Adaptive airport runway underground structure disease automatic detection method

Abstract

The application discloses an adaptive airport runway underground structure disease automatic detection method, which comprises the following steps: obtaining different airport runway underground B-scan images; constructing a runway underground structure disease detection model; extracting features p1 to p5 of the B-scan images; obtaining new features p*5; performing multi-scale feature fusion on the features, and then inputting the features into a detection head to perform positioning and classification of underground diseases; calculating an inter-domain distance; determining the optimal weight of the runway underground structure disease detection model; inputting the B-scan images of the airport runway underground to be detected into the runway underground structure disease detection model with the optimal weight determined, and outputting positioning and classification results of disease targets and the like. The application can train a model by using an unlabeled data set. On one hand, more effective features can be extracted, the recognition accuracy of the airport runway underground disease target is improved, the system can more accurately and effectively recognize the target type. On the other hand, the extracted feature information is more common, and can be applied to different airport data sets.

Description

Technical Field

[0001] This invention belongs to the field of civil aviation technology, and specifically relates to an adaptive automatic detection method for defects in underground structures of airport runways. Background Technology

[0002] Runways are a crucial component of airports, serving as the direct operational area for aircraft takeoffs and landings. Therefore, the structural integrity of airport runways plays a vital role in ensuring aircraft safety during flight. However, due to factors such as the loads caused by high-frequency aircraft operation, the seepage of groundwater, and the contraction and expansion effects of extreme temperature changes, airport runways inevitably develop various underground structural defects. These defects reduce the structural strength and integrity of the runway, severely impacting its service life and even posing a threat to aviation and passenger safety. Therefore, it is necessary to inspect these underground defects to ensure the safety of daily aircraft operations.

[0003] Currently, commonly used airport pavement inspection methods include manual inspection, core sampling, ultrasonic testing, and falling weight deflectometer testing. Ground-penetrating radar (GPR), as a non-destructive technology, is fast, safe, and accurate, and is a commonly used detection device in airport runway inspections. However, many challenges remain in using GPR to detect defects in the underground structures of airport runways.

[0004] Initially, ground-penetrating radar (GPR) data was primarily interpreted manually. Due to the complexity and heterogeneity of the underground environment, coupled with the noise generated by the GPR equipment itself, the acquired GPR data was both complex and prone to distortion and defects. These interferences made manual interpretation challenging, resulting in high costs and time consumption.

[0005] The raw data collected by ground-penetrating radar (GPR) is signal data, so much of the work on interpreting GPR data is based on standard signal processing methods. For example, identifying defects by extracting the time-frequency characteristics of the GPR signal, identifying voids using continuous wavelet transform, and detecting shallow pavement through transverse wave characteristics. However, the biggest problem with these signal processing-based methods is their susceptibility to clutter and noise.

[0006] Besides signal processing-based methods, many studies also rely on traditional machine learning. Examples include using Support Vector Machines (SVMs) to detect hyperbolic features, designing a radial basis function-based classifier to identify GPR targets, and applying open-scan clustering algorithms to extract hyperbolic reflection feature points. However, these traditional machine learning algorithms often require manual feature design, which is time-consuming and labor-intensive. Furthermore, these algorithms are often designed for single targets like hyperbolic reflections, as these targets exhibit distinct hyperbolic characteristics that differentiate them from other underground targets. This simplifies practical applications.

[0007] Thanks to the powerful automated feature extraction capabilities of deep learning networks, an increasing number of research studies are choosing object detection algorithms for disease detection based on B-SCAN images. Some of these studies are based on classic one-stage and two-stage models. Besides 2D object detection, researchers are also focusing on 3D detection, mining 3D feature information across multiple B-SCAN images and combining it with other dimensions of GPR data, such as A-SCAN images, to detect underground targets. These algorithms have demonstrated excellent performance, indicating the effectiveness and trend of using deep learning for interpreting GPR data.

[0008] Therefore, deep learning technology is gradually being widely applied to the automatic identification of underground diseases.

[0009] Using deep learning models to detect runway subsurface defects can significantly improve detection efficiency, and there are currently many studies on defect detection based on GPR data. However, these studies have some drawbacks: (1) Most of these algorithms are based on supervised model training using labeled data, and cannot utilize most of the unlabeled data in real-world scenarios. (2) They cannot solve the problem of extremely low transfer and generalization ability caused by the differences in the distribution of GPR data samples collected from different airports.

