Aircraft tire tread defect automatic identification method

Abstract

The invention discloses an aircraft tire tread defect automatic identification method. The method comprises the steps of obtaining an original image data set; dividing a training set, a test set and a verification set; an aircraft tire tread defect detection model is constructed and trained; outputting and displaying an aircraft tire tread defect detection result and the like. According to the method, through a mode of fusing machine vision and deep learning, defect detection can be rapidly and accurately carried out on the aircraft tire tread, and an intelligent and efficient solution is provided for the field of aviation maintenance; a C2fMAEFS module is constructed in the backbone network, global context information can be reserved, the recognition capability of the model on irregular boundary defects is enhanced, and positioning deviation caused by boundary fuzziness is reduced; an FPSCCPCA module is designed in the backbone network, and the feature extraction capacity of the model for defect areas with complex textures and low contrast is enhanced; the neck network adopts a multi-branch auxiliary feature pyramid MAFPN structure, through cross-layer bidirectional feature transfer, the fusion capability of multi-size features can be enhanced, and the omission ratio of small targets is reduced.

Description

Technical Field

[0001] This invention belongs to the field of aviation inspection equipment technology, and specifically relates to an automatic identification method for aircraft tire tread defects. Background Technology

[0002] Aircraft tires, as a critical component of the aircraft landing gear system, are the only part of the aircraft in contact with the ground, bearing the entire load during takeoff, landing, and taxiing. Due to the complex operating environment, aircraft tires are highly susceptible to defects. Failure to detect and address these defects in a timely manner can seriously compromise aircraft safety. Therefore, regular inspection of aircraft tires is necessary to prevent accidents caused by tire damage. Furthermore, aircraft tires are consumable components with high replacement and maintenance costs. Failure to detect and address tire defects promptly can accelerate tire wear, increasing repair and replacement costs. Currently, airline maintenance personnel primarily rely on manual visual inspection for the inspection of various aircraft components. This method is highly dependent on manpower, has low inspection efficiency, and its results are highly dependent on the experience and expertise of the inspectors. Different personnel may have different judgments about the same defect, easily leading to false positives or false negatives. Especially under complex conditions such as insufficient lighting or concealed locations, manual inspection cannot cover all defect areas, posing safety hazards. Therefore, a standardized defect detection method is needed to meet the needs of rapid pre-flight and post-flight inspections and improve the accuracy of defect identification.

[0003] In recent years, with the development of machine vision and deep learning technologies, object detection models have been widely applied in industrial inspection and safety assurance, and have found good results in the inspection of components such as aircraft skin, fuse orientation, and engines, providing a feasible technical path for aircraft tire defect detection. The YOLO series of algorithms, with their end-to-end structure and efficient inference capabilities, are widely used in object detection tasks with high real-time requirements. Among them, the YOLOv8 model performs well in balancing detection accuracy and speed. However, when directly applying the YOLOv8 model to aircraft tire defect detection, the challenge of balancing performance and efficiency remains, necessitating optimization of the existing algorithm structure to achieve more efficient and accurate identification. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide an automatic identification method for aircraft tire tread defects.

[0005] To achieve the above objectives, the automatic identification method for aircraft tire tread defects provided by the present invention includes the following steps performed in sequence:

[0006] Step 1: Collect images of the defective aircraft tire tread to obtain the raw image dataset;

[0007] Step 2: Perform data augmentation on the original aircraft tire tread images in the above original image dataset, and then perform manual annotation to obtain the aircraft tire tread defect information dataset and divide it into training set, test set and validation set according to the proportion.

[0008] Step 3: Based on the YOLOv8 model, construct an aircraft tire tread defect detection model;

[0009] Step 4: Use the training set, test set and validation set obtained in Step 2 to train, test and validate the above aircraft tire tread defect detection model, and finally obtain the trained aircraft tire tread defect detection model.

[0010] Step 5: Deploy the trained aircraft tire tread defect detection model on the server. Deploy the embedded acquisition device, consisting of a Raspberry Pi, an industrial camera, and a display screen, on the unmanned inspection equipment. Use the industrial camera to take pictures of the aircraft tire tread at the aircraft maintenance site, and upload the captured images to the server under the control of the Raspberry Pi. The server receives the images to be inspected and automatically loads them into the trained aircraft tire tread defect detection model, outputs the aircraft tire tread defect detection results, and displays the results on the display screen.

[0011] In step 1, the method for acquiring images of defective aircraft tire treads to obtain the original image dataset is as follows:

[0012] Multiple original aircraft tire tread images, including three types of defects—cuts, cracks, and wear lines—were collected from multiple angles and under different lighting conditions at aircraft maintenance sites and on training aircraft in aviation colleges. The original image dataset is composed of all the original aircraft tire tread images.

[0013] In step 2, the method for performing data augmentation on the original aircraft tire tread images in the aforementioned original image dataset, followed by manual annotation to obtain an aircraft tire tread defect information dataset, and then dividing it into training, testing, and validation sets proportionally, is as follows:

[0014] Complete and relatively clear original aircraft tire tread images are selected from the original image dataset obtained in step 1. Then, they are processed using methods including translation, rotation, and mirroring. At the same time, the Albumentations data augmentation library is used to perform blurring, brightness adjustment, and the addition of light spots, rain, snow, and shadows. This process augments the original aircraft tire tread images to obtain enhanced images. Then, the labelimg annotation tool is used to annotate the original aircraft tire tread images and enhanced images according to the defect type in three categories. All annotated images constitute the aircraft tire tread defect information dataset and are divided into training set, test set, and validation set in a ratio of 8:1:1.

