CN116778260A - Aerospace rivet flushness detection method, device and system based on AdaBoost integrated learning - Google Patents

Abstract

The invention discloses a method for detecting the flushness of aviation rivets based on AdaBoost integrated learning, which includes: obtaining a point cloud model training set of marked rivet areas , calculate the local features of a single point; based on the AdaBoost integrated learning algorithm, using the training set Generate a strong classifier for segmenting rivet areas; obtain a point cloud model of the rivet-containing area to be tested , for the point cloud model every single point of Perform nearby point query and calculate local features, recorded as a data set ;will gather Input the strong classifier to segment the rivet area and obtain the segmented rivet area. and non-rivet areas ;For rivet area and non-rivet areas Calculate the flushness of the target surface. The present invention uses a strong classifier formed by multiple iterations of weak classifiers to semantically segment point cloud data on the skin surface, thereby achieving high-precision rivet flushness calculation, thereby better controlling aircraft riveting quality.

Description

Aerospace rivet flushness detection method and device based on AdaBoost integrated learning system

Technical field

The present invention relates to the technical field of digital detection of aircraft rivets, and specifically relates to an aviation rivet flushness detection method, device and system based on AdaBoost integrated learning.

Background technique

In aircraft manufacturing, riveting is a key process used to securely join different parts or structural elements together. However, due to the complexity of manufacturing and assembly, rivet flushness issues may exist in aircraft riveting. Rivet flushness refers to the difference in flatness or verticality between the riveting surface and the rivet head. This has an important impact on the aerodynamic shape and performance of the aircraft. At present, there are a large number of rivets on the aircraft skin. The traditional manual method for detection is inefficient and reliable. It cannot achieve quantitative detection and relies too much on experience. Although methods based on image processing can identify rivets, the images lack three-dimensional information and cannot detect flushness.

Three-dimensional laser scanning technology can efficiently obtain three-dimensional information on the aircraft skin surface, with the advantages of high precision and accurate reflection of the true shape. However, due to the small difference between rivet points and non-rivet points on the skin surface, it is difficult for traditional point cloud segmentation algorithms to effectively distinguish rivet areas to calculate flushness. Therefore, this research aims to solve the problem of quantitative detection of rivet flushness on the aircraft skin surface; therefore, the present invention proposes a real-time measurement method of assembly gaps based on digital twins to solve the above problems.

Contents of the invention

In order to solve the above problems, a method, device and system for aerospace rivet flushness detection based on AdaBoost integrated learning are proposed, aiming to solve the small differences between rivet points and non-rivet points on the skin surface caused by the traditional point cloud segmentation algorithm. It is difficult to effectively distinguish the rivet area to calculate flushness; this invention is based on the AdaBoost integrated learning algorithm and uses a strong classifier formed by multiple iterations of weak classifiers to semantically segment the point cloud data on the skin surface, and then Achieve high-precision rivet flush calculation to better control aircraft riveting quality and improve aircraft performance.

In order to achieve the above objectives, the present invention provides the following technical solutions: The invention proposes an aviation rivet flushness detection method based on AdaBoost integrated learning, which includes the following steps:

S1. Obtain the point cloud model training set with marked rivet area , calculate the training set/> Single point local features;

S2, based on the AdaBoost integrated learning algorithm, using the training set The single-point local feature in the medium generates a strong classifier for segmenting rivet regions;

S3. Obtain the point cloud model of the rivet-containing area to be tested. , for point cloud model/> Every single point/> Perform nearby point query and calculate point cloud model/> The single-point local feature of is recorded as a data set/> ;

S4. Collect data Input the strong classifier to segment the rivet area, and obtain the segmented rivet area and non-rivet area;

S5, rivet area and non-rivet areas/> Calculate the flushness of the target surface to obtain the detection results.

Further, in step S1, the point cloud model training set of the marked rivet area is obtained. , calculate the training set/> Single-point local features include the following steps:

S11. For the training set The i-th point/> Use kd-tree for nearby point query;

S12. Calculate target points based on nearby points local features, including Gaussian curvature/> , PFH descriptor/> and density/> ;

S13. Record single points The sample characteristics are/> , record the training set/> The data set is , of which/> is the number of samples in the training set,/> , is the sample feature of point i,/> is the mark set of point i,/> .

