Road disease 3D data set construction and data enhancement method and system

Abstract

The invention provides a road disease 3D data set construction and data enhancement method and system, and the method comprises the steps: obtaining the internal waveform data of a target road, and generating radar waveform images of a longitudinal section and a horizontal plane; preprocessing the images of the two cross sections, storing the slice images of different cross sections in different files, and sorting the slice images of each cross section according to a sequence; according to the slice image of each section, determining a disease waveform position, judging a disease category, adding a label, generating a model by using training data, further generating three-dimensional image data of each disease category of a vertical section and a horizontal plane, and adding three-dimensional radar echo data to form a final three-dimensional ground penetrating radar disease data set. According to the method, the problems of difficulty in making three-dimensional disease data, tedious construction process and low data set construction efficiency at present are solved, and the problem of insufficient deep learning model training caused by too few three-dimensional road disease samples is effectively solved.

Description

Technical Field

[0001] The present invention belongs to the field of data enhancement, and in particular relates to a road damage 3D data set construction and data enhancement method and system. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, with the country's continuous investment in basic transportation facilities, urban roads have continued to develop, and transportation infrastructure has also entered a stage of large-scale maintenance. Invisible diseases such as cracks, voids, and looseness generated during the use of urban roads, bridges and other infrastructure are sudden, hidden, and destructive, and cannot be observed and located artificially.

[0004] As a new type of non-destructive road testing technology, 3D ground penetrating radar (3DGPR) has gradually become an important tool for road disease detection. With the widespread application of artificial intelligence, deep learning is used to identify the echo images of 3D ground penetrating radar to realize the intelligent recognition of road internal diseases. However, the intelligent recognition model requires a large number of learning samples for model training. Currently, there are few public 3D disease datasets, and there is a lack of detailed and convenient dataset construction methods.

[0005] On the other hand, most studies on 3D GPR defect recognition have the problem of insufficient defect data samples, which leads to insufficient model training and low recognition accuracy of the model for voids. The common data enhancement method for radar defect waveforms is to physically transform the defect image or use forward modeling to generate simulated image data. However, traditional methods of physical changes include rotation, scaling, flipping, Gaussian noise addition, and elastic transformation of image data. Although these methods are simple and effective, they have some limitations. They cannot add nonlinear and hierarchical features for classification, and they do not have the underlying features of image data. The forward modeling method of generating images can generate data similar to real image features, but the generated images are mostly too ideal and cannot be fully simulated on the background clutter noise.

[0006] Therefore, a data enhancement method for 3D ground penetrating radar is urgently needed to meet the training requirements of intelligent recognition models. Summary of the invention

[0007] In order to solve the above problems, the present invention proposes a method and system for constructing and enhancing a 3D road damage dataset. The present invention solves the current problems of difficulty in producing three-dimensional damage data, cumbersome construction process and low efficiency of dataset construction, and effectively solves the problem of insufficient deep learning model training due to too few three-dimensional road damage samples.

[0008] According to some embodiments, the present invention adopts the following technical solutions:

[0009] A method for constructing a road damage 3D dataset and enhancing data includes the following steps:

[0010] Acquire the internal waveform data of the target road and generate radar waveform images of the longitudinal section and the horizontal plane respectively;

[0011] Preprocess the images of the two sections, and save the slice images of different sections in different files, and sort the slice images of each section in order;

[0012] According to the slice images of each section, the waveform position of the defect is determined, the defect category is judged and labeled, and the three-dimensional radar echo data of each defect category in the longitudinal section and horizontal plane are obtained;

[0013] Based on the three-dimensional radar echo data, a data generation model is trained, and the pre-trained data generation model is used to generate three-dimensional image data of various disease categories in longitudinal sections and horizontal planes, and the three-dimensional radar echo data is added to form a final three-dimensional ground penetrating radar disease data set.

[0014] As an optional implementation, the process of obtaining the internal waveform data of the target road includes using a three-dimensional ground penetrating radar device to collect the internal waveform data of the road on-site. The three-dimensional ground penetrating radar device moves on the ground for scanning, collects reflection signals at different positions, and generates radar waveform images of two sections, specifically including a B-scan image of the longitudinal section and a C-scan image of the horizontal plane.

