A method for detecting cracks in aircraft rivets
By improving the YOLOv5 model, introducing the DMSA module and attention mechanism, and optimizing the dataset, the problem of low accuracy in aircraft rivet crack detection was solved, achieving more efficient detection results.
Patent Information
- Application Number
- CN202410042360.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-11
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-01-11
AI Technical Summary
Existing methods for detecting cracks in aircraft rivets suffer from low accuracy in identification and calculation, making it difficult to adapt to the small, dense, and complex nature of aircraft rivets. This results in slow detection speed, limited range, and high false negative rate.
An improved YOLOv5 model is adopted. By introducing the C3 layer of the DMSA module and combining different structures of the C3 module, the feature interaction performance is enhanced. NAM and SA attention mechanisms are added, and a probe is added to the head structure. The dataset is optimized to improve the model's ability to detect small targets.
It significantly improves the identification accuracy and detection rate of aircraft rivet crack defects, better adapts to the diversity and complex scenarios of aircraft rivets, reduces the risk of overfitting, and enhances the detection capability of small targets.
Smart Images

Figure CN117934397B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image classification, and in particular to a method for detecting cracks and defects in aircraft rivets. Background Technology
[0002] Currently, potential defects in aircraft rivets mainly include cracks, deformation, and misshapenness. These potential defects can lead to serious safety accidents during flight, especially when encountering turbulence, posing a potential threat to flight safety and potentially causing incalculable losses.
[0003] Traditional methods for detecting cracks in aircraft rivets include manual inspection, ultrasonic testing, eddy current testing, pulsed eddy current testing, eyepiece testing, and magneto-optical image detection. However, these methods suffer from drawbacks such as slow detection speed, limited detection range, and high false negative rate. Deep learning-based methods, such as the invention with publication number CN202210146476.0, determine the tightness of the rivets through detection and analysis. However, this algorithm only infers the deviation factor from the gap value, making it difficult to adapt to the small, dense, and complex nature of aircraft rivets, resulting in low recognition and calculation accuracy.
[0004] Therefore, there is an urgent need for a method to detect cracks and defects in aircraft rivets that can adapt to the small, dense, and complex nature of aircraft rivets. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the existing technology in terms of low identification and calculation accuracy and to provide a method for detecting cracks in aircraft rivets.
[0006] The objective of this invention can be achieved through the following technical solutions:
[0007] A method for detecting crack defects in aircraft rivets includes the following steps:
[0008] S1: Collect existing and testable data on aircraft rivet crack defects;
[0009] S2: Build an improved YOLOv5 model;
[0010] S3: By training an improved YOLOv5 model using existing data, a defect detection model for aircraft rivet cracks is obtained.
[0011] S4: Classify the test data using the aircraft rivet crack defect detection model to obtain the aircraft rivet crack defect detection results;
[0012] The improved YOLOv5 model includes a backbone network, a neck structure, and a head structure. The neck structure includes an upsampling module and a downsampling module. The backbone network connects to the upsampling module of the neck structure through branches, and the downsampling module of the neck structure connects to the head structure through branches. Both the backbone network and the neck structure include a C3 layer with an introduced DMSA module to improve the receptive field of neurons in the YOLOv5 model.
[0013] Furthermore, the C3 layer of the DMSA module includes layers C31, C32, and C33. Layer C31 is set in the backbone network, layer C32 is set in the upsampling module of the neck structure, and layer C33 is set in the downsampling module of the neck structure.
[0014] Furthermore, layers C31, C32, and C33 are any of Structure 1, Structure 2, and Structure 3. Structure 1, Structure 2, and Structure 3 all include an input terminal, a residual branch, a bottleneck branch, a splicing function layer, and a DMSA module. The residual branch includes a first convolutional layer connecting the input terminal and the splicing function layer. The bottleneck branch includes a second convolutional layer and a bottleneck layer connected sequentially from the input terminal to the splicing function layer. The splicing function layer splices the residual branch and the bottleneck branch and then connects them to the third convolutional layer. In Structure 1, the DMSA module is connected in series between the input terminal and the first convolutional layer. In Structure 2, the DMSA module is connected in series between the splicing function layer and the third convolutional layer. In Structure 3, the DMSA module is connected in series between the second convolutional layer and the bottleneck layer.
