Ground penetrating radar training dataset simulation augmentation and road non-destructive testing method and system

Abstract

The present application relates to a kind of ground penetrating radar training data set simulation amplification and road nondestructive testing method and system, belong to ground penetrating radar detection technical field, first obtain multiple simulation parameters, simulation parameters include the number of structural layer in road, the structural layer information of each structural layer and the disease body information of disease body in road, then according to the number, position, thickness of structural layer and the size of disease body generation structural layer combination model, according to the composition of the material of structural layer and the proportion and dielectric constant of each component component in structural layer combination model generate each structural layer, according to disease body information in structural layer combination model generation disease body, obtain three-dimensional simulation model, finally each three-dimensional simulation model is scanned, and ground penetrating radar simulation image is obtained, so that various complex working conditions under ground penetrating radar simulation image can be randomly, large batch simulation generation, training data set is amplified, and road nondestructive testing precision is further improved.

Description

Technical Field

[0001] This invention relates to the field of ground penetrating radar (GPR) detection technology, and in particular to a method and system for simulating and amplifying training dataset for complex GPR operating conditions and performing non-destructive road inspection. Background Technology

[0002] Currently, the total length of highways in China has reached 5.2807 million kilometers, of which 5.2516 million kilometers are under maintenance, including 169,100 kilometers of expressways, greatly promoting economic and social development. However, with the increasing sophistication and saturation of the transportation network, road surface structures will inevitably suffer damage over time, and highway construction will gradually enter the maintenance phase. Faced with the maintenance work of such a vast road network, employing effective detection methods to quickly, efficiently, and accurately locate the location, type, and severity of surface and internal defects, and scientifically and rationally determine maintenance and repair plans, has significant economic and social value.

[0003] 3D ground-penetrating radar (GPR) is a representative technology in road non-destructive testing. It primarily obtains relevant information about the object being tested by analyzing the propagation of electromagnetic waves within it, including information on internal road defects such as cracks and voids. It is an effective means of detecting internal damage to road structures. However, in the interpretation of 3D GPR defect images, there is currently a lack of mature automated identification methods, with manual interpretation being the primary method. Manual interpretation of radar images has several problems, such as: the interpretation process requires a high level of expertise, and there is a shortage of sufficient data interpreters; the interpretation process is subjective, and the interpretation results of the same radar image often differ depending on the interpreter; manual interpretation is time-consuming, labor-intensive, and inefficient. These problems have, to some extent, limited the application and promotion of GPR technology.

[0004] To address the problem of automated interpretation of ground-penetrating radar (GPR) images, researchers have proposed numerous radar signal processing algorithms, such as traditional machine learning algorithms and deep learning methods. Traditional machine learning methods utilize machine vision technology to detect hyperbolic features in B-SCAN images. Commonly used algorithms include those based on Hough transform and those based on feature representation. Hough transform is an effective method for detecting and locating straight lines and analytical curves, but it has a large parameter space and high computational complexity. Feature representation-based methods, such as the Viola-Jones algorithm based on Haar-like wavelet features and hyperbolic feature detection algorithms combining gradient direction histograms and edge histogram descriptors, require manual feature design during application, and the accuracy of the detection results is not high.

[0005] In recent years, convolutional neural networks (CNNs) have emerged as a technological advancement, enabling them to learn from correctly labeled images and thus identify similar features in other unlabeled images. For example, the two-stage recognition method of GPR-RCNN can achieve high-precision identification of targets such as voids, pipelines, and subsidence. ResNet50 and YOLOv2 networks can detect the characteristic hyperbolic curves of water damage targets on asphalt pavements. These CNN models primarily learn from training samples to identify target features, significantly improving detection accuracy, but still heavily rely on the representativeness and size of the training dataset.

[0006] Convolutional neural network (CNN) models require at least several thousand labeled training images for various working conditions to achieve good training results. Furthermore, the larger the number of accurately labeled training images and the more diverse the working condition types, the higher the model's recognition accuracy, robustness, and generalization ability. However, due to the limited availability of ground-penetrating radar (GPR) data, the number of images for each working condition is often in the hundreds or even tens, and the number of training images for various working conditions is uneven; for example, the number of cracked defects is often far greater than that of voids and subsidence. Insufficient dataset size, lack of representativeness, and imbalanced sample size are among the main reasons why the accuracy of current CNN models in interpreting 3D GPR data fails to meet engineering requirements.

[0007] Therefore, there is an urgent need for a technology that can randomly and in large quantities simulate and generate ground-penetrating radar images under various complex working conditions. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for simulating and expanding ground-penetrating radar training datasets and for non-destructive road inspection. This method can randomly and in large batches generate simulated ground-penetrating radar images under various complex working conditions, thereby expanding the training dataset and further improving the accuracy of non-destructive road inspection.

[0009] To achieve the above objectives, the present invention provides the following solution:

[0010] A method for simulating augmentation of a ground-penetrating radar training dataset, the method comprising:

[0011] Multiple simulation parameters are obtained; the simulation parameters include the number of structural layers within the road, structural layer information for each structural layer, and disease information for road defects; the structural layer information includes the location, thickness, material composition, and proportion and dielectric constant of each component of the structural layer; the disease information includes the type, shape, location, and size of the disease.