[0015] In step 3, the method for constructing an aircraft tire tread defect detection model based on the YOLOv8 model is as follows:

[0016] The aircraft tire tread defect detection model includes a backbone network, a neck network, and a detection head. In the backbone network, the Bottleneck module in the C2f module of the YOLOv8 model backbone network is replaced with the MAEFS module to construct the C2f_MAEFS module; the SPPF module is replaced with the FPSC_CPCA module; in the neck network, the path aggregation feature pyramid network PAFPN structure of the YOLOv8 model is improved into a multi-branch auxiliary feature pyramid MAFPN structure.

[0017] In the backbone network, the labeled image obtained in step 2 is used as the input image. First, a convolution operation is performed using the first convolutional layer with a stride of 2, thereby reducing the size of the input image from 640×640 to 320×320, while increasing the number of channels, to obtain the initial feature map P1.

[0018] Then, the initial feature map P1 is downsampled through the second convolutional layer with a stride of 2, and processed by the first C2f_MAEFS module to reduce the size of the initial feature map P1 from 320×320 to 160×160, resulting in feature map P2.

[0019] The feature map P2 is then downsampled using the third convolutional layer and processed by the second C2f_MAEFS module to reduce the size of the feature map P2 from 160×160 to 80×80, resulting in the feature map P3.

[0020] Then, feature map P3 is downsampled using the fourth convolutional layer and processed by the third C2f_MAEFS module to reduce the size of feature map P3 from 80×80 to 40×40, resulting in feature map P4.

[0021] The feature map P4 is then downsampled using the fifth convolutional layer and processed by the fourth C2f_MAEFS module to reduce the size of the feature map P4 from 40×40 to 20×20, resulting in the feature map P5.

[0022] Finally, feature map P5 is input into the FPSC_CPCA module to enhance the receptive field without changing the feature map size, and outputs a high-level semantic feature map P6 with a size of 20×20.

[0023] At this point, the backbone network outputs multiple layers of feature maps with different sizes for subsequent feature fusion and detection.

[0024] The main function of the neck network is to perform bidirectional fusion of feature maps of different sizes output by the backbone network from top to bottom and bottom to top.

[0025] In the top-down feature fusion path, the 40×40 scale P4 feature map output by the backbone network is first input into the first convolutional layer of the neck. It is downsampled by a convolution operation with a stride of 2, reducing its size from 40×40 to 20×20. It is then concatenated with the high-level semantic feature map P6 of the same size in the channel dimension. The concatenated feature map is processed by the C2f module to obtain the intermediate fusion feature map of the 20×20 scale.

[0026] The 20×20 scale intermediate fused feature map is upsampled to increase its size from 20×20 to 40×40. The 80×80 scale feature map P3 output from the backbone network is input into the second convolutional layer at the neck. It is downsampled by a convolution operation with a stride of 2 to reduce its size from 80×80 to 40×40. The two 40×40 scale feature maps are then concatenated along the channel dimension. The concatenated feature map is processed by the C2f module to obtain the 40×40 scale intermediate fused feature map.

[0027] Then, the 40×40 scale intermediate fused feature map is further upsampled to increase its size from 40×40 to 80×80. The 160×160 scale feature map P2 output by the backbone network is input into the third convolutional layer of the neck. It is downsampled by a convolution operation with a stride of 2 to reduce its size from 160×160 to 80×80. The two 80×80 scale feature maps are then concatenated along the channel dimension. The concatenated feature map is processed by the C2f module to obtain the 80×80 scale intermediate fused feature map.

[0028] In the bottom-up feature fusion path, the 80×80 scale intermediate fusion feature map is concatenated with the 80×80 scale feature map obtained by upsampling the 40×40 scale intermediate fusion feature map. The concatenated feature map is processed by the C2f module to obtain an 80×80 output fusion feature map, which is used as the first layer output of the neck network.

[0029] Subsequently, the intermediate fusion feature map of scale 80×80 and the output fusion feature map of scale 80×80 are input into the fourth and fifth convolutional layers of the neck for downsampling, so that their size is reduced from 80×80 to 40×40. Then, the two 40×40 feature maps obtained above are concatenated with the 40×40 intermediate fusion feature map. The concatenated feature map is processed by the C2f module to obtain the 40×40 output fusion feature map, which is used as the output of the second layer of the neck network.

[0030] Furthermore, the 40×40 scale intermediate fusion feature map and the 40×40 scale output fusion feature map are input into the sixth and seventh convolutional layers of the neck for downsampling. Then, the two 20×20 scale feature maps obtained above are concatenated with the 20×20 scale intermediate fusion feature map. The concatenated feature map is processed by the C2f module to obtain the 20×20 scale output fusion feature map, which is used as the third layer output of the neck network.

[0031] Finally, the three detection heads receive the three-layer output fusion feature map from the neck network for target detection prediction; among them, the 80×80 output fusion feature map is used for small-sized target detection, the 40×40 output fusion feature map is used for medium-sized target detection, and the 20×20 output fusion feature map is used for large-sized target detection.

[0032] The MAEFS module includes one main branch and four parallel auxiliary branches, and its processing method is as follows:

[0033] The feature map input to the MAEFS module first enters a 3×3 convolutional layer in the main branch to obtain the main branch output feature map. Simultaneously, the above input feature map is fed into four parallel auxiliary branches. These four auxiliary branches consist of the following structural units: the first processing unit uses 3×3, 6×6, 9×9, and 12×12 adaptive average pooling layers, respectively; the second processing unit is a 1×1 convolutional layer; the third processing unit is a 3×3 depthwise convolutional layer; the fourth processing unit is an upsampling unit; and the fifth processing unit is an EB edge enhancement module. Each auxiliary branch outputs its own feature map.

[0034] Subsequently, the output feature maps of the four auxiliary branches are concatenated with the output feature map of the main branch to obtain a multi-scale fused feature map; then, the fused feature map is input into the DSM dual-domain selection mechanism module; finally, the feature map output by the DSM dual-domain selection mechanism module is input into a 1×1 convolutional layer to obtain the feature map output by the MAEFS module.

