Composite component surface defect recognition and repair value calculation method based on structure perception

Abstract

The invention relates to the technical field of digital repair of composite material components, in particular to a composite material component surface defect recognition and repair value calculation method based on structure perception, which comprises the following steps: acquiring three-dimensional measurement point cloud data of the surface of a composite material component; constructing an end-to-end network and a mixed loss function, establishing a data set, and optimizing the end-to-end network based on an adaptive moment estimation method; and taking the three-dimensional measurement point cloud data as input, and obtaining a repair value corresponding to each three-dimensional measurement point through the composite component surface repair value end-to-end calculation model. According to the method, the three-dimensional point cloud on the surface of the composite material component is taken as input, and the repair value of each measuring point is output, so that deep fusion of defect perception, curved surface reconstruction and repair amount calculation is realized, the engineering requirements of high precision and high efficiency are met, and the method is particularly suitable for an online repair process of large components such as aviation composite material skins and wind power blades.

Description

Technical Field

[0001] The present invention relates to the technical field of digital repair of composite components, and in particular to a method for surface defect recognition and repair value calculation of composite components based on structural perception. Background Art

[0002] With the widespread use of advanced composite materials in the aviation industry, the repair and maintenance of composite structures has become increasingly important. The lightweight and high-strength properties of composite materials make them an ideal choice for aircraft manufacturing. However, during use and maintenance, composite components may suffer damage or wear. Prompt repair and maintenance of composite components is crucial to ensure aircraft safety, performance, and longevity.

[0003] For repairing damage to composite components, the mainstream international approach is to use a stepped scarf repair technique, which removes the damaged area, lays a patch, and then cures it to restore structural integrity. However, controlling the geometric morphology of the repair area is difficult. The curing shrinkage of the resin matrix and the anisotropy of the fiber layup result in a 0.1-1.5mm step-difference protrusion on the surface of the repair area, which damages the aerodynamic shape and stealth performance. Academia and industry have proposed a variety of digital repair methods, but they still have significant shortcomings, such as:

[0004] 1. Point cloud processing methods based on threshold segmentation use curvature filtering and region growing algorithms to extract defect areas from 3D point clouds. These methods lack quantitative capabilities and can only locate damaged areas, but cannot calculate repair values. They also have poor adaptability, rely on manual adjustment of threshold parameters, and cannot adapt to the surface characteristics of different composite components.

[0005] 2. Some studies have attempted to apply networks such as PointNet and PointCNN to defect detection, but their generalization is poor and they are unable to regress the repair value of each measurement point. Summary of the Invention

[0006] To address the above technical issues, the present invention proposes a method for surface defect identification and repair value calculation for composite components based on structural perception. This method solves the technical problem that traditional methods cannot strike a balance between "accuracy, efficiency, and generalizability" when calculating repair values.

[0007] The technical problem to be solved by the present invention is achieved by adopting the following technical solutions:

[0008] A method for identifying surface defects of composite components and calculating repair values ​​based on structural perception includes the following steps:

[0009] S1. Obtain M three-dimensional measurement points p on the surface of the composite component. i 3D measurement point cloud data

[0010] S2. Build an end-to-end network for automatic identification of surface defect areas and quantitative calculation of repair values ​​for composite components;

[0011] S3. Construct a hybrid loss function L for the end-to-end network, including the defect area classification loss term L cls , repair value regression loss term L reg And the surface consistency loss term L consist ;

[0012] S4. Establish a data set and optimize the end-to-end network based on the adaptive moment estimation method to obtain an end-to-end calculation model for the surface repair value of composite components;

[0013] S5. 3D measurement point cloud data As input, the repair value Δp corresponding to each 3D measurement point is obtained through the end-to-end calculation model of the composite component surface repair value i .

[0014] As a further improvement of the present invention, the end-to-end network includes a method for converting any three-dimensional measurement point p on the surface of the composite component into a i Encoded as high-dimensional feature expression F i , and obtain 3D measurement point cloud data 3D point cloud feature encoding module of global context information g;

[0015] And, according to the high-dimensional feature expression F i and global context information to identify any 3D measurement point p i A three-dimensional structure perception module where the area is a defect area or a non-defect area;

[0016] And, according to the high-dimensional feature expression F i and global context information g to obtain any 3D measurement point p i Repair value Δp i 3D measurement point repair value regression module.

