale feature map;

[0057] Step S104: Use a defect detection network to perform defect detection on the two-dimensional grayscale feature map to obtain a two-dimensional defect detection box;

[0058] Step S105: Map the two-dimensional defect detection box to the three-dimensional point cloud to obtain the three-dimensional defect-enclosed region; trim the defect point cloud fragments according to the three-dimensional defect-enclosed region;

[0059] Step S106: Input the defective point cloud fragment into the multi-scale fractal reconstruction network, and generate a complete point cloud through multi-resolution encoding and point pyramid decoding;

[0060] Step S107: Merge the completed point cloud with the original blade point cloud to obtain a complete 3D blade point cloud;

[0061] Step 3: Project the original aligned blade point cloud of the damaged blade into a two-dimensional grayscale feature map. Use the trained optimal defect detection network to process the two-dimensional grayscale feature map of the damaged blade to obtain a two-dimensional defect detection box. Map the two-dimensional defect detection box to a three-dimensional point cloud to obtain a three-dimensional defect-enclosed region. Trim the defect point cloud fragments according to the three-dimensional defect-enclosed region, and then input the defect point cloud fragments into the trained optimal multi-scale fractal reconstruction network to generate the complete point cloud of the damaged blade. Merge the complete point cloud of the damaged blade with the original aligned blade point cloud to obtain a complete three-dimensional blade point cloud.

[0062] The preprocessing in step S102 includes: performing statistical outlier removal, voxel mesh downsampling, and blade base removal on the 3D point cloud of the damaged blade obtained by scanning, and performing solidified volume filling based on principal component analysis to obtain a normalized set of original aligned blade points.

[0063] Statistical outlier removal involves calculating the average distance from each point in the point cloud to its neighboring points and then using an adaptive distance threshold to eliminate isolated noise points. Voxel mesh downsampling utilizes a preset 3D mesh size to spatially simplify and homogenize the density of the denoised point cloud.

[0064] In one specific embodiment, when removing outliers, outliers are identified based on the number of neighboring points and the standard deviation multiple; the preset three-dimensional voxel size for voxel downsampling is determined by the average nearest neighbor distance of the point cloud, so that scan data of different densities have a consistent spatial sampling scale before entering subsequent projection and reconstruction.

[0065] Specifically, the solidified volume filling based on principal component analysis involves aligning the point cloud in the principal component analysis coordinate system, projecting the aligned blade point cloud onto a horizontal plane to construct a two-dimensional mesh index plane, extracting the local extrema of the normal depth of the upper and lower surfaces within each mesh cell, and performing equidistant normal interpolation within the local extrema interval at a preset step size to generate an internal filling lattice; subsequently, using the inverse matrix of orthogonal transformation, the internal filling lattice is mapped back to the original space, transforming the empty shell point cloud of a single surface into a three-dimensional lattice with internal solid thickness.

[0066] In a specific embodiment, during the principal component analysis pose unification stage, to eliminate pose differences caused by different scanning devices, this embodiment uses principal component analysis (PCA) to align the pose of the point cloud. Let the point cloud set be... The point cloud's geometric centroid is First, calculate the covariance matrix of the point cloud. :

[0067] (1);

[0068] Subsequently, the covariance matrix was analyzed. Perform eigenvalue decomposition:

[0069] (2);

[0070] in, For eigenvalues, These are the eigenvectors. The eigenvalues ​​are sorted by size, and the eigenvectors corresponding to the three largest eigenvalues ​​are selected. As a new coordinate system, the blades are aligned with the main axis direction.

[0071] For scanned point clouds containing only the outer surface, this embodiment further performs volume filling. Let the average nearest neighbor distance of the point cloud be... ,by As the initial voxel downsampling scale, the point cloud is projected onto the principal component coordinate system. For a plane, establish a two-dimensional indexed grid and record the local lower surface height at each grid location. and the height of the upper surface Then according to The step size is used to generate filling points along the normal direction. Finally, the inverse principal component transformation is performed to return to the original principal component analysis coordinate system, thus obtaining a solidified blade point cloud with thickness information.

[0072] To improve the adaptability of the defect detection and point cloud completion model to different damage morphologies, this application constructs training samples containing surface defects, edge defects, blade tip defects, and composite defects using a template elimination method. The specific steps of the template elimination method in step S102 include: reading the complete undamaged blade point cloud and the defect template point cloud; selecting the defect template rotation angle from 0°, 90°, 180°, and 270°, scaling the defect template according to a scaling factor, and selecting the translation position of the defect template in the principal component analysis coordinate system of the complete undamaged blade point cloud; searching for defect template coverage points based on the KD-Tree and subtracting them from the complete undamaged blade point cloud to obtain damaged blade point cloud samples; the defect templates include surface wedge templates, edge wedge templates, and blade tip cutting templates.