[0012] For each simulation parameter, a structural layer combination model is generated based on the number, position, thickness, and size of the structural layers; each structural layer is generated within the structural layer combination model based on the material composition of the structural layers and the proportion and dielectric constant of each component; the disease is generated within the structural layer combination model based on the disease information, thus obtaining a three-dimensional simulation model.

[0013] Each of the three-dimensional simulation models is scanned to obtain a ground-penetrating radar simulation image.

[0014] A ground-penetrating radar training dataset simulation augmentation system, the simulation augmentation system comprising:

[0015] The simulation parameter acquisition module is used to acquire multiple simulation parameters; the simulation parameters include the number of structural layers within the road, structural layer information of each structural layer, and disease information of road defects; the structural layer information includes the location, thickness, material composition, and proportion and dielectric constant of each component of the structural layer; the disease information includes the type, shape, location, and size of the disease.

[0016] A three-dimensional simulation model generation module is used to generate a structural layer combination model for each simulation parameter based on the number, position, thickness, and size of the structural layers; generate each structural layer within the structural layer combination model based on the material composition of the structural layers and the proportion and dielectric constant of each component; and generate the disease within the structural layer combination model based on the disease information, thereby obtaining a three-dimensional simulation model.

[0017] The image generation module is used to scan each of the three-dimensional simulation models to obtain ground-penetrating radar simulation images.

[0018] A method for non-destructive testing of roads, the method comprising:

[0019] A training dataset is constructed; the training dataset includes multiple ground-penetrating radar images and a label corresponding to each ground-penetrating radar image; the label is the type, size, and location of the disease in the ground-penetrating radar image; the ground-penetrating radar images include ground-penetrating radar measured images collected in the field using ground-penetrating radar and ground-penetrating radar simulated images generated using the above-mentioned simulation amplification method;

[0020] Construct a convolutional neural network model;

[0021] The convolutional neural network model is trained using the training dataset to obtain a detection model;

[0022] The aforementioned detection model is used for non-destructive testing of roads.

[0023] A road non-destructive testing system, the road non-destructive testing system comprising:

[0024] A dataset construction module is used to construct a training dataset; the training dataset includes multiple ground-penetrating radar images and a label corresponding to each ground-penetrating radar image; the label is the type, size, and location of the disease in the ground-penetrating radar image; the ground-penetrating radar images include ground-penetrating radar measured images collected in the field using ground-penetrating radar and ground-penetrating radar simulated images generated using the above-mentioned simulation amplification method;

[0025] The model building module is used to build convolutional neural network models;

[0026] The training module is used to train the convolutional neural network model using the training dataset to obtain a detection model;

[0027] The non-destructive testing module is used to perform non-destructive testing of roads using the aforementioned testing model.

[0028] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0029] This invention provides a method and system for simulating and expanding a ground-penetrating radar (GPR) training dataset and performing non-destructive road inspection. First, multiple simulation parameters are acquired, including the number of structural layers within the road, structural layer information for each layer, and information on road defects. Then, a structural layer combination model is generated based on the number, location, thickness, and size of the structural layers and defects. Each structural layer is then generated within the combination model based on its material composition, proportion, and dielectric constant. Defects are generated within the combination model based on their information, resulting in a three-dimensional simulation model. Finally, each three-dimensional simulation model is scanned to obtain a GPR simulation image. This allows for the random and large-scale generation of GPR simulation images under various complex working conditions, expanding the training dataset and further improving the accuracy of non-destructive road inspection. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of the simulated amplification method provided in Embodiment 1 of the present invention;

[0032] Figure 2This is a schematic diagram of the structural layer combination model provided in Embodiment 1 of the present invention;

[0033] Figure 3 This is a schematic diagram of the model obtained after generating the structural layers according to Embodiment 1 of the present invention;

[0034] Figure 4 This is a schematic diagram of the three-dimensional simulation model provided in Embodiment 1 of the present invention;

[0035] Figure 5 This is a schematic diagram of a ground-penetrating radar simulated image provided in Embodiment 1 of the present invention;

[0036] Figure 6 This is a system block diagram of the simulated amplification system provided in Embodiment 2 of the present invention;

[0037] Figure 7 This is a flowchart of the road non-destructive testing method provided in Embodiment 3 of the present invention;

[0038] Figure 8 This is a schematic diagram of the loss function curve provided in Embodiment 3 of the present invention;

[0039] Figure 9 This is a system block diagram of the road non-destructive testing system provided in Embodiment 4 of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] The purpose of this invention is to provide a method and system for simulating and expanding ground-penetrating radar training datasets and for non-destructive road inspection. This method can randomly and in large batches generate simulated ground-penetrating radar images under various complex working conditions, thereby expanding the training dataset and further improving the accuracy of non-destructive road inspection.