[0017] As a further improvement of the present invention, in the three-dimensional point cloud feature encoding module, the specific process includes the following steps:

[0018] S21, find each three-dimensional measurement point p based on KNN algorithm i l neighboring points N(i);

[0019] S22. Constructing graph convolutional layer And through three graph convolution layers To encode the three-dimensional measurement point p i The local structure of the feature set is obtained Then the three-dimensional measuring point p i The high-dimensional feature expression is Concat represents the concatenation of feature dimensions, and the expression is:

[0020]

[0021] in, l represents the l-th graph convolution layer; is an intermediate variable; and Represents the three-dimensional measurement point p i and p j characteristics;

[0022] S23. Based on the global maximum pooling operation, obtain the three-dimensional measurement point cloud data The global context information g, that is,

[0023] As a further improvement of the present invention, the graph convolution layer It consists of a 1D convolutional layer, a batch normalization layer, and a ReLU nonlinear activation layer.

[0024] As a further improvement of the present invention, in the three-dimensional structure perception module, the specific process is as follows:

[0025] Splice the global context information g into the high-dimensional feature representation F i Later, it is fed into a multi-layer perceptron MLP1, and the output is a 2-dimensional area value y i , representing the three-dimensional measurement point p i the region where you are located;

[0026] The multilayer perceptron MLP1 is composed of 3 fully connected layers and 2 ReLU nonlinear activation layers.

[0027] As a further improvement of the present invention, in the three-dimensional measurement point repair value regression module, the specific process is as follows:

[0028] Splice the global context information g into the high-dimensional feature representation F i Later, it is fed into a multi-layer perceptron MLP2, and the output is a 3-dimensional repair value Δp i , representing the three-dimensional measurement point p i Repair value Δp i =(Δx i , Δy i , Δz i );

[0029] The multilayer perceptron MLP2 consists of 4 fully connected layers and 3 ReLU nonlinear activation layers.

[0030] As a further improvement of the present invention, the expression of the hybrid loss function L is:

[0031] L=αL cls +βL reg +γL consist ,

[0032] Among them, α, β, and corpse are hyperparameters used to balance the impact of each loss term;

[0033] The defect area classification loss term L cls The expression is:

[0034]

[0035] Among them, y i The regional value output by MLP1; is the classification truth value;

[0036] The repair value regression loss term L reg The expression is:

[0037]

[0038] Among them, δ is a hyperparameter; Δp i is the repair value; is the repaired truth value;

[0039] The surface consistency loss term L consist The expression is:

[0040]

[0041] Among them, P intact is the set of all three-dimensional measurement points in the non-defective area.

[0042] As a further improvement of the present invention, in step S4, the specific process includes the following steps:

[0043] S41. Based on CAD modeling software, first establish 3D models of 50 different composite components and discretize them into 50 3D point cloud data;

[0044] S42. Randomly sample 20 points on each 3D point cloud data, and segment 20 surface areas with the sampling point as the center and r as the radius to obtain 3D measurement data. The measurement data is the true value.

[0045] S43. After obtaining 3D measurement data of 1000 composite component surface areas, randomly select an area on each 3D measurement data and perform random deformation processing. The 3D point cloud of the randomly selected area is the defect area, and the remaining point clouds are non-defect areas.

[0046] S44. Based on the adaptive moment estimation method, the parameters of the 3D point cloud feature encoding module, the 3D structure perception module, and the 3D measurement point repair value regression module are optimized until the hybrid loss function L converges, and finally an end-to-end calculation model of the surface repair value of the composite component is obtained.

[0047] The beneficial effects of the present invention are:

[0048] The present invention provides a method for surface defect identification and repair value calculation of composite components based on structural perception. It takes the three-dimensional point cloud of the composite component surface as input and outputs the repair value of each measuring point, thereby realizing the deep integration of defect perception, surface reconstruction and repair amount calculation, meeting the engineering requirements of high precision and high efficiency, and is particularly suitable for the online repair process of large components such as aviation composite skins and wind turbine blades. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0050] Figure 1 Flowchart of the present invention;

[0051] Figure 2 Schematic diagram of the structure of the graph convolution layer in the present invention;

[0052] Figure 3 Schematic diagram of the structure of the multilayer perceptron MLP1 in the present invention;

[0053] Figure 4 Schematic diagram of the structure of the multilayer perceptron MLP2 in the present invention;

[0054] Figure 5 Schematic diagram of the output results of the three-dimensional structure perception module in the present invention. DETAILED DESCRIPTION

[0055] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below with reference to the accompanying drawings and embodiments.