[0015] As a further step, a slice image is acquired at a certain interval on the horizontal plane.

[0016] As an optional implementation, the process of preprocessing the images of the two cross sections includes noise removal, filtering and image gain.

[0017] As an optional implementation, all B-scan slice images of the longitudinal section are named and sorted according to the channel order and stored in a folder, and all C-scan slice images of the horizontal plane are named and sorted according to the position from top to bottom and stored in a folder.

[0018] As an optional implementation, the process of determining the location of the disease waveform, judging the disease category and adding labels according to the slice images of each section includes:

[0019] Crop and segment all images within the drawing box, and unify the size of the cropped sliced ​​images to make them square;

[0020] For the B-scan disease image of the longitudinal section, the slice images with unified size are interpolated in the depth direction to make the number of slices in the depth direction consistent with the image size after interpolation, and finally stacked in the depth direction to become the cubic three-dimensional voxel data of B-scan;

[0021] For the C-scan disease image on the horizontal plane, the slice image with unified size is vertically extracted with the same number of slices as the image size, and all the slices are stacked to form the cubic three-dimensional voxel data of C-scan;

[0022] All three-dimensional voxel data are classified and summarized. The B-scan data and C-scan data of the same disease waveform are placed in the same folder and defined as one disease data. All disease data of the same category are summarized into an overall folder, and finally the original data set containing various disease types is obtained.

[0023] As an optional embodiment, the types of defects include poor layer adhesion, voids, water abundance, voids and looseness.

[0024] As an optional implementation, the data generation model includes a two-branch EncoderCNN encoder, a two-branch self-attention generator and a discriminator connected in sequence.

[0025] As an optional implementation, the data generation model uses the B-scan and C-scan three-dimensional voxel data in the data set as input data. The model adopts a dual-branch architecture. The encoders of the two branches extract potential features from the real image data of the two sections respectively, and use an information fusion module to fuse the feature information extracted by the encoder and use it as the input of the generator. An ELSA self-attention module is added after sampling each layer of the generator, and the discriminator is used to extract features, and the output value is compressed between 0 and 1 for binary classification.

[0026] As an optional implementation, during the training process of the data generation model, the cosine annealing method is used to dynamically adjust the learning rate, and a feature consistency loss function is used to stabilize the training process by matching the statistical characteristics of the generated samples and the real samples at the feature layer, thereby ensuring that the generated image is similar to the target image in the feature space.

[0027] A road damage 3D dataset construction and data enhancement system, including:

[0028] A data acquisition module is configured to acquire internal waveform data of the target road and generate radar waveform images of the longitudinal section and the horizontal plane respectively;

[0029] A data preprocessing module is configured to preprocess the images of the two sections and save the slice images of different sections in different files, and the slice images of each section are sorted in order;

[0030] The disease judgment module is configured to determine the location of the disease waveform according to the slice images of each section, judge the disease category and add labels, and obtain the three-dimensional radar echo data of each disease category in the longitudinal section and the horizontal plane;

[0031] The data enhancement module is configured to train a data generation model based on the three-dimensional radar echo data, use the pre-trained data generation model to generate three-dimensional image data of various disease categories in the longitudinal section and the horizontal plane, and add the three-dimensional radar echo data to form a final three-dimensional ground penetrating radar disease data set.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present invention provides a feasible method for constructing a three-dimensional data set for road damage, solving the current problems of difficulty in producing three-dimensional damage data, cumbersome construction process and low efficiency of data set construction. At the same time, it provides a feasible three-dimensional damage data enhancement method, which effectively solves the problem of insufficient deep learning model training due to too few three-dimensional road damage samples.

[0034] The present invention can effectively expand the three-dimensional ground penetrating radar road disease data samples, thereby more fully training the disease recognition model and improving the accuracy and efficiency of disease recognition.