[0015] Furthermore, layers C31, C32, and C33 have the same structure.
[0016] Furthermore, layer C31 is structure one, layer C32 is structure two, and layer C33 is structure three.
[0017] Furthermore, the feature processing of Structure 1 is as follows: after receiving features at the input end, they are input into the residual branch and the bottleneck branch respectively. In the bottleneck branch, the second convolutional layer convolves the features and then processes them through the bottleneck layer to obtain the first bottleneck feature. In the residual branch, the features are sequentially input into the DMSA module and the first convolutional layer to obtain the first residual feature. The first bottleneck feature and the first residual feature are fused through the concatenation function layer and then input into the third convolutional layer to obtain the output result of the C31 layer. The features in the residual branch are enhanced through the DMSA module.
[0018] Furthermore, the data processing procedure for Structure 2 is as follows: After receiving the features at the input end, they are input into the residual branch and the bottleneck branch respectively. In the bottleneck branch, the second convolutional layer convolves the features and then processes them to obtain the second bottleneck features. In the residual branch, the features are convolved by the first convolutional layer to obtain the second residual features. The second bottleneck features and the second residual features are fused by the concatenation function layer and then input into the DMSA module. After that, they are input into the third convolutional layer to obtain the output result of the C32 layer. The DMSA module selectively focuses on the second bottleneck features and the second residual features.
[0019] Furthermore, the feature processing of Structure 3 is as follows: after receiving features at the input end, they are input into the residual branch and the bottleneck branch respectively. In the bottleneck branch, the second convolutional layer convolves the features and then inputs them into the DMSA module. After being processed by the bottleneck layer, the third bottleneck feature is obtained. In the residual branch, the features are convolved by the first convolutional layer to obtain the third residual feature. The third bottleneck feature and the third residual feature are fused by the concatenation function layer and then input into the third convolutional layer to obtain the output result of the C33 layer. The features in the bottleneck branch are enhanced by the DMSA module.
[0020] Furthermore, the backbone network sequentially includes a FOCUS input layer, a first CBS layer, a first C31 layer, a second CBS layer, a second C31 layer, a third CBS layer, a third C31 layer, a fourth CBS layer, a fourth C31 layer, and an SPPF pooling layer. The neck structure sequentially includes a fifth CBS layer connected to the SPPF pooling layer, a first upsampling layer, a first fusion layer, a first C32 layer, a sixth CBS layer, a second upsampling layer, a second fusion layer, a second C32 layer, a NAM attention mechanism layer, a seventh CBS layer, a third upsampling layer, a third fusion layer, a third C32 layer, an SA attention mechanism layer, an eighth CBS layer, and a fourth fusion layer. The structure comprises a first C33 layer, a ninth CBS layer, a fifth fusion layer, a second C33 layer, a tenth CBS layer, a sixth fusion layer, and a third C33 layer. The head structure includes a convolutional layer connecting the third C32 layer and a first probe, a convolutional layer connecting the first C33 layer and a second probe, a convolutional layer connecting the second C33 layer and a third probe, a convolutional layer connecting the third C33 layer and a fourth probe, a first C31 layer also connecting to the third fusion layer, a second C31 layer also connecting to the second fusion layer, a third C31 layer also connecting to the first fusion layer, a fifth CBS layer also connecting to the sixth fusion layer, a sixth CBS layer also connecting to the fifth fusion layer, and a seventh CBS layer also connecting to the fourth fusion layer.
[0021] Layers 1, 2, 3, and 4 are all structure one; layers 1, 2, and 3 are all structure two; layers 1, 2, and 3 are all structure three.
[0022] Furthermore, the first, second, third, and fourth probes each contain anchor frames of various sizes, including 5x6, 8x13, and 15x11 pixels.