[0056] Based on existing technologies, this embodiment addresses the high reliability and digital maintenance requirements of the next generation of aviation equipment. Composite component repair technology urgently needs to overcome the following core issues:

[0057] (1) High-precision repair value calculation: Achieve sub-millimeter (≤0.1mm) repair value quantification;

[0058] (2) Full process automation, reducing manual intervention to less than 10%;

[0059] (3) Field applicability: supports deployment of embedded devices and adapts to real-time processing requirements under complex working conditions.

[0060] Regarding the above technical bottlenecks and requirements, please refer to Figure 1-Figure 5 This embodiment proposes a method for surface defect recognition and repair value calculation of composite components based on structural perception. By deeply integrating the functions of defect perception, surface reconstruction and repair value calculation, it overcomes the problem of "accuracy-efficiency-generalization" that cannot be achieved simultaneously in traditional methods, and provides an innovative solution for the digital repair of composite components. The method includes the following steps:

[0061] S1. Obtain M three-dimensional measurement points p on the surface of the composite component. i 3D measurement point cloud data In this embodiment, a three-dimensional laser measurement sensor or other three-dimensional measurement equipment is used to collect point cloud data.

[0062] S2. Construct an end-to-end network for realizing automatic recognition of surface defect areas of composite components and quantitative calculation of repair values. The end-to-end network includes a method for automatically identifying any three-dimensional measurement point p on the surface of the composite component. i Encoded as high-dimensional feature expression F i , and obtain 3D measurement point cloud data The 3D point cloud feature encoding module of the global context information g. In the 3D point cloud feature encoding module, the specific process includes the following steps:

[0063] S21, find each three-dimensional measurement point p based on KNN algorithm i The l neighboring points N(i).

[0064] S22. Constructing graph convolutional layer And through three graph convolution layers To encode the three-dimensional measurement point p i The local structure of the feature set is obtained Then the three-dimensional measuring point p i The high-dimensional feature expression is Where concat represents the concatenation of feature dimensions.

[0065] In this embodiment, the three-dimensional measurement point cloud data of the composite component surface is First, based on the KNN algorithm, find the k neighboring points of each 3D measurement point, such as the neighboring points and Then design the graph convolution layer To encode the local structure of 3D measurement points, such as Figure 2 Specifically, the expression is:

[0066]

[0067] in, l represents the l-th graph convolution layer; is an intermediate variable; and Represents the three-dimensional measurement point p i and p j The superscript represents the input of the lth convolutional layer of the feature, and the subscript represents the index number of the corresponding 3D measurement point.

[0068] Among them, the graph convolution layer It consists of a 1D convolutional layer, a batch normalization layer, and a ReLU nonlinear activation layer.

[0069] S23. Based on the global maximum pooling operation, obtain the three-dimensional measurement point cloud data The global context information g, that is,

[0070] And, according to the high-dimensional feature expression F i and global context information g to identify any 3D measurement point p i The three-dimensional structure perception module is located in a defect area or a non-defect area. In the three-dimensional structure perception module, the specific process is: splicing the global context information g into the high-dimensional feature expression F i Then, it is fed into a multi-layer perceptron MLP1, such as Figure 3 The multilayer perceptron MLP1 is composed of 3 fully connected layers and 2 ReLU nonlinear activation layers. The output region value y of the multilayer perceptron MLP1 is i The dimension is 2D, representing the three-dimensional measurement point p i The area where it is located. For example: (1, 0) is the defect area and (0, 1) is the non-defect area, such as Figure 5 shown.

[0071] And, according to the high-dimensional feature expression F i and global context information g to obtain any 3D measurement point p i Repair value Δp i =(Δx i , Δy i , Δz i ) of the three-dimensional measurement point repair value regression module. In the three-dimensional measurement point repair value regression module, the specific process is: splicing the global context information g into the high-dimensional feature expression F i Then, it is fed into a multi-layer perceptron MLP2, such as Figure 4 The multilayer perceptron MLP2 is composed of 4 fully connected layers and 3 ReLU nonlinear activation layers. The output of the multilayer perceptron MLP2 is a 3-dimensional repair value Δp i , representing the repair value of the three-dimensional measuring point.

[0072] S3. Construct a hybrid loss function L for the end-to-end network, including the defect area classification loss term L cls, repair value regression loss term L reg And the surface consistency loss term L consist .

[0073] The expression of the hybrid loss function L is:

[0074] L=αL cls +βL reg +γL consist ,

[0075] Wherein, α, β, and γ are hyperparameters used to balance the impact of each loss term. In this embodiment, the values ​​are adjusted based on the convergence effect of the hybrid loss function.