[0035] The feature map input to the EB edge enhancement module is first subjected to a 3×3 average pooling layer for average pooling. Then, the average pooled feature map is subtracted from the input feature map to generate a difference feature map. This difference feature map is then input into a 3×3 convolutional layer with a sigmoid activation function to output an enhanced feature map. Finally, the enhanced feature map is added to the input feature map to output the edge-enhanced result feature map.

[0036] The DSM dual-domain selection mechanism module includes a spatial domain selection module (SSM) and a frequency domain selection module (FSM), and its processing method is as follows:

[0037] First, in the Spatial Domain Selection Module (SSM), the input feature map F is subjected to max pooling and average pooling operations respectively using max pooling and average pooling layers. Then, a spatial attention map F′ is generated through a 3×3 convolutional layer. Subsequently, the input feature map F is subjected to two sets of depthwise convolutional layers: one set consists of a 5×5 kernel with a dilation of 2 and a 7×7 kernel with a dilation of 3; the other set consists of a 3×3 depthwise convolutional layer. Finally, the two sets of depthwise convolutional results are multiplied and added element-wise with the spatial attention map F′ respectively to obtain the spatially enhanced feature map F. S In FSM, the spatially enhanced feature map F is first obtained through a mean filter. S Extract low-frequency feature map Mean(F) S And obtain the corresponding high-frequency feature map through subtraction operation. Then the high-frequency feature map Compared with the above spatial enhancement feature map F S Element-wise multiplication is performed, and the results are fused through residual connections to finally output the enhanced feature map.

[0038] The processing method of the FPSC_CPCA module is as follows:

[0039] First, the input feature map is convolved through a 1×1 convolutional layer, and then sequentially fed into three 3×3 dilated convolutional layers with dilation rates of 1, 3, and 5. Next, the input feature map is concatenated with the feature maps output from all the dilated convolutions to obtain a fused feature map. This fused feature map is then fed into the CPCA attention mechanism module, and finally, it is convolved through a 1×1 convolutional layer to obtain the feature map output by the FPSC_CPCA module.

[0040] The CPCA attention mechanism module includes channel attention and spatial attention branches;

[0041] First, the input feature map F is processed by average pooling and max pooling layers respectively before being fed into a shared multilayer perceptron (MLP). Then, the two outputs of the MLP are summed and activated by a sigmoid activation function to obtain the spatial attention feature map F. CA Then, the spatial attention feature map F CA The channel prior feature map F is obtained by element-wise multiplication with the input feature map F. C Subsequently, the channel prior feature map F CThe input is fed into a 5×5 depthwise convolutional module, and the convolutional output is fed into the following three sets of depthwise convolutional branches: 1×7 depthwise convolutional layer + 7×1 depthwise convolutional layer; 1×11 depthwise convolutional layer + 11×1 depthwise convolutional layer; 1×21 depthwise convolutional layer + 21×1 depthwise convolutional layer.

[0042] The feature maps output from the three sets of deep convolution branches are added to the feature map output from the 5×5 deep convolution module to obtain the spatial attention feature map F. SA Then, a 1×1 convolutional layer is used to perform channel fusion to obtain a fused feature map; finally, the fused feature map is combined with the channel prior feature map F. C Multiplication yields the final output feature map F. OUT .

[0043] In step 4, the method for training, testing, and validating the aircraft tire tread defect detection model using the training set, test set, and validation set obtained in step 2, and finally obtaining the trained aircraft tire tread defect detection model, is as follows:

[0044] The hyperparameters were configured as follows: batch size of 16, training epochs of 300, learning rate of 0.02, input image size of 640×640, and SGD optimizer. First, the aircraft tire tread defect detection model was trained using the training set. Mosaic was enabled for data augmentation during training, and disabled in the last 10 epochs. Then, the training results were tested using the test set. Training was stopped when accuracy (P), recall (R), and mean precision (mAP) stopped increasing, and the loss function gradually converged. The final training weights were retained as the weights for the aircraft tire tread defect detection model, resulting in a trained model. Finally, the aircraft tire tread defect detection model was evaluated using the validation set to determine its performance in real-world scenarios.

[0045] In step 5, the detection results include the defect category, bounding box, number of defects, confidence level, and inference time for each image to be detected for the aircraft tire tread.

[0046] The automatic identification method for aircraft tire tread defects provided by this invention has the following beneficial effects:

[0047] 1. This invention, through the fusion of machine vision and deep learning, can quickly and accurately detect defects in aircraft tire treads, assisting maintenance personnel in rapidly determining the damage status of aircraft tires during line maintenance, and providing an intelligent and efficient solution for the aviation maintenance field;

[0048] 2. A C2f_MAEFS module is constructed in the backbone network. Multi-size adaptive pooling and convolution operations are used to extract multi-size information. Combined with edge information enhancement and dual-domain selection mechanism, global context information can be preserved, the model's ability to identify irregular boundary defects can be strengthened, and the localization deviation caused by boundary ambiguity can be reduced.

[0049] 3. An FPSC_CPCA module was designed in the backbone network, which combines shared convolution with different dilation rates and channel prior attention mechanism to enhance the model's ability to extract features from defective regions with complex textures and low contrast.

[0050] 4. The neck network adopts a multi-branch auxiliary feature pyramid (MAFPN) structure. Through bidirectional feature transfer across layers, it can enhance the fusion capability of multi-size features and reduce the false negative rate of small targets. Attached Figure Description

[0051] Figure 1 The flowchart of the automatic identification method for aircraft tire tread defects provided by the present invention is shown.

[0052] Figure 2 This is a diagram of the original YOLOv8 model structure.

