Road underground disease detection model based on medium inversion and deep learning

Abstract

The invention provides a road underground disease detection model based on medium inversion and deep learning. A generation method of the model comprises the following steps: 1) collecting road underground radar data by using a ground penetrating radar; 2) obtaining the distribution condition of the underground dielectric constant corresponding to the B-Scan, and constructing a B-Scan-dielectric constant distribution data set; 3) constructing a deep learning medium dielectric property inversion model, and performing training by optimizing a comprehensive loss function; 4) utilizing the trained dielectric dielectric property inversion model to perform inversion on the actually measured radar map to generate an underground dielectric property distribution map; 5) performing disease classification marking according to the underground dielectric property distribution diagram, and constructing and training a disease target identification model; and 6) inputting the to-be-detected dielectric property distribution data into the disease target identification model, and outputting a prediction result of the underground hidden disease. Compared with the prior art, the method provided by the invention can realize accurate disease identification and positioning under complex underground medium distribution, and has better accuracy advantage and stronger generalization ability.

Description

Technical Field

[0001] This invention relates to the field of ground-penetrating radar data processing and underground disease detection technology, and in particular to a road underground disease detection model based on media inversion and deep learning. Background Technology

[0002] With the continuous increase in road traffic volume, the safety and stability of road structures have become particularly important. Especially in complex geological environments, underground road structures may suffer from various defects due to natural factors and long-term use. These defects, such as cavities, cracks, delamination, leakage, and loose concrete, are often difficult to identify using conventional testing methods. To ensure the safe operation of roads, efficient and accurate non-destructive testing techniques must be employed to assess the health status of underground structures.

[0003] Ground-penetrating radar (GPR) technology is widely used in the detection of underground road defects because it can penetrate deep underground and generate images reflecting the underground structure. However, the application of GPR also faces some challenges. First, due to the complex morphology and uneven distribution of underground defects, traditional GPR image processing methods often struggle to accurately extract defect information from complex background noise. Second, highly reflective materials (such as concrete and steel reinforcement) may obscure defect signals, affecting the accurate identification of underground structural defects. Current GPR data processing methods, including common inversion techniques such as ray tracing, reverse time migration (RTM), and full waveform inversion (FWI), while providing some solutions, still fall short of achieving ideal results in practical applications, especially in the detection of complex underground defects, due to their high computational complexity and dependence on data quality.

[0004] In recent years, deep neural networks (DNNs), as an advanced machine learning technique, have made significant progress in fields such as computer vision and image processing. Especially in tasks like object detection and image segmentation, DNNs have demonstrated exceptional capabilities. However, existing deep learning methods are mostly used for object detection or image classification, and have not yet been effectively applied to the inversion of complex GPR data. In particular, GPR images have unique spatial characteristics, where the relationship between signal reflection and reception is relatively fixed, while the effective signal is concentrated near the lesion area, and signal propagation is affected by various factors. Existing deep learning methods are mainly designed for lesions with regular geometric shapes, making them unsuitable for inversion tasks involving complex lesions.

[0005] Currently, deep learning-based GPR data processing methods mainly focus on target detection and signal classification. While these studies have advanced the application of GPR data to some extent, many shortcomings remain. For example, existing models often neglect the nonlinear relationships between different defects when dealing with underground defects in complex geological structures, making accurate defect inversion still quite difficult. Therefore, developing a deep neural network optimized for the characteristics of GPR data, capable of accurately inverting the dielectric properties of complex underground defects, is crucial for improving the accuracy and reliability of road underground structure detection.

[0006] In summary, although existing GPR data processing methods can provide detection results for underground defects to some extent, their limitations prevent them from meeting the high-precision requirements of practical applications. Therefore, there is an urgent need for a road underground defect detection model based on media inversion and deep learning that can accurately identify complex underground defects in roads, in order to overcome the shortcomings of existing technologies. Summary of the Invention

[0007] The purpose of this invention is to provide a method for detecting underground road defects that can accurately invert the dielectric properties of complex underground defects and improve the accuracy of underground road defect detection.