[0023] Furthermore, the existing data is optimized through rotation, cropping, and translation, and augmented using a data augmentation algorithm to ensure that the number of images in each category is equal.
[0024] Compared with the prior art, the present invention has the following beneficial effects:
[0025] 1) A YOLOv5 model incorporating DMSA is proposed to enhance the multi-scale feature interaction performance of the model, thereby extracting more comprehensive deep feature information. DMSA is fused with the C3 module to form a C3 layer with DMSA module, which is applied to the backbone network and neck structure. Parallel branches and dilated convolutions are introduced through the DMSA module, which helps to expand the receptive field of each neuron and flexibly switch within the receptive field, making the network more sensitive when processing information in a wider area, thereby better adapting to the diversity of targets such as aircraft rivets and improving recognition accuracy.
[0026] 2) Three fusion structures of DMSA and C3 module are proposed, which can respectively enhance the residual branch performance of C3 model and improve the generalization performance of model; enable model to pay more selective attention to the information of different branches and learn the feature representation of different branches more effectively according to the attention distribution; enhance the bottleneck branch features in the network and increase the model's perception range of input features.
[0027] 3) By introducing NAM and SA attention mechanisms, the system focuses on key feature information in different parts, thereby achieving more comprehensive and faster extraction of effective feature information.
[0028] 4) A new probe dedicated to the detection of small targets has been added to the head structure, which helps to more easily match the bounding box size with the size of cracks and defects, thereby significantly improving the detection accuracy of small targets.
[0029] 5) To preserve and learn important features in the task, the backbone uses the C3_1 module. This module helps the network selectively retain information from different branches and effectively learn the feature representations of different branches. The neck structure upsampling module further processes the feature maps based on the backbone to improve network performance. To enable the model to selectively focus on information from different branches, the neck structure upsampling module uses the C3_2 module. This module introduces an attention distribution, which helps the network learn the feature representations of different branches more effectively and improves the network's performance in the task. For the neck structure downsampling module, judgment and output of results are required. To enable the model to selectively focus on information from different branches, this process uses the C3_3 module, allowing the model to learn the feature representations of different branches more effectively, thus helping to make more accurate judgments and generate the final output results. Attached Figure Description
[0030] Figure 1 This is a flowchart of the present invention.
[0031] Figure 2a This is a C3 layer diagram of the first method of introducing a DMSA module according to the present invention.
[0032] Figure 2b This is a C3 layer diagram of the second type of DMSA module introduced in this invention.
[0033] Figure 2c This is a C3 layer diagram of the third type of DMSA module introduced in this invention.
[0034] Figure 3 This is a structural diagram of the DMSA module of the present invention.
[0035] Figure 4 This is a structural diagram of the YOLOv5 model of the present invention.
[0036] Figure 5 This is a flowchart illustrating the model training process in an embodiment of the present invention.
[0037] Figure 6 The image shows the results of a comparative experiment. Detailed Implementation
[0038] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0039] Example 1
[0040] like Figure 1 As shown, the present invention is a method for detecting crack defects in aircraft rivets, comprising the following steps:
[0041] S1: Collect existing and testable data on aircraft rivet crack defects;
[0042] S2: Build an improved YOLOv5 model;
[0043] S3: By training an improved YOLOv5 model using existing data, a defect detection model for aircraft rivet cracks is obtained.
[0044] S4: Classify the test data using the aircraft rivet crack defect detection model to obtain the aircraft rivet crack defect detection results;
[0045] The improved YOLOv5 model includes a backbone network, a neck structure, and a head structure. The neck structure includes an upsampling module and a downsampling module. The backbone network connects to the upsampling module of the neck structure through branches, and the downsampling module of the neck structure connects to the head structure through branches. Both the backbone network and the neck structure include a C3 layer that incorporates the DMSA module. The role of the C3 layer is to extract features.
[0046] The C3 layer of the DMSA module includes layers C31, C32, and C33. Layer C31 is set in the backbone network, layer C32 is set in the upsampling module of the neck structure, and layer C33 is set in the downsampling module of the neck structure.