[0073] In one specific embodiment, surface defect samples are generated by reading a defect template, selecting a candidate midpoint in the blade principal component analysis coordinate system, rotating the template by angles of 0°, 90°, 180°, and 270°, and scaling the defect template height by a ratio of 0.4 to 0.6. Then, a KD-Tree search is used to search for blade points covered by the defect template and subtract them from the complete, undamaged blade point cloud. Composite defect samples are generated by combining three types of defects, and overlapping points are removed using a KD-Tree search with a radius threshold of twice the average nearest neighbor distance. In the sample enhancement stage, the damaged point cloud and corresponding missing label points are uniformly transformed to the principal component analysis coordinate system, and scale enhancement is performed along the first principal component direction at a random ratio of 1.00 to 1.08; two enhanced samples are generated for each original sample.

[0074] The specific steps for generating the two-dimensional grayscale feature map in step S103 include: retaining the horizontal projection coordinates of the leaf point cloud training samples; subtracting the minimum points of the corresponding dimensions from the horizontal projection coordinates to translate them to a non-negative interval; and multiplying them by a preset scaling factor to convert them into discrete two-dimensional pixel coordinates. The frequency of spatial points coinciding with each pixel coordinate is counted, and the frequency of each pixel is linearly stretched to the grayscale interval to output the final two-dimensional grayscale feature map.

[0075] In step S104, the YOLOv5 defect detection network is used to detect defects in the two-dimensional grayscale feature map. In step S105, the specific steps of mapping the two-dimensional defect detection box to the three-dimensional point cloud include: inputting the two-dimensional grayscale feature map into the pre-trained YOLOv5 network, outputting a two-dimensional defect detection box containing normalized center coordinates and width and height dimensions, using inverse transformation to convert the boundary of the two-dimensional defect detection box into three-dimensional absolute physical coordinates, and extracting the three-dimensional defect-enclosed region from the original aligned leaf point cloud.

[0076] In one specific embodiment, a single-class defect detection network (YOLOv5) is used in the defect detection stage. Training samples consist of a 3D point cloud projection map and defect labels; the labels are obtained by clustering bounding boxes of the missing point clouds in the 2D projection plane. In one implementation, DBSCAN clustering is used for the missing point clouds, with the cluster radius being twice the average nearest neighbor distance of the missing points and a minimum number of points being 5. During network training, the input image size is 640×640, the maximum number of training epochs is set to 300, the batch size is 16, the optimizer is SGD, the initial learning rate is 0.01, the momentum is 0.937, and the weight decay is 0.0005; during inference, the confidence threshold is 0.25, the non-maximum suppression (NMS) IoU threshold is 0.45, and the maximum number of detection boxes per image is 1000.

[0077] To adapt the 3D structure to the YOLOv5 2D detection network, the physical coordinates are mapped to pixel coordinates by calculating the scale factor of the 2D projection. Let the leaf point cloud be... The minimum boundary of the plane is The maximum boundary is The image width and height are both 640 pixels (i.e. Then the pixel coordinates satisfy:

[0078] (3);

[0079] (4);

[0080] Among them, the scale scaling factor and The calculation formula is:

[0081] (5);

[0082] (6);

[0083] A density map is formed by counting the number of points falling within each pixel. The density map is then normalized to an 8-bit grayscale image (0-255):

[0084] (7);

[0085] in, It is an 8-bit grayscale image. and These are the minimum and maximum pixel counts in the grayscale density image, respectively.

[0086] After the two-dimensional defect detection box is output, the scaling factor saved during the dimensionality reduction projection stage is used. , The physical translation amount is used to inversely map the 2D defect detection box to 3D physical space. Specifically, firstly, the diagonal limit pixel boundaries of the 2D defect detection box are extracted, and then, combined with the center position and width and height dimensions of the detection box, the pixel boundaries of the detection area are appropriately expanded outward to retain sufficient local geometric context features. Subsequently, the extreme values ​​of the expanded 2D discrete pixel boundaries are divided by a preset 2D projection scale scaling factor. and The corresponding 3D physical space translation is then added as compensation, thereby reversibly resolving and restoring the absolute physical boundary on the 3D principal plane. Finally, using the 3D principal plane physical boundary obtained from this inverse calculation, combined with the global minimum and maximum depth boundaries of the entire lossless blade point cloud in the vertical normal direction, a 3D spatial logical bounding box that spans the blade thickness is jointly constructed. Spatial truncation is then performed in the original aligned solidified point cloud to output a local 3D defect bounding box.