[0008] To achieve the above objectives, this invention proposes a road underground defect detection model based on media inversion and deep learning. The method for generating the road underground defect detection model includes the following steps:

[0009] S1: Use ground-penetrating radar (GPR) to collect underground radar data of roads and establish a dataset of underground defects; the radar data includes B-Scan image data.

[0010] S2: By combining indoor sandbox experiments and numerical simulation experiments, the distribution of underground dielectric constant corresponding to B-Scan was obtained, and a B-scan dielectric constant distribution dataset was constructed.

[0011] S3: Construct a deep learning dielectric property inversion model, using the B-scan dielectric constant distribution dataset as input, and train it by optimizing the comprehensive loss function to establish the mapping relationship between radar B-scan data and the dielectric constant distribution of underground media;

[0012] S4: Use the trained deep learning dielectric property inversion model to invert the measured radar spectrum and generate a distribution map of underground dielectric properties.

[0013] S5: Classify and label diseases based on the underground dielectric property distribution map, add category labels, obtain the dielectric constant distribution map dataset corresponding to different diseases, and build and train the disease target recognition model.

[0014] S6: Input the dielectric property distribution data to be detected into the disease target identification model, and output the prediction results of underground hidden diseases.

[0015] Furthermore, in step S1, the underground defects include cavities, looseness, uneven settlement, water-rich defects, non-compactness, and pipeline defects; the radar includes two-dimensional or three-dimensional ground-penetrating radar, vehicle-mounted or handheld ground-penetrating radar, as well as multi-frequency and fully polarized radar; the radar operates in a frequency range of 0.1 GHz to 3.0 GHz, and adopts vertical, horizontal, or dual-polarization working modes to adapt to different detection scenarios and underground media.

[0016] Further, step S2 specifically involves: in the indoor sand river experiment, constructing experimental scenarios with different underground medium characteristics to simulate the underground environment of roads, and acquiring B-Scan images through GPR equipment; based on the experimental data, using numerical simulation methods to model different underground dielectric properties and generate corresponding dielectric constant distribution maps; constructing a corresponding dataset of B-Scan and underground dielectric constant distribution through experimental data and simulation data to obtain the B-scan-dielectric constant distribution dataset, providing reliable basic data for training the deep learning inversion model.

[0017] Furthermore, in step S3, the B-Scan image is standardized and its size is unified by the preprocessing module, and then used as the input data for the inversion model to ensure data consistency.

[0018] In the deep learning dielectric property inversion model, the encoder adopts a "Trace-to-Trace Encoder" architecture, consisting of 5 convolutional layers and 5 fully connected layers. It extracts and condenses the information of each GPR trajectory, maintaining the spatial alignment between the B-Scan image and the dielectric constant map. The decoder maps the feature map to the same size as the dielectric constant map through 4×4 upper convolutional layers, 3×3 convolutional layers, and upsampling operations. It also compresses feature channels through convolutional layers to prevent overfitting. A composite loss function is designed, combining L2 norm and multi-scale structural similarity (MSSIM) to minimize error and maximize structural similarity, thereby optimizing the inversion results. The model parameters are optimized by minimizing the loss function to complete training, ultimately obtaining the trained dielectric inversion model, which is used to invert the dielectric constant distribution.

[0019] Furthermore, the formula for calculating the loss function is as follows:

[0020]

[0021] Among them, P i and Let x((h,w),r) and y((h,w),r) represent the inversion result and the true value of the i-th pair of data, respectively; x((h,w),r) and y((h,w),r) are two corresponding windows of size r centered at (h,w), where h∈[1,H] and w∈[1,W]; R is the total number of scales; λ r The weights are for scale r.