[0047] As shown in Figure 2, this invention provides three different C3 layers that incorporate the DMSA module, named Structure 1 C3_1, Structure 2 C3_2, and Structure 3 C3_3, respectively, and are specifically configured as follows:
[0048] C3_1, C3_2, and C3_3 all include an input layer, a residual branch, a bottleneck branch, a concatenation function layer, and a DMSA module. The residual branch includes a first convolutional layer connecting the input layer and the concatenation function layer. The bottleneck branch includes a second convolutional layer and a bottleneck layer connected sequentially from the input layer to the concatenation function layer. The concatenation function layer concatenates the residual branch and the bottleneck branch and then connects them to the third convolutional layer. The difference lies in:
[0049] C3_1: The DMSA module connects the input and the first convolutional layer. This structure adds DMSA to the residual branch, which allows the network to learn features that are important to the task more specifically, thereby improving the model's generalization performance and helping to reduce the risk of overfitting.
[0050] C3_2: The DMSA module connects the concatenation function layer and the third convolutional layer. This structure adds DMSA to the position after the two branches are concatenated. In the case of multiple branches, by introducing DMSA after the concatenation operation, the model can more selectively focus on the information of different branches. The model can learn the feature representation of different branches more effectively according to the attention distribution.
[0051] C3_3: The DMSA module connects the second convolutional layer and the bottleneck layer. This structure adds DMSA to the bottleneck branch, which can help the network focus more on the important features in the input in the bottleneck layer, increase the model's perception range of input features, and enable the model to learn and use specific features more selectively.
[0052] The specific structure of the DMSA module is as follows: Figure 3 As shown, DMSA employs numerous parallel branches to achieve flexible receptive field transformation. In this design, the network structure is divided into two main branches. One half uses 3x3 convolutional kernels, while the other half uses 1x1 convolutional kernels. These two branches undergo multi-directional average pooling, followed by feature fusion and multiplication operations, and finally linearization and normalization. The 3x3 convolutional kernels can extract a wider range and deeper level of features, while the 1x1 convolutional kernels emphasize the perception of smaller-scale features. The combination of these two allows the feature extraction models to compensate for each other's shortcomings in feature extraction, thereby achieving more comprehensive feature information extraction.
[0053] In this embodiment of the invention, the DMSA module is designed for two main purposes: First, to increase the receptive field. By introducing dilated convolution, the DMSA module helps to expand the receptive field of each neuron, making the network more sensitive when processing information over a wider area. Second, for multi-scale extraction, the design of the DMSA module allows the network to consider information at multiple scales simultaneously, thereby better adapting to the diversity of targets such as aircraft rivets and improving the multi-scale representation capability of depth feature information. This improvement enables modules such as C3_1, C3_2, and C3_3 to more effectively capture the depth feature information of aircraft rivets, providing a more powerful feature representation capability for the entire model. This structural adjustment aims to optimize the network architecture, enabling it to better handle targets of various scenarios and scales in the aircraft rivet crack defect detection task.
[0054] In this invention, the structure of each C31 layer, C32 layer, and C33 layer can be selected from any one of C3_1, C3_2, and C3_3, and any combination thereof should fall within the protection scope of this invention.
[0055] In this embodiment, a specific combination is used: layer C31 is structured as C3_1, layer C32 as C3_2, and layer C33 as C3_3. The backbone is responsible for extracting features from the input image to ensure that more feature information is extracted at deeper levels, thereby reducing the loss of feature information. To retain and learn important features in the task, the backbone uses the C3_1 module, which helps the network selectively retain information from different branches and effectively learn the feature representations of different branches. The neck structure upsampling module further processes the feature maps based on the backbone to improve network performance. To enable the model to selectively focus on information from different branches, the neck structure upsampling module uses the C3_2 module. This module introduces an attention distribution, which helps the network learn the feature representations of different branches more effectively and improves the network's performance in the task. For the neck structure downsampling module, judgment and output of results are required. To enable the model to selectively focus on information from different branches, this process uses the C3_3 module. This allows the model to learn the feature representations of different branches more effectively, thereby helping to make more accurate judgments and generate the final output results.