[0035] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0037] Figure 1 is a structural diagram of a defect waveform slice processing system of an embodiment;

[0038] Figure 2 is a flowchart of processing B-scan data in a disease waveform slice processing system of an embodiment;

[0039] Figure 3 is a flow chart of processing C-scan data in a disease waveform slice processing system of an embodiment;

[0040] Figure 4is an original three-dimensional ground penetrating radar disease dataset of an embodiment;

[0041] Figure 5 is a data enhancement model structure diagram of an embodiment;

[0042] Figure 6 The present invention is an embodiment of the structure composition of a ground penetrating radar three-dimensional disease data set for deep learning classification and identification. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0044] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0045] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0046] In the absence of conflict, the embodiments in this application and the features in the embodiments may be combined with each other.

[0047] Embodiment 1

[0048] A three-dimensional disease data enhancement method comprises the following steps:

[0049] S1: Data acquisition. In this embodiment, the specific steps are as follows: use the vehicle-mounted three-dimensional ground penetrating radar equipment to travel along the lateral line direction. The antenna of the ground penetrating radar emits high-frequency electromagnetic waves of 200-600Mhz. When the electromagnetic waves encounter different media (such as soil, rock, water, metal, etc.) during the propagation process, reflections will be generated. The receiving antenna receives the electromagnetic wave signals reflected from different depths underground and different material interfaces and performs signal processing to generate longitudinal section images (B-scan) and horizontal slice images (C-scan). The scanning width of the vehicle-mounted three-dimensional ground penetrating radar equipment is 1.58m, and the number of channels is 15 channels. The antenna array equipped with the three-dimensional ground penetrating radar equipment enables it to collect slice image data of 15 channels at the same time. Therefore, slice images equivalent to the number of channels of the three-dimensional ground penetrating radar equipment can be obtained on the total section, and a slice image can be obtained every 1cm on the horizontal plane. In this embodiment, the road type where the disease data is collected is a four-lane municipal road, and the length of the collected section is 20km.

[0050] S2: Data processing, pre-processing the double-section slice images obtained in step S1: to improve the image quality and achieve the purpose of enhancing the underground disease signal.

[0051] The specific steps are as follows:

[0052] (1) Noise reduction processing: Noise reduction processing is to eliminate the noise of radar echo signals and enhance signal quality. Gaussian filtering uses a Gaussian function (normal distribution) as the convolution kernel to perform convolution operations on the image to achieve a smoothing effect. The formula is as follows:

[0053]

[0054] Where: (x, y) is the coordinate of the filter, σ is the standard deviation of the Gaussian function, which determines the width of the filter.

[0055] (2) Filtering: The purpose is to highlight specific frequency components in the image to enhance the edge or smooth the image. The edge and detail features of the defect waveform are enhanced by a high-pass filter. The Laplace filter high-pass filter is selected for filtering, and the Laplace filter kernel is convolved with the 3D GPR data.

[0056] Assume that the original 3D-GPR data is I(x, y) and the filtered data is I′(x, y), then the convolution formula is:

[0057]

[0058] L(i, j) is the weight of the Laplace kernel.

[0059] (3) Gain processing: Gain processing can compensate for signal attenuation, improve the signal-to-noise ratio, and increase the visibility of deep targets. The gain function is applied to the echo signal through multiplication. Assuming that the original echo signal is S(t), the signal after gain adjustment is S′(t), then the gain adjustment can be expressed as:

[0060] S′(t)=G(t)·S(t);

[0061] Where G(t) is the gain function that varies with time.

[0062] After image preprocessing, all B-scan slice images of the longitudinal section are named and sorted according to the channel order and stored in a folder, and all C-scan slice images of the horizontal plane are named and sorted according to the position from top to bottom and stored in a folder. The data format is defined as:

[0063] Dataset·B=[B1, B2, B3···B 15 ];

[0064] Dataset·C=[C1, C2, C3···C n ];

[0065] S3: Import the defect waveform slice processing system. The present invention also proposes a defect waveform slice processing system for displaying and processing the echo image generated by the three-dimensional ground penetrating radar. The B-scan data folder and the C-scan data folder in step S2: are imported into the defect waveform processing system. The defect waveform system is an image processing software installed on a computer, and its structure is as follows: Figure 1 The specific steps are as follows:

[0066] (1) Import all slice images, including B-scan slice images of longitudinal sections and C-scan slice images of horizontal sections, of which there are 15 B-scan slice images and 120 C-scan slice images. The disease waveform slice processing system displays all slices of B / C-scan in two windows respectively.