Further, in step S2, the training set is used The single-point local feature in the medium generates a strong classifier for segmenting the rivet area, which specifically includes the following steps:

S21. Initialize the sample weight and set the weight of each sample in the training set Initialized to/> ;

S22. Start iterative training of the weak classifier model. is the weak classifier model of the t-th iteration, where/> is a set of sample features,/> , assuming it needs to be iterated T times, then/> ;

S23. Calculate the classification error rate of the weak classifier model , its calculation formula is as follows:

;

in, Is the sample feature/> The true category of /> Is a weak classifier model/> Characteristics of the sample classification results,/> is the current sample weight;

S24. Calculate the weight of the weak classifier model , its calculation formula is as follows:

;

S25. Update the sample weight according to the classification accuracy of the current weak classifier model. , its calculation formula is as follows:

;

S26. Repeat S22~S25 for T rounds of iterations, and get weak classifier model and corresponding weights/> , and then combine them into a strong classifier, whose calculation formula is as follows:

;

in, Indicates that/> is its symbol, that is, when/> +1 when /> time is -1.

Further, the weak classifier model in step S22 includes two convolutional layers and/> , there is a max pooling layer after each convolutional layer/> ,/> , in/> Then configure a fully connected layer FC.

Further, the convolutional layer The convolution kernel size is 3x3, the convolution kernel depth is 6, the stride is 1, and the maximum pooling layer/> The convolution kernel size is 2x2, the stride is 2, and the convolution layer/> The convolution kernel size is 3x3, the convolution kernel depth is 6, the stride is 1, and the maximum pooling layer/> The convolution kernel size is 2x2 and the stride is 2.

Further, in step S3, the point cloud model The single-point local characteristics include calculating the Gaussian curvature of the point/> and PFH descriptor/> and local density/> , recorded as data set/> .

Furthermore, for the rivet area and non-rivet areas/> Calculate the flushness of the target surface to obtain the detection results, which includes the following steps:

S51. Use the region growing algorithm to map the rivet area To split, ///> Split into/> independent rivet area ,/> ;

S52. Use RANSAC for non-rivet areas Perform plane fitting, and the fitted plane is recorded as/> ;

S53. Calculation every point in/> To the plane/> vertical distance/> ,/> ,/> for/> number of midpoints;

S54, calculation Each point in the region is aligned with the plane/> vertical distance/> average/> , that is, get the i-th rivet/> the height of.

The technical solution also provides a device for implementing the aviation rivet flushness detection method based on AdaBoost integrated learning, including:

A training set local feature extraction module. The training set local feature extraction module is used to obtain a point cloud model training set with annotated rivet areas. , calculate the training set/> Single point local features;

Strong classifier generation module, the strong classifier generation module is used based on the AdaBoost integrated learning algorithm, using the training set The single-point local feature in the medium generates a strong classifier for segmenting rivet regions;

The point cloud model local feature extraction module is used to obtain the point cloud model of the rivet-containing area to be measured. , for point cloud model/> Every single point/> Perform nearby point query and calculate point cloud model The single-point local feature of is recorded as a data set/> ;

The rivet area segmentation module is used to input the data set into a strong classifier to segment the rivet area and obtain the segmented rivet area. and non-rivet areas/> ;

Acquire the detection result module, which is used to detect the rivet area and non-rivet areas Calculate the flushness of the target surface to obtain the detection results.

The technical solution also provides a system for implementing the aviation rivet flushness detection method based on AdaBoost integrated learning, including:

processor;

memory;

and one or more programs, wherein the one or more programs are stored in a memory and configured to be executed by the processor, the program being used by the computer to perform the above-mentioned method.