[0076] Accurately identifying and classifying defect regions is crucial for defect detection. However, because defects vary in shape, size, and location, and can be affected by factors like image quality and lighting conditions, traditional classification methods often struggle to achieve optimal results. The defect region classification loss calculates the difference between the predicted defect category and the true category, guiding the network to learn more discriminative features. This improves the accuracy of defect detection and classification, reducing false positives and missed detections.

[0077] The defect area classification loss term L cls The expression is:

[0078]

[0079] Among them, y i The regional value output by MLP1; is the true classification value, which comes from the dataset in step S4.

[0080] In defect repair tasks, it is necessary to accurately predict the repair value of the defective area to achieve high-quality repairs. However, due to the complex characteristics of the defective area and its differences from the characteristics of the surrounding normal areas, directly predicting the repair value is difficult. The repair value regression loss term calculates the error between the predicted repair value and the true repair value, guiding the network to learn the mapping relationship between the defective area and the repair value. This improves the accuracy of the repair value regression, produces repair results that are closer to the true defect-free area, and improves the repair quality.

[0081] The repair value regression loss term L reg The expression is:

[0082]

[0083] Among them, δ is a hyperparameter; Δp i is the repair value; is the restored true value, which comes from the dataset in step S4.

[0084] During defect repair, it's crucial not only to ensure the accuracy of the repaired values ​​but also to ensure that the surface features of the repaired area are consistent with those of the surrounding healthy areas. Significant surface differences between the repaired area and the surrounding area can result in unnatural results, impacting visual quality. The surface consistency loss term calculates the surface differences between the repaired area and the surrounding healthy areas, guiding the network to learn features that maintain surface consistency. This ensures that the repaired area maintains consistent surface features with the surrounding healthy areas, avoiding noticeable repair artifacts and resulting in a more natural-looking repair result.

[0085] The surface consistency loss term L consist The expression is:

[0086]

[0087] Among them, P intact is the set of all 3D measurement points in the non-defective area. This item forces the repair value of the non-defective area to approach 0.

[0088] S4. Establish a data set and optimize the end-to-end network based on the adaptive moment estimation method to obtain an end-to-end calculation model for the surface repair value of composite components.

[0089] The specific process includes the following steps:

[0090] S41. Based on CAD modeling software, first establish 3D models of 50 different composite components and discretize them into 50 3D point cloud data.

[0091] S42. Randomly sample 20 points on each 3D point cloud data, and segment 20 surface areas with the sampling point as the center and r as the radius to obtain 3D measurement data. The measurement data is the true value.

[0092] S43. After obtaining 1000 three-dimensional measurement data of the surface area of ​​the composite component, a random area is randomly selected on each three-dimensional measurement data for random deformation processing. The three-dimensional point cloud of the randomly selected area is the defect area, and the remaining point clouds are non-defect areas. Therefore, the classification truth value is provided. The direction of the deformation is the true value of the repair value

[0093] S44. Using the adaptive moment estimation method (Adam), the parameters of the 3D point cloud feature encoding module, the 3D structure perception module, and the 3D measurement point repair value regression module are optimized until the hybrid loss function L converges. Finally, an end-to-end calculation model for composite component surface repair values, consisting of these three modules, is obtained. For any 3D measurement data of a defective composite component surface, this model can accurately calculate and output the required repair value for each measurement point.

[0094] S5. 3D measurement point cloud data As input, the repair value Δp corresponding to each 3D measurement point is obtained through the end-to-end calculation model of the composite component surface repair value i By taking the 3D point cloud of the composite component surface as input, the repair value of each measuring point is output, where the intact area is 0 and the defect area is the height difference relative to the theoretical surface. This achieves a deep integration of defect perception, surface reconstruction and repair amount calculation to meet the high-precision and high-efficiency engineering requirements.

[0095] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and description merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying surface defects and calculating repair values ​​of composite components based on structural perception, characterized by: The following steps are involved: S1. Obtain M three-dimensional measurement points p on the surface of the composite component. i 3D measurement point cloud data S2. Build an end-to-end network for automatic identification of surface defect areas and quantitative calculation of repair values ​​for composite components; S3. Construct a hybrid loss function L for the end-to-end network, including the defect area classification loss term L cls , repair value regression loss term L reg And the surface consistency loss term L consist ; S4. Establish a data set and optimize the end-to-end network based on the adaptive moment estimation method to obtain an end-to-end calculation model for the surface repair value of composite components; S5. 3D measurement point cloud data As input, the repair value Δp corresponding to each 3D measurement point is obtained through the end-to-end calculation model of the composite component surface repair value i .