[0022] Furthermore, in step S3, the deep learning dielectric property inversion model includes inversion models based on waveform propagation characteristics, typically using full waveform inversion and reverse time offset; inversion models based on time-frequency domain signal processing, typically using short-time Fourier transform or wavelet transform; and inversion models based on deep learning, typically using convolutional neural networks (CNN), recurrent neural networks (RNN), and generative adversarial networks (GAN).

[0023] Furthermore, in step S5, the method for classifying and labeling diseases based on the underground dielectric property distribution map is as follows: Based on the output of the underground dielectric property distribution map, and combined with known disease types and locations, labels are created to form a training dataset containing different types of underground diseases. Each dielectric constant distribution map corresponds to one or more disease target areas, and these target areas need to be classified and labeled according to the changes in dielectric constant.

[0024] Furthermore, a disease target recognition model is constructed and trained using a dielectric constant distribution map dataset as input to establish the relationship between the dielectric constant distribution map and the classification and recognition of underground road diseases. The disease target recognition model uses the YOLOv5s target detection framework as its base model to enhance its ability to recognize underground disease images. In the YOLOv5s backbone network, the Dense-C3 module is used to process feature information from complex backgrounds. The CBAM attention module is used to enable the model to adaptively focus on regions containing key information, thereby improving detection accuracy. During the training process of the disease target recognition model, a multi-task loss function is used, which includes location loss, confidence loss, and classification loss. Hyperparameters (such as learning rate, training epochs, batch size, etc.) are adjusted to maximize detection performance. The model's effectiveness is evaluated using a validation set, and optimization is performed based on the model's precision and recall.

[0025] Furthermore, the location loss adopts GIoU loss, and the confidence loss and classification loss adopt the Focal Loss function to alleviate sample imbalance.

[0026] Furthermore, in step S5, the disease target identification model includes models based on commonly used target detection algorithms such as CNN, Faster-RCNN, YOLO, SSD, or GAN.

[0027] Furthermore, in step S6, the new underground dielectric property distribution map data is input into the trained disease identification model. The input data should undergo the same preprocessing steps as the training data to ensure data consistency. Based on the features learned during training, the disease identification model will identify potential underground disease areas in the image, and the output prediction results will include the category, location, and prediction confidence value of each detected area.

[0028] Furthermore, based on the aforementioned road underground disease detection model based on medium inversion and deep learning, the method for detecting road underground diseases is as follows: 1) Obtain the radar signal data to be processed; 2) Use the trained deep learning medium dielectric property inversion model to invert the radar signal data to be processed into a dielectric constant distribution map; 3) Input the dielectric constant distribution map into the trained disease target recognition model and output the prediction results of underground hidden diseases, thereby realizing the target detection and classification of road underground diseases.

[0029] Compared with the prior art, the advantages of the present invention are:

[0030] 1. This invention comprehensively utilizes the advantages of ground-penetrating radar (GPR) data inversion technology and deep neural networks to effectively improve the accuracy and reliability of road underground disease detection. Through the integration of deep learning algorithms and media inversion technology, it can not only accurately identify the distribution of complex diseases, but also fully preserve the detailed features of underground structures without relying on traditional complex inversion algorithms, thereby providing accurate data support for the maintenance and repair of road structures.

[0031] 2. This invention creates a deep learning medium inversion model to efficiently extract complex features from radar signals and accurately reconstruct the dielectric properties of underground road defects. By optimizing the training process, the model can accurately invert underground structural defects under complex geological conditions, overcoming the limitations of traditional inversion methods in processing complex defect signals and significantly improving the accuracy and reliability of underground defect inversion.

[0032] 3. This invention achieves high-precision disease target detection by constructing a dataset based on dielectric constant distribution maps and combining it with deep learning algorithms. This method can not only effectively identify common underground disease types, but also handle diseases with complex underground structures and irregular geometric shapes, providing more reliable technical support for the classification and identification of underground diseases.