[0067] (2) The operator uses the mouse wheel to browse in different slice images, clicks and drags the mouse or slides the scroll bar to drag the slice image, searches for the defect waveform through the above operations, draws a rectangular frame with the mouse to mark the defect area and selects the defect label. Furthermore, the defect categories are mainly divided into five categories: poor layer adhesion, voids, water-rich, voids, and looseness. Therefore, after framing the defect area, the available defect labels are: poor layer adhesion, voids, water-rich, voids, and looseness.

[0068] (3) After the entire image is labeled, the system automatically crops and segments all images within the drawing frame, and unifies the size of the cropped slice images to make them square slices of 64×64 pixels, obtaining 15 B-scan slices containing the disease waveform and 120 C-scan slice images of the same disease at the same position.

[0069] (4) For the longitudinal section B-scan disease image, 15 slice images are interpolated into 64 slices in the depth direction using trilinear interpolation. The 64 slice images are then stacked in order to form the B-scan three-dimensional voxel data of the disease in the longitudinal section direction. Furthermore, the principle of trilinear interpolation is: to estimate the value of an unknown point by linearly interpolating eight known points in a three-dimensional grid. The processing flow of the disease waveform slice processing system for B-scan slices is as follows: Figure 2 shown.

[0070] (5) For the C-scan disease images on the horizontal plane, 64 images containing the disease waveform are extracted vertically from the 120 slice images, and all slices are stacked to form the C-scan three-dimensional voxel data of the disease on the horizontal plane. The processing flow of the disease waveform slice processing system for C-scan slices is as follows: Figure 3 shown.

[0071] (6) All three-dimensional voxel data are classified and summarized. The B-scan data and C-scan data of the same defect waveform are placed in the same folder and defined as one defect data. All defect data of the same category are summarized into a whole folder. Finally, the original data set containing five types of defects is obtained. The structure of the original data set is as follows: Figure 4 shown.

[0072] S4: Build a data enhancement model. The data enhancement model architecture is as follows Figure 5 As shown in the figure, it consists of three modules: 1. Dual-branch EncoderCNN encoder, 2. Dual-branch self-attention generator. 3. Discriminator.

[0073] In this embodiment, the EncoderCNN in module 1 is designed to extract features from the input real image and generate a random vector with features for use as the input of the generator, thereby improving the quality and accuracy of the three-dimensional voxel data generated by the generator. It consists of four layers of three-dimensional convolution blocks, where the convolution kernel size Kernel_size = 4, the step length stride = 2, and the padding padding = 1. The number of filters in each layer is 32, 64, 128, and 256, respectively. After four convolutional layers, the size of the feature map is gradually reduced from 64x64x64 to 4x4x4, and the number of channels is increased to 256. The 256x4x4x4 feature map is then flattened into a 16384-dimensional one-dimensional vector through a flattening operation. Finally, the high-dimensional feature vector is converted into a 100-dimensional random vector through a fully connected layer as the input of the generator in module 2. The final output 100-dimensional random vector contains feature information extracted from the input three-dimensional voxel data, which can effectively improve the generation quality of the generative adversarial network.

[0074] The dual-branch self-attention generator in module 2 uses a dual-branch architecture to generate three-dimensional defect waveform image pseudo data of B-scan and C-scan through two generators with self-attention modules. The generator of each branch consists of four layers of upsampling convolution blocks, and the convolution blocks of each layer of the generator consist of a three-dimensional deconvolution, a three-dimensional batch normalization, and a ReLu activation function. Furthermore, each layer of self-attention convolution block is followed by an ELSA self-attention module, which can more effectively extract local detail features while reducing the amount of calculation and memory consumption.

[0075] The ELSA self-attention module can better extract local detail features. Its structure is as follows Figure 5 As shown in the figure, 3D convolution is first used instead of linear layer to generate query (Q), key (K) and value (V) vectors. Then Hadamard product is introduced to replace the traditional dot product to calculate the attention score. The Hadamard product multiplies each corresponding position of the matrix instead of calculating the inner product of the entire matrix.