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Example 1:

[0044] To address the issues of poor representativeness and insufficient quantity in training datasets for deep learning models, the dielectric constants of each component material in the road structural layers were measured using a Percometer instrument. Based on the proportions and dielectric constants of each component material, as well as the volume and location of the defects, a random amplification algorithm with pre-defined constraints was developed using MATLAB to generate a three-dimensional simulation model composed of the road structural layers and irregular defects. This model simulates actual engineering conditions. The generated three-dimensional simulation model is then saved as an HDF5 file and called by gprMax, a ground-penetrating radar electromagnetic wave simulation software based on finite-difference time domain (FDTD). This simulates the propagation model of electromagnetic waves in irregular defects, generating ground-penetrating radar simulated images. Defect signals in the ground-penetrating radar simulated images are manually labeled and combined with ground-penetrating radar measured images from actual working conditions to form a hybrid dataset. This dataset is then used to train a convolutional neural network model, thereby improving the recognition accuracy.

[0045] Specifically, this embodiment provides a method for simulating and expanding a training dataset for ground-penetrating radar under complex operating conditions, such as... Figure 1 As shown, the simulated amplification method includes:

[0046] S1: Obtain multiple simulation parameters; the simulation parameters include the number of structural layers within the road, structural layer information for each structural layer, and disease information for road defects; the structural layer information includes the location, thickness, material composition, and proportion and dielectric constant of each component of the structural layer; the disease information includes the type, shape, location, and size of the disease.

[0047] Different roads have different structural layers, and the method in this embodiment can be applied to any type of road. For example, the internal structural layers of a road may include an asphalt surface layer, a base layer, and a soil layer arranged sequentially from top to bottom. In this case, the number of internal structural layers is three, with the asphalt surface layer at the top, the base layer in the middle, and the soil layer at the bottom. Therefore, this embodiment can set the number, position, and thickness of the internal structural layers of the road according to actual needs or circumstances.

[0048] This embodiment collects the composition of the materials in each structural layer and the proportion of each component. If no relevant design data is available, relevant tests should be conducted. The method for obtaining the composition of the structural layer materials and the proportion of each component includes: determining whether design data for the structural layer exists, obtaining a first determination result; if the first determination result is yes, determining the composition of the structural layer materials and the proportion of each component based on the design data; if the first determination result is no, determining whether a standard measurement method for the structural layer materials exists, obtaining a second determination result; if the second determination result is yes, measuring the structural layer materials according to the standard measurement method to obtain the composition of the structural layer materials and the proportion of each component; if the second determination result is no, treating the structural layer as a single homogeneous material.

[0049] For asphalt pavement, based on the design data of the asphalt mixture, if the designed asphalt content is 4.5% and the designed void ratio is 4.5%, then the composition and proportion of the asphalt pavement material can be determined as aggregate:asphalt:air = 91:4.5:4.5. If there is no relevant design data to determine the composition and proportion of the asphalt pavement material, the void ratio, asphalt content, and aggregate ratio can be determined by referring to the "Test Procedures for Asphalt and Asphalt Mixtures in Highway Engineering" (JTG E20-2011). For other structural layers, if the relevant specifications provide methods for determining the composition of the structural layer, those methods should be followed; otherwise, the entire structural layer should be considered as a homogeneous material.

[0050] This embodiment uses a dielectric constant measuring instrument to determine the dielectric constant of each component of the road structure layer, thereby obtaining the material's dielectric constant parameters. The components of the road structure layer are generally measured using the capacitance method. One suitable instrument is the Percometer, which consists of a main unit and a sensor probe. The testing principle utilizes the capacitance formed by the inner and outer metal rings of the sensor probe. By comparing the capacitance changes under conditions where the material under test is the dielectric and under conditions where air is the dielectric, the relative dielectric constant of the material specimen is calculated using the following formula, thus obtaining the dielectric constant of the material under test.

[0051]

[0052] In the formula, ε is the dielectric constant of the material to be tested; C′ is the measurement capacitance under the condition that the material to be tested is the medium; and C0 is the measurement capacitance under the condition that air is the medium.

[0053] The Percometer instrument has a relative permittivity testing range of 1–32, an accuracy of ±0.1%, a detection frequency of 40–50 MHz, and an effective detection depth of 2–3 cm. Before testing, the top and bottom surfaces of the material specimen to be tested must be cut smooth and flat to ensure that the Percometer instrument's sensor probe can fit tightly against the specimen. During testing, the sensor probe is first suspended in the air to test the relative permittivity of air, calibrate the instrument's built-in parameters, and confirm that the instrument is working properly. Then, the sensor probe can be tightly bonded to the smooth surface of the material specimen to be tested, and the permittivity of each specimen is tested and recorded.

[0054] The disease information in this embodiment is determined according to actual needs, and the types of diseases include cracks, voids, and loosening.

[0055] S2: For each simulation parameter, generate a structural layer combination model based on the number, position, thickness, and size of the structural layers; generate each structural layer within the structural layer combination model based on the material composition of the structural layers and the proportion and dielectric constant of each component; generate the disease within the structural layer combination model based on the disease information, thus obtaining a three-dimensional simulation model.

[0056] In S2, generating a structural layer combination model based on the number, location, thickness, and size of the structural layers can include:

[0057] (1) Determine the height based on the number and thickness of the structural layers, and determine the length and width based on the size of the diseased body to generate a cuboid model;

[0058] The height of the cuboid model is determined based on the number and thickness of the structural layers. For a single structural layer, if the actual thickness is no more than 1 meter, its simulated thickness is set to be equal to the actual thickness of the structural layer; if the actual thickness is greater than 1 meter, its simulated thickness is set to 1 meter. This determines the simulated thickness of each structural layer, and the sum of the simulated thicknesses of all structural layers is the height of the cuboid model. The length and width of the cuboid model are determined based on the size of the diseased organism, specifically the length and width of the diseased organism. The length of the cuboid model is equal to 10 times the length of the diseased organism, and the width of the cuboid model is equal to 10 times the width of the diseased organism, to avoid the influence of the model's UPML boundary signal on the diseased organism signal.