[0033] 4. This invention demonstrates good generalization ability and can be successfully applied to the processing of actual radar data; by introducing training data with different background noise and multiple underground links, the model can adapt to the detection of different types of underground road defects, improving its stability and practicality in practical applications. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating the generation process of a road underground disease detection model based on media inversion and deep learning, according to an embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram of the overall architecture of the GPRInvNet model for inverting the dielectric properties of a deep learning medium, constructed in the road underground disease detection method of this invention.

[0036] Figure 3 This is a schematic diagram of the overall architecture of the YOLO model for identifying underground diseases constructed in the road underground disease detection method of this invention.

[0037] Figure 4 This is a schematic diagram comparing the original C3 module and the dense-C3 module of the YOLO model for identifying underground diseases in an embodiment of the present invention.

[0038] Figure 5 This is a comparison of B-scan images, actual conditions, and results of inversion identification in the road underground disease detection method of this invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described below.

[0040] This embodiment proposes a road underground disease detection model based on media inversion and deep learning, such as... Figure 1 As shown, its generation method includes the following steps:

[0041] S1: Use ground-penetrating radar (GPR) to scan the underground area of ​​the target road and collect underground radar data, mainly including underground B-Scan image data of the target road;

[0042] Underground defects include cavities, looseness, uneven settlement, water-rich defects, non-compactness, and pipeline defects; radar includes two-dimensional or three-dimensional ground-penetrating radar, vehicle-mounted or handheld ground-penetrating radar, as well as multi-frequency and fully polarized radar; the radar operates in a frequency range of 0.1 GHz to 3.0 GHz, and adopts vertical, horizontal, or dual-polarization working modes, which can be flexibly adjusted according to the characteristics of different underground media.

[0043] S2: Training the medium inversion model requires a dataset. The dataset is obtained from two sources, namely, sufficient underground B-Scan images and corresponding underground dielectric constant distribution maps, through a combination of indoor sandbox experiments and numerical simulations. This B-scan image-dielectric constant distribution dataset is then constructed as the training set for the medium inversion model in step S3.

[0044] Relying solely on indoor sandbox experiments to collect training data is too time-consuming and costly. Therefore, the dataset is provided partly by indoor sandbox experiments and partly by numerical simulations. In the indoor sandbox experiments, underground road environments are simulated by constructing underground media scenarios with different dielectric constant distributions, and electromagnetic wave reflection signals are collected using GPR equipment to obtain B-scan images of the indoor models and their corresponding dielectric constant distributions. Numerical simulations are used to construct electromagnetic response models for different underground media, obtaining B-scan images and their corresponding dielectric constant distributions under numerical simulation. The two sources are combined to construct the B-scan-dielectric constant distribution dataset.

[0045] S3: Construct a deep learning model for inverting the dielectric properties of a medium. Use the B-scan-dielectric constant distribution dataset constructed in S2 as the training set. Train the model by optimizing the comprehensive loss function and establish the mapping relationship between radar B-scan data and the dielectric constant distribution of underground media.

[0046] The preprocessing module standardizes and unifies the size of the B-Scan images to ensure they can be adapted to subsequent deep learning models. After standardization and size unification, the processed GPR images are used as input data for the network and fed into the subsequent feature extraction and inversion models.

