Aviation rivet classification and anomaly detection method based on deep learning

Abstract

The present invention discloses a method for aviation rivet classification and anomaly detection based on deep learning, comprising the following steps: step 1, collecting aviation rivet image data, and preprocessing the image data; step 2, labeling the aviation rivet image data with target box categories and data enhancement processing; step 3, training the aviation rivet classification and anomaly detection model by using transfer training plus frozen training; step 4, using a convolutional neural network with a DarkNet‑53 structure as a baseline as the backbone network to extract features for neural network reasoning, using the Focus layer, cross-stage local network layer, spatial pyramid pooling structure module, upsampling layer, and connection layer as the neck network for feature fusion, and finally using a decoupled detection head to generate a detection result. Compared with the prior art, the positive effect of the present invention is that the aviation rivet classification and abnormal target detection accuracy and speed of the present invention have absolute advantages.

Description

Technical Field

[0001] The present invention relates to an aviation rivet classification and anomaly detection method based on deep learning. Background Art

[0002] The main material of modern large civil airliners is aluminum alloy, and riveting technology is usually used to reduce the weight of the aircraft. Rivets, as the lightest fasteners with the best connection strength, are naturally favored by major aircraft manufacturers. According to statistics, each aircraft is equipped with about one million rivets, mainly including solid rivets and special rivets. The main reason for rivet loosening is flight vibration. If it falls on the runway during takeoff and landing, it may be sucked into the following aircraft as runway foreign object debris (FOD), causing tires to be punctured, engines and fuselage damage, and even causing major aviation accidents. Therefore, how to quickly and accurately locate the position of rivets, identify rivet classification, and whether there are abnormal conditions is of great significance to improving flight safety.

[0003] At present, the research on aviation rivet detection at home and abroad is still in its infancy. The existing achievements can be roughly divided into three categories: (1) Detection of abnormal conditions of aviation rivets based on magneto-optical imaging and pulsed eddy current technology. Among them, ZHIWEI ZENG et al. proposed a finite element model for automatic detection of aviation rivets, YUNZE HE et al. proposed a non-destructive detection method for riveted structural defects, and GAO Qingji et al. and VALENTYN UCHANIN respectively realized the detection of rivet surface crack defects. (2) Detection of abnormal conditions of aviation rivets based on optical technology. Among them, GAIL OVERTON conducted a rapid detection of the three-dimensional coordinates of aircraft rivets, LI Hongwei et al. proposed an adaptive local point cloud density calculation method to realize the detection of riveting quality; BI Chao et al. developed a cross-line structured light vision detection system to realize the detection of rivet hole orientation. (3) Detection of aviation rivets based on machine vision. The main application scenarios are as follows: ① Detection of rivet damage. Chen Jianfei et al. and Yang Fei et al. used support vector machines and noise-resistant local binary pattern coding algorithms to detect rivet structural defects and rivet surface defects respectively. ② Detection of rivet position. Xing Xueliang et al. proposed a detection method based on machine vision to detect rivet position and rivet size; Yu Qiang et al. proposed a detection method based on monocular vision to solve the problem of rivet identification and positioning difficulties. ③ Detection of rivet holes. Tian Qinglian et al. proposed a rivet hole machine vision recognition algorithm based on scattered point clouds.

[0004] In summary, although existing research has achieved certain results in aviation rivet detection, the accuracy and speed of the detection results still have a lot of room for improvement. In addition, there have been no research results on the classification of aviation rivets and the rapid identification and positioning of abnormal conditions. Therefore, the present invention proposes an aviation rivet classification and anomaly detection algorithm based on deep learning, which realizes the target detection of rivet classification and abnormal conditions during the inspection process, in order to help maintenance engineers locate various rivet-related faults and ensure the safety of aviation operations. Summary of the invention

[0005] In order to overcome the above-mentioned shortcomings of the prior art, the present invention proposes an aviation rivet classification and anomaly detection method based on deep learning.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a method for aviation rivet classification and anomaly detection based on deep learning, comprising the following steps:

[0007] Step 1: Collect aviation rivet image data and pre-process the image data;

[0008] Step 2: label the target box category and perform data enhancement processing on the aviation rivet image data;

[0009] Step 3: Use transfer training plus frozen training to train the aviation rivet classification and anomaly detection model;

[0010] Step 4: Use the convolutional neural network with the DarkNet-53 structure as the baseline as the backbone network to extract features for neural network inference. Use the Focus layer, cross-stage local network layer, spatial pyramid pooling structure module, upsampling layer, and connection layer as the neck network to perform feature fusion. Finally, use the decoupled detection head to generate detection results.