[0087] The multi-scale fractal reconstruction network (PF-net) in step S106 includes a multi-resolution encoder (MRE) and a point pyramid decoder (PPD). The multi-resolution encoder extracts local geometric features from defect point cloud fragments of different resolution scales from the input and splices them into a global feature vector. The point pyramid decoder receives the global feature vector and performs a progressive decoding operation from coarse to fine, generating primary center points, secondary center points, and final detail points in sequence as the output of the completed point cloud.

[0088] In one specific embodiment, the local completion stage performs point number alignment, principal component alignment, and normalization on the damaged point cloud fragments within the 3D defect bounding box. During point number alignment, if the number of points is greater than the target number, the voxel scale is adaptively selected and downsampled in the principal component analysis coordinate system based on the median nearest neighbor distance; if the number of points is less than the target number, it is supplemented by random repeated sampling. The input point cloud alignment is 4096 points, and two low-resolution inputs of 2048 points and 1024 points are obtained through iterative farthest point sampling (IFPS), thus forming a three-scale point cloud input of 4096 / 2048 / 1024.

[0089] The multi-resolution encoder processes inputs of 4096, 2048, and 1024 points respectively, and extracts 64-dimensional, 128-dimensional, 256-dimensional, 512-dimensional, and 1024-dimensional features sequentially through a parallel multilayer perceptron, which are then fused into 1920-dimensional global geometric features. The point pyramid decoder generates primary center points step by step through fully connected layers and one-dimensional convolutional layers. Secondary center point Details of completing the target points .

[0090] When training the multi-scale fractal reconstruction network, hierarchical chamfer distance (CD) is used to constrain the final detail points, primary center points, and secondary center points. Two point cloud sets are used. and The chamfer distance between them is defined as:

[0091] (8).

[0092] Let the final detail output be The primary center point output is The secondary center point output is The corresponding real point set is , and The overall joint reconstruction loss satisfy:

[0093] (9).

[0094] In one implementation, the first 30 rounds of training are taken... , Rounds 30 to 79 , After the 80th round, take , The optimizer is Adam, the learning rate is 0.0001, the weight decay is 0.001, and a validation loss, platform decay learning rate, and early stopping strategy are adopted.

[0095] The specific steps in step S107 to fuse the completed point cloud with the original blade point cloud to obtain a complete 3D blade point cloud include: performing an inverse normalization operation on the predicted completed point cloud; using the orthogonal inverse transformation matrix of principal component analysis to restore the inverse normalized completed point cloud to its original spatial pose; and using spatial nearest neighbor search to find and remove overlapping redundant points between the completed point cloud and the undamaged point cloud truncated section, and outputting a complete 3D blade point cloud.

[0096] In one specific embodiment, after outputting the final completed point cloud, the physical scale of the completed point cloud is restored by inverse normalization using the scale scaling factor and geometric centroid of the truncated record. Then, the inverse orthogonal transformation matrix of principal component analysis is used to rotate the completed point cloud inversely to the original scan space pose. Finally, the overlapping and redundant points at the truncated section are removed by the spatial nearest neighbor search algorithm, and the undamaged matrix region of the completed point cloud and the original blade point cloud are stitched together using 3D topological smoothing, ultimately outputting a high-fidelity complete 3D blade point cloud model.

[0097] The present application will be further described below with reference to the accompanying drawings and specific embodiments.

[0098] The objects of detection, analysis, and reconstruction in this embodiment are five sets of aero-engine compressor blades that have suffered complex damage after service. The base material of these blades is titanium alloy. Figure 1 and Figure 2As shown, this embodiment provides a point cloud repair method for aero-engine blades based on damage perception and multi-scale fractal reconstruction. The entire process software and hardware operating environment configuration is as follows: Ubuntu 20.04 operating system, PyTorch 1.10 deep learning framework, and the underlying core parallel module adopts the unified computing architecture (CUDA) for accelerated compilation; the hardware platform is equipped with an Intel Core i9-13900K processor, 64GB of system memory, and a single RTX4090 graphics processor.