[0059] Preferably, the minimum length, width, and height of the cuboid model are 1m to avoid the influence of the boundary on the electromagnetic wave echo of the diseased body. When the sum of the simulated thicknesses of each structural layer is less than 1m, the height of the cuboid model is set to 1m. At this time, the simulated thickness of each structural layer is increased proportionally (the ratio of the actual sum to 1m), and the subsequent simulated thickness is used as the new thickness of the structural layer.

[0060] (2) The cuboid model is divided into layers according to the position and thickness of the structural layers, and the space corresponding to each structural layer is determined to obtain the structural layer combination model.

[0061] Based on the spatial size of the cuboid model and the position and thickness of each structural layer, a regular structural layer combination model is generated layer by layer, such as... Figure 2 As shown, Figure 2 This is a structural layer combination model generated using a three-layer structure as an example.

[0062] In S2, generating each structural layer within the structural layer combination model based on the material composition of the structural layer, the proportion of each component, and the dielectric constant includes: for each structural layer, based on the material composition of the structural layer, the proportion of each component, and the dielectric constant, one component is used as a binder (also called a binder material), and the remaining components are used as particle materials; a particle model of each particle material is generated inside the space corresponding to the structural layer; after each particle material is filled in sequence, the gaps between the particle models in the space corresponding to the structural layer are filled with the binder material to generate the structural layer.

[0063] Specifically, dielectric constant is an important electromagnetic property of materials and the most important parameter for simulating ground-penetrating radar electromagnetic wave images. When constructing a 3D simulation model, it is necessary to record the dielectric constant of each component to facilitate subsequent scanning processes. Generally, the components used as bonding materials are determined according to the design documents; for road construction materials, the bonding materials are typically asphalt and cement. Generating a particle model for each material particle within the space corresponding to the structural layer can include: for each material particle, calculating the filling quantity based on the length and width of the structural layer combination model, the thickness of the structural layer to which the material belongs, and the proportion and radius of the material; randomly generating several spherical particles within the space corresponding to the structural layer, where the radius of each particle is the particle radius and the number of spherical particles is the filling quantity; determining whether the distance between the centers of any two spherical particles is less than the product of the particle radius and 2; if so, moving the positions of spherical particles whose center-to-center distance is less than the product of the particle radius and 2, until the distance between the centers of any two spherical particles is greater than or equal to the product of the particle radius and 2, thus generating the particle model of the material.

[0064] The formula for calculating the fill quantity is as follows:

[0065]

[0066] Where N is the filling quantity; l is the model length of the structural layer combination model; w is the model width of the structural layer combination model; h′ is the structural layer thickness of the structural layer to which the particle material belongs (this thickness is the simulated thickness of the structural layer); γ is the proportion of the particle material, i.e., the component ratio; φ is the particle radius, which is a preset value, obtained based on the particle size in actual working conditions. The specific method for obtaining it is as follows: first, determine the particle size based on the particle diameter of the structural layer constituent particles, and then determine the particle radius based on the particle size using the equal volume method.

[0067] In the structural layer combination model, several spherical particles are randomly generated. The distance between the centers of these particles is examined. If the distance is less than twice the particle radius, inter-particle interference is considered to exist. The centers of the interfering spherical particles are then randomly re-determined until interference is eliminated. If two types of particle materials are involved, interference between them must also be considered, ensuring that the distance between the centers of any two spherical particles is greater than or equal to twice the radius. When the spherical particles belong to different particle materials, twice the radius becomes the sum of the radii of one particle material and the radii of the other.

[0068] After filling each structural layer sequentially, the disease organisms are then filled into the model with the already filled structural layers based on the disease information, forming a three-dimensional simulation model, which can be completed in MATLAB. The generated three-dimensional simulation model data matrix file is saved as an HDF5 file, and the model is rendered using three-dimensional simulation software, allowing for a visual display of the three-dimensional simulation model.

[0069] As an example, this embodiment takes a road with asphalt surface layer, base course, and subgrade as its structural layers from top to bottom as an example to give a method for generating a three-dimensional simulation model:

[0070] The parameters of each structural layer are shown in Table 1.

[0071] Table 1

[0072]

[0073] The three-dimensional simulation model generated based on the above parameters is as follows: Figure 3 and Figure 4 As shown, the structural layers from top to bottom are asphalt surface layer, base layer, and subgrade. Large particles are coarse aggregate, small particles are air voids, and the rest are filled with bonding material. Irregular prismatic shapes are void defects.

[0074] In this embodiment, the simulation parameters obtained by S1 are used as the basic parameters for generating the structural layer combination model. A three-dimensional simulation model containing the disease body is randomly generated. By repeating the random model generation algorithm, a large number of three-dimensional simulation models that meet certain requirements can be generated in batches according to the set simulation parameters.