2. The method for surface defect identification and repair value calculation of composite components based on structural perception according to claim 1, characterized in that: The end-to-end network includes a method for converting any three-dimensional measurement point p on the surface of the composite component into i Encoded as high-dimensional feature expression F i , and obtain 3D measurement point cloud data 3D point cloud feature encoding module of global context information g; And, according to the high-dimensional feature expression F i and global context information to identify any 3D measurement point p i A three-dimensional structure perception module where the area is a defect area or a non-defect area; And, according to the high-dimensional feature expression F i and global context information g to obtain any 3D measurement point p i Repair value Δp i 3D measurement point repair value regression module.

3. The method for surface defect identification and repair value calculation of composite components based on structural perception according to claim 2, characterized in that: In the three-dimensional point cloud feature encoding module, the specific process includes the following steps: S21, find each three-dimensional measurement point p based on KNN algorithm i k neighboring points N(i); S22. Constructing graph convolutional layer And through three graph convolution layers To encode the three-dimensional measurement point p i The local structure of the feature set is obtained Then the three-dimensional measuring point p i The high-dimensional feature expression is Among them, concat represents the concatenation of feature dimensions, and the expression is: in, 1≤l≤3, l represents the graph convolution layer of the lth layer; is an intermediate variable; and Represents the three-dimensional measurement point p i and p j characteristics; S23. Based on the global maximum pooling operation, obtain the three-dimensional measurement point cloud data The global context information g, that is, 4. The method for surface defect identification and repair value calculation of composite components based on structural perception according to claim 3, characterized in that: The graph convolutional layer It consists of a 1D convolutional layer, a batch normalization layer, and a ReLU nonlinear activation layer.

5. The method for surface defect identification and repair value calculation of composite components based on structural perception according to claim 2, characterized in that: In the three-dimensional structure perception module, the specific process is as follows: Splice the global context information g into the high-dimensional feature representation F i Later, it is fed into a multi-layer perceptron MLP1, and the output is a 2-dimensional area value y i , representing the three-dimensional measurement point p i the region where you are located; The multilayer perceptron MLP1 is composed of 3 fully connected layers and 2 ReLU nonlinear activation layers.

6. The method for surface defect identification and repair value calculation of composite components based on structural perception according to claim 2, characterized in that: In the three-dimensional measurement point repair value regression module, the specific process is as follows: Splice the global context information g into the high-dimensional feature representation F i Later, it is fed into a multi-layer perceptron MLP2, and the output is a 3-dimensional repair value Δp i , representing the three-dimensional measurement point p i Repair value Δp i =(Δx i , Δy i , Δz i ); The multilayer perceptron MLP2 consists of 4 fully connected layers and 3 ReLU nonlinear activation layers.

7. The method for surface defect identification and repair value calculation of composite components based on structural perception according to claim 1, characterized in that: The expression of the hybrid loss function L is: L=αL cls +βL reg +γL consist , Among them, α, β, and γ are hyperparameters used to balance the impact of each loss term; The defect area classification loss term L cls The expression is: Among them, y i The regional value output by MLP1; is the classification truth value; The repair value regression loss term L reg The expression is: Among them, δ is a hyperparameter; Δp i is the repair value; is the repaired truth value; The surface consistency loss term L consist The expression is: Among them, P intact is the set of all three-dimensional measurement points in the non-defective area.

8. The method for surface defect identification and repair value calculation of composite components based on structural perception according to claim 1, characterized in that: In step S4, the specific process includes the following steps: S41. Based on CAD modeling software, first establish 3D models of 50 different composite components and discretize them into 50 3D point cloud data; S42. Randomly sample 20 points on each 3D point cloud data, and segment 20 surface areas with the sampling point as the center and r as the radius to obtain 3D measurement data. The measurement data is the true value. S43. After obtaining 3D measurement data of 1000 composite component surface areas, randomly select an area on each 3D measurement data and perform random deformation processing. The 3D point cloud of the randomly selected area is the defect area, and the remaining point clouds are non-defect areas. S44. Based on the adaptive moment estimation method, the parameters of the 3D point cloud feature encoding module, the 3D structure perception module, and the 3D measurement point repair value regression module are optimized until the hybrid loss function L converges, and finally an end-to-end calculation model of the surface repair value of the composite component is obtained.

Images