[0099] Example 1: Damaged point cloud acquisition, spatial pose alignment, and adaptive solidification volume filling. This example involves... Figure 1 Steps S101 to S102, and Figure 2 The system includes a point cloud acquisition module and a preprocessing module. The specific execution flow is as follows:

[0100] A high-precision non-contact structured light 3D scanner was used to perform a full-field morphological scan of the blade. The initial state of the disordered 3D discrete point cloud output from a single scan is as follows: Figure 3 As shown. Taking the second group of blades as an example, the initial number of points is 245,320. The K-nearest neighbor algorithm is used to construct spatial topological relationships and estimate the initial surface normal vector. The entire field of 245,320 points is traversed. For each coordinate point, its 50 nearest neighbors are retrieved, and the average Euclidean distance from that point to all 50 neighbors is calculated. After the traversal, the average distance of the global point cloud is calculated to be 0.124 mm, and the standard deviation of the spatial distance distribution is 0.042 mm. A standard deviation threshold of 0.5 is set. The adaptive maximum distance truncation threshold for denoising is calculated. Boolean mask filtering was performed on the entire point cloud to remove 3412 isolated outliers with an average neighborhood distance greater than 0.145 mm, leaving a total of 241908 clean point clouds that satisfy the Gaussian distance field distribution.

[0101] Adaptive downsampling and attitude alignment are performed on the clean point cloud. The spatial distance from each discrete point in the point cloud to its nearest neighbor is calculated, and the global average nearest neighbor distance is found to be 0.120 mm, which is defined as the base distance threshold. A 3D mesh is constructed using twice the base distance threshold (i.e., 0.240 mm) as the voxel edge length. Voxel mesh downsampling is performed on the clean point cloud, and the coordinates within the mesh are merged to output a normalized damaged leaf point cloud, which converges to 32415 points. Let the set of normalized point clouds after sampling be... Calculate the arithmetic mean of the spatial point set along the three coordinate axes, and extract the coordinates of the geometric centroid. mm. After subtracting the geometric centroid coordinates from all points, calculate the 3×3 three-dimensional spatial covariance matrix. Perform eigenvalue decomposition on this covariance matrix, and jointly construct an orthogonal attitude transformation matrix using the three orthogonal eigenvalues ​​and their corresponding unit eigenvectors. Use this orthogonal attitude transformation matrix to perform rigid body rotation on the point cloud, making the blade axial length direction completely coincide with the first principal component axis of the principal axis coordinate system, thus eliminating pose deviation.

[0102] Perform adaptive solidification volume filling. This embodiment uses... Figure 4 The process described is as follows for point cloud completion: In the aligned principal axis coordinate system, the horizontal physical coordinates are discretized with a scaling factor of 4.167px / mm, and a 2D mesh index plane with a size of 512×120 is constructed by combining the origin translation. Then, the upper and lower boundary matrices of the global depth are initialized, and the sampling points are traversed. By comparing the normal depth of the current point with the matrix retention value, the local depth extrema of the lower and upper surfaces of each effective mesh region are dynamically overwritten and extracted. Based on this, with an equal step size of twice the base distance threshold, normal arithmetic interpolation is performed within the extreme value interval of the upper and lower surfaces, automatically deriving a total of 12540 internal filling points. Finally, the above filling points are restored to the original physical space through inverse translation, scaling, and inverse orthogonal transformation, ultimately outputting a high-fidelity solid point cloud that combines surface morphology and solid thickness, expanding the total number of points to 44955.

[0103] Example 2: Dimensionality Reduction Feature Projection, 2D Density Map Construction, and Automatic Generation of Defect Detection Labels. This example involves... Figure 1 Step S103 in the process, and Figure 2 The two-dimensional projection module in the middle, combined with Figure 5 , Figure 6 and Figure 7 The process of constructing training samples and generating datasets is shown.

[0104] Training samples of damaged blades and corresponding defect-labeled point clouds are generated using template elimination. The complete, undamaged blade point cloud and preset defect templates are read, including surface wedge templates, edge wedge templates, and tip cut templates. When generating a single damaged sample, a preset horizontal rotation angle is randomly selected, and the defect template is scaled in 3D using a randomly generated scaling factor. A translation position conforming to boundary constraints is randomly selected within the principal component analysis coordinate system of the complete, undamaged blade point cloud. Subsequently, a spatial search tree is constructed to retrieve the blade matrix point cloud enveloped by the geometric space of the defect template. The retrieved covered point set is forcibly stripped from the complete, undamaged blade point cloud, and the remaining main body constitutes the materialized damaged blade point cloud sample simulating real-world working conditions. Simultaneously, the stripped covered point set directly serves as the corresponding 3D defect-labeled point cloud for the damaged sample, thus enabling the batch generation of thousands of 3D damaged datasets.

[0105] Size extraction and adaptive dimensionality reduction scaling are performed, and the coordinates of the minimum boundary points in the 3D space of the 3D defect label point cloud are obtained through extreme value retrieval. mm, coordinates of the maximum boundary point mm. Subtracting the two, the spatial physical dimensions of the leaf point cloud are calculated as follows: lateral length 124.50 mm, longitudinal width 38.20 mm, and normal thickness 3.40 mm. The standard pixel resolution of the target 2D feature image is set to 640×640. A buffer margin of 1 pixel is reserved at the left, right, top, and bottom boundaries of the image, resulting in an effective pixel area of ​​638×638. The affine scaling factors (scale scaling factors) from the horizontal and vertical spatial dimensions are obtained: .