[0047] In deep learning dielectric property inversion models, such as Figure 2 As shown, the model's input is the B-Scan image D. i ,in and For B-Scan image D i Two adjacent GPR trajectories are shown. A GPR trajectory refers to the time-varying electromagnetic wave amplitude data recorded by a ground-penetrating radar (GPR) device after emitting electromagnetic waves at a fixed location and receiving the reflected signals. The encoder adopts a "Trace-to-Trace Encoder" architecture, consisting of 5 convolutional layers and 5 fully connected layers. It uses five 5×5 convolutional kernels to generate data from D... i Extracting feature F i ,in and It is F i The features of the (r-1)th and rth columns are spatially related to and Alignment; then, five fully connected layers are used from F i The features obtained by medium compression are G i , and It is G i Two adjacent features in the middle, by and The time dimension (time step) is compressed from T to C; information of each GPR trajectory is extracted and condensed while maintaining spatial alignment between the B-Scan image and the dielectric constant map; the decoder maps the feature map to the same size as the dielectric constant map through 4×4 upper convolutional layers, 3×3 convolutional layers, and upsampling operations, and compresses feature channels through convolutional layers to prevent overfitting. The final model output is the dielectric constant distribution map P. i , among which and It is P i The features of the (r-1)th and rth columns are spatially related to and Alignment; design a composite loss function that combines L2 norm and multi-scale structural similarity (MSSIM) to minimize error and maximize structural similarity, thereby optimizing the inversion results; optimize model parameters by minimizing the loss function to complete training, and finally obtain a trained medium inversion model for inverting the dielectric constant distribution.

[0048] In this embodiment, the input data is GPR B-Scan. First, the feature map is obtained through five convolutional layers, each using a 5×5 kernel with a stride of 1. Generated feature maps With input GPRB-Scan They have the same spatial dimensions, but the features at each location also include neighborhood information. In this embodiment, Let E represent the feature map obtained from the i-th GPR B-Scan encoding, where E represents the number of feature channels. For each GPR trajectory data... With dimensions [T, 1], the convolutional layer transforms the data into a feature vector. The dimensions are [T, E].

[0049] Next, for each encoded GPR trajectory Five fully connected layers were used to fuse the temporal features of each trajectory and combine them into a feature map with [C,E] dimensions.

[0050] Furthermore, each fully connected layer includes activation and batch normalization operations. In this way, new feature maps are generated. Its dimension is proportional to that of the dielectric constant map. The r-th column vector of the new feature map is represented as... in The size is C×E.

[0051] Then, each The space is forced to align to a column of the dielectric constant map to be inverted. Finally, in the encoder, a map corresponding to the dielectric constant map to be inverted is generated. Feature maps with the same size ratio More importantly, the characteristics of each GPR survey line Sensitive regions in spatial distribution and dielectric constant diagram Alignment, where r*∈[1,W].

[0052] Modify the decoder parameters to adapt to the "Trace-to-Trace" encoder. The decoder consists of one 4×4 upconvolutional layer, six 3×3 convolutional layers, and one upsampling operation.

[0053] First, a 4×4 upconvolutional layer with a stride of 2 is used to enlarge the size of the feature map. Next, a 3×3 convolutional layer with a stride of 1 is used to stabilize the information. Then, an upsampling operation is used to generate a feature map of the same size as the dielectric constant image. Finally, four 3×3 convolutional kernels with a stride of 1 are added to compress the size of the feature channels. To avoid overfitting and improve the robustness of the network, a dropout method is also used to randomly discard some feature maps.

[0054] Each feature map is of size C×E Used to accurately invert a dielectric constant model fragment of size H×1 By all By stitching the images together, this invention successfully reconstructed the entire dielectric constant image P. i .

[0055] Regarding the loss function, this embodiment employs a combination of L2 norm and multi-scale structural similarity (MSSIM) to minimize the error between the input and output images. The formula for calculating the loss function is as follows:

[0056]

[0057] Among them, P i and Let x((h,w),r) and y((h,w),r) represent the inversion result and the true value of the i-th pair of data, respectively; x((h,w),r) and y((h,w),r) are two corresponding windows of size r centered at (h,w), where h∈[1,H] and w∈[1,W]; R is the total number of scales; λ r The weights are for scale r. By simultaneously minimizing the norm metric and maximizing MSSIM, the model is optimized for structural similarity and pixel-wise error rate of the output image, ultimately yielding a trained dielectric inversion model capable of accurately inverting the dielectric constant distribution.