[0053] Figure 3 This is a structural diagram of the aircraft tire tread defect detection model constructed in this invention.

[0054] Figure 4 The structural diagram of the EB module, MAEFS module, and DSM module in the C2f_MAEFS module provided by this invention.

[0055] Figure 5 This is a structural diagram of the original SPPF module, the FPSC_CPCA module, and the CPCA module in this invention. Detailed Implementation

[0056] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0057] like Figure 1 As shown, the automatic identification method for aircraft tire tread defects provided by the present invention includes the following steps performed in sequence:

[0058] Step 1: Collect images of the defective aircraft tire tread to obtain the raw image dataset;

[0059] At the aircraft maintenance site and on the training aircraft of aviation colleges, multiple original aircraft tire tread images, including three types of defects such as cuts, cracks and wear lines, were collected by cameras from multiple angles and under different lighting conditions. The original image dataset is composed of all the original aircraft tire tread images.

[0060] Step 2: Perform data augmentation on the original aircraft tire tread images in the above original image dataset, and then perform manual annotation to obtain the aircraft tire tread defect information dataset and divide it into training set, test set and validation set according to the proportion.

[0061] Complete and relatively clear original aircraft tire tread images are selected from the original image dataset obtained in step 1. Then, they are processed using methods including translation, rotation, and mirroring. At the same time, the Albumentations data augmentation library is used to perform blurring, brightness adjustment, and the addition of light spots, rain, snow, and shadows. This process augments the original aircraft tire tread images to obtain enhanced images. Then, the labelimg annotation tool is used to annotate the original aircraft tire tread images and enhanced images according to the defect type in three categories. All annotated images constitute the aircraft tire tread defect information dataset and are divided into training set, test set, and validation set in a ratio of 8:1:1.

[0062] Step 3: Based on the YOLOv8 model, construct an aircraft tire tread defect detection model;

[0063] like Figure 2 As shown, the YOLOv8 model includes a backbone network, a neck network, and a detection head; the backbone network is used for feature extraction, the neck network is used for multi-size feature fusion, and the detection head is used for predicting the target category and location.

[0064] like Figure 3 As shown, the aircraft tire tread defect detection model includes a backbone network, a neck network, and a detection head. In the backbone network, the Bottleneck module in the C2f module of the YOLOv8 model backbone network is replaced with the MAEFS module to construct the C2f_MAEFS module; the SPPF module is replaced with the FPSC_CPCA module; in the neck network, the path aggregation feature pyramid network PAPPN structure of the YOLOv8 model is improved into a multi-branch auxiliary feature pyramid MAFPN structure.

[0065] In the backbone network, the labeled image obtained in step 2 is used as the input image. First, a convolution operation is performed using the first convolutional layer with a stride of 2, thereby reducing the size of the input image from 640×640 to 320×320, while increasing the number of channels, to obtain the initial feature map P1.

[0066] Then, the initial feature map P1 is downsampled through the second convolutional layer with a stride of 2, and processed by the first C2f_MAEFS module to reduce the size of the initial feature map P1 from 320×320 to 160×160, resulting in feature map P2.

[0067] The feature map P2 is then downsampled using the third convolutional layer and processed by the second C2f_MAEFS module to reduce the size of the feature map P2 from 160×160 to 80×80, resulting in the feature map P3.

[0068] Then, feature map P3 is downsampled using the fourth convolutional layer and processed by the third C2f_MAEFS module to reduce the size of feature map P3 from 80×80 to 40×40, resulting in feature map P4.

[0069] The feature map P4 is then downsampled using the fifth convolutional layer and processed by the fourth C2f_MAEFS module to reduce the size of the feature map P4 from 40×40 to 20×20, resulting in the feature map P5.

[0070] Finally, feature map P5 is input into the FPSC_CPCA module to enhance the receptive field without changing the feature map size, and outputs a high-level semantic feature map P6 with a size of 20×20.

[0071] At this point, the backbone network outputs multiple layers of feature maps with different sizes for subsequent feature fusion and detection.

[0072] The main function of the neck network is to perform bidirectional fusion of feature maps of different sizes output by the backbone network from top to bottom and bottom to top.

[0073] In the top-down feature fusion path, the 40×40 scale P4 feature map output by the backbone network is first input into the first convolutional layer of the neck. It is downsampled by a convolution operation with a stride of 2, reducing its size from 40×40 to 20×20. It is then concatenated with the high-level semantic feature map P6 of the same size in the channel dimension. The concatenated feature map is processed by the C2f module to obtain the intermediate fusion feature map of the 20×20 scale.

[0074] The 20×20 scale intermediate fused feature map is upsampled to increase its size from 20×20 to 40×40. The 80×80 scale feature map P3 output from the backbone network is input into the second convolutional layer at the neck. It is downsampled by a convolution operation with a stride of 2 to reduce its size from 80×80 to 40×40. The two 40×40 scale feature maps are then concatenated along the channel dimension. The concatenated feature map is processed by the C2f module to obtain the 40×40 scale intermediate fused feature map.

[0075] Then, the 40×40 scale intermediate fused feature map is further upsampled to increase its size from 40×40 to 80×80. The 160×160 scale feature map P2 output by the backbone network is input into the third convolutional layer of the neck. It is downsampled by a convolution operation with a stride of 2 to reduce its size from 160×160 to 80×80. The two 80×80 scale feature maps are then concatenated along the channel dimension. The concatenated feature map is processed by the C2f module to obtain the 80×80 scale intermediate fused feature map.

[0076] In the bottom-up feature fusion path, the 80×80 scale intermediate fusion feature map is concatenated with the 80×80 scale feature map obtained by upsampling the 40×40 scale intermediate fusion feature map. The concatenated feature map is processed by the C2f module to obtain an 80×80 output fusion feature map, which is used as the first layer output of the neck network.