[0106] Detection labels are automatically generated by performing the same non-negative translation, scaling factor multiplication, and rounding on the real defect label point cloud, converting it into a two-dimensional pixel coordinate set of defects. The global average nearest neighbor distance of the defect point cloud is calculated to be 0.115 mm, and twice this value (i.e., 0.230 mm) is used as the neighborhood search radius. The minimum density threshold for core points is set to 5, and the data is input into a density-based spatial clustering algorithm for unsupervised clustering, automatically separating independent local defect clusters. The absolute boundary extrema of each independent local defect cluster in the pixel plane are extracted, respectively. px, px, px, The normalized target detection bounding box (i.e., the 2D defect detection box) center coordinates and relative width and height dimensions are calculated using the following formula: , The coordinates of the minimum boundary point of the current sample and the scale factor are serialized, packaged, and saved in a file. 5000 sample files are generated using a random number seed and strictly divided into 3500 training images, 1000 validation images, and 500 test images according to a 7:2:1 ratio.

[0107] Example 3: Defect-aware network training, feature inference, and 3D spatial inverse extraction. This example involves... Figure 1 Steps S104 to S105, and Figure 2 The defect perception module and 3D localization module are included. The specific processing flow relies on... Figure 8 The defect detection module process shown is executed as follows:

[0108] A single-class object detection network based on a feature pyramid path aggregation structure was constructed, with the class localized as defect. During model training, the input image size was 640×640 pixels, the batch size for a single forward propagation was configured to be 16, and the maximum number of training iterations was set to 300. A stochastic gradient descent optimizer was used to dynamically update the network weight matrix, with a momentum factor of 0.937 and a damped weight decay regularization coefficient of 0.0005. The initial learning rate was set to 0.01, with the first three iterations configured as a linear warm-up period, followed by linear learning rate scheduling, and an early stopping strategy was used to preserve the optimal defect-aware inference model. During the training loop, the bounding box regression loss, confidence loss, and classification loss were jointly calculated, with the relevant loss weights executed according to the YOLOv5 training hyperparameter configuration. After 300 complete backpropagation iterations, the network regression loss steadily converged from the initial 0.085 to 0.012, solidifying the optimal defect-aware inference model. The persistent spatial registration parameter dictionary corresponding to the sample was read to parse the minimum physical boundary coordinates. mm, scale factor px / mm. Using a fixed pixel width of 640 and height of 640, the predicted bounding box is first calculated as a two-dimensional discrete pixel boundary extremum using the formula: , Similarly, the calculation yields... , Substituting the discrete pixel boundaries into the inverse affine transformation formula, we can inversely calculate the corresponding three-dimensional absolute physical coordinate boundaries on the principal plane: and .

[0109] Spatial segmentation and separation are performed. The coordinates of the solidified point cloud of the original damaged blade are traversed to construct a cylindrical spatial bounding box that runs through the vertical normal. A local 3D defect point cloud fragment containing 5640 points is obtained and used as the local input of the fractal reconstruction network. Spatial points that do not meet the conditions are stored in the non-defect matrix point cloud array, realizing accurate segmentation of local spatial features and data decoupling.

[0110] Example 4: Multi-resolution coding and fractal geometry reconstruction of local defect point cloud fragments. This example involves... Figure 1 Step S106 in the process, and Figure 2 The fractal reconstruction module within it. The specific processing architecture is... Figure 10 The multi-scale fractal reconstruction module workflow shown in the exhibition supports the following:

[0111] Cascaded feature encoding extraction is performed. Through iterative farthest point sampling, 2048 and 1024 points are extracted from the baseline feature stream, respectively, forming medium- and low-resolution subsets. The feature streams at the three resolution scales are input into a parallel multilayer perceptron for point-level dimensionality upscaling. Successive convolutional operations expand the channel dimensions to 64, 128, 256, 512, and 1024 dimensions sequentially. When the dimensions reach 128, 256, 512, and 1024, global max pooling is performed on the extracted feature tensors, and the four extracted features are directly concatenated along the channel dimensions to form a 1920-dimensional global feature vector. Subsequently, this vector enters the fully connected feature convergence module, undergoing sequential 1024-dimensional, 512-dimensional, and 256-dimensional feature mapping to output a 256-dimensional deep global latent variable.