[0076] Specifically, the ELSA attention mechanism can be expressed as follows:

[0077] Attention(Q,K,V)=Softmax(f(Q⊙K))V;

[0078] Among them, Attention represents the attention score, f(x) represents the 3D convolution operation on the feature map obtained by the upsampling convolution block, and ⊙ represents the Hadamard product.

[0079] The discriminator in module 3 of this embodiment is composed of four layers of downsampling convolution blocks, which extract features from 64x64x64 three-dimensional voxel data, and finally use the sigmoid activation function to compress the output value between 0 and 1 for binary classification. The data enhancement model structure is as follows Figure 5 shown.

[0080] S5: Model training and evaluation. The original data set containing five types of diseases produced by the disease waveform slicing processing system in step S3 is input into the designed data augmentation model and trained. During the training process, the cosine annealing method is used to dynamically adjust the learning rate, and the feature consistency loss function (Feature Consistency Loss) is used to stabilize the training process by matching the statistical characteristics of the generated samples and the real samples on the feature layer in the discriminator to ensure that the generated image is similar to the target image in the feature space, rather than in the pixel space. The formula of the feature consistency loss function is as follows:

[0081]

[0082] in Represents the distribution of real data p data The expectation of the samples sampled in the discriminator feature space. Denotes the noise distribution p input from the generator z The sample sampled in is the expectation of the sample generated by the generator G(z) in the discriminator feature space.

[0083] The loss function of the generator consists of adversarial loss and feature consistency loss, which can be expressed as:

[0084] LG =L adv +λL FM ;

[0085] Where λ is a weight coefficient used to balance the two losses.

[0086] After training for 200 epochs, the data generated by the model is evaluated. Since the main focus is on the similarity between features rather than the similarity between images in the evaluation index, the FID index is used.

[0087] S6: Dataset enhancement. The trained data enhancement model is used to generate three-dimensional voxel data of five types of disease waveforms to achieve data expansion effect on the original data set. The data set generated by the data enhancement model is defined as the GAN data set, and the data collected on the spot using the three-dimensional ground penetrating radar equipment is defined as the original data set. The original data set and the GAN data set are integrated in a ratio of 6:4 to obtain the final ground penetrating radar three-dimensional disease data set for deep learning classification and recognition. The final data set structure is as follows: Figure 6 shown.

[0088] Embodiment 2

[0089] A system corresponding to the first embodiment is provided, including:

[0090] A data acquisition module is configured to acquire internal waveform data of the target road and generate radar waveform images of the longitudinal section and the horizontal plane respectively;

[0091] A data preprocessing module is configured to preprocess the images of the two sections and save the slice images of different sections in different files, and the slice images of each section are sorted in order;

[0092] The disease judgment module is configured to determine the location of the disease waveform according to the slice images of each section, judge the disease category and add labels, and obtain the three-dimensional radar echo data of each disease category in the longitudinal section and the horizontal plane;

[0093] The data enhancement module is configured to train a data generation model based on the three-dimensional radar echo data, use the pre-trained data generation model to generate three-dimensional image data of various disease categories in the longitudinal section and the horizontal plane, and add the three-dimensional radar echo data to form a final three-dimensional ground penetrating radar disease data set.

[0094] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0096] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0097] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principle of the present invention without creative labor shall be included in the protection scope of the present invention.

Claims

1. A method for constructing a 3D road damage dataset and enhancing data, characterized in that: The following steps are involved: Acquire the internal waveform data of the target road and generate radar waveform images of the longitudinal section and the horizontal plane respectively; Preprocess the images of the two sections, and save the slice images of different sections in different files, and sort the slice images of each section in order; According to the slice images of each section, the waveform position of the defect is determined, the defect category is judged and labeled, and the three-dimensional radar echo data of each defect category in the longitudinal section and horizontal plane are obtained; Based on the three-dimensional radar echo data, a data generation model is trained, and the pre-trained data generation model is used to generate three-dimensional image data of various disease categories in longitudinal sections and horizontal planes, and the three-dimensional radar echo data is added to form a final three-dimensional ground penetrating radar disease data set.