[0056] The specific structure of the algorithm used in this embodiment is as follows: Figure 4 As shown, the network is divided into a backbone network, a neck structure, and a head structure. The backbone network sequentially includes the FOCUS input layer, the first CBS layer, the first C31 layer, the second CBS layer, the second C31 layer, the third CBS layer, the third C31 layer, the fourth CBS layer, the fourth C31 layer, and the SPPF pooling layer. The neck structure sequentially includes the fifth CBS layer connected to the SPPF pooling layer, the first upsampling layer, the first fusion layer, the first C32 layer, the sixth CBS layer, the second upsampling layer, the second fusion layer, the second C32 layer, the NAM attention mechanism layer, the seventh CBS layer, the third upsampling layer, the third fusion layer, the third C32 layer, the SA attention mechanism layer, and the eighth CB layer. The structure consists of an S layer, a fourth fusion layer, a first C33 layer, a ninth CBS layer, a fifth fusion layer, a second C33 layer, a tenth CBS layer, a sixth fusion layer, and a third C33 layer. The head structure includes a convolutional layer connecting the third C32 layer and a first probe, a convolutional layer connecting the first C33 layer and a second probe, a convolutional layer connecting the second C33 layer and a third probe, a convolutional layer connecting the third C33 layer and a fourth probe, a first C31 layer connecting to the third fusion layer, a second C31 layer connecting to the second fusion layer, a third C31 layer connecting to the first fusion layer, a fifth CBS layer connecting to the sixth fusion layer, a sixth CBS layer connecting to the fifth fusion layer, and a seventh CBS layer connecting to the fourth fusion layer.
[0057] The feature processing of the C31 layer using the C3_1 structure is as follows: After receiving the features at the input end, they are input into the residual branch and the bottleneck branch respectively. In the bottleneck branch, the second convolutional layer convolves the features and then processes them through the bottleneck layer to obtain the first bottleneck feature. In the residual branch, the features are first enhanced by the DMSA module and then convolved through the first convolutional layer to obtain the first residual feature. The first bottleneck feature and the first residual feature are fused by the concatenation function layer and then input into the third convolutional layer to obtain the output result of the C31 layer.
[0058] The data processing procedure of the C32 layer using the C3_2 structure is as follows: After receiving features at the input end, they are input into the residual branch and the bottleneck branch respectively. In the bottleneck branch, the second convolutional layer convolves the features and then processes them to obtain the second bottleneck feature. In the residual branch, the features are convolved by the first convolutional layer to obtain the second residual feature. The second bottleneck feature and the second residual feature are fused by the concatenation function layer and then input into the DMSA module. After that, they are input into the third convolutional layer to obtain the output result of the C32 layer. The DMSA module is used to selectively focus on the second bottleneck feature and the second residual feature.
[0059] The feature processing of the C33 layer using the C3_3 structure is as follows: After receiving features at the input end, they are input into the residual branch and the bottleneck branch respectively. In the bottleneck branch, the second convolutional layer convolves the features and then inputs them into the DMSA module. After being processed by the bottleneck layer, the third bottleneck feature is obtained. In the residual branch, the features are convolved by the first convolutional layer to obtain the third residual feature. The third bottleneck feature and the third residual feature are fused by the concatenation function layer and then input into the third convolutional layer to obtain the output result of the C33 layer. The DMSA module is used to enhance the features in the bottleneck branch.
[0060] Meanwhile, the NAM attention mechanism and SA attention mechanism added to the neck structure can focus on the feature information of different parts and enhance the extracted feature information. By enhancing the extracted feature information, the overall model can improve its ability to detect cracks in small targets such as aircraft rivets.
[0061] The YOLOv5 model's head structure incorporates a first probe with three anchor boxes of 5x6, 8x13, and 15x11 pixels. This first probe connects to the last feature-fusion convolutional layer in the neck structure to enhance effective detection of small targets. This change aims to improve the model's sensitivity to small targets, thereby improving detection performance.