[0075] S3: Scan each of the three-dimensional simulation models to obtain a ground-penetrating radar simulation image.

[0076] The model parameters of the randomly generated 3D simulation model were imported into gprMax to construct the experimental model. The experimental model was surrounded by a 20-element PML boundary. Parameters such as the wave source, dominant frequency, polarization direction, transmit / receive antenna spacing, sampling interval, time window length, survey line position, and direction were set. The experimental model was then scanned to obtain simulated ground-penetrating radar images. Figure 4 The model shown uses a Ricker wavelet as the wave source, with a dominant frequency of 1 GHz, a polarization direction of z, and perpendicular to the survey line. The distance between the transmitting and receiving antennas is 0.1 m, the sampling interval is 0.06 m, the time window length is 30 ns, and the survey line is located in the center of the void defect. The survey line is scanned from left to right along the X direction. The resulting ground-penetrating radar simulated image is shown below. Figure 5 As shown in the image, following the steps above will generate a large number of simulated ground-penetrating radar images.

[0077] Three-dimensional ground-penetrating radar (GPR) data acquisition is costly, and the process of manually interpreting and labeling data with defect signals is inefficient, time-consuming, and labor-intensive. GPR signal simulation algorithms have emerged to address this issue, generating simulated GPR images. However, current GPR signal simulation algorithms primarily focus on simulating regular defects, lacking effective modeling methods for irregular defects. Furthermore, current GPR signal simulation algorithms mainly simulate the propagation of GPR signals in a homogeneous medium. Road construction materials are composite materials composed of multiple components with a degree of randomness; using a homogeneous medium to represent the composite material medium does not align with actual working conditions. The simulation amplification method provided in this embodiment addresses these problems. It proposes a method for determining the components and proportions of the road structure layer and uses a dielectric constant measuring instrument to determine the dielectric constant of each component material in the road structure layer, serving as a fundamental parameter for simulating the material properties of the structure layer. The simulation process can consider the influence of composite materials, making it more consistent with actual working conditions and capable of simulating irregular defects, thus significantly improving the applicability and accuracy of the simulation.

[0078] Example 2:

[0079] This embodiment provides a ground-penetrating radar training dataset simulation augmentation system, such as... Figure 6 As shown, the simulated amplification system includes:

[0080] The simulation parameter acquisition module M1 is used to acquire multiple simulation parameters; the simulation parameters include the number of structural layers in the road, structural layer information of each structural layer, and disease information of road defects; the structural layer information includes the location, thickness, material composition, and proportion and dielectric constant of each component of the structural layer; the disease information includes the type, shape, location, and size of the disease.

[0081] The three-dimensional simulation model generation module M2 is used to generate a structural layer combination model for each simulation parameter based on the number, position, thickness, and size of the structural layers; generate each structural layer within the structural layer combination model based on the material composition of the structural layers and the proportion and dielectric constant of each component; and generate the disease within the structural layer combination model based on the disease information, thereby obtaining a three-dimensional simulation model.

[0082] The image generation module M3 is used to scan each of the three-dimensional simulation models to obtain ground-penetrating radar simulation images.

[0083] Example 3:

[0084] This embodiment provides a method for non-destructive testing of roads, such as... Figure 7 As shown, the road non-destructive testing method includes:

[0085] T1: Construct a training dataset; the training dataset includes multiple ground-penetrating radar images and a label corresponding to each ground-penetrating radar image; the label is the type, size and location of the disease in the ground-penetrating radar image; the ground-penetrating radar images include ground-penetrating radar measured images collected in the field using ground-penetrating radar and ground-penetrating radar simulated images generated using the simulation amplification method described in Example 1;

[0086] In this embodiment, the Labelme tool can be used to annotate the ground penetrating radar image to obtain the label of the ground penetrating radar image.

[0087] T2: Construct a convolutional neural network model;

[0088] T3: Train the convolutional neural network model using the training dataset to obtain the detection model;

[0089] During training, random horizontal flipping, translation, and scaling are used to augment the training dataset. The optimizer can be Adam, with an initial learning rate of 10⁻⁵, a batch size of 3, and 200 training epochs. A mechanism for automatic learning rate reduction and early termination is used: monitoring the validation set loss, halving the learning rate if the validation set loss does not decrease within 6 epochs, and terminating training if the validation set loss still does not decrease after 10 epochs.

[0090] Due to the elongated and thin nature of the cracks, their pixels account for a relatively small proportion of the image. Therefore, in this embodiment, Dice Loss and Focal Loss were used as loss functions when training the convolutional neural network model using the training dataset.

[0091] Dice Loss, proposed in VNet, was initially used to address the problem of imbalanced samples in medical image segmentation, and is defined as follows:

[0092]

[0093] In the formula, Loss dice p is the Dice Loss value. i y represents the prediction result for the i-th pixel; i Let be the label of the i-th pixel; N is the number of pixels in the image.

[0094] Focal Loss is also used to address the problems of imbalanced training samples and varying sample difficulty. It is a variant of cross-entropy. The Focal Loss for a single pixel in semantic segmentation is defined as follows:

[0095] Loss focal =-(1-p t ) γ log(p t );

[0096] In the formula, Loss focal FocalLoss value; γ is a constant, which can take the value 2; p t It is the probability that the model predicts a positive sample, defined as follows:

[0097]

[0098] Where y is the label of the pixel, and its value is equal to the Boolean value of the sample. Positive samples (i.e., pixels predicted to be diseased) are 1, and other samples (i.e., pixels predicted not to be diseased) are 0.