Based on the above technical solution, the present invention provides an aviation rivet flushness detection method, device and system based on AdaBoost integrated learning. It has at least the following beneficial effects:

The flushness of rivets on the aircraft skin has an important impact on the aerodynamic shape and performance of the aircraft. The use of traditional manual methods for inspection has low efficiency and low reliability, and it is impossible to achieve quantitative inspection and relies too much on experience. Methods based on image processing can identify rivets, but the image lacks three-dimensional information and cannot detect flushness; three-dimensional laser scanning technology can efficiently obtain three-dimensional information on the aircraft skin surface, and has the advantages of high precision and accurately reflecting the true shape; however, , due to the small difference between rivet points and non-rivet points on the skin surface, it is difficult for the traditional point cloud segmentation algorithm to effectively distinguish the rivet area to calculate the flushness; therefore, the present invention uses three-dimensional laser scanning technology to perform rivet flushness calculation. Flatness detection, based on the AdaBoost integrated learning algorithm, uses a strong classifier formed by multiple iterations of weak classifiers to semantically segment the point cloud data on the skin surface, which can achieve accurate segmentation of the rivet area and improve detection accuracy and efficiency. Through this method and flushness calculation, high-precision rivet flushness calculation can be achieved, thereby better controlling the quality of aircraft riveting and improving aircraft performance.

Description of drawings

The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. In the attached picture:

Figure 1 is a flow chart of a rivet flushness detection method based on AdaBoost integrated learning of the present invention;

Figure 2 is a structural diagram of the weak classifier network proposed by the present invention;

Figure 3 is a rivet extraction result diagram provided by the implementation of the present invention;

Figure 4 is a diagram showing the calculation results of rivet flushness provided by the implementation of the present invention;

Figure 5 is a schematic block diagram of the device of the rivet flushness detection method based on AdaBoost integrated learning provided by the implementation of the present invention.

In the picture: 100 training set local feature extraction module; 200 top classifier generation module; 300 point cloud model local feature extraction module; 400 rivet area segmentation module; 500 detection result acquisition module.

Detailed ways

In order to make the above objects, features and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. In this way, the implementation process of how this application applies technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

Those of ordinary skill in the art can understand that all or part of the steps in implementing the methods of the above embodiments can be completed by instructing relevant hardware through programs. Therefore, this application can adopt a complete hardware embodiment, a complete software embodiment, or a combination of software and Hardware embodiments. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

Please refer to Figures 1 to 5, which illustrate a specific implementation of this embodiment. In this embodiment, the method is based on the AdaBoost ensemble learning algorithm and uses a strong classifier formed by multiple iterations of weak classifiers to classify the skin surface. The point cloud data is semantically segmented to achieve high-precision rivet flushness calculation, thereby better controlling the quality of aircraft riveting, improving aircraft performance, and solving the problem of traditional point cloud segmentation algorithms between rivet points and non-rivet points on the skin surface. It is difficult to effectively distinguish the rivet area due to the small differences between the rivets, so as to calculate the flushness.

Please refer to Figure 1, a rivet flushness detection method based on AdaBoost integrated learning, including the following steps:

S1. Obtain the point cloud model training set with marked rivet area , calculate the training set/> Single point local features;

Specifically, step S1 includes the following steps:

S11. For the training set The i-th point/> Use kd-tree to query nearby points; query/> The nearest neighbor points are recorded as point sets/> ,of which/> Is the target point/> nearest neighbor point;

S12. Calculate target points based on nearby points local features, including Gaussian curvature/> , PFH descriptor/> and density/> ;

Among them, calculate the target point Gaussian curvature/> , PFH descriptor/> and density/> , specifically including the following steps:

S121. For The jth neighborhood point/> , calculate its relative to the target point/> coordinate offset vector/> :

;

S122. Calculate covariance matrix for neighborhood points :

;

S123. Covariance matrix Perform eigenvalue decomposition to obtain eigenvalues/> and/> , and their corresponding feature vectors/> and/> , the Gaussian curvature can be calculated/> :

;

S124, at the target point Define the coordinate system above:

;

in as target point/> The normal vector of ,/> for/> The three axes of the coordinate system defined above;

;

in for point/> normal vector;

S125, use Constitute the PFH operator/> ;

S126. Offset the coordinate vector calculated in step S121 , take/> , , calculate/> Point cloud density at/> , its calculation formula is expressed as follows:

;

S13. Record single points The sample characteristics are/> , record the training set/> The data set is , of which/> is the number of samples in the training set,/> , is the sample feature of point i,/> is the mark set of point i,/> .