[0010] Currently, to address the performance degradation caused by data distribution discrepancies, most existing technical solutions involve expanding the size of the training dataset. The aim is to improve detection performance on unknown datasets by using a large number of labeled data samples and supervised training methods to allow the model to learn more disease features.

[0011] However, collecting large amounts of underground data from airport runways is extremely difficult, and as mentioned earlier, manually interpreting GPR data is challenging and inefficient, while labeling datasets is time-consuming and costly. Therefore, addressing the problem of sample distribution differences by expanding the size of the training dataset is costly and challenging. Summary of the Invention

[0012] To address the aforementioned problems, the present invention aims to provide an adaptive automatic detection method for defects in underground structures of airport runways.

[0013] To achieve the above objectives, the adaptive automatic detection method for defects in underground airport runways provided by the present invention includes the following steps performed in sequence:

[0014] Step 1: Obtain B-scan images of the underground of different airport runways using a robot equipped with a multi-channel ground-penetrating radar, and then divide them into source domain datasets and target domain datasets.

[0015] Step 2: Construct a YOLOv5-based model for detecting defects in underground runway structures. This model includes a shared weight, a two-stream backbone network for feature extraction and inter-domain distance calculation, a self-attention module for acquiring global information, an FPN neck network for multi-scale feature fusion, and three detection heads for different feature sizes.

[0016] Step 3: Input the source domain dataset and target domain dataset obtained in Step 1 into the runway underground structure defect detection model constructed in Step 2, and use the dual-stream backbone network to extract the features p1 to p5 of the B-scan images in the source domain dataset and target domain dataset respectively;

[0017] Step 4: Input the feature p5 obtained in Step 3 into the self-attention module to obtain a new feature p*5;

[0018] Step 5: Input the features p3 and p4 obtained in Step 3 and the new feature p*5 obtained in Step 4 into the FPN neck network for multi-scale feature fusion. Then, input the three fused features at different scales into the corresponding three detector heads to locate and classify underground diseases.

[0019] Step 6: Calculate the inter-domain distance between the two features p3 extracted from the source domain dataset and the target domain dataset in Step 3 using unsupervised domain adaptation (MMD);

[0020] Step 7: Calculate the classification loss and regression loss by combining the localization and classification results obtained in Step 5 with the actual disease target labels. Then, calculate the total loss function of the runway underground structure disease detection model by combining the inter-domain distance obtained in Step 6. Then, optimize the model through backpropagation and iterate until the total loss value converges, thereby determining the final weight of the runway underground structure disease detection model.

[0021] Step 8: After the B-scan image of the airport runway underground structure to be detected is preprocessed as described in Step 1, it is input into the runway underground structure defect detection model whose final weights have been determined in Step 7. The runway underground structure defect detection model outputs the final location and classification results of the defect targets.

[0022] In step 1, the method of obtaining B-scan images of the underside of different airport runways using a robot equipped with a multi-channel ground-penetrating radar, and then dividing them into source domain datasets and target domain datasets, is as follows:

[0023] Using a robot equipped with a multi-channel ground-penetrating radar, multiple raw B-scan images of the underground of real airport runways are collected according to a pre-planned route. The raw B-scan images are then preprocessed sequentially, including zero-bias correction, zero-point adjustment, background removal, data filtering, gain adjustment, and image visualization, to obtain B-scan images. All B-scan images are then divided into source domain datasets and target domain datasets. Both the source domain dataset and the target domain dataset include four typical underground defects or objects of airport runways: voids, cracks, steel reinforcement, and settlement.

[0024] In step 2, the backbone network includes a Focus module, an L-Dark2 module, an L-Dark3 module, an L-Dark4 module, and an L-Dark5 module; wherein the L-Dark module includes the C3 module in YOLOv5 and a deep row attention module, used to acquire rich strata information and establish the correlation between runway underground defects and runway underground structures; the deep row attention module includes a learnable position encoding module and a row attention module.