[0077] Subsequently, the intermediate fusion feature map of scale 80×80 and the output fusion feature map of scale 80×80 are input into the fourth and fifth convolutional layers of the neck for downsampling, so that their size is reduced from 80×80 to 40×40. Then, the two 40×40 feature maps obtained above are concatenated with the 40×40 intermediate fusion feature map. The concatenated feature map is processed by the C2f module to obtain the 40×40 output fusion feature map, which is used as the output of the second layer of the neck network.

[0078] Furthermore, the 40×40 scale intermediate fusion feature map and the 40×40 scale output fusion feature map are input into the sixth and seventh convolutional layers of the neck for downsampling. Then, the two 20×20 scale feature maps obtained above are concatenated with the 20×20 scale intermediate fusion feature map. The concatenated feature map is processed by the C2f module to obtain the 20×20 scale output fusion feature map, which is used as the third layer output of the neck network.

[0079] Finally, the three detection heads receive the three-layer output fusion feature map from the neck network for target detection prediction; among them, the 80×80 output fusion feature map is used for small-sized target detection, the 40×40 output fusion feature map is used for medium-sized target detection, and the 20×20 output fusion feature map is used for large-sized target detection.

[0080] like Figure 4 As shown, the MAEFS module includes one main branch and four parallel auxiliary branches, and its processing method is as follows:

[0081] The feature map input to the MAEFS module first enters a 3×3 convolutional layer in the main branch to obtain the main branch output feature map. Simultaneously, the above input feature map is fed into four parallel auxiliary branches. These four auxiliary branches consist of the following structural units: the first processing unit uses 3×3, 6×6, 9×9, and 12×12 adaptive average pooling layers, respectively; the second processing unit is a 1×1 convolutional layer; the third processing unit is a 3×3 depthwise convolutional layer; the fourth processing unit is an upsampling unit; and the fifth processing unit is an EB edge enhancement module. Each auxiliary branch outputs its own feature map.

[0082] Subsequently, the output feature maps of the four auxiliary branches are concatenated with the output feature map of the main branch to obtain a multi-scale fused feature map; then, the fused feature map is input into the DSM dual-domain selection mechanism module; finally, the feature map output by the DSM dual-domain selection mechanism module is input into a 1×1 convolutional layer to obtain the feature map output by the MAEFS module.

[0083] The feature map input to the EB edge enhancement module is first subjected to a 3×3 average pooling layer for average pooling. Then, the average pooled feature map is subtracted from the input feature map to generate a difference feature map. This difference feature map is then input into a 3×3 convolutional layer with a sigmoid activation function to output an enhanced feature map. Finally, the enhanced feature map is added to the input feature map to output the edge-enhanced result feature map.

[0084] The DSM dual-domain selection mechanism module includes a spatial domain selection module (SSM) and a frequency domain selection module (FSM), and its processing method is as follows:

[0085] First, in the Spatial Domain Selection Module (SSM), the input feature map F is subjected to max pooling and average pooling operations respectively using max pooling and average pooling layers. Then, a spatial attention map F′ is generated through a 3×3 convolutional layer. Subsequently, the input feature map F is subjected to two sets of depthwise convolutional layers: one set consists of a 5×5 kernel with a dilation of 2 and a 7×7 kernel with a dilation of 3; the other set consists of a 3×3 depthwise convolutional layer. Finally, the two sets of depthwise convolutional results are multiplied and added element-wise with the spatial attention map F′ respectively to obtain the spatially enhanced feature map F. S In FSM, the spatially enhanced feature map F is first obtained through a mean filter. S Extract low-frequency feature map Mean(F) S And obtain the corresponding high-frequency feature map through subtraction operation. Then the high-frequency feature map Compared with the above spatial enhancement feature map F S Element-wise multiplication is performed, and the results are fused through residual connections to finally output the enhanced feature map.

[0086] like Figure 5 As shown, the processing method of the FPSC_CPCA module is as follows:

[0087] First, the input feature map is convolved through a 1×1 convolutional layer, and then sequentially fed into three 3×3 dilated convolutional layers with dilation rates of 1, 3, and 5. Next, the input feature map is concatenated with the feature maps output from all the dilated convolutions to obtain a fused feature map. This fused feature map is then fed into the CPCA attention mechanism module, and finally, it is convolved through a 1×1 convolutional layer to obtain the feature map output by the FPSC_CPCA module.

[0088] The CPCA attention mechanism module includes channel attention and spatial attention branches;

[0089] First, the input feature map F is processed by average pooling and max pooling layers respectively before being fed into a shared multilayer perceptron (MLP). Then, the two outputs of the MLP are summed and activated by a sigmoid activation function to obtain the spatial attention feature map F. CA Then, the spatial attention feature map F CA The channel prior feature map F is obtained by element-wise multiplication with the input feature map F. C Subsequently, the channel prior feature map F C The input is fed into a 5×5 depthwise convolutional module, and the convolutional output is fed into the following three sets of depthwise convolutional branches: 1×7 depthwise convolutional layer + 7×1 depthwise convolutional layer; 1×11 depthwise convolutional layer + 11×1 depthwise convolutional layer; 1×21 depthwise convolutional layer + 21×1 depthwise convolutional layer.

[0090] The feature maps output from the three sets of deep convolution branches are added to the feature map output from the 5×5 deep convolution module to obtain the spatial attention feature map F. SA Then, a 1×1 convolutional layer is used to perform channel fusion to obtain a fused feature map; finally, the fused feature map is combined with the channel prior feature map F. C Multiplication yields the final output feature map F. OUT ;

[0091] Step 4: Use the training set, test set and validation set obtained in Step 2 to train, test and validate the above aircraft tire tread defect detection model, and finally obtain the trained aircraft tire tread defect detection model.