2. A method for constructing and enhancing a 3D road damage dataset as claimed in claim 1, characterized in that: The process of obtaining the internal waveform data of the target road includes using a three-dimensional ground-penetrating radar device to collect the internal waveform data of the road on-site. The three-dimensional ground-penetrating radar device moves on the ground for scanning, collects reflection signals at different positions, and generates radar waveform images of two sections, including a B-scan image of the longitudinal section and a C-scan image of the horizontal plane.

3. The method for constructing and enhancing a 3D road damage dataset as claimed in claim 2, wherein: A slice image is acquired at a certain interval on the horizontal plane.

4. The method for constructing and enhancing a 3D road damage dataset according to claim 1, wherein: The process of preprocessing the images of the two sections includes noise removal, filtering and image gain.

5. The method for constructing and enhancing a 3D road damage dataset according to claim 1, wherein: All B-scan slice images of the longitudinal section are named and sorted according to the channel order and stored in a folder, and all C-scan slice images of the horizontal plane are named and sorted according to the position from top to bottom and stored in a folder.

6. The method for constructing and enhancing a 3D road damage dataset as claimed in claim 1, characterized in that: The process of determining the location of the disease waveform, judging the disease category and adding labels based on the slice images of each section includes: Crop and segment all images within the drawing box, and unify the size of the cropped sliced ​​images to make them square; For the B-scan disease image of the longitudinal section, the slice images with unified size are interpolated in the depth direction to make the number of slices in the depth direction consistent with the image size after interpolation, and finally stacked in the depth direction to become the cubic three-dimensional voxel data of B-scan; For the C-scan disease image on the horizontal plane, the slice image with unified size is vertically extracted with the same number of slices as the image size, and all the slices are stacked to form the cubic three-dimensional voxel data of C-scan; All three-dimensional voxel data are classified and summarized. The B-scan data and C-scan data of the same disease waveform are placed in the same folder and defined as one disease data. All disease data of the same category are summarized into an overall folder, and finally the original data set containing various disease types is obtained.

7. A method for constructing and enhancing a road damage 3D dataset as claimed in claim 1 or 6, characterized in that: The types of defects include poor layer bonding, voids, water abundance, voids and looseness.

8. The method for constructing and enhancing a 3D road damage dataset according to claim 1, wherein: The data generation model includes a two-branch EncoderCNN encoder, a two-branch self-attention generator and a discriminator connected in sequence; Alternatively, the data generation model uses the B-scan and C-scan three-dimensional voxel data in the data set as input data. The model adopts a dual-branch architecture. The encoders of the two branches extract potential features from the real image data of the two sections respectively, and use an information fusion module to fuse the feature information extracted by the encoder and use it as the input of the generator. After sampling each layer of the generator, an ELSA self-attention module is added, and the discriminator is used to extract features, and the output value is compressed between 0 and 1 for binary classification.

9. A method for constructing and enhancing a 3D road damage dataset as claimed in claim 1 or 8, characterized in that: During the training process of the data generation model, the cosine annealing method is used to dynamically adjust the learning rate, and the feature consistency loss function is used to stabilize the training process by matching the statistical characteristics of the generated samples and the real samples on the feature layer, ensuring that the generated image is similar to the target image in the feature space.

10. A road disease 3D data set construction and data enhancement system, characterized by: include: A data acquisition module is configured to acquire internal waveform data of the target road and generate radar waveform images of the longitudinal section and the horizontal plane respectively; A data preprocessing module is configured to preprocess the images of the two sections and save the slice images of different sections in different files, and the slice images of each section are sorted in order; The disease judgment module is configured to determine the location of the disease waveform according to the slice images of each section, judge the disease category and add labels, and obtain the three-dimensional radar echo data of each disease category in the longitudinal section and the horizontal plane; The data enhancement module is configured to train a data generation model based on the three-dimensional radar echo data, use the pre-trained data generation model to generate three-dimensional image data of various disease categories in the longitudinal section and the horizontal plane, and add the three-dimensional radar echo data to form a final three-dimensional ground penetrating radar disease data set.

Images