[0025] In step 3, the method of inputting the source domain dataset and target domain dataset obtained in step 1 into the runway underground structure defect detection model constructed in step 2, and extracting features p1 to p5 of the B-scan images in the source domain dataset and target domain dataset respectively using a dual-stream backbone network is as follows:

[0026] First, the B-scan images in the source and target domain datasets are unified to 448*448 pixels. Then, they are input into the backbone networks of the runway underground structure defect detection model. After that, they are passed through the Focus module, L-Dark2 module, L-Dark3 module, L-Dark4 module and L-Dark5 module in sequence to obtain features with lengths and widths of 224*224 pixels, 112*112 pixels, 56*56 pixels, 28*28 pixels and 14*14 pixels, respectively, which are denoted as features p1, features p2, features p3, features p4 and features p5.

[0027] The learningable positional encoding module in the L-Dark module processes information as follows: A one-dimensional positional encoding is embedded into the algorithm model, representing only depth positional information. Specifically, using the row feature vectors in the features as units, the absolute depth positional information is integrated into the extracted features using a learningable positional encoding method. Let the feature size be c*h*w, and the learned depth positional information size be c*h*1. Then, the depth positional information is copied and expanded to a size matching the feature size c*h*w. Finally, a simple addition operation is used to directly add it to the features, thus completing the fusion of depth positional information and features.

[0028] The row attention module processes the data as follows: During computation, each row is kept independent while information within the row is mixed. That is, each row of features is used as a computational unit, and then self-attention is applied for computation.

[0029] The formula for calculating self-attention is as follows:

[0030] (1);

[0031] In the formula, Q, K, and V represent the query vector, key vector, and value vector, respectively, and d k The dimension of the key vector is represented; the softmax function is used to normalize the relevance scores between different positions, and the final output is a weighted vector sum of the attention-based score values.

[0032] In step 5, the method of inputting the features p3 and p4 obtained in step 3 and the new feature p*5 obtained in step 4 into the FPN neck network for multi-scale feature fusion, and then inputting the three fused features at different scales into the corresponding three detector heads to locate and classify underground diseases is as follows:

[0033] The features p3 and p4 obtained in step 3 and the new feature p*5 obtained in step 4 are input into the FPN neck network. After calculation by the FPN neck network, fused features P3 to P5 with lengths of 56*56 pixels, widths of 28*28 pixels, and widths of 14*14 pixels are obtained. Finally, the fused features P3 to P5 are sent to three detection heads respectively for the detection of large, medium, and small targets, thereby completing the localization and classification of disease targets.

[0034] In step 6, the method for calculating the inter-domain distance between the two features p3 extracted from the source domain dataset and the target domain dataset in step 3 using unsupervised domain adaptation techniques is as follows:

[0035] The inter-domain distance between the two features p3 extracted from the source domain dataset and the target domain dataset in step 3 is calculated using the following formula:

[0036] (2);

[0037] in, Represents source domain data, Represents target domain data. and Let represent samples from the source domain and the target domain, respectively. It is a mapping function, usually a kernel function.

[0038] In step 7, the method of calculating classification loss and regression loss by combining the localization and classification results obtained in step 5 with the actual disease target labels, and then calculating the total loss function of the runway underground structure disease detection model by combining the inter-domain distance obtained in step 6, and then optimizing the model through backpropagation, and iterating until the total loss value converges, thereby determining the final weights of the runway underground structure disease detection model, is as follows:

[0039] The localization and classification results obtained in step 5 are compared with the actual disease target labels to calculate the classification loss and regression loss, respectively, to obtain the classification loss value. and regression loss value Then, combining the inter-domain distance MMD obtained in step 6, the total loss value of the runway underground structure defect detection model is calculated using the following formula:

[0040] (3);

[0041] The hyperparameter λ represents the importance that the model assigns to the differences between the two domains; according to experimental verification, the best results are achieved when the hyperparameter λ is set to 0.05.

[0042] After obtaining the total loss value, the model is optimized through backpropagation. The optimization is iterated until the total loss value converges, thereby determining the final weights of the runway underground structure defect detection model.

[0043] The adaptive automatic detection method for defects in underground airport runway structures provided by this invention has the following beneficial effects:

[0044] 1. By using unsupervised domain adaptation (MMD), the requirement for GPR data labeling in the model can be reduced, thus lowering the time and financial costs in this regard. It enables model training using unlabeled datasets, thereby addressing the problem of low model generalization ability caused by differences in sample distribution and achieving better generalization across different datasets.