[0058] S4: Use the trained deep learning dielectric property inversion model to invert the measured radar spectrum and generate a distribution map of underground dielectric properties.

[0059] The GPRInvNet deep learning model, trained with a large amount of underground disease data, can effectively extract features from GPR images and invert the dielectric constant distribution of underground structures. In this step, B-Scan images are used as input, processed by the model's encoder and decoder to finally generate the corresponding dielectric constant image.

[0060] S5: Classify and label diseases based on the underground dielectric property distribution map, add category labels, obtain the dielectric constant distribution map dataset corresponding to different diseases, and build and train the disease target recognition model.

[0061] In this embodiment, the method for classifying and labeling diseases based on the underground dielectric property distribution map is as follows: The output of the underground dielectric property distribution map is combined with known disease types and locations for labeling, forming a training dataset containing different types of underground diseases. Each dielectric constant distribution map corresponds to one or more disease target areas, and these target areas need to be classified and labeled according to the changes in dielectric constant.

[0062] In the process of constructing the disease target identification model, the dielectric constant distribution map dataset is used as input for training to establish the relationship between the dielectric constant distribution map and the classification and identification of road underground diseases.

[0063] In this embodiment, as Figure 3 As shown, this disease target recognition model uses the YOLOv5s target detection framework as its base model to enhance the model's ability to recognize underground disease images, as detailed below:

[0064] 1) In the YOLOv5s backbone network, the Dense-C3 module is used to replace the traditional original C3 module, such as... Figure 4 As shown, the Dense-C3 module enhances feature reuse through dense connections, helping to capture complex disease features in underground dielectric property images. DenseNet is a convolutional neural network composed of multiple dense blocks, employing a dense connection structure to enhance feature reuse. In a traditional L-layer convolutional network, there are typically L connections; however, in the dense connection structure within a dense block, there are L*(L+1) / 2 connections. Dense blocks facilitate feature reuse in the channel dimension and help alleviate the gradient vanishing problem. In the backbone, following the DenseNet concept, dense blocks are constructed by integrating two Conv modules to the bottleneck of the C3 module, establishing dense connections; the densely connected C3 module is also called the Dense-C3 module.

[0065] In the original bottleneck, the output of the Lth layer is defined as the sum of the nonlinear transformations of the output of the previous layer and the output of the previous layer itself:

[0066] xL =F L (x L-1 )+x L-1

[0067] Where, x L F represents the output of the Lth layer. L It is a nonlinear transformation function.

[0068] In the Dense-C3 module, the original bottleneck is replaced by a dense block. The dense block consists of four densely connected Conv modules, and the output of the Lth layer comes from a nonlinear operation performed by concatenating the output feature maps of all previous layers, as shown in the following equation:

[0069] x L =F L ([x0,x1,...,x L-1 ])

[0070] Where [x0, x1, ..., xL-1] represents the output combination of the previous L-1 layer.

[0071] 2) Applying the Convolutional Block Attention Module (CBAM) to the backbone and adding it after each Dense-C3 module allows the model to adaptively focus on important disease areas, improving detection accuracy. CBAM is a typical hybrid attention mechanism module, including a channel attention module (CAM) and a spatial attention module (SAM). The channel attention matrix is ​​calculated as follows:

[0072]

[0073] In the formula, σ represents the sigmoid activation function, and W0 and W1 represent the two layers of parameters of the MLP.

[0074] The spatial attention matrix is ​​calculated as follows:

[0075]

[0076] In the formula, f 7×7 This represents a 7×7 convolutional layer.

[0077] 3) During the training of the disease target identification model, a focal loss function is introduced to address the sample imbalance problem; multi-task loss functions (including location loss, confidence loss, and classification loss) are used for model training. Specifically, the location loss uses GIoU loss, and the confidence loss and classification loss use cross-entropy (CE) loss functions to alleviate sample imbalance.