[0112] Pyramid fractal folding decoding generation is performed. The 256-dimensional global latent variable input point pyramid decoder undergoes a three-order progressive reconstruction: First, a first-level fully connected network with 192 neurons maps and outputs 64 primary center points to establish the macroscopic skeleton; then, the primary center point features are concatenated with the latent variables, and a second-level decoding network maps and outputs 128 secondary center points to construct the transition surface; finally, the features generated in the first two levels are concatenated again and fed into the final detail decoding path, and the number of channels is gradually reduced from 512 to 48 through multiple one-dimensional convolutions, directly reconstructing and generating a high-fidelity final detail point cloud tensor containing 2048 three-dimensional coordinates, completing the geometric reconstruction of locally damaged defect segments.

[0113] Example 5: Optimized Training of Multi-Scale Fractal Reconstruction Networks. This example involves the fast convergence and optimal approximation of global error for multi-scale fractal reconstruction networks. The specific implementation method is as follows:

[0114] Multi-threaded parallel processing is configured to accelerate the underlying distance metric. The bidirectional chamfer distance is rapidly reduced by calculating the mean bidirectional nearest neighbor Euclidean distance between the predicted point cloud and the labeled point cloud. A multi-stage composite loss function is used for dynamic updates. Furthest point sampling is used to extract subsets of 64 and 128 points from the complete labeled point cloud as supervision at each level. The weights of the primary and secondary centroid loss are dynamically adjusted over 1001 training epochs: 0.01 and 0.02 are used for the first 30 epochs to prioritize fitting the macroscopic topology; from epochs 30 to 79, they are increased to 0.05 and 0.10 to optimize the transition surface; from epoch 80 onwards, they are fixed at 0.10 and 0.20 to approximate local high-frequency details.

[0115] Hyperparameter control and model solidification are implemented. The single forward batch size is configured to be 16, an adaptive moment estimation optimizer is used, the initial learning rate is 0.0001, and the weight decay coefficient is 0.001. Combined with a dynamic learning rate decay and early stopping mechanism based on validation set loss (100 rounds), the final detail point cloud chamfer distance loss is evaluated after each iteration, and the optimal network state file is persistently saved.

[0116] Example 6: Closed-loop forward inference, spatial coordinate reconstruction, and two-way geometric error evaluation. This example involves... Figure 1 Step S107 in the middle, and Figure 2 The fusion output module in the system. For a real titanium alloy blade that suffered a large-curvature tip impact defect, closed-loop forward inference, reverse calculation, point cloud smoothing and stitching, and morphological restoration assessment were performed. The input-output comparison and results at the system level are shown in [reference needed]. Figure 11 .

[0117] The process involves closed-loop forward inference and inverse solving. The damaged blade is input into the solidified defect perception module, which precisely extracts a local 3D defect fragment containing 5640 points. Point alignment and unit sphere space normalization are performed, and the geometric centroid and scale factor of the current fragment are recorded. The normalized tensor is input into the optimal multi-scale fractal reconstruction model, and forward propagation directly outputs a spatial coordinate matrix containing 2048 points. Gradient calculation is disabled, and the coordinates of the 2048 predicted detail points are multiplied by a pre-stored scale factor of 11.45, followed by center point compensation translation to back-calculate the absolute physical scale. Subsequently, a rigid body inverse rotation is performed on the completed point cloud using the principal component orthogonal inverse transformation matrix to accurately restore the original spatial pose.

[0118] Precision quantitative verification and evaluation. In order to conduct a rigorous and specific quantitative verification of the repair results, the repair model and the high-fidelity digital nominal CAD discrete standard point cloud (as the real physical surface label value) of this type of blade are mapped to each other in a bidirectional nearest neighbor distance field: accuracy spatial deviation measurement, traversing each point in the predicted repair point cloud set, and using the nearest neighbor algorithm to calculate its nearest neighbor Euclidean distance to the standard point cloud. The mathematical relationship strictly satisfies the following formula (10), and the average one-way geometric projection error of the whole field is obtained as follows. mm:

[0119] (10);

[0120] in, To predict and complete the point cloud, This corresponds to a real, complete point cloud or a labeled point cloud.

[0121] Integrity Coverage Calculation: Traverse each point in the standard point cloud and use the nearest neighbor algorithm to calculate its nearest Euclidean distance to the predicted repair point cloud set. The mathematical relationship strictly satisfies the following formula (11). The average one-way geometric projection error of the entire field is then obtained. mm:

[0122] (11).

[0123] Comprehensive reconstruction accuracy verification: The sum of the above two bidirectional projection errors and the arithmetic mean are calculated to obtain the final comprehensive repair geometric reconstruction error of the current blade. mm. Simultaneously, using the obtained distance matrix from each point to its nearest neighbor, the maximum bound of the bidirectional chamfer distance is calculated, which strictly satisfies the following equation (12): the bidirectional Hausdorff distance, with a calculated value of 0.739 mm.