[0092] The hyperparameters were configured as follows: batch size of 16, training epochs of 300, learning rate of 0.02, input image size of 640×640, and SGD optimizer. First, the aircraft tire tread defect detection model was trained using the training set. Mosaic was enabled for data augmentation during training, and disabled in the last 10 epochs. Then, the training results were tested using the test set. Training was stopped when precision (P), recall (R), and mean precision (mAP) stopped increasing, and the loss function gradually converged. The final training weights were retained as the weights for the aircraft tire tread defect detection model, resulting in a trained model. Finally, the aircraft tire tread defect detection model was evaluated using the validation set to determine its performance in real-world scenarios.

[0093] Step 5: Deploy the trained aircraft tire tread defect detection model on the server. Deploy the embedded acquisition device, consisting of a Raspberry Pi, an industrial camera, and a display screen, on the unmanned inspection equipment. Use the industrial camera to take pictures of the aircraft tire tread at the aircraft maintenance site, and upload the captured images to the server under the control of the Raspberry Pi. The server receives the images to be inspected and automatically loads them into the trained aircraft tire tread defect detection model, outputs the aircraft tire tread defect detection results, and displays the results on the display screen.

[0094] The detection results include the defect category, bounding box, number of defects, confidence level, and inference time for each image to be detected for the aircraft tire tread.

[0095] To verify the impact of each improvement on the overall performance of the model, under the same training and testing conditions, the original YOLOv8 model without any improvements was used as the original model. The FPSC_CPCA module, MAFPN structure and C2f_MAEFS module were gradually introduced to build various model structures for ablation comparison experiments. The experimental results are shown in Table 1, where the best performing result is marked in bold.

[0096] Table 1 Ablation Experiment Results

[0097] method FPSC_CPCA MAFPN C2f_MAEFS P％ R％ mAP@0.5% YOLOv8 84.1 70.8 79.1 Method 1 √ 86.4 74 81 Method 2 √ 84.2 73.2 79.6 Method 3 √ 85.4 71.4 79.5 Method 4 √ √ 87.1 75.1 81.6 Method 5 √ √ √ 88 75.5 82.7

[0098] Table 1 shows that the original model achieved accuracy (P), recall (R), and mean precision (mAP@0.5%) of 84.1%, 70.8%, and 79.1%, respectively, on the fabricated aircraft tire defect detection dataset. Based on this, Methods 1, 2, and 3 each added three improvements to the original model, each enhancing the model's detection performance to varying degrees. First, replacing the original SPPF module with the FPSC_CPCA module improved the accuracy (P), recall (R), and mean precision (mAP@0.5%) of Method 1 by 2.3%, 3.2%, and 1.9%, respectively. Second, introducing the MAFPN structure into Method 4 improved the accuracy (P), recall (R), and mean precision (mAP@0.5%) by 3%, 4.3%, and 2.5%, respectively, compared to the original model, indicating that the synergy between the two methods further enhanced the model's feature extraction and multi-scale fusion capabilities. Finally, by replacing all C2f modules in the backbone network with C2f_MAEFS modules, the model achieved its highest performance metrics, improving accuracy (P), recall (R), and mean precision (mAP@0.5%) by 3.9%, 4.7%, and 3.6%, respectively, compared to the original model. These experimental results demonstrate that each improved module proposed in this invention effectively enhances target detection performance, and combining multiple improvement methods yields even more significant performance gains.

[0099] In addition, to more comprehensively evaluate the superiority of the method of the present invention, the model proposed in this invention was compared with a variety of mainstream YOLO series object detection algorithms under the same dataset and experimental conditions, including YOLOv7-tiny, YOLOv8n, YOLOv9t, YOLOv10n and YOLO11n. The experimental results are shown in Table 2.

[0100] Table 2 Comparison of experimental results

[0101]

[0102]

[0103] As shown in Table 2, the improved model proposed in this invention achieves higher values ​​in multiple evaluation indicators, including accuracy (P), recall (R), mean precision (mAP@0.5%), and mean precision (mAP@0.5:0.95%). The improvement is particularly significant in the mean precision (mAP@0.5:0.95), indicating that the improved model proposed in this invention has better overall detection performance in complex detection scenarios.

Claims

1. A method for automatic identification of aircraft tire tread defects, characterized in that: The automatic identification method for aircraft tire tread defects includes the following steps performed in sequence: Step 1: Collect images of the defective aircraft tire tread to obtain the raw image dataset; Step 2: Perform data augmentation on the original aircraft tire tread images in the above original image dataset, and then perform manual annotation to obtain the aircraft tire tread defect information dataset and divide it into training set, test set and validation set according to the proportion. Step 3: Based on the YOLOv8 model, construct an aircraft tire tread defect detection model; Step 4: Use the training set, test set and validation set obtained in Step 2 to train, test and validate the above aircraft tire tread defect detection model, and finally obtain the trained aircraft tire tread defect detection model. Step 5: Deploy the trained aircraft tire tread defect detection model on the server, deploy the embedded acquisition device consisting of Raspberry Pi, industrial camera and display screen on the unmanned inspection device, use the industrial camera to take pictures of the aircraft tire tread at the aircraft maintenance site, and upload the pictures to the server under the control of Raspberry Pi. The server receives the image to be detected and automatically loads it into the trained aircraft tire tread defect detection model, outputs the aircraft tire tread defect detection results, and displays the results on the screen.

2. The automatic identification method for aircraft tire tread defects according to claim 1, characterized in that: In step 1, the method for acquiring images of defective aircraft tire treads to obtain the original image dataset is as follows: Multiple original aircraft tire tread images, including three types of defects—cuts, cracks, and wear lines—were collected from multiple angles and under different lighting conditions at aircraft maintenance sites and on training aircraft in aviation colleges. The original image dataset is composed of all the original aircraft tire tread images.