S2, based on the AdaBoost integrated learning algorithm, using the training set The single-point local feature in the medium generates a strong classifier for segmenting rivet regions;

Specifically, step S2 includes the following steps:

S21. Initialize the sample weight and set the weight of each sample in the training set Initialized to/> , of which/> is the number of samples in the training set;

S22. Start iteratively training the weak classifier model and normalize the current sample weight. , using the current sample weight/> Train a weak classifier model,/> is the weak classifier model of the t-th iteration, where/> is a set of sample features,/> , assuming it needs to be iterated T times, then/> ;

Specifically, the weak classifier model includes two convolutional layers and/> , as shown in Figure 2, there is a maximum pooling layer after each convolutional layer/> ,/> , in/> Then configure a full connection layer FC;

Specifically, the convolutional layer The convolution kernel size is 3x3, the convolution kernel depth is 6, the stride is 1, and the maximum pooling layer/> The convolution kernel size is 2x2, the stride is 2, and the convolution layer/> The convolution kernel size is 3x3, the convolution kernel depth is 6, the stride is 1, and the maximum pooling layer/> The convolution kernel size is 2x2 and the stride is 2;

S23. Calculate the classification error rate of the weak classifier model , its calculation formula is as follows:

;

in, Is the sample feature/> The true category of /> Is a weak classifier model/> Characteristics of the sample classification results,/> is the current sample weight;

S24. Calculate the weight of the weak classifier model , its calculation formula is as follows:

;

S25. Update the sample weight according to the classification accuracy of the current weak classifier model. , its calculation formula is as follows:

;

S26. Repeat S22~S25 for T rounds of iterations, and get weak classifier model and corresponding weights/> , and then combine them into a strong classifier, whose calculation formula is as follows:

;

in, Indicates that/> is its symbol, that is, when/> +1 when /> time is -1.

S3. Obtain the point cloud model of the rivet-containing area to be tested. , for point cloud model/> Every single point/> Perform nearby point query and calculate point cloud model/> The single-point local feature of is recorded as a data set/> ;

Specifically, the point cloud model is calculated in step S3 The single-point local characteristics include calculating the Gaussian curvature of the point/> and PFH descriptor/> and local density/> , recorded as data set/> ;

First, use a laser scanner to obtain a point cloud model of the aircraft skin in the rivet area. , secondly, for the point cloud model Each single point uses the kd-tree algorithm for a fixed number/> Query the adjacent points, calculate the Gaussian curvature of the point based on the adjacent points/> , PFH descriptor/> and local density/> , the specific calculation process is the same as S121~S126, and the calculation results are recorded as sets/> .

S4. Collect data Input the trained strong classifier to perform rivet area segmentation, and obtain the classification result of each single point sample, which is the rivet segmentation result, as shown in Figure 3. Figure 3 is a rivet extraction result diagram provided by the implementation of the present invention, and the output Segmented point cloud group/> , point cloud group/> Includes rivet area/> and non-rivet areas/> , finally get the divided rivet area/> and non-rivet areas/> , as shown in Figure 2, the Rivet region in the figure is the rivet area/> , Non-rivet region is the non-rivet region/> .

S5, rivet area and non-rivet areas/> Calculate the flushness of the target surface to obtain the detection results;

Specifically, for the rivet area and non-rivet areas/> Calculating the flushness of the target surface to obtain the detection results includes the following steps:

S51. Use the region growing algorithm to map the rivet area To split, ///> Split into/> independent rivet area ,/> ;

S52. Use RANSAC for non-rivet areas Perform plane fitting, and the fitted plane is recorded as/> , note/> The equation of is:/> ;

S53, for independent rivet area every point in/> Calculate its distance from the plane/> vertical distance/> ,/> ,/> for/> The number of midpoints and their vertical distance/> The calculation formula is as follows:

;