[0124] (12).

[0125] Extensive experimental statistics show that the maximum local geometric deviation of the comprehensive reconstruction error for over 95% of the damaged blade samples in the test set strictly converges to within 1.070 mm. This application demonstrates extremely high fidelity accuracy on five types of blade samples with different curvatures and torsion angles, successfully reducing the median geometric reconstruction error of 7.148 mm generated by conventional single-stage networks in the absence of spatial priors to 0.189 mm. On a single-card graphics processor computing device, the entire process takes only 95.4 ms, well below the 100 ms threshold for industrial real-time online detection. The final output of a complete 3D high-fidelity point cloud model is exported in a standard general point cloud format, directly packaged, and uploaded to the CNC control terminal of the laser cladding intelligent remanufacturing machine tool. This serves as the geometric reference for executing high-precision additive cladding adaptive slicing path planning and five-axis linkage grinding and polishing trajectory generation, realizing high-precision, closed-loop, automated intelligent remanufacturing repair of damaged thin-walled freeform surface parts.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for repairing leaf point clouds based on damage perception and multi-scale fractal reconstruction, characterized in that: Includes the following steps: Step 1: Collect 3D point cloud data of the damaged blades of the aero-engine to be repaired and preprocess them to obtain the original aligned blade point cloud, and construct training samples containing various blade defects. Step 2: Construct a defect detection network and a multi-scale fractal reconstruction network, and train the defect detection network and the multi-scale fractal reconstruction network based on training samples; Step 3: Project the original aligned blade point cloud of the damaged blade into a two-dimensional grayscale feature map. Use the trained optimal defect detection network to process the two-dimensional grayscale feature map of the damaged blade to obtain a two-dimensional defect detection box. Then map the two-dimensional defect detection box to the three-dimensional point cloud to obtain the three-dimensional defect-enclosed region. The defect point cloud fragments are cropped based on the 3D defect-enclosed region, and then the defect point cloud fragments are input into the trained optimal multi-scale fractal reconstruction network to generate the complete point cloud of the damaged blade. The complete point cloud of the damaged blade is then fused with the original aligned blade point cloud to obtain the complete 3D blade point cloud.

2. The leaf point cloud repair method based on damage perception and multi-scale fractal reconstruction according to claim 1, characterized in that: In step one, a structured light 3D vision device is used to acquire the 3D point cloud of the damaged blade of the aero-engine to be repaired. The 3D point cloud is preprocessed, including statistical outlier removal, voxel mesh downsampling and blade base removal. Then, solidified volume filling based on principal component analysis is performed to obtain a normalized original aligned blade point set.

3. The leaf point cloud repair method based on damage perception and multi-scale fractal reconstruction according to claim 2, characterized in that: Statistical outlier removal involves calculating the average distance from each point in the point cloud to its neighboring points and then using an adaptive distance threshold to remove isolated noise points. Voxel grid downsampling uses a preset 3D grid size to spatially simplify and homogenize the density of the denoised point cloud, ensuring that scan data of different densities have a consistent spatial sampling scale before being projected and reconstructed.

4. A method for repairing leaf point clouds based on damage perception and multi-scale fractal reconstruction according to claim 2 or 3, characterized in that: The steps for solidification volume filling based on principal component analysis include: In the principal component analysis coordinate system, the point cloud after denoising and spatial sampling scale unification is aligned in attitude, and the aligned point cloud is projected onto the horizontal plane to construct a two-dimensional mesh index plane. The local extreme values ​​of the normal depth of the upper and lower surfaces in each mesh cell are extracted, and within the local extreme value interval, normal equidistant interpolation is performed at a preset step size to generate an internal filling lattice. Then, the internal filling lattice is mapped back to the original space using the inverse matrix of orthogonal transformation, and the empty shell point cloud of one side surface is converted into a three-dimensional lattice with internal solid thickness.

5. The method for repairing leaf point clouds based on damage perception and multi-scale fractal reconstruction according to claim 1, characterized in that: Step one involves constructing training samples containing surface defects, edge defects, blade tip defects, and combined defects using template elimination. The specific steps include: Read the complete undamaged blade point cloud and the defect template point cloud; select the defect template rotation angle from 0°, 90°, 180° and 270°, scale the defect template according to the scaling factor, and select the translation position of the defect template in the principal component analysis coordinate system of the complete undamaged blade point cloud; search for the defect template coverage points based on KD-Tree and subtract them from the complete undamaged blade point cloud to obtain the damaged blade point cloud sample; the defect template includes surface wedge template, edge wedge template and blade tip cutting template.