[0011] Compared with the prior art, the present invention has the following positive effects:

[0012] Aviation rivets are of great significance for reducing aircraft weight, preventing metal fatigue, and improving flight reliability and stability. Rivet loosening and falling off will seriously threaten flight safety. Aiming at the problem of low accuracy of small target detection of aviation rivets, the present invention proposes an aviation rivet classification and anomaly detection method based on deep learning. Firstly, aviation rivets are classified according to the appearance of the rivet head, a data set of aviation rivets is produced, and an aviation rivet classification and anomaly detection model is constructed. The detection results are evaluated from the indicators such as confidence, precision, recall rate, mAP, etc., and the algorithm is compared with the detection results of Yolox-s, Yolox-m, Yolov5-s, Yolov5-m, and Yolov4. The results show that:

[0013] (1) The algorithm can classify aviation rivets and detect abnormal conditions. The confidence level of the detection results is greater than 90%, and the average precision, recall rate and AP value are above 95%, 85% and 88% respectively.

[0014] (2) Because the algorithm adopts a lighter network architecture and has made improvements in the detection end, anchor frame mechanism and label assignment method, it has absolute advantages in the accuracy and speed of aviation rivet classification and abnormal target detection compared with other Yolox algorithms of the same version and Yolov5 and Yolov4 algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be described with reference to the accompanying drawings, in which:

[0016] Figure 1 Flow chart for aviation rivet identification;

[0017] Figure 2 This is a classification table for rivets. DETAILED DESCRIPTION

[0018] A deep learning-based aviation rivet classification and anomaly detection method, such as Figure 1 As shown, it includes five steps: aviation rivet image data collection and classification, data set preparation and data enhancement, model training, aviation rivet classification identification and anomaly detection and recognition, and detection result evaluation, among which:

[0019] Step S1: The maintenance engineer wears smart glasses (or other video shooting devices) to perform a walk-around inspection, collects videos and pictures of rivets at specific locations on the fuselage, and at the same time connects to a mobile phone or computer terminal through a 5G network to transmit and save videos and images in real time. Then, the collected video and image data are cleaned and sorted, and the obtained rivet samples are divided.

[0020] Step S2: sorting the collected aviation rivet data, labeling the category target frame and dividing the data set; after labeling, enhancing the target data set by using Mosaic data enhancement and Mix Up data enhancement methods;

[0021] Step S3: training the aviation rivet classification and anomaly detection model by using transfer training plus frozen training;

[0022] Step S4: A convolutional neural network based on the DarkNet-53 structure is used as the backbone network to extract features for neural network calculation. The Focus layer, the cross-stage partial network layer (CSP), the spatial pyramid pooling structure module (SPP), the upsampling layer, and the connection layer are used as the neck network to perform feature fusion. Finally, the decoupled detection head is used to generate detection data.

[0023] Step S5: To verify the effectiveness of the proposed algorithm in target detection, the recall rate, precision rate, AP (Average Precision) value, and mean average precision (mAP) are used to evaluate the detection results.

[0024] The step S1 specifically further includes:

[0025] Step S1.1: Collect pictures of aviation rivets at different angles, distances, and brightness, and extract appropriate samples frame by frame from the collected video data.

[0026] Step S1.2: The classification of aviation rivets is based on the national standards commonly used in aerospace, and can be divided into semicircular head, 90° countersunk head, 120° countersunk head, blind rivets and flat head rivets, with codes GB867, GB869, GB954, B12615 and GB109 respectively. Since the present invention will adopt the target detection method based on deep learning, the rivets are reclassified according to their appearance images, such as Figure 2 As shown: (1) Countersunk Head Rivets, marked as 'CHr', are mainly used for rivets on the aircraft surface that need to be smooth and not subject to light loads, and can effectively reduce wind resistance; (2) Flat Head Rivets, marked as 'FHr', are mainly used to strengthen seams; (3) Round Head Rivets, marked as 'RHr', are mainly used for rivets that bear large lateral loads, and play a role in improving the strength of the fuselage and maintaining structural stability. They are widely used in the fuselage; (4) Blind Rivets, marked as 'Br', are mainly used for single-sided riveting, and are suitable for parts of the fuselage that are not convenient for two-sided riveting; (5) Phillips Screws, marked as 'Ps', are mainly used in the fuselage and its interior for fastening; (6) Abnormal, marked as 'Abn'.