S54, calculation Each point in the region is aligned with the plane/> vertical distance/> average/> , which is the i-th rivet The height, as shown in Figure 4, Figure 4 is a diagram of the calculation results of rivet flushness provided by the implementation of the present invention, where the average value/> The calculation formula is as follows:

;

The technical solution also provides a device for implementing the aerospace rivet flushness detection method based on AdaBoost integrated learning, as shown in Figure 5. Figure 5 shows the rivet flushness detection method based on AdaBoost integrated learning provided by the implementation of the present invention. The device principle block diagram of the homogeneity detection method includes:

Training set local feature extraction module 100. The training set local feature extraction module 100 is used to obtain a point cloud model training set with labeled rivet areas. , calculate the training set/> Single point local features;

Strong classifier generation module 200. The strong classifier generation module 200 is used to use the training set based on the AdaBoost integrated learning algorithm. The single-point local feature in the medium generates a strong classifier for segmenting rivet regions;

Point cloud model local feature extraction module 300. The point cloud model local feature extraction module 300 is used to obtain a point cloud model of the rivet-containing area to be measured. , for point cloud model/> Every single point/> Perform nearby point query and calculate point cloud model/> The single-point local feature of is recorded as a data set/> ;

Rivet area segmentation module 400. The rivet area segmentation module 400 is used to input the data set into a strong classifier to perform rivet area segmentation and obtain the segmented rivet area. and non-rivet areas/> ;

Acquire the detection result module 500, which is used to obtain the detection result module 500 for rivet area and non-rivet areas/> Calculate the flushness of the target surface to obtain the detection results.

The technical solution also provides a system for implementing the aviation rivet flushness detection method based on AdaBoost integrated learning, including:

processor;

memory;

and one or more programs, wherein the one or more programs are stored in a memory and configured to be executed by the processor, the program being used by the computer to perform the above-mentioned method.

This invention is based on the AdaBoost integrated learning algorithm and uses a strong classifier formed by multiple iterations of weak classifiers to semantically segment point cloud data on the skin surface, thereby achieving high-precision rivet flushness calculation, thereby better controlling the aircraft. riveting quality, improving aircraft performance.

Each embodiment in this specification is described in a progressive manner. Each embodiment focuses on its differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. As for the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple. For relevant details, please refer to the description in the method section.

This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method and the core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the present invention There will be changes in the specific implementation methods and application scope of the ideas. In summary, the contents of this description should not be construed as limitations of the present invention.

Claims (9)

1. An aircraft rivet flushness detection method based on AdaBoost integrated learning is characterized by comprising the following steps:

s1, acquiring a point cloud model training set of a marked rivet regionComputing training set +.>A medium single point local feature;

s2, based on AdaBoost integrated learning algorithm, training set is usedThe medium single-point local features generate a strong classifier for segmenting rivet regions;

s3, acquiring a point cloud model of the rivet-containing region to be testedPoint cloud model->Is>Performing a nearby point query and calculating a point cloud model +.>Is recorded as a data set +.>；

S4, gathering the dataInputting the rivet region segmentation into a strong classifier to obtain segmented rivet regions and non-rivet regions;

s5, aligning rivet areasAnd non-rivet area->And carrying out the calculation of the flushness of the target surface to obtain a detection result.

2. The method for detecting the level of the aviation rivet based on AdaBoost ensemble learning according to claim 1, wherein in step S1, a training set of point cloud models of marked rivet areas is obtainedComputing training set +.>The single-point local feature of the medium specifically comprises the following steps:

s11, training setI < th > point->Using kd-tree to inquire the adjacent points;

s12, calculating a target point based on the adjacent pointsComprises a Gaussian curvature>PFH descriptor->And Density->；

S13, single point of recordIs characterized by->Memory training set->Is as follows, wherein />Is the number of samples in the training set, +.>，For the sample feature of point i, +.>For the marker set of point i, +.>。

3. The method for detecting the level of the avionic rivet based on AdaBoost ensemble learning according to claim 1, wherein a training set is used in step S2The strong classifier for dividing the rivet region is generated by the single-point local features, and specifically comprises the following steps:

s21, initializing sample weights, and weighting each sample in the training setInitialized to->；