6. The leaf point cloud repair method based on damage perception and multi-scale fractal reconstruction according to claim 1, characterized in that: The defect detection network in step two uses a single-class defect detection network, and the multi-scale fractal reconstruction network includes a multi-resolution encoder and a point pyramid decoder. In step three, the cropped defect point cloud fragments are processed by point number alignment, principal component alignment and normalization, and defect point cloud fragments at different resolution scales are obtained by iterative sampling of the farthest point. The multi-resolution encoder extracts local geometric features from defect point cloud fragments at different resolution scales from the input and splices and fuses them into a global feature vector. The point pyramid decoder receives the global feature vector and performs a progressive decoding operation from coarse to fine, sequentially generating primary center points, secondary center points, and final detail points as the output of the completed point cloud.

7. The method for repairing leaf point clouds based on damage perception and multi-scale fractal reconstruction according to claim 1, characterized in that: Step three involves projecting the original aligned point cloud of the damaged leaf into a two-dimensional grayscale feature map, which includes: The horizontal projection coordinates of the original aligned leaf point cloud are retained. The minimum points of the corresponding dimensions are subtracted from the horizontal projection coordinates to translate them to the non-negative interval. The coordinates are then multiplied by a preset scaling factor to convert them into discrete two-dimensional pixel coordinates. The frequency of spatial points that coincide with each pixel coordinate is counted. The frequency of each pixel is linearly stretched to the grayscale interval, and the final two-dimensional grayscale feature map is output.

8. The leaf point cloud repair method based on damage perception and multi-scale fractal reconstruction according to claim 1, characterized in that: Step 3, mapping the 2D defect detection box to the 3D point cloud to obtain the 3D defect bounding region, includes the following steps: First, extract the diagonal limit pixel boundary of the 2D defect detection box, and combine the center position and width and height of the 2D defect detection box to expand the pixel boundary of the detection region outward; then, divide the extreme value of the expanded 2D discrete pixel boundary by a preset scale factor and compensate for the corresponding 3D physical space translation, thereby reversibly solving and restoring it to the absolute physical boundary on the 3D spatial principal plane; finally, using the reverse-calculated 3D spatial principal plane physical boundary, combined with the global minimum depth boundary and maximum depth boundary of the 3D point cloud of the damaged blade in the vertical normal direction, jointly construct a 3D spatial logical bounding box that runs through the blade thickness, and perform spatial truncation in the original aligned blade point cloud to output the local 3D defect bounding box.

9. The method for repairing leaf point clouds based on damage perception and multi-scale fractal reconstruction according to claim 1, characterized in that: The specific steps for fusing the completed point cloud of the damaged blade with the original aligned blade point cloud in step three to obtain a complete 3D blade point cloud include: performing an inverse normalization operation on the predicted completed point cloud; using the orthogonal inverse transformation matrix of principal component analysis to restore the inverse normalized completed point cloud to its original spatial pose; and using spatial nearest neighbor search to find and remove overlapping redundant points between the completed point cloud and the undamaged point cloud truncated section, outputting a complete 3D blade point cloud.

10. A system for performing the leaf point cloud repair method based on damage perception and multi-scale fractal reconstruction as described in any one of claims 1-9, characterized in that: The system includes the following modules: Point cloud acquisition module: used to acquire three-dimensional point cloud data of the damaged blades of the aero-engine to be repaired; Preprocessing module: Used to process the 3D point cloud data of the damaged blade into the original aligned blade point cloud; 2D projection module: used to project the original aligned leaf point cloud into a 2D grayscale feature map; Defect perception module: Used to process the two-dimensional grayscale feature map of the damaged blade through the trained optimal defect detection network to obtain a two-dimensional defect detection box; 3D positioning module: used to map the 2D defect detection box to the 3D point cloud to obtain the 3D defect-enclosed region; Defect clipping module: It is used to clip defect point cloud fragments based on the 3D defect-enclosed region, and perform point number alignment, principal component alignment and normalization on the clipped defect point cloud fragments. It obtains defect point cloud fragments at different resolution scales by iterative sampling of the farthest point. Fractal Reconstruction Module: Used to process defect point cloud segments at different resolution scales through the trained optimal multi-scale fractal reconstruction network to obtain the complete point cloud of the damaged leaf. The fusion output module is used to fuse the completed point cloud of the damaged blade with the original aligned point cloud to obtain a complete 3D blade point cloud. It is then exported in a standard general point cloud format, directly packaged, and uploaded and pushed to the CNC control terminal of the laser cladding intelligent remanufacturing machine tool.

Images