[0078] To address the imbalanced sample problem, a focal loss function is introduced. Focal Loss assigns greater weight to hard-to-classify samples during training, improving the identification of difficult-to-classify diseases. The confidence loss is modified as follows:

[0079]

[0080] The model was trained using labeled underground dielectric property image data. A conventional backpropagation algorithm was employed for optimization, and the Adam optimizer was used for hyperparameter tuning, adjusting hyperparameters (such as learning rate, training epochs, batch size, etc.) to maximize detection performance. The model was evaluated using metrics such as accuracy, recall, and F1 score, ultimately yielding an underground hazard target recognition model based on the distribution of dielectric properties.

[0081] S6: Input the dielectric property distribution data to be detected into the disease target identification model, and output the prediction results of underground hidden diseases; specifically: input the new underground dielectric property distribution map data into the trained disease identification model. The input data should undergo the same preprocessing steps as the training data to ensure data consistency; the disease identification model will identify the possible underground disease areas in the image based on the features learned during training, and the output prediction results include the category, location, and prediction confidence value of each detected area.

[0082] Based on the aforementioned road underground disease detection model based on medium inversion and deep learning, the method for detecting underground diseases in target roads is as follows: 1) Scan the underground area of ​​the target road using ground-penetrating radar equipment to obtain radar signal data to be processed, i.e., B-Scan scan images; 2) Use a trained deep learning medium dielectric property inversion model to invert the radar signal data to be processed into a dielectric constant distribution map; 3) Input the dielectric constant distribution map into a trained disease target recognition model to output the prediction results of hidden underground diseases, such as... Figure 5 As shown in the figure, a comparison diagram of the B-Scan scan image, the actual underground conditions of the target road, and the identification results after inversion is presented. Figure 5 As shown, the first column of images is a B-Scan; the second column is the actual situation; and the third column is the result image after medium inversion and defect target identification. The identification results are: (a) underground has non-compact and layered defects; (b) underground contains reinforcing steel; (c) there is a waterless defect below the underground reinforcing steel; (d) there is a water-bearing crack below the underground reinforcing steel; and (e) there is a water-bearing non-compact defect below the underground reinforcing steel. Figure 5 The comparison results show that the recognition results obtained by the method of the present invention are consistent with the actual situation and the recognition effect is excellent.

[0083] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A road subsurface defect detection model based on media inversion and deep learning, characterized in that, The method for generating the road underground defect detection model includes the following steps: S1: Use ground-penetrating radar (GPR) to collect underground radar data of roads and establish a dataset of underground defects; the radar data includes B-Scan image data. S2: By combining indoor sandbox experiments and numerical simulation experiments, the distribution of underground dielectric constant corresponding to B-Scan was obtained, and a B-scan dielectric constant distribution dataset was constructed. S3: Construct a deep learning dielectric property inversion model, using the B-scan dielectric constant distribution dataset as input, and train it by optimizing the comprehensive loss function to establish the mapping relationship between radar B-scan data and the dielectric constant distribution of underground media; S4: Use the trained deep learning dielectric property inversion model to invert the measured radar spectrum and generate a distribution map of underground dielectric properties. S5: Classify and label diseases based on the underground dielectric property distribution map, and build and train a disease target recognition model; S6: Input the dielectric property distribution data to be detected into the disease target identification model, and output the prediction results of underground hidden diseases.

2. The road subsurface defect detection model based on media inversion and deep learning according to claim 1, characterized in that, In step S1, the underground defects include cavities, looseness, uneven settlement, water-rich defects, non-compactness, and pipeline defects; the radar includes two-dimensional or three-dimensional ground penetrating radar, vehicle-mounted or handheld ground penetrating radar, as well as multi-frequency and fully polarized radar; the radar operates in a frequency range of 0.1 GHz to 3.0 GHz, and adopts vertical, horizontal, or dual-polarization working modes to adapt to different detection scenarios and underground media.