[0062] Preferably, the present invention optimizes the dataset by methods such as rotation, cropping, and translation, and expands the dataset composed of existing data by expansion algorithms such as Mixup, CutMix, and Mosaic, so as to balance the amount of data in each category of the dataset, so as to conduct sufficient and accurate training on rivet crack defects.
[0063] like Figure 5As shown, the model training process of an aircraft rivet crack defect detection method based on the improved YOLOv5 model in this embodiment includes the following steps:
[0064] Construct an aircraft crack dataset;
[0065] The dataset is optimized by methods such as rotation, cropping, and translation, and the dataset composed of existing data is expanded by expansion algorithms such as Mixup, CutMix, and Mosaic. The dataset is then input into the improved YOLOv5 model.
[0066] The improved YOLOv5 model includes a C3 layer backbone network and neck structure with the introduction of the DMSA module;
[0067] Introducing NAM attention mechanism and SA attention mechanism;
[0068] A small target detection layer is introduced, namely the head structure of the anchor point;
[0069] The expanded dataset was divided into a training set and a validation set in an 8:2 ratio. The YOLOv5 model was trained and validated to obtain the aircraft rivet crack defect detection model.
[0070] Conduct an evaluation of the results.
[0071] The evaluation results of the aircraft rivet crack defect detection model in this embodiment are as follows: Figure 6 As shown in the figure, the improved model, which incorporates NAM and SA attention mechanisms, demonstrates superior performance in recognizing complex scenes and small-sized targets, outperforming CBAM and ECA attention mechanisms. This confirms the effectiveness of the improved model in enhancing target detection accuracy.
[0072] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A method of rivet crack defect detection for an aircraft, characterized by, Includes the following steps: S1: Collect existing and testable data on aircraft rivet crack defects; S2: Build an improved YOLOv5 model; S3: By training an improved YOLOv5 model using existing data, a defect detection model for aircraft rivet cracks is obtained. S4: Classify the test data using the aircraft rivet crack defect detection model to obtain the aircraft rivet crack defect detection results; The improved YOLOv5 model includes a backbone network, a neck structure, and a head structure. The neck structure includes an upsampling module and a downsampling module. The backbone network connects to the upsampling module of the neck structure through branches, and the downsampling module of the neck structure connects to the head structure through branches. Both the backbone network and the neck structure include a C3 layer with an introduced DMSA module to improve the receptive field of neurons in the YOLOv5 model. The C3 layer of the DMSA module includes layers C31, C32 and C33. Layer C31 is set in the backbone network, layer C32 is set in the upsampling module of the neck structure, and layer C33 is set in the downsampling module of the neck structure. Layers C31, C32, and C33 are any one of Structure 1, Structure 2, and Structure 3. Structure 1, Structure 2, and Structure 3 all include an input terminal, a residual branch, a bottleneck branch, a splicing function layer, and a DMSA module. The residual branch includes a first convolutional layer connecting the input terminal and the splicing function layer. The bottleneck branch includes a second convolutional layer and a bottleneck layer connected sequentially from the input terminal to the splicing function layer. The splicing function layer splices the residual branch and the bottleneck branch and then connects them to a third convolutional layer. In Structure 1, the DMSA module is connected in series between the input terminal and the first convolutional layer. In Structure 2, the DMSA module is connected in series between the splicing function layer and the third convolutional layer. In Structure 3, the DMSA module is connected in series between the second convolutional layer and the bottleneck layer. The backbone network sequentially includes a FOCUS input layer, a first CBS layer, a first C31 layer, a second CBS layer, a second C31 layer, a third CBS layer, a third C31 layer, a fourth CBS layer, a fourth C31 layer, and an SPPF pooling layer. The neck structure sequentially includes a fifth CBS layer connected to the SPPF pooling layer, a first upsampling layer, a first fusion layer, a first C32 layer, a sixth CBS layer, a second upsampling layer, a second fusion layer, a second C32 layer, a NAM attention mechanism layer, a seventh CBS layer, a third upsampling layer, a third fusion