[0045] 2. Mining two domain knowledge areas in airport runway underground defect detection: By employing deep row attention and self-attention mechanisms, the correlation between defects and runway underground structures, as well as the dimensional differences between different defects, are integrated into the deep learning algorithm model. This establishes the model's ability to perceive runway layer information and the surrounding environment. On one hand, this allows for the extraction of more effective features, improving the accuracy of identifying airport runway underground defect targets and enabling the system to more accurately and effectively identify target types. On the other hand, the extracted feature information is more common and can adapt to different airport datasets. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the YOLOv5-based runway underground structure defect detection model in this invention.

[0047] Figure 2 This is a schematic diagram of the deep row attention module structure in this invention.

[0048] Figure 3 This is a visual comparison chart of the detection results of different algorithms in this invention. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0050] The adaptive automatic detection method for defects in underground airport runway structures provided by this invention includes the following steps performed in sequence:

[0051] Step 1: Obtain B-scan images of the underground of different airport runways using a robot equipped with a multi-channel ground-penetrating radar, and then divide them into source domain datasets and target domain datasets.

[0052] Using a robot equipped with a multi-channel ground-penetrating radar, the robot collects raw B-scan images of multiple real airport runways along a pre-planned route. The raw B-scan images are then preprocessed sequentially, including zero-bias correction, zero-point adjustment, background removal, data filtering, gain adjustment, and image visualization, to obtain B-scan images. After that, all B-scan images are divided into source domain datasets and target domain datasets.

[0053] This invention collected raw B-scan images of the subsurface of seven real airport runways; the ground-penetrating radar operated at frequencies of 200MHz, 300MHz, and 500MHz. A 14-channel ground-penetrating radar with equal spacing was used, with each channel spaced 0.7m apart. Both the source and target domain datasets included four typical subsurface defects or objects found on airport runways: voids, cracks, reinforcing steel reinforcement, and settlement.

[0054] Step 2: Construct as follows Figure 1 The YOLOv5-based runway underground structure defect detection model shown includes a shared weight, a two-stream backbone network for feature extraction and inter-domain distance calculation, a self-attention module for acquiring global information, an FPN neck network for multi-scale feature fusion, and three detection heads for different feature sizes.

[0055] The backbone network includes a Focus module, an L-Dark2 module, an L-Dark3 module, an L-Dark4 module, and an L-Dark5 module; among which, the L-Dark module includes the C3 module from YOLOv5 and a deep row attention module, used to acquire rich stratigraphic information and establish the correlation between runway subsurface defects and runway subsurface structures; such as Figure 2 As shown, the deep row attention module includes a learnable position encoding module and a row attention module.

[0056] Step 3: Input the source domain dataset and target domain dataset obtained in Step 1 into the runway underground structure defect detection model constructed in Step 2, and use the dual-stream backbone network to extract the features p1 to p5 of the B-scan images in the source domain dataset and target domain dataset respectively;

[0057] First, the B-scan images in the source and target domain datasets are unified to 448*448 pixels. Then, they are input into the backbone networks of the runway underground structure defect detection model. After that, they are passed through the Focus module, L-Dark2 module, L-Dark3 module, L-Dark4 module and L-Dark5 module in sequence to obtain features with lengths and widths of 224*224 pixels, 112*112 pixels, 56*56 pixels, 28*28 pixels and 14*14 pixels, respectively, which are denoted as features p1, features p2, features p3, features p4 and features p5.

[0058] The learningable positional encoding module in the L-Dark module processes information as follows: A one-dimensional positional encoding is embedded into the algorithm model, representing only depth positional information. Specifically, using the row feature vectors in the features as units, a learningable positional encoding method is used to integrate absolute depth positional information into the extracted features; let the feature size be c*h*w, and the size of the learned depth positional information be c*h*1; then, the depth positional information is copied and expanded to a size matching the feature size c*h*w; finally, a simple addition operation is used to directly add it to the features, thus completing the fusion of depth positional information and features.

[0059] The row attention module processes data as follows: Row attention is an axial attention mechanism, essentially a self-attention mechanism. During computation, each row remains independent while mixing information within the row; that is, each row of features is treated as a computational unit, and self-attention is then applied for computation.

[0060] The formula for calculating self-attention is as follows:

[0061] (1);

[0062] In the formula, Q, K, and V represent the query vector, key vector, and value vector, respectively, and d k The dimension of the key vector is represented; the softmax function is used to normalize the relevance scores between different positions, and the final output is a weighted vector sum of the attention-based score values.