[0099] The loss function used by the model is Loss. dice +Loss focal .

[0100] To improve the generalization ability and training efficiency of the trained model, this embodiment performs image standardization by centering the image using the mean. The centered data better conforms to the distribution pattern. The processing procedure is shown in the following formula:

[0101]

[0102] Where X′ is the standardized image; X is the original image; μ is the mean of the image; and σ is the standard deviation of the image.

[0103] T4: Use the aforementioned detection model to perform non-destructive road inspection.

[0104] Here, this embodiment provides a comparative experiment to demonstrate that the detection model trained in this embodiment has better detection accuracy:

[0105] The ground-penetrating radar (GPR) field acquisition dataset used in this embodiment contains 8000 images, each with a resolution of 604×604. Each image shows different types of radar defects, with the ratio of cracks, voids, and loosening defects being approximately 5:1:1. Simultaneously, a dataset of 6000 images with different types of defects was generated using a GPR simulation augmentation method, where the ratio of cracks, voids, and loosening defects is consistent with the GPR field acquisition dataset. All images were labeled with defect areas using the Labelme tool.

[0106] The training data was prepared according to three schemes. Scheme 1 used 8,000 images from the ground-penetrating radar (GPR) field acquisition dataset, randomly selecting 6,000 field-acquired images as the training and validation sets (8:2 ratio). The validation set was not used in training but was used to verify the training results in each round. The remaining 2,000 field-acquired images were used as the test set. Scheme 2 used only 6,000 simulated images generated by the GPR simulation augmentation method as the training and validation sets (8:2 ratio), and randomly selected 2,000 field-acquired images from the GPR field acquisition dataset as the test set. Scheme 3 used both the GPR field acquisition dataset and the dataset generated by the GPR simulation augmentation method, totaling 14,000 images. 2,000 field-acquired images were first selected as the test set, and the remaining 12,000 field-acquired and simulated images were randomly assigned to the training and validation sets (8:2 ratio). The test sets for schemes one through three are all extracted from ground-penetrating radar field-collected datasets, and the number of sets is the same. This is to ensure that the model performance is evaluated using actual ground-penetrating radar images, and that the evaluation criteria are consistent.

[0107] The training configuration used in the experiment was an NVIDIA 1080Ti GPU with 11GB of VRAM, running CenterOS 7. The deep learning framework used was Keras with TensorFlow-gpu as the backend, and the recognition model employed was the mainstream Unet model. During training, random horizontal flipping, translation, and scaling were used to augment the dataset. The optimizer chosen was Adam, with an initial learning rate of 10⁻⁵, a batch size of 3, and 200 training epochs. A mechanism for automatic learning rate reduction and early termination was implemented: monitoring the loss on the validation set, halving the learning rate if the validation loss did not decrease within 6 epochs, and terminating training if this continued for 10 epochs.

[0108] The descent curves of the training set loss and validation set loss for Scheme 1 are as follows: Figure 8 (a) and Figure 8 As shown in (b), the model converged at the 63rd epoch due to the automatic stopping strategy.

[0109] This embodiment uses standard semantic segmentation metrics such as Mean Pixel Accuracy (MPA), Mean Intersection over Union (MIoU), Floating-point Operations (FLOPs), and Faster Per Second (FPS) as evaluation metrics. MPA is the proportion of correctly classified pixels out of all pixels, and it can be used to evaluate the accuracy of pixel-level classification. Each pixel in the segmentation result corresponds to one of the following four categories:

[0110] (1) True Positive (TP): The model predicts that the example is positive and the label is also positive;

[0111] (2) False Positive (FP): The model predicts it as a positive example, but the label is a negative example;

[0112] (3) False Negative (FN): The model predicts it as a negative example, but the label is a positive example;

[0113] (4) True Negative (TN): The model predicts a negative example, and the label is also a negative example.

[0114] For a single category in semantic segmentation, the pixel accuracy (PA) is:

[0115]

[0116] MPA is the mean of the sum of PA for all categories. In addition to the target category to be segmented, semantic segmentation also has a default category called background. Therefore, the segmentation categories in this embodiment are crack and background.

[0117] IoU is the intersection of the labeled region and the predicted region divided by their union. That is, for labeled region A and predicted region B, IoU is expressed as:

[0118]

[0119] MIoU is the average of the sum of IoU across all categories. IoU assesses the degree of overlap between the segmented region and the expected label. FLOPs represent the number of floating-point operations performed by the neural network and can be used to evaluate the computational cost or complexity of the model. The FLOPs of a single convolutional layer are calculated as follows:

[0120] FLOPs = (2 × C in ×K 2 -1)×H×W×C out ;

[0121] In the formula, C in Where K is the number of input channels, K is the kernel size, H and W are the height and width of the output feature map, respectively, and C is the number of input channels. out This represents the number of output channels.