layer, a third C32 layer, an SA attention mechanism layer, an eighth CBS layer, a fourth fusion layer, a first C33 layer, and a fourth C31 layer. The structure comprises nine CBS layers, a fifth fusion layer, a second C33 layer, a tenth CBS layer, a sixth fusion layer, and a third C33 layer. The head structure includes a convolutional layer connecting the third C32 layer and a first probe, a convolutional layer connecting the first C33 layer and a second probe, a convolutional layer connecting the second C33 layer and a third probe, and a convolutional layer connecting the third C33 layer and a fourth probe. The first C31 layer is also connected to the third fusion layer, the second C31 layer is also connected to the second fusion layer, the third C31 layer is also connected to the first fusion layer, the fifth CBS layer is also connected to the sixth fusion layer, the sixth CBS layer is also connected to the fifth fusion layer, and the seventh CBS layer is also connected to the fourth fusion layer. The first C31 layer, the second C31 layer, the third C31 layer, and the fourth C31 layer are all structure one; the first C32 layer, the second C32 layer, and the third C32 layer are all structure two; and the first C33 layer, the second C33 layer, and the third C33 layer are all structure three.
2. A method of rivet crack defect detection for an aircraft as defined in claim 1, wherein Layers C31, C32, and C33 have the same structure.
3. The method of rivet crack detection of an aircraft as defined in claim 1, wherein, The C31 layer is structure one, the C32 layer is structure two, and the C33 layer is structure three.
4. The method of rivet crack detection of an aircraft as defined in claim 1, wherein, The feature processing of Structure 1 is as follows: After receiving the features, the input end inputs them into the residual branch and the bottleneck branch respectively. In the bottleneck branch, the second convolutional layer convolves the features and then processes them through the bottleneck layer to obtain the first bottleneck feature. In the residual branch, the features are sequentially input into the DMSA module and the first convolutional layer to obtain the first residual feature. The first bottleneck feature and the first residual feature are fused through the concatenation function layer and then input into the third convolutional layer to obtain the output result of the C31 layer. The features in the residual branch are enhanced through the DMSA module.
5. The method of rivet crack detection of an aircraft as defined in claim 1, wherein, The data processing procedure of Structure 2 is as follows: After receiving the features, the input end inputs the residual branch and the bottleneck branch respectively. In the bottleneck branch, the second convolutional layer convolves the features and then processes them to obtain the second bottleneck feature. In the residual branch, the features are convolved by the first convolutional layer to obtain the second residual feature. The second bottleneck feature and the second residual feature are fused by the concatenation function layer and then input into the DMSA module. After that, they are input into the third convolutional layer to obtain the output result of the C32 layer. The DMSA module selectively focuses on the second bottleneck feature and the second residual feature.
6. The method for detecting crack defects in aircraft rivets according to claim 1, characterized in that, The feature processing of Structure 3 is as follows: After receiving the features, the input end inputs them into the residual branch and the bottleneck branch respectively. In the bottleneck branch, the second convolutional layer convolves the features and then inputs them into the DMSA module. After processing by the bottleneck layer, the third bottleneck feature is obtained. In the residual branch, the features are convolved by the first convolutional layer to obtain the third residual feature. The third bottleneck feature and the third residual feature are fused by the concatenation function layer and then input into the third convolutional layer to obtain the output result of the C33 layer. The features in the bottleneck branch are enhanced by the DMSA module.
7. The method for detecting crack defects in aircraft rivets according to claim 1, characterized in that, The first, second, third, and fourth probes each contain anchor frames of various sizes, including 5x6, 8x13, and 15x11 pixels.
Citation Information
Patent Citations
Rivet tightness detection device and use method thereof
CN114593672A
Aircraft detection method in aerial image based on OfficientDet and Transformer
CN113283409A
X-ray weld defect detection method based on convolutional neural network
CN113674247A