[0063] Step 4: Input the feature p5 obtained in Step 3 into the self-attention module to obtain a new feature p*5;

[0064] The deepest feature p5 extracted from the dual-stream backbone network is processed using a self-attention module to obtain a new feature p*5. This enables the runway underground structure defect detection model to acquire global information, strengthens the correlation between defects and the surrounding environment, and thus solves the problem of weak continuity of large defects in B-scan images.

[0065] Step 5: Input the features p3 and p4 obtained in Step 3 and the new feature p*5 obtained in Step 4 into the FPN neck network for multi-scale feature fusion. Then, input the three fused features at different scales into the corresponding three detector heads to locate and classify underground diseases.

[0066] The features p3 and p4 obtained in step 3 and the new feature p*5 obtained in step 4 are input into the FPN neck network. After calculation by the FPN neck network, fused features P3 to P5 with lengths of 56*56 pixels, widths of 28*28 pixels, and widths of 14*14 pixels are obtained. Finally, the fused features P3 to P5 are sent to three detection heads respectively for the detection of large, medium, and small targets, thereby completing the localization and classification of disease targets.

[0067] Step 6: Calculate the inter-domain distance between the two features p3 extracted from the source domain dataset and the target domain dataset in Step 3 using unsupervised domain adaptation (MMD);

[0068] MMD is widely used to measure the distance between two distributed data sets. It is used to find a mapping function that maps both source and target domain data to a reproducing kernel Hilbert space.

[0069] The inter-domain distance between the two features p3 extracted from the source domain dataset and the target domain dataset in step 3 is calculated using the following formula:

[0070] (2);

[0071] in, Represents source domain data, Represents target domain data. and Let represent samples from the source domain and the target domain, respectively. It is a mapping function, usually a kernel function.

[0072] Step 7: Calculate the classification loss and regression loss by combining the localization and classification results obtained in Step 5 with the actual disease target labels. Then, calculate the total loss function of the runway underground structure disease detection model by combining the inter-domain distance obtained in Step 6. Then, optimize the model through backpropagation and iterate until the total loss value converges, thereby determining the final weight of the runway underground structure disease detection model.

[0073] The localization and classification results obtained in step 5 are compared with the actual disease target labels to calculate the classification loss and regression loss, respectively, to obtain the classification loss value. and regression loss value Then, combining the inter-domain distance MMD obtained in step 6, the total loss value of the runway underground structure defect detection model is calculated using the following formula:

[0074] (3);

[0075] The hyperparameter λ represents the importance that the model assigns to the differences between the two domains; according to experimental verification, the best results are achieved when the hyperparameter λ is set to 0.05.

[0076] After obtaining the total loss value, the model is optimized through backpropagation. The optimization is iterated until the total loss value converges, thereby determining the final weights of the runway underground structure defect detection model.

[0077] Step 8: After the B-scan image of the airport runway underground structure to be detected is preprocessed as described in Step 1, it is input into the runway underground structure defect detection model whose final weights have been determined in Step 7. The runway underground structure defect detection model outputs the final location and classification results of the defect targets.

[0078] To verify the effectiveness of the present invention, the inventors conducted the following experiments:

[0079] (1) Selection of training parameters for the runway underground structure defect detection model: The initial learning rate is set to 0.01 (it will change during training, but the minimum will not be less than 0.0001). The momentum is set to 0.937 to optimize the gradient descent method. The batch size of the source domain and the target domain is set to 8. During training, the batch data of the two different domains are passed together through a dual-stream backbone network with shared weights.

[0080] (2) Evaluation index description: This invention uses five evaluation indices widely used in the field of target detection: precision, recall, F1-measure (F1), AP, and mAP (mean average precision). The formulas for each are as follows:

[0081] ;

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] This invention sets the Intersection over Union (IoU) ratio to 0.5. When the overlap ratio between the predicted bounding box and the labeled bounding box is greater than the IoU ratio, it is counted as a true positive (TP); otherwise, it is a false positive (FP). By adding the false negative (FN) ratio, the specific values ​​of these five evaluation indicators can be calculated.

[0087] (3) Comparison method description: The method of the present invention will be compared with the following four algorithms in terms of performance.