[0122] Comparing the three training dataset construction schemes, all test environments in this embodiment are the same as the training environment, and the results are shown in Table 2:

[0123] Table 2 Comparison of Indicators for Different Models

[0124]

[0125] As shown in Table 2, the model trained using a dataset combining simulated images generated by the ground-penetrating radar (GPR) simulation amplification method and measured images from actual operating conditions has higher MPA and MIoU than models trained solely using actual operating condition radar images or solely using simulated radar images. Specifically, the MPA is 7.39% and 13.17% higher than models trained using actual operating condition radar images, and the MIoU is 5.62% and 11.12% higher, respectively, all showing significant improvements. Therefore, the training dataset establishment method in this embodiment significantly improves the accuracy of the Unet model compared to the commonly used method of training with actual operating condition radar datasets.

[0126] This embodiment proposes to combine images acquired by ground penetrating radar in the field and simulated images generated by ground penetrating radar complex working condition simulation amplification algorithm in three ways to form different datasets, which are divided into training set, validation set and test set. The same model is trained on the same dataset. According to scientific and objective model accuracy measurement indicators, the improvement of model accuracy by the three dataset division methods is verified. Obviously, the hybrid dataset constructed in this embodiment can greatly improve the detection accuracy.

[0127] This embodiment proposes a ground-penetrating radar (GPR) data augmentation method for various complex working conditions. First, the road structure layer combination and the proportions of each component are determined. Then, the electromagnetic properties of each component are measured using dielectric constant measuring instruments such as the Percometer. Shape parameters of each structural layer and defect are set, and combined with the measured component proportions and dielectric constants, a structural layer combination model including the defect is generated. A large number of the generated models conform to the set parameters, but the particle distribution in the models exhibits randomness. The models are imported into gprMax to construct experimental models, and detection parameters are set to calculate simulated GPR images. Experimental results show that training the Unet artificial intelligence model using the dataset established by the above GPR data augmentation method, mixed with a real-world GPR training dataset, can significantly improve the model accuracy.

[0128] Example 4:

[0129] This embodiment provides a road non-destructive testing system, such as... Figure 9 As shown, the road non-destructive testing system includes:

[0130] The dataset construction module M4 is used to construct a training dataset; the training dataset includes multiple ground-penetrating radar images and a label corresponding to each ground-penetrating radar image; the label is the type, size and location of the disease in the ground-penetrating radar image; the ground-penetrating radar images include ground-penetrating radar measured images collected in the field using ground-penetrating radar and ground-penetrating radar simulated images generated using the simulation amplification method described in Example 1.

[0131] The model building module M5 is used to build convolutional neural network models;

[0132] The training module M6 is used to train the convolutional neural network model using the training dataset to obtain a detection model;

[0133] The non-destructive testing module M7 is used to perform non-destructive testing of roads using the aforementioned testing model.

[0134] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0135] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for simulating and expanding a ground-penetrating radar training dataset, characterized in that, The simulated amplification method includes: Multiple simulation parameters are obtained; the simulation parameters include the number of structural layers within the road, structural layer information for each structural layer, and disease information for road defects; the structural layer information includes the location, thickness, material composition, and proportion and dielectric constant of each component of the structural layer; the disease information includes the type, shape, location, and size of the disease. For each simulation parameter, a structural layer combination model is generated based on the number, position, thickness, and size of the structural layers; each structural layer is generated within the structural layer combination model based on the material composition of the structural layers and the proportion and dielectric constant of each component; the disease is generated within the structural layer combination model based on the disease information, thus obtaining a three-dimensional simulation model. Each of the three-dimensional simulation models is scanned to obtain a ground-penetrating radar simulation image. Specifically, the three-dimensional simulation model is called by the ground-penetrating radar electromagnetic wave simulation software based on the finite-difference time domain to simulate the propagation of electromagnetic waves and generate a ground-penetrating radar simulation image. Generating a structural layer combination model based on the number, location, thickness, and size of the structural layers specifically includes: The height is determined based on the number and thickness of the structural layers, and the length and width are determined based on the size of the diseased organism, thus generating a cuboid model. The cuboid model is divided into layers according to the position and thickness of the structural layers, and the space corresponding to each structural layer is determined to obtain a structural layer combination model. Generating each structural layer within the structural layer assembly model based on the material composition of the structural layer and the proportion and dielectric constant of each component specifically includes: For each of the structural layers, based on the composition of the material of the structural layer and the proportion and dielectric constant of each of the components, one component is used as a binder and the remaining components are used as particulate materials. A particle model of each of the particle materials is generated inside the space corresponding to the structural layer, and the gaps between the particle models in the space corresponding to the structural layer are filled with the adhesive material to generate the structural layer. The dielectric constant of each component of the structural layer is used as the basic parameter for simulating the material properties of the structural layer. The influence of the combined materials is considered in the simulation process, which is more consistent with the actual working conditions. Generating a particle model for each of the particle materials within the space corresponding to the structural layer specifically includes: For each particle material, the filling quantity of the particle material is calculated based on the length and width of the structural layer combination model, the thickness of the structural layer to which the particle material belongs, and the proportion and particle radius of the particle material. Several spherical particles are randomly generated within the space corresponding to the structural layer. The radius of each spherical particle is the particle radius, and the number of spherical particles is the filling quantity. Determine whether the distance between the centers of any two spherical particles is less than the product of the particle radius and 2; If so, move the positions of spherical particles whose center-to-center distance is less than the product of the particle radius and 2, until the center-to-center distance between any two spherical particles is greater than or equal to the product of the particle radius and 2, thereby generating a particle model of the particle material.

2. The simulated amplification method according to claim 1, characterized in that, The method for obtaining the composition of the material of the structural layer and the proportion of each of the components includes: Determine whether design data for the structural layer exists, and obtain a first determination result; If the first determination result is yes, then the composition of the material of the structural layer and the proportion of each of the components are determined according to the design data; If the first judgment result is negative, then determine whether there is a standard method for measuring the material of the structural layer, and obtain the second judgment result; If the second judgment result is yes, then the material of the structural layer is measured according to the standard measurement method to obtain the composition of the material of the structural layer and the proportion of each of the components.

3. The simulated amplification method according to claim 1, characterized in that, The minimum length, width, and height of the cuboid model are 1m.

4. A ground-penetrating radar training dataset simulation augmentation system, characterized in that, The simulated amplification system includes: The simulation parameter acquisition module is used to acquire multiple simulation parameters; the simulation parameters include the number of structural layers within the road, structural layer information of each structural layer, and disease information of road defects; the structural layer information includes the location, thickness, material composition, and proportion and dielectric constant of each component of the structural layer; the disease information includes the type, shape, location, and size of the disease. A three-dimensional simulation model generation module is used to generate a structural layer combination model for each simulation parameter based on the number, position, thickness, and size of the structural layers; generate each structural layer within the structural layer combination model based on the material composition of the structural layers and the proportion and dielectric constant of each component; and generate the disease within the structural layer combination model based on the disease information, thereby obtaining a three-dimensional simulation model. The image generation module is used to scan each of the three-dimensional simulation models to obtain ground-penetrating radar simulation images. Specifically, the three-dimensional simulation model is called by the ground-penetrating radar electromagnetic wave simulation software based on the finite-difference time domain to simulate the propagation of electromagnetic waves and generate ground-penetrating radar simulation images. Generating a structural layer combination model based on the number, location, thickness, and size of the structural layers specifically includes: The height is determined based on the number and thickness of the structural layers, and the length and width are determined based on the size of the diseased organism, thus generating a cuboid model. The cuboid model is divided into layers according to the position and thickness of the structural layers, and the space corresponding to each structural layer is determined to obtain a structural layer combination model. Generating each structural layer within the structural layer assembly model based on the material composition of the structural layer and the proportion and dielectric constant of each component specifically includes: For each of the structural layers, based on the composition of the material of the structural layer and the proportion and dielectric constant of each of the components, one component is used as a binder and the remaining components are used as particulate materials. A particle model of each of the particle materials is generated inside the space corresponding to the structural layer, and the gaps between the particle models in the space corresponding to the structural layer are filled with the adhesive material to generate the structural layer. The dielectric constant of each component of the structural layer is used as the basic parameter for simulating the material properties of the structural layer. The influence of the combined materials is considered in the simulation process, which is more consistent with the actual working conditions. Generating a particle model for each of the particle materials within the space corresponding to the structural layer specifically includes: For each particle material, the filling quantity of the particle material is calculated based on the length and width of the structural layer combination model, the thickness of the structural layer to which the particle material belongs, and the proportion and particle radius of the particle material. Several spherical particles are randomly generated within the space corresponding to the structural layer. The radius of each spherical particle is the particle radius, and the number of spherical particles is the filling quantity. Determine whether the distance between the centers of any two spherical particles is less than the product of the particle radius and 2; If so, move the positions of spherical particles whose center-to-center distance is less than the product of the particle radius and 2, until the center-to-center distance between any two spherical particles is greater than or equal to the product of the particle radius and 2, thereby generating a particle model of the particle material.

5. A method for non-destructive testing of roads, characterized in that, The road non-destructive testing method includes: Construct a training dataset; the training dataset includes multiple ground-penetrating radar images and a label corresponding to each ground-penetrating radar image; the label is the type, size and location of the disease in the ground-penetrating radar image; the ground-penetrating radar images include ground-penetrating radar measured images collected in the field using ground-penetrating radar and ground-penetrating radar simulated images generated using the simulation amplification method according to any one of claims 1-3. Construct a convolutional neural network model; The convolutional neural network model is trained using the training dataset to obtain a detection model; The aforementioned detection model is used for non-destructive testing of roads.

6. The road non-destructive testing method according to claim 5, characterized in that, When training the convolutional neural network model using the training dataset, the loss functions used include Dice Loss and Focal Loss.

7. A road non-destructive testing system, characterized in that, The road non-destructive testing system includes: A dataset construction module is used to construct a training dataset; the training dataset includes multiple ground-penetrating radar images and a label corresponding to each ground-penetrating radar image; the label is the type, size and location of the disease in the ground-penetrating radar image; the ground-penetrating radar images include ground-penetrating radar measured images collected in the field using ground-penetrating radar and ground-penetrating radar simulated images generated using the simulation amplification method according to any one of claims 1-3. The model building module is used to build convolutional neural network models; The training module is used to train the convolutional neural network model using the training dataset to obtain a detection model; The non-destructive testing module is used to perform non-destructive testing of roads using the aforementioned testing model.

Images