3. The automatic identification method for aircraft tire tread defects according to claim 1, characterized in that: In step 2, the method for performing data augmentation on the original aircraft tire tread images in the aforementioned original image dataset, followed by manual annotation to obtain an aircraft tire tread defect information dataset, and then dividing it into training, testing, and validation sets proportionally, is as follows: Complete and relatively clear original aircraft tire tread images are selected from the original image dataset obtained in step 1. Then, they are processed using methods including translation, rotation, and mirroring. At the same time, the Albumentations data augmentation library is used to perform blurring, brightness adjustment, and the addition of light spots, rain, snow, and shadows. This process augments the original aircraft tire tread images to obtain enhanced images. Then, the labelimg annotation tool is used to annotate the original aircraft tire tread images and enhanced images according to the defect type in three categories. All annotated images constitute the aircraft tire tread defect information dataset and are divided into training set, test set, and validation set in a ratio of 8:1:

1.

4. The automatic identification method for aircraft tire tread defects according to claim 1, characterized in that: In step 3, the method for constructing an aircraft tire tread defect detection model based on the YOLOv8 model is as follows: The aircraft tire tread defect detection model includes a backbone network, a neck network, and a detection head. In the backbone network, the Bottleneck module in the C2f module of the YOLOv8 model backbone network is replaced with the MAEFS module to construct the C2f_MAEFS module; the SPPF module is replaced with the FPSC_CPCA module; in the neck network, the path aggregation feature pyramid network PAFPN structure of the YOLOv8 model is improved into a multi-branch auxiliary feature pyramid MAFPN structure. In the backbone network, the labeled image obtained in step 2 is used as the input image. First, a convolution operation is performed using the first convolutional layer with a stride of 2, thereby reducing the size of the input image from 640×640 to 320×320, while increasing the number of channels, to obtain the initial feature map P1. Then, the initial feature map P1 is downsampled through the second convolutional layer with a stride of 2, and processed by the first C2f_MAEFS module to reduce the size of the initial feature map P1 from 320×320 to 160×160, resulting in feature map P2. The feature map P2 is then downsampled using the third convolutional layer and processed by the second C2f_MAEFS module to reduce the size of the feature map P2 from 160×160 to 80×80, resulting in the feature map P3. Then, feature map P3 is downsampled using the fourth convolutional layer and processed by the third C2f_MAEFS module to reduce the size of feature map P3 from 80×80 to 40×40, resulting in feature map P4. The feature map P4 is then downsampled using the fifth convolutional layer and processed by the fourth C2f_MAEFS module to reduce the size of the feature map P4 from 40×40 to 20×20, resulting in the feature map P5. Finally, feature map P5 is input into the FPSC_CPCA module to enhance the receptive field without changing the feature map size, and outputs a high-level semantic feature map P6 with a size of 20×20. At this point, the backbone network outputs multiple layers of feature maps with different sizes for subsequent feature fusion and detection. The main function of the neck network is to perform bidirectional fusion of feature maps of different sizes output by the backbone network from top to bottom and bottom to top. In the top-down feature fusion path, the 40×40 scale P4 feature map output by the backbone network is first input into the first convolutional layer of the neck. It is downsampled by a convolution operation with a stride of 2, reducing its size from 40×40 to 20×20. It is then concatenated with the high-level semantic feature map P6 of the same size in the channel dimension. The concatenated feature map is processed by the C2f module to obtain the intermediate fusion feature map of the 20×20 scale. The 20×20 scale intermediate fused feature map is upsampled to increase its size from 20×20 to 40×40. The 80×80 scale feature map P3 output from the backbone network is input into the second convolutional layer at the neck. It is downsampled by a convolution operation with a stride of 2 to reduce its size from 80×80 to 40×40. The two 40×40 scale feature maps are then concatenated along the channel dimension. The concatenated feature map is processed by the C2f module to obtain the 40×40 scale intermediate fused feature map. Then, the 40×40 scale intermediate fused feature map is further upsampled to increase its size from 40×40 to 80×80. The 160×160 scale feature map P2 output by the backbone network is input into the third convolutional layer of the neck. It is downsampled by a convolution operation with a stride of 2 to reduce its size from 160×160 to 80×80. The two 80×80 scale feature maps are then concatenated along the channel dimension. The concatenated feature map is processed by the C2f module to obtain the 80×80 scale intermediate fused feature map. In the bottom-up feature fusion path, the 80×80 scale intermediate fusion feature map is concatenated with the 80×80 scale feature map obtained by upsampling the 40×40 scale intermediate fusion feature map. The concatenated feature map is processed by the C2f module to obtain an 80×80 output fusion feature map, which is used as the first layer output of the neck network. Subsequently, the intermediate fusion feature map of scale 80×80 and the output fusion feature map of scale 80×80 are input into the fourth and fifth convolutional layers of the neck for downsampling, so that their size is reduced from 80×80 to 40×40. Then, the two 40×40 feature maps obtained above are concatenated with the 40×40 intermediate fusion feature map. The concatenated feature map is processed by the C2f module to obtain the 40×40 output fusion feature map, which is used as the output of the second layer of the neck network. Furthermore, the 40×40 scale intermediate fusion feature map and the 40×40 scale output fusion feature map are input into the sixth and seventh convolutional layers of the neck for downsampling. Then, the two 20×20 scale feature maps obtained above are concatenated with the 20×20 scale intermediate fusion feature map. The concatenated feature map is processed by the C2f module to obtain the 20×20 scale output fusion feature map, which is used as the third layer output of the neck network. Finally, the three detection heads receive the three-layer output fusion feature map from the neck network for target detection prediction; among them, the 80×80 output fusion feature map is used for small-sized target detection, the 40×40 output fusion feature map is used for medium-sized target detection, and the 20×20 output fusion feature map is used for large-sized target detection.

5. The automatic identification method for aircraft tire tread defects according to claim 4, characterized in that: The MAEFS module includes one main branch and four parallel auxiliary branches, and its processing method is as follows: The feature map input to the MAEFS module first enters a 3×3 convolutional layer in the main branch to obtain the main branch output feature map. Simultaneously, the above input feature map is fed into four parallel auxiliary branches. These four auxiliary branches consist of the following structural units: the first processing unit uses 3×3, 6×6, 9×9, and 12×12 adaptive average pooling layers, respectively; the second processing unit is a 1×1 convolutional layer; the third processing unit is a 3×3 depthwise convolutional layer; the fourth processing unit is an upsampling unit; and the fifth processing unit is an EB edge enhancement module. Each auxiliary branch outputs its own feature map. Subsequently, the output feature maps of the four auxiliary branches are concatenated with the output feature map of the main branch to obtain a multi-scale fused feature map; then, the fused feature map is input into the DSM dual-domain selection mechanism module; finally, the feature map output by the DSM dual-domain selection mechanism module is input into a 1×1 convolutional layer to obtain the feature map output by the MAEFS module.

6. The automatic identification method for aircraft tire tread defects according to claim 5, characterized in that: The feature map input to the EB edge enhancement module is first subjected to an average pooling operation through a 3×3 average pooling layer. Then, the average pooled feature map is subtracted from the input feature map to generate a difference feature map. The difference feature map is then fed into a 3×3 convolutional layer with a Sigmoid activation function, and the output is an enhanced feature map. Finally, the enhanced feature map is added to the input feature map to output the resulting feature map with enhanced edges.

7. The automatic identification method for aircraft tire tread defects according to claim 5, characterized in that: The DSM dual-domain selection mechanism module includes a spatial domain selection module (SSM) and a frequency domain selection module (FSM), and its processing method is as follows: First, in the Spatial Domain Selection Module (SSM), the input feature map F is subjected to max pooling and average pooling operations respectively using max pooling and average pooling layers. Then, a spatial attention map F′ is generated through a 3×3 convolutional layer. Subsequently, the input feature map F is subjected to two sets of depthwise convolutional layers: one set consists of a 5×5 kernel with a dilation of 2 and a 7×7 kernel with a dilation of 3; the other set consists of a 3×3 depthwise convolutional layer. Finally, the two sets of depthwise convolutional results are multiplied and added element-wise with the spatial attention map F′ respectively to obtain the spatially enhanced feature map F. S In FSM, the spatially enhanced feature map F is first obtained through a mean filter. S Extract the low-frequency feature map Mean(F) S And obtain the corresponding high-frequency feature map through subtraction operation. Then the high-frequency feature map Compared with the above spatial enhancement feature map F S Element-wise multiplication is performed, and the results are fused through residual connections to finally output the enhanced feature map.

8. The automatic identification method for aircraft tire tread defects according to claim 4, characterized in that: The processing method of the FPSC_CPCA module is as follows: First, the input feature map is convolved through a 1×1 convolutional layer, and then sequentially fed into three 3×3 dilated convolutional layers with dilation rates of 1, 3, and 5. Next, the input feature map is concatenated with the feature maps output from all the dilated convolutions to obtain a fused feature map. This fused feature map is then fed into the CPCA attention mechanism module, and finally, it is convolved through a 1×1 convolutional layer to obtain the feature map output by the FPSC_CPCA module. The CPCA attention mechanism module includes channel attention and spatial attention branches; First, the input feature map F is processed by average pooling and max pooling layers, respectively, before being fed into a shared multilayer perceptron (MLP). Then, the two outputs of the MLP are summed and activated using a sigmoid activation function to obtain the spatial attention feature map F. CA Then, the spatial attention feature map F CA The channel prior feature map F is obtained by element-wise multiplication with the input feature map F. C Subsequently, the channel prior feature map F C The input is fed into a 5×5 depthwise convolutional module, and the convolutional output is fed into the following three sets of depthwise convolutional branches: 1×7 depthwise convolutional layer + 7×1 depthwise convolutional layer; 1×11 depthwise convolutional layer + 11×1 depthwise convolutional layer; 1×21 depthwise convolutional layer + 21×1 depthwise convolutional layer. The feature maps output from the three sets of deep convolution branches are added to the feature map output from the 5×5 deep convolution module to obtain the spatial attention feature map F. SA Then, a 1×1 convolutional layer is used to perform channel fusion to obtain a fused feature map; finally, the fused feature map is combined with the channel prior feature map F. C Multiplication yields the final output feature map F. OUT .

9. The automatic identification method for aircraft tire tread defects according to claim 1, characterized in that: In step 4, the method for training, testing, and validating the aircraft tire tread defect detection model using the training set, test set, and validation set obtained in step 2, and finally obtaining the trained aircraft tire tread defect detection model, is as follows: The hyperparameters were configured as follows: batch size of 16, training epochs of 300, learning rate of 0.02, input image size of 640×640, and SGD optimizer. First, the aircraft tire tread defect detection model was trained using the training set. Mosaic was enabled for data augmentation during training, and disabled in the last 10 epochs. Then, the training results were tested using the test set. Training was stopped when accuracy (P), recall (R), and mean precision (mAP) stopped increasing, and the loss function gradually converged. The final training weights were retained as the weights for the aircraft tire tread defect detection model, resulting in a trained model. Finally, the aircraft tire tread defect detection model was evaluated using the validation set to determine its performance in real-world scenarios.

10. The automatic identification method for aircraft tire tread defects according to claim 1, characterized in that: In step 5, the detection results include the defect category, bounding box, number of defects, confidence level, and inference time for each image to be detected for the aircraft tire tread.

Images