3. The road subsurface defect detection model based on media inversion and deep learning according to claim 1, characterized in that, Step S2 specifically involves: in the indoor sandy river experiment, constructing experimental scenarios with different underground medium characteristics to simulate the underground environment of a road, and acquiring B-Scan images through a GPR device; based on the experimental data, using numerical simulation methods to model different underground dielectric properties and generate corresponding dielectric constant distribution maps; constructing a corresponding dataset of B-Scan and underground dielectric constant distribution through experimental data and simulation data to obtain the B-scan-dielectric constant distribution dataset.

4. The road subsurface defect detection model based on media inversion and deep learning according to claim 1, characterized in that, In step S3, the B-Scan images are standardized and resized by the preprocessing module and then used as input data for the model. In the deep learning dielectric property inversion model, the encoder adopts a "Trace-to-Trace Encoder" architecture, consisting of 5 convolutional layers and 5 fully connected layers, which extracts and condenses the information of each GPR trajectory, maintaining the spatial alignment between the B-Scan image and the dielectric constant map. The decoder maps the feature map to the same size as the dielectric constant map through 4×4 upconvolutional layers, 3×3 convolutional layers, and upsampling operations, and compresses the feature channels through convolutional layers; A composite loss function is designed, combining L2 norm and multi-scale structural similarity, to minimize the error and maximize structural similarity, thereby optimizing the inversion results. The model parameters are optimized by minimizing the loss function to complete the training, and finally a trained medium inversion model is obtained, which is used to invert the dielectric constant distribution.

5. The road subsurface defect detection model based on media inversion and deep learning according to claim 4, characterized in that, The formula for calculating the loss function is as follows: Among them, P i and Let x((h,w),r) and y((h,w),r) represent the inversion result and the true value of the i-th pair of data, respectively; x((h,w),r) and y((h,w),r) are two corresponding windows of size r centered at (h,w), where h∈[1,H] and w∈[1,W]; R is the total number of scales; λ r The weights are for scale r.

6. The road subsurface defect detection model based on media inversion and deep learning according to claim 1, characterized in that, In step S3, the deep learning dielectric property inversion model includes inversion models based on waveform propagation characteristics, typically using full waveform inversion and reverse time offset; inversion models based on time-frequency domain signal processing, typically using short-time Fourier transform or wavelet transform; and inversion models based on deep learning, typically using convolutional neural networks (CNN), recurrent neural networks (RNN), and generative adversarial networks (GAN).

7. The road subsurface defect detection model based on media inversion and deep learning according to claim 1, characterized in that, In step S5, a disease target recognition model is constructed and trained using a dielectric constant distribution map dataset as input to establish the relationship between the dielectric constant distribution map and the classification and recognition of road underground diseases. The disease target recognition model uses the YOLOv5s target detection framework as its basic model. In the YOLOv5s backbone network, the Dense-C3 module is used to process feature information of complex backgrounds, and the CBAM attention module is used to enable the model to adaptively focus on regions containing key information. During the training process of the disease target recognition model, a multi-task loss function is used, which includes location loss, confidence loss, and classification loss.

8. The road subsurface defect detection model based on media inversion and deep learning according to claim 7, characterized in that, The location loss uses GIoU loss, while the confidence loss and classification loss use the Focal Loss function.

9. The road subsurface defect detection model based on media inversion and deep learning according to claim 1, characterized in that, In step S5, the disease target identification model includes models based on commonly used target detection algorithms such as CNN, Faster-RCNN, YOLO, SSD, or GAN.

10. The road subsurface defect detection model based on media inversion and deep learning according to claim 1, characterized in that, The method for detecting underground road defects is as follows: 1) Acquire radar signal data to be processed; 2) Use a trained deep learning dielectric property inversion model to invert the radar signal data to be processed into a dielectric constant distribution map; 3) Input the dielectric constant distribution map into the trained defect target recognition model and output the prediction results of underground hidden defects, thereby realizing the target detection and classification of underground road defects.

Images