[0088] ① The classic anchor-free algorithm Centernet. See the reference: Duan K, Bai S, Xie L, et al. Centernet: Keypoint triplets for object detection[C] / / Proceedings of the IEEE / CVF international conference on computer vision. 2019: 6569-6578.

[0089] ② Swin transformer based on hierarchical design and shifted windows. Reference: Liu Z, Lin Y, Cao Y, et al. Swin transformer: Hierarchical vision transformer using shifted windows[C] / / Proceedings of the IEEE / CVF international conference on computer vision. 2021: 10012-10022.

[0090] ③ The two latest models of the most popular one-stage YOLO series algorithms are YOLOv8 and YOLOv9. References are: Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics YOLO, January 2023. and Wang CY, Yeh IH, Liao HY M. YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information[J]. arXiv preprint arXiv:2402.13616,2024.

[0091] The method of this invention achieves F1 scores of 24%, 30%, 38%, and 72% for four types of underground defects or objects (cracks, voids, settlement, and reinforcing steel), and mAP of 38%, respectively, all surpassing the other four mainstream and latest algorithms. The visualization results of each model are shown below. Figure 3 As shown. Application to real-world airport data demonstrates that the method of this invention can effectively detect these four types of defects or objects.

Claims

1. An adaptive automatic detection method for defects in underground structures of airport runways, characterized in that: The adaptive automatic detection method for defects in underground structures of airport runways includes the following steps performed in sequence: Step 1: Obtain B-scan images of the underground of different airport runways using a robot equipped with a multi-channel ground-penetrating radar, and then divide them into source domain datasets and target domain datasets. Step 2: Construct a YOLOv5-based model for detecting defects in underground runway structures. This model includes a shared weight, a two-stream backbone network for feature extraction and inter-domain distance calculation, a self-attention module for acquiring global information, an FPN neck network for multi-scale feature fusion, and three detection heads for different feature sizes. Step 3: Input the source domain dataset and target domain dataset obtained in Step 1 into the runway underground structure defect detection model constructed in Step 2, and use the dual-stream backbone network to extract the features p1 to p5 of the B-scan images in the source domain dataset and target domain dataset respectively; Step 4: Input the feature p5 obtained in Step 3 into the self-attention module to obtain a new feature p*5; Step 5: Input the features p3 and p4 obtained in Step 3 and the new feature p*5 obtained in Step 4 into the FPN neck network for multi-scale feature fusion. Then, input the three fused features at different scales into the corresponding three detector heads to locate and classify underground diseases. Step 6: Calculate the inter-domain distance between the two features p3 extracted from the source domain dataset and the target domain dataset in Step 3 using unsupervised domain adaptation techniques; Step 7: Calculate the classification loss and regression loss by combining the localization and classification results obtained in Step 5 with the actual disease target labels. Then, calculate the total loss function of the runway underground structure disease detection model by combining the inter-domain distance obtained in Step 6. Then, optimize the model through backpropagation and iterate until the total loss value converges, thereby determining the final weight of the runway underground structure disease detection model. Step 8: After the B-scan image of the airport runway underground structure to be detected is preprocessed as described in Step 1, it is input into the runway underground structure defect detection model whose final weights have been determined in Step 7. The runway underground structure defect detection model outputs the final location and classification results of the defect targets.

2. The adaptive automatic detection method for defects in underground airport runway structures according to claim 1, characterized in that: In step 1, the method of obtaining B-scan images of the underside of different airport runways using a robot equipped with a multi-channel ground-penetrating radar, and then dividing them into source domain datasets and target domain datasets, is as follows: Using a robot equipped with a multi-channel ground-penetrating radar, multiple raw B-scan images of the underground of real airport runways are collected according to a pre-planned route. The raw B-scan images are then preprocessed sequentially, including zero-bias correction, zero-point adjustment, background removal, data filtering, gain adjustment, and image visualization, to obtain B-scan images. All B-scan images are then divided into source domain datasets and target domain datasets. Both the source domain dataset and the target domain dataset include four typical underground defects or objects of airport runways: voids, cracks, steel reinforcement, and settlement.

3. The adaptive automatic detection method for defects in underground airport runway structures according to claim 1, characterized in that: In step 2, the backbone network includes a Focus module, an L-Dark2 module, an L-Dark3 module, an L-Dark4 module, and an L-Dark5 module; wherein the L-Dark module includes the C3 module in YOLOv5 and a deep row attention module, used to acquire rich strata information and establish the correlation between runway underground defects and runway underground structures; the deep row attention module includes a learnable position encoding module and a row attention module.