S22, starting iterative training of a weak classifier model,a weak classifier model for the t-th iteration, wherein +.>Is a set of sample features, +.>Let T iterations be needed ∈ ->；

S23, calculating the classification error rate of the weak classifier modelThe calculation formula is as follows:

；

wherein ,is a sample feature->Is (are) true category->Is a weak classifier model->Sample characterization->Classification result of->Weighting the current sample;

s24, calculating the weight of the weak classifier modelThe calculation formula is as follows:

；

s25, updating the sample weight according to the classification accuracy of the current weak classifier modelThe calculation formula is as follows:

；

s26, repeating S22-S25 for T-round iteration to obtainWeak classifier model and corresponding weights +.>And then combining them into a strong classifier, the calculation formula of which is as follows:

；

wherein ,the representation will->For its sign, i.e. when->Is +1 when->And is-1.

4. An aircraft rivet flushness detection method based on AdaBoost ensemble learning as set forth in claim 3, wherein the weak classifier model in step S22 includes two convolution layers and />Each convolution layer is followed by a maximum pooling layer->、/>In->And a full connection layer FC is arranged.

5. The method for detecting the level of the blind rivet based on AdaBoost ensemble learning according to claim 4, wherein the convolution layer isThe convolution kernel size of (2) is 3x3, the convolution kernel depth is 6, the step size is 1, and the maximum pooling layer is +.>The convolution kernel size of (2 x 2), step size of 2, said convolution layer +.>The convolution kernel size of (2) is 3x3, the convolution kernel depth is 6, the step size is 1, and the maximum pooling layer is +.>The convolution kernel size of (2) is 2x2, the step size is 2.

6. The method for detecting the level of the avionic rivet based on AdaBoost ensemble learning according to claim 1, wherein the method is characterized in that a point cloud model is adopted in the step S3Comprises calculating the point Gaussian curvature +.>And PFH descriptor->And local Density->Record as data set->。

7. An aerospace rivet flushness detection method based on AdaBoost ensemble learning as set forth in claim 1, wherein for rivet regionsAnd non-rivet area->The method comprises the following steps of:

s51, using a region growing algorithm to perform rivet regionDividing ∈>Divided into->Individual rivet areas->，/>；

S52, using RANSAC to the non-rivet areaPerforming plane fitting, wherein the fitted plane is marked as +.>；

S53, calculatingEach point of->For plane->Is>，/>，/>Is->The number of midpoints;

s54, calculatingEach point in the region is +.>Is>Mean value of>Obtaining the ith rivetIs a high level of (2).

8. An apparatus for implementing the AdaBoost ensemble learning-based method for detecting the flatness of an aircraft rivet according to any one of the preceding claims 1 to 7, characterized by comprising:

the training set local feature extraction module (100), wherein the training set local feature extraction module (100) is used for obtaining a point cloud model training set marked with a rivet regionComputing training set +.>A medium single point local feature;

a strong classifier generating module (200), wherein the strong classifier generating module (200) is used for using a training set based on an AdaBoost ensemble learning algorithmThe medium single-point local features generate a strong classifier for segmenting rivet regions;

the device comprises a point cloud model local feature extraction module (300), wherein the point cloud model local feature extraction module (300) is used for obtaining a point cloud model of a rivet-containing region to be detectedPoint cloud model->Is>Performing a nearby point query and calculating a point cloud model +.>Is recorded as a data set +.>；

A rivet region segmentation module (400), wherein the rivet region segmentation module (400) is used for inputting the data set into a strong classifier for rivet region segmentation,obtaining segmented rivet regionsAnd non-rivet area->；

A detection result acquisition module (500), wherein the detection result acquisition module (500) is used for the rivet regionAnd non-rivet area->And carrying out the calculation of the flushness of the target surface to obtain a detection result.

9. A system for implementing an AdaBoost ensemble learning-based aerial rivet flushness detection method according to any one of the preceding claims 1-7, characterized in that it comprises:

a processor;

a memory;

and one or more programs, wherein the one or more programs are stored in memory and configured to be executed by the processor, the program for a computer to perform the method of any of claims 1-7.