[0027] In order to better implement the present invention, the step S2 specifically further includes the following steps:

[0028] Step S2.1: After sorting the sample data, use the labeling tool to label various types of data. The target box categories are: CHr, FHr, Br, RHr, Ps, and Abn. The labeled image data is divided into training set, validation set, and test set in a ratio of 7:1:2.

[0029] Step S2.2: Mosaic data enhancement and Mix Up data enhancement are performed on the image data annotated in step S2.1. Specifically, mosaic data enhancement is to randomly read 4 pictures each time from the data set prepared in step S2.1, and then flip (flip the original picture left and right), scale (scale the original picture), and change the color gamut (change the brightness, saturation, and hue of the original picture) and other operations on the 4 pictures. After the operation is completed, the original pictures are placed in four directions according to the first picture on the upper left, the second picture on the lower left, the third picture on the lower right, and the fourth picture on the upper right. Then, the fixed areas of the four pictures are cut out using a matrix method and spliced ​​into a new picture, which contains a series of contents such as a labeling box. This operation enriches the background information of the image, and stitching the four images together also increases the number of data set samples in disguise. When performing batch normalization, it is equivalent to calculating the four images at the same time, which greatly improves the training efficiency.

[0030] Hybrid data enhancement is an algorithm used in computer vision to enhance images by mixing different classes. It can mix images of different classes to expand the training data set. The formula is as follows:

[0031] λ=Beat(α,β)

[0032]

[0033]

[0034] in is the mixed image, Beat is the Beta distribution, α, β are constants, λ is the random sampling number calculated from the Beta distribution, (x i ,y i )、(x j ,y j ) are two randomly selected samples and corresponding labels in the same batch.

[0035] In order to further implement the present invention, the step S3 specifically comprises the following steps:

[0036] Step S3.1: Pre-train based on the VOC dataset to obtain pre-training weights. The training parameters are set to score threshold of 0.5, batcH_size of 10, max_boxes of 20, model_image_size of (640×640), and initial learning rate of 0.001. When the loss value does not decrease for multiple times, the training is automatically terminated and the unfreezing training of step S3.2 is performed.

[0037] Step S3.2: While performing step S3.1, freeze training is performed to freeze some common pre-trained weights, such as the backbone network, and put more resources into the network parameter calculation in the later part of the training. This can effectively reduce the training time and greatly improve resource utilization. When the loss value does not decrease for many times, the frozen training is automatically ended, and then the unfrozen training is started, and all layers of the model are trained together. The frozen training parameters are score threshold of 0.5, batch_size of 8, max_boxes value of 20, and initial learning rate of 0.0001.

[0038] Step S3.3: After the neural network is used to extract features in step S3.1, the network outputs the predicted coordinate information (x, y, w, h), as well as the confidence value of the target prediction result. Among them, x, y represent the coordinate point of the upper left corner of the predicted data; w, h represent the width and height of the predicted data. The difference between the coordinate value of the predicted result and the coordinate value of the real target can use CIoU Loss as the loss function, and the formula is as follows:

[0039]

[0040] Among them, b and b gt Represents the center point of the anchor box and the target box; ρ represents the Euclidean distance between the two center points; c represents the diagonal distance of the minimum rectangle covering the anchor box and the target box; α is a parameter used for trade-off; v is a parameter used to measure the consistency of the aspect ratio.

[0041] CIoULoss takes into account the distance, overlap rate, scale, and penalty term between the target and the predicted box, which makes the target box regression more stable and does not have problems such as divergence during training like IoU and GIoU. In addition to considering the center point distance between the real box and the predicted box, CIoU also considers the diagonal distance of the minimum enclosing box of the two boxes, so the predicted box can be closer to the position of the real box, making the prediction result more accurate.

[0042] In order to further implement the present invention, the step S4 specifically includes the following steps:

[0043] Step S4.1: Based on the weights of aviation rivet classification obtained after training in step S3, aviation rivet data is classified and anomalies are identified. The convolutional neural network reasoning process can be summarized into three stages: feature extraction stage, feature information collection and processing stage, and prediction stage. Feature extraction stage: The input end inputs a picture of size 416×416×3. After cropping, splicing and convolution processing in the Focus layer, a feature map of 208×208×24 is obtained. After layer-by-layer reasoning of the convolution layer, CSP layer and SPP module, feature maps of sizes 52×52×96, 26×26×192, and 13×13×384 are obtained in the middle layer, lower middle layer and bottom layer of the backbone network, respectively, and then participate in the network neck feature fusion.

[0044] Anchor-Free mechanism is adopted in the reasoning process. After the backbone network, information of different sizes is input into the prediction branch after downsampling. Each central anchor point only generates one candidate box, which reduces the computational complexity.

[0045] Step S4.2: Insert the SE (Squeeze-and-Excitation) attention mechanism at the end of the backbone network. This module automatically obtains the importance of each feature channel through autonomous learning, and then promotes useful features and suppresses features that are not very useful for the current task according to this importance. The squeeze module compresses features along the spatial dimension, converting each two-dimensional feature channel into a real number, which has a global receptive field to some extent, and the output dimension matches the number of input feature channels. It represents the global distribution of responses on the feature channel, and enables the layer close to the input to obtain a global receptive field. The specific operation is as follows: perform global average pooling on the original feature maps 52×52×96, 26×26×192, and 13×13×384, and then obtain three feature maps of sizes 1×1×96, 1×1×192, and 1×1×384, which can have a global receptive field.

[0046] The 1×1×96, 1×1×192, and 1×1×384 feature maps output by the Excitation module are then passed through two fully connected neural networks, and finally a mechanism similar to the gate in a recurrent neural network is used to generate weights for each feature channel through parameter adjustment, where the parameters are learned to explicitly model the correlation between feature channels.

[0047] The last step is feature recalibration, which uses the result obtained by Excitation as the weight, and then weights it channel by channel to the channels corresponding to the three feature maps through multiplication, completing the recalibration of the original features in the channel dimension and serving as the input data for the next level.

[0048] Step S4.3: In the stage of collecting and processing feature information, the neck of the network adopts the feature pyramid network (FPN) + path aggregation network structure (PAN) to fuse the three shallow semantic features with the deeper semantic features, so that the overall network can better extract the target feature information. Finally, feature maps of sizes 52×52×96, 26×26×192, and 13×13×384 are obtained in the middle layer, lower middle layer, and bottom layer of the neck of the network, respectively.

[0049] Step S4.4: Input to the decoupled detection head for inference prediction to obtain the predicted prospect, predicted type and predicted position of the target. The decoupled detection head (Decoupled Head) uses a 1×1 convolution for channel dimensionality reduction, which reduces the time consumption of inference, followed by two 3×3 convolutions for regression or classification processing. The Cls branch is mainly used to distinguish the category of the target box and calculate the target confidence; the Obj branch is mainly used to determine whether the target box is the identified object; the Reg branch is mainly used to infer the coordinate information (x, y, w, h) of the target box. Finally, the feature information obtained by the three branches is spliced ​​and fused together to obtain complete target recognition information. In the face of multi-channel recognition tasks, since the Decoupled Head increases the number of prediction branches and convolutions, regression and classification are inferred separately, which improves the expression ability of the prediction head.

[0050] Step S4.5: When assigning labels, use the advanced positive and negative sample assignment strategy (Sim-Optimal Transport Assignment for Object Detection, SimOTA). First, use the central prior method to determine the candidate region, calculate the loss values ​​of Reg. and Cls. between the anchor box and the true target in the candidate region, and then construct the cost function, as shown in the following formula.

[0051]

[0052] According to the size of the cost function, k candidate boxes with the lowest cost are selected, and finally the final prediction box is obtained after removing duplicate candidate boxes according to the global situation.

[0053] In order to further implement the present invention, the step S5 specifically includes the following steps:

[0054] Step S5.1: To verify the effectiveness of the proposed algorithm in target detection, recall rate, precision rate, AP (Average Precision) value, and mean average precision (mAP) are used for evaluation.

[0055]

[0056]

[0057] Among them, recall rate R refers to the probability of being detected correctly among all detected targets, and precision rate P refers to the probability of actually being the correct target among all detected targets. TP (true positive) is the number of correctly detected targets to be detected, FN (false positive) is the number of undetected targets to be detected, and FP (false negative) is the number of incorrect targets detected as targets to be detected.

[0058] Both AP and mAP are used to measure the overall performance of the model. AP does not refer to the average precision of all detected targets of a single type, but the area of ​​the graph enclosed by the PR curve with recall as the horizontal axis and precision as the vertical axis. The value can be obtained by integration method in actual calculation, as shown in the following formula. The average AP value of all types of targets can be used to measure the overall detection performance of the model, that is, the mAP value.

[0059]

[0060] Step S5.2: According to the target detection related evaluation indicators mentioned in S5.1, the recall rate (R), precision rate (P), and AP value of aviation rivet target category detection are obtained as shown in the following table.

[0061] Precision, recall and AP values ​​for different types of rivets

[0062]

[0063] It can be seen that the model can detect normal and abnormal conditions of various types of rivets. Among them, the detection accuracy of the targets is higher than 95%, especially the recognition accuracy of flat head rivets and semicircular head rivets is as high as 99%, and the accuracy of blind rivets is relatively low, up to 95.08%; the recall rate of the targets is above 85%, especially the recognition accuracy of semicircular head rivets and cross groove head rivets is as high as 96%, and the recall rate of blind rivets and abnormal conditions is relatively low, about 86%; the recognition accuracy of the AP value of the targets is greater than 88%, except for abnormal conditions, the AP of other types of rivets is above 92%.

[0064] In order to verify the performance of the algorithm of the present invention, a comparative experiment was conducted with the traditional Yolov4, Yolov5 algorithms and other versions of the Yolox algorithm in the same scenario, and the performance comparison of the detection results is shown in the following table.

[0065] Performance comparison of 6 algorithms

[0066]

[0067]

[0068] It can be seen that all algorithms can realize the classification of aviation rivets and the detection of abnormal targets. The overall performance is ranked from high to low as follows: the algorithm proposed in this paper, Yolox-s, Yolox-m, Yolov5-s, Yolov4, and Yolov5-m. Among them, the precision, recall, and average precision of the results obtained by the algorithm in this paper are 97.80%, 92.09%, and 95.07%, respectively, which are 4.81%, 8.06%, 6.49% and 2.51%, 6.06%, and 5.62% higher than other versions of the same algorithm, Yolox-m and Yolox-s algorithms, respectively. It is also a significant improvement over the results obtained by Yolov5-m, Yolov5-s, and Yolov4 algorithms. The GPU processing speed of the proposed algorithm is 84.56f / s, which is 50.97% and 16.39% higher than other versions of the same algorithm, Yolox-m and Yolox-s, respectively, and also has different degrees of improvement over Yolov5-m, Yolov5-s, and Yolov4. Compared with the same version of the algorithm, the speed improvement of the deep learning-based aviation rivet classification and abnormal target detection algorithm should be attributed more to the use of a lightweight network structure framework; the improvement of the accuracy of aviation rivet classification and abnormal target detection with different versions of the algorithm should be attributed to the improvement of the detection head, anchor-free frame mechanism, and attention mechanism.

Claims

1. A method for aviation rivet classification and anomaly detection based on deep learning, characterized in that: The steps include: Step 1: Collect aviation rivet image data and pre-process the image data; Step 2: label the target box category and perform data enhancement processing on the aviation rivet image data; Step 3: Train the aviation rivet classification and anomaly detection model by using transfer training plus frozen training to obtain the weights of the aviation rivet classification and anomaly detection model; Step 4: Based on the weights of the aviation rivet classification and anomaly detection model obtained after training in step 3, classify the aviation rivet data and identify anomalies; The reasoning process of the aviation rivet classification and anomaly detection model includes the feature extraction stage, the feature information collection and processing stage, and the prediction stage, among which: Step 1, feature extraction stage: The input image of size 416×416×3 is cropped, spliced ​​and convolved in the Focus layer to obtain a feature map of 208×208×24. After layer-by-layer reasoning of the convolution layer, the cross-stage local network layer CSP and the spatial pyramid pooling structure module SPP, feature maps of sizes 52×52×96, 26×26×192 and 13×13×384 are obtained in the middle layer, lower middle layer and bottom layer of the backbone network respectively, and then participate in the network neck feature fusion; In the inference process, the anchor-free frame mechanism is adopted. After the backbone network, information of different sizes is input into the prediction branch through downsampling. Each central anchor point only generates one candidate frame. The SE attention mechanism is inserted at the end of the backbone network to automatically obtain the importance of each feature channel through autonomous learning. Then, according to this importance, useful features are promoted and features that are not very useful for the current task are suppressed. Step 2: Collect and process feature information: The neck of the network uses a feature graph pyramid network and a path aggregation network structure to fuse shallow semantic features with deeper semantic features. Finally, feature graphs of sizes 52×52×96, 26×26×192, and 13×13×384 are obtained in the middle, lower-middle, and bottom layers of the neck of the network, respectively. The third step is to input it into the decoupled detection head for inference prediction to obtain the predicted prospect, predicted type and predicted position of the target; The fourth step is to use the positive and negative sample allocation strategy SimOTA when assigning labels. First, use the center prior method to determine the candidate area, build the cost function, and then select k candidate boxes with the lowest cost according to the size of the cost function. Finally, remove the duplicate candidate boxes according to the global situation to get the final prediction box.

2. The aviation rivet classification and anomaly detection method based on deep learning according to claim 1 is characterized in that: The methods for performing data enhancement processing on aviation rivet image data in step 2 include mosaic data enhancement and hybrid data enhancement.

3. The aviation rivet classification and anomaly detection method based on deep learning according to claim 2 is characterized in that: The mosaic data enhancement method is as follows: four pictures are randomly read from the data set each time, and the four pictures are flipped, scaled and color gamut changed respectively; the four pictures are then placed in the upper left, lower left, lower right and upper right positions in sequence, and then the fixed areas of the four pictures are cut out using a matrix method and spliced ​​into a new picture.

4. The aviation rivet classification and anomaly detection method based on deep learning according to claim 2 is characterized in that: Mixed data enhancement is performed according to the following formula: λ=Beat(α,β) in is the mixed image, Beat is the Beta distribution, α, β are constants, λ is the random sampling number calculated from the Beta distribution, (x i ,y i )、(x j ,y j ) are two randomly selected samples and corresponding labels in the same batch.

5. The aviation rivet classification and anomaly detection method based on deep learning according to claim 1 is characterized in that: The method of training the aviation rivet classification and anomaly detection model by using transfer training plus frozen training in step 3 is: pre-training based on the VOC dataset to obtain the pre-training weight, and the training parameter score threshold is set to 0.

5. The batch_size is 10, the max_boxes value is 20, the model_image_size is 640×640, and the initial learning rate is 0.

001. At the same time, frozen training is used to freeze the pre-trained weights of the general part. When the loss value does not decrease for many times, the frozen training is automatically ended, and then the unfreezing training is started. Then all layers of the model are trained together. The score threshold of the frozen training parameters is set to 0.5, the batch_size is 8, the max_boxes value is 20, and the initial learning rate is 0.0001.

6. The aviation rivet classification and anomaly detection method based on deep learning according to claim 5 is characterized in that: The difference between the coordinate values ​​of the prediction results of the aviation rivet classification and anomaly detection model in step 3 and the coordinate values ​​of the real target uses CIoU Loss as the loss function, and the formula is as follows: Among them, b and b gt represents the center point of the anchor box and the target box; ρ represents the Euclidean distance between the two center points; c represents the diagonal distance of the minimum rectangle covering the anchor box and the target box; α is the parameter used for trade-off; v is the parameter used to measure the consistency of the aspect ratio.

7. The aviation rivet classification and anomaly detection method based on deep learning according to claim 1 is characterized in that: In the SE attention mechanism, the Squeeze module is used to compress features along the spatial dimension, converting each two-dimensional feature channel into a real number, and the output dimension matches the number of input feature channels; The result obtained by using the Excitation module is used as the weight, and then weighted to the channels corresponding to the three feature maps channel by channel through multiplication, completing the recalibration of the original features in the channel dimension and serving as the input data for the next level.

8. The aviation rivet classification and anomaly detection method based on deep learning according to claim 1 is characterized in that: The decoupled detection head uses a 1×1 convolution for channel dimensionality reduction, followed by two 3×3 convolutions for regression or classification processing. The Cls branch is used to distinguish the category of the target box and calculate the target confidence. The Obj branch is used to determine whether the target box is the identified object. The Reg branch is used to infer the coordinate information (x, y, w, h) of the target box. Finally, the feature information obtained by the three branches is spliced ​​and fused together to obtain complete target recognition information.

9. The aviation rivet classification and anomaly detection method based on deep learning according to claim 1 is characterized in that: When preprocessing aviation rivet image data, rivets are classified as follows: (1) Countersunk rivet, marked as CHr; (2) Flat head rivets, marked as FHr; (3) Round head rivets, marked as RHr; (4) Blind rivets, marked as Br; (5) Cross-slot rivets, marked as Ps; (6) Abnormal situation, marked as Abn.

Images