4. The adaptive automatic detection method for defects in underground airport runway structures according to claim 1, characterized in that: In step 3, the method of inputting the source domain dataset and target domain dataset obtained in step 1 into the runway underground structure defect detection model constructed in step 2, and extracting features p1 to p5 of the B-scan images in the source domain dataset and target domain dataset respectively using a dual-stream backbone network is as follows: First, the B-scan images in the source and target domain datasets are unified to 448*448 pixels. Then, they are input into the backbone networks of the runway underground structure defect detection model. After that, they are passed through the Focus module, L-Dark2 module, L-Dark3 module, L-Dark4 module and L-Dark5 module in sequence to obtain features with lengths and widths of 224*224 pixels, 112*112 pixels, 56*56 pixels, 28*28 pixels and 14*14 pixels, respectively, which are denoted as features p1, features p2, features p3, features p4 and features p5. The learningable position encoding module in the L-Dark module is processed as follows: a one-dimensional position encoding is embedded in the algorithm model, which only represents depth position information; specifically, the absolute depth position information is integrated into the extracted features using the row feature vectors in the features and a learningable position encoding method is used. set up The feature size is c*h*w, and the learned depth location information is c*h*1. Then, the depth location information is copied and expanded to a size that matches the feature size c*h*w. Finally, it is added directly to the feature using a simple addition operation, thus completing the fusion of depth location information and feature. The row attention module processes the data as follows: During computation, each row is kept independent while information within the row is mixed. That is, each row of features is used as a computational unit, and then self-attention is applied for computation. The formula for calculating self-attention is as follows: （1）； In the formula, Q, K, and V represent the query vector, key vector, and value vector, respectively, and d k The dimension of the key vector is represented; the softmax function is used to normalize the relevance scores between different positions, and the final output is a weighted vector sum of the attention-based score values.

5. The adaptive automatic detection method for defects in underground airport runway structures according to claim 1, characterized in that: In step 5, the method of inputting the features p3 and p4 obtained in step 3 and the new feature p*5 obtained in step 4 into the FPN neck network for multi-scale feature fusion, and then inputting the three fused features at different scales into the corresponding three detector heads to locate and classify underground diseases is as follows: The features p3 and p4 obtained in step 3 and the new feature p*5 obtained in step 4 are input into the FPN neck network. After calculation by the FPN neck network, fused features P3 to P5 with lengths of 56*56 pixels, widths of 28*28 pixels, and widths of 14*14 pixels are obtained. Finally, the fused features P3 to P5 are sent to three detection heads respectively for the detection of large, medium, and small targets, thereby completing the localization and classification of disease targets.

6. The adaptive automatic detection method for defects in underground airport runway structures according to claim 1, characterized in that: In step 6, the method for calculating the inter-domain distance between the two features p3 extracted from the source domain dataset and the target domain dataset in step 3 using unsupervised domain adaptation techniques is as follows: The inter-domain distance between the two features p3 extracted from the source domain dataset and the target domain dataset in step 3 is calculated using the following formula: （2）； in, Represents source domain data, Represents target domain data. and Let represent samples from the source domain and the target domain, respectively. It is a mapping function.

7. The adaptive automatic detection method for defects in underground airport runway structures according to claim 1, characterized in that: In step 7, the method of calculating classification loss and regression loss by combining the localization and classification results obtained in step 5 with the actual disease target labels, and then calculating the total loss function of the runway underground structure disease detection model by combining the inter-domain distance obtained in step 6, and then optimizing the model through backpropagation, and iterating until the total loss value converges, thereby determining the final weights of the runway underground structure disease detection model, is as follows: The localization and classification results obtained in step 5 are compared with the actual disease target labels to calculate the classification loss and regression loss, respectively, to obtain the classification loss value. and regression loss value Then, combining the inter-domain distance MMD obtained in step 6, the total loss value of the runway underground structure defect detection model is calculated using the following formula: （3）； Here, the hyperparameter λ represents the importance that the model assigns to the differences between the two domains; according to experimental verification, the hyperparameter λ is set to 0.05; After obtaining the total loss value, the model is optimized through backpropagation. The optimization is iterated until the total loss value converges, thereby determining the final weights of the runway underground structure